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/ptrace.h>
89 #include <linux/swap.h>
90 #include <linux/seq_file.h>
91 #include <linux/proc_fs.h>
92 #include <linux/migrate.h>
93 #include <linux/ksm.h>
94 #include <linux/rmap.h>
95 #include <linux/security.h>
96 #include <linux/syscalls.h>
97 #include <linux/ctype.h>
98 #include <linux/mm_inline.h>
99 #include <linux/mmu_notifier.h>
100 #include <linux/printk.h>
101 #include <linux/swapops.h>
103 #include <asm/tlbflush.h>
104 #include <linux/uaccess.h>
106 #include "internal.h"
109 #define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
110 #define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
112 static struct kmem_cache
*policy_cache
;
113 static struct kmem_cache
*sn_cache
;
115 /* Highest zone. An specific allocation for a zone below that is not
117 enum zone_type policy_zone
= 0;
120 * run-time system-wide default policy => local allocation
122 static struct mempolicy default_policy
= {
123 .refcnt
= ATOMIC_INIT(1), /* never free it */
124 .mode
= MPOL_PREFERRED
,
125 .flags
= MPOL_F_LOCAL
,
128 static struct mempolicy preferred_node_policy
[MAX_NUMNODES
];
130 struct mempolicy
*get_task_policy(struct task_struct
*p
)
132 struct mempolicy
*pol
= p
->mempolicy
;
138 node
= numa_node_id();
139 if (node
!= NUMA_NO_NODE
) {
140 pol
= &preferred_node_policy
[node
];
141 /* preferred_node_policy is not initialised early in boot */
146 return &default_policy
;
149 static const struct mempolicy_operations
{
150 int (*create
)(struct mempolicy
*pol
, const nodemask_t
*nodes
);
151 void (*rebind
)(struct mempolicy
*pol
, const nodemask_t
*nodes
);
152 } mpol_ops
[MPOL_MAX
];
154 static inline int mpol_store_user_nodemask(const struct mempolicy
*pol
)
156 return pol
->flags
& MPOL_MODE_FLAGS
;
159 static void mpol_relative_nodemask(nodemask_t
*ret
, const nodemask_t
*orig
,
160 const nodemask_t
*rel
)
163 nodes_fold(tmp
, *orig
, nodes_weight(*rel
));
164 nodes_onto(*ret
, tmp
, *rel
);
167 static int mpol_new_interleave(struct mempolicy
*pol
, const nodemask_t
*nodes
)
169 if (nodes_empty(*nodes
))
171 pol
->v
.nodes
= *nodes
;
175 static int mpol_new_preferred(struct mempolicy
*pol
, const nodemask_t
*nodes
)
178 pol
->flags
|= MPOL_F_LOCAL
; /* local allocation */
179 else if (nodes_empty(*nodes
))
180 return -EINVAL
; /* no allowed nodes */
182 pol
->v
.preferred_node
= first_node(*nodes
);
186 static int mpol_new_bind(struct mempolicy
*pol
, const nodemask_t
*nodes
)
188 if (nodes_empty(*nodes
))
190 pol
->v
.nodes
= *nodes
;
195 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
196 * any, for the new policy. mpol_new() has already validated the nodes
197 * parameter with respect to the policy mode and flags. But, we need to
198 * handle an empty nodemask with MPOL_PREFERRED here.
200 * Must be called holding task's alloc_lock to protect task's mems_allowed
201 * and mempolicy. May also be called holding the mmap_semaphore for write.
203 static int mpol_set_nodemask(struct mempolicy
*pol
,
204 const nodemask_t
*nodes
, struct nodemask_scratch
*nsc
)
208 /* if mode is MPOL_DEFAULT, pol is NULL. This is right. */
212 nodes_and(nsc
->mask1
,
213 cpuset_current_mems_allowed
, node_states
[N_MEMORY
]);
216 if (pol
->mode
== MPOL_PREFERRED
&& nodes_empty(*nodes
))
217 nodes
= NULL
; /* explicit local allocation */
219 if (pol
->flags
& MPOL_F_RELATIVE_NODES
)
220 mpol_relative_nodemask(&nsc
->mask2
, nodes
, &nsc
->mask1
);
222 nodes_and(nsc
->mask2
, *nodes
, nsc
->mask1
);
224 if (mpol_store_user_nodemask(pol
))
225 pol
->w
.user_nodemask
= *nodes
;
227 pol
->w
.cpuset_mems_allowed
=
228 cpuset_current_mems_allowed
;
232 ret
= mpol_ops
[pol
->mode
].create(pol
, &nsc
->mask2
);
234 ret
= mpol_ops
[pol
->mode
].create(pol
, NULL
);
239 * This function just creates a new policy, does some check and simple
240 * initialization. You must invoke mpol_set_nodemask() to set nodes.
