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/sched/mm.h>
77 #include <linux/sched/numa_balancing.h>
78 #include <linux/sched/task.h>
79 #include <linux/nodemask.h>
80 #include <linux/cpuset.h>
81 #include <linux/slab.h>
82 #include <linux/string.h>
83 #include <linux/export.h>
84 #include <linux/nsproxy.h>
85 #include <linux/interrupt.h>
86 #include <linux/init.h>
87 #include <linux/compat.h>
88 #include <linux/swap.h>
89 #include <linux/seq_file.h>
90 #include <linux/proc_fs.h>
91 #include <linux/migrate.h>
92 #include <linux/ksm.h>
93 #include <linux/rmap.h>
94 #include <linux/security.h>
95 #include <linux/syscalls.h>
96 #include <linux/ctype.h>
97 #include <linux/mm_inline.h>
98 #include <linux/mmu_notifier.h>
99 #include <linux/printk.h>
101 #include <asm/tlbflush.h>
102 #include <linux/uaccess.h>
104 #include "internal.h"
107 #define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
108 #define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
110 static struct kmem_cache *policy_cache;
111 static struct kmem_cache *sn_cache;
113 /* Highest zone. An specific allocation for a zone below that is not
115 enum zone_type policy_zone = 0;
118 * run-time system-wide default policy => local allocation
120 static struct mempolicy default_policy = {
121 .refcnt = ATOMIC_INIT(1), /* never free it */
122 .mode = MPOL_PREFERRED,
123 .flags = MPOL_F_LOCAL,
126 static struct mempolicy preferred_node_policy[MAX_NUMNODES];
128 struct mempolicy *get_task_policy(struct task_struct *p)
130 struct mempolicy *pol = p->mempolicy;
136 node = numa_node_id();
137 if (node != NUMA_NO_NODE) {
138 pol = &preferred_node_policy[node];
139 /* preferred_node_policy is not initialised early in boot */
144 return &default_policy;
147 static const struct mempolicy_operations {
148 int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
150 * If read-side task has no lock to protect task->mempolicy, write-side
151 * task will rebind the task->mempolicy by two step. The first step is
152 * setting all the newly nodes, and the second step is cleaning all the
153 * disallowed nodes. In this way, we can avoid finding no node to alloc
155 * If we have a lock to protect task->mempolicy in read-side, we do
159 * MPOL_REBIND_ONCE - do rebind work at once
160 * MPOL_REBIND_STEP1 - set all the newly nodes
161 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
163 void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes,
164 enum mpol_rebind_step step);
165 } mpol_ops[MPOL_MAX];
167 static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
169 return pol->flags & MPOL_MODE_FLAGS;
172 static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
173 const nodemask_t *rel)
176 nodes_fold(tmp, *orig, nodes_weight(*rel));
177 nodes_onto(*ret, tmp, *rel);
180 static int mpol_new_interleave(struct mempolicy *pol, const nodemask_t *nodes)
182 if (nodes_empty(*nodes))
184 pol->v.nodes = *nodes;
188 static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
191 pol->flags |= MPOL_F_LOCAL; /* local allocation */
192 else if (nodes_empty(*nodes))
193 return -EINVAL; /* no allowed nodes */
195 pol->v.preferred_node = first_node(*nodes);
199 static int mpol_new_bind(struct mempolicy *pol, const nodemask_t *nodes)
201 if (nodes_empty(*nodes))
203 pol->v.nodes = *nodes;
208 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
209 * any, for the new policy. mpol_new() has already validated the nodes
210 * parameter with respect to the policy mode and flags. But, we need to
211 * handle an empty nodemask with MPOL_PREFERRED here.
213 * Must be called holding task's alloc_lock to protect task's mems_allowed
214 * and mempolicy. May also be called holding the mmap_semaphore for write.
216 static int mpol_set_nodemask(struct mempolicy *pol,
217 const nodemask_t *nodes, struct nodemask_scratch *nsc)
221 /* if mode is MPOL_DEFAULT, pol is NULL. This is right. */
225 nodes_and(nsc->mask1,
226 cpuset_current_mems_allowed, node_states[N_MEMORY]);
229 if (pol->mode == MPOL_PREFERRED && nodes_empty(*nodes))
230 nodes = NULL; /* explicit local allocation */
232 if (pol->flags & MPOL_F_RELATIVE_NODES)
233 mpol_relative_nodemask(&nsc->mask2, nodes, &nsc->mask1);
235 nodes_and(nsc->mask2, *nodes, nsc->mask1);
237 if (mpol_store_user_nodemask(pol))
238 pol->w.user_nodemask = *nodes;
240 pol->w.cpuset_mems_allowed =
241 cpuset_current_mems_allowed;
245 ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
247 ret = mpol_ops[pol->mode].create(pol, NULL);
252 * This function just creates a new policy, does some check and simple
253 * initialization. You must invoke mpol_set_nodemask() to set nodes.
255 static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
258 struct mempolicy *policy;
260 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
261 mode, flags, nodes ? nodes_addr(*nodes)[0] : NUMA_NO_NODE);
263 if (mode == MPOL_DEFAULT) {
264 if (nodes && !nodes_empty(*nodes))
265 return ERR_PTR(-EINVAL);
271 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
272 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
273 * All other modes require a valid pointer to a non-empty nodemask.
275 if (mode == MPOL_PREFERRED) {
276 if (nodes_empty(*nodes)) {
277 if (((flags & MPOL_F_STATIC_NODES) ||
278 (flags & MPOL_F_RELATIVE_NODES)))
279 return ERR_PTR(-EINVAL);
281 } else if (mode == MPOL_LOCAL) {
282 if (!nodes_empty(*nodes) ||
283 (flags & MPOL_F_STATIC_NODES) ||
284 (flags & MPOL_F_RELATIVE_NODES))
285 return ERR_PTR(-EINVAL);
286 mode = MPOL_PREFERRED;
287 } else if (nodes_empty(*nodes))
288 return ERR_PTR(-EINVAL);
289 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
291 return ERR_PTR(-ENOMEM);
292 atomic_set(&policy->refcnt, 1);
294 policy->flags = flags;
299 /* Slow path of a mpol destructor. */
300 void __mpol_put(struct mempolicy *p)
302 if (!atomic_dec_and_test(&p->refcnt))
304 kmem_cache_free(policy_cache, p);
307 static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes,
308 enum mpol_rebind_step step)
314 * MPOL_REBIND_ONCE - do rebind work at once
315 * MPOL_REBIND_STEP1 - set all the newly nodes
316 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
318 static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes,
319 enum mpol_rebind_step step)
323 if (pol->flags & MPOL_F_STATIC_NODES)
324 nodes_and(tmp, pol->w.user_nodemask, *nodes);
325 else if (pol->flags & MPOL_F_RELATIVE_NODES)
326 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
329 * if step == 1, we use ->w.cpuset_mems_allowed to cache the
332 if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP1) {
333 nodes_remap(tmp, pol->v.nodes,
334 pol->w.cpuset_mems_allowed, *nodes);
335 pol->w.cpuset_mems_allowed = step ? tmp : *nodes;
336 } else if (step == MPOL_REBIND_STEP2) {
337 tmp = pol->w.cpuset_mems_allowed;
338 pol->w.cpuset_mems_allowed = *nodes;
343 if (nodes_empty(tmp))
346 if (step == MPOL_REBIND_STEP1)
347 nodes_or(pol->v.nodes, pol->v.nodes, tmp);
348 else if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP2)
354 static void mpol_rebind_preferred(struct mempolicy *pol,
355 const nodemask_t *nodes,
356 enum mpol_rebind_step step)
360 if (pol->flags & MPOL_F_STATIC_NODES) {
361 int node = first_node(pol->w.user_nodemask);
363 if (node_isset(node, *nodes)) {
364 pol->v.preferred_node = node;
365 pol->flags &= ~MPOL_F_LOCAL;
367 pol->flags |= MPOL_F_LOCAL;
368 } else if (pol->flags & MPOL_F_RELATIVE_NODES) {
369 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
370 pol->v.preferred_node = first_node(tmp);
371 } else if (!(pol->flags & MPOL_F_LOCAL)) {
372 pol->v.preferred_node = node_remap(pol->v.preferred_node,
373 pol->w.cpuset_mems_allowed,
375 pol->w.cpuset_mems_allowed = *nodes;
380 * mpol_rebind_policy - Migrate a policy to a different set of nodes
382 * If read-side task has no lock to protect task->mempolicy, write-side
383 * task will rebind the task->mempolicy by two step. The first step is
384 * setting all the newly nodes, and the second step is cleaning all the
385 * disallowed nodes. In this way, we can avoid finding no node to alloc
387 * If we have a lock to protect task->mempolicy in read-side, we do
391 * MPOL_REBIND_ONCE - do rebind work at once
392 * MPOL_REBIND_STEP1 - set all the newly nodes
393 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
395 static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask,
396 enum mpol_rebind_step step)
400 if (!mpol_store_user_nodemask(pol) && step == MPOL_REBIND_ONCE &&
401 nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
404 if (step == MPOL_REBIND_STEP1 && (pol->flags & MPOL_F_REBINDING))
407 if (step == MPOL_REBIND_STEP2 && !(pol->flags & MPOL_F_REBINDING))
410 if (step == MPOL_REBIND_STEP1)
411 pol->flags |= MPOL_F_REBINDING;
412 else if (step == MPOL_REBIND_STEP2)
413 pol->flags &= ~MPOL_F_REBINDING;
414 else if (step >= MPOL_REBIND_NSTEP)
417 mpol_ops[pol->mode].rebind(pol, newmask, step);
421 * Wrapper for mpol_rebind_policy() that just requires task
422 * pointer, and updates task mempolicy.
