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>
93 #include <linux/mmu_notifier.h>
95 #include <asm/tlbflush.h>
96 #include <asm/uaccess.h>
97 #include <linux/random.h>
102 #define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
103 #define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
105 static struct kmem_cache *policy_cache;
106 static struct kmem_cache *sn_cache;
108 /* Highest zone. An specific allocation for a zone below that is not
110 enum zone_type policy_zone = 0;
113 * run-time system-wide default policy => local allocation
115 static struct mempolicy default_policy = {
116 .refcnt = ATOMIC_INIT(1), /* never free it */
117 .mode = MPOL_PREFERRED,
118 .flags = MPOL_F_LOCAL,
121 static const struct mempolicy_operations {
122 int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
124 * If read-side task has no lock to protect task->mempolicy, write-side
125 * task will rebind the task->mempolicy by two step. The first step is
126 * setting all the newly nodes, and the second step is cleaning all the
127 * disallowed nodes. In this way, we can avoid finding no node to alloc
129 * If we have a lock to protect task->mempolicy in read-side, we do
133 * MPOL_REBIND_ONCE - do rebind work at once
134 * MPOL_REBIND_STEP1 - set all the newly nodes
135 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
137 void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes,
138 enum mpol_rebind_step step);
139 } mpol_ops[MPOL_MAX];
141 /* Check that the nodemask contains at least one populated zone */
142 static int is_valid_nodemask(const nodemask_t *nodemask)
146 for_each_node_mask(nd, *nodemask) {
149 for (k = 0; k <= policy_zone; k++) {
150 z = &NODE_DATA(nd)->node_zones[k];
151 if (z->present_pages > 0)
159 static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
161 return pol->flags & MPOL_MODE_FLAGS;
164 static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
165 const nodemask_t *rel)
168 nodes_fold(tmp, *orig, nodes_weight(*rel));
169 nodes_onto(*ret, tmp, *rel);
172 static int mpol_new_interleave(struct mempolicy *pol, const nodemask_t *nodes)
174 if (nodes_empty(*nodes))
176 pol->v.nodes = *nodes;
180 static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
183 pol->flags |= MPOL_F_LOCAL; /* local allocation */
184 else if (nodes_empty(*nodes))
185 return -EINVAL; /* no allowed nodes */
187 pol->v.preferred_node = first_node(*nodes);
191 static int mpol_new_bind(struct mempolicy *pol, const nodemask_t *nodes)
193 if (!is_valid_nodemask(nodes))
195 pol->v.nodes = *nodes;
200 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
201 * any, for the new policy. mpol_new() has already validated the nodes
202 * parameter with respect to the policy mode and flags. But, we need to
203 * handle an empty nodemask with MPOL_PREFERRED here.
205 * Must be called holding task's alloc_lock to protect task's mems_allowed
206 * and mempolicy. May also be called holding the mmap_semaphore for write.
208 static int mpol_set_nodemask(struct mempolicy *pol,
209 const nodemask_t *nodes, struct nodemask_scratch *nsc)
213 /* if mode is MPOL_DEFAULT, pol is NULL. This is right. */
216 /* Check N_HIGH_MEMORY */
217 nodes_and(nsc->mask1,
218 cpuset_current_mems_allowed, node_states[N_HIGH_MEMORY]);
221 if (pol->mode == MPOL_PREFERRED && nodes_empty(*nodes))
222 nodes = NULL; /* explicit local allocation */
224 if (pol->flags & MPOL_F_RELATIVE_NODES)
225 mpol_relative_nodemask(&nsc->mask2, nodes,&nsc->mask1);
227 nodes_and(nsc->mask2, *nodes, nsc->mask1);
229 if (mpol_store_user_nodemask(pol))
230 pol->w.user_nodemask = *nodes;
232 pol->w.cpuset_mems_allowed =
233 cpuset_current_mems_allowed;
237 ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
239 ret = mpol_ops[pol->mode].create(pol, NULL);
244 * This function just creates a new policy, does some check and simple
245 * initialization. You must invoke mpol_set_nodemask() to set nodes.
247 static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
250 struct mempolicy *policy;
252 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
253 mode, flags, nodes ? nodes_addr(*nodes)[0] : -1);
255 if (mode == MPOL_DEFAULT || mode == MPOL_NOOP) {
256 if (nodes && !nodes_empty(*nodes))
257 return ERR_PTR(-EINVAL);
263 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
264 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
265 * All other modes require a valid pointer to a non-empty nodemask.
267 if (mode == MPOL_PREFERRED) {
268 if (nodes_empty(*nodes)) {
269 if (((flags & MPOL_F_STATIC_NODES) ||
270 (flags & MPOL_F_RELATIVE_NODES)))
271 return ERR_PTR(-EINVAL);
273 } else if (mode == MPOL_LOCAL) {
274 if (!nodes_empty(*nodes))
275 return ERR_PTR(-EINVAL);
276 mode = MPOL_PREFERRED;
277 } else if (nodes_empty(*nodes))
278 return ERR_PTR(-EINVAL);
279 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
281 return ERR_PTR(-ENOMEM);
282 atomic_set(&policy->refcnt, 1);
284 policy->flags = flags;
289 /* Slow path of a mpol destructor. */
290 void __mpol_put(struct mempolicy *p)
292 if (!atomic_dec_and_test(&p->refcnt))
294 kmem_cache_free(policy_cache, p);
297 static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes,
298 enum mpol_rebind_step step)
304 * MPOL_REBIND_ONCE - do rebind work at once
305 * MPOL_REBIND_STEP1 - set all the newly nodes
306 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
308 static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes,
309 enum mpol_rebind_step step)
313 if (pol->flags & MPOL_F_STATIC_NODES)
314 nodes_and(tmp, pol->w.user_nodemask, *nodes);
315 else if (pol->flags & MPOL_F_RELATIVE_NODES)
316 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
319 * if step == 1, we use ->w.cpuset_mems_allowed to cache the
322 if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP1) {
323 nodes_remap(tmp, pol->v.nodes,
324 pol->w.cpuset_mems_allowed, *nodes);
325 pol->w.cpuset_mems_allowed = step ? tmp : *nodes;
326 } else if (step == MPOL_REBIND_STEP2) {
327 tmp = pol->w.cpuset_mems_allowed;
328 pol->w.cpuset_mems_allowed = *nodes;
333 if (nodes_empty(tmp))
336 if (step == MPOL_REBIND_STEP1)
337 nodes_or(pol->v.nodes, pol->v.nodes, tmp);
338 else if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP2)
343 if (!node_isset(current->il_next, tmp)) {
344 current->il_next = next_node(current->il_next, tmp);
345 if (current->il_next >= MAX_NUMNODES)
346 current->il_next = first_node(tmp);
347 if (current->il_next >= MAX_NUMNODES)
348 current->il_next = numa_node_id();
352 static void mpol_rebind_preferred(struct mempolicy *pol,
353 const nodemask_t *nodes,
354 enum mpol_rebind_step step)
358 if (pol->flags & MPOL_F_STATIC_NODES) {
359 int node = first_node(pol->w.user_nodemask);
361 if (node_isset(node, *nodes)) {
362 pol->v.preferred_node = node;
363 pol->flags &= ~MPOL_F_LOCAL;
365 pol->flags |= MPOL_F_LOCAL;
366 } else if (pol->flags & MPOL_F_RELATIVE_NODES) {
367 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
368 pol->v.preferred_node = first_node(tmp);
369 } else if (!(pol->flags & MPOL_F_LOCAL)) {
370 pol->v.preferred_node = node_remap(pol->v.preferred_node,
371 pol->w.cpuset_mems_allowed,
373 pol->w.cpuset_mems_allowed = *nodes;
378 * mpol_rebind_policy - Migrate a policy to a different set of nodes
380 * If read-side task has no lock to protect task->mempolicy, write-side
381 * task will rebind the task->mempolicy by two step. The first step is
382 * setting all the newly nodes, and the second step is cleaning all the
383 * disallowed nodes. In this way, we can avoid finding no node to alloc
385 * If we have a lock to protect task->mempolicy in read-side, we do
389 * MPOL_REBIND_ONCE - do rebind work at once
390 * MPOL_REBIND_STEP1 - set all the newly nodes
391 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
393 static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask,
394 enum mpol_rebind_step step)
398 if (!mpol_store_user_nodemask(pol) && step == MPOL_REBIND_ONCE &&
399 nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
402 if (step == MPOL_REBIND_STEP1 && (pol->flags & MPOL_F_REBINDING))
405 if (step == MPOL_REBIND_STEP2 && !(pol->flags & MPOL_F_REBINDING))
408 if (step == MPOL_REBIND_STEP1)
409 pol->flags |= MPOL_F_REBINDING;
410 else if (step == MPOL_REBIND_STEP2)
411 pol->flags &= ~MPOL_F_REBINDING;
412 else if (step >= MPOL_REBIND_NSTEP)
415 mpol_ops[pol->mode].rebind(pol, newmask, step);
419 * Wrapper for mpol_rebind_policy() that just requires task
420 * pointer, and updates task mempolicy.