242 static struct mempolicy
*mpol_new(unsigned short mode
, unsigned short flags
,
245 struct mempolicy
*policy
;
247 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
248 mode
, flags
, nodes
? nodes_addr(*nodes
)[0] : NUMA_NO_NODE
);
250 if (mode
== MPOL_DEFAULT
) {
251 if (nodes
&& !nodes_empty(*nodes
))
252 return ERR_PTR(-EINVAL
);
258 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
259 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
260 * All other modes require a valid pointer to a non-empty nodemask.
262 if (mode
== MPOL_PREFERRED
) {
263 if (nodes_empty(*nodes
)) {
264 if (((flags
& MPOL_F_STATIC_NODES
) ||
265 (flags
& MPOL_F_RELATIVE_NODES
)))
266 return ERR_PTR(-EINVAL
);
268 } else if (mode
== MPOL_LOCAL
) {
269 if (!nodes_empty(*nodes
) ||
270 (flags
& MPOL_F_STATIC_NODES
) ||
271 (flags
& MPOL_F_RELATIVE_NODES
))
272 return ERR_PTR(-EINVAL
);
273 mode
= MPOL_PREFERRED
;
274 } else if (nodes_empty(*nodes
))
275 return ERR_PTR(-EINVAL
);
276 policy
= kmem_cache_alloc(policy_cache
, GFP_KERNEL
);
278 return ERR_PTR(-ENOMEM
);
279 atomic_set(&policy
->refcnt
, 1);
281 policy
->flags
= flags
;
286 /* Slow path of a mpol destructor. */
287 void __mpol_put(struct mempolicy
*p
)
289 if (!atomic_dec_and_test(&p
->refcnt
))
291 kmem_cache_free(policy_cache
, p
);
294 static void mpol_rebind_default(struct mempolicy
*pol
, const nodemask_t
*nodes
)
298 static void mpol_rebind_nodemask(struct mempolicy
*pol
, const nodemask_t
*nodes
)
302 if (pol
->flags
& MPOL_F_STATIC_NODES
)
303 nodes_and(tmp
, pol
->w
.user_nodemask
, *nodes
);
304 else if (pol
->flags
& MPOL_F_RELATIVE_NODES
)
305 mpol_relative_nodemask(&tmp
, &pol
->w
.user_nodemask
, nodes
);
307 nodes_remap(tmp
, pol
->v
.nodes
,pol
->w
.cpuset_mems_allowed
,
309 pol
->w
.cpuset_mems_allowed
= *nodes
;
312 if (nodes_empty(tmp
))
318 static void mpol_rebind_preferred(struct mempolicy
*pol
,
319 const nodemask_t
*nodes
)
323 if (pol
->flags
& MPOL_F_STATIC_NODES
) {
324 int node
= first_node(pol
->w
.user_nodemask
);
326 if (node_isset(node
, *nodes
)) {
327 pol
->v
.preferred_node
= node
;
328 pol
->flags
&= ~MPOL_F_LOCAL
;
330 pol
->flags
|= MPOL_F_LOCAL
;
331 } else if (pol
->flags
& MPOL_F_RELATIVE_NODES
) {
332 mpol_relative_nodemask(&tmp
, &pol
->w
.user_nodemask
, nodes
);
333 pol
->v
.preferred_node
= first_node(tmp
);
334 } else if (!(pol
->flags
& MPOL_F_LOCAL
)) {
335 pol
->v
.preferred_node
= node_remap(pol
->v
.preferred_node
,
336 pol
->w
.cpuset_mems_allowed
,
338 pol
->w
.cpuset_mems_allowed
= *nodes
;
343 * mpol_rebind_policy - Migrate a policy to a different set of nodes
345 * Per-vma policies are protected by mmap_sem. Allocations using per-task
346 * policies are protected by task->mems_allowed_seq to prevent a premature
347 * OOM/allocation failure due to parallel nodemask modification.
