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 NUMA_NO_NODE 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 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
70 #include <linux/mempolicy.h>
72 #include <linux/highmem.h>
73 #include <linux/hugetlb.h>
74 #include <linux/kernel.h>
75 #include <linux/sched.h>
76 #include <linux/nodemask.h>
77 #include <linux/cpuset.h>
78 #include <linux/slab.h>
79 #include <linux/string.h>
80 #include <linux/export.h>
81 #include <linux/nsproxy.h>
82 #include <linux/interrupt.h>
83 #include <linux/init.h>
84 #include <linux/compat.h>
85 #include <linux/swap.h>
86 #include <linux/seq_file.h>
87 #include <linux/proc_fs.h>
88 #include <linux/migrate.h>
89 #include <linux/ksm.h>
90 #include <linux/rmap.h>
91 #include <linux/security.h>
92 #include <linux/syscalls.h>
93 #include <linux/ctype.h>
94 #include <linux/mm_inline.h>
95 #include <linux/mmu_notifier.h>
96 #include <linux/printk.h>
98 #include <asm/tlbflush.h>
99 #include <asm/uaccess.h>
100 #include <linux/random.h>
102 #include "internal.h"
105 #define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
106 #define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
108 static struct kmem_cache *policy_cache;
109 static struct kmem_cache *sn_cache;
111 /* Highest zone. An specific allocation for a zone below that is not
113 enum zone_type policy_zone = 0;
116 * run-time system-wide default policy => local allocation
118 static struct mempolicy default_policy = {
119 .refcnt = ATOMIC_INIT(1), /* never free it */
120 .mode = MPOL_PREFERRED,
121 .flags = MPOL_F_LOCAL,
124 static struct mempolicy preferred_node_policy[MAX_NUMNODES];
126 struct mempolicy *get_task_policy(struct task_struct *p)
128 struct mempolicy *pol = p->mempolicy;
134 node = numa_node_id();
135 if (node != NUMA_NO_NODE) {
136 pol = &preferred_node_policy[node];
137 /* preferred_node_policy is not initialised early in boot */
142 return &default_policy;
145 static const struct mempolicy_operations {
146 int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
148 * If read-side task has no lock to protect task->mempolicy, write-side
149 * task will rebind the task->mempolicy by two step. The first step is
150 * setting all the newly nodes, and the second step is cleaning all the
151 * disallowed nodes. In this way, we can avoid finding no node to alloc
153 * If we have a lock to protect task->mempolicy in read-side, we do
157 * MPOL_REBIND_ONCE - do rebind work at once
158 * MPOL_REBIND_STEP1 - set all the newly nodes
159 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
161 void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes,
162 enum mpol_rebind_step step);
163 } mpol_ops[MPOL_MAX];
165 /* Check that the nodemask contains at least one populated zone */
166 static int is_valid_nodemask(const nodemask_t *nodemask)
168 return nodes_intersects(*nodemask, node_states[N_MEMORY]);
171 static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
173 return pol->flags & MPOL_MODE_FLAGS;
176 static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
177 const nodemask_t *rel)
180 nodes_fold(tmp, *orig, nodes_weight(*rel));
181 nodes_onto(*ret, tmp, *rel);
184 static int mpol_new_interleave(struct mempolicy *pol, const nodemask_t *nodes)
186 if (nodes_empty(*nodes))
188 pol->v.nodes = *nodes;
192 static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
195 pol->flags |= MPOL_F_LOCAL; /* local allocation */
196 else if (nodes_empty(*nodes))
197 return -EINVAL; /* no allowed nodes */
199 pol->v.preferred_node = first_node(*nodes);
203 static int mpol_new_bind(struct mempolicy *pol, const nodemask_t *nodes)
205 if (!is_valid_nodemask(nodes))
207 pol->v.nodes = *nodes;
212 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
213 * any, for the new policy. mpol_new() has already validated the nodes
214 * parameter with respect to the policy mode and flags. But, we need to
215 * handle an empty nodemask with MPOL_PREFERRED here.
217 * Must be called holding task's alloc_lock to protect task's mems_allowed
218 * and mempolicy. May also be called holding the mmap_semaphore for write.
220 static int mpol_set_nodemask(struct mempolicy *pol,
221 const nodemask_t *nodes, struct nodemask_scratch *nsc)
225 /* if mode is MPOL_DEFAULT, pol is NULL. This is right. */
229 nodes_and(nsc->mask1,
230 cpuset_current_mems_allowed, node_states[N_MEMORY]);
233 if (pol->mode == MPOL_PREFERRED && nodes_empty(*nodes))
234 nodes = NULL; /* explicit local allocation */
236 if (pol->flags & MPOL_F_RELATIVE_NODES)
237 mpol_relative_nodemask(&nsc->mask2, nodes,&nsc->mask1);
239 nodes_and(nsc->mask2, *nodes, nsc->mask1);
241 if (mpol_store_user_nodemask(pol))
242 pol->w.user_nodemask = *nodes;
244 pol->w.cpuset_mems_allowed =
245 cpuset_current_mems_allowed;
249 ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
251 ret = mpol_ops[pol->mode].create(pol, NULL);
256 * This function just creates a new policy, does some check and simple
257 * initialization. You must invoke mpol_set_nodemask() to set nodes.
259 static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
262 struct mempolicy *policy;
264 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
265 mode, flags, nodes ? nodes_addr(*nodes)[0] : NUMA_NO_NODE);
267 if (mode == MPOL_DEFAULT) {
268 if (nodes && !nodes_empty(*nodes))
269 return ERR_PTR(-EINVAL);
275 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
276 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
277 * All other modes require a valid pointer to a non-empty nodemask.