422 * Called with task's alloc_lock held.
425 void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new,
426 enum mpol_rebind_step step)
428 mpol_rebind_policy(tsk->mempolicy, new, step);
432 * Rebind each vma in mm to new nodemask.
434 * Call holding a reference to mm. Takes mm->mmap_sem during call.
437 void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
439 struct vm_area_struct *vma;
441 down_write(&mm->mmap_sem);
442 for (vma = mm->mmap; vma; vma = vma->vm_next)
443 mpol_rebind_policy(vma->vm_policy, new, MPOL_REBIND_ONCE);
444 up_write(&mm->mmap_sem);
447 static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
449 .rebind = mpol_rebind_default,
451 [MPOL_INTERLEAVE] = {
452 .create = mpol_new_interleave,
453 .rebind = mpol_rebind_nodemask,
456 .create = mpol_new_preferred,
457 .rebind = mpol_rebind_preferred,
460 .create = mpol_new_bind,
461 .rebind = mpol_rebind_nodemask,
465 static void migrate_page_add(struct page *page, struct list_head *pagelist,
466 unsigned long flags);
468 /* Scan through pages checking if pages follow certain conditions. */
469 static int check_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
470 unsigned long addr, unsigned long end,
471 const nodemask_t *nodes, unsigned long flags,
478 orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
483 if (!pte_present(*pte))
485 page = vm_normal_page(vma, addr, *pte);
489 * vm_normal_page() filters out zero pages, but there might
490 * still be PageReserved pages to skip, perhaps in a VDSO.
491 * And we cannot move PageKsm pages sensibly or safely yet.
493 if (PageReserved(page) || PageKsm(page))
495 nid = page_to_nid(page);
496 if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
499 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
500 migrate_page_add(page, private, flags);
503 } while (pte++, addr += PAGE_SIZE, addr != end);
504 pte_unmap_unlock(orig_pte, ptl);
508 static inline int check_pmd_range(struct vm_area_struct *vma, pud_t *pud,
509 unsigned long addr, unsigned long end,
510 const nodemask_t *nodes, unsigned long flags,
516 pmd = pmd_offset(pud, addr);
518 next = pmd_addr_end(addr, end);
519 split_huge_page_pmd(vma->vm_mm, pmd);
520 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
522 if (check_pte_range(vma, pmd, addr, next, nodes,
525 } while (pmd++, addr = next, addr != end);
529 static inline int check_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
530 unsigned long addr, unsigned long end,
531 const nodemask_t *nodes, unsigned long flags,
537 pud = pud_offset(pgd, addr);
539 next = pud_addr_end(addr, end);
540 if (pud_none_or_clear_bad(pud))
542 if (check_pmd_range(vma, pud, addr, next, nodes,
545 } while (pud++, addr = next, addr != end);
549 static inline int check_pgd_range(struct vm_area_struct *vma,
550 unsigned long addr, unsigned long end,
551 const nodemask_t *nodes, unsigned long flags,
557 pgd = pgd_offset(vma->vm_mm, addr);
559 next = pgd_addr_end(addr, end);
560 if (pgd_none_or_clear_bad(pgd))
562 if (check_pud_range(vma, pgd, addr, next, nodes,
565 } while (pgd++, addr = next, addr != end);
569 #ifdef CONFIG_ARCH_USES_NUMA_PROT_NONE
571 * This is used to mark a range of virtual addresses to be inaccessible.
572 * These are later cleared by a NUMA hinting fault. Depending on these
573 * faults, pages may be migrated for better NUMA placement.
575 * This is assuming that NUMA faults are handled using PROT_NONE. If
576 * an architecture makes a different choice, it will need further
577 * changes to the core.
579 unsigned long change_prot_numa(struct vm_area_struct *vma,
580 unsigned long addr, unsigned long end)
583 BUILD_BUG_ON(_PAGE_NUMA != _PAGE_PROTNONE);
585 nr_updated = change_protection(vma, addr, end, vma->vm_page_prot, 0, 1);
590 static unsigned long change_prot_numa(struct vm_area_struct *vma,
591 unsigned long addr, unsigned long end)
595 #endif /* CONFIG_ARCH_USES_NUMA_PROT_NONE */
598 * Check if all pages in a range are on a set of nodes.
599 * If pagelist != NULL then isolate pages from the LRU and
600 * put them on the pagelist.
602 static struct vm_area_struct *
603 check_range(struct mm_struct *mm, unsigned long start, unsigned long end,
604 const nodemask_t *nodes, unsigned long flags, void *private)
607 struct vm_area_struct *first, *vma, *prev;
610 first = find_vma(mm, start);
612 return ERR_PTR(-EFAULT);
614 for (vma = first; vma && vma->vm_start < end; vma = vma->vm_next) {
615 unsigned long endvma = vma->vm_end;
619 if (vma->vm_start > start)
620 start = vma->vm_start;
622 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
623 if (!vma->vm_next && vma->vm_end < end)
624 return ERR_PTR(-EFAULT);
625 if (prev && prev->vm_end < vma->vm_start)
626 return ERR_PTR(-EFAULT);
629 if (is_vm_hugetlb_page(vma))
632 if (flags & MPOL_MF_LAZY) {
633 change_prot_numa(vma, start, endvma);
637 if ((flags & MPOL_MF_STRICT) ||
638 ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) &&
639 vma_migratable(vma))) {
641 err = check_pgd_range(vma, start, endvma, nodes,
644 first = ERR_PTR(err);
655 * Apply policy to a single VMA
656 * This must be called with the mmap_sem held for writing.
