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>
100 #include <linux/swapops.h>
102 #include <asm/tlbflush.h>
103 #include <linux/uaccess.h>
105 #include "internal.h"
108 #define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
109 #define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
111 static struct kmem_cache
*policy_cache
;
112 static struct kmem_cache
*sn_cache
;
114 /* Highest zone. An specific allocation for a zone below that is not
116 enum zone_type policy_zone
= 0;
119 * run-time system-wide default policy => local allocation
121 static struct mempolicy default_policy
= {
122 .refcnt
= ATOMIC_INIT(1), /* never free it */
123 .mode
= MPOL_PREFERRED
,
124 .flags
= MPOL_F_LOCAL
,
127 static struct mempolicy preferred_node_policy
[MAX_NUMNODES
];
129 struct mempolicy
*get_task_policy(struct task_struct
*p
)
131 struct mempolicy
*pol
= p
->mempolicy
;
137 node
= numa_node_id();
138 if (node
!= NUMA_NO_NODE
) {
139 pol
= &preferred_node_policy
[node
];
140 /* preferred_node_policy is not initialised early in boot */
145 return &default_policy
;
148 static const struct mempolicy_operations
{
149 int (*create
)(struct mempolicy
*pol
, const nodemask_t
*nodes
);
150 void (*rebind
)(struct mempolicy
*pol
, const nodemask_t
*nodes
);
151 } mpol_ops
[MPOL_MAX
];
153 static inline int mpol_store_user_nodemask(const struct mempolicy
*pol
)
155 return pol
->flags
& MPOL_MODE_FLAGS
;
158 static void mpol_relative_nodemask(nodemask_t
*ret
, const nodemask_t
*orig
,
159 const nodemask_t
*rel
)
162 nodes_fold(tmp
, *orig
, nodes_weight(*rel
));
163 nodes_onto(*ret
, tmp
, *rel
);
166 static int mpol_new_interleave(struct mempolicy
*pol
, const nodemask_t
*nodes
)
168 if (nodes_empty(*nodes
))
170 pol
->v
.nodes
= *nodes
;
174 static int mpol_new_preferred(struct mempolicy
*pol
, const nodemask_t
*nodes
)
177 pol
->flags
|= MPOL_F_LOCAL
; /* local allocation */
178 else if (nodes_empty(*nodes
))
179 return -EINVAL
; /* no allowed nodes */
181 pol
->v
.preferred_node
= first_node(*nodes
);
185 static int mpol_new_bind(struct mempolicy
*pol
, const nodemask_t
*nodes
)
187 if (nodes_empty(*nodes
))
189 pol
->v
.nodes
= *nodes
;
194 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
195 * any, for the new policy. mpol_new() has already validated the nodes
196 * parameter with respect to the policy mode and flags. But, we need to
197 * handle an empty nodemask with MPOL_PREFERRED here.
199 * Must be called holding task's alloc_lock to protect task's mems_allowed
200 * and mempolicy. May also be called holding the mmap_semaphore for write.
202 static int mpol_set_nodemask(struct mempolicy
*pol
,
203 const nodemask_t
*nodes
, struct nodemask_scratch
*nsc
)
207 /* if mode is MPOL_DEFAULT, pol is NULL. This is right. */
211 nodes_and(nsc
->mask1
,
212 cpuset_current_mems_allowed
, node_states
[N_MEMORY
]);
215 if (pol
->mode
== MPOL_PREFERRED
&& nodes_empty(*nodes
))
216 nodes
= NULL
; /* explicit local allocation */
218 if (pol
->flags
& MPOL_F_RELATIVE_NODES
)
219 mpol_relative_nodemask(&nsc
->mask2
, nodes
, &nsc
->mask1
);
221 nodes_and(nsc
->mask2
, *nodes
, nsc
->mask1
);
223 if (mpol_store_user_nodemask(pol
))
224 pol
->w
.user_nodemask
= *nodes
;
226 pol
->w
.cpuset_mems_allowed
=
227 cpuset_current_mems_allowed
;
231 ret
= mpol_ops
[pol
->mode
].create(pol
, &nsc
->mask2
);
233 ret
= mpol_ops
[pol
->mode
].create(pol
, NULL
);
238 * This function just creates a new policy, does some check and simple
239 * initialization. You must invoke mpol_set_nodemask() to set nodes.
241 static struct mempolicy
*mpol_new(unsigned short mode
, unsigned short flags
,
244 struct mempolicy
*policy
;
246 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
247 mode
, flags
, nodes
? nodes_addr(*nodes
)[0] : NUMA_NO_NODE
);
249 if (mode
== MPOL_DEFAULT
) {
250 if (nodes
&& !nodes_empty(*nodes
))
251 return ERR_PTR(-EINVAL
);
257 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
258 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
259 * All other modes require a valid pointer to a non-empty nodemask.
261 if (mode
== MPOL_PREFERRED
) {
262 if (nodes_empty(*nodes
)) {
263 if (((flags
& MPOL_F_STATIC_NODES
) ||
264 (flags
& MPOL_F_RELATIVE_NODES
)))
265 return ERR_PTR(-EINVAL
);
267 } else if (mode
== MPOL_LOCAL
) {
268 if (!nodes_empty(*nodes
) ||
269 (flags
& MPOL_F_STATIC_NODES
) ||
270 (flags
& MPOL_F_RELATIVE_NODES
))
271 return ERR_PTR(-EINVAL
);
272 mode
= MPOL_PREFERRED
;
273 } else if (nodes_empty(*nodes
))
274 return ERR_PTR(-EINVAL
);
275 policy
= kmem_cache_alloc(policy_cache
, GFP_KERNEL
);
277 return ERR_PTR(-ENOMEM
);
278 atomic_set(&policy
->refcnt
, 1);
280 policy
->flags
= flags
;
285 /* Slow path of a mpol destructor. */
286 void __mpol_put(struct mempolicy
*p
)
288 if (!atomic_dec_and_test(&p
->refcnt
))
290 kmem_cache_free(policy_cache
, p
);
293 static void mpol_rebind_default(struct mempolicy
*pol
, const nodemask_t
*nodes
)
297 static void mpol_rebind_nodemask(struct mempolicy
*pol
, const nodemask_t
*nodes
)
301 if (pol
->flags
& MPOL_F_STATIC_NODES
)
302 nodes_and(tmp
, pol
->w
.user_nodemask
, *nodes
);
303 else if (pol
->flags
& MPOL_F_RELATIVE_NODES
)
304 mpol_relative_nodemask(&tmp
, &pol
->w
.user_nodemask
, nodes
);
306 nodes_remap(tmp
, pol
->v
.nodes
,pol
->w
.cpuset_mems_allowed
,
308 pol
->w
.cpuset_mems_allowed
= tmp
;
311 if (nodes_empty(tmp
))
317 static void mpol_rebind_preferred(struct mempolicy
*pol
,
318 const nodemask_t
*nodes
)
322 if (pol
->flags
& MPOL_F_STATIC_NODES
) {
323 int node
= first_node(pol
->w
.user_nodemask
);
325 if (node_isset(node
, *nodes
)) {
326 pol
->v
.preferred_node
= node
;
327 pol
->flags
&= ~MPOL_F_LOCAL
;
329 pol
->flags
|= MPOL_F_LOCAL
;
330 } else if (pol
->flags
& MPOL_F_RELATIVE_NODES
) {
331 mpol_relative_nodemask(&tmp
, &pol
->w
.user_nodemask
, nodes
);
332 pol
->v
.preferred_node
= first_node(tmp
);
333 } else if (!(pol
->flags
& MPOL_F_LOCAL
)) {
334 pol
->v
.preferred_node
= node_remap(pol
->v
.preferred_node
,
335 pol
->w
.cpuset_mems_allowed
,
337 pol
->w
.cpuset_mems_allowed
= *nodes
;
342 * mpol_rebind_policy - Migrate a policy to a different set of nodes
344 * Per-vma policies are protected by mmap_sem. Allocations using per-task
345 * policies are protected by task->mems_allowed_seq to prevent a premature
346 * OOM/allocation failure due to parallel nodemask modification.
348 static void mpol_rebind_policy(struct mempolicy
*pol
, const nodemask_t
*newmask
)
352 if (!mpol_store_user_nodemask(pol
) && !(pol
->flags
& MPOL_F_LOCAL
) &&
353 nodes_equal(pol
->w
.cpuset_mems_allowed
, *newmask
))
356 mpol_ops
[pol
->mode
].rebind(pol
, newmask
);
360 * Wrapper for mpol_rebind_policy() that just requires task
361 * pointer, and updates task mempolicy.
363 * Called with task's alloc_lock held.
366 void mpol_rebind_task(struct task_struct
*tsk
, const nodemask_t
*new)
368 mpol_rebind_policy(tsk
->mempolicy
, new);
372 * Rebind each vma in mm to new nodemask.
374 * Call holding a reference to mm. Takes mm->mmap_sem during call.
377 void mpol_rebind_mm(struct mm_struct
*mm
, nodemask_t
*new)
379 struct vm_area_struct
*vma
;
381 down_write(&mm
->mmap_sem
);
382 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
)
383 mpol_rebind_policy(vma
->vm_policy
, new);
384 up_write(&mm
->mmap_sem
);
387 static const struct mempolicy_operations mpol_ops
[MPOL_MAX
] = {
389 .rebind
= mpol_rebind_default
,
391 [MPOL_INTERLEAVE
] = {
392 .create
= mpol_new_interleave
,
393 .rebind
= mpol_rebind_nodemask
,
396 .create
= mpol_new_preferred
,
397 .rebind
= mpol_rebind_preferred
,
400 .create
= mpol_new_bind
,
401 .rebind
= mpol_rebind_nodemask
,
405 static void migrate_page_add(struct page
*page
, struct list_head
*pagelist
,
406 unsigned long flags
);
409 struct list_head
*pagelist
;
412 struct vm_area_struct
*prev
;
416 * Check if the page's nid is in qp->nmask.