279 if (mode == MPOL_PREFERRED) {
280 if (nodes_empty(*nodes)) {
281 if (((flags & MPOL_F_STATIC_NODES) ||
282 (flags & MPOL_F_RELATIVE_NODES)))
283 return ERR_PTR(-EINVAL);
285 } else if (mode == MPOL_LOCAL) {
286 if (!nodes_empty(*nodes))
287 return ERR_PTR(-EINVAL);
288 mode = MPOL_PREFERRED;
289 } else if (nodes_empty(*nodes))
290 return ERR_PTR(-EINVAL);
291 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
293 return ERR_PTR(-ENOMEM);
294 atomic_set(&policy->refcnt, 1);
296 policy->flags = flags;
301 /* Slow path of a mpol destructor. */
302 void __mpol_put(struct mempolicy *p)
304 if (!atomic_dec_and_test(&p->refcnt))
306 kmem_cache_free(policy_cache, p);
309 static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes,
310 enum mpol_rebind_step step)
316 * MPOL_REBIND_ONCE - do rebind work at once
317 * MPOL_REBIND_STEP1 - set all the newly nodes
318 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
320 static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes,
321 enum mpol_rebind_step step)
325 if (pol->flags & MPOL_F_STATIC_NODES)
326 nodes_and(tmp, pol->w.user_nodemask, *nodes);
327 else if (pol->flags & MPOL_F_RELATIVE_NODES)
328 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
331 * if step == 1, we use ->w.cpuset_mems_allowed to cache the
334 if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP1) {
335 nodes_remap(tmp, pol->v.nodes,
336 pol->w.cpuset_mems_allowed, *nodes);
337 pol->w.cpuset_mems_allowed = step ? tmp : *nodes;
338 } else if (step == MPOL_REBIND_STEP2) {
339 tmp = pol->w.cpuset_mems_allowed;
340 pol->w.cpuset_mems_allowed = *nodes;
345 if (nodes_empty(tmp))
348 if (step == MPOL_REBIND_STEP1)
349 nodes_or(pol->v.nodes, pol->v.nodes, tmp);
350 else if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP2)
355 if (!node_isset(current->il_next, tmp)) {
356 current->il_next = next_node(current->il_next, tmp);
357 if (current->il_next >= MAX_NUMNODES)
358 current->il_next = first_node(tmp);
359 if (current->il_next >= MAX_NUMNODES)
360 current->il_next = numa_node_id();
364 static void mpol_rebind_preferred(struct mempolicy *pol,
365 const nodemask_t *nodes,
366 enum mpol_rebind_step step)
370 if (pol->flags & MPOL_F_STATIC_NODES) {
371 int node = first_node(pol->w.user_nodemask);
373 if (node_isset(node, *nodes)) {
374 pol->v.preferred_node = node;
375 pol->flags &= ~MPOL_F_LOCAL;
377 pol->flags |= MPOL_F_LOCAL;
378 } else if (pol->flags & MPOL_F_RELATIVE_NODES) {
379 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
380 pol->v.preferred_node = first_node(tmp);
381 } else if (!(pol->flags & MPOL_F_LOCAL)) {
382 pol->v.preferred_node = node_remap(pol->v.preferred_node,
383 pol->w.cpuset_mems_allowed,
385 pol->w.cpuset_mems_allowed = *nodes;
390 * mpol_rebind_policy - Migrate a policy to a different set of nodes
392 * If read-side task has no lock to protect task->mempolicy, write-side
393 * task will rebind the task->mempolicy by two step. The first step is
394 * setting all the newly nodes, and the second step is cleaning all the
395 * disallowed nodes. In this way, we can avoid finding no node to alloc
397 * If we have a lock to protect task->mempolicy in read-side, we do
401 * MPOL_REBIND_ONCE - do rebind work at once
402 * MPOL_REBIND_STEP1 - set all the newly nodes
403 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
405 static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask,
406 enum mpol_rebind_step step)
410 if (!mpol_store_user_nodemask(pol) && step == MPOL_REBIND_ONCE &&
411 nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
414 if (step == MPOL_REBIND_STEP1 && (pol->flags & MPOL_F_REBINDING))
417 if (step == MPOL_REBIND_STEP2 && !(pol->flags & MPOL_F_REBINDING))
420 if (step == MPOL_REBIND_STEP1)
421 pol->flags |= MPOL_F_REBINDING;
422 else if (step == MPOL_REBIND_STEP2)
423 pol->flags &= ~MPOL_F_REBINDING;
424 else if (step >= MPOL_REBIND_NSTEP)
427 mpol_ops[pol->mode].rebind(pol, newmask, step);
431 * Wrapper for mpol_rebind_policy() that just requires task
432 * pointer, and updates task mempolicy.
434 * Called with task's alloc_lock held.
437 void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new,
438 enum mpol_rebind_step step)
440 mpol_rebind_policy(tsk->mempolicy, new, step);
444 * Rebind each vma in mm to new nodemask.
446 * Call holding a reference to mm. Takes mm->mmap_sem during call.
449 void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
451 struct vm_area_struct *vma;
453 down_write(&mm->mmap_sem);
454 for (vma = mm->mmap; vma; vma = vma->vm_next)
455 mpol_rebind_policy(vma->vm_policy, new, MPOL_REBIND_ONCE);
456 up_write(&mm->mmap_sem);
459 static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
461 .rebind = mpol_rebind_default,
463 [MPOL_INTERLEAVE] = {
464 .create = mpol_new_interleave,
465 .rebind = mpol_rebind_nodemask,
468 .create = mpol_new_preferred,
469 .rebind = mpol_rebind_preferred,
472 .create = mpol_new_bind,
473 .rebind = mpol_rebind_nodemask,
477 static void migrate_page_add(struct page *page, struct list_head *pagelist,
478 unsigned long flags);
481 * Scan through pages checking if pages follow certain conditions,
482 * and move them to the pagelist if they do.
484 static int queue_pages_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.
507 if (PageReserved(page))
509 nid = page_to_nid(page);
510 if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
513 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
514 migrate_page_add(page, private, flags);
517 } while (pte++, addr += PAGE_SIZE, addr != end);
518 pte_unmap_unlock(orig_pte, ptl);
522 static void queue_pages_hugetlb_pmd_range(struct vm_area_struct *vma,
523 pmd_t *pmd, const nodemask_t *nodes, unsigned long flags,
526 #ifdef CONFIG_HUGETLB_PAGE
532 ptl = huge_pte_lock(hstate_vma(vma), vma->vm_mm, (pte_t *)pmd);
533 entry = huge_ptep_get((pte_t *)pmd);
534 if (!pte_present(entry))
536 page = pte_page(entry);
537 nid = page_to_nid(page);
538 if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
540 /* With MPOL_MF_MOVE, we migrate only unshared hugepage. */
541 if (flags & (MPOL_MF_MOVE_ALL) ||
542 (flags & MPOL_MF_MOVE && page_mapcount(page) == 1))
543 isolate_huge_page(page, private);
551 static inline int queue_pages_pmd_range(struct vm_area_struct *vma, pud_t *pud,
552 unsigned long addr, unsigned long end,
553 const nodemask_t *nodes, unsigned long flags,
559 pmd = pmd_offset(pud, addr);
561 next = pmd_addr_end(addr, end);
562 if (!pmd_present(*pmd))
564 if (pmd_huge(*pmd) && is_vm_hugetlb_page(vma)) {
565 queue_pages_hugetlb_pmd_range(vma, pmd, nodes,
569 split_huge_page_pmd(vma, addr, pmd);
570 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
572 if (queue_pages_pte_range(vma, pmd, addr, next, nodes,
575 } while (pmd++, addr = next, addr != end);
579 static inline int queue_pages_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
580 unsigned long addr, unsigned long end,
581 const nodemask_t *nodes, unsigned long flags,
587 pud = pud_offset(pgd, addr);
589 next = pud_addr_end(addr, end);
590 if (pud_huge(*pud) && is_vm_hugetlb_page(vma))
592 if (pud_none_or_clear_bad(pud))
594 if (queue_pages_pmd_range(vma, pud, addr, next, nodes,
597 } while (pud++, addr = next, addr != end);
601 static inline int queue_pages_pgd_range(struct vm_area_struct *vma,
602 unsigned long addr, unsigned long end,
603 const nodemask_t *nodes, unsigned long flags,
609 pgd = pgd_offset(vma->vm_mm, addr);
611 next = pgd_addr_end(addr, end);
612 if (pgd_none_or_clear_bad(pgd))
614 if (queue_pages_pud_range(vma, pgd, addr, next, nodes,
617 } while (pgd++, addr = next, addr != end);
621 #ifdef CONFIG_NUMA_BALANCING
623 * This is used to mark a range of virtual addresses to be inaccessible.