658 static int vma_replace_policy(struct vm_area_struct *vma,
659 struct mempolicy *pol)
662 struct mempolicy *old;
663 struct mempolicy *new;
665 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
666 vma->vm_start, vma->vm_end, vma->vm_pgoff,
667 vma->vm_ops, vma->vm_file,
668 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
674 if (vma->vm_ops && vma->vm_ops->set_policy) {
675 err = vma->vm_ops->set_policy(vma, new);
680 old = vma->vm_policy;
681 vma->vm_policy = new; /* protected by mmap_sem */
690 /* Step 2: apply policy to a range and do splits. */
691 static int mbind_range(struct mm_struct *mm, unsigned long start,
692 unsigned long end, struct mempolicy *new_pol)
694 struct vm_area_struct *next;
695 struct vm_area_struct *prev;
696 struct vm_area_struct *vma;
699 unsigned long vmstart;
702 vma = find_vma(mm, start);
703 if (!vma || vma->vm_start > start)
707 if (start > vma->vm_start)
710 for (; vma && vma->vm_start < end; prev = vma, vma = next) {
712 vmstart = max(start, vma->vm_start);
713 vmend = min(end, vma->vm_end);
715 if (mpol_equal(vma_policy(vma), new_pol))
718 pgoff = vma->vm_pgoff +
719 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
720 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
721 vma->anon_vma, vma->vm_file, pgoff,
728 if (vma->vm_start != vmstart) {
729 err = split_vma(vma->vm_mm, vma, vmstart, 1);
733 if (vma->vm_end != vmend) {
734 err = split_vma(vma->vm_mm, vma, vmend, 0);
738 err = vma_replace_policy(vma, new_pol);
748 * Update task->flags PF_MEMPOLICY bit: set iff non-default
749 * mempolicy. Allows more rapid checking of this (combined perhaps
750 * with other PF_* flag bits) on memory allocation hot code paths.
752 * If called from outside this file, the task 'p' should -only- be
753 * a newly forked child not yet visible on the task list, because
754 * manipulating the task flags of a visible task is not safe.
756 * The above limitation is why this routine has the funny name
757 * mpol_fix_fork_child_flag().
759 * It is also safe to call this with a task pointer of current,
760 * which the static wrapper mpol_set_task_struct_flag() does,
761 * for use within this file.
764 void mpol_fix_fork_child_flag(struct task_struct *p)
767 p->flags |= PF_MEMPOLICY;
769 p->flags &= ~PF_MEMPOLICY;
772 static void mpol_set_task_struct_flag(void)
774 mpol_fix_fork_child_flag(current);
777 /* Set the process memory policy */
778 static long do_set_mempolicy(unsigned short mode, unsigned short flags,
781 struct mempolicy *new, *old;
782 struct mm_struct *mm = current->mm;
783 NODEMASK_SCRATCH(scratch);
789 new = mpol_new(mode, flags, nodes);
795 * prevent changing our mempolicy while show_numa_maps()
797 * Note: do_set_mempolicy() can be called at init time
801 down_write(&mm->mmap_sem);
803 ret = mpol_set_nodemask(new, nodes, scratch);
805 task_unlock(current);
807 up_write(&mm->mmap_sem);
811 old = current->mempolicy;
812 current->mempolicy = new;
813 mpol_set_task_struct_flag();
814 if (new && new->mode == MPOL_INTERLEAVE &&
815 nodes_weight(new->v.nodes))
816 current->il_next = first_node(new->v.nodes);
817 task_unlock(current);
819 up_write(&mm->mmap_sem);
824 NODEMASK_SCRATCH_FREE(scratch);
829 * Return nodemask for policy for get_mempolicy() query
831 * Called with task's alloc_lock held
833 static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
836 if (p == &default_policy)
842 case MPOL_INTERLEAVE:
846 if (!(p->flags & MPOL_F_LOCAL))
847 node_set(p->v.preferred_node, *nodes);
848 /* else return empty node mask for local allocation */
855 static int lookup_node(struct mm_struct *mm, unsigned long addr)
860 err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL);
862 err = page_to_nid(p);
868 /* Retrieve NUMA policy */
869 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
870 unsigned long addr, unsigned long flags)
873 struct mm_struct *mm = current->mm;
874 struct vm_area_struct *vma = NULL;
875 struct mempolicy *pol = current->mempolicy;
878 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
881 if (flags & MPOL_F_MEMS_ALLOWED) {
882 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
884 *policy = 0; /* just so it's initialized */
886 *nmask = cpuset_current_mems_allowed;
887 task_unlock(current);
891 if (flags & MPOL_F_ADDR) {
893 * Do NOT fall back to task policy if the
894 * vma/shared policy at addr is NULL. We
895 * want to return MPOL_DEFAULT in this case.
897 down_read(&mm->mmap_sem);
898 vma = find_vma_intersection(mm, addr, addr+1);
900 up_read(&mm->mmap_sem);
903 if (vma->vm_ops && vma->vm_ops->get_policy)
904 pol = vma->vm_ops->get_policy(vma, addr);
906 pol = vma->vm_policy;
911 pol = &default_policy; /* indicates default behavior */
913 if (flags & MPOL_F_NODE) {
914 if (flags & MPOL_F_ADDR) {
915 err = lookup_node(mm, addr);
919 } else if (pol == current->mempolicy &&
920 pol->mode == MPOL_INTERLEAVE) {
921 *policy = current->il_next;
927 *policy = pol == &default_policy ? MPOL_DEFAULT :
930 * Internal mempolicy flags must be masked off before exposing
931 * the policy to userspace.
933 *policy |= (pol->flags & MPOL_MODE_FLAGS);
937 up_read(¤t->mm->mmap_sem);
943 if (mpol_store_user_nodemask(pol)) {
944 *nmask = pol->w.user_nodemask;
947 get_policy_nodemask(pol, nmask);
948 task_unlock(current);
955 up_read(¤t->mm->mmap_sem);
959 #ifdef CONFIG_MIGRATION
963 static void migrate_page_add(struct page *page, struct list_head *pagelist,
967 * Avoid migrating a page that is shared with others.
969 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) {
970 if (!isolate_lru_page(page)) {
971 list_add_tail(&page->lru, pagelist);
972 inc_zone_page_state(page, NR_ISOLATED_ANON +
973 page_is_file_cache(page));
978 static struct page *new_node_page(struct page *page, unsigned long node, int **x)
980 return alloc_pages_exact_node(node, GFP_HIGHUSER_MOVABLE, 0);
984 * Migrate pages from one node to a target node.
985 * Returns error or the number of pages not migrated.
987 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
995 node_set(source, nmask);
998 * This does not "check" the range but isolates all pages that
999 * need migration. Between passing in the full user address
1000 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1002 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
1003 check_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
1004 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
1006 if (!list_empty(&pagelist)) {
1007 err = migrate_pages(&pagelist, new_node_page, dest,
1008 false, MIGRATE_SYNC,
1011 putback_lru_pages(&pagelist);
1018 * Move pages between the two nodesets so as to preserve the physical
1019 * layout as much as possible.
1021 * Returns the number of page that could not be moved.
1023 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1024 const nodemask_t *to, int flags)
1030 err = migrate_prep();
1034 down_read(&mm->mmap_sem);
1036 err = migrate_vmas(mm, from, to, flags);
1041 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1042 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1043 * bit in 'tmp', and return that <source, dest> pair for migration.
1044 * The pair of nodemasks 'to' and 'from' define the map.
1046 * If no pair of bits is found that way, fallback to picking some
1047 * pair of 'source' and 'dest' bits that are not the same. If the
1048 * 'source' and 'dest' bits are the same, this represents a node
1049 * that will be migrating to itself, so no pages need move.
1051 * If no bits are left in 'tmp', or if all remaining bits left
1052 * in 'tmp' correspond to the same bit in 'to', return false
1053 * (nothing left to migrate).