418 * If MPOL_MF_INVERT is set in qp->flags, check if the nid is
419 * in the invert of qp->nmask.
421 static inline bool queue_pages_required(struct page
*page
,
422 struct queue_pages
*qp
)
424 int nid
= page_to_nid(page
);
425 unsigned long flags
= qp
->flags
;
427 return node_isset(nid
, *qp
->nmask
) == !(flags
& MPOL_MF_INVERT
);
431 * queue_pages_pmd() has three possible return values:
432 * 1 - pages are placed on the right node or queued successfully.
434 * -EIO - is migration entry or MPOL_MF_STRICT was specified and an existing
435 * page was already on a node that does not follow the policy.
437 static int queue_pages_pmd(pmd_t
*pmd
, spinlock_t
*ptl
, unsigned long addr
,
438 unsigned long end
, struct mm_walk
*walk
)
442 struct queue_pages
*qp
= walk
->private;
445 if (unlikely(is_pmd_migration_entry(*pmd
))) {
449 page
= pmd_page(*pmd
);
450 if (is_huge_zero_page(page
)) {
452 __split_huge_pmd(walk
->vma
, pmd
, addr
, false, NULL
);
455 if (!thp_migration_supported()) {
459 ret
= split_huge_page(page
);
464 if (!queue_pages_required(page
, qp
)) {
471 /* go to thp migration */
472 if (flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
)) {
473 if (!vma_migratable(walk
->vma
)) {
478 migrate_page_add(page
, qp
->pagelist
, flags
);
488 * Scan through pages checking if pages follow certain conditions,
489 * and move them to the pagelist if they do.
491 static int queue_pages_pte_range(pmd_t
*pmd
, unsigned long addr
,
492 unsigned long end
, struct mm_walk
*walk
)
494 struct vm_area_struct
*vma
= walk
->vma
;
496 struct queue_pages
*qp
= walk
->private;
497 unsigned long flags
= qp
->flags
;
502 ptl
= pmd_trans_huge_lock(pmd
, vma
);
504 ret
= queue_pages_pmd(pmd
, ptl
, addr
, end
, walk
);
511 if (pmd_trans_unstable(pmd
))
514 pte
= pte_offset_map_lock(walk
->mm
, pmd
, addr
, &ptl
);
515 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
) {
516 if (!pte_present(*pte
))
518 page
= vm_normal_page(vma
, addr
, *pte
);
522 * vm_normal_page() filters out zero pages, but there might
523 * still be PageReserved pages to skip, perhaps in a VDSO.
525 if (PageReserved(page
))
527 if (!queue_pages_required(page
, qp
))
529 if (PageTransCompound(page
) && !thp_migration_supported()) {
531 pte_unmap_unlock(pte
, ptl
);
533 ret
= split_huge_page(page
);
536 /* Failed to split -- skip. */
538 pte
= pte_offset_map_lock(walk
->mm
, pmd
,
545 if (flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
)) {
546 if (!vma_migratable(vma
))
548 migrate_page_add(page
, qp
->pagelist
, flags
);
552 pte_unmap_unlock(pte
- 1, ptl
);
554 return addr
!= end
? -EIO
: 0;
557 static int queue_pages_hugetlb(pte_t
*pte
, unsigned long hmask
,
558 unsigned long addr
, unsigned long end
,
559 struct mm_walk
*walk
)
561 #ifdef CONFIG_HUGETLB_PAGE
562 struct queue_pages
*qp
= walk
->private;
563 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 if (!queue_pages_required(page
, qp
))
575 /* With MPOL_MF_MOVE, we migrate only unshared hugepage. */
576 if (flags
& (MPOL_MF_MOVE_ALL
) ||
577 (flags
& MPOL_MF_MOVE
&& page_mapcount(page
) == 1))
578 isolate_huge_page(page
, qp
->pagelist
);
587 #ifdef CONFIG_NUMA_BALANCING
589 * This is used to mark a range of virtual addresses to be inaccessible.
590 * These are later cleared by a NUMA hinting fault. Depending on these
591 * faults, pages may be migrated for better NUMA placement.
593 * This is assuming that NUMA faults are handled using PROT_NONE. If
594 * an architecture makes a different choice, it will need further
595 * changes to the core.
597 unsigned long change_prot_numa(struct vm_area_struct
*vma
,
598 unsigned long addr
, unsigned long end
)
602 nr_updated
= change_protection(vma
, addr
, end
, PAGE_NONE
, 0, 1);
604 count_vm_numa_events(NUMA_PTE_UPDATES
, nr_updated
);
609 static unsigned long change_prot_numa(struct vm_area_struct
*vma
,
610 unsigned long addr
, unsigned long end
)
614 #endif /* CONFIG_NUMA_BALANCING */
616 static int queue_pages_test_walk(unsigned long start
, unsigned long end
,
617 struct mm_walk
*walk
)
619 struct vm_area_struct
*vma
= walk
->vma
;
620 struct queue_pages
*qp
= walk
->private;
621 unsigned long endvma
= vma
->vm_end
;
622 unsigned long flags
= qp
->flags
;
625 * Need check MPOL_MF_STRICT to return -EIO if possible
626 * regardless of vma_migratable
628 if (!vma_migratable(vma
) &&
629 !(flags
& MPOL_MF_STRICT
))
634 if (vma
->vm_start
> start
)
635 start
= vma
->vm_start
;
637 if (!(flags
& MPOL_MF_DISCONTIG_OK
)) {
638 if (!vma
->vm_next
&& vma
->vm_end
< end
)
640 if (qp
->prev
&& qp
->prev
->vm_end
< vma
->vm_start
)
646 if (flags
& MPOL_MF_LAZY
) {
647 /* Similar to task_numa_work, skip inaccessible VMAs */
648 if (!is_vm_hugetlb_page(vma
) &&
649 (vma
->vm_flags
& (VM_READ
| VM_EXEC
| VM_WRITE
)) &&
650 !(vma
->vm_flags
& VM_MIXEDMAP
))
651 change_prot_numa(vma
, start
, endvma
);
655 /* queue pages from current vma */
656 if (flags
& MPOL_MF_VALID
)
662 * Walk through page tables and collect pages to be migrated.
664 * If pages found in a given range are on a set of nodes (determined by
665 * @nodes and @flags,) it's isolated and queued to the pagelist which is
666 * passed via @private.)
669 queue_pages_range(struct mm_struct
*mm
, unsigned long start
, unsigned long end
,
670 nodemask_t
*nodes
, unsigned long flags
,
671 struct list_head
*pagelist
)
673 struct queue_pages qp
= {
674 .pagelist
= pagelist
,
679 struct mm_walk queue_pages_walk
= {
680 .hugetlb_entry
= queue_pages_hugetlb
,
681 .pmd_entry
= queue_pages_pte_range
,
682 .test_walk
= queue_pages_test_walk
,
687 return walk_page_range(start
, end
, &queue_pages_walk
);
691 * Apply policy to a single VMA
692 * This must be called with the mmap_sem held for writing.
694 static int vma_replace_policy(struct vm_area_struct
*vma
,
695 struct mempolicy
*pol
)
698 struct mempolicy
*old
;
699 struct mempolicy
*new;
701 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
702 vma
->vm_start
, vma
->vm_end
, vma
->vm_pgoff
,
703 vma
->vm_ops
, vma
->vm_file
,
704 vma
->vm_ops
? vma
->vm_ops
->set_policy
: NULL
);
710 if (vma
->vm_ops
&& vma
->vm_ops
->set_policy
) {
711 err
= vma
->vm_ops
->set_policy(vma
, new);
716 old
= vma
->vm_policy
;
717 vma
->vm_policy
= new; /* protected by mmap_sem */
726 /* Step 2: apply policy to a range and do splits. */
727 static int mbind_range(struct mm_struct
*mm
, unsigned long start
,
728 unsigned long end
, struct mempolicy
*new_pol
)
730 struct vm_area_struct
*next
;
731 struct vm_area_struct
*prev
;
732 struct vm_area_struct
*vma
;
735 unsigned long vmstart
;
738 vma
= find_vma(mm
, start
);
739 if (!vma
|| vma
->vm_start
> start
)
743 if (start
> vma
->vm_start
)
746 for (; vma
&& vma
->vm_start
< end
; prev
= vma
, vma
= next
) {
748 vmstart
= max(start
, vma
->vm_start
);
749 vmend
= min(end
, vma
->vm_end
);
751 if (mpol_equal(vma_policy(vma
), new_pol
))
754 pgoff
= vma
->vm_pgoff
+
755 ((vmstart
- vma
->vm_start
) >> PAGE_SHIFT
);
756 prev
= vma_merge(mm
, prev
, vmstart
, vmend
, vma
->vm_flags
,
757 vma
->anon_vma
, vma
->vm_file
, pgoff
,
758 new_pol
, vma
->vm_userfaultfd_ctx
,
759 vma_get_anon_name(vma
));
763 if (mpol_equal(vma_policy(vma
), new_pol
))
765 /* vma_merge() joined vma && vma->next, case 8 */
768 if (vma
->vm_start
!= vmstart
) {
769 err
= split_vma(vma
->vm_mm
, vma
, vmstart
, 1);
773 if (vma
->vm_end
!= vmend
) {
774 err
= split_vma(vma
->vm_mm
, vma
, vmend
, 0);
779 err
= vma_replace_policy(vma
, new_pol
);
788 /* Set the process memory policy */
789 static long do_set_mempolicy(unsigned short mode
, unsigned short flags
,
792 struct mempolicy
*new, *old
;
793 NODEMASK_SCRATCH(scratch
);
799 new = mpol_new(mode
, flags
, nodes
);
806 ret
= mpol_set_nodemask(new, nodes
, scratch
);
808 task_unlock(current
);
812 old
= current
->mempolicy
;
813 current
->mempolicy
= new;
814 if (new && new->mode
== MPOL_INTERLEAVE
)
815 current
->il_prev
= MAX_NUMNODES
-1;
816 task_unlock(current
);
820 NODEMASK_SCRATCH_FREE(scratch
);
825 * Return nodemask for policy for get_mempolicy() query
827 * Called with task's alloc_lock held
829 static void get_policy_nodemask(struct mempolicy
*p
, nodemask_t
*nodes
)
832 if (p
== &default_policy
)
838 case MPOL_INTERLEAVE
:
842 if (!(p
->flags
& MPOL_F_LOCAL
))
843 node_set(p
->v
.preferred_node
, *nodes
);
844 /* else return empty node mask for local allocation */
851 static int lookup_node(unsigned long addr
)
856 err
= get_user_pages(addr
& PAGE_MASK
, 1, 0, &p
, NULL
);
858 err
= page_to_nid(p
);
864 /* Retrieve NUMA policy */
865 static long do_get_mempolicy(int *policy
, nodemask_t
*nmask
,
866 unsigned long addr
, unsigned long flags
)
869 struct mm_struct
*mm
= current
->mm
;
870 struct vm_area_struct
*vma
= NULL
;
871 struct mempolicy
*pol
= current
->mempolicy
;
874 ~(unsigned long)(MPOL_F_NODE
|MPOL_F_ADDR
|MPOL_F_MEMS_ALLOWED
))
877 if (flags
& MPOL_F_MEMS_ALLOWED
) {
878 if (flags
& (MPOL_F_NODE
|MPOL_F_ADDR
))
880 *policy
= 0; /* just so it's initialized */
882 *nmask
= cpuset_current_mems_allowed
;
883 task_unlock(current
);
887 if (flags
& MPOL_F_ADDR
) {
889 * Do NOT fall back to task policy if the
890 * vma/shared policy at addr is NULL. We
891 * want to return MPOL_DEFAULT in this case.