424 * Called with task's alloc_lock held.
427 void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new,
428 enum mpol_rebind_step step)
430 mpol_rebind_policy(tsk->mempolicy, new, step);
434 * Rebind each vma in mm to new nodemask.
436 * Call holding a reference to mm. Takes mm->mmap_sem during call.
439 void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
441 struct vm_area_struct *vma;
443 down_write(&mm->mmap_sem);
444 for (vma = mm->mmap; vma; vma = vma->vm_next)
445 mpol_rebind_policy(vma->vm_policy, new, MPOL_REBIND_ONCE);
446 up_write(&mm->mmap_sem);
449 static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
451 .rebind = mpol_rebind_default,
453 [MPOL_INTERLEAVE] = {
454 .create = mpol_new_interleave,
455 .rebind = mpol_rebind_nodemask,
458 .create = mpol_new_preferred,
459 .rebind = mpol_rebind_preferred,
462 .create = mpol_new_bind,
463 .rebind = mpol_rebind_nodemask,
467 static void migrate_page_add(struct page *page, struct list_head *pagelist,
468 unsigned long flags);
471 struct list_head *pagelist;
474 struct vm_area_struct *prev;
478 * Scan through pages checking if pages follow certain conditions,
479 * and move them to the pagelist if they do.
481 static int queue_pages_pte_range(pmd_t *pmd, unsigned long addr,
482 unsigned long end, struct mm_walk *walk)
484 struct vm_area_struct *vma = walk->vma;
486 struct queue_pages *qp = walk->private;
487 unsigned long flags = qp->flags;
492 if (pmd_trans_huge(*pmd)) {
493 ptl = pmd_lock(walk->mm, pmd);
494 if (pmd_trans_huge(*pmd)) {
495 page = pmd_page(*pmd);
496 if (is_huge_zero_page(page)) {
498 __split_huge_pmd(vma, pmd, addr, false, NULL);
503 ret = split_huge_page(page);
514 if (pmd_trans_unstable(pmd))
517 pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
518 for (; addr != end; pte++, addr += PAGE_SIZE) {
519 if (!pte_present(*pte))
521 page = vm_normal_page(vma, addr, *pte);
525 * vm_normal_page() filters out zero pages, but there might
526 * still be PageReserved pages to skip, perhaps in a VDSO.
528 if (PageReserved(page))
530 nid = page_to_nid(page);
531 if (node_isset(nid, *qp->nmask) == !!(flags & MPOL_MF_INVERT))
533 if (PageTransCompound(page)) {
535 pte_unmap_unlock(pte, ptl);
537 ret = split_huge_page(page);
540 /* Failed to split -- skip. */
542 pte = pte_offset_map_lock(walk->mm, pmd,
549 migrate_page_add(page, qp->pagelist, flags);
551 pte_unmap_unlock(pte - 1, ptl);
556 static int queue_pages_hugetlb(pte_t *pte, unsigned long hmask,
557 unsigned long addr, unsigned long end,
558 struct mm_walk *walk)
560 #ifdef CONFIG_HUGETLB_PAGE
561 struct queue_pages *qp = walk->private;
562 unsigned long flags = qp->flags;
568 ptl = huge_pte_lock(hstate_vma(walk->vma), walk->mm, pte);
569 entry = huge_ptep_get(pte);
570 if (!pte_present(entry))
572 page = pte_page(entry);
573 nid = page_to_nid(page);
574 if (node_isset(nid, *qp->nmask) == !!(flags & MPOL_MF_INVERT))
576 /* With MPOL_MF_MOVE, we migrate only unshared hugepage. */
577 if (flags & (MPOL_MF_MOVE_ALL) ||
578 (flags & MPOL_MF_MOVE && page_mapcount(page) == 1))
579 isolate_huge_page(page, qp->pagelist);
588 #ifdef CONFIG_NUMA_BALANCING
590 * This is used to mark a range of virtual addresses to be inaccessible.
591 * These are later cleared by a NUMA hinting fault. Depending on these
592 * faults, pages may be migrated for better NUMA placement.
594 * This is assuming that NUMA faults are handled using PROT_NONE. If
595 * an architecture makes a different choice, it will need further
596 * changes to the core.
598 unsigned long change_prot_numa(struct vm_area_struct *vma,
599 unsigned long addr, unsigned long end)
603 nr_updated = change_protection(vma, addr, end, PAGE_NONE, 0, 1);
605 count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
610 static unsigned long change_prot_numa(struct vm_area_struct *vma,
611 unsigned long addr, unsigned long end)
615 #endif /* CONFIG_NUMA_BALANCING */
617 static int queue_pages_test_walk(unsigned long start, unsigned long end,
618 struct mm_walk *walk)
620 struct vm_area_struct *vma = walk->vma;
621 struct queue_pages *qp = walk->private;
622 unsigned long endvma = vma->vm_end;
623 unsigned long flags = qp->flags;
625 if (!vma_migratable(vma))
630 if (vma->vm_start > start)
631 start = vma->vm_start;
633 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
634 if (!vma->vm_next && vma->vm_end < end)
636 if (qp->prev && qp->prev->vm_end < vma->vm_start)
642 if (flags & MPOL_MF_LAZY) {
643 /* Similar to task_numa_work, skip inaccessible VMAs */
644 if (!is_vm_hugetlb_page(vma) &&
645 (vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)) &&
646 !(vma->vm_flags & VM_MIXEDMAP))
647 change_prot_numa(vma, start, endvma);
651 /* queue pages from current vma */
652 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
658 * Walk through page tables and collect pages to be migrated.
660 * If pages found in a given range are on a set of nodes (determined by
661 * @nodes and @flags,) it's isolated and queued to the pagelist which is
662 * passed via @private.)