624 * These are later cleared by a NUMA hinting fault. Depending on these
625 * faults, pages may be migrated for better NUMA placement.
627 * This is assuming that NUMA faults are handled using PROT_NONE. If
628 * an architecture makes a different choice, it will need further
629 * changes to the core.
631 unsigned long change_prot_numa(struct vm_area_struct *vma,
632 unsigned long addr, unsigned long end)
636 nr_updated = change_protection(vma, addr, end, vma->vm_page_prot, 0, 1);
638 count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
643 static unsigned long change_prot_numa(struct vm_area_struct *vma,
644 unsigned long addr, unsigned long end)
648 #endif /* CONFIG_NUMA_BALANCING */
651 * Walk through page tables and collect pages to be migrated.
653 * If pages found in a given range are on a set of nodes (determined by
654 * @nodes and @flags,) it's isolated and queued to the pagelist which is
655 * passed via @private.)
658 queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
659 const nodemask_t *nodes, unsigned long flags, void *private)
662 struct vm_area_struct *vma, *prev;
664 vma = find_vma(mm, start);
668 for (; vma && vma->vm_start < end; vma = vma->vm_next) {
669 unsigned long endvma = vma->vm_end;
673 if (vma->vm_start > start)
674 start = vma->vm_start;
676 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
677 if (!vma->vm_next && vma->vm_end < end)
679 if (prev && prev->vm_end < vma->vm_start)
683 if (flags & MPOL_MF_LAZY) {
684 /* Similar to task_numa_work, skip inaccessible VMAs */
685 if (vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))
686 change_prot_numa(vma, start, endvma);
690 if ((flags & MPOL_MF_STRICT) ||
691 ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) &&
692 vma_migratable(vma))) {
694 err = queue_pages_pgd_range(vma, start, endvma, nodes,
706 * Apply policy to a single VMA
707 * This must be called with the mmap_sem held for writing.
709 static int vma_replace_policy(struct vm_area_struct *vma,
710 struct mempolicy *pol)
713 struct mempolicy *old;
714 struct mempolicy *new;
716 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
717 vma->vm_start, vma->vm_end, vma->vm_pgoff,
718 vma->vm_ops, vma->vm_file,
719 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
725 if (vma->vm_ops && vma->vm_ops->set_policy) {
726 err = vma->vm_ops->set_policy(vma, new);
731 old = vma->vm_policy;
732 vma->vm_policy = new; /* protected by mmap_sem */
741 /* Step 2: apply policy to a range and do splits. */
742 static int mbind_range(struct mm_struct *mm, unsigned long start,
743 unsigned long end, struct mempolicy *new_pol)
745 struct vm_area_struct *next;
746 struct vm_area_struct *prev;
747 struct vm_area_struct *vma;
750 unsigned long vmstart;
753 vma = find_vma(mm, start);
754 if (!vma || vma->vm_start > start)
758 if (start > vma->vm_start)
761 for (; vma && vma->vm_start < end; prev = vma, vma = next) {
763 vmstart = max(start, vma->vm_start);
764 vmend = min(end, vma->vm_end);
766 if (mpol_equal(vma_policy(vma), new_pol))
769 pgoff = vma->vm_pgoff +
770 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
771 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
772 vma->anon_vma, vma->vm_file, pgoff,
777 if (mpol_equal(vma_policy(vma), new_pol))
779 /* vma_merge() joined vma && vma->next, case 8 */
782 if (vma->vm_start != vmstart) {
783 err = split_vma(vma->vm_mm, vma, vmstart, 1);
787 if (vma->vm_end != vmend) {
788 err = split_vma(vma->vm_mm, vma, vmend, 0);
793 err = vma_replace_policy(vma, new_pol);
802 /* Set the process memory policy */
803 static long do_set_mempolicy(unsigned short mode, unsigned short flags,
806 struct mempolicy *new, *old;
807 NODEMASK_SCRATCH(scratch);
813 new = mpol_new(mode, flags, nodes);
820 ret = mpol_set_nodemask(new, nodes, scratch);
822 task_unlock(current);
826 old = current->mempolicy;
827 current->mempolicy = new;
828 if (new && new->mode == MPOL_INTERLEAVE &&
829 nodes_weight(new->v.nodes))
830 current->il_next = first_node(new->v.nodes);
831 task_unlock(current);
835 NODEMASK_SCRATCH_FREE(scratch);
840 * Return nodemask for policy for get_mempolicy() query
842 * Called with task's alloc_lock held
844 static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
847 if (p == &default_policy)
853 case MPOL_INTERLEAVE:
857 if (!(p->flags & MPOL_F_LOCAL))
858 node_set(p->v.preferred_node, *nodes);
859 /* else return empty node mask for local allocation */
866 static int lookup_node(struct mm_struct *mm, unsigned long addr)
871 err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL);
873 err = page_to_nid(p);
879 /* Retrieve NUMA policy */
880 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
881 unsigned long addr, unsigned long flags)
884 struct mm_struct *mm = current->mm;
885 struct vm_area_struct *vma = NULL;
886 struct mempolicy *pol = current->mempolicy;
889 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
892 if (flags & MPOL_F_MEMS_ALLOWED) {
893 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
895 *policy = 0; /* just so it's initialized */
897 *nmask = cpuset_current_mems_allowed;
898 task_unlock(current);
902 if (flags & MPOL_F_ADDR) {
904 * Do NOT fall back to task policy if the
905 * vma/shared policy at addr is NULL. We
906 * want to return MPOL_DEFAULT in this case.
908 down_read(&mm->mmap_sem);
909 vma = find_vma_intersection(mm, addr, addr+1);
911 up_read(&mm->mmap_sem);
914 if (vma->vm_ops && vma->vm_ops->get_policy)
915 pol = vma->vm_ops->get_policy(vma, addr);
917 pol = vma->vm_policy;
922 pol = &default_policy; /* indicates default behavior */
924 if (flags & MPOL_F_NODE) {
925 if (flags & MPOL_F_ADDR) {
926 err = lookup_node(mm, addr);
930 } else if (pol == current->mempolicy &&
931 pol->mode == MPOL_INTERLEAVE) {
932 *policy = current->il_next;
938 *policy = pol == &default_policy ? MPOL_DEFAULT :
941 * Internal mempolicy flags must be masked off before exposing
942 * the policy to userspace.
944 *policy |= (pol->flags & MPOL_MODE_FLAGS);
948 up_read(¤t->mm->mmap_sem);
954 if (mpol_store_user_nodemask(pol)) {
955 *nmask = pol->w.user_nodemask;
958 get_policy_nodemask(pol, nmask);
959 task_unlock(current);
966 up_read(¤t->mm->mmap_sem);
970 #ifdef CONFIG_MIGRATION
974 static void migrate_page_add(struct page *page, struct list_head *pagelist,
978 * Avoid migrating a page that is shared with others.