893 down_read(&mm
->mmap_sem
);
894 vma
= find_vma_intersection(mm
, addr
, addr
+1);
896 up_read(&mm
->mmap_sem
);
899 if (vma
->vm_ops
&& vma
->vm_ops
->get_policy
)
900 pol
= vma
->vm_ops
->get_policy(vma
, addr
);
902 pol
= vma
->vm_policy
;
907 pol
= &default_policy
; /* indicates default behavior */
909 if (flags
& MPOL_F_NODE
) {
910 if (flags
& MPOL_F_ADDR
) {
911 err
= lookup_node(addr
);
915 } else if (pol
== current
->mempolicy
&&
916 pol
->mode
== MPOL_INTERLEAVE
) {
917 *policy
= next_node_in(current
->il_prev
, pol
->v
.nodes
);
923 *policy
= pol
== &default_policy
? MPOL_DEFAULT
:
926 * Internal mempolicy flags must be masked off before exposing
927 * the policy to userspace.
929 *policy
|= (pol
->flags
& MPOL_MODE_FLAGS
);
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
952 * page migration, thp tail pages can be passed.
954 static void migrate_page_add(struct page
*page
, struct list_head
*pagelist
,
957 struct page
*head
= compound_head(page
);
959 * Avoid migrating a page that is shared with others.
961 if ((flags
& MPOL_MF_MOVE_ALL
) || page_mapcount(head
) == 1) {
962 if (!isolate_lru_page(head
)) {
963 list_add_tail(&head
->lru
, pagelist
);
964 mod_node_page_state(page_pgdat(head
),
965 NR_ISOLATED_ANON
+ page_is_file_cache(head
),
966 hpage_nr_pages(head
));
971 static struct page
*new_node_page(struct page
*page
, unsigned long node
, int **x
)
974 return alloc_huge_page_node(page_hstate(compound_head(page
)),
976 else if (thp_migration_supported() && PageTransHuge(page
)) {
979 thp
= alloc_pages_node(node
,
980 (GFP_TRANSHUGE
| __GFP_THISNODE
),
984 prep_transhuge_page(thp
);
987 return __alloc_pages_node(node
, GFP_HIGHUSER_MOVABLE
|
992 * Migrate pages from one node to a target node.
993 * Returns error or the number of pages not migrated.
995 static int migrate_to_node(struct mm_struct
*mm
, int source
, int dest
,
1003 node_set(source
, nmask
);
1006 * This does not "check" the range but isolates all pages that
1007 * need migration. Between passing in the full user address
1008 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1010 VM_BUG_ON(!(flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
)));
1011 queue_pages_range(mm
, mm
->mmap
->vm_start
, mm
->task_size
, &nmask
,
1012 flags
| MPOL_MF_DISCONTIG_OK
, &pagelist
);
1014 if (!list_empty(&pagelist
)) {
1015 err
= migrate_pages(&pagelist
, new_node_page
, NULL
, dest
,
1016 MIGRATE_SYNC
, MR_SYSCALL
);
1018 putback_movable_pages(&pagelist
);
1025 * Move pages between the two nodesets so as to preserve the physical
1026 * layout as much as possible.
1028 * Returns the number of page that could not be moved.
1030 int do_migrate_pages(struct mm_struct
*mm
, const nodemask_t
*from
,
1031 const nodemask_t
*to
, int flags
)
1037 err
= migrate_prep();
1041 down_read(&mm
->mmap_sem
);
1044 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1045 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1046 * bit in 'tmp', and return that <source, dest> pair for migration.
1047 * The pair of nodemasks 'to' and 'from' define the map.
1049 * If no pair of bits is found that way, fallback to picking some
1050 * pair of 'source' and 'dest' bits that are not the same. If the
1051 * 'source' and 'dest' bits are the same, this represents a node
1052 * that will be migrating to itself, so no pages need move.
1054 * If no bits are left in 'tmp', or if all remaining bits left
1055 * in 'tmp' correspond to the same bit in 'to', return false
1056 * (nothing left to migrate).
1058 * This lets us pick a pair of nodes to migrate between, such that
1059 * if possible the dest node is not already occupied by some other
1060 * source node, minimizing the risk of overloading the memory on a
1061 * node that would happen if we migrated incoming memory to a node
1062 * before migrating outgoing memory source that same node.
1064 * A single scan of tmp is sufficient. As we go, we remember the
1065 * most recent <s, d> pair that moved (s != d). If we find a pair
1066 * that not only moved, but what's better, moved to an empty slot
1067 * (d is not set in tmp), then we break out then, with that pair.
1068 * Otherwise when we finish scanning from_tmp, we at least have the
1069 * most recent <s, d> pair that moved. If we get all the way through
1070 * the scan of tmp without finding any node that moved, much less
1071 * moved to an empty node, then there is nothing left worth migrating.
1075 while (!nodes_empty(tmp
)) {
1077 int source
= NUMA_NO_NODE
;
1080 for_each_node_mask(s
, tmp
) {
1083 * do_migrate_pages() tries to maintain the relative
1084 * node relationship of the pages established between
1085 * threads and memory areas.
1087 * However if the number of source nodes is not equal to
1088 * the number of destination nodes we can not preserve
1089 * this node relative relationship. In that case, skip
1090 * copying memory from a node that is in the destination
1093 * Example: [2,3,4] -> [3,4,5] moves everything.
1094 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1097 if ((nodes_weight(*from
) != nodes_weight(*to
)) &&
1098 (node_isset(s
, *to
)))
1101 d
= node_remap(s
, *from
, *to
);
1105 source
= s
; /* Node moved. Memorize */
1108 /* dest not in remaining from nodes? */
1109 if (!node_isset(dest
, tmp
))
1112 if (source
== NUMA_NO_NODE
)
1115 node_clear(source
, tmp
);
1116 err
= migrate_to_node(mm
, source
, dest
, flags
);
1122 up_read(&mm
->mmap_sem
);
1130 * Allocate a new page for page migration based on vma policy.
1131 * Start by assuming the page is mapped by the same vma as contains @start.
1132 * Search forward from there, if not. N.B., this assumes that the
1133 * list of pages handed to migrate_pages()--which is how we get here--
1134 * is in virtual address order.
