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 * Here we search for not shared page mappings (mapcount == 1) and we
572 * set up the pmd/pte_numa on those mappings so the very next access
573 * will fire a NUMA hinting page fault.
576 change_prot_numa_range(struct mm_struct *mm, struct vm_area_struct *vma,
577 unsigned long address)
584 unsigned long _address, end;
588 VM_BUG_ON(address & ~PAGE_MASK);
590 pgd = pgd_offset(mm, address);
591 if (!pgd_present(*pgd))
594 pud = pud_offset(pgd, address);
595 if (!pud_present(*pud))
598 pmd = pmd_offset(pud, address);
602 if (pmd_trans_huge_lock(pmd, vma) == 1) {
606 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
608 if (pmd_numa(*pmd)) {
609 spin_unlock(&mm->page_table_lock);
613 page = pmd_page(*pmd);
615 /* only check non-shared pages */
616 if (page_mapcount(page) != 1) {
617 spin_unlock(&mm->page_table_lock);
621 page_nid = page_to_nid(page);
623 if (pmd_numa(*pmd)) {
624 spin_unlock(&mm->page_table_lock);
628 set_pmd_at(mm, address, pmd, pmd_mknuma(*pmd));
630 /* defer TLB flush to lower the overhead */
631 spin_unlock(&mm->page_table_lock);
635 if (pmd_trans_unstable(pmd))
637 VM_BUG_ON(!pmd_present(*pmd));
639 end = min(vma->vm_end, (address + PMD_SIZE) & PMD_MASK);
640 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
641 for (_address = address, _pte = pte; _address < end;
642 _pte++, _address += PAGE_SIZE) {
643 pte_t pteval = *_pte;
644 if (!pte_present(pteval))
646 if (pte_numa(pteval))
648 page = vm_normal_page(vma, _address, pteval);
651 /* only check non-shared pages */
652 if (page_mapcount(page) != 1)
655 set_pte_at(mm, _address, _pte, pte_mknuma(pteval));
657 /* defer TLB flush to lower the overhead */
660 pte_unmap_unlock(pte, ptl);
662 if (ret && !pmd_numa(*pmd)) {
663 spin_lock(&mm->page_table_lock);
664 set_pmd_at(mm, address, pmd, pmd_mknuma(*pmd));
665 spin_unlock(&mm->page_table_lock);
666 /* defer TLB flush to lower the overhead */
673 /* Assumes mmap_sem is held */
675 change_prot_numa(struct vm_area_struct *vma,
676 unsigned long address, unsigned long end)
678 struct mm_struct *mm = vma->vm_mm;
681 while (address < end) {
682 VM_BUG_ON(address < vma->vm_start ||
683 address + PAGE_SIZE > vma->vm_end);
685 progress += change_prot_numa_range(mm, vma, address);
686 address = (address + PMD_SIZE) & PMD_MASK;
690 * Flush the TLB for the mm to start the NUMA hinting
691 * page faults after we finish scanning this vma part
692 * if there were any PTE updates
695 mmu_notifier_invalidate_range_start(vma->vm_mm, address, end);
696 flush_tlb_range(vma, address, end);
697 mmu_notifier_invalidate_range_end(vma->vm_mm, address, end);
701 static unsigned long change_prot_numa(struct vm_area_struct *vma,
702 unsigned long addr, unsigned long end)
706 #endif /* CONFIG_ARCH_USES_NUMA_PROT_NONE */
709 * Check if all pages in a range are on a set of nodes.
710 * If pagelist != NULL then isolate pages from the LRU and
711 * put them on the pagelist.
713 static struct vm_area_struct *
714 check_range(struct mm_struct *mm, unsigned long start, unsigned long end,
715 const nodemask_t *nodes, unsigned long flags, void *private)
718 struct vm_area_struct *first, *vma, *prev;
721 first = find_vma(mm, start);
723 return ERR_PTR(-EFAULT);
725 for (vma = first; vma && vma->vm_start < end; vma = vma->vm_next) {
726 unsigned long endvma = vma->vm_end;
730 if (vma->vm_start > start)
731 start = vma->vm_start;
733 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
734 if (!vma->vm_next && vma->vm_end < end)
735 return ERR_PTR(-EFAULT);
736 if (prev && prev->vm_end < vma->vm_start)
737 return ERR_PTR(-EFAULT);
740 if (is_vm_hugetlb_page(vma))
743 if (flags & MPOL_MF_LAZY) {
744 change_prot_numa(vma, start, endvma);
748 if ((flags & MPOL_MF_STRICT) ||
749 ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) &&
750 vma_migratable(vma))) {
752 err = check_pgd_range(vma, start, endvma, nodes,
755 first = ERR_PTR(err);
766 * Apply policy to a single VMA
767 * This must be called with the mmap_sem held for writing.
769 static int vma_replace_policy(struct vm_area_struct *vma,
770 struct mempolicy *pol)
773 struct mempolicy *old;
774 struct mempolicy *new;
776 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
777 vma->vm_start, vma->vm_end, vma->vm_pgoff,
778 vma->vm_ops, vma->vm_file,
779 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
785 if (vma->vm_ops && vma->vm_ops->set_policy) {
786 err = vma->vm_ops->set_policy(vma, new);
791 old = vma->vm_policy;
792 vma->vm_policy = new; /* protected by mmap_sem */
801 /* Step 2: apply policy to a range and do splits. */
802 static int mbind_range(struct mm_struct *mm, unsigned long start,
803 unsigned long end, struct mempolicy *new_pol)
805 struct vm_area_struct *next;
806 struct vm_area_struct *prev;
807 struct vm_area_struct *vma;
810 unsigned long vmstart;
813 vma = find_vma(mm, start);
814 if (!vma || vma->vm_start > start)
818 if (start > vma->vm_start)
821 for (; vma && vma->vm_start < end; prev = vma, vma = next) {
823 vmstart = max(start, vma->vm_start);
824 vmend = min(end, vma->vm_end);
826 if (mpol_equal(vma_policy(vma), new_pol))
829 pgoff = vma->vm_pgoff +
830 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
831 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
832 vma->anon_vma, vma->vm_file, pgoff,
839 if (vma->vm_start != vmstart) {
840 err = split_vma(vma->vm_mm, vma, vmstart, 1);
844 if (vma->vm_end != vmend) {
845 err = split_vma(vma->vm_mm, vma, vmend, 0);
849 err = vma_replace_policy(vma, new_pol);
859 * Update task->flags PF_MEMPOLICY bit: set iff non-default
860 * mempolicy. Allows more rapid checking of this (combined perhaps
861 * with other PF_* flag bits) on memory allocation hot code paths.