1055 * This lets us pick a pair of nodes to migrate between, such that
1056 * if possible the dest node is not already occupied by some other
1057 * source node, minimizing the risk of overloading the memory on a
1058 * node that would happen if we migrated incoming memory to a node
1059 * before migrating outgoing memory source that same node.
1061 * A single scan of tmp is sufficient. As we go, we remember the
1062 * most recent <s, d> pair that moved (s != d). If we find a pair
1063 * that not only moved, but what's better, moved to an empty slot
1064 * (d is not set in tmp), then we break out then, with that pair.
1065 * Otherwise when we finish scanning from_tmp, we at least have the
1066 * most recent <s, d> pair that moved. If we get all the way through
1067 * the scan of tmp without finding any node that moved, much less
1068 * moved to an empty node, then there is nothing left worth migrating.
1072 while (!nodes_empty(tmp)) {
1077 for_each_node_mask(s, tmp) {
1080 * do_migrate_pages() tries to maintain the relative
1081 * node relationship of the pages established between
1082 * threads and memory areas.
1084 * However if the number of source nodes is not equal to
1085 * the number of destination nodes we can not preserve
1086 * this node relative relationship. In that case, skip
1087 * copying memory from a node that is in the destination
1090 * Example: [2,3,4] -> [3,4,5] moves everything.
1091 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1094 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1095 (node_isset(s, *to)))
1098 d = node_remap(s, *from, *to);
1102 source = s; /* Node moved. Memorize */
1105 /* dest not in remaining from nodes? */
1106 if (!node_isset(dest, tmp))
1112 node_clear(source, tmp);
1113 err = migrate_to_node(mm, source, dest, flags);
1120 up_read(&mm->mmap_sem);
1128 * Allocate a new page for page migration based on vma policy.
1129 * Start assuming that page is mapped by vma pointed to by @private.
1130 * Search forward from there, if not. N.B., this assumes that the
1131 * list of pages handed to migrate_pages()--which is how we get here--
1132 * is in virtual address order.
1134 static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
1136 struct vm_area_struct *vma = (struct vm_area_struct *)private;
1137 unsigned long uninitialized_var(address);
1140 address = page_address_in_vma(page, vma);
1141 if (address != -EFAULT)
1147 * if !vma, alloc_page_vma() will use task or system default policy
1149 return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1153 static void migrate_page_add(struct page *page, struct list_head *pagelist,
1154 unsigned long flags)
1158 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1159 const nodemask_t *to, int flags)
1164 static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
1170 static long do_mbind(unsigned long start, unsigned long len,
1171 unsigned short mode, unsigned short mode_flags,
1172 nodemask_t *nmask, unsigned long flags)
1174 struct vm_area_struct *vma;
1175 struct mm_struct *mm = current->mm;
1176 struct mempolicy *new;
1179 LIST_HEAD(pagelist);
1181 if (flags & ~(unsigned long)MPOL_MF_VALID)
1183 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1186 if (start & ~PAGE_MASK)
1189 if (mode == MPOL_DEFAULT || mode == MPOL_NOOP)
1190 flags &= ~MPOL_MF_STRICT;
1192 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1200 new = mpol_new(mode, mode_flags, nmask);
1202 return PTR_ERR(new);
1204 if (flags & MPOL_MF_LAZY)
1205 new->flags |= MPOL_F_MOF;
1208 * If we are using the default policy then operation
1209 * on discontinuous address spaces is okay after all
1212 flags |= MPOL_MF_DISCONTIG_OK;
1214 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1215 start, start + len, mode, mode_flags,
1216 nmask ? nodes_addr(*nmask)[0] : -1);
1218 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1220 err = migrate_prep();
1225 NODEMASK_SCRATCH(scratch);
1227 down_write(&mm->mmap_sem);
1229 err = mpol_set_nodemask(new, nmask, scratch);
1230 task_unlock(current);
1232 up_write(&mm->mmap_sem);
1235 NODEMASK_SCRATCH_FREE(scratch);
1240 vma = check_range(mm, start, end, nmask,
1241 flags | MPOL_MF_INVERT, &pagelist);
1243 err = PTR_ERR(vma); /* maybe ... */
1244 if (!IS_ERR(vma) && mode != MPOL_NOOP)
1245 err = mbind_range(mm, start, end, new);
1250 if (!list_empty(&pagelist)) {
1251 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1252 nr_failed = migrate_pages(&pagelist, new_vma_page,
1254 false, MIGRATE_SYNC,
1255 MR_MEMPOLICY_MBIND);
1257 putback_lru_pages(&pagelist);
1260 if (nr_failed && (flags & MPOL_MF_STRICT))
1263 putback_lru_pages(&pagelist);
1265 up_write(&mm->mmap_sem);
1272 * User space interface with variable sized bitmaps for nodelists.
1275 /* Copy a node mask from user space. */
1276 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1277 unsigned long maxnode)
1280 unsigned long nlongs;
1281 unsigned long endmask;
1284 nodes_clear(*nodes);
1285 if (maxnode == 0 || !nmask)
1287 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1290 nlongs = BITS_TO_LONGS(maxnode);
1291 if ((maxnode % BITS_PER_LONG) == 0)
1294 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1296 /* When the user specified more nodes than supported just check
1297 if the non supported part is all zero. */
1298 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1299 if (nlongs > PAGE_SIZE/sizeof(long))
1301 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1303 if (get_user(t, nmask + k))
1305 if (k == nlongs - 1) {
1311 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1315 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1317 nodes_addr(*nodes)[nlongs-1] &= endmask;
1321 /* Copy a kernel node mask to user space */
1322 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1325 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1326 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
1328 if (copy > nbytes) {
1329 if (copy > PAGE_SIZE)
1331 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1335 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1338 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1339 unsigned long, mode, unsigned long __user *, nmask,
1340 unsigned long, maxnode, unsigned, flags)
1344 unsigned short mode_flags;
1346 mode_flags = mode & MPOL_MODE_FLAGS;
1347 mode &= ~MPOL_MODE_FLAGS;
1348 if (mode >= MPOL_MAX)
1350 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1351 (mode_flags & MPOL_F_RELATIVE_NODES))
1353 err = get_nodes(&nodes, nmask, maxnode);
1356 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1359 /* Set the process memory policy */
1360 SYSCALL_DEFINE3(set_mempolicy, int, mode, unsigned long __user *, nmask,
1361 unsigned long, maxnode)
1365 unsigned short flags;
1367 flags = mode & MPOL_MODE_FLAGS;
1368 mode &= ~MPOL_MODE_FLAGS;
1369 if ((unsigned int)mode >= MPOL_MAX)
1371 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1373 err = get_nodes(&nodes, nmask, maxnode);
1376 return do_set_mempolicy(mode, flags, &nodes);
1379 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1380 const unsigned long __user *, old_nodes,
1381 const unsigned long __user *, new_nodes)
1383 const struct cred *cred = current_cred(), *tcred;
1384 struct mm_struct *mm = NULL;
1385 struct task_struct *task;
1386 nodemask_t task_nodes;
1390 NODEMASK_SCRATCH(scratch);
1395 old = &scratch->mask1;
1396 new = &scratch->mask2;
1398 err = get_nodes(old, old_nodes, maxnode);
1402 err = get_nodes(new, new_nodes, maxnode);
1406 /* Find the mm_struct */
1408 task = pid ? find_task_by_vpid(pid) : current;
1414 get_task_struct(task);
1419 * Check if this process has the right to modify the specified
1420 * process. The right exists if the process has administrative
1421 * capabilities, superuser privileges or the same
1422 * userid as the target process.