980 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) {
981 if (!isolate_lru_page(page)) {
982 list_add_tail(&page->lru, pagelist);
983 inc_zone_page_state(page, NR_ISOLATED_ANON +
984 page_is_file_cache(page));
989 static struct page *new_node_page(struct page *page, unsigned long node, int **x)
992 return alloc_huge_page_node(page_hstate(compound_head(page)),
995 return alloc_pages_exact_node(node, GFP_HIGHUSER_MOVABLE, 0);
999 * Migrate pages from one node to a target node.
1000 * Returns error or the number of pages not migrated.
1002 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
1006 LIST_HEAD(pagelist);
1010 node_set(source, nmask);
1013 * This does not "check" the range but isolates all pages that
1014 * need migration. Between passing in the full user address
1015 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1017 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
1018 queue_pages_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
1019 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
1021 if (!list_empty(&pagelist)) {
1022 err = migrate_pages(&pagelist, new_node_page, NULL, dest,
1023 MIGRATE_SYNC, MR_SYSCALL);
1025 putback_movable_pages(&pagelist);
1032 * Move pages between the two nodesets so as to preserve the physical
1033 * layout as much as possible.
1035 * Returns the number of page that could not be moved.
1037 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1038 const nodemask_t *to, int flags)
1044 err = migrate_prep();
1048 down_read(&mm->mmap_sem);
1051 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1052 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1053 * bit in 'tmp', and return that <source, dest> pair for migration.
1054 * The pair of nodemasks 'to' and 'from' define the map.
1056 * If no pair of bits is found that way, fallback to picking some
1057 * pair of 'source' and 'dest' bits that are not the same. If the
1058 * 'source' and 'dest' bits are the same, this represents a node
1059 * that will be migrating to itself, so no pages need move.
1061 * If no bits are left in 'tmp', or if all remaining bits left
1062 * in 'tmp' correspond to the same bit in 'to', return false
1063 * (nothing left to migrate).
1065 * This lets us pick a pair of nodes to migrate between, such that
1066 * if possible the dest node is not already occupied by some other
1067 * source node, minimizing the risk of overloading the memory on a
1068 * node that would happen if we migrated incoming memory to a node
1069 * before migrating outgoing memory source that same node.
1071 * A single scan of tmp is sufficient. As we go, we remember the
1072 * most recent <s, d> pair that moved (s != d). If we find a pair
1073 * that not only moved, but what's better, moved to an empty slot
1074 * (d is not set in tmp), then we break out then, with that pair.
1075 * Otherwise when we finish scanning from_tmp, we at least have the
1076 * most recent <s, d> pair that moved. If we get all the way through
1077 * the scan of tmp without finding any node that moved, much less
1078 * moved to an empty node, then there is nothing left worth migrating.
1082 while (!nodes_empty(tmp)) {
1084 int source = NUMA_NO_NODE;
1087 for_each_node_mask(s, tmp) {
1090 * do_migrate_pages() tries to maintain the relative
1091 * node relationship of the pages established between
1092 * threads and memory areas.
1094 * However if the number of source nodes is not equal to
1095 * the number of destination nodes we can not preserve
1096 * this node relative relationship. In that case, skip
1097 * copying memory from a node that is in the destination
1100 * Example: [2,3,4] -> [3,4,5] moves everything.
1101 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1104 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1105 (node_isset(s, *to)))
1108 d = node_remap(s, *from, *to);
1112 source = s; /* Node moved. Memorize */
1115 /* dest not in remaining from nodes? */
1116 if (!node_isset(dest, tmp))
1119 if (source == NUMA_NO_NODE)
1122 node_clear(source, tmp);
1123 err = migrate_to_node(mm, source, dest, flags);
1129 up_read(&mm->mmap_sem);
1137 * Allocate a new page for page migration based on vma policy.
1138 * Start by assuming the page is mapped by the same vma as contains @start.
1139 * Search forward from there, if not. N.B., this assumes that the
1140 * list of pages handed to migrate_pages()--which is how we get here--
1141 * is in virtual address order.
1143 static struct page *new_page(struct page *page, unsigned long start, int **x)
1145 struct vm_area_struct *vma;
1146 unsigned long uninitialized_var(address);
1148 vma = find_vma(current->mm, start);
1150 address = page_address_in_vma(page, vma);
1151 if (address != -EFAULT)
1156 if (PageHuge(page)) {
1158 return alloc_huge_page_noerr(vma, address, 1);
1161 * if !vma, alloc_page_vma() will use task or system default policy
1163 return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1167 static void migrate_page_add(struct page *page, struct list_head *pagelist,
1168 unsigned long flags)
1172 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1173 const nodemask_t *to, int flags)
1178 static struct page *new_page(struct page *page, unsigned long start, int **x)
1184 static long do_mbind(unsigned long start, unsigned long len,
1185 unsigned short mode, unsigned short mode_flags,
1186 nodemask_t *nmask, unsigned long flags)
1188 struct mm_struct *mm = current->mm;
1189 struct mempolicy *new;
1192 LIST_HEAD(pagelist);
1194 if (flags & ~(unsigned long)MPOL_MF_VALID)
1196 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1199 if (start & ~PAGE_MASK)
1202 if (mode == MPOL_DEFAULT)
1203 flags &= ~MPOL_MF_STRICT;
1205 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1213 new = mpol_new(mode, mode_flags, nmask);
1215 return PTR_ERR(new);
1217 if (flags & MPOL_MF_LAZY)
1218 new->flags |= MPOL_F_MOF;
1221 * If we are using the default policy then operation
1222 * on discontinuous address spaces is okay after all
1225 flags |= MPOL_MF_DISCONTIG_OK;
1227 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1228 start, start + len, mode, mode_flags,
1229 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
1231 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1233 err = migrate_prep();
1238 NODEMASK_SCRATCH(scratch);
1240 down_write(&mm->mmap_sem);
1242 err = mpol_set_nodemask(new, nmask, scratch);
1243 task_unlock(current);
1245 up_write(&mm->mmap_sem);
1248 NODEMASK_SCRATCH_FREE(scratch);
1253 err = queue_pages_range(mm, start, end, nmask,
1254 flags | MPOL_MF_INVERT, &pagelist);
1256 err = mbind_range(mm, start, end, new);
1261 if (!list_empty(&pagelist)) {
1262 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1263 nr_failed = migrate_pages(&pagelist, new_page, NULL,
1264 start, MIGRATE_SYNC, MR_MEMPOLICY_MBIND);
1266 putback_movable_pages(&pagelist);
1269 if (nr_failed && (flags & MPOL_MF_STRICT))
1272 putback_movable_pages(&pagelist);
1274 up_write(&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, const 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, const 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_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 COMPAT_SYSCALL_DEFINE5(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 COMPAT_SYSCALL_DEFINE3(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 COMPAT_SYSCALL_DEFINE6(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);
1590 struct mempolicy *__get_vma_policy(struct vm_area_struct *vma,
1593 struct mempolicy *pol = NULL;
1596 if (vma->vm_ops && vma->vm_ops->get_policy) {
1597 pol = vma->vm_ops->get_policy(vma, addr);
1598 } else if (vma->vm_policy) {
1599 pol = vma->vm_policy;
1602 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1603 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1604 * count on these policies which will be dropped by
1605 * mpol_cond_put() later
1607 if (mpol_needs_cond_ref(pol))
1616 * get_vma_policy(@vma, @addr)
1617 * @vma: virtual memory area whose policy is sought
1618 * @addr: address in @vma for shared policy lookup
1620 * Returns effective policy for a VMA at specified address.