1136 static struct page
*new_page(struct page
*page
, unsigned long start
, int **x
)
1138 struct vm_area_struct
*vma
;
1139 unsigned long uninitialized_var(address
);
1141 vma
= find_vma(current
->mm
, start
);
1143 address
= page_address_in_vma(page
, vma
);
1144 if (address
!= -EFAULT
)
1149 if (PageHuge(page
)) {
1151 return alloc_huge_page_noerr(vma
, address
, 1);
1152 } else if (thp_migration_supported() && PageTransHuge(page
)) {
1155 thp
= alloc_hugepage_vma(GFP_TRANSHUGE
, vma
, address
,
1159 prep_transhuge_page(thp
);
1163 * if !vma, alloc_page_vma() will use task or system default policy
1165 return alloc_page_vma(GFP_HIGHUSER_MOVABLE
| __GFP_RETRY_MAYFAIL
,
1170 static void migrate_page_add(struct page
*page
, struct list_head
*pagelist
,
1171 unsigned long flags
)
1175 int do_migrate_pages(struct mm_struct
*mm
, const nodemask_t
*from
,
1176 const nodemask_t
*to
, int flags
)
1181 static struct page
*new_page(struct page
*page
, unsigned long start
, int **x
)
1187 static long do_mbind(unsigned long start
, unsigned long len
,
1188 unsigned short mode
, unsigned short mode_flags
,
1189 nodemask_t
*nmask
, unsigned long flags
)
1191 struct mm_struct
*mm
= current
->mm
;
1192 struct mempolicy
*new;
1195 LIST_HEAD(pagelist
);
1197 if (flags
& ~(unsigned long)MPOL_MF_VALID
)
1199 if ((flags
& MPOL_MF_MOVE_ALL
) && !capable(CAP_SYS_NICE
))
1202 if (start
& ~PAGE_MASK
)
1205 if (mode
== MPOL_DEFAULT
)
1206 flags
&= ~MPOL_MF_STRICT
;
1208 len
= (len
+ PAGE_SIZE
- 1) & PAGE_MASK
;
1216 new = mpol_new(mode
, mode_flags
, nmask
);
1218 return PTR_ERR(new);
1220 if (flags
& MPOL_MF_LAZY
)
1221 new->flags
|= MPOL_F_MOF
;
1224 * If we are using the default policy then operation
1225 * on discontinuous address spaces is okay after all
1228 flags
|= MPOL_MF_DISCONTIG_OK
;
1230 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1231 start
, start
+ len
, mode
, mode_flags
,
1232 nmask
? nodes_addr(*nmask
)[0] : NUMA_NO_NODE
);
1234 if (flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
)) {
1236 err
= migrate_prep();
1241 NODEMASK_SCRATCH(scratch
);
1243 down_write(&mm
->mmap_sem
);
1245 err
= mpol_set_nodemask(new, nmask
, scratch
);
1246 task_unlock(current
);
1248 up_write(&mm
->mmap_sem
);
1251 NODEMASK_SCRATCH_FREE(scratch
);
1256 err
= queue_pages_range(mm
, start
, end
, nmask
,
1257 flags
| MPOL_MF_INVERT
, &pagelist
);
1259 err
= mbind_range(mm
, start
, end
, new);
1264 if (!list_empty(&pagelist
)) {
1265 WARN_ON_ONCE(flags
& MPOL_MF_LAZY
);
1266 nr_failed
= migrate_pages(&pagelist
, new_page
, NULL
,
1267 start
, MIGRATE_SYNC
, MR_MEMPOLICY_MBIND
);
1269 putback_movable_pages(&pagelist
);
1272 if (nr_failed
&& (flags
& MPOL_MF_STRICT
))
1275 putback_movable_pages(&pagelist
);
1277 up_write(&mm
->mmap_sem
);
1284 * User space interface with variable sized bitmaps for nodelists.
1287 /* Copy a node mask from user space. */
1288 static int get_nodes(nodemask_t
*nodes
, const unsigned long __user
*nmask
,
1289 unsigned long maxnode
)
1293 unsigned long nlongs
;
1294 unsigned long endmask
;
1297 nodes_clear(*nodes
);
1298 if (maxnode
== 0 || !nmask
)
1300 if (maxnode
> PAGE_SIZE
*BITS_PER_BYTE
)
1303 nlongs
= BITS_TO_LONGS(maxnode
);
1304 if ((maxnode
% BITS_PER_LONG
) == 0)
1307 endmask
= (1UL << (maxnode
% BITS_PER_LONG
)) - 1;
1310 * When the user specified more nodes than supported just check
1311 * if the non supported part is all zero.
1313 * If maxnode have more longs than MAX_NUMNODES, check
1314 * the bits in that area first. And then go through to
1315 * check the rest bits which equal or bigger than MAX_NUMNODES.
1316 * Otherwise, just check bits [MAX_NUMNODES, maxnode).
1318 if (nlongs
> BITS_TO_LONGS(MAX_NUMNODES
)) {
1319 if (nlongs
> PAGE_SIZE
/sizeof(long))
1321 for (k
= BITS_TO_LONGS(MAX_NUMNODES
); k
< nlongs
; k
++) {
1322 if (get_user(t
, nmask
+ k
))
1324 if (k
== nlongs
- 1) {
1330 nlongs
= BITS_TO_LONGS(MAX_NUMNODES
);
1334 if (maxnode
> MAX_NUMNODES
&& MAX_NUMNODES
% BITS_PER_LONG
!= 0) {
1335 unsigned long valid_mask
= endmask
;
1337 valid_mask
&= ~((1UL << (MAX_NUMNODES
% BITS_PER_LONG
)) - 1);
1338 if (get_user(t
, nmask
+ nlongs
- 1))
1344 if (copy_from_user(nodes_addr(*nodes
), nmask
, nlongs
*sizeof(unsigned long)))
1346 nodes_addr(*nodes
)[nlongs
-1] &= endmask
;
1350 /* Copy a kernel node mask to user space */
1351 static int copy_nodes_to_user(unsigned long __user
*mask
, unsigned long maxnode
,
1354 unsigned long copy
= ALIGN(maxnode
-1, 64) / 8;
1355 unsigned int nbytes
= BITS_TO_LONGS(nr_node_ids
) * sizeof(long);
1357 if (copy
> nbytes
) {
1358 if (copy
> PAGE_SIZE
)
1360 if (clear_user((char __user
*)mask
+ nbytes
, copy
- nbytes
))
1364 return copy_to_user(mask
, nodes_addr(*nodes
), copy
) ? -EFAULT
: 0;
1367 SYSCALL_DEFINE6(mbind
, unsigned long, start
, unsigned long, len
,
1368 unsigned long, mode
, const unsigned long __user
*, nmask
,
1369 unsigned long, maxnode
, unsigned, flags
)
1373 unsigned short mode_flags
;
1375 mode_flags
= mode
& MPOL_MODE_FLAGS
;
1376 mode
&= ~MPOL_MODE_FLAGS
;
1377 if (mode
>= MPOL_MAX
)
1379 if ((mode_flags
& MPOL_F_STATIC_NODES
) &&
1380 (mode_flags
& MPOL_F_RELATIVE_NODES
))
1382 err
= get_nodes(&nodes
, nmask
, maxnode
);
1385 return do_mbind(start
, len
, mode
, mode_flags
, &nodes
, flags
);
1388 /* Set the process memory policy */
1389 SYSCALL_DEFINE3(set_mempolicy
, int, mode
, const unsigned long __user
*, nmask
,
1390 unsigned long, maxnode
)
1394 unsigned short flags
;
1396 flags
= mode
& MPOL_MODE_FLAGS
;
1397 mode
&= ~MPOL_MODE_FLAGS
;
1398 if ((unsigned int)mode
>= MPOL_MAX
)
1400 if ((flags
& MPOL_F_STATIC_NODES
) && (flags
& MPOL_F_RELATIVE_NODES
))
1402 err
= get_nodes(&nodes
, nmask
, maxnode
);
1405 return do_set_mempolicy(mode
, flags
, &nodes
);
1408 SYSCALL_DEFINE4(migrate_pages
, pid_t
, pid
, unsigned long, maxnode
,
1409 const unsigned long __user
*, old_nodes
,
1410 const unsigned long __user
*, new_nodes
)
1412 const struct cred
*cred
= current_cred(), *tcred
;
1413 struct mm_struct
*mm
= NULL
;
1414 struct task_struct
*task
;
1415 nodemask_t task_nodes
;
1419 NODEMASK_SCRATCH(scratch
);
1424 old
= &scratch
->mask1
;
1425 new = &scratch
->mask2
;
1427 err
= get_nodes(old
, old_nodes
, maxnode
);
1431 err
= get_nodes(new, new_nodes
, maxnode
);
1435 /* Find the mm_struct */
1437 task
= pid
? find_task_by_vpid(pid
) : current
;
1443 get_task_struct(task
);
1448 * Check if this process has the right to modify the specified
1449 * process. The right exists if the process has administrative
1450 * capabilities, superuser privileges or the same
1451 * userid as the target process.