863 * If called from outside this file, the task 'p' should -only- be
864 * a newly forked child not yet visible on the task list, because
865 * manipulating the task flags of a visible task is not safe.
867 * The above limitation is why this routine has the funny name
868 * mpol_fix_fork_child_flag().
870 * It is also safe to call this with a task pointer of current,
871 * which the static wrapper mpol_set_task_struct_flag() does,
872 * for use within this file.
875 void mpol_fix_fork_child_flag(struct task_struct *p)
878 p->flags |= PF_MEMPOLICY;
880 p->flags &= ~PF_MEMPOLICY;
883 static void mpol_set_task_struct_flag(void)
885 mpol_fix_fork_child_flag(current);
888 /* Set the process memory policy */
889 static long do_set_mempolicy(unsigned short mode, unsigned short flags,
892 struct mempolicy *new, *old;
893 struct mm_struct *mm = current->mm;
894 NODEMASK_SCRATCH(scratch);
900 new = mpol_new(mode, flags, nodes);
906 * prevent changing our mempolicy while show_numa_maps()
908 * Note: do_set_mempolicy() can be called at init time
912 down_write(&mm->mmap_sem);
914 ret = mpol_set_nodemask(new, nodes, scratch);
916 task_unlock(current);
918 up_write(&mm->mmap_sem);
922 old = current->mempolicy;
923 current->mempolicy = new;
924 mpol_set_task_struct_flag();
925 if (new && new->mode == MPOL_INTERLEAVE &&
926 nodes_weight(new->v.nodes))
927 current->il_next = first_node(new->v.nodes);
928 task_unlock(current);
930 up_write(&mm->mmap_sem);
935 NODEMASK_SCRATCH_FREE(scratch);
940 * Return nodemask for policy for get_mempolicy() query
942 * Called with task's alloc_lock held
944 static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
947 if (p == &default_policy)
953 case MPOL_INTERLEAVE:
957 if (!(p->flags & MPOL_F_LOCAL))
958 node_set(p->v.preferred_node, *nodes);
959 /* else return empty node mask for local allocation */
966 static int lookup_node(struct mm_struct *mm, unsigned long addr)
971 err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL);
973 err = page_to_nid(p);
979 /* Retrieve NUMA policy */
980 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
981 unsigned long addr, unsigned long flags)
984 struct mm_struct *mm = current->mm;
985 struct vm_area_struct *vma = NULL;
986 struct mempolicy *pol = current->mempolicy;
989 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
992 if (flags & MPOL_F_MEMS_ALLOWED) {
993 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
995 *policy = 0; /* just so it's initialized */
997 *nmask = cpuset_current_mems_allowed;
998 task_unlock(current);
1002 if (flags & MPOL_F_ADDR) {
1004 * Do NOT fall back to task policy if the
1005 * vma/shared policy at addr is NULL. We
1006 * want to return MPOL_DEFAULT in this case.
1008 down_read(&mm->mmap_sem);
1009 vma = find_vma_intersection(mm, addr, addr+1);
1011 up_read(&mm->mmap_sem);
1014 if (vma->vm_ops && vma->vm_ops->get_policy)
1015 pol = vma->vm_ops->get_policy(vma, addr);
1017 pol = vma->vm_policy;
1022 pol = &default_policy; /* indicates default behavior */
1024 if (flags & MPOL_F_NODE) {
1025 if (flags & MPOL_F_ADDR) {
1026 err = lookup_node(mm, addr);
1030 } else if (pol == current->mempolicy &&
1031 pol->mode == MPOL_INTERLEAVE) {
1032 *policy = current->il_next;
1038 *policy = pol == &default_policy ? MPOL_DEFAULT :
1041 * Internal mempolicy flags must be masked off before exposing
1042 * the policy to userspace.
1044 *policy |= (pol->flags & MPOL_MODE_FLAGS);
1048 up_read(¤t->mm->mmap_sem);
1054 if (mpol_store_user_nodemask(pol)) {
1055 *nmask = pol->w.user_nodemask;
1058 get_policy_nodemask(pol, nmask);
1059 task_unlock(current);
1066 up_read(¤t->mm->mmap_sem);
1070 #ifdef CONFIG_MIGRATION
1074 static void migrate_page_add(struct page *page, struct list_head *pagelist,
1075 unsigned long flags)
1078 * Avoid migrating a page that is shared with others.
1080 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) {
1081 if (!isolate_lru_page(page)) {
1082 list_add_tail(&page->lru, pagelist);
1083 inc_zone_page_state(page, NR_ISOLATED_ANON +
1084 page_is_file_cache(page));
1089 static struct page *new_node_page(struct page *page, unsigned long node, int **x)
1091 return alloc_pages_exact_node(node, GFP_HIGHUSER_MOVABLE, 0);
1095 * Migrate pages from one node to a target node.
1096 * Returns error or the number of pages not migrated.
1098 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
1102 LIST_HEAD(pagelist);
1106 node_set(source, nmask);
1109 * This does not "check" the range but isolates all pages that
1110 * need migration. Between passing in the full user address
1111 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1113 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
1114 check_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
1115 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
1117 if (!list_empty(&pagelist)) {
1118 err = migrate_pages(&pagelist, new_node_page, dest,
1119 false, MIGRATE_SYNC,
1122 putback_lru_pages(&pagelist);
1129 * Move pages between the two nodesets so as to preserve the physical
1130 * layout as much as possible.
1132 * Returns the number of page that could not be moved.
1134 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1135 const nodemask_t *to, int flags)
1141 err = migrate_prep();
1145 down_read(&mm->mmap_sem);
1147 err = migrate_vmas(mm, from, to, flags);
1152 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1153 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1154 * bit in 'tmp', and return that <source, dest> pair for migration.
1155 * The pair of nodemasks 'to' and 'from' define the map.
1157 * If no pair of bits is found that way, fallback to picking some
1158 * pair of 'source' and 'dest' bits that are not the same. If the
1159 * 'source' and 'dest' bits are the same, this represents a node
1160 * that will be migrating to itself, so no pages need move.