1424 tcred = __task_cred(task);
1425 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1426 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
1427 !capable(CAP_SYS_NICE)) {
1434 task_nodes = cpuset_mems_allowed(task);
1435 /* Is the user allowed to access the target nodes? */
1436 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1441 if (!nodes_subset(*new, node_states[N_HIGH_MEMORY])) {
1446 err = security_task_movememory(task);
1450 mm = get_task_mm(task);
1451 put_task_struct(task);
1458 err = do_migrate_pages(mm, old, new,
1459 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1463 NODEMASK_SCRATCH_FREE(scratch);
1468 put_task_struct(task);
1474 /* Retrieve NUMA policy */
1475 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1476 unsigned long __user *, nmask, unsigned long, maxnode,
1477 unsigned long, addr, unsigned long, flags)
1480 int uninitialized_var(pval);
1483 if (nmask != NULL && maxnode < MAX_NUMNODES)
1486 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1491 if (policy && put_user(pval, policy))
1495 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1500 #ifdef CONFIG_COMPAT
1502 asmlinkage long compat_sys_get_mempolicy(int __user *policy,
1503 compat_ulong_t __user *nmask,
1504 compat_ulong_t maxnode,
1505 compat_ulong_t addr, compat_ulong_t flags)
1508 unsigned long __user *nm = NULL;
1509 unsigned long nr_bits, alloc_size;
1510 DECLARE_BITMAP(bm, MAX_NUMNODES);
1512 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1513 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1516 nm = compat_alloc_user_space(alloc_size);
1518 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1520 if (!err && nmask) {
1521 unsigned long copy_size;
1522 copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1523 err = copy_from_user(bm, nm, copy_size);
1524 /* ensure entire bitmap is zeroed */
1525 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1526 err |= compat_put_bitmap(nmask, bm, nr_bits);
1532 asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask,
1533 compat_ulong_t maxnode)
1536 unsigned long __user *nm = NULL;
1537 unsigned long nr_bits, alloc_size;
1538 DECLARE_BITMAP(bm, MAX_NUMNODES);
1540 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1541 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1544 err = compat_get_bitmap(bm, nmask, nr_bits);
1545 nm = compat_alloc_user_space(alloc_size);
1546 err |= copy_to_user(nm, bm, alloc_size);
1552 return sys_set_mempolicy(mode, nm, nr_bits+1);
1555 asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len,
1556 compat_ulong_t mode, compat_ulong_t __user *nmask,
1557 compat_ulong_t maxnode, compat_ulong_t flags)
1560 unsigned long __user *nm = NULL;
1561 unsigned long nr_bits, alloc_size;
1564 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1565 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1568 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
1569 nm = compat_alloc_user_space(alloc_size);
1570 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
1576 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1582 * get_vma_policy(@task, @vma, @addr)
1583 * @task - task for fallback if vma policy == default
1584 * @vma - virtual memory area whose policy is sought
1585 * @addr - address in @vma for shared policy lookup
1587 * Returns effective policy for a VMA at specified address.
1588 * Falls back to @task or system default policy, as necessary.
1589 * Current or other task's task mempolicy and non-shared vma policies must be
1590 * protected by task_lock(task) by the caller.
1591 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1592 * count--added by the get_policy() vm_op, as appropriate--to protect against
1593 * freeing by another task. It is the caller's responsibility to free the
1594 * extra reference for shared policies.
1596 struct mempolicy *get_vma_policy(struct task_struct *task,
1597 struct vm_area_struct *vma, unsigned long addr)
1599 struct mempolicy *pol = task->mempolicy;
1602 if (vma->vm_ops && vma->vm_ops->get_policy) {
1603 struct mempolicy *vpol = vma->vm_ops->get_policy(vma,
1607 } else if (vma->vm_policy) {
1608 pol = vma->vm_policy;
1611 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1612 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1613 * count on these policies which will be dropped by
1614 * mpol_cond_put() later
1616 if (mpol_needs_cond_ref(pol))
1621 pol = &default_policy;
1626 * Return a nodemask representing a mempolicy for filtering nodes for
1629 static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1631 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1632 if (unlikely(policy->mode == MPOL_BIND) &&
1633 gfp_zone(gfp) >= policy_zone &&
1634 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1635 return &policy->v.nodes;
1640 /* Return a zonelist indicated by gfp for node representing a mempolicy */
1641 static struct zonelist *policy_zonelist(gfp_t gfp, struct mempolicy *policy,
1644 switch (policy->mode) {
1645 case MPOL_PREFERRED:
1646 if (!(policy->flags & MPOL_F_LOCAL))
1647 nd = policy->v.preferred_node;
1651 * Normally, MPOL_BIND allocations are node-local within the
1652 * allowed nodemask. However, if __GFP_THISNODE is set and the
1653 * current node isn't part of the mask, we use the zonelist for
1654 * the first node in the mask instead.
1656 if (unlikely(gfp & __GFP_THISNODE) &&
1657 unlikely(!node_isset(nd, policy->v.nodes)))
1658 nd = first_node(policy->v.nodes);
1663 return node_zonelist(nd, gfp);
1666 /* Do dynamic interleaving for a process */
1667 static unsigned interleave_nodes(struct mempolicy *policy)
1670 struct task_struct *me = current;
1673 next = next_node(nid, policy->v.nodes);
1674 if (next >= MAX_NUMNODES)
1675 next = first_node(policy->v.nodes);
1676 if (next < MAX_NUMNODES)
1682 * Depending on the memory policy provide a node from which to allocate the
1684 * @policy must be protected by freeing by the caller. If @policy is
1685 * the current task's mempolicy, this protection is implicit, as only the
1686 * task can change it's policy. The system default policy requires no
1689 unsigned slab_node(void)
1691 struct mempolicy *policy;
1694 return numa_node_id();
1696 policy = current->mempolicy;
1697 if (!policy || policy->flags & MPOL_F_LOCAL)
1698 return numa_node_id();
1700 switch (policy->mode) {
1701 case MPOL_PREFERRED:
1703 * handled MPOL_F_LOCAL above
1705 return policy->v.preferred_node;
1707 case MPOL_INTERLEAVE:
1708 return interleave_nodes(policy);
1712 * Follow bind policy behavior and start allocation at the
1715 struct zonelist *zonelist;
1717 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1718 zonelist = &NODE_DATA(numa_node_id())->node_zonelists[0];
1719 (void)first_zones_zonelist(zonelist, highest_zoneidx,
1722 return zone ? zone->node : numa_node_id();
1730 /* Do static interleaving for a VMA with known offset. */
1731 static unsigned offset_il_node(struct mempolicy *pol,
1732 struct vm_area_struct *vma, unsigned long off)
1734 unsigned nnodes = nodes_weight(pol->v.nodes);
1740 return numa_node_id();
1741 target = (unsigned int)off % nnodes;
1744 nid = next_node(nid, pol->v.nodes);
1746 } while (c <= target);
1750 /* Determine a node number for interleave */
1751 static inline unsigned interleave_nid(struct mempolicy *pol,
1752 struct vm_area_struct *vma, unsigned long addr, int shift)
1758 * for small pages, there is no difference between
1759 * shift and PAGE_SHIFT, so the bit-shift is safe.
1760 * for huge pages, since vm_pgoff is in units of small
1761 * pages, we need to shift off the always 0 bits to get
1764 BUG_ON(shift < PAGE_SHIFT);
1765 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1766 off += (addr - vma->vm_start) >> shift;
1767 return offset_il_node(pol, vma, off);
1769 return interleave_nodes(pol);
1773 * Return the bit number of a random bit set in the nodemask.