1453 tcred
= __task_cred(task
);
1454 if (!uid_eq(cred
->euid
, tcred
->suid
) && !uid_eq(cred
->euid
, tcred
->uid
) &&
1455 !uid_eq(cred
->uid
, tcred
->suid
) && !uid_eq(cred
->uid
, tcred
->uid
) &&
1456 !capable(CAP_SYS_NICE
)) {
1463 task_nodes
= cpuset_mems_allowed(task
);
1464 /* Is the user allowed to access the target nodes? */
1465 if (!nodes_subset(*new, task_nodes
) && !capable(CAP_SYS_NICE
)) {
1470 task_nodes
= cpuset_mems_allowed(current
);
1471 nodes_and(*new, *new, task_nodes
);
1472 if (nodes_empty(*new))
1475 nodes_and(*new, *new, node_states
[N_MEMORY
]);
1476 if (nodes_empty(*new))
1479 err
= security_task_movememory(task
);
1483 mm
= get_task_mm(task
);
1484 put_task_struct(task
);
1491 err
= do_migrate_pages(mm
, old
, new,
1492 capable(CAP_SYS_NICE
) ? MPOL_MF_MOVE_ALL
: MPOL_MF_MOVE
);
1496 NODEMASK_SCRATCH_FREE(scratch
);
1501 put_task_struct(task
);
1507 /* Retrieve NUMA policy */
1508 SYSCALL_DEFINE5(get_mempolicy
, int __user
*, policy
,
1509 unsigned long __user
*, nmask
, unsigned long, maxnode
,
1510 unsigned long, addr
, unsigned long, flags
)
1513 int uninitialized_var(pval
);
1516 if (nmask
!= NULL
&& maxnode
< nr_node_ids
)
1519 err
= do_get_mempolicy(&pval
, &nodes
, addr
, flags
);
1524 if (policy
&& put_user(pval
, policy
))
1528 err
= copy_nodes_to_user(nmask
, maxnode
, &nodes
);
1533 #ifdef CONFIG_COMPAT
1535 COMPAT_SYSCALL_DEFINE5(get_mempolicy
, int __user
*, policy
,
1536 compat_ulong_t __user
*, nmask
,
1537 compat_ulong_t
, maxnode
,
1538 compat_ulong_t
, addr
, compat_ulong_t
, flags
)
1541 unsigned long __user
*nm
= NULL
;
1542 unsigned long nr_bits
, alloc_size
;
1543 DECLARE_BITMAP(bm
, MAX_NUMNODES
);
1545 nr_bits
= min_t(unsigned long, maxnode
-1, nr_node_ids
);
1546 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
1549 nm
= compat_alloc_user_space(alloc_size
);
1551 err
= sys_get_mempolicy(policy
, nm
, nr_bits
+1, addr
, flags
);
1553 if (!err
&& nmask
) {
1554 unsigned long copy_size
;
1555 copy_size
= min_t(unsigned long, sizeof(bm
), alloc_size
);
1556 err
= copy_from_user(bm
, nm
, copy_size
);
1557 /* ensure entire bitmap is zeroed */
1558 err
|= clear_user(nmask
, ALIGN(maxnode
-1, 8) / 8);
1559 err
|= compat_put_bitmap(nmask
, bm
, nr_bits
);
1565 COMPAT_SYSCALL_DEFINE3(set_mempolicy
, int, mode
, compat_ulong_t __user
*, nmask
,
1566 compat_ulong_t
, maxnode
)
1568 unsigned long __user
*nm
= NULL
;
1569 unsigned long nr_bits
, alloc_size
;
1570 DECLARE_BITMAP(bm
, MAX_NUMNODES
);
1572 nr_bits
= min_t(unsigned long, maxnode
-1, MAX_NUMNODES
);
1573 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
1576 if (compat_get_bitmap(bm
, nmask
, nr_bits
))
1578 nm
= compat_alloc_user_space(alloc_size
);
1579 if (copy_to_user(nm
, bm
, alloc_size
))
1583 return sys_set_mempolicy(mode
, nm
, nr_bits
+1);
1586 COMPAT_SYSCALL_DEFINE6(mbind
, compat_ulong_t
, start
, compat_ulong_t
, len
,
1587 compat_ulong_t
, mode
, compat_ulong_t __user
*, nmask
,
1588 compat_ulong_t
, maxnode
, compat_ulong_t
, flags
)
1590 unsigned long __user
*nm
= NULL
;
1591 unsigned long nr_bits
, alloc_size
;
1594 nr_bits
= min_t(unsigned long, maxnode
-1, MAX_NUMNODES
);
1595 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
1598 if (compat_get_bitmap(nodes_addr(bm
), nmask
, nr_bits
))
1600 nm
= compat_alloc_user_space(alloc_size
);
1601 if (copy_to_user(nm
, nodes_addr(bm
), alloc_size
))
1605 return sys_mbind(start
, len
, mode
, nm
, nr_bits
+1, flags
);
1610 struct mempolicy
*__get_vma_policy(struct vm_area_struct
*vma
,
1613 struct mempolicy
*pol
= NULL
;
1616 if (vma
->vm_ops
&& vma
->vm_ops
->get_policy
) {
1617 pol
= vma
->vm_ops
->get_policy(vma
, addr
);
1618 } else if (vma
->vm_policy
) {
1619 pol
= vma
->vm_policy
;
1622 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1623 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1624 * count on these policies which will be dropped by
1625 * mpol_cond_put() later
1627 if (mpol_needs_cond_ref(pol
))
1636 * get_vma_policy(@vma, @addr)
1637 * @vma: virtual memory area whose policy is sought
1638 * @addr: address in @vma for shared policy lookup
1640 * Returns effective policy for a VMA at specified address.
1641 * Falls back to current->mempolicy or system default policy, as necessary.
1642 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1643 * count--added by the get_policy() vm_op, as appropriate--to protect against
1644 * freeing by another task. It is the caller's responsibility to free the
1645 * extra reference for shared policies.
1647 static struct mempolicy
*get_vma_policy(struct vm_area_struct
*vma
,
1650 struct mempolicy
*pol
= __get_vma_policy(vma
, addr
);
1653 pol
= get_task_policy(current
);
1658 bool vma_policy_mof(struct vm_area_struct
*vma
)
1660 struct mempolicy
*pol
;
1662 if (vma
->vm_ops
&& vma
->vm_ops
->get_policy
) {
1665 pol
= vma
->vm_ops
->get_policy(vma
, vma
->vm_start
);
1666 if (pol
&& (pol
->flags
& MPOL_F_MOF
))
1673 pol
= vma
->vm_policy
;
1675 pol
= get_task_policy(current
);
1677 return pol
->flags
& MPOL_F_MOF
;
1680 static int apply_policy_zone(struct mempolicy
*policy
, enum zone_type zone
)
1682 enum zone_type dynamic_policy_zone
= policy_zone
;
1684 BUG_ON(dynamic_policy_zone
== ZONE_MOVABLE
);
1687 * if policy->v.nodes has movable memory only,
1688 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1690 * policy->v.nodes is intersect with node_states[N_MEMORY].
1691 * so if the following test faile, it implies
1692 * policy->v.nodes has movable memory only.
1694 if (!nodes_intersects(policy
->v
.nodes
, node_states
[N_HIGH_MEMORY
]))
1695 dynamic_policy_zone
= ZONE_MOVABLE
;
1697 return zone
>= dynamic_policy_zone
;
1701 * Return a nodemask representing a mempolicy for filtering nodes for
1704 static nodemask_t
*policy_nodemask(gfp_t gfp
, struct mempolicy
*policy
)
1706 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1707 if (unlikely(policy
->mode
== MPOL_BIND
) &&
1708 apply_policy_zone(policy
, gfp_zone(gfp
)) &&
1709 cpuset_nodemask_valid_mems_allowed(&policy
->v
.nodes
))
1710 return &policy
->v
.nodes
;
1715 /* Return the node id preferred by the given mempolicy, or the given id */
1716 static int policy_node(gfp_t gfp
, struct mempolicy
*policy
,
1719 if (policy
->mode
== MPOL_PREFERRED
&& !(policy
->flags
& MPOL_F_LOCAL
))
1720 nd
= policy
->v
.preferred_node
;
1723 * __GFP_THISNODE shouldn't even be used with the bind policy
1724 * because we might easily break the expectation to stay on the
1725 * requested node and not break the policy.
1727 WARN_ON_ONCE(policy
->mode
== MPOL_BIND
&& (gfp
& __GFP_THISNODE
));
1733 /* Do dynamic interleaving for a process */
1734 static unsigned interleave_nodes(struct mempolicy
*policy
)
1737 struct task_struct
*me
= current
;
1739 next
= next_node_in(me
->il_prev
, policy
->v
.nodes
);
1740 if (next
< MAX_NUMNODES
)
1746 * Depending on the memory policy provide a node from which to allocate the
1749 unsigned int mempolicy_slab_node(void)
1751 struct mempolicy
*policy
;
1752 int node
= numa_mem_id();
1757 policy
= current
->mempolicy
;
1758 if (!policy
|| policy
->flags
& MPOL_F_LOCAL
)
1761 switch (policy
->mode
) {
1762 case MPOL_PREFERRED
:
1764 * handled MPOL_F_LOCAL above
1766 return policy
->v
.preferred_node
;
1768 case MPOL_INTERLEAVE
:
1769 return interleave_nodes(policy
);
1775 * Follow bind policy behavior and start allocation at the
1778 struct zonelist
*zonelist
;
1779 enum zone_type highest_zoneidx
= gfp_zone(GFP_KERNEL
);
1780 zonelist
= &NODE_DATA(node
)->node_zonelists
[ZONELIST_FALLBACK
];
1781 z
= first_zones_zonelist(zonelist
, highest_zoneidx
,
1783 return z
->zone
? z
->zone
->node
: node
;
1792 * Do static interleaving for a VMA with known offset @n. Returns the n'th
1793 * node in pol->v.nodes (starting from n=0), wrapping around if n exceeds the
1794 * number of present nodes.
1796 static unsigned offset_il_node(struct mempolicy
*pol
, unsigned long n
)
1798 unsigned nnodes
= nodes_weight(pol
->v
.nodes
);
1804 return numa_node_id();
1805 target
= (unsigned int)n
% nnodes
;
1806 nid
= first_node(pol
->v
.nodes
);
1807 for (i
= 0; i
< target
; i
++)
1808 nid
= next_node(nid
, pol
->v
.nodes
);
1812 /* Determine a node number for interleave */
1813 static inline unsigned interleave_nid(struct mempolicy
*pol
,
1814 struct vm_area_struct
*vma
, unsigned long addr
, int shift
)
1820 * for small pages, there is no difference between
1821 * shift and PAGE_SHIFT, so the bit-shift is safe.
1822 * for huge pages, since vm_pgoff is in units of small
1823 * pages, we need to shift off the always 0 bits to get
1826 BUG_ON(shift
< PAGE_SHIFT
);
1827 off
= vma
->vm_pgoff
>> (shift
- PAGE_SHIFT
);
1828 off
+= (addr
- vma
->vm_start
) >> shift
;
1829 return offset_il_node(pol
, off
);
1831 return interleave_nodes(pol
);
1834 #ifdef CONFIG_HUGETLBFS
1836 * huge_node(@vma, @addr, @gfp_flags, @mpol)
1837 * @vma: virtual memory area whose policy is sought
1838 * @addr: address in @vma for shared policy lookup and interleave policy
1839 * @gfp_flags: for requested zone
1840 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
1841 * @nodemask: pointer to nodemask pointer for MPOL_BIND nodemask
1843 * Returns a nid suitable for a huge page allocation and a pointer
1844 * to the struct mempolicy for conditional unref after allocation.