665 queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
666 nodemask_t *nodes, unsigned long flags,
667 struct list_head *pagelist)
669 struct queue_pages qp = {
670 .pagelist = pagelist,
675 struct mm_walk queue_pages_walk = {
676 .hugetlb_entry = queue_pages_hugetlb,
677 .pmd_entry = queue_pages_pte_range,
678 .test_walk = queue_pages_test_walk,
683 return walk_page_range(start, end, &queue_pages_walk);
687 * Apply policy to a single VMA
688 * This must be called with the mmap_sem held for writing.
690 static int vma_replace_policy(struct vm_area_struct *vma,
691 struct mempolicy *pol)
694 struct mempolicy *old;
695 struct mempolicy *new;
697 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
698 vma->vm_start, vma->vm_end, vma->vm_pgoff,
699 vma->vm_ops, vma->vm_file,
700 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
706 if (vma->vm_ops && vma->vm_ops->set_policy) {
707 err = vma->vm_ops->set_policy(vma, new);
712 old = vma->vm_policy;
713 vma->vm_policy = new; /* protected by mmap_sem */
722 /* Step 2: apply policy to a range and do splits. */
723 static int mbind_range(struct mm_struct *mm, unsigned long start,
724 unsigned long end, struct mempolicy *new_pol)
726 struct vm_area_struct *next;
727 struct vm_area_struct *prev;
728 struct vm_area_struct *vma;
731 unsigned long vmstart;
734 vma = find_vma(mm, start);
735 if (!vma || vma->vm_start > start)
739 if (start > vma->vm_start)
742 for (; vma && vma->vm_start < end; prev = vma, vma = next) {
744 vmstart = max(start, vma->vm_start);
745 vmend = min(end, vma->vm_end);
747 if (mpol_equal(vma_policy(vma), new_pol))
750 pgoff = vma->vm_pgoff +
751 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
752 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
753 vma->anon_vma, vma->vm_file, pgoff,
754 new_pol, vma->vm_userfaultfd_ctx);
758 if (mpol_equal(vma_policy(vma), new_pol))
760 /* vma_merge() joined vma && vma->next, case 8 */
763 if (vma->vm_start != vmstart) {
764 err = split_vma(vma->vm_mm, vma, vmstart, 1);
768 if (vma->vm_end != vmend) {
769 err = split_vma(vma->vm_mm, vma, vmend, 0);
774 err = vma_replace_policy(vma, new_pol);
783 /* Set the process memory policy */
784 static long do_set_mempolicy(unsigned short mode, unsigned short flags,
787 struct mempolicy *new, *old;
788 NODEMASK_SCRATCH(scratch);
794 new = mpol_new(mode, flags, nodes);
801 ret = mpol_set_nodemask(new, nodes, scratch);
803 task_unlock(current);
807 old = current->mempolicy;
808 current->mempolicy = new;
809 if (new && new->mode == MPOL_INTERLEAVE)
810 current->il_prev = MAX_NUMNODES-1;
811 task_unlock(current);
815 NODEMASK_SCRATCH_FREE(scratch);
820 * Return nodemask for policy for get_mempolicy() query
822 * Called with task's alloc_lock held
824 static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
827 if (p == &default_policy)
833 case MPOL_INTERLEAVE:
837 if (!(p->flags & MPOL_F_LOCAL))
838 node_set(p->v.preferred_node, *nodes);
839 /* else return empty node mask for local allocation */
846 static int lookup_node(unsigned long addr)
851 err = get_user_pages(addr & PAGE_MASK, 1, 0, &p, NULL);
853 err = page_to_nid(p);
859 /* Retrieve NUMA policy */
860 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
861 unsigned long addr, unsigned long flags)
864 struct mm_struct *mm = current->mm;
865 struct vm_area_struct *vma = NULL;
866 struct mempolicy *pol = current->mempolicy;
869 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
872 if (flags & MPOL_F_MEMS_ALLOWED) {
873 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
875 *policy = 0; /* just so it's initialized */
877 *nmask = cpuset_current_mems_allowed;
878 task_unlock(current);
882 if (flags & MPOL_F_ADDR) {
884 * Do NOT fall back to task policy if the
885 * vma/shared policy at addr is NULL. We
886 * want to return MPOL_DEFAULT in this case.
888 down_read(&mm->mmap_sem);
889 vma = find_vma_intersection(mm, addr, addr+1);
891 up_read(&mm->mmap_sem);
894 if (vma->vm_ops && vma->vm_ops->get_policy)
895 pol = vma->vm_ops->get_policy(vma, addr);
897 pol = vma->vm_policy;
902 pol = &default_policy; /* indicates default behavior */
904 if (flags & MPOL_F_NODE) {
905 if (flags & MPOL_F_ADDR) {
906 err = lookup_node(addr);
910 } else if (pol == current->mempolicy &&
911 pol->mode == MPOL_INTERLEAVE) {
912 *policy = next_node_in(current->il_prev, pol->v.nodes);
918 *policy = pol == &default_policy ? MPOL_DEFAULT :
921 * Internal mempolicy flags must be masked off before exposing
922 * the policy to userspace.
924 *policy |= (pol->flags & MPOL_MODE_FLAGS);
928 up_read(¤t->mm->mmap_sem);
934 if (mpol_store_user_nodemask(pol)) {
935 *nmask = pol->w.user_nodemask;
938 get_policy_nodemask(pol, nmask);
939 task_unlock(current);
946 up_read(¤t->mm->mmap_sem);
950 #ifdef CONFIG_MIGRATION
954 static void migrate_page_add(struct page *page, struct list_head *pagelist,
958 * Avoid migrating a page that is shared with others.
960 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) {
961 if (!isolate_lru_page(page)) {
962 list_add_tail(&page->lru, pagelist);
963 inc_node_page_state(page, NR_ISOLATED_ANON +
964 page_is_file_cache(page));
969 static struct page *new_node_page(struct page *page, unsigned long node, int **x)
972 return alloc_huge_page_node(page_hstate(compound_head(page)),
975 return __alloc_pages_node(node, GFP_HIGHUSER_MOVABLE |
980 * Migrate pages from one node to a target node.
981 * Returns error or the number of pages not migrated.
983 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
991 node_set(source, nmask);
994 * This does not "check" the range but isolates all pages that
995 * need migration. Between passing in the full user address
996 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
998 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
999 queue_pages_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
1000 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
1002 if (!list_empty(&pagelist)) {
1003 err = migrate_pages(&pagelist, new_node_page, NULL, dest,
1004 MIGRATE_SYNC, MR_SYSCALL);
1006 putback_movable_pages(&pagelist);
1013 * Move pages between the two nodesets so as to preserve the physical
1014 * layout as much as possible.
1016 * Returns the number of page that could not be moved.
1018 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1019 const nodemask_t *to, int flags)
1025 err = migrate_prep();
1029 down_read(&mm->mmap_sem);
1032 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1033 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1034 * bit in 'tmp', and return that <source, dest> pair for migration.
1035 * The pair of nodemasks 'to' and 'from' define the map.
1037 * If no pair of bits is found that way, fallback to picking some
1038 * pair of 'source' and 'dest' bits that are not the same. If the
1039 * 'source' and 'dest' bits are the same, this represents a node
1040 * that will be migrating to itself, so no pages need move.
1042 * If no bits are left in 'tmp', or if all remaining bits left
1043 * in 'tmp' correspond to the same bit in 'to', return false
1044 * (nothing left to migrate).
1046 * This lets us pick a pair of nodes to migrate between, such that
1047 * if possible the dest node is not already occupied by some other
1048 * source node, minimizing the risk of overloading the memory on a
1049 * node that would happen if we migrated incoming memory to a node
1050 * before migrating outgoing memory source that same node.
1052 * A single scan of tmp is sufficient. As we go, we remember the
1053 * most recent <s, d> pair that moved (s != d). If we find a pair
1054 * that not only moved, but what's better, moved to an empty slot
1055 * (d is not set in tmp), then we break out then, with that pair.