1162 * If no bits are left in 'tmp', or if all remaining bits left
1163 * in 'tmp' correspond to the same bit in 'to', return false
1164 * (nothing left to migrate).
1166 * This lets us pick a pair of nodes to migrate between, such that
1167 * if possible the dest node is not already occupied by some other
1168 * source node, minimizing the risk of overloading the memory on a
1169 * node that would happen if we migrated incoming memory to a node
1170 * before migrating outgoing memory source that same node.
1172 * A single scan of tmp is sufficient. As we go, we remember the
1173 * most recent <s, d> pair that moved (s != d). If we find a pair
1174 * that not only moved, but what's better, moved to an empty slot
1175 * (d is not set in tmp), then we break out then, with that pair.
1176 * Otherwise when we finish scanning from_tmp, we at least have the
1177 * most recent <s, d> pair that moved. If we get all the way through
1178 * the scan of tmp without finding any node that moved, much less
1179 * moved to an empty node, then there is nothing left worth migrating.
1183 while (!nodes_empty(tmp)) {
1188 for_each_node_mask(s, tmp) {
1191 * do_migrate_pages() tries to maintain the relative
1192 * node relationship of the pages established between
1193 * threads and memory areas.
1195 * However if the number of source nodes is not equal to
1196 * the number of destination nodes we can not preserve
1197 * this node relative relationship. In that case, skip
1198 * copying memory from a node that is in the destination
1201 * Example: [2,3,4] -> [3,4,5] moves everything.
1202 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1205 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1206 (node_isset(s, *to)))
1209 d = node_remap(s, *from, *to);
1213 source = s; /* Node moved. Memorize */
1216 /* dest not in remaining from nodes? */
1217 if (!node_isset(dest, tmp))
1223 node_clear(source, tmp);
1224 err = migrate_to_node(mm, source, dest, flags);
1231 up_read(&mm->mmap_sem);
1239 * Allocate a new page for page migration based on vma policy.
1240 * Start assuming that page is mapped by vma pointed to by @private.
1241 * Search forward from there, if not. N.B., this assumes that the
1242 * list of pages handed to migrate_pages()--which is how we get here--
1243 * is in virtual address order.
1245 static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
1247 struct vm_area_struct *vma = (struct vm_area_struct *)private;
1248 unsigned long uninitialized_var(address);
1251 address = page_address_in_vma(page, vma);
1252 if (address != -EFAULT)
1258 * if !vma, alloc_page_vma() will use task or system default policy
1260 return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1264 static void migrate_page_add(struct page *page, struct list_head *pagelist,
1265 unsigned long flags)
1269 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1270 const nodemask_t *to, int flags)
1275 static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
1281 static long do_mbind(unsigned long start, unsigned long len,
1282 unsigned short mode, unsigned short mode_flags,
1283 nodemask_t *nmask, unsigned long flags)
1285 struct vm_area_struct *vma;
1286 struct mm_struct *mm = current->mm;
1287 struct mempolicy *new;
1290 LIST_HEAD(pagelist);
1292 if (flags & ~(unsigned long)MPOL_MF_VALID)
1294 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1297 if (start & ~PAGE_MASK)
1300 if (mode == MPOL_DEFAULT || mode == MPOL_NOOP)
1301 flags &= ~MPOL_MF_STRICT;
1303 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1311 new = mpol_new(mode, mode_flags, nmask);
1313 return PTR_ERR(new);
1315 if (flags & MPOL_MF_LAZY)
1316 new->flags |= MPOL_F_MOF;
1319 * If we are using the default policy then operation
1320 * on discontinuous address spaces is okay after all
1323 flags |= MPOL_MF_DISCONTIG_OK;
1325 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1326 start, start + len, mode, mode_flags,
1327 nmask ? nodes_addr(*nmask)[0] : -1);
1329 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1331 err = migrate_prep();
1336 NODEMASK_SCRATCH(scratch);
1338 down_write(&mm->mmap_sem);
1340 err = mpol_set_nodemask(new, nmask, scratch);
1341 task_unlock(current);
1343 up_write(&mm->mmap_sem);
1346 NODEMASK_SCRATCH_FREE(scratch);
1351 vma = check_range(mm, start, end, nmask,
1352 flags | MPOL_MF_INVERT, &pagelist);
1354 err = PTR_ERR(vma); /* maybe ... */
1355 if (!IS_ERR(vma) && mode != MPOL_NOOP)
1356 err = mbind_range(mm, start, end, new);
1361 if (!list_empty(&pagelist)) {
1362 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1363 nr_failed = migrate_pages(&pagelist, new_vma_page,
1365 false, MIGRATE_SYNC,
1366 MR_MEMPOLICY_MBIND);
1368 putback_lru_pages(&pagelist);
1371 if (nr_failed && (flags & MPOL_MF_STRICT))
1374 putback_lru_pages(&pagelist);
1376 up_write(&mm->mmap_sem);
1383 * User space interface with variable sized bitmaps for nodelists.
1386 /* Copy a node mask from user space. */
1387 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1388 unsigned long maxnode)
1391 unsigned long nlongs;
1392 unsigned long endmask;
1395 nodes_clear(*nodes);
1396 if (maxnode == 0 || !nmask)
1398 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1401 nlongs = BITS_TO_LONGS(maxnode);
1402 if ((maxnode % BITS_PER_LONG) == 0)
1405 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1407 /* When the user specified more nodes than supported just check
1408 if the non supported part is all zero. */
1409 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1410 if (nlongs > PAGE_SIZE/sizeof(long))
1412 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1414 if (get_user(t, nmask + k))
1416 if (k == nlongs - 1) {
1422 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1426 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1428 nodes_addr(*nodes)[nlongs-1] &= endmask;
1432 /* Copy a kernel node mask to user space */
1433 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1436 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1437 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
1439 if (copy > nbytes) {
1440 if (copy > PAGE_SIZE)
1442 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1446 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1449 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1450 unsigned long, mode, unsigned long __user *, nmask,
1451 unsigned long, maxnode, unsigned, flags)
1455 unsigned short mode_flags;
1457 mode_flags = mode & MPOL_MODE_FLAGS;
1458 mode &= ~MPOL_MODE_FLAGS;
1459 if (mode >= MPOL_MAX)
1461 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1462 (mode_flags & MPOL_F_RELATIVE_NODES))
1464 err = get_nodes(&nodes, nmask, maxnode);
1467 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1470 /* Set the process memory policy */
1471 SYSCALL_DEFINE3(set_mempolicy, int, mode, unsigned long __user *, nmask,
1472 unsigned long, maxnode)
1476 unsigned short flags;
1478 flags = mode & MPOL_MODE_FLAGS;
1479 mode &= ~MPOL_MODE_FLAGS;
1480 if ((unsigned int)mode >= MPOL_MAX)
1482 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1484 err = get_nodes(&nodes, nmask, maxnode);
1487 return do_set_mempolicy(mode, flags, &nodes);
1490 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1491 const unsigned long __user *, old_nodes,
1492 const unsigned long __user *, new_nodes)
1494 const struct cred *cred = current_cred(), *tcred;
1495 struct mm_struct *mm = NULL;
1496 struct task_struct *task;
1497 nodemask_t task_nodes;
1501 NODEMASK_SCRATCH(scratch);
1506 old = &scratch->mask1;
1507 new = &scratch->mask2;
1509 err = get_nodes(old, old_nodes, maxnode);
1513 err = get_nodes(new, new_nodes, maxnode);
1517 /* Find the mm_struct */
1519 task = pid ? find_task_by_vpid(pid) : current;
1525 get_task_struct(task);
1530 * Check if this process has the right to modify the specified
1531 * process. The right exists if the process has administrative
1532 * capabilities, superuser privileges or the same
1533 * userid as the target process.