1621 * Falls back to current->mempolicy or system default policy, as necessary.
1622 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1623 * count--added by the get_policy() vm_op, as appropriate--to protect against
1624 * freeing by another task. It is the caller's responsibility to free the
1625 * extra reference for shared policies.
1627 static struct mempolicy *get_vma_policy(struct vm_area_struct *vma,
1630 struct mempolicy *pol = __get_vma_policy(vma, addr);
1633 pol = get_task_policy(current);
1638 bool vma_policy_mof(struct vm_area_struct *vma)
1640 struct mempolicy *pol;
1642 if (vma->vm_ops && vma->vm_ops->get_policy) {
1645 pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1646 if (pol && (pol->flags & MPOL_F_MOF))
1653 pol = vma->vm_policy;
1655 pol = get_task_policy(current);
1657 return pol->flags & MPOL_F_MOF;
1660 static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1662 enum zone_type dynamic_policy_zone = policy_zone;
1664 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1667 * if policy->v.nodes has movable memory only,
1668 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1670 * policy->v.nodes is intersect with node_states[N_MEMORY].
1671 * so if the following test faile, it implies
1672 * policy->v.nodes has movable memory only.
1674 if (!nodes_intersects(policy->v.nodes, node_states[N_HIGH_MEMORY]))
1675 dynamic_policy_zone = ZONE_MOVABLE;
1677 return zone >= dynamic_policy_zone;
1681 * Return a nodemask representing a mempolicy for filtering nodes for
1684 static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1686 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1687 if (unlikely(policy->mode == MPOL_BIND) &&
1688 apply_policy_zone(policy, gfp_zone(gfp)) &&
1689 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1690 return &policy->v.nodes;
1695 /* Return a zonelist indicated by gfp for node representing a mempolicy */
1696 static struct zonelist *policy_zonelist(gfp_t gfp, struct mempolicy *policy,
1699 switch (policy->mode) {
1700 case MPOL_PREFERRED:
1701 if (!(policy->flags & MPOL_F_LOCAL))
1702 nd = policy->v.preferred_node;
1706 * Normally, MPOL_BIND allocations are node-local within the
1707 * allowed nodemask. However, if __GFP_THISNODE is set and the
1708 * current node isn't part of the mask, we use the zonelist for
1709 * the first node in the mask instead.
1711 if (unlikely(gfp & __GFP_THISNODE) &&
1712 unlikely(!node_isset(nd, policy->v.nodes)))
1713 nd = first_node(policy->v.nodes);
1718 return node_zonelist(nd, gfp);
1721 /* Do dynamic interleaving for a process */
1722 static unsigned interleave_nodes(struct mempolicy *policy)
1725 struct task_struct *me = current;
1728 next = next_node(nid, policy->v.nodes);
1729 if (next >= MAX_NUMNODES)
1730 next = first_node(policy->v.nodes);
1731 if (next < MAX_NUMNODES)
1737 * Depending on the memory policy provide a node from which to allocate the
1740 unsigned int mempolicy_slab_node(void)
1742 struct mempolicy *policy;
1743 int node = numa_mem_id();
1748 policy = current->mempolicy;
1749 if (!policy || policy->flags & MPOL_F_LOCAL)
1752 switch (policy->mode) {
1753 case MPOL_PREFERRED:
1755 * handled MPOL_F_LOCAL above
1757 return policy->v.preferred_node;
1759 case MPOL_INTERLEAVE:
1760 return interleave_nodes(policy);
1764 * Follow bind policy behavior and start allocation at the
1767 struct zonelist *zonelist;
1769 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1770 zonelist = &NODE_DATA(node)->node_zonelists[0];
1771 (void)first_zones_zonelist(zonelist, highest_zoneidx,
1774 return zone ? zone->node : node;
1782 /* Do static interleaving for a VMA with known offset. */
1783 static unsigned offset_il_node(struct mempolicy *pol,
1784 struct vm_area_struct *vma, unsigned long off)
1786 unsigned nnodes = nodes_weight(pol->v.nodes);
1789 int nid = NUMA_NO_NODE;
1792 return numa_node_id();
1793 target = (unsigned int)off % nnodes;
1796 nid = next_node(nid, pol->v.nodes);
1798 } while (c <= target);
1802 /* Determine a node number for interleave */
1803 static inline unsigned interleave_nid(struct mempolicy *pol,
1804 struct vm_area_struct *vma, unsigned long addr, int shift)
1810 * for small pages, there is no difference between
1811 * shift and PAGE_SHIFT, so the bit-shift is safe.
1812 * for huge pages, since vm_pgoff is in units of small
1813 * pages, we need to shift off the always 0 bits to get
1816 BUG_ON(shift < PAGE_SHIFT);
1817 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1818 off += (addr - vma->vm_start) >> shift;
1819 return offset_il_node(pol, vma, off);
1821 return interleave_nodes(pol);
1825 * Return the bit number of a random bit set in the nodemask.
1826 * (returns NUMA_NO_NODE if nodemask is empty)
1828 int node_random(const nodemask_t *maskp)
1830 int w, bit = NUMA_NO_NODE;
1832 w = nodes_weight(*maskp);
1834 bit = bitmap_ord_to_pos(maskp->bits,
1835 get_random_int() % w, MAX_NUMNODES);
1839 #ifdef CONFIG_HUGETLBFS
1841 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1842 * @vma: virtual memory area whose policy is sought
1843 * @addr: address in @vma for shared policy lookup and interleave policy
1844 * @gfp_flags: for requested zone
1845 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
1846 * @nodemask: pointer to nodemask pointer for MPOL_BIND nodemask
1848 * Returns a zonelist suitable for a huge page allocation and a pointer
1849 * to the struct mempolicy for conditional unref after allocation.
1850 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1851 * @nodemask for filtering the zonelist.
1853 * Must be protected by read_mems_allowed_begin()
1855 struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
1856 gfp_t gfp_flags, struct mempolicy **mpol,
1857 nodemask_t **nodemask)
1859 struct zonelist *zl;
1861 *mpol = get_vma_policy(vma, addr);
1862 *nodemask = NULL; /* assume !MPOL_BIND */
1864 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1865 zl = node_zonelist(interleave_nid(*mpol, vma, addr,
1866 huge_page_shift(hstate_vma(vma))), gfp_flags);
1868 zl = policy_zonelist(gfp_flags, *mpol, numa_node_id());
1869 if ((*mpol)->mode == MPOL_BIND)
1870 *nodemask = &(*mpol)->v.nodes;
1876 * init_nodemask_of_mempolicy
1878 * If the current task's mempolicy is "default" [NULL], return 'false'
1879 * to indicate default policy. Otherwise, extract the policy nodemask
1880 * for 'bind' or 'interleave' policy into the argument nodemask, or
1881 * initialize the argument nodemask to contain the single node for
1882 * 'preferred' or 'local' policy and return 'true' to indicate presence
1883 * of non-default mempolicy.