1056 * Otherwise when we finish scanning from_tmp, we at least have the
1057 * most recent <s, d> pair that moved. If we get all the way through
1058 * the scan of tmp without finding any node that moved, much less
1059 * moved to an empty node, then there is nothing left worth migrating.
1063 while (!nodes_empty(tmp)) {
1065 int source = NUMA_NO_NODE;
1068 for_each_node_mask(s, tmp) {
1071 * do_migrate_pages() tries to maintain the relative
1072 * node relationship of the pages established between
1073 * threads and memory areas.
1075 * However if the number of source nodes is not equal to
1076 * the number of destination nodes we can not preserve
1077 * this node relative relationship. In that case, skip
1078 * copying memory from a node that is in the destination
1081 * Example: [2,3,4] -> [3,4,5] moves everything.
1082 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1085 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1086 (node_isset(s, *to)))
1089 d = node_remap(s, *from, *to);
1093 source = s; /* Node moved. Memorize */
1096 /* dest not in remaining from nodes? */
1097 if (!node_isset(dest, tmp))
1100 if (source == NUMA_NO_NODE)
1103 node_clear(source, tmp);
1104 err = migrate_to_node(mm, source, dest, flags);
1110 up_read(&mm->mmap_sem);
1118 * Allocate a new page for page migration based on vma policy.
1119 * Start by assuming the page is mapped by the same vma as contains @start.
1120 * Search forward from there, if not. N.B., this assumes that the
1121 * list of pages handed to migrate_pages()--which is how we get here--
1122 * is in virtual address order.
1124 static struct page *new_page(struct page *page, unsigned long start, int **x)
1126 struct vm_area_struct *vma;
1127 unsigned long uninitialized_var(address);
1129 vma = find_vma(current->mm, start);
1131 address = page_address_in_vma(page, vma);
1132 if (address != -EFAULT)
1137 if (PageHuge(page)) {
1139 return alloc_huge_page_noerr(vma, address, 1);
1142 * if !vma, alloc_page_vma() will use task or system default policy
1144 return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1148 static void migrate_page_add(struct page *page, struct list_head *pagelist,
1149 unsigned long flags)
1153 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1154 const nodemask_t *to, int flags)
1159 static struct page *new_page(struct page *page, unsigned long start, int **x)
1165 static long do_mbind(unsigned long start, unsigned long len,
1166 unsigned short mode, unsigned short mode_flags,
1167 nodemask_t *nmask, unsigned long flags)
1169 struct mm_struct *mm = current->mm;
1170 struct mempolicy *new;
1173 LIST_HEAD(pagelist);
1175 if (flags & ~(unsigned long)MPOL_MF_VALID)
1177 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1180 if (start & ~PAGE_MASK)
1183 if (mode == MPOL_DEFAULT)
1184 flags &= ~MPOL_MF_STRICT;
1186 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1194 new = mpol_new(mode, mode_flags, nmask);
1196 return PTR_ERR(new);
1198 if (flags & MPOL_MF_LAZY)
1199 new->flags |= MPOL_F_MOF;
1202 * If we are using the default policy then operation
1203 * on discontinuous address spaces is okay after all
1206 flags |= MPOL_MF_DISCONTIG_OK;
1208 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1209 start, start + len, mode, mode_flags,
1210 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
1212 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1214 err = migrate_prep();
1219 NODEMASK_SCRATCH(scratch);
1221 down_write(&mm->mmap_sem);
1223 err = mpol_set_nodemask(new, nmask, scratch);
1224 task_unlock(current);
1226 up_write(&mm->mmap_sem);
1229 NODEMASK_SCRATCH_FREE(scratch);
1234 err = queue_pages_range(mm, start, end, nmask,
1235 flags | MPOL_MF_INVERT, &pagelist);
1237 err = mbind_range(mm, start, end, new);
1242 if (!list_empty(&pagelist)) {
1243 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1244 nr_failed = migrate_pages(&pagelist, new_page, NULL,
1245 start, MIGRATE_SYNC, MR_MEMPOLICY_MBIND);
1247 putback_movable_pages(&pagelist);
1250 if (nr_failed && (flags & MPOL_MF_STRICT))
1253 putback_movable_pages(&pagelist);
1255 up_write(&mm->mmap_sem);
1262 * User space interface with variable sized bitmaps for nodelists.
1265 /* Copy a node mask from user space. */
1266 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1267 unsigned long maxnode)
1270 unsigned long nlongs;
1271 unsigned long endmask;
1274 nodes_clear(*nodes);
1275 if (maxnode == 0 || !nmask)
1277 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1280 nlongs = BITS_TO_LONGS(maxnode);
1281 if ((maxnode % BITS_PER_LONG) == 0)
1284 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1286 /* When the user specified more nodes than supported just check
1287 if the non supported part is all zero. */
1288 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1289 if (nlongs > PAGE_SIZE/sizeof(long))
1291 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1293 if (get_user(t, nmask + k))
1295 if (k == nlongs - 1) {
1301 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1305 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1307 nodes_addr(*nodes)[nlongs-1] &= endmask;
1311 /* Copy a kernel node mask to user space */
1312 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1315 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1316 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
1318 if (copy > nbytes) {
1319 if (copy > PAGE_SIZE)
1321 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1325 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1328 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1329 unsigned long, mode, const unsigned long __user *, nmask,
1330 unsigned long, maxnode, unsigned, flags)
1334 unsigned short mode_flags;
1336 mode_flags = mode & MPOL_MODE_FLAGS;
1337 mode &= ~MPOL_MODE_FLAGS;
1338 if (mode >= MPOL_MAX)
1340 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1341 (mode_flags & MPOL_F_RELATIVE_NODES))
1343 err = get_nodes(&nodes, nmask, maxnode);
1346 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1349 /* Set the process memory policy */
1350 SYSCALL_DEFINE3(set_mempolicy, int, mode, const unsigned long __user *, nmask,
1351 unsigned long, maxnode)
1355 unsigned short flags;
1357 flags = mode & MPOL_MODE_FLAGS;
1358 mode &= ~MPOL_MODE_FLAGS;
1359 if ((unsigned int)mode >= MPOL_MAX)
1361 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1363 err = get_nodes(&nodes, nmask, maxnode);
1366 return do_set_mempolicy(mode, flags, &nodes);
1369 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1370 const unsigned long __user *, old_nodes,
1371 const unsigned long __user *, new_nodes)
1373 const struct cred *cred = current_cred(), *tcred;
1374 struct mm_struct *mm = NULL;
1375 struct task_struct *task;
1376 nodemask_t task_nodes;
1380 NODEMASK_SCRATCH(scratch);
1385 old = &scratch->mask1;
1386 new = &scratch->mask2;
1388 err = get_nodes(old, old_nodes, maxnode);
1392 err = get_nodes(new, new_nodes, maxnode);
1396 /* Find the mm_struct */
1398 task = pid ? find_task_by_vpid(pid) : current;
1404 get_task_struct(task);
1409 * Check if this process has the right to modify the specified
1410 * process. The right exists if the process has administrative
1411 * capabilities, superuser privileges or the same
1412 * userid as the target process.