1535 tcred = __task_cred(task);
1536 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1537 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
1538 !capable(CAP_SYS_NICE)) {
1545 task_nodes = cpuset_mems_allowed(task);
1546 /* Is the user allowed to access the target nodes? */
1547 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1552 if (!nodes_subset(*new, node_states[N_HIGH_MEMORY])) {
1557 err = security_task_movememory(task);
1561 mm = get_task_mm(task);
1562 put_task_struct(task);
1569 err = do_migrate_pages(mm, old, new,
1570 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1574 NODEMASK_SCRATCH_FREE(scratch);
1579 put_task_struct(task);
1585 /* Retrieve NUMA policy */
1586 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1587 unsigned long __user *, nmask, unsigned long, maxnode,
1588 unsigned long, addr, unsigned long, flags)
1591 int uninitialized_var(pval);
1594 if (nmask != NULL && maxnode < MAX_NUMNODES)
1597 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1602 if (policy && put_user(pval, policy))
1606 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1611 #ifdef CONFIG_COMPAT
1613 asmlinkage long compat_sys_get_mempolicy(int __user *policy,
1614 compat_ulong_t __user *nmask,
1615 compat_ulong_t maxnode,
1616 compat_ulong_t addr, compat_ulong_t flags)
1619 unsigned long __user *nm = NULL;
1620 unsigned long nr_bits, alloc_size;
1621 DECLARE_BITMAP(bm, MAX_NUMNODES);
1623 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1624 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1627 nm = compat_alloc_user_space(alloc_size);
1629 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1631 if (!err && nmask) {
1632 unsigned long copy_size;
1633 copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1634 err = copy_from_user(bm, nm, copy_size);
1635 /* ensure entire bitmap is zeroed */
1636 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1637 err |= compat_put_bitmap(nmask, bm, nr_bits);
1643 asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask,
1644 compat_ulong_t maxnode)
1647 unsigned long __user *nm = NULL;
1648 unsigned long nr_bits, alloc_size;
1649 DECLARE_BITMAP(bm, MAX_NUMNODES);
1651 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1652 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1655 err = compat_get_bitmap(bm, nmask, nr_bits);
1656 nm = compat_alloc_user_space(alloc_size);
1657 err |= copy_to_user(nm, bm, alloc_size);
1663 return sys_set_mempolicy(mode, nm, nr_bits+1);
1666 asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len,
1667 compat_ulong_t mode, compat_ulong_t __user *nmask,
1668 compat_ulong_t maxnode, compat_ulong_t flags)
1671 unsigned long __user *nm = NULL;
1672 unsigned long nr_bits, alloc_size;
1675 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1676 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1679 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
1680 nm = compat_alloc_user_space(alloc_size);
1681 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
1687 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1693 * get_vma_policy(@task, @vma, @addr)
1694 * @task - task for fallback if vma policy == default
1695 * @vma - virtual memory area whose policy is sought
1696 * @addr - address in @vma for shared policy lookup
1698 * Returns effective policy for a VMA at specified address.
1699 * Falls back to @task or system default policy, as necessary.
1700 * Current or other task's task mempolicy and non-shared vma policies must be
1701 * protected by task_lock(task) by the caller.
1702 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1703 * count--added by the get_policy() vm_op, as appropriate--to protect against
1704 * freeing by another task. It is the caller's responsibility to free the
1705 * extra reference for shared policies.
1707 struct mempolicy *get_vma_policy(struct task_struct *task,
1708 struct vm_area_struct *vma, unsigned long addr)
1710 struct mempolicy *pol = task->mempolicy;
1713 if (vma->vm_ops && vma->vm_ops->get_policy) {
1714 struct mempolicy *vpol = vma->vm_ops->get_policy(vma,
1718 } else if (vma->vm_policy) {
1719 pol = vma->vm_policy;
1722 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1723 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1724 * count on these policies which will be dropped by
1725 * mpol_cond_put() later
1727 if (mpol_needs_cond_ref(pol))
1732 pol = &default_policy;
1737 * Return a nodemask representing a mempolicy for filtering nodes for
1740 static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1742 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1743 if (unlikely(policy->mode == MPOL_BIND) &&
1744 gfp_zone(gfp) >= policy_zone &&
1745 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1746 return &policy->v.nodes;
1751 /* Return a zonelist indicated by gfp for node representing a mempolicy */
1752 static struct zonelist *policy_zonelist(gfp_t gfp, struct mempolicy *policy,
1755 switch (policy->mode) {
1756 case MPOL_PREFERRED:
1757 if (!(policy->flags & MPOL_F_LOCAL))
1758 nd = policy->v.preferred_node;
1762 * Normally, MPOL_BIND allocations are node-local within the
1763 * allowed nodemask. However, if __GFP_THISNODE is set and the
1764 * current node isn't part of the mask, we use the zonelist for
1765 * the first node in the mask instead.