349 static void mpol_rebind_policy(struct mempolicy
*pol
, const nodemask_t
*newmask
)
353 if (!mpol_store_user_nodemask(pol
) && !(pol
->flags
& MPOL_F_LOCAL
) &&
354 nodes_equal(pol
->w
.cpuset_mems_allowed
, *newmask
))
357 mpol_ops
[pol
->mode
].rebind(pol
, newmask
);
361 * Wrapper for mpol_rebind_policy() that just requires task
362 * pointer, and updates task mempolicy.
364 * Called with task's alloc_lock held.
367 void mpol_rebind_task(struct task_struct
*tsk
, const nodemask_t
*new)
369 mpol_rebind_policy(tsk
->mempolicy
, new);
373 * Rebind each vma in mm to new nodemask.
375 * Call holding a reference to mm. Takes mm->mmap_sem during call.
378 void mpol_rebind_mm(struct mm_struct
*mm
, nodemask_t
*new)
380 struct vm_area_struct
*vma
;
382 down_write(&mm
->mmap_sem
);
383 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
)
384 mpol_rebind_policy(vma
->vm_policy
, new);
385 up_write(&mm
->mmap_sem
);
388 static const struct mempolicy_operations mpol_ops
[MPOL_MAX
] = {
390 .rebind
= mpol_rebind_default
,
392 [MPOL_INTERLEAVE
] = {
393 .create
= mpol_new_interleave
,
394 .rebind
= mpol_rebind_nodemask
,
397 .create
= mpol_new_preferred
,
398 .rebind
= mpol_rebind_preferred
,
401 .create
= mpol_new_bind
,
402 .rebind
= mpol_rebind_nodemask
,
406 static void migrate_page_add(struct page
*page
, struct list_head
*pagelist
,
407 unsigned long flags
);
410 struct list_head
*pagelist
;
413 struct vm_area_struct
*prev
;
417 * Check if the page's nid is in qp->nmask.
419 * If MPOL_MF_INVERT is set in qp->flags, check if the nid is
420 * in the invert of qp->nmask.
422 static inline bool queue_pages_required(struct page
*page
,
423 struct queue_pages
*qp
)
425 int nid
= page_to_nid(page
);
426 unsigned long flags
= qp
->flags
;
428 return node_isset(nid
, *qp
->nmask
) == !(flags
& MPOL_MF_INVERT
);
432 * queue_pages_pmd() has three possible return values:
433 * 1 - pages are placed on the right node or queued successfully.
435 * -EIO - is migration entry or MPOL_MF_STRICT was specified and an existing
436 * page was already on a node that does not follow the policy.
438 static int queue_pages_pmd(pmd_t
*pmd
, spinlock_t
*ptl
, unsigned long addr
,
439 unsigned long end
, struct mm_walk
*walk
)
443 struct queue_pages
*qp
= walk
->private;
446 if (unlikely(is_pmd_migration_entry(*pmd
))) {
450 page
= pmd_page(*pmd
);
451 if (is_huge_zero_page(page
)) {
453 __split_huge_pmd(walk
->vma
, pmd
, addr
, false, NULL
);
456 if (!thp_migration_supported()) {
460 ret
= split_huge_page(page
);
465 if (!queue_pages_required(page
, qp
)) {
472 /* go to thp migration */
473 if (flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
)) {
474 if (!vma_migratable(walk
->vma
)) {
479 migrate_page_add(page
, qp
->pagelist
, flags
);
489 * Scan through pages checking if pages follow certain conditions,
490 * and move them to the pagelist if they do.
492 static int queue_pages_pte_range(pmd_t
*pmd
, unsigned long addr
,
493 unsigned long end
, struct mm_walk
*walk
)
495 struct vm_area_struct
*vma
= walk
->vma
;
497 struct queue_pages
*qp
= walk
->private;
498 unsigned long flags
= qp
->flags
;
503 ptl
= pmd_trans_huge_lock(pmd
, vma
);
505 ret
= queue_pages_pmd(pmd
, ptl
, addr
, end
, walk
);
512 if (pmd_trans_unstable(pmd
))
515 pte
= pte_offset_map_lock(walk
->mm
, pmd
, addr
, &ptl
);
516 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
) {
517 if (!pte_present(*pte
))
519 page
= vm_normal_page(vma
, addr
, *pte
);
523 * vm_normal_page() filters out zero pages, but there might
524 * still be PageReserved pages to skip, perhaps in a VDSO.