1414 tcred = __task_cred(task);
1415 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1416 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
1417 !capable(CAP_SYS_NICE)) {
1424 task_nodes = cpuset_mems_allowed(task);
1425 /* Is the user allowed to access the target nodes? */
1426 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1431 if (!nodes_subset(*new, node_states[N_MEMORY])) {
1436 err = security_task_movememory(task);
1440 mm = get_task_mm(task);
1441 put_task_struct(task);
1448 err = do_migrate_pages(mm, old, new,
1449 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1453 NODEMASK_SCRATCH_FREE(scratch);
1458 put_task_struct(task);
1464 /* Retrieve NUMA policy */
1465 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1466 unsigned long __user *, nmask, unsigned long, maxnode,
1467 unsigned long, addr, unsigned long, flags)
1470 int uninitialized_var(pval);
1473 if (nmask != NULL && maxnode < MAX_NUMNODES)
1476 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1481 if (policy && put_user(pval, policy))
1485 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1490 #ifdef CONFIG_COMPAT
1492 COMPAT_SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1493 compat_ulong_t __user *, nmask,
1494 compat_ulong_t, maxnode,
1495 compat_ulong_t, addr, compat_ulong_t, flags)
1498 unsigned long __user *nm = NULL;
1499 unsigned long nr_bits, alloc_size;
1500 DECLARE_BITMAP(bm, MAX_NUMNODES);
1502 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1503 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1506 nm = compat_alloc_user_space(alloc_size);
1508 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1510 if (!err && nmask) {
1511 unsigned long copy_size;
1512 copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1513 err = copy_from_user(bm, nm, copy_size);
1514 /* ensure entire bitmap is zeroed */
1515 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1516 err |= compat_put_bitmap(nmask, bm, nr_bits);
1522 COMPAT_SYSCALL_DEFINE3(set_mempolicy, int, mode, compat_ulong_t __user *, nmask,
1523 compat_ulong_t, maxnode)
1525 unsigned long __user *nm = NULL;
1526 unsigned long nr_bits, alloc_size;
1527 DECLARE_BITMAP(bm, MAX_NUMNODES);
1529 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1530 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1533 if (compat_get_bitmap(bm, nmask, nr_bits))
1535 nm = compat_alloc_user_space(alloc_size);
1536 if (copy_to_user(nm, bm, alloc_size))
1540 return sys_set_mempolicy(mode, nm, nr_bits+1);
1543 COMPAT_SYSCALL_DEFINE6(mbind, compat_ulong_t, start, compat_ulong_t, len,
1544 compat_ulong_t, mode, compat_ulong_t __user *, nmask,
1545 compat_ulong_t, maxnode, compat_ulong_t, flags)
1547 unsigned long __user *nm = NULL;
1548 unsigned long nr_bits, alloc_size;
1551 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1552 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1555 if (compat_get_bitmap(nodes_addr(bm), nmask, nr_bits))
1557 nm = compat_alloc_user_space(alloc_size);
1558 if (copy_to_user(nm, nodes_addr(bm), alloc_size))
1562 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1567 struct mempolicy *__get_vma_policy(struct vm_area_struct *vma,
1570 struct mempolicy *pol = NULL;
1573 if (vma->vm_ops && vma->vm_ops->get_policy) {
1574 pol = vma->vm_ops->get_policy(vma, addr);
1575 } else if (vma->vm_policy) {
1576 pol = vma->vm_policy;
1579 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1580 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1581 * count on these policies which will be dropped by
1582 * mpol_cond_put() later
1584 if (mpol_needs_cond_ref(pol))
1593 * get_vma_policy(@vma, @addr)
1594 * @vma: virtual memory area whose policy is sought
1595 * @addr: address in @vma for shared policy lookup
1597 * Returns effective policy for a VMA at specified address.
1598 * Falls back to current->mempolicy or system default policy, as necessary.
1599 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1600 * count--added by the get_policy() vm_op, as appropriate--to protect against
1601 * freeing by another task. It is the caller's responsibility to free the
1602 * extra reference for shared policies.
1604 static struct mempolicy *get_vma_policy(struct vm_area_struct *vma,
1607 struct mempolicy *pol = __get_vma_policy(vma, addr);
1610 pol = get_task_policy(current);
1615 bool vma_policy_mof(struct vm_area_struct *vma)
1617 struct mempolicy *pol;
1619 if (vma->vm_ops && vma->vm_ops->get_policy) {
1622 pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1623 if (pol && (pol->flags & MPOL_F_MOF))
1630 pol = vma->vm_policy;
1632 pol = get_task_policy(current);
1634 return pol->flags & MPOL_F_MOF;
1637 static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1639 enum zone_type dynamic_policy_zone = policy_zone;
1641 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1644 * if policy->v.nodes has movable memory only,
1645 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1647 * policy->v.nodes is intersect with node_states[N_MEMORY].
1648 * so if the following test faile, it implies
1649 * policy->v.nodes has movable memory only.
1651 if (!nodes_intersects(policy->v.nodes, node_states[N_HIGH_MEMORY]))
1652 dynamic_policy_zone = ZONE_MOVABLE;
1654 return zone >= dynamic_policy_zone;
1658 * Return a nodemask representing a mempolicy for filtering nodes for
1661 static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1663 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1664 if (unlikely(policy->mode == MPOL_BIND) &&
1665 apply_policy_zone(policy, gfp_zone(gfp)) &&
1666 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1667 return &policy->v.nodes;
1672 /* Return the node id preferred by the given mempolicy, or the given id */
1673 static int policy_node(gfp_t gfp, struct mempolicy *policy,
1676 if (policy->mode == MPOL_PREFERRED && !(policy->flags & MPOL_F_LOCAL))
1677 nd = policy->v.preferred_node;
1680 * __GFP_THISNODE shouldn't even be used with the bind policy
1681 * because we might easily break the expectation to stay on the
1682 * requested node and not break the policy.
1684 WARN_ON_ONCE(policy->mode == MPOL_BIND && (gfp & __GFP_THISNODE));
1690 /* Do dynamic interleaving for a process */
1691 static unsigned interleave_nodes(struct mempolicy *policy)
1694 struct task_struct *me = current;
1696 next = next_node_in(me->il_prev, policy->v.nodes);
1697 if (next < MAX_NUMNODES)
1703 * Depending on the memory policy provide a node from which to allocate the
1706 unsigned int mempolicy_slab_node(void)
1708 struct mempolicy *policy;
1709 int node = numa_mem_id();
1714 policy = current->mempolicy;
1715 if (!policy || policy->flags & MPOL_F_LOCAL)
1718 switch (policy->mode) {
1719 case MPOL_PREFERRED:
1721 * handled MPOL_F_LOCAL above
1723 return policy->v.preferred_node;
1725 case MPOL_INTERLEAVE:
1726 return interleave_nodes(policy);
1732 * Follow bind policy behavior and start allocation at the
1735 struct zonelist *zonelist;
1736 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1737 zonelist = &NODE_DATA(node)->node_zonelists[ZONELIST_FALLBACK];
1738 z = first_zones_zonelist(zonelist, highest_zoneidx,
1740 return z->zone ? z->zone->node : node;
1749 * Do static interleaving for a VMA with known offset @n. Returns the n'th
1750 * node in pol->v.nodes (starting from n=0), wrapping around if n exceeds the
1751 * number of present nodes.
1753 static unsigned offset_il_node(struct mempolicy *pol,
1754 struct vm_area_struct *vma, unsigned long n)
1756 unsigned nnodes = nodes_weight(pol->v.nodes);
1762 return numa_node_id();
1763 target = (unsigned int)n % nnodes;
1764 nid = first_node(pol->v.nodes);
1765 for (i = 0; i < target; i++)
1766 nid = next_node(nid, pol->v.nodes);
1770 /* Determine a node number for interleave */
1771 static inline unsigned interleave_nid(struct mempolicy *pol,
1772 struct vm_area_struct *vma, unsigned long addr, int shift)
1778 * for small pages, there is no difference between
1779 * shift and PAGE_SHIFT, so the bit-shift is safe.
1780 * for huge pages, since vm_pgoff is in units of small
1781 * pages, we need to shift off the always 0 bits to get
1784 BUG_ON(shift < PAGE_SHIFT);
1785 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1786 off += (addr - vma->vm_start) >> shift;
1787 return offset_il_node(pol, vma, off);
1789 return interleave_nodes(pol);
1792 #ifdef CONFIG_HUGETLBFS
1794 * huge_node(@vma, @addr, @gfp_flags, @mpol)
1795 * @vma: virtual memory area whose policy is sought
1796 * @addr: address in @vma for shared policy lookup and interleave policy
1797 * @gfp_flags: for requested zone
1798 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
1799 * @nodemask: pointer to nodemask pointer for MPOL_BIND nodemask
1801 * Returns a nid suitable for a huge page allocation and a pointer
1802 * to the struct mempolicy for conditional unref after allocation.