1767 if (unlikely(gfp & __GFP_THISNODE) &&
1768 unlikely(!node_isset(nd, policy->v.nodes)))
1769 nd = first_node(policy->v.nodes);
1774 return node_zonelist(nd, gfp);
1777 /* Do dynamic interleaving for a process */
1778 static unsigned interleave_nodes(struct mempolicy *policy)
1781 struct task_struct *me = current;
1784 next = next_node(nid, policy->v.nodes);
1785 if (next >= MAX_NUMNODES)
1786 next = first_node(policy->v.nodes);
1787 if (next < MAX_NUMNODES)
1793 * Depending on the memory policy provide a node from which to allocate the
1795 * @policy must be protected by freeing by the caller. If @policy is
1796 * the current task's mempolicy, this protection is implicit, as only the
1797 * task can change it's policy. The system default policy requires no
1800 unsigned slab_node(void)
1802 struct mempolicy *policy;
1805 return numa_node_id();
1807 policy = current->mempolicy;
1808 if (!policy || policy->flags & MPOL_F_LOCAL)
1809 return numa_node_id();
1811 switch (policy->mode) {
1812 case MPOL_PREFERRED:
1814 * handled MPOL_F_LOCAL above
1816 return policy->v.preferred_node;
1818 case MPOL_INTERLEAVE:
1819 return interleave_nodes(policy);
1823 * Follow bind policy behavior and start allocation at the
1826 struct zonelist *zonelist;
1828 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1829 zonelist = &NODE_DATA(numa_node_id())->node_zonelists[0];
1830 (void)first_zones_zonelist(zonelist, highest_zoneidx,
1833 return zone ? zone->node : numa_node_id();
1841 /* Do static interleaving for a VMA with known offset. */
1842 static unsigned offset_il_node(struct mempolicy *pol,
1843 struct vm_area_struct *vma, unsigned long off)
1845 unsigned nnodes = nodes_weight(pol->v.nodes);
1851 return numa_node_id();
1852 target = (unsigned int)off % nnodes;
1855 nid = next_node(nid, pol->v.nodes);
1857 } while (c <= target);
1861 /* Determine a node number for interleave */
1862 static inline unsigned interleave_nid(struct mempolicy *pol,
1863 struct vm_area_struct *vma, unsigned long addr, int shift)
1869 * for small pages, there is no difference between
1870 * shift and PAGE_SHIFT, so the bit-shift is safe.
1871 * for huge pages, since vm_pgoff is in units of small
1872 * pages, we need to shift off the always 0 bits to get
1875 BUG_ON(shift < PAGE_SHIFT);
1876 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1877 off += (addr - vma->vm_start) >> shift;
1878 return offset_il_node(pol, vma, off);
1880 return interleave_nodes(pol);
1884 * Return the bit number of a random bit set in the nodemask.
1885 * (returns -1 if nodemask is empty)
1887 int node_random(const nodemask_t *maskp)
1891 w = nodes_weight(*maskp);
1893 bit = bitmap_ord_to_pos(maskp->bits,
1894 get_random_int() % w, MAX_NUMNODES);
1898 #ifdef CONFIG_HUGETLBFS
1900 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1901 * @vma = virtual memory area whose policy is sought
1902 * @addr = address in @vma for shared policy lookup and interleave policy
1903 * @gfp_flags = for requested zone
1904 * @mpol = pointer to mempolicy pointer for reference counted mempolicy
1905 * @nodemask = pointer to nodemask pointer for MPOL_BIND nodemask
1907 * Returns a zonelist suitable for a huge page allocation and a pointer
1908 * to the struct mempolicy for conditional unref after allocation.
1909 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1910 * @nodemask for filtering the zonelist.
1912 * Must be protected by get_mems_allowed()
1914 struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
1915 gfp_t gfp_flags, struct mempolicy **mpol,
1916 nodemask_t **nodemask)
1918 struct zonelist *zl;
1920 *mpol = get_vma_policy(current, vma, addr);
1921 *nodemask = NULL; /* assume !MPOL_BIND */
1923 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1924 zl = node_zonelist(interleave_nid(*mpol, vma, addr,
1925 huge_page_shift(hstate_vma(vma))), gfp_flags);
1927 zl = policy_zonelist(gfp_flags, *mpol, numa_node_id());
1928 if ((*mpol)->mode == MPOL_BIND)
1929 *nodemask = &(*mpol)->v.nodes;
1935 * init_nodemask_of_mempolicy
1937 * If the current task's mempolicy is "default" [NULL], return 'false'
1938 * to indicate default policy. Otherwise, extract the policy nodemask
1939 * for 'bind' or 'interleave' policy into the argument nodemask, or
1940 * initialize the argument nodemask to contain the single node for
1941 * 'preferred' or 'local' policy and return 'true' to indicate presence
1942 * of non-default mempolicy.
1944 * We don't bother with reference counting the mempolicy [mpol_get/put]
1945 * because the current task is examining it's own mempolicy and a task's
1946 * mempolicy is only ever changed by the task itself.
1948 * N.B., it is the caller's responsibility to free a returned nodemask.
1950 bool init_nodemask_of_mempolicy(nodemask_t *mask)
1952 struct mempolicy *mempolicy;
1955 if (!(mask && current->mempolicy))
1959 mempolicy = current->mempolicy;
1960 switch (mempolicy->mode) {
1961 case MPOL_PREFERRED:
1962 if (mempolicy->flags & MPOL_F_LOCAL)
1963 nid = numa_node_id();
1965 nid = mempolicy->v.preferred_node;
1966 init_nodemask_of_node(mask, nid);
1971 case MPOL_INTERLEAVE:
1972 *mask = mempolicy->v.nodes;
1978 task_unlock(current);
1985 * mempolicy_nodemask_intersects
1987 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1988 * policy. Otherwise, check for intersection between mask and the policy
1989 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1990 * policy, always return true since it may allocate elsewhere on fallback.
1992 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1994 bool mempolicy_nodemask_intersects(struct task_struct *tsk,
1995 const nodemask_t *mask)
1997 struct mempolicy *mempolicy;
2003 mempolicy = tsk->mempolicy;
2007 switch (mempolicy->mode) {
2008 case MPOL_PREFERRED:
2010 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
2011 * allocate from, they may fallback to other nodes when oom.
2012 * Thus, it's possible for tsk to have allocated memory from
2017 case MPOL_INTERLEAVE:
2018 ret = nodes_intersects(mempolicy->v.nodes, *mask);
2028 /* Allocate a page in interleaved policy.