1774 * (returns -1 if nodemask is empty)
1776 int node_random(const nodemask_t *maskp)
1780 w = nodes_weight(*maskp);
1782 bit = bitmap_ord_to_pos(maskp->bits,
1783 get_random_int() % w, MAX_NUMNODES);
1787 #ifdef CONFIG_HUGETLBFS
1789 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1790 * @vma = virtual memory area whose policy is sought
1791 * @addr = address in @vma for shared policy lookup and interleave policy
1792 * @gfp_flags = for requested zone
1793 * @mpol = pointer to mempolicy pointer for reference counted mempolicy
1794 * @nodemask = pointer to nodemask pointer for MPOL_BIND nodemask
1796 * Returns a zonelist suitable for a huge page allocation and a pointer
1797 * to the struct mempolicy for conditional unref after allocation.
1798 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1799 * @nodemask for filtering the zonelist.
1801 * Must be protected by get_mems_allowed()
1803 struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
1804 gfp_t gfp_flags, struct mempolicy **mpol,
1805 nodemask_t **nodemask)
1807 struct zonelist *zl;
1809 *mpol = get_vma_policy(current, vma, addr);
1810 *nodemask = NULL; /* assume !MPOL_BIND */
1812 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1813 zl = node_zonelist(interleave_nid(*mpol, vma, addr,
1814 huge_page_shift(hstate_vma(vma))), gfp_flags);
1816 zl = policy_zonelist(gfp_flags, *mpol, numa_node_id());
1817 if ((*mpol)->mode == MPOL_BIND)
1818 *nodemask = &(*mpol)->v.nodes;
1824 * init_nodemask_of_mempolicy
1826 * If the current task's mempolicy is "default" [NULL], return 'false'
1827 * to indicate default policy. Otherwise, extract the policy nodemask
1828 * for 'bind' or 'interleave' policy into the argument nodemask, or
1829 * initialize the argument nodemask to contain the single node for
1830 * 'preferred' or 'local' policy and return 'true' to indicate presence
1831 * of non-default mempolicy.
1833 * We don't bother with reference counting the mempolicy [mpol_get/put]
1834 * because the current task is examining it's own mempolicy and a task's
1835 * mempolicy is only ever changed by the task itself.
1837 * N.B., it is the caller's responsibility to free a returned nodemask.
1839 bool init_nodemask_of_mempolicy(nodemask_t *mask)
1841 struct mempolicy *mempolicy;
1844 if (!(mask && current->mempolicy))
1848 mempolicy = current->mempolicy;
1849 switch (mempolicy->mode) {
1850 case MPOL_PREFERRED:
1851 if (mempolicy->flags & MPOL_F_LOCAL)
1852 nid = numa_node_id();
1854 nid = mempolicy->v.preferred_node;
1855 init_nodemask_of_node(mask, nid);
1860 case MPOL_INTERLEAVE:
1861 *mask = mempolicy->v.nodes;
1867 task_unlock(current);
1874 * mempolicy_nodemask_intersects
1876 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1877 * policy. Otherwise, check for intersection between mask and the policy
1878 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1879 * policy, always return true since it may allocate elsewhere on fallback.
1881 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1883 bool mempolicy_nodemask_intersects(struct task_struct *tsk,
1884 const nodemask_t *mask)
1886 struct mempolicy *mempolicy;
1892 mempolicy = tsk->mempolicy;
1896 switch (mempolicy->mode) {
1897 case MPOL_PREFERRED:
1899 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1900 * allocate from, they may fallback to other nodes when oom.
1901 * Thus, it's possible for tsk to have allocated memory from
1906 case MPOL_INTERLEAVE:
1907 ret = nodes_intersects(mempolicy->v.nodes, *mask);
1917 /* Allocate a page in interleaved policy.
1918 Own path because it needs to do special accounting. */
1919 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1922 struct zonelist *zl;
1925 zl = node_zonelist(nid, gfp);
1926 page = __alloc_pages(gfp, order, zl);
1927 if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0]))
1928 inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
1933 * alloc_pages_vma - Allocate a page for a VMA.
1936 * %GFP_USER user allocation.
1937 * %GFP_KERNEL kernel allocations,
1938 * %GFP_HIGHMEM highmem/user allocations,
1939 * %GFP_FS allocation should not call back into a file system.
1940 * %GFP_ATOMIC don't sleep.
1942 * @order:Order of the GFP allocation.
1943 * @vma: Pointer to VMA or NULL if not available.
1944 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1946 * This function allocates a page from the kernel page pool and applies
1947 * a NUMA policy associated with the VMA or the current process.
1948 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1949 * mm_struct of the VMA to prevent it from going away. Should be used for
1950 * all allocations for pages that will be mapped into
1951 * user space. Returns NULL when no page can be allocated.
1953 * Should be called with the mm_sem of the vma hold.
1956 alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
1957 unsigned long addr, int node)
1959 struct mempolicy *pol;
1960 struct zonelist *zl;
1962 unsigned int cpuset_mems_cookie;
1965 pol = get_vma_policy(current, vma, addr);
1966 cpuset_mems_cookie = get_mems_allowed();
1968 if (unlikely(pol->mode == MPOL_INTERLEAVE)) {
1971 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
1973 page = alloc_page_interleave(gfp, order, nid);
1974 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
1979 zl = policy_zonelist(gfp, pol, node);
1980 if (unlikely(mpol_needs_cond_ref(pol))) {
1982 * slow path: ref counted shared policy
1984 struct page *page = __alloc_pages_nodemask(gfp, order,
1985 zl, policy_nodemask(gfp, pol));
1987 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
1992 * fast path: default or task policy
1994 page = __alloc_pages_nodemask(gfp, order, zl,
1995 policy_nodemask(gfp, pol));
1996 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2002 * alloc_pages_current - Allocate pages.
2005 * %GFP_USER user allocation,
2006 * %GFP_KERNEL kernel allocation,
2007 * %GFP_HIGHMEM highmem allocation,
2008 * %GFP_FS don't call back into a file system.
2009 * %GFP_ATOMIC don't sleep.
2010 * @order: Power of two of allocation size in pages. 0 is a single page.
2012 * Allocate a page from the kernel page pool. When not in
2013 * interrupt context and apply the current process NUMA policy.
2014 * Returns NULL when no page can be allocated.
2016 * Don't call cpuset_update_task_memory_state() unless
2017 * 1) it's ok to take cpuset_sem (can WAIT), and
2018 * 2) allocating for current task (not interrupt).
2020 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
2022 struct mempolicy *pol = current->mempolicy;
2024 unsigned int cpuset_mems_cookie;
2026 if (!pol || in_interrupt() || (gfp & __GFP_THISNODE))
2027 pol = &default_policy;
2030 cpuset_mems_cookie = get_mems_allowed();
2033 * No reference counting needed for current->mempolicy
2034 * nor system default_policy
2036 if (pol->mode == MPOL_INTERLEAVE)
2037 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2039 page = __alloc_pages_nodemask(gfp, order,
2040 policy_zonelist(gfp, pol, numa_node_id()),
2041 policy_nodemask(gfp, pol));
2043 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2048 EXPORT_SYMBOL(alloc_pages_current);
2051 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2052 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2053 * with the mems_allowed returned by cpuset_mems_allowed(). This
2054 * keeps mempolicies cpuset relative after its cpuset moves. See
2055 * further kernel/cpuset.c update_nodemask().
2057 * current's mempolicy may be rebinded by the other task(the task that changes
2058 * cpuset's mems), so we needn't do rebind work for current task.