1803 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1804 * @nodemask for filtering the zonelist.
1806 * Must be protected by read_mems_allowed_begin()
1808 int huge_node(struct vm_area_struct *vma, unsigned long addr, gfp_t gfp_flags,
1809 struct mempolicy **mpol, nodemask_t **nodemask)
1813 *mpol = get_vma_policy(vma, addr);
1814 *nodemask = NULL; /* assume !MPOL_BIND */
1816 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1817 nid = interleave_nid(*mpol, vma, addr,
1818 huge_page_shift(hstate_vma(vma)));
1820 nid = policy_node(gfp_flags, *mpol, numa_node_id());
1821 if ((*mpol)->mode == MPOL_BIND)
1822 *nodemask = &(*mpol)->v.nodes;
1828 * init_nodemask_of_mempolicy
1830 * If the current task's mempolicy is "default" [NULL], return 'false'
1831 * to indicate default policy. Otherwise, extract the policy nodemask
1832 * for 'bind' or 'interleave' policy into the argument nodemask, or
1833 * initialize the argument nodemask to contain the single node for
1834 * 'preferred' or 'local' policy and return 'true' to indicate presence
1835 * of non-default mempolicy.
1837 * We don't bother with reference counting the mempolicy [mpol_get/put]
1838 * because the current task is examining it's own mempolicy and a task's
1839 * mempolicy is only ever changed by the task itself.
1841 * N.B., it is the caller's responsibility to free a returned nodemask.
1843 bool init_nodemask_of_mempolicy(nodemask_t *mask)
1845 struct mempolicy *mempolicy;
1848 if (!(mask && current->mempolicy))
1852 mempolicy = current->mempolicy;
1853 switch (mempolicy->mode) {
1854 case MPOL_PREFERRED:
1855 if (mempolicy->flags & MPOL_F_LOCAL)
1856 nid = numa_node_id();
1858 nid = mempolicy->v.preferred_node;
1859 init_nodemask_of_node(mask, nid);
1864 case MPOL_INTERLEAVE:
1865 *mask = mempolicy->v.nodes;
1871 task_unlock(current);
1878 * mempolicy_nodemask_intersects
1880 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1881 * policy. Otherwise, check for intersection between mask and the policy
1882 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1883 * policy, always return true since it may allocate elsewhere on fallback.
1885 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1887 bool mempolicy_nodemask_intersects(struct task_struct *tsk,
1888 const nodemask_t *mask)
1890 struct mempolicy *mempolicy;
1896 mempolicy = tsk->mempolicy;
1900 switch (mempolicy->mode) {
1901 case MPOL_PREFERRED:
1903 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1904 * allocate from, they may fallback to other nodes when oom.
1905 * Thus, it's possible for tsk to have allocated memory from
1910 case MPOL_INTERLEAVE:
1911 ret = nodes_intersects(mempolicy->v.nodes, *mask);
1921 /* Allocate a page in interleaved policy.
1922 Own path because it needs to do special accounting. */
1923 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1928 page = __alloc_pages(gfp, order, nid);
1929 if (page && page_to_nid(page) == nid)
1930 inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
1935 * alloc_pages_vma - Allocate a page for a VMA.
1938 * %GFP_USER user allocation.
1939 * %GFP_KERNEL kernel allocations,
1940 * %GFP_HIGHMEM highmem/user allocations,
1941 * %GFP_FS allocation should not call back into a file system.
1942 * %GFP_ATOMIC don't sleep.
1944 * @order:Order of the GFP allocation.
1945 * @vma: Pointer to VMA or NULL if not available.
1946 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1947 * @node: Which node to prefer for allocation (modulo policy).
1948 * @hugepage: for hugepages try only the preferred node if possible
1950 * This function allocates a page from the kernel page pool and applies
1951 * a NUMA policy associated with the VMA or the current process.
1952 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1953 * mm_struct of the VMA to prevent it from going away. Should be used for
1954 * all allocations for pages that will be mapped into user space. Returns
1955 * NULL when no page can be allocated.
1958 alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
1959 unsigned long addr, int node, bool hugepage)
1961 struct mempolicy *pol;
1964 unsigned int cpuset_mems_cookie;
1968 pol = get_vma_policy(vma, addr);
1969 cpuset_mems_cookie = read_mems_allowed_begin();
1971 if (pol->mode == MPOL_INTERLEAVE) {
1974 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
1976 page = alloc_page_interleave(gfp, order, nid);
1980 if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && hugepage)) {
1981 int hpage_node = node;
1984 * For hugepage allocation and non-interleave policy which
1985 * allows the current node (or other explicitly preferred
1986 * node) we only try to allocate from the current/preferred
1987 * node and don't fall back to other nodes, as the cost of
1988 * remote accesses would likely offset THP benefits.
1990 * If the policy is interleave, or does not allow the current
1991 * node in its nodemask, we allocate the standard way.
1993 if (pol->mode == MPOL_PREFERRED &&
1994 !(pol->flags & MPOL_F_LOCAL))
1995 hpage_node = pol->v.preferred_node;
1997 nmask = policy_nodemask(gfp, pol);
1998 if (!nmask || node_isset(hpage_node, *nmask)) {
2000 page = __alloc_pages_node(hpage_node,
2001 gfp | __GFP_THISNODE, order);
2006 nmask = policy_nodemask(gfp, pol);
2007 preferred_nid = policy_node(gfp, pol, node);
2008 page = __alloc_pages_nodemask(gfp, order, preferred_nid, nmask);
2011 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
2017 * alloc_pages_current - Allocate pages.
2020 * %GFP_USER user allocation,
2021 * %GFP_KERNEL kernel allocation,
2022 * %GFP_HIGHMEM highmem allocation,
2023 * %GFP_FS don't call back into a file system.
2024 * %GFP_ATOMIC don't sleep.
2025 * @order: Power of two of allocation size in pages. 0 is a single page.
2027 * Allocate a page from the kernel page pool. When not in
2028 * interrupt context and apply the current process NUMA policy.
2029 * Returns NULL when no page can be allocated.
2031 * Don't call cpuset_update_task_memory_state() unless
2032 * 1) it's ok to take cpuset_sem (can WAIT), and
2033 * 2) allocating for current task (not interrupt).
2035 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
2037 struct mempolicy *pol = &default_policy;
2039 unsigned int cpuset_mems_cookie;
2041 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2042 pol = get_task_policy(current);
2045 cpuset_mems_cookie = read_mems_allowed_begin();
2048 * No reference counting needed for current->mempolicy
2049 * nor system default_policy
2051 if (pol->mode == MPOL_INTERLEAVE)
2052 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2054 page = __alloc_pages_nodemask(gfp, order,
2055 policy_node(gfp, pol, numa_node_id()),
2056 policy_nodemask(gfp, pol));
2058 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
2063 EXPORT_SYMBOL(alloc_pages_current);
2065 int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2067 struct mempolicy *pol = mpol_dup(vma_policy(src));
2070 return PTR_ERR(pol);
2071 dst->vm_policy = pol;
2076 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2077 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2078 * with the mems_allowed returned by cpuset_mems_allowed(). This
2079 * keeps mempolicies cpuset relative after its cpuset moves. See
2080 * further kernel/cpuset.c update_nodemask().
2082 * current's mempolicy may be rebinded by the other task(the task that changes
2083 * cpuset's mems), so we needn't do rebind work for current task.