2029 Own path because it needs to do special accounting. */
2030 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
2033 struct zonelist *zl;
2036 zl = node_zonelist(nid, gfp);
2037 page = __alloc_pages(gfp, order, zl);
2038 if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0]))
2039 inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
2044 * alloc_pages_vma - Allocate a page for a VMA.
2047 * %GFP_USER user allocation.
2048 * %GFP_KERNEL kernel allocations,
2049 * %GFP_HIGHMEM highmem/user allocations,
2050 * %GFP_FS allocation should not call back into a file system.
2051 * %GFP_ATOMIC don't sleep.
2053 * @order:Order of the GFP allocation.
2054 * @vma: Pointer to VMA or NULL if not available.
2055 * @addr: Virtual Address of the allocation. Must be inside the VMA.
2057 * This function allocates a page from the kernel page pool and applies
2058 * a NUMA policy associated with the VMA or the current process.
2059 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
2060 * mm_struct of the VMA to prevent it from going away. Should be used for
2061 * all allocations for pages that will be mapped into
2062 * user space. Returns NULL when no page can be allocated.
2064 * Should be called with the mm_sem of the vma hold.
2067 alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
2068 unsigned long addr, int node)
2070 struct mempolicy *pol;
2071 struct zonelist *zl;
2073 unsigned int cpuset_mems_cookie;
2076 pol = get_vma_policy(current, vma, addr);
2077 cpuset_mems_cookie = get_mems_allowed();
2079 if (unlikely(pol->mode == MPOL_INTERLEAVE)) {
2082 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
2084 page = alloc_page_interleave(gfp, order, nid);
2085 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2090 zl = policy_zonelist(gfp, pol, node);
2091 if (unlikely(mpol_needs_cond_ref(pol))) {
2093 * slow path: ref counted shared policy
2095 struct page *page = __alloc_pages_nodemask(gfp, order,
2096 zl, policy_nodemask(gfp, pol));
2098 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2103 * fast path: default or task policy
2105 page = __alloc_pages_nodemask(gfp, order, zl,
2106 policy_nodemask(gfp, pol));
2107 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2113 * alloc_pages_current - Allocate pages.
2116 * %GFP_USER user allocation,
2117 * %GFP_KERNEL kernel allocation,
2118 * %GFP_HIGHMEM highmem allocation,
2119 * %GFP_FS don't call back into a file system.
2120 * %GFP_ATOMIC don't sleep.
2121 * @order: Power of two of allocation size in pages. 0 is a single page.
2123 * Allocate a page from the kernel page pool. When not in
2124 * interrupt context and apply the current process NUMA policy.
2125 * Returns NULL when no page can be allocated.
2127 * Don't call cpuset_update_task_memory_state() unless
2128 * 1) it's ok to take cpuset_sem (can WAIT), and
2129 * 2) allocating for current task (not interrupt).
2131 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
2133 struct mempolicy *pol = current->mempolicy;
2135 unsigned int cpuset_mems_cookie;
2137 if (!pol || in_interrupt() || (gfp & __GFP_THISNODE))
2138 pol = &default_policy;
2141 cpuset_mems_cookie = get_mems_allowed();
2144 * No reference counting needed for current->mempolicy
2145 * nor system default_policy
2147 if (pol->mode == MPOL_INTERLEAVE)
2148 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2150 page = __alloc_pages_nodemask(gfp, order,
2151 policy_zonelist(gfp, pol, numa_node_id()),
2152 policy_nodemask(gfp, pol));
2154 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2159 EXPORT_SYMBOL(alloc_pages_current);
2162 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2163 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2164 * with the mems_allowed returned by cpuset_mems_allowed(). This
2165 * keeps mempolicies cpuset relative after its cpuset moves. See
2166 * further kernel/cpuset.c update_nodemask().
2168 * current's mempolicy may be rebinded by the other task(the task that changes
2169 * cpuset's mems), so we needn't do rebind work for current task.
2172 /* Slow path of a mempolicy duplicate */
2173 struct mempolicy *__mpol_dup(struct mempolicy *old)
2175 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2178 return ERR_PTR(-ENOMEM);
2180 /* task's mempolicy is protected by alloc_lock */
2181 if (old == current->mempolicy) {
2184 task_unlock(current);
2189 if (current_cpuset_is_being_rebound()) {
2190 nodemask_t mems = cpuset_mems_allowed(current);
2191 if (new->flags & MPOL_F_REBINDING)
2192 mpol_rebind_policy(new, &mems, MPOL_REBIND_STEP2);
2194 mpol_rebind_policy(new, &mems, MPOL_REBIND_ONCE);
2197 atomic_set(&new->refcnt, 1);
2202 * If *frompol needs [has] an extra ref, copy *frompol to *tompol ,
2203 * eliminate the * MPOL_F_* flags that require conditional ref and
2204 * [NOTE!!!] drop the extra ref. Not safe to reference *frompol directly
2205 * after return. Use the returned value.
2207 * Allows use of a mempolicy for, e.g., multiple allocations with a single
2208 * policy lookup, even if the policy needs/has extra ref on lookup.
2209 * shmem_readahead needs this.
2211 struct mempolicy *__mpol_cond_copy(struct mempolicy *tompol,
2212 struct mempolicy *frompol)
2214 if (!mpol_needs_cond_ref(frompol))
2218 tompol->flags &= ~MPOL_F_SHARED; /* copy doesn't need unref */
2219 __mpol_put(frompol);
2223 /* Slow path of a mempolicy comparison */
2224 bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2228 if (a->mode != b->mode)
2230 if (a->flags != b->flags)
2232 if (mpol_store_user_nodemask(a))
2233 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2239 case MPOL_INTERLEAVE:
2240 return !!nodes_equal(a->v.nodes, b->v.nodes);
2241 case MPOL_PREFERRED:
2242 return a->v.preferred_node == b->v.preferred_node;
2250 * Shared memory backing store policy support.
2252 * Remember policies even when nobody has shared memory mapped.
2253 * The policies are kept in Red-Black tree linked from the inode.
2254 * They are protected by the sp->lock spinlock, which should be held
2255 * for any accesses to the tree.