526 if (PageReserved(page
))
528 if (!queue_pages_required(page
, qp
))
530 if (PageTransCompound(page
) && !thp_migration_supported()) {
532 pte_unmap_unlock(pte
, ptl
);
534 ret
= split_huge_page(page
);
537 /* Failed to split -- skip. */
539 pte
= pte_offset_map_lock(walk
->mm
, pmd
,
546 if (flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
)) {
547 if (!vma_migratable(vma
))
549 migrate_page_add(page
, qp
->pagelist
, flags
);
553 pte_unmap_unlock(pte
- 1, ptl
);
555 return addr
!= end
? -EIO
: 0;
558 static int queue_pages_hugetlb(pte_t
*pte
, unsigned long hmask
,
559 unsigned long addr
, unsigned long end
,
560 struct mm_walk
*walk
)
562 #ifdef CONFIG_HUGETLB_PAGE
563 struct queue_pages
*qp
= walk
->private;
564 unsigned long flags
= qp
->flags
;
569 ptl
= huge_pte_lock(hstate_vma(walk
->vma
), walk
->mm
, pte
);
570 entry
= huge_ptep_get(pte
);
571 if (!pte_present(entry
))
573 page
= pte_page(entry
);
574 if (!queue_pages_required(page
, qp
))
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
;
626 * Need check MPOL_MF_STRICT to return -EIO if possible
627 * regardless of vma_migratable
629 if (!vma_migratable(vma
) &&
630 !(flags
& MPOL_MF_STRICT
))
635 if (vma
->vm_start
> start
)
636 start
= vma
->vm_start
;
638 if (!(flags
& MPOL_MF_DISCONTIG_OK
)) {
639 if (!vma
->vm_next
&& vma
->vm_end
< end
)
641 if (qp
->prev
&& qp
->prev
->vm_end
< vma
->vm_start
)
647 if (flags
& MPOL_MF_LAZY
) {
648 /* Similar to task_numa_work, skip inaccessible VMAs */
649 if (!is_vm_hugetlb_page(vma
) &&
650 (vma
->vm_flags
& (VM_READ
| VM_EXEC
| VM_WRITE
)) &&
651 !(vma
->vm_flags
& VM_MIXEDMAP
))
652 change_prot_numa(vma
, start
, endvma
);
656 /* queue pages from current vma */
657 if (flags
& MPOL_MF_VALID
)
663 * Walk through page tables and collect pages to be migrated.
665 * If pages found in a given range are on a set of nodes (determined by
666 * @nodes and @flags,) it's isolated and queued to the pagelist which is
667 * passed via @private.)
670 queue_pages_range(struct mm_struct
*mm
, unsigned long start
, unsigned long end
,
671 nodemask_t
*nodes
, unsigned long flags
,
672 struct list_head
*pagelist
)
674 struct queue_pages qp
= {
675 .pagelist
= pagelist
,
680 struct mm_walk queue_pages_walk
= {
681 .hugetlb_entry
= queue_pages_hugetlb
,
682 .pmd_entry
= queue_pages_pte_range
,
683 .test_walk
= queue_pages_test_walk
,
688 return walk_page_range(start
, end
, &queue_pages_walk
);
692 * Apply policy to a single VMA
693 * This must be called with the mmap_sem held for writing.