1885 * We don't bother with reference counting the mempolicy [mpol_get/put]
1886 * because the current task is examining it's own mempolicy and a task's
1887 * mempolicy is only ever changed by the task itself.
1889 * N.B., it is the caller's responsibility to free a returned nodemask.
1891 bool init_nodemask_of_mempolicy(nodemask_t *mask)
1893 struct mempolicy *mempolicy;
1896 if (!(mask && current->mempolicy))
1900 mempolicy = current->mempolicy;
1901 switch (mempolicy->mode) {
1902 case MPOL_PREFERRED:
1903 if (mempolicy->flags & MPOL_F_LOCAL)
1904 nid = numa_node_id();
1906 nid = mempolicy->v.preferred_node;
1907 init_nodemask_of_node(mask, nid);
1912 case MPOL_INTERLEAVE:
1913 *mask = mempolicy->v.nodes;
1919 task_unlock(current);
1926 * mempolicy_nodemask_intersects
1928 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1929 * policy. Otherwise, check for intersection between mask and the policy
1930 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1931 * policy, always return true since it may allocate elsewhere on fallback.
1933 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1935 bool mempolicy_nodemask_intersects(struct task_struct *tsk,
1936 const nodemask_t *mask)
1938 struct mempolicy *mempolicy;
1944 mempolicy = tsk->mempolicy;
1948 switch (mempolicy->mode) {
1949 case MPOL_PREFERRED:
1951 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1952 * allocate from, they may fallback to other nodes when oom.
1953 * Thus, it's possible for tsk to have allocated memory from
1958 case MPOL_INTERLEAVE:
1959 ret = nodes_intersects(mempolicy->v.nodes, *mask);
1969 /* Allocate a page in interleaved policy.
1970 Own path because it needs to do special accounting. */
1971 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1974 struct zonelist *zl;
1977 zl = node_zonelist(nid, gfp);
1978 page = __alloc_pages(gfp, order, zl);
1979 if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0]))
1980 inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
1985 * alloc_pages_vma - Allocate a page for a VMA.
1988 * %GFP_USER user allocation.
1989 * %GFP_KERNEL kernel allocations,
1990 * %GFP_HIGHMEM highmem/user allocations,
1991 * %GFP_FS allocation should not call back into a file system.
1992 * %GFP_ATOMIC don't sleep.
1994 * @order:Order of the GFP allocation.
1995 * @vma: Pointer to VMA or NULL if not available.
1996 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1998 * This function allocates a page from the kernel page pool and applies
1999 * a NUMA policy associated with the VMA or the current process.
2000 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
2001 * mm_struct of the VMA to prevent it from going away. Should be used for
2002 * all allocations for pages that will be mapped into
2003 * user space. Returns NULL when no page can be allocated.
2005 * Should be called with the mm_sem of the vma hold.
2008 alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
2009 unsigned long addr, int node)
2011 struct mempolicy *pol;
2013 unsigned int cpuset_mems_cookie;
2016 pol = get_vma_policy(vma, addr);
2017 cpuset_mems_cookie = read_mems_allowed_begin();
2019 if (unlikely(pol->mode == MPOL_INTERLEAVE)) {
2022 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
2024 page = alloc_page_interleave(gfp, order, nid);
2025 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
2030 page = __alloc_pages_nodemask(gfp, order,
2031 policy_zonelist(gfp, pol, node),
2032 policy_nodemask(gfp, pol));
2034 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
2040 * alloc_pages_current - Allocate pages.
2043 * %GFP_USER user allocation,
2044 * %GFP_KERNEL kernel allocation,
2045 * %GFP_HIGHMEM highmem allocation,
2046 * %GFP_FS don't call back into a file system.
2047 * %GFP_ATOMIC don't sleep.
2048 * @order: Power of two of allocation size in pages. 0 is a single page.
2050 * Allocate a page from the kernel page pool. When not in
2051 * interrupt context and apply the current process NUMA policy.
2052 * Returns NULL when no page can be allocated.
2054 * Don't call cpuset_update_task_memory_state() unless
2055 * 1) it's ok to take cpuset_sem (can WAIT), and
2056 * 2) allocating for current task (not interrupt).
2058 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
2060 struct mempolicy *pol = &default_policy;
2062 unsigned int cpuset_mems_cookie;
2064 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2065 pol = get_task_policy(current);
2068 cpuset_mems_cookie = read_mems_allowed_begin();
2071 * No reference counting needed for current->mempolicy
2072 * nor system default_policy
2074 if (pol->mode == MPOL_INTERLEAVE)
2075 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2077 page = __alloc_pages_nodemask(gfp, order,
2078 policy_zonelist(gfp, pol, numa_node_id()),
2079 policy_nodemask(gfp, pol));
2081 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
2086 EXPORT_SYMBOL(alloc_pages_current);
2088 int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2090 struct mempolicy *pol = mpol_dup(vma_policy(src));
2093 return PTR_ERR(pol);
2094 dst->vm_policy = pol;
2099 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2100 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2101 * with the mems_allowed returned by cpuset_mems_allowed(). This
2102 * keeps mempolicies cpuset relative after its cpuset moves. See
2103 * further kernel/cpuset.c update_nodemask().
2105 * current's mempolicy may be rebinded by the other task(the task that changes
2106 * cpuset's mems), so we needn't do rebind work for current task.
2109 /* Slow path of a mempolicy duplicate */
2110 struct mempolicy *__mpol_dup(struct mempolicy *old)
2112 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2115 return ERR_PTR(-ENOMEM);
2117 /* task's mempolicy is protected by alloc_lock */
2118 if (old == current->mempolicy) {
2121 task_unlock(current);
2125 if (current_cpuset_is_being_rebound()) {
2126 nodemask_t mems = cpuset_mems_allowed(current);
2127 if (new->flags & MPOL_F_REBINDING)
2128 mpol_rebind_policy(new, &mems, MPOL_REBIND_STEP2);
2130 mpol_rebind_policy(new, &mems, MPOL_REBIND_ONCE);
2132 atomic_set(&new->refcnt, 1);
2136 /* Slow path of a mempolicy comparison */
2137 bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2141 if (a->mode != b->mode)
2143 if (a->flags != b->flags)
2145 if (mpol_store_user_nodemask(a))
2146 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2152 case MPOL_INTERLEAVE:
2153 return !!nodes_equal(a->v.nodes, b->v.nodes);
2154 case MPOL_PREFERRED:
2155 return a->v.preferred_node == b->v.preferred_node;
2163 * Shared memory backing store policy support.
2165 * Remember policies even when nobody has shared memory mapped.
2166 * The policies are kept in Red-Black tree linked from the inode.
2167 * They are protected by the sp->lock spinlock, which should be held
2168 * for any accesses to the tree.