2258 /* lookup first element intersecting start-end */
2259 /* Caller holds sp->mutex */
2260 static struct sp_node *
2261 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2263 struct rb_node *n = sp->root.rb_node;
2266 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2268 if (start >= p->end)
2270 else if (end <= p->start)
2278 struct sp_node *w = NULL;
2279 struct rb_node *prev = rb_prev(n);
2282 w = rb_entry(prev, struct sp_node, nd);
2283 if (w->end <= start)
2287 return rb_entry(n, struct sp_node, nd);
2290 /* Insert a new shared policy into the list. */
2291 /* Caller holds sp->lock */
2292 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2294 struct rb_node **p = &sp->root.rb_node;
2295 struct rb_node *parent = NULL;
2300 nd = rb_entry(parent, struct sp_node, nd);
2301 if (new->start < nd->start)
2303 else if (new->end > nd->end)
2304 p = &(*p)->rb_right;
2308 rb_link_node(&new->nd, parent, p);
2309 rb_insert_color(&new->nd, &sp->root);
2310 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2311 new->policy ? new->policy->mode : 0);
2314 /* Find shared policy intersecting idx */
2316 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2318 struct mempolicy *pol = NULL;
2321 if (!sp->root.rb_node)
2323 mutex_lock(&sp->mutex);
2324 sn = sp_lookup(sp, idx, idx+1);
2326 mpol_get(sn->policy);
2329 mutex_unlock(&sp->mutex);
2333 static void sp_free(struct sp_node *n)
2335 mpol_put(n->policy);
2336 kmem_cache_free(sn_cache, n);
2340 * mpol_misplaced - check whether current page node is valid in policy
2342 * @page - page to be checked
2343 * @vma - vm area where page mapped
2344 * @addr - virtual address where page mapped
2346 * Lookup current policy node id for vma,addr and "compare to" page's
2350 * -1 - not misplaced, page is in the right node
2351 * node - node id where the page should be
2353 * Policy determination "mimics" alloc_page_vma().
2354 * Called from fault path where we know the vma and faulting address.
2356 int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2358 struct mempolicy *pol;
2360 int curnid = page_to_nid(page);
2361 unsigned long pgoff;
2367 pol = get_vma_policy(current, vma, addr);
2368 if (!(pol->flags & MPOL_F_MOF))
2371 switch (pol->mode) {
2372 case MPOL_INTERLEAVE:
2373 BUG_ON(addr >= vma->vm_end);
2374 BUG_ON(addr < vma->vm_start);
2376 pgoff = vma->vm_pgoff;
2377 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2378 polnid = offset_il_node(pol, vma, pgoff);
2381 case MPOL_PREFERRED:
2382 if (pol->flags & MPOL_F_LOCAL)
2383 polnid = numa_node_id();
2385 polnid = pol->v.preferred_node;
2390 * allows binding to multiple nodes.
2391 * use current page if in policy nodemask,
2392 * else select nearest allowed node, if any.
2393 * If no allowed nodes, use current [!misplaced].
2395 if (node_isset(curnid, pol->v.nodes))
2397 (void)first_zones_zonelist(
2398 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2399 gfp_zone(GFP_HIGHUSER),
2400 &pol->v.nodes, &zone);
2401 polnid = zone->node;
2407 if (curnid != polnid)
2415 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2417 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2418 rb_erase(&n->nd, &sp->root);
2422 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2423 struct mempolicy *pol)
2426 struct mempolicy *newpol;
2428 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2432 newpol = mpol_dup(pol);
2433 if (IS_ERR(newpol)) {
2434 kmem_cache_free(sn_cache, n);
2437 newpol->flags |= MPOL_F_SHARED;
2446 /* Replace a policy range. */
2447 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2448 unsigned long end, struct sp_node *new)
2453 mutex_lock(&sp->mutex);
2454 n = sp_lookup(sp, start, end);
2455 /* Take care of old policies in the same range. */
2456 while (n && n->start < end) {
2457 struct rb_node *next = rb_next(&n->nd);
2458 if (n->start >= start) {
2464 /* Old policy spanning whole new range. */
2466 struct sp_node *new2;
2467 new2 = sp_alloc(end, n->end, n->policy);
2473 sp_insert(sp, new2);
2480 n = rb_entry(next, struct sp_node, nd);
2485 mutex_unlock(&sp->mutex);
2490 * mpol_shared_policy_init - initialize shared policy for inode
2491 * @sp: pointer to inode shared policy
2492 * @mpol: struct mempolicy to install
2494 * Install non-NULL @mpol in inode's shared policy rb-tree.
2495 * On entry, the current task has a reference on a non-NULL @mpol.
2496 * This must be released on exit.
2497 * This is called at get_inode() calls and we can use GFP_KERNEL.
2499 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2503 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2504 mutex_init(&sp->mutex);
2507 struct vm_area_struct pvma;
2508 struct mempolicy *new;
2509 NODEMASK_SCRATCH(scratch);
2513 /* contextualize the tmpfs mount point mempolicy */
2514 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2516 goto free_scratch; /* no valid nodemask intersection */
2519 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2520 task_unlock(current);
2524 /* Create pseudo-vma that contains just the policy */
2525 memset(&pvma, 0, sizeof(struct vm_area_struct));
2526 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2527 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2530 mpol_put(new); /* drop initial ref */
2532 NODEMASK_SCRATCH_FREE(scratch);
2534 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2538 int mpol_set_shared_policy(struct shared_policy *info,
2539 struct vm_area_struct *vma, struct mempolicy *npol)
2542 struct sp_node *new = NULL;
2543 unsigned long sz = vma_pages(vma);
2545 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2547 sz, npol ? npol->mode : -1,
2548 npol ? npol->flags : -1,
2549 npol ? nodes_addr(npol->v.nodes)[0] : -1);
2552 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2556 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2562 /* Free a backing policy store on inode delete. */
2563 void mpol_free_shared_policy(struct shared_policy *p)
2566 struct rb_node *next;
2568 if (!p->root.rb_node)
2570 mutex_lock(&p->mutex);
2571 next = rb_first(&p->root);
2573 n = rb_entry(next, struct sp_node, nd);
2574 next = rb_next(&n->nd);
2577 mutex_unlock(&p->mutex);
2580 /* assumes fs == KERNEL_DS */
2581 void __init numa_policy_init(void)
2583 nodemask_t interleave_nodes;
2584 unsigned long largest = 0;
2585 int nid, prefer = 0;
2587 policy_cache = kmem_cache_create("numa_policy",
2588 sizeof(struct mempolicy),
2589 0, SLAB_PANIC, NULL);
2591 sn_cache = kmem_cache_create("shared_policy_node",
2592 sizeof(struct sp_node),
2593 0, SLAB_PANIC, NULL);
2596 * Set interleaving policy for system init. Interleaving is only
2597 * enabled across suitably sized nodes (default is >= 16MB), or
2598 * fall back to the largest node if they're all smaller.