695 static int vma_replace_policy(struct vm_area_struct
*vma
,
696 struct mempolicy
*pol
)
699 struct mempolicy
*old
;
700 struct mempolicy
*new;
702 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
703 vma
->vm_start
, vma
->vm_end
, vma
->vm_pgoff
,
704 vma
->vm_ops
, vma
->vm_file
,
705 vma
->vm_ops
? vma
->vm_ops
->set_policy
: NULL
);
711 if (vma
->vm_ops
&& vma
->vm_ops
->set_policy
) {
712 err
= vma
->vm_ops
->set_policy(vma
, new);
717 old
= vma
->vm_policy
;
718 vma
->vm_policy
= new; /* protected by mmap_sem */
727 /* Step 2: apply policy to a range and do splits. */
728 static int mbind_range(struct mm_struct
*mm
, unsigned long start
,
729 unsigned long end
, struct mempolicy
*new_pol
)
731 struct vm_area_struct
*next
;
732 struct vm_area_struct
*prev
;
733 struct vm_area_struct
*vma
;
736 unsigned long vmstart
;
739 vma
= find_vma(mm
, start
);
740 if (!vma
|| vma
->vm_start
> start
)
744 if (start
> vma
->vm_start
)
747 for (; vma
&& vma
->vm_start
< end
; prev
= vma
, vma
= next
) {
749 vmstart
= max(start
, vma
->vm_start
);
750 vmend
= min(end
, vma
->vm_end
);
752 if (mpol_equal(vma_policy(vma
), new_pol
))
755 pgoff
= vma
->vm_pgoff
+
756 ((vmstart
- vma
->vm_start
) >> PAGE_SHIFT
);
757 prev
= vma_merge(mm
, prev
, vmstart
, vmend
, vma
->vm_flags
,
758 vma
->anon_vma
, vma
->vm_file
, pgoff
,
759 new_pol
, vma
->vm_userfaultfd_ctx
);
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 struct mm_struct
*mm
= NULL
;
1413 struct task_struct
*task
;
1414 nodemask_t task_nodes
;
1418 NODEMASK_SCRATCH(scratch
);
1423 old
= &scratch
->mask1
;
1424 new = &scratch
->mask2
;
1426 err
= get_nodes(old
, old_nodes
, maxnode
);
1430 err
= get_nodes(new, new_nodes
, maxnode
);
1434 /* Find the mm_struct */
1436 task
= pid
? find_task_by_vpid(pid
) : current
;
1442 get_task_struct(task
);
1447 * Check if this process has the right to modify the specified process.
1448 * Use the regular "ptrace_may_access()" checks.
1450 if (!ptrace_may_access(task
, PTRACE_MODE_READ_REALCREDS
)) {
1457 task_nodes
= cpuset_mems_allowed(task
);
1458 /* Is the user allowed to access the target nodes? */
1459 if (!nodes_subset(*new, task_nodes
) && !capable(CAP_SYS_NICE
)) {
1464 task_nodes
= cpuset_mems_allowed(current
);
1465 nodes_and(*new, *new, task_nodes
);
1466 if (nodes_empty(*new))
1469 nodes_and(*new, *new, node_states
[N_MEMORY
]);
1470 if (nodes_empty(*new))
1473 err
= security_task_movememory(task
);
1477 mm
= get_task_mm(task
);
1478 put_task_struct(task
);
1485 err
= do_migrate_pages(mm
, old
, new,
1486 capable(CAP_SYS_NICE
) ? MPOL_MF_MOVE_ALL
: MPOL_MF_MOVE
);
1490 NODEMASK_SCRATCH_FREE(scratch
);
1495 put_task_struct(task
);
1501 /* Retrieve NUMA policy */
1502 SYSCALL_DEFINE5(get_mempolicy
, int __user
*, policy
,
1503 unsigned long __user
*, nmask
, unsigned long, maxnode
,
1504 unsigned long, addr
, unsigned long, flags
)
1507 int uninitialized_var(pval
);
1510 if (nmask
!= NULL
&& maxnode
< nr_node_ids
)
1513 err
= do_get_mempolicy(&pval
, &nodes
, addr
, flags
);
1518 if (policy
&& put_user(pval
, policy
))
1522 err
= copy_nodes_to_user(nmask
, maxnode
, &nodes
);
1527 #ifdef CONFIG_COMPAT
1529 COMPAT_SYSCALL_DEFINE5(get_mempolicy
, int __user
*, policy
,
1530 compat_ulong_t __user
*, nmask
,
1531 compat_ulong_t
, maxnode
,
1532 compat_ulong_t
, addr
, compat_ulong_t
, flags
)
1535 unsigned long __user
*nm
= NULL
;
1536 unsigned long nr_bits
, alloc_size
;
1537 DECLARE_BITMAP(bm
, MAX_NUMNODES
);
1539 nr_bits
= min_t(unsigned long, maxnode
-1, nr_node_ids
);
1540 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
1543 nm
= compat_alloc_user_space(alloc_size
);
1545 err