2171 /* lookup first element intersecting start-end */
2172 /* Caller holds sp->lock */
2173 static struct sp_node *
2174 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2176 struct rb_node *n = sp->root.rb_node;
2179 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2181 if (start >= p->end)
2183 else if (end <= p->start)
2191 struct sp_node *w = NULL;
2192 struct rb_node *prev = rb_prev(n);
2195 w = rb_entry(prev, struct sp_node, nd);
2196 if (w->end <= start)
2200 return rb_entry(n, struct sp_node, nd);
2203 /* Insert a new shared policy into the list. */
2204 /* Caller holds sp->lock */
2205 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2207 struct rb_node **p = &sp->root.rb_node;
2208 struct rb_node *parent = NULL;
2213 nd = rb_entry(parent, struct sp_node, nd);
2214 if (new->start < nd->start)
2216 else if (new->end > nd->end)
2217 p = &(*p)->rb_right;
2221 rb_link_node(&new->nd, parent, p);
2222 rb_insert_color(&new->nd, &sp->root);
2223 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2224 new->policy ? new->policy->mode : 0);
2227 /* Find shared policy intersecting idx */
2229 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2231 struct mempolicy *pol = NULL;
2234 if (!sp->root.rb_node)
2236 spin_lock(&sp->lock);
2237 sn = sp_lookup(sp, idx, idx+1);
2239 mpol_get(sn->policy);
2242 spin_unlock(&sp->lock);
2246 static void sp_free(struct sp_node *n)
2248 mpol_put(n->policy);
2249 kmem_cache_free(sn_cache, n);
2253 * mpol_misplaced - check whether current page node is valid in policy
2255 * @page: page to be checked
2256 * @vma: vm area where page mapped
2257 * @addr: virtual address where page mapped
2259 * Lookup current policy node id for vma,addr and "compare to" page's
2263 * -1 - not misplaced, page is in the right node
2264 * node - node id where the page should be
2266 * Policy determination "mimics" alloc_page_vma().
2267 * Called from fault path where we know the vma and faulting address.
2269 int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2271 struct mempolicy *pol;
2273 int curnid = page_to_nid(page);
2274 unsigned long pgoff;
2275 int thiscpu = raw_smp_processor_id();
2276 int thisnid = cpu_to_node(thiscpu);
2282 pol = get_vma_policy(vma, addr);
2283 if (!(pol->flags & MPOL_F_MOF))
2286 switch (pol->mode) {
2287 case MPOL_INTERLEAVE:
2288 BUG_ON(addr >= vma->vm_end);
2289 BUG_ON(addr < vma->vm_start);
2291 pgoff = vma->vm_pgoff;
2292 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2293 polnid = offset_il_node(pol, vma, pgoff);
2296 case MPOL_PREFERRED:
2297 if (pol->flags & MPOL_F_LOCAL)
2298 polnid = numa_node_id();
2300 polnid = pol->v.preferred_node;
2305 * allows binding to multiple nodes.
2306 * use current page if in policy nodemask,
2307 * else select nearest allowed node, if any.
2308 * If no allowed nodes, use current [!misplaced].
2310 if (node_isset(curnid, pol->v.nodes))
2312 (void)first_zones_zonelist(
2313 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2314 gfp_zone(GFP_HIGHUSER),
2315 &pol->v.nodes, &zone);
2316 polnid = zone->node;
2323 /* Migrate the page towards the node whose CPU is referencing it */
2324 if (pol->flags & MPOL_F_MORON) {
2327 if (!should_numa_migrate_memory(current, page, curnid, thiscpu))
2331 if (curnid != polnid)
2339 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2341 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2342 rb_erase(&n->nd, &sp->root);
2346 static void sp_node_init(struct sp_node *node, unsigned long start,
2347 unsigned long end, struct mempolicy *pol)
2349 node->start = start;
2354 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2355 struct mempolicy *pol)
2358 struct mempolicy *newpol;
2360 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2364 newpol = mpol_dup(pol);
2365 if (IS_ERR(newpol)) {
2366 kmem_cache_free(sn_cache, n);
2369 newpol->flags |= MPOL_F_SHARED;
2370 sp_node_init(n, start, end, newpol);
2375 /* Replace a policy range. */
2376 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2377 unsigned long end, struct sp_node *new)
2380 struct sp_node *n_new = NULL;
2381 struct mempolicy *mpol_new = NULL;
2385 spin_lock(&sp->lock);
2386 n = sp_lookup(sp, start, end);
2387 /* Take care of old policies in the same range. */
2388 while (n && n->start < end) {
2389 struct rb_node *next = rb_next(&n->nd);
2390 if (n->start >= start) {
2396 /* Old policy spanning whole new range. */
2401 *mpol_new = *n->policy;
2402 atomic_set(&mpol_new->refcnt, 1);
2403 sp_node_init(n_new, end, n->end, mpol_new);
2405 sp_insert(sp, n_new);
2414 n = rb_entry(next, struct sp_node, nd);
2418 spin_unlock(&sp->lock);
2425 kmem_cache_free(sn_cache, n_new);
2430 spin_unlock(&sp->lock);
2432 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2435 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2442 * mpol_shared_policy_init - initialize shared policy for inode
2443 * @sp: pointer to inode shared policy
2444 * @mpol: struct mempolicy to install
2446 * Install non-NULL @mpol in inode's shared policy rb-tree.
2447 * On entry, the current task has a reference on a non-NULL @mpol.
2448 * This must be released on exit.
2449 * This is called at get_inode() calls and we can use GFP_KERNEL.