2086 /* Slow path of a mempolicy duplicate */
2087 struct mempolicy *__mpol_dup(struct mempolicy *old)
2089 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2092 return ERR_PTR(-ENOMEM);
2094 /* task's mempolicy is protected by alloc_lock */
2095 if (old == current->mempolicy) {
2098 task_unlock(current);
2102 if (current_cpuset_is_being_rebound()) {
2103 nodemask_t mems = cpuset_mems_allowed(current);
2104 if (new->flags & MPOL_F_REBINDING)
2105 mpol_rebind_policy(new, &mems, MPOL_REBIND_STEP2);
2107 mpol_rebind_policy(new, &mems, MPOL_REBIND_ONCE);
2109 atomic_set(&new->refcnt, 1);
2113 /* Slow path of a mempolicy comparison */
2114 bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2118 if (a->mode != b->mode)
2120 if (a->flags != b->flags)
2122 if (mpol_store_user_nodemask(a))
2123 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2129 case MPOL_INTERLEAVE:
2130 return !!nodes_equal(a->v.nodes, b->v.nodes);
2131 case MPOL_PREFERRED:
2132 return a->v.preferred_node == b->v.preferred_node;
2140 * Shared memory backing store policy support.
2142 * Remember policies even when nobody has shared memory mapped.
2143 * The policies are kept in Red-Black tree linked from the inode.
2144 * They are protected by the sp->lock rwlock, which should be held
2145 * for any accesses to the tree.
2149 * lookup first element intersecting start-end. Caller holds sp->lock for
2150 * reading or for writing
2152 static struct sp_node *
2153 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2155 struct rb_node *n = sp->root.rb_node;
2158 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2160 if (start >= p->end)
2162 else if (end <= p->start)
2170 struct sp_node *w = NULL;
2171 struct rb_node *prev = rb_prev(n);
2174 w = rb_entry(prev, struct sp_node, nd);
2175 if (w->end <= start)
2179 return rb_entry(n, struct sp_node, nd);
2183 * Insert a new shared policy into the list. Caller holds sp->lock for
2186 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2188 struct rb_node **p = &sp->root.rb_node;
2189 struct rb_node *parent = NULL;
2194 nd = rb_entry(parent, struct sp_node, nd);
2195 if (new->start < nd->start)
2197 else if (new->end > nd->end)
2198 p = &(*p)->rb_right;
2202 rb_link_node(&new->nd, parent, p);
2203 rb_insert_color(&new->nd, &sp->root);
2204 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2205 new->policy ? new->policy->mode : 0);
2208 /* Find shared policy intersecting idx */
2210 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2212 struct mempolicy *pol = NULL;
2215 if (!sp->root.rb_node)
2217 read_lock(&sp->lock);
2218 sn = sp_lookup(sp, idx, idx+1);
2220 mpol_get(sn->policy);
2223 read_unlock(&sp->lock);
2227 static void sp_free(struct sp_node *n)
2229 mpol_put(n->policy);
2230 kmem_cache_free(sn_cache, n);
2234 * mpol_misplaced - check whether current page node is valid in policy
2236 * @page: page to be checked
2237 * @vma: vm area where page mapped
2238 * @addr: virtual address where page mapped
2240 * Lookup current policy node id for vma,addr and "compare to" page's
2244 * -1 - not misplaced, page is in the right node
2245 * node - node id where the page should be
2247 * Policy determination "mimics" alloc_page_vma().
2248 * Called from fault path where we know the vma and faulting address.
2250 int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2252 struct mempolicy *pol;
2254 int curnid = page_to_nid(page);
2255 unsigned long pgoff;
2256 int thiscpu = raw_smp_processor_id();
2257 int thisnid = cpu_to_node(thiscpu);
2263 pol = get_vma_policy(vma, addr);
2264 if (!(pol->flags & MPOL_F_MOF))
2267 switch (pol->mode) {
2268 case MPOL_INTERLEAVE:
2269 BUG_ON(addr >= vma->vm_end);
2270 BUG_ON(addr < vma->vm_start);
2272 pgoff = vma->vm_pgoff;
2273 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2274 polnid = offset_il_node(pol, vma, pgoff);
2277 case MPOL_PREFERRED:
2278 if (pol->flags & MPOL_F_LOCAL)
2279 polnid = numa_node_id();
2281 polnid = pol->v.preferred_node;
2287 * allows binding to multiple nodes.
2288 * use current page if in policy nodemask,
2289 * else select nearest allowed node, if any.
2290 * If no allowed nodes, use current [!misplaced].
2292 if (node_isset(curnid, pol->v.nodes))
2294 z = first_zones_zonelist(
2295 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2296 gfp_zone(GFP_HIGHUSER),
2298 polnid = z->zone->node;
2305 /* Migrate the page towards the node whose CPU is referencing it */
2306 if (pol->flags & MPOL_F_MORON) {
2309 if (!should_numa_migrate_memory(current, page, curnid, thiscpu))
2313 if (curnid != polnid)
2322 * Drop the (possibly final) reference to task->mempolicy. It needs to be
2323 * dropped after task->mempolicy is set to NULL so that any allocation done as
2324 * part of its kmem_cache_free(), such as by KASAN, doesn't reference a freed
2327 void mpol_put_task_policy(struct task_struct *task)
2329 struct mempolicy *pol;
2332 pol = task->mempolicy;
2333 task->mempolicy = NULL;
2338 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2340 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2341 rb_erase(&n->nd, &sp->root);
2345 static void sp_node_init(struct sp_node *node, unsigned long start,
2346 unsigned long end, struct mempolicy *pol)
2348 node->start = start;
2353 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2354 struct mempolicy *pol)
2357 struct mempolicy *newpol;
2359 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2363 newpol = mpol_dup(pol);
2364 if (IS_ERR(newpol)) {
2365 kmem_cache_free(sn_cache, n);
2368 newpol->flags |= MPOL_F_SHARED;
2369 sp_node_init(n, start, end, newpol);
2374 /* Replace a policy range. */
2375 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2376 unsigned long end, struct sp_node *new)
2379 struct sp_node *n_new = NULL;
2380 struct mempolicy *mpol_new = NULL;
2384 write_lock(&sp->lock);
2385 n = sp_lookup(sp, start, end);
2386 /* Take care of old policies in the same range. */
2387 while (n && n->start < end) {
2388 struct rb_node *next = rb_next(&n->nd);
2389 if (n->start >= start) {
2395 /* Old policy spanning whole new range. */
2400 *mpol_new = *n->policy;
2401 atomic_set(&mpol_new->refcnt, 1);
2402 sp_node_init(n_new, end, n->end, mpol_new);
2404 sp_insert(sp, n_new);
2413 n = rb_entry(next, struct sp_node, nd);
2417 write_unlock(&sp->lock);
2424 kmem_cache_free(sn_cache, n_new);
2429 write_unlock(&sp->lock);
2431 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2434 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2441 * mpol_shared_policy_init - initialize shared policy for inode
2442 * @sp: pointer to inode shared policy
2443 * @mpol: struct mempolicy to install
2445 * Install non-NULL @mpol in inode's shared policy rb-tree.
2446 * On entry, the current task has a reference on a non-NULL @mpol.
2447 * This must be released on exit.
2448 * This is called at get_inode() calls and we can use GFP_KERNEL.