= sys_get_mempolicy(policy
, nm
, nr_bits
+1, addr
, flags
);
1547 if (!err
&& nmask
) {
1548 unsigned long copy_size
;
1549 copy_size
= min_t(unsigned long, sizeof(bm
), alloc_size
);
1550 err
= copy_from_user(bm
, nm
, copy_size
);
1551 /* ensure entire bitmap is zeroed */
1552 err
|= clear_user(nmask
, ALIGN(maxnode
-1, 8) / 8);
1553 err
|= compat_put_bitmap(nmask
, bm
, nr_bits
);
1559 COMPAT_SYSCALL_DEFINE3(set_mempolicy
, int, mode
, compat_ulong_t __user
*, nmask
,
1560 compat_ulong_t
, maxnode
)
1562 unsigned long __user
*nm
= NULL
;
1563 unsigned long nr_bits
, alloc_size
;
1564 DECLARE_BITMAP(bm
, MAX_NUMNODES
);
1566 nr_bits
= min_t(unsigned long, maxnode
-1, MAX_NUMNODES
);
1567 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
1570 if (compat_get_bitmap(bm
, nmask
, nr_bits
))
1572 nm
= compat_alloc_user_space(alloc_size
);
1573 if (copy_to_user(nm
, bm
, alloc_size
))
1577 return sys_set_mempolicy(mode
, nm
, nr_bits
+1);
1580 COMPAT_SYSCALL_DEFINE6(mbind
, compat_ulong_t
, start
, compat_ulong_t
, len
,
1581 compat_ulong_t
, mode
, compat_ulong_t __user
*, nmask
,
1582 compat_ulong_t
, maxnode
, compat_ulong_t
, flags
)
1584 unsigned long __user
*nm
= NULL
;
1585 unsigned long nr_bits
, alloc_size
;
1588 nr_bits
= min_t(unsigned long, maxnode
-1, MAX_NUMNODES
);
1589 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
1592 if (compat_get_bitmap(nodes_addr(bm
), nmask
, nr_bits
))
1594 nm
= compat_alloc_user_space(alloc_size
);
1595 if (copy_to_user(nm
, nodes_addr(bm
), alloc_size
))
1599 return sys_mbind(start
, len
, mode
, nm
, nr_bits
+1, flags
);
1604 struct mempolicy
*__get_vma_policy(struct vm_area_struct
*vma
,
1607 struct mempolicy
*pol
= NULL
;
1610 if (vma
->vm_ops
&& vma
->vm_ops
->get_policy
) {
1611 pol
= vma
->vm_ops
->get_policy(vma
, addr
);
1612 } else if (vma
->vm_policy
) {
1613 pol
= vma
->vm_policy
;
1616 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1617 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1618 * count on these policies which will be dropped by
1619 * mpol_cond_put() later
1621 if (mpol_needs_cond_ref(pol
))
1630 * get_vma_policy(@vma, @addr)
1631 * @vma: virtual memory area whose policy is sought
1632 * @addr: address in @vma for shared policy lookup
1634 * Returns effective policy for a VMA at specified address.
1635 * Falls back to current->mempolicy or system default policy, as necessary.
1636 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1637 * count--added by the get_policy() vm_op, as appropriate--to protect against
1638 * freeing by another task. It is the caller's responsibility to free the
1639 * extra reference for shared policies.
1641 static struct mempolicy
*get_vma_policy(struct vm_area_struct
*vma
,
1644 struct mempolicy
*pol
= __get_vma_policy(vma
, addr
);
1647 pol
= get_task_policy(current
);
1652 bool vma_policy_mof(struct vm_area_struct
*vma
)
1654 struct mempolicy
*pol
;
1656 if (vma
->vm_ops
&& vma
->vm_ops
->get_policy
) {
1659 pol
= vma
->vm_ops
->get_policy(vma
, vma
->vm_start
);
1660 if (pol
&& (pol
->flags
& MPOL_F_MOF
))
1667 pol
= vma
->vm_policy
;
1669 pol
= get_task_policy(current
);
1671 return pol
->flags
& MPOL_F_MOF
;
1674 static int apply_policy_zone(struct mempolicy
*policy
, enum zone_type zone
)
1676 enum zone_type dynamic_policy_zone
= policy_zone
;
1678 BUG_ON(dynamic_policy_zone
== ZONE_MOVABLE
);
1681 * if policy->v.nodes has movable memory only,
1682 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1684 * policy->v.nodes is intersect with node_states[N_MEMORY].
1685 * so if the following test faile, it implies
1686 * policy->v.nodes has movable memory only.