1845 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1846 * @nodemask for filtering the zonelist.
1848 * Must be protected by read_mems_allowed_begin()
1850 int huge_node(struct vm_area_struct
*vma
, unsigned long addr
, gfp_t gfp_flags
,
1851 struct mempolicy
**mpol
, nodemask_t
**nodemask
)
1855 *mpol
= get_vma_policy(vma
, addr
);
1856 *nodemask
= NULL
; /* assume !MPOL_BIND */
1858 if (unlikely((*mpol
)->mode
== MPOL_INTERLEAVE
)) {
1859 nid
= interleave_nid(*mpol
, vma
, addr
,
1860 huge_page_shift(hstate_vma(vma
)));
1862 nid
= policy_node(gfp_flags
, *mpol
, numa_node_id());
1863 if ((*mpol
)->mode
== MPOL_BIND
)
1864 *nodemask
= &(*mpol
)->v
.nodes
;
1870 * init_nodemask_of_mempolicy
1872 * If the current task's mempolicy is "default" [NULL], return 'false'
1873 * to indicate default policy. Otherwise, extract the policy nodemask
1874 * for 'bind' or 'interleave' policy into the argument nodemask, or
1875 * initialize the argument nodemask to contain the single node for
1876 * 'preferred' or 'local' policy and return 'true' to indicate presence
1877 * of non-default mempolicy.
1879 * We don't bother with reference counting the mempolicy [mpol_get/put]
1880 * because the current task is examining it's own mempolicy and a task's
1881 * mempolicy is only ever changed by the task itself.
1883 * N.B., it is the caller's responsibility to free a returned nodemask.
1885 bool init_nodemask_of_mempolicy(nodemask_t
*mask
)
1887 struct mempolicy
*mempolicy
;
1890 if (!(mask
&& current
->mempolicy
))
1894 mempolicy
= current
->mempolicy
;
1895 switch (mempolicy
->mode
) {
1896 case MPOL_PREFERRED
:
1897 if (mempolicy
->flags
& MPOL_F_LOCAL
)
1898 nid
= numa_node_id();
1900 nid
= mempolicy
->v
.preferred_node
;
1901 init_nodemask_of_node(mask
, nid
);
1906 case MPOL_INTERLEAVE
:
1907 *mask
= mempolicy
->v
.nodes
;
1913 task_unlock(current
);
1920 * mempolicy_nodemask_intersects
1922 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1923 * policy. Otherwise, check for intersection between mask and the policy
1924 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1925 * policy, always return true since it may allocate elsewhere on fallback.
1927 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1929 bool mempolicy_nodemask_intersects(struct task_struct
*tsk
,
1930 const nodemask_t
*mask
)
1932 struct mempolicy
*mempolicy
;
1938 mempolicy
= tsk
->mempolicy
;
1942 switch (mempolicy
->mode
) {
1943 case MPOL_PREFERRED
:
1945 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1946 * allocate from, they may fallback to other nodes when oom.
1947 * Thus, it's possible for tsk to have allocated memory from
1952 case MPOL_INTERLEAVE
:
1953 ret
= nodes_intersects(mempolicy
->v
.nodes
, *mask
);
1963 /* Allocate a page in interleaved policy.
1964 Own path because it needs to do special accounting. */
1965 static struct page
*alloc_page_interleave(gfp_t gfp
, unsigned order
,
1970 page
= __alloc_pages(gfp
, order
, nid
);
1971 if (page
&& page_to_nid(page
) == nid
) {
1973 __inc_numa_state(page_zone(page
), NUMA_INTERLEAVE_HIT
);
1980 * alloc_pages_vma - Allocate a page for a VMA.
1983 * %GFP_USER user allocation.
1984 * %GFP_KERNEL kernel allocations,
1985 * %GFP_HIGHMEM highmem/user allocations,
1986 * %GFP_FS allocation should not call back into a file system.
1987 * %GFP_ATOMIC don't sleep.
1989 * @order:Order of the GFP allocation.
1990 * @vma: Pointer to VMA or NULL if not available.
1991 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1992 * @node: Which node to prefer for allocation (modulo policy).
1993 * @hugepage: for hugepages try only the preferred node if possible
1995 * This function allocates a page from the kernel page pool and applies
1996 * a NUMA policy associated with the VMA or the current process.
1997 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1998 * mm_struct of the VMA to prevent it from going away. Should be used for
1999 * all allocations for pages that will be mapped into user space. Returns
2000 * NULL when no page can be allocated.
2003 alloc_pages_vma(gfp_t gfp
, int order
, struct vm_area_struct
*vma
,
2004 unsigned long addr
, int node
, bool hugepage
)
2006 struct mempolicy
*pol
;
2011 pol
= get_vma_policy(vma
, addr
);
2013 if (pol
->mode
== MPOL_INTERLEAVE
) {
2016 nid
= interleave_nid(pol
, vma
, addr
, PAGE_SHIFT
+ order
);
2018 page
= alloc_page_interleave(gfp
, order
, nid
);
2022 if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE
) && hugepage
)) {
2023 int hpage_node
= node
;
2026 * For hugepage allocation and non-interleave policy which
2027 * allows the current node (or other explicitly preferred
2028 * node) we only try to allocate from the current/preferred
2029 * node and don't fall back to other nodes, as the cost of
2030 * remote accesses would likely offset THP benefits.
2032 * If the policy is interleave, or does not allow the current
2033 * node in its nodemask, we allocate the standard way.
2035 if (pol
->mode
== MPOL_PREFERRED
&&
2036 !(pol
->flags
& MPOL_F_LOCAL
))
2037 hpage_node
= pol
->v
.preferred_node
;
2039 nmask
= policy_nodemask(gfp
, pol
);
2040 if (!nmask
|| node_isset(hpage_node
, *nmask
)) {
2043 * We cannot invoke reclaim if __GFP_THISNODE
2044 * is set. Invoking reclaim with
2045 * __GFP_THISNODE set, would cause THP
2046 * allocations to trigger heavy swapping
2047 * despite there may be tons of free memory
2048 * (including potentially plenty of THP
2049 * already available in the buddy) on all the
2052 * At most we could invoke compaction when
2053 * __GFP_THISNODE is set (but we would need to
2054 * refrain from invoking reclaim even if
2055 * compaction returned COMPACT_SKIPPED because
2056 * there wasn't not enough memory to succeed
2057 * compaction). For now just avoid
2058 * __GFP_THISNODE instead of limiting the
2059 * allocation path to a strict and single
2060 * compaction invocation.
2062 * Supposedly if direct reclaim was enabled by
2063 * the caller, the app prefers THP regardless
2064 * of the node it comes from so this would be
2065 * more desiderable behavior than only
2066 * providing THP originated from the local
2067 * node in such case.
2069 if (!(gfp
& __GFP_DIRECT_RECLAIM
))
2070 gfp
|= __GFP_THISNODE
;
2071 page
= __alloc_pages_node(hpage_node
, gfp
, order
);
2076 nmask
= policy_nodemask(gfp
, pol
);
2077 preferred_nid
= policy_node(gfp
, pol
, node
);
2078 page
= __alloc_pages_nodemask(gfp
, order
, preferred_nid
, nmask
);
2085 * alloc_pages_current - Allocate pages.
2088 * %GFP_USER user allocation,
2089 * %GFP_KERNEL kernel allocation,
2090 * %GFP_HIGHMEM highmem allocation,
2091 * %GFP_FS don't call back into a file system.
2092 * %GFP_ATOMIC don't sleep.
2093 * @order: Power of two of allocation size in pages. 0 is a single page.
2095 * Allocate a page from the kernel page pool. When not in
2096 * interrupt context and apply the current process NUMA policy.
2097 * Returns NULL when no page can be allocated.
2099 struct page
*alloc_pages_current(gfp_t gfp
, unsigned order
)
2101 struct mempolicy
*pol
= &default_policy
;
2104 if (!in_interrupt() && !(gfp
& __GFP_THISNODE
))
2105 pol
= get_task_policy(current
);
2108 * No reference counting needed for current->mempolicy
2109 * nor system default_policy
2111 if (pol
->mode
== MPOL_INTERLEAVE
)
2112 page
= alloc_page_interleave(gfp
, order
, interleave_nodes(pol
));
2114 page
= __alloc_pages_nodemask(gfp
, order
,
2115 policy_node(gfp
, pol
, numa_node_id()),
2116 policy_nodemask(gfp
, pol
));
2120 EXPORT_SYMBOL(alloc_pages_current
);
2122 int vma_dup_policy(struct vm_area_struct
*src
, struct vm_area_struct
*dst
)
2124 struct mempolicy
*pol
= mpol_dup(vma_policy(src
));
2127 return PTR_ERR(pol
);
2128 dst
->vm_policy
= pol
;
2133 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2134 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2135 * with the mems_allowed returned by cpuset_mems_allowed(). This
2136 * keeps mempolicies cpuset relative after its cpuset moves. See
2137 * further kernel/cpuset.c update_nodemask().
2139 * current's mempolicy may be rebinded by the other task(the task that changes
2140 * cpuset's mems), so we needn't do rebind work for current task.