2451 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2455 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2456 spin_lock_init(&sp->lock);
2459 struct vm_area_struct pvma;
2460 struct mempolicy *new;
2461 NODEMASK_SCRATCH(scratch);
2465 /* contextualize the tmpfs mount point mempolicy */
2466 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2468 goto free_scratch; /* no valid nodemask intersection */
2471 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2472 task_unlock(current);
2476 /* Create pseudo-vma that contains just the policy */
2477 memset(&pvma, 0, sizeof(struct vm_area_struct));
2478 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2479 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2482 mpol_put(new); /* drop initial ref */
2484 NODEMASK_SCRATCH_FREE(scratch);
2486 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2490 int mpol_set_shared_policy(struct shared_policy *info,
2491 struct vm_area_struct *vma, struct mempolicy *npol)
2494 struct sp_node *new = NULL;
2495 unsigned long sz = vma_pages(vma);
2497 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2499 sz, npol ? npol->mode : -1,
2500 npol ? npol->flags : -1,
2501 npol ? nodes_addr(npol->v.nodes)[0] : NUMA_NO_NODE);
2504 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2508 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2514 /* Free a backing policy store on inode delete. */
2515 void mpol_free_shared_policy(struct shared_policy *p)
2518 struct rb_node *next;
2520 if (!p->root.rb_node)
2522 spin_lock(&p->lock);
2523 next = rb_first(&p->root);
2525 n = rb_entry(next, struct sp_node, nd);
2526 next = rb_next(&n->nd);
2529 spin_unlock(&p->lock);
2532 #ifdef CONFIG_NUMA_BALANCING
2533 static int __initdata numabalancing_override;
2535 static void __init check_numabalancing_enable(void)
2537 bool numabalancing_default = false;
2539 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2540 numabalancing_default = true;
2542 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2543 if (numabalancing_override)
2544 set_numabalancing_state(numabalancing_override == 1);
2546 if (nr_node_ids > 1 && !numabalancing_override) {
2547 pr_info("%s automatic NUMA balancing. "
2548 "Configure with numa_balancing= or the "
2549 "kernel.numa_balancing sysctl",
2550 numabalancing_default ? "Enabling" : "Disabling");
2551 set_numabalancing_state(numabalancing_default);
2555 static int __init setup_numabalancing(char *str)
2561 if (!strcmp(str, "enable")) {
2562 numabalancing_override = 1;
2564 } else if (!strcmp(str, "disable")) {
2565 numabalancing_override = -1;
2570 pr_warn("Unable to parse numa_balancing=\n");
2574 __setup("numa_balancing=", setup_numabalancing);
2576 static inline void __init check_numabalancing_enable(void)
2579 #endif /* CONFIG_NUMA_BALANCING */
2581 /* assumes fs == KERNEL_DS */
2582 void __init numa_policy_init(void)
2584 nodemask_t interleave_nodes;
2585 unsigned long largest = 0;
2586 int nid, prefer = 0;
2588 policy_cache = kmem_cache_create("numa_policy",
2589 sizeof(struct mempolicy),
2590 0, SLAB_PANIC, NULL);
2592 sn_cache = kmem_cache_create("shared_policy_node",
2593 sizeof(struct sp_node),
2594 0, SLAB_PANIC, NULL);
2596 for_each_node(nid) {
2597 preferred_node_policy[nid] = (struct mempolicy) {
2598 .refcnt = ATOMIC_INIT(1),
2599 .mode = MPOL_PREFERRED,
2600 .flags = MPOL_F_MOF | MPOL_F_MORON,
2601 .v = { .preferred_node = nid, },
2606 * Set interleaving policy for system init. Interleaving is only
2607 * enabled across suitably sized nodes (default is >= 16MB), or
2608 * fall back to the largest node if they're all smaller.
2610 nodes_clear(interleave_nodes);
2611 for_each_node_state(nid, N_MEMORY) {
2612 unsigned long total_pages = node_present_pages(nid);
2614 /* Preserve the largest node */
2615 if (largest < total_pages) {
2616 largest = total_pages;
2620 /* Interleave this node? */
2621 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2622 node_set(nid, interleave_nodes);
2625 /* All too small, use the largest */
2626 if (unlikely(nodes_empty(interleave_nodes)))
2627 node_set(prefer, interleave_nodes);
2629 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2630 pr_err("%s: interleaving failed\n", __func__);
2632 check_numabalancing_enable();
2635 /* Reset policy of current process to default */
2636 void numa_default_policy(void)
2638 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2642 * Parse and format mempolicy from/to strings
2646 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
2648 static const char * const policy_modes[] =
2650 [MPOL_DEFAULT] = "default",
2651 [MPOL_PREFERRED] = "prefer",
2652 [MPOL_BIND] = "bind",
2653 [MPOL_INTERLEAVE] = "interleave",
2654 [MPOL_LOCAL] = "local",
2660 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2661 * @str: string containing mempolicy to parse
2662 * @mpol: pointer to struct mempolicy pointer, returned on success.
2665 * <mode>[=<flags>][:<nodelist>]
2667 * On success, returns 0, else 1
2669 int mpol_parse_str(char *str, struct mempolicy **mpol)
2671 struct mempolicy *new = NULL;
2672 unsigned short mode;
2673 unsigned short mode_flags;
2675 char *nodelist = strchr(str, ':');
2676 char *flags = strchr(str, '=');
2680 /* NUL-terminate mode or flags string */
2682 if (nodelist_parse(nodelist, nodes))
2684 if (!nodes_subset(nodes, node_states[N_MEMORY]))
2690 *flags++ = '\0'; /* terminate mode string */
2692 for (mode = 0; mode < MPOL_MAX; mode++) {
2693 if (!strcmp(str, policy_modes[mode])) {
2697 if (mode >= MPOL_MAX)
2701 case MPOL_PREFERRED:
2703 * Insist on a nodelist of one node only
2706 char *rest = nodelist;
2707 while (isdigit(*rest))
2713 case MPOL_INTERLEAVE:
2715 * Default to online nodes with memory if no nodelist
2718 nodes = node_states[N_MEMORY];
2722 * Don't allow a nodelist; mpol_new() checks flags
2726 mode = MPOL_PREFERRED;
2730 * Insist on a empty nodelist
2737 * Insist on a nodelist
2746 * Currently, we only support two mutually exclusive
2749 if (!strcmp(flags, "static"))
2750 mode_flags |= MPOL_F_STATIC_NODES;
2751 else if (!strcmp(flags, "relative"))
2752 mode_flags |= MPOL_F_RELATIVE_NODES;
2757 new = mpol_new(mode, mode_flags, &nodes);
2762 * Save nodes for mpol_to_str() to show the tmpfs mount options
2763 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2765 if (mode != MPOL_PREFERRED)
2766 new->v.nodes = nodes;
2768 new->v.preferred_node = first_node(nodes);
2770 new->flags |= MPOL_F_LOCAL;
2773 * Save nodes for contextualization: this will be used to "clone"
2774 * the mempolicy in a specific context [cpuset] at a later time.
2776 new->w.user_nodemask = nodes;
2781 /* Restore string for error message */
2790 #endif /* CONFIG_TMPFS */
2793 * mpol_to_str - format a mempolicy structure for printing
2794 * @buffer: to contain formatted mempolicy string
2795 * @maxlen: length of @buffer
2796 * @pol: pointer to mempolicy to be formatted
2798 * Convert @pol into a string. If @buffer is too short, truncate the string.
2799 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
2800 * longest flag, "relative", and to display at least a few node ids.
2802 void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
2805 nodemask_t nodes = NODE_MASK_NONE;
2806 unsigned short mode = MPOL_DEFAULT;
2807 unsigned short flags = 0;
2809 if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
2817 case MPOL_PREFERRED:
2818 if (flags & MPOL_F_LOCAL)
2821 node_set(pol->v.preferred_node, nodes);
2824 case MPOL_INTERLEAVE:
2825 nodes = pol->v.nodes;
2829 snprintf(p, maxlen, "unknown");
2833 p += snprintf(p, maxlen, "%s", policy_modes[mode]);
2835 if (flags & MPOL_MODE_FLAGS) {
2836 p += snprintf(p, buffer + maxlen - p, "=");
2839 * Currently, the only defined flags are mutually exclusive
2841 if (flags & MPOL_F_STATIC_NODES)
2842 p += snprintf(p, buffer + maxlen - p, "static");
2843 else if (flags & MPOL_F_RELATIVE_NODES)
2844 p += snprintf(p, buffer + maxlen - p, "relative");
2847 if (!nodes_empty(nodes)) {
2848 p += snprintf(p, buffer + maxlen - p, ":");
2849 p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);