2061 /* Slow path of a mempolicy duplicate */
2062 struct mempolicy *__mpol_dup(struct mempolicy *old)
2064 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2067 return ERR_PTR(-ENOMEM);
2069 /* task's mempolicy is protected by alloc_lock */
2070 if (old == current->mempolicy) {
2073 task_unlock(current);
2078 if (current_cpuset_is_being_rebound()) {
2079 nodemask_t mems = cpuset_mems_allowed(current);
2080 if (new->flags & MPOL_F_REBINDING)
2081 mpol_rebind_policy(new, &mems, MPOL_REBIND_STEP2);
2083 mpol_rebind_policy(new, &mems, MPOL_REBIND_ONCE);
2086 atomic_set(&new->refcnt, 1);
2091 * If *frompol needs [has] an extra ref, copy *frompol to *tompol ,
2092 * eliminate the * MPOL_F_* flags that require conditional ref and
2093 * [NOTE!!!] drop the extra ref. Not safe to reference *frompol directly
2094 * after return. Use the returned value.
2096 * Allows use of a mempolicy for, e.g., multiple allocations with a single
2097 * policy lookup, even if the policy needs/has extra ref on lookup.
2098 * shmem_readahead needs this.
2100 struct mempolicy *__mpol_cond_copy(struct mempolicy *tompol,
2101 struct mempolicy *frompol)
2103 if (!mpol_needs_cond_ref(frompol))
2107 tompol->flags &= ~MPOL_F_SHARED; /* copy doesn't need unref */
2108 __mpol_put(frompol);
2112 /* Slow path of a mempolicy comparison */
2113 bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2117 if (a->mode != b->mode)
2119 if (a->flags != b->flags)
2121 if (mpol_store_user_nodemask(a))
2122 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2128 case MPOL_INTERLEAVE:
2129 return !!nodes_equal(a->v.nodes, b->v.nodes);
2130 case MPOL_PREFERRED:
2131 return a->v.preferred_node == b->v.preferred_node;
2139 * Shared memory backing store policy support.
2141 * Remember policies even when nobody has shared memory mapped.
2142 * The policies are kept in Red-Black tree linked from the inode.
2143 * They are protected by the sp->lock spinlock, which should be held
2144 * for any accesses to the tree.
2147 /* lookup first element intersecting start-end */
2148 /* Caller holds sp->mutex */
2149 static struct sp_node *
2150 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2152 struct rb_node *n = sp->root.rb_node;
2155 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2157 if (start >= p->end)
2159 else if (end <= p->start)
2167 struct sp_node *w = NULL;
2168 struct rb_node *prev = rb_prev(n);
2171 w = rb_entry(prev, struct sp_node, nd);
2172 if (w->end <= start)
2176 return rb_entry(n, struct sp_node, nd);
2179 /* Insert a new shared policy into the list. */
2180 /* Caller holds sp->lock */
2181 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2183 struct rb_node **p = &sp->root.rb_node;
2184 struct rb_node *parent = NULL;
2189 nd = rb_entry(parent, struct sp_node, nd);
2190 if (new->start < nd->start)
2192 else if (new->end > nd->end)
2193 p = &(*p)->rb_right;
2197 rb_link_node(&new->nd, parent, p);
2198 rb_insert_color(&new->nd, &sp->root);
2199 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2200 new->policy ? new->policy->mode : 0);
2203 /* Find shared policy intersecting idx */
2205 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2207 struct mempolicy *pol = NULL;
2210 if (!sp->root.rb_node)
2212 mutex_lock(&sp->mutex);
2213 sn = sp_lookup(sp, idx, idx+1);
2215 mpol_get(sn->policy);
2218 mutex_unlock(&sp->mutex);
2222 static void sp_free(struct sp_node *n)
2224 mpol_put(n->policy);
2225 kmem_cache_free(sn_cache, n);
2229 * mpol_misplaced - check whether current page node is valid in policy
2231 * @page - page to be checked
2232 * @vma - vm area where page mapped
2233 * @addr - virtual address where page mapped
2235 * Lookup current policy node id for vma,addr and "compare to" page's
2239 * -1 - not misplaced, page is in the right node
2240 * node - node id where the page should be
2242 * Policy determination "mimics" alloc_page_vma().
2243 * Called from fault path where we know the vma and faulting address.
2245 int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2247 struct mempolicy *pol;
2249 int curnid = page_to_nid(page);
2250 unsigned long pgoff;
2256 pol = get_vma_policy(current, vma, addr);
2257 if (!(pol->flags & MPOL_F_MOF))
2260 switch (pol->mode) {
2261 case MPOL_INTERLEAVE:
2262 BUG_ON(addr >= vma->vm_end);
2263 BUG_ON(addr < vma->vm_start);
2265 pgoff = vma->vm_pgoff;
2266 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2267 polnid = offset_il_node(pol, vma, pgoff);
2270 case MPOL_PREFERRED:
2271 if (pol->flags & MPOL_F_LOCAL)
2272 polnid = numa_node_id();
2274 polnid = pol->v.preferred_node;
2279 * allows binding to multiple nodes.
2280 * use current page if in policy nodemask,
2281 * else select nearest allowed node, if any.
2282 * If no allowed nodes, use current [!misplaced].
2284 if (node_isset(curnid, pol->v.nodes))
2286 (void)first_zones_zonelist(
2287 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2288 gfp_zone(GFP_HIGHUSER),
2289 &pol->v.nodes, &zone);
2290 polnid = zone->node;
2296 if (curnid != polnid)
2304 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2306 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2307 rb_erase(&n->nd, &sp->root);
2311 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2312 struct mempolicy *pol)
2315 struct mempolicy *newpol;
2317 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2321 newpol = mpol_dup(pol);
2322 if (IS_ERR(newpol)) {
2323 kmem_cache_free(sn_cache, n);
2326 newpol->flags |= MPOL_F_SHARED;
2335 /* Replace a policy range. */
2336 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2337 unsigned long end, struct sp_node *new)
2342 mutex_lock(&sp->mutex);
2343 n = sp_lookup(sp, start, end);
2344 /* Take care of old policies in the same range. */
2345 while (n && n->start < end) {
2346 struct rb_node *next = rb_next(&n->nd);
2347 if (n->start >= start) {
2353 /* Old policy spanning whole new range. */
2355 struct sp_node *new2;
2356 new2 = sp_alloc(end, n->end, n->policy);
2362 sp_insert(sp, new2);
2369 n = rb_entry(next, struct sp_node, nd);
2374 mutex_unlock(&sp->mutex);
2379 * mpol_shared_policy_init - initialize shared policy for inode
2380 * @sp: pointer to inode shared policy
2381 * @mpol: struct mempolicy to install
2383 * Install non-NULL @mpol in inode's shared policy rb-tree.
2384 * On entry, the current task has a reference on a non-NULL @mpol.
2385 * This must be released on exit.
2386 * This is called at get_inode() calls and we can use GFP_KERNEL.