2143 /* Slow path of a mempolicy duplicate */
2144 struct mempolicy
*__mpol_dup(struct mempolicy
*old
)
2146 struct mempolicy
*new = kmem_cache_alloc(policy_cache
, GFP_KERNEL
);
2149 return ERR_PTR(-ENOMEM
);
2151 /* task's mempolicy is protected by alloc_lock */
2152 if (old
== current
->mempolicy
) {
2155 task_unlock(current
);
2159 if (current_cpuset_is_being_rebound()) {
2160 nodemask_t mems
= cpuset_mems_allowed(current
);
2161 mpol_rebind_policy(new, &mems
);
2163 atomic_set(&new->refcnt
, 1);
2167 /* Slow path of a mempolicy comparison */
2168 bool __mpol_equal(struct mempolicy
*a
, struct mempolicy
*b
)
2172 if (a
->mode
!= b
->mode
)
2174 if (a
->flags
!= b
->flags
)
2176 if (mpol_store_user_nodemask(a
))
2177 if (!nodes_equal(a
->w
.user_nodemask
, b
->w
.user_nodemask
))
2183 case MPOL_INTERLEAVE
:
2184 return !!nodes_equal(a
->v
.nodes
, b
->v
.nodes
);
2185 case MPOL_PREFERRED
:
2186 /* a's ->flags is the same as b's */
2187 if (a
->flags
& MPOL_F_LOCAL
)
2189 return a
->v
.preferred_node
== b
->v
.preferred_node
;
2197 * Shared memory backing store policy support.
2199 * Remember policies even when nobody has shared memory mapped.
2200 * The policies are kept in Red-Black tree linked from the inode.
2201 * They are protected by the sp->lock rwlock, which should be held
2202 * for any accesses to the tree.
2206 * lookup first element intersecting start-end. Caller holds sp->lock for
2207 * reading or for writing
2209 static struct sp_node
*
2210 sp_lookup(struct shared_policy
*sp
, unsigned long start
, unsigned long end
)
2212 struct rb_node
*n
= sp
->root
.rb_node
;
2215 struct sp_node
*p
= rb_entry(n
, struct sp_node
, nd
);
2217 if (start
>= p
->end
)
2219 else if (end
<= p
->start
)
2227 struct sp_node
*w
= NULL
;
2228 struct rb_node
*prev
= rb_prev(n
);
2231 w
= rb_entry(prev
, struct sp_node
, nd
);
2232 if (w
->end
<= start
)
2236 return rb_entry(n
, struct sp_node
, nd
);
2240 * Insert a new shared policy into the list. Caller holds sp->lock for
2243 static void sp_insert(struct shared_policy
*sp
, struct sp_node
*new)
2245 struct rb_node
**p
= &sp
->root
.rb_node
;
2246 struct rb_node
*parent
= NULL
;
2251 nd
= rb_entry(parent
, struct sp_node
, nd
);
2252 if (new->start
< nd
->start
)
2254 else if (new->end
> nd
->end
)
2255 p
= &(*p
)->rb_right
;
2259 rb_link_node(&new->nd
, parent
, p
);
2260 rb_insert_color(&new->nd
, &sp
->root
);
2261 pr_debug("inserting %lx-%lx: %d\n", new->start
, new->end
,
2262 new->policy
? new->policy
->mode
: 0);
2265 /* Find shared policy intersecting idx */
2267 mpol_shared_policy_lookup(struct shared_policy
*sp
, unsigned long idx
)
2269 struct mempolicy
*pol
= NULL
;
2272 if (!sp
->root
.rb_node
)
2274 read_lock(&sp
->lock
);
2275 sn
= sp_lookup(sp
, idx
, idx
+1);
2277 mpol_get(sn
->policy
);
2280 read_unlock(&sp
->lock
);
2284 static void sp_free(struct sp_node
*n
)
2286 mpol_put(n
->policy
);
2287 kmem_cache_free(sn_cache
, n
);
2291 * mpol_misplaced - check whether current page node is valid in policy
2293 * @page: page to be checked
2294 * @vma: vm area where page mapped
2295 * @addr: virtual address where page mapped
2297 * Lookup current policy node id for vma,addr and "compare to" page's
2301 * -1 - not misplaced, page is in the right node
2302 * node - node id where the page should be
2304 * Policy determination "mimics" alloc_page_vma().
2305 * Called from fault path where we know the vma and faulting address.
2307 int mpol_misplaced(struct page
*page
, struct vm_area_struct
*vma
, unsigned long addr
)
2309 struct mempolicy
*pol
;
2311 int curnid
= page_to_nid(page
);
2312 unsigned long pgoff
;
2313 int thiscpu
= raw_smp_processor_id();
2314 int thisnid
= cpu_to_node(thiscpu
);
2318 pol
= get_vma_policy(vma
, addr
);
2319 if (!(pol
->flags
& MPOL_F_MOF
))
2322 switch (pol
->mode
) {
2323 case MPOL_INTERLEAVE
:
2324 pgoff
= vma
->vm_pgoff
;
2325 pgoff
+= (addr
- vma
->vm_start
) >> PAGE_SHIFT
;
2326 polnid
= offset_il_node(pol
, pgoff
);
2329 case MPOL_PREFERRED
:
2330 if (pol
->flags
& MPOL_F_LOCAL
)
2331 polnid
= numa_node_id();
2333 polnid
= pol
->v
.preferred_node
;
2339 * allows binding to multiple nodes.
2340 * use current page if in policy nodemask,
2341 * else select nearest allowed node, if any.
2342 * If no allowed nodes, use current [!misplaced].
2344 if (node_isset(curnid
, pol
->v
.nodes
))
2346 z
= first_zones_zonelist(
2347 node_zonelist(numa_node_id(), GFP_HIGHUSER
),
2348 gfp_zone(GFP_HIGHUSER
),
2350 polnid
= z
->zone
->node
;
2357 /* Migrate the page towards the node whose CPU is referencing it */
2358 if (pol
->flags
& MPOL_F_MORON
) {
2361 if (!should_numa_migrate_memory(current
, page
, curnid
, thiscpu
))
2365 if (curnid
!= polnid
)
2374 * Drop the (possibly final) reference to task->mempolicy. It needs to be
2375 * dropped after task->mempolicy is set to NULL so that any allocation done as
2376 * part of its kmem_cache_free(), such as by KASAN, doesn't reference a freed
2379 void mpol_put_task_policy(struct task_struct
*task
)
2381 struct mempolicy
*pol
;
2384 pol
= task
->mempolicy
;
2385 task
->mempolicy
= NULL
;
2390 static void sp_delete(struct shared_policy
*sp
, struct sp_node
*n
)
2392 pr_debug("deleting %lx-l%lx\n", n
->start
, n
->end
);
2393 rb_erase(&n
->nd
, &sp
->root
);
2397 static void sp_node_init(struct sp_node
*node
, unsigned long start
,
2398 unsigned long end
, struct mempolicy
*pol
)
2400 node
->start
= start
;
2405 static struct sp_node
*sp_alloc(unsigned long start
, unsigned long end
,
2406 struct mempolicy
*pol
)
2409 struct mempolicy
*newpol
;
2411 n
= kmem_cache_alloc(sn_cache
, GFP_KERNEL
);
2415 newpol
= mpol_dup(pol
);
2416 if (IS_ERR(newpol
)) {
2417 kmem_cache_free(sn_cache
, n
);
2420 newpol
->flags
|= MPOL_F_SHARED
;
2421 sp_node_init(n
, start
, end
, newpol
);
2426 /* Replace a policy range. */
2427 static int shared_policy_replace(struct shared_policy
*sp
, unsigned long start
,
2428 unsigned long end
, struct sp_node
*new)
2431 struct sp_node
*n_new
= NULL
;
2432 struct mempolicy
*mpol_new
= NULL
;
2436 write_lock(&sp
->lock
);
2437 n
= sp_lookup(sp
, start
, end
);
2438 /* Take care of old policies in the same range. */
2439 while (n
&& n
->start
< end
) {
2440 struct rb_node
*next
= rb_next(&n
->nd
);
2441 if (n
->start
>= start
) {
2447 /* Old policy spanning whole new range. */
2452 *mpol_new
= *n
->policy
;
2453 atomic_set(&mpol_new
->refcnt
, 1);
2454 sp_node_init(n_new
, end
, n
->end
, mpol_new
);
2456 sp_insert(sp
, n_new
);
2465 n
= rb_entry(next
, struct sp_node
, nd
);
2469 write_unlock(&sp
->lock
);
2476 kmem_cache_free(sn_cache
, n_new
);
2481 write_unlock(&sp
->lock
);
2483 n_new
= kmem_cache_alloc(sn_cache
, GFP_KERNEL
);
2486 mpol_new
= kmem_cache_alloc(policy_cache
, GFP_KERNEL
);
2493 * mpol_shared_policy_init - initialize shared policy for inode
2494 * @sp: pointer to inode shared policy
2495 * @mpol: struct mempolicy to install
2497 * Install non-NULL @mpol in inode's shared policy rb-tree.
2498 * On entry, the current task has a reference on a non-NULL @mpol.
2499 * This must be released on exit.
2500 * This is called at get_inode() calls and we can use GFP_KERNEL.