1688 if (!nodes_intersects(policy
->v
.nodes
, node_states
[N_HIGH_MEMORY
]))
1689 dynamic_policy_zone
= ZONE_MOVABLE
;
1691 return zone
>= dynamic_policy_zone
;
1695 * Return a nodemask representing a mempolicy for filtering nodes for
1698 static nodemask_t
*policy_nodemask(gfp_t gfp
, struct mempolicy
*policy
)
1700 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1701 if (unlikely(policy
->mode
== MPOL_BIND
) &&
1702 apply_policy_zone(policy
, gfp_zone(gfp
)) &&
1703 cpuset_nodemask_valid_mems_allowed(&policy
->v
.nodes
))
1704 return &policy
->v
.nodes
;
1709 /* Return the node id preferred by the given mempolicy, or the given id */
1710 static int policy_node(gfp_t gfp
, struct mempolicy
*policy
,
1713 if (policy
->mode
== MPOL_PREFERRED
&& !(policy
->flags
& MPOL_F_LOCAL
))
1714 nd
= policy
->v
.preferred_node
;
1717 * __GFP_THISNODE shouldn't even be used with the bind policy
1718 * because we might easily break the expectation to stay on the
1719 * requested node and not break the policy.
1721 WARN_ON_ONCE(policy
->mode
== MPOL_BIND
&& (gfp
& __GFP_THISNODE
));
1727 /* Do dynamic interleaving for a process */
1728 static unsigned interleave_nodes(struct mempolicy
*policy
)
1731 struct task_struct
*me
= current
;
1733 next
= next_node_in(me
->il_prev
, policy
->v
.nodes
);
1734 if (next
< MAX_NUMNODES
)
1740 * Depending on the memory policy provide a node from which to allocate the
1743 unsigned int mempolicy_slab_node(void)
1745 struct mempolicy
*policy
;
1746 int node
= numa_mem_id();
1751 policy
= current
->mempolicy
;
1752 if (!policy
|| policy
->flags
& MPOL_F_LOCAL
)
1755 switch (policy
->mode
) {
1756 case MPOL_PREFERRED
:
1758 * handled MPOL_F_LOCAL above
1760 return policy
->v
.preferred_node
;
1762 case MPOL_INTERLEAVE
:
1763 return interleave_nodes(policy
);
1769 * Follow bind policy behavior and start allocation at the
1772 struct zonelist
*zonelist
;
1773 enum zone_type highest_zoneidx
= gfp_zone(GFP_KERNEL
);
1774 zonelist
= &NODE_DATA(node
)->node_zonelists
[ZONELIST_FALLBACK
];
1775 z
= first_zones_zonelist(zonelist
, highest_zoneidx
,
1777 return z
->zone
? z
->zone
->node
: node
;
1786 * Do static interleaving for a VMA with known offset @n. Returns the n'th
1787 * node in pol->v.nodes (starting from n=0), wrapping around if n exceeds the
1788 * number of present nodes.
1790 static unsigned offset_il_node(struct mempolicy
*pol
, unsigned long n
)
1792 unsigned nnodes
= nodes_weight(pol
->v
.nodes
);
1798 return numa_node_id();
1799 target
= (unsigned int)n
% nnodes
;
1800 nid
= first_node(pol
->v
.nodes
);
1801 for (i
= 0; i
< target
; i
++)
1802 nid
= next_node(nid
, pol
->v
.nodes
);
1806 /* Determine a node number for interleave */
1807 static inline unsigned interleave_nid(struct mempolicy
*pol
,
1808 struct vm_area_struct
*vma
, unsigned long addr
, int shift
)
1814 * for small pages, there is no difference between
1815 * shift and PAGE_SHIFT, so the bit-shift is safe.
1816 * for huge pages, since vm_pgoff is in units of small
1817 * pages, we need to shift off the always 0 bits to get
1820 BUG_ON(shift
< PAGE_SHIFT
);
1821 off
= vma
->vm_pgoff
>> (shift
- PAGE_SHIFT
);
1822 off
+= (addr
- vma
->vm_start
) >> shift
;
1823 return offset_il_node(pol
, off
);
1825 return interleave_nodes(pol
);
1828 #ifdef CONFIG_HUGETLBFS
1830 * huge_node(@vma, @addr, @gfp_flags, @mpol)
1831 * @vma: virtual memory area whose policy is sought
1832 * @addr: address in @vma for shared policy lookup and interleave policy
1833 * @gfp_flags: for requested zone
1834 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
1835 * @nodemask: pointer to nodemask pointer for MPOL_BIND nodemask
1837 * Returns a nid suitable for a huge page allocation and a pointer
1838 * to the struct mempolicy for conditional unref after allocation.