2600 nodes_clear(interleave_nodes);
2601 for_each_node_state(nid, N_HIGH_MEMORY) {
2602 unsigned long total_pages = node_present_pages(nid);
2604 /* Preserve the largest node */
2605 if (largest < total_pages) {
2606 largest = total_pages;
2610 /* Interleave this node? */
2611 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2612 node_set(nid, interleave_nodes);
2615 /* All too small, use the largest */
2616 if (unlikely(nodes_empty(interleave_nodes)))
2617 node_set(prefer, interleave_nodes);
2619 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2620 printk("numa_policy_init: interleaving failed\n");
2623 /* Reset policy of current process to default */
2624 void numa_default_policy(void)
2626 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2630 * Parse and format mempolicy from/to strings
2634 * "local" is pseudo-policy: MPOL_PREFERRED with MPOL_F_LOCAL flag
2635 * Used only for mpol_parse_str() and mpol_to_str()
2637 static const char * const policy_modes[] =
2639 [MPOL_DEFAULT] = "default",
2640 [MPOL_PREFERRED] = "prefer",
2641 [MPOL_BIND] = "bind",
2642 [MPOL_INTERLEAVE] = "interleave",
2643 [MPOL_LOCAL] = "local",
2644 [MPOL_NOOP] = "noop", /* should not actually be used */
2650 * mpol_parse_str - parse string to mempolicy
2651 * @str: string containing mempolicy to parse
2652 * @mpol: pointer to struct mempolicy pointer, returned on success.
2653 * @no_context: flag whether to "contextualize" the mempolicy
2656 * <mode>[=<flags>][:<nodelist>]
2658 * if @no_context is true, save the input nodemask in w.user_nodemask in
2659 * the returned mempolicy. This will be used to "clone" the mempolicy in
2660 * a specific context [cpuset] at a later time. Used to parse tmpfs mpol
2661 * mount option. Note that if 'static' or 'relative' mode flags were
2662 * specified, the input nodemask will already have been saved. Saving
2663 * it again is redundant, but safe.
2665 * On success, returns 0, else 1
2667 int mpol_parse_str(char *str, struct mempolicy **mpol, int no_context)
2669 struct mempolicy *new = NULL;
2670 unsigned short mode;
2671 unsigned short uninitialized_var(mode_flags);
2673 char *nodelist = strchr(str, ':');
2674 char *flags = strchr(str, '=');
2678 /* NUL-terminate mode or flags string */
2680 if (nodelist_parse(nodelist, nodes))
2682 if (!nodes_subset(nodes, node_states[N_HIGH_MEMORY]))
2688 *flags++ = '\0'; /* terminate mode string */
2690 for (mode = 0; mode < MPOL_MAX; mode++) {
2691 if (!strcmp(str, policy_modes[mode])) {
2695 if (mode >= MPOL_MAX || mode == MPOL_NOOP)
2699 case MPOL_PREFERRED:
2701 * Insist on a nodelist of one node only
2704 char *rest = nodelist;
2705 while (isdigit(*rest))
2711 case MPOL_INTERLEAVE:
2713 * Default to online nodes with memory if no nodelist
2716 nodes = node_states[N_HIGH_MEMORY];
2720 * Don't allow a nodelist; mpol_new() checks flags
2724 mode = MPOL_PREFERRED;
2728 * Insist on a empty nodelist
2735 * Insist on a nodelist
2744 * Currently, we only support two mutually exclusive
2747 if (!strcmp(flags, "static"))
2748 mode_flags |= MPOL_F_STATIC_NODES;
2749 else if (!strcmp(flags, "relative"))
2750 mode_flags |= MPOL_F_RELATIVE_NODES;
2755 new = mpol_new(mode, mode_flags, &nodes);
2760 /* save for contextualization */
2761 new->w.user_nodemask = nodes;
2764 NODEMASK_SCRATCH(scratch);
2767 ret = mpol_set_nodemask(new, &nodes, scratch);
2768 task_unlock(current);
2771 NODEMASK_SCRATCH_FREE(scratch);
2780 /* Restore string for error message */
2789 #endif /* CONFIG_TMPFS */
2792 * mpol_to_str - format a mempolicy structure for printing
2793 * @buffer: to contain formatted mempolicy string
2794 * @maxlen: length of @buffer
2795 * @pol: pointer to mempolicy to be formatted
2796 * @no_context: "context free" mempolicy - use nodemask in w.user_nodemask
2798 * Convert a mempolicy into a string.
2799 * Returns the number of characters in buffer (if positive)
2800 * or an error (negative)
2802 int mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol, int no_context)
2807 unsigned short mode;
2808 unsigned short flags = pol ? pol->flags : 0;
2811 * Sanity check: room for longest mode, flag and some nodes
2813 VM_BUG_ON(maxlen < strlen("interleave") + strlen("relative") + 16);
2815 if (!pol || pol == &default_policy)
2816 mode = MPOL_DEFAULT;
2825 case MPOL_PREFERRED:
2827 if (flags & MPOL_F_LOCAL)
2828 mode = MPOL_LOCAL; /* pseudo-policy */
2830 node_set(pol->v.preferred_node, nodes);
2835 case MPOL_INTERLEAVE:
2837 nodes = pol->w.user_nodemask;
2839 nodes = pol->v.nodes;
2846 l = strlen(policy_modes[mode]);
2847 if (buffer + maxlen < p + l + 1)
2850 strcpy(p, policy_modes[mode]);
2853 if (flags & MPOL_MODE_FLAGS) {
2854 if (buffer + maxlen < p + 2)
2859 * Currently, the only defined flags are mutually exclusive
2861 if (flags & MPOL_F_STATIC_NODES)
2862 p += snprintf(p, buffer + maxlen - p, "static");
2863 else if (flags & MPOL_F_RELATIVE_NODES)
2864 p += snprintf(p, buffer + maxlen - p, "relative");
2867 if (!nodes_empty(nodes)) {
2868 if (buffer + maxlen < p + 2)
2871 p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);