2502 void mpol_shared_policy_init(struct shared_policy
*sp
, struct mempolicy
*mpol
)
2506 sp
->root
= RB_ROOT
; /* empty tree == default mempolicy */
2507 rwlock_init(&sp
->lock
);
2510 struct vm_area_struct pvma
;
2511 struct mempolicy
*new;
2512 NODEMASK_SCRATCH(scratch
);
2516 /* contextualize the tmpfs mount point mempolicy */
2517 new = mpol_new(mpol
->mode
, mpol
->flags
, &mpol
->w
.user_nodemask
);
2519 goto free_scratch
; /* no valid nodemask intersection */
2522 ret
= mpol_set_nodemask(new, &mpol
->w
.user_nodemask
, scratch
);
2523 task_unlock(current
);
2527 /* Create pseudo-vma that contains just the policy */
2528 memset(&pvma
, 0, sizeof(struct vm_area_struct
));
2529 pvma
.vm_end
= TASK_SIZE
; /* policy covers entire file */
2530 mpol_set_shared_policy(sp
, &pvma
, new); /* adds ref */
2533 mpol_put(new); /* drop initial ref */
2535 NODEMASK_SCRATCH_FREE(scratch
);
2537 mpol_put(mpol
); /* drop our incoming ref on sb mpol */
2541 int mpol_set_shared_policy(struct shared_policy
*info
,
2542 struct vm_area_struct
*vma
, struct mempolicy
*npol
)
2545 struct sp_node
*new = NULL
;
2546 unsigned long sz
= vma_pages(vma
);
2548 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2550 sz
, npol
? npol
->mode
: -1,
2551 npol
? npol
->flags
: -1,
2552 npol
? nodes_addr(npol
->v
.nodes
)[0] : NUMA_NO_NODE
);
2555 new = sp_alloc(vma
->vm_pgoff
, vma
->vm_pgoff
+ sz
, npol
);
2559 err
= shared_policy_replace(info
, vma
->vm_pgoff
, vma
->vm_pgoff
+sz
, new);
2565 /* Free a backing policy store on inode delete. */
2566 void mpol_free_shared_policy(struct shared_policy
*p
)
2569 struct rb_node
*next
;
2571 if (!p
->root
.rb_node
)
2573 write_lock(&p
->lock
);
2574 next
= rb_first(&p
->root
);
2576 n
= rb_entry(next
, struct sp_node
, nd
);
2577 next
= rb_next(&n
->nd
);
2580 write_unlock(&p
->lock
);
2583 #ifdef CONFIG_NUMA_BALANCING
2584 static int __initdata numabalancing_override
;
2586 static void __init
check_numabalancing_enable(void)
2588 bool numabalancing_default
= false;
2590 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED
))
2591 numabalancing_default
= true;
2593 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2594 if (numabalancing_override
)
2595 set_numabalancing_state(numabalancing_override
== 1);
2597 if (num_online_nodes() > 1 && !numabalancing_override
) {
2598 pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
2599 numabalancing_default
? "Enabling" : "Disabling");
2600 set_numabalancing_state(numabalancing_default
);
2604 static int __init
setup_numabalancing(char *str
)
2610 if (!strcmp(str
, "enable")) {
2611 numabalancing_override
= 1;
2613 } else if (!strcmp(str
, "disable")) {
2614 numabalancing_override
= -1;
2619 pr_warn("Unable to parse numa_balancing=\n");
2623 __setup("numa_balancing=", setup_numabalancing
);
2625 static inline void __init
check_numabalancing_enable(void)
2628 #endif /* CONFIG_NUMA_BALANCING */
2630 /* assumes fs == KERNEL_DS */
2631 void __init
numa_policy_init(void)
2633 nodemask_t interleave_nodes
;
2634 unsigned long largest
= 0;
2635 int nid
, prefer
= 0;
2637 policy_cache
= kmem_cache_create("numa_policy",
2638 sizeof(struct mempolicy
),
2639 0, SLAB_PANIC
, NULL
);
2641 sn_cache
= kmem_cache_create("shared_policy_node",
2642 sizeof(struct sp_node
),
2643 0, SLAB_PANIC
, NULL
);
2645 for_each_node(nid
) {
2646 preferred_node_policy
[nid
] = (struct mempolicy
) {
2647 .refcnt
= ATOMIC_INIT(1),
2648 .mode
= MPOL_PREFERRED
,
2649 .flags
= MPOL_F_MOF
| MPOL_F_MORON
,
2650 .v
= { .preferred_node
= nid
, },
2655 * Set interleaving policy for system init. Interleaving is only
2656 * enabled across suitably sized nodes (default is >= 16MB), or
2657 * fall back to the largest node if they're all smaller.
2659 nodes_clear(interleave_nodes
);
2660 for_each_node_state(nid
, N_MEMORY
) {
2661 unsigned long total_pages
= node_present_pages(nid
);
2663 /* Preserve the largest node */
2664 if (largest
< total_pages
) {
2665 largest
= total_pages
;
2669 /* Interleave this node? */
2670 if ((total_pages
<< PAGE_SHIFT
) >= (16 << 20))
2671 node_set(nid
, interleave_nodes
);
2674 /* All too small, use the largest */
2675 if (unlikely(nodes_empty(interleave_nodes
)))
2676 node_set(prefer
, interleave_nodes
);
2678 if (do_set_mempolicy(MPOL_INTERLEAVE
, 0, &interleave_nodes
))
2679 pr_err("%s: interleaving failed\n", __func__
);
2681 check_numabalancing_enable();
2684 /* Reset policy of current process to default */
2685 void numa_default_policy(void)
2687 do_set_mempolicy(MPOL_DEFAULT
, 0, NULL
);
2691 * Parse and format mempolicy from/to strings
2695 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
2697 static const char * const policy_modes
[] =
2699 [MPOL_DEFAULT
] = "default",
2700 [MPOL_PREFERRED
] = "prefer",
2701 [MPOL_BIND
] = "bind",
2702 [MPOL_INTERLEAVE
] = "interleave",
2703 [MPOL_LOCAL
] = "local",
2709 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2710 * @str: string containing mempolicy to parse
2711 * @mpol: pointer to struct mempolicy pointer, returned on success.
2714 * <mode>[=<flags>][:<nodelist>]
2716 * On success, returns 0, else 1
2718 int mpol_parse_str(char *str
, struct mempolicy
**mpol
)
2720 struct mempolicy
*new = NULL
;
2721 unsigned short mode
;
2722 unsigned short mode_flags
;
2724 char *nodelist
= strchr(str
, ':');
2725 char *flags
= strchr(str
, '=');
2729 /* NUL-terminate mode or flags string */
2731 if (nodelist_parse(nodelist
, nodes
))
2733 if (!nodes_subset(nodes
, node_states
[N_MEMORY
]))
2739 *flags
++ = '\0'; /* terminate mode string */
2741 for (mode
= 0; mode
< MPOL_MAX
; mode
++) {
2742 if (!strcmp(str
, policy_modes
[mode
])) {
2746 if (mode
>= MPOL_MAX
)
2750 case MPOL_PREFERRED
:
2752 * Insist on a nodelist of one node only
2755 char *rest
= nodelist
;
2756 while (isdigit(*rest
))
2762 case MPOL_INTERLEAVE
:
2764 * Default to online nodes with memory if no nodelist
2767 nodes
= node_states
[N_MEMORY
];
2771 * Don't allow a nodelist; mpol_new() checks flags
2775 mode
= MPOL_PREFERRED
;
2779 * Insist on a empty nodelist
2786 * Insist on a nodelist
2795 * Currently, we only support two mutually exclusive
2798 if (!strcmp(flags
, "static"))
2799 mode_flags
|= MPOL_F_STATIC_NODES
;
2800 else if (!strcmp(flags
, "relative"))
2801 mode_flags
|= MPOL_F_RELATIVE_NODES
;
2806 new = mpol_new(mode
, mode_flags
, &nodes
);
2811 * Save nodes for mpol_to_str() to show the tmpfs mount options
2812 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2814 if (mode
!= MPOL_PREFERRED
)
2815 new->v
.nodes
= nodes
;
2817 new->v
.preferred_node
= first_node(nodes
);
2819 new->flags
|= MPOL_F_LOCAL
;
2822 * Save nodes for contextualization: this will be used to "clone"
2823 * the mempolicy in a specific context [cpuset] at a later time.
2825 new->w
.user_nodemask
= nodes
;
2830 /* Restore string for error message */
2839 #endif /* CONFIG_TMPFS */
2842 * mpol_to_str - format a mempolicy structure for printing
2843 * @buffer: to contain formatted mempolicy string
2844 * @maxlen: length of @buffer
2845 * @pol: pointer to mempolicy to be formatted
2847 * Convert @pol into a string. If @buffer is too short, truncate the string.
2848 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
2849 * longest flag, "relative", and to display at least a few node ids.
2851 void mpol_to_str(char *buffer
, int maxlen
, struct mempolicy
*pol
)
2854 nodemask_t nodes
= NODE_MASK_NONE
;
2855 unsigned short mode
= MPOL_DEFAULT
;
2856 unsigned short flags
= 0;
2858 if (pol
&& pol
!= &default_policy
&& !(pol
->flags
& MPOL_F_MORON
)) {
2866 case MPOL_PREFERRED
:
2867 if (flags
& MPOL_F_LOCAL
)
2870 node_set(pol
->v
.preferred_node
, nodes
);
2873 case MPOL_INTERLEAVE
:
2874 nodes
= pol
->v
.nodes
;
2878 snprintf(p
, maxlen
, "unknown");
2882 p
+= snprintf(p
, maxlen
, "%s", policy_modes
[mode
]);
2884 if (flags
& MPOL_MODE_FLAGS
) {
2885 p
+= snprintf(p
, buffer
+ maxlen
- p
, "=");
2888 * Currently, the only defined flags are mutually exclusive
2890 if (flags
& MPOL_F_STATIC_NODES
)
2891 p
+= snprintf(p
, buffer
+ maxlen
- p
, "static");
2892 else if (flags
& MPOL_F_RELATIVE_NODES
)
2893 p
+= snprintf(p
, buffer
+ maxlen
- p
, "relative");
2896 if (!nodes_empty(nodes
))
2897 p
+= scnprintf(p
, buffer
+ maxlen
- p
, ":%*pbl",
2898 nodemask_pr_args(&nodes
));