2450 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2454 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2455 rwlock_init(&sp->lock);
2458 struct vm_area_struct pvma;
2459 struct mempolicy *new;
2460 NODEMASK_SCRATCH(scratch);
2464 /* contextualize the tmpfs mount point mempolicy */
2465 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2467 goto free_scratch; /* no valid nodemask intersection */
2470 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2471 task_unlock(current);
2475 /* Create pseudo-vma that contains just the policy */
2476 memset(&pvma, 0, sizeof(struct vm_area_struct));
2477 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2478 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2481 mpol_put(new); /* drop initial ref */
2483 NODEMASK_SCRATCH_FREE(scratch);
2485 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2489 int mpol_set_shared_policy(struct shared_policy *info,
2490 struct vm_area_struct *vma, struct mempolicy *npol)
2493 struct sp_node *new = NULL;
2494 unsigned long sz = vma_pages(vma);
2496 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2498 sz, npol ? npol->mode : -1,
2499 npol ? npol->flags : -1,
2500 npol ? nodes_addr(npol->v.nodes)[0] : NUMA_NO_NODE);
2503 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2507 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2513 /* Free a backing policy store on inode delete. */
2514 void mpol_free_shared_policy(struct shared_policy *p)
2517 struct rb_node *next;
2519 if (!p->root.rb_node)
2521 write_lock(&p->lock);
2522 next = rb_first(&p->root);
2524 n = rb_entry(next, struct sp_node, nd);
2525 next = rb_next(&n->nd);
2528 write_unlock(&p->lock);
2531 #ifdef CONFIG_NUMA_BALANCING
2532 static int __initdata numabalancing_override;
2534 static void __init check_numabalancing_enable(void)
2536 bool numabalancing_default = false;
2538 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2539 numabalancing_default = true;
2541 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2542 if (numabalancing_override)
2543 set_numabalancing_state(numabalancing_override == 1);
2545 if (num_online_nodes() > 1 && !numabalancing_override) {
2546 pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
2547 numabalancing_default ? "Enabling" : "Disabling");
2548 set_numabalancing_state(numabalancing_default);
2552 static int __init setup_numabalancing(char *str)
2558 if (!strcmp(str, "enable")) {
2559 numabalancing_override = 1;
2561 } else if (!strcmp(str, "disable")) {
2562 numabalancing_override = -1;
2567 pr_warn("Unable to parse numa_balancing=\n");
2571 __setup("numa_balancing=", setup_numabalancing);
2573 static inline void __init check_numabalancing_enable(void)
2576 #endif /* CONFIG_NUMA_BALANCING */
2578 /* assumes fs == KERNEL_DS */
2579 void __init numa_policy_init(void)
2581 nodemask_t interleave_nodes;
2582 unsigned long largest = 0;
2583 int nid, prefer = 0;
2585 policy_cache = kmem_cache_create("numa_policy",
2586 sizeof(struct mempolicy),
2587 0, SLAB_PANIC, NULL);
2589 sn_cache = kmem_cache_create("shared_policy_node",
2590 sizeof(struct sp_node),
2591 0, SLAB_PANIC, NULL);
2593 for_each_node(nid) {
2594 preferred_node_policy[nid] = (struct mempolicy) {
2595 .refcnt = ATOMIC_INIT(1),
2596 .mode = MPOL_PREFERRED,
2597 .flags = MPOL_F_MOF | MPOL_F_MORON,
2598 .v = { .preferred_node = nid, },
2603 * Set interleaving policy for system init. Interleaving is only
2604 * enabled across suitably sized nodes (default is >= 16MB), or
2605 * fall back to the largest node if they're all smaller.
2607 nodes_clear(interleave_nodes);
2608 for_each_node_state(nid, N_MEMORY) {
2609 unsigned long total_pages = node_present_pages(nid);
2611 /* Preserve the largest node */
2612 if (largest < total_pages) {
2613 largest = total_pages;
2617 /* Interleave this node? */
2618 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2619 node_set(nid, interleave_nodes);
2622 /* All too small, use the largest */
2623 if (unlikely(nodes_empty(interleave_nodes)))
2624 node_set(prefer, interleave_nodes);
2626 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2627 pr_err("%s: interleaving failed\n", __func__);
2629 check_numabalancing_enable();
2632 /* Reset policy of current process to default */
2633 void numa_default_policy(void)
2635 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2639 * Parse and format mempolicy from/to strings
2643 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
2645 static const char * const policy_modes[] =
2647 [MPOL_DEFAULT] = "default",
2648 [MPOL_PREFERRED] = "prefer",
2649 [MPOL_BIND] = "bind",
2650 [MPOL_INTERLEAVE] = "interleave",
2651 [MPOL_LOCAL] = "local",
2657 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2658 * @str: string containing mempolicy to parse
2659 * @mpol: pointer to struct mempolicy pointer, returned on success.
2662 * <mode>[=<flags>][:<nodelist>]
2664 * On success, returns 0, else 1
2666 int mpol_parse_str(char *str, struct mempolicy **mpol)
2668 struct mempolicy *new = NULL;
2669 unsigned short mode;
2670 unsigned short mode_flags;
2672 char *nodelist = strchr(str, ':');
2673 char *flags = strchr(str, '=');
2677 /* NUL-terminate mode or flags string */
2679 if (nodelist_parse(nodelist, nodes))
2681 if (!nodes_subset(nodes, node_states[N_MEMORY]))
2687 *flags++ = '\0'; /* terminate mode string */
2689 for (mode = 0; mode < MPOL_MAX; mode++) {
2690 if (!strcmp(str, policy_modes[mode])) {
2694 if (mode >= MPOL_MAX)
2698 case MPOL_PREFERRED:
2700 * Insist on a nodelist of one node only
2703 char *rest = nodelist;
2704 while (isdigit(*rest))
2710 case MPOL_INTERLEAVE:
2712 * Default to online nodes with memory if no nodelist
2715 nodes = node_states[N_MEMORY];
2719 * Don't allow a nodelist; mpol_new() checks flags
2723 mode = MPOL_PREFERRED;
2727 * Insist on a empty nodelist
2734 * Insist on a nodelist
2743 * Currently, we only support two mutually exclusive
2746 if (!strcmp(flags, "static"))
2747 mode_flags |= MPOL_F_STATIC_NODES;
2748 else if (!strcmp(flags, "relative"))
2749 mode_flags |= MPOL_F_RELATIVE_NODES;
2754 new = mpol_new(mode, mode_flags, &nodes);
2759 * Save nodes for mpol_to_str() to show the tmpfs mount options
2760 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2762 if (mode != MPOL_PREFERRED)
2763 new->v.nodes = nodes;
2765 new->v.preferred_node = first_node(nodes);
2767 new->flags |= MPOL_F_LOCAL;
2770 * Save nodes for contextualization: this will be used to "clone"
2771 * the mempolicy in a specific context [cpuset] at a later time.
2773 new->w.user_nodemask = nodes;
2778 /* Restore string for error message */
2787 #endif /* CONFIG_TMPFS */
2790 * mpol_to_str - format a mempolicy structure for printing
2791 * @buffer: to contain formatted mempolicy string
2792 * @maxlen: length of @buffer
2793 * @pol: pointer to mempolicy to be formatted
2795 * Convert @pol into a string. If @buffer is too short, truncate the string.
2796 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
2797 * longest flag, "relative", and to display at least a few node ids.
2799 void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
2802 nodemask_t nodes = NODE_MASK_NONE;
2803 unsigned short mode = MPOL_DEFAULT;
2804 unsigned short flags = 0;
2806 if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
2814 case MPOL_PREFERRED:
2815 if (flags & MPOL_F_LOCAL)
2818 node_set(pol->v.preferred_node, nodes);
2821 case MPOL_INTERLEAVE:
2822 nodes = pol->v.nodes;
2826 snprintf(p, maxlen, "unknown");
2830 p += snprintf(p, maxlen, "%s", policy_modes[mode]);
2832 if (flags & MPOL_MODE_FLAGS) {
2833 p += snprintf(p, buffer + maxlen - p, "=");
2836 * Currently, the only defined flags are mutually exclusive
2838 if (flags & MPOL_F_STATIC_NODES)
2839 p += snprintf(p, buffer + maxlen - p, "static");
2840 else if (flags & MPOL_F_RELATIVE_NODES)
2841 p += snprintf(p, buffer + maxlen - p, "relative");
2844 if (!nodes_empty(nodes))
2845 p += scnprintf(p, buffer + maxlen - p, ":%*pbl",
2846 nodemask_pr_args(&nodes));