2388 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2392 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2393 mutex_init(&sp->mutex);
2396 struct vm_area_struct pvma;
2397 struct mempolicy *new;
2398 NODEMASK_SCRATCH(scratch);
2402 /* contextualize the tmpfs mount point mempolicy */
2403 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2405 goto free_scratch; /* no valid nodemask intersection */
2408 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2409 task_unlock(current);
2413 /* Create pseudo-vma that contains just the policy */
2414 memset(&pvma, 0, sizeof(struct vm_area_struct));
2415 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2416 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2419 mpol_put(new); /* drop initial ref */
2421 NODEMASK_SCRATCH_FREE(scratch);
2423 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2427 int mpol_set_shared_policy(struct shared_policy *info,
2428 struct vm_area_struct *vma, struct mempolicy *npol)
2431 struct sp_node *new = NULL;
2432 unsigned long sz = vma_pages(vma);
2434 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2436 sz, npol ? npol->mode : -1,
2437 npol ? npol->flags : -1,
2438 npol ? nodes_addr(npol->v.nodes)[0] : -1);
2441 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2445 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2451 /* Free a backing policy store on inode delete. */
2452 void mpol_free_shared_policy(struct shared_policy *p)
2455 struct rb_node *next;
2457 if (!p->root.rb_node)
2459 mutex_lock(&p->mutex);
2460 next = rb_first(&p->root);
2462 n = rb_entry(next, struct sp_node, nd);
2463 next = rb_next(&n->nd);
2466 mutex_unlock(&p->mutex);
2469 /* assumes fs == KERNEL_DS */
2470 void __init numa_policy_init(void)
2472 nodemask_t interleave_nodes;
2473 unsigned long largest = 0;
2474 int nid, prefer = 0;
2476 policy_cache = kmem_cache_create("numa_policy",
2477 sizeof(struct mempolicy),
2478 0, SLAB_PANIC, NULL);
2480 sn_cache = kmem_cache_create("shared_policy_node",
2481 sizeof(struct sp_node),
2482 0, SLAB_PANIC, NULL);
2485 * Set interleaving policy for system init. Interleaving is only
2486 * enabled across suitably sized nodes (default is >= 16MB), or
2487 * fall back to the largest node if they're all smaller.
2489 nodes_clear(interleave_nodes);
2490 for_each_node_state(nid, N_HIGH_MEMORY) {
2491 unsigned long total_pages = node_present_pages(nid);
2493 /* Preserve the largest node */
2494 if (largest < total_pages) {
2495 largest = total_pages;
2499 /* Interleave this node? */
2500 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2501 node_set(nid, interleave_nodes);
2504 /* All too small, use the largest */
2505 if (unlikely(nodes_empty(interleave_nodes)))
2506 node_set(prefer, interleave_nodes);
2508 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2509 printk("numa_policy_init: interleaving failed\n");
2512 /* Reset policy of current process to default */
2513 void numa_default_policy(void)
2515 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2519 * Parse and format mempolicy from/to strings
2523 * "local" is pseudo-policy: MPOL_PREFERRED with MPOL_F_LOCAL flag
2524 * Used only for mpol_parse_str() and mpol_to_str()
2526 static const char * const policy_modes[] =
2528 [MPOL_DEFAULT] = "default",
2529 [MPOL_PREFERRED] = "prefer",
2530 [MPOL_BIND] = "bind",
2531 [MPOL_INTERLEAVE] = "interleave",
2532 [MPOL_LOCAL] = "local",
2533 [MPOL_NOOP] = "noop", /* should not actually be used */
2539 * mpol_parse_str - parse string to mempolicy
2540 * @str: string containing mempolicy to parse
2541 * @mpol: pointer to struct mempolicy pointer, returned on success.
2542 * @no_context: flag whether to "contextualize" the mempolicy
2545 * <mode>[=<flags>][:<nodelist>]
2547 * if @no_context is true, save the input nodemask in w.user_nodemask in
2548 * the returned mempolicy. This will be used to "clone" the mempolicy in
2549 * a specific context [cpuset] at a later time. Used to parse tmpfs mpol
2550 * mount option. Note that if 'static' or 'relative' mode flags were
2551 * specified, the input nodemask will already have been saved. Saving
2552 * it again is redundant, but safe.
2554 * On success, returns 0, else 1
2556 int mpol_parse_str(char *str, struct mempolicy **mpol, int no_context)
2558 struct mempolicy *new = NULL;
2559 unsigned short mode;
2560 unsigned short uninitialized_var(mode_flags);
2562 char *nodelist = strchr(str, ':');
2563 char *flags = strchr(str, '=');
2567 /* NUL-terminate mode or flags string */
2569 if (nodelist_parse(nodelist, nodes))
2571 if (!nodes_subset(nodes, node_states[N_HIGH_MEMORY]))
2577 *flags++ = '\0'; /* terminate mode string */
2579 for (mode = 0; mode < MPOL_MAX; mode++) {
2580 if (!strcmp(str, policy_modes[mode])) {
2584 if (mode >= MPOL_MAX || mode == MPOL_NOOP)
2588 case MPOL_PREFERRED:
2590 * Insist on a nodelist of one node only
2593 char *rest = nodelist;
2594 while (isdigit(*rest))
2600 case MPOL_INTERLEAVE:
2602 * Default to online nodes with memory if no nodelist
2605 nodes = node_states[N_HIGH_MEMORY];
2609 * Don't allow a nodelist; mpol_new() checks flags
2613 mode = MPOL_PREFERRED;
2617 * Insist on a empty nodelist
2624 * Insist on a nodelist
2633 * Currently, we only support two mutually exclusive
2636 if (!strcmp(flags, "static"))
2637 mode_flags |= MPOL_F_STATIC_NODES;
2638 else if (!strcmp(flags, "relative"))
2639 mode_flags |= MPOL_F_RELATIVE_NODES;
2644 new = mpol_new(mode, mode_flags, &nodes);
2649 /* save for contextualization */
2650 new->w.user_nodemask = nodes;
2653 NODEMASK_SCRATCH(scratch);
2656 ret = mpol_set_nodemask(new, &nodes, scratch);
2657 task_unlock(current);
2660 NODEMASK_SCRATCH_FREE(scratch);
2669 /* Restore string for error message */
2678 #endif /* CONFIG_TMPFS */
2681 * mpol_to_str - format a mempolicy structure for printing
2682 * @buffer: to contain formatted mempolicy string
2683 * @maxlen: length of @buffer
2684 * @pol: pointer to mempolicy to be formatted
2685 * @no_context: "context free" mempolicy - use nodemask in w.user_nodemask
2687 * Convert a mempolicy into a string.
2688 * Returns the number of characters in buffer (if positive)
2689 * or an error (negative)
2691 int mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol, int no_context)
2696 unsigned short mode;
2697 unsigned short flags = pol ? pol->flags : 0;
2700 * Sanity check: room for longest mode, flag and some nodes
2702 VM_BUG_ON(maxlen < strlen("interleave") + strlen("relative") + 16);
2704 if (!pol || pol == &default_policy)
2705 mode = MPOL_DEFAULT;
2714 case MPOL_PREFERRED:
2716 if (flags & MPOL_F_LOCAL)
2717 mode = MPOL_LOCAL; /* pseudo-policy */
2719 node_set(pol->v.preferred_node, nodes);
2724 case MPOL_INTERLEAVE:
2726 nodes = pol->w.user_nodemask;
2728 nodes = pol->v.nodes;
2735 l = strlen(policy_modes[mode]);
2736 if (buffer + maxlen < p + l + 1)
2739 strcpy(p, policy_modes[mode]);
2742 if (flags & MPOL_MODE_FLAGS) {
2743 if (buffer + maxlen < p + 2)
2748 * Currently, the only defined flags are mutually exclusive
2750 if (flags & MPOL_F_STATIC_NODES)
2751 p += snprintf(p, buffer + maxlen - p, "static");
2752 else if (flags & MPOL_F_RELATIVE_NODES)
2753 p += snprintf(p, buffer + maxlen - p, "relative");
2756 if (!nodes_empty(nodes)) {
2757 if (buffer + maxlen < p + 2)
2760 p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);