1839 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1840 * @nodemask for filtering the zonelist.
1842 * Must be protected by read_mems_allowed_begin()
1844 int huge_node(struct vm_area_struct
*vma
, unsigned long addr
, gfp_t gfp_flags
,
1845 struct mempolicy
**mpol
, nodemask_t
**nodemask
)
1849 *mpol
= get_vma_policy(vma
, addr
);
1850 *nodemask
= NULL
; /* assume !MPOL_BIND */
1852 if (unlikely((*mpol
)->mode
== MPOL_INTERLEAVE
)) {
1853 nid
= interleave_nid(*mpol
, vma
, addr
,
1854 huge_page_shift(hstate_vma(vma
)));
1856 nid
= policy_node(gfp_flags
, *mpol
, numa_node_id());
1857 if ((*mpol
)->mode
== MPOL_BIND
)
1858 *nodemask
= &(*mpol
)->v
.nodes
;
1864 * init_nodemask_of_mempolicy
1866 * If the current task's mempolicy is "default" [NULL], return 'false'
1867 * to indicate default policy. Otherwise, extract the policy nodemask
1868 * for 'bind' or 'interleave' policy into the argument nodemask, or
1869 * initialize the argument nodemask to contain the single node for
1870 * 'preferred' or 'local' policy and return 'true' to indicate presence
1871 * of non-default mempolicy.
1873 * We don't bother with reference counting the mempolicy [mpol_get/put]
1874 * because the current task is examining it's own mempolicy and a task's
1875 * mempolicy is only ever changed by the task itself.
1877 * N.B., it is the caller's responsibility to free a returned nodemask.
1879 bool init_nodemask_of_mempolicy(nodemask_t
*mask
)
1881 struct mempolicy
*mempolicy
;
1884 if (!(mask
&& current
->mempolicy
))
1888 mempolicy
= current
->mempolicy
;
1889 switch (mempolicy
->mode
) {
1890 case MPOL_PREFERRED
:
1891 if (mempolicy
->flags
& MPOL_F_LOCAL
)
1892 nid
= numa_node_id();
1894 nid
= mempolicy
->v
.preferred_node
;
1895 init_nodemask_of_node(mask
, nid
);
1900 case MPOL_INTERLEAVE
:
1901 *mask
= mempolicy
->v
.nodes
;
1907 task_unlock(current
);
1914 * mempolicy_nodemask_intersects
1916 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1917 * policy. Otherwise, check for intersection between mask and the policy
1918 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1919 * policy, always return true since it may allocate elsewhere on fallback.
1921 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1923 bool mempolicy_nodemask_intersects(struct task_struct
*tsk
,
1924 const nodemask_t
*mask
)
1926 struct mempolicy
*mempolicy
;
1932 mempolicy
= tsk
->mempolicy
;
1936 switch (mempolicy
->mode
) {
1937 case MPOL_PREFERRED
:
1939 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1940 * allocate from, they may fallback to other nodes when oom.
1941 * Thus, it's possible for tsk to have allocated memory from
1946 case MPOL_INTERLEAVE
:
1947 ret
= nodes_intersects(mempolicy
->v
.nodes
, *mask
);
1957 /* Allocate a page in interleaved policy.
1958 Own path because it needs to do special accounting. */
1959 static struct page
*alloc_page_interleave(gfp_t gfp
, unsigned order
,
1964 page
= __alloc_pages(gfp
, order
, nid
);
1965 /* skip NUMA_INTERLEAVE_HIT counter update if numa stats is disabled */
1966 if (!static_branch_likely(&vm_numa_stat_key
))
1968 if (page
&& page_to_nid(page
) == nid
) {
1970 __inc_numa_state(page_zone(page
), NUMA_INTERLEAVE_HIT
);
1977 * alloc_pages_vma - Allocate a page for a VMA.
1980 * %GFP_USER user allocation.
1981 * %GFP_KERNEL kernel allocations,
1982 * %GFP_HIGHMEM highmem/user allocations,
1983 * %GFP_FS allocation should not call back into a file system.
1984 * %GFP_ATOMIC don't sleep.
1986 * @order:Order of the GFP allocation.
1987 * @vma: Pointer to VMA or NULL if not available.
1988 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1989 * @node: Which node to prefer for allocation (modulo policy).
1990 * @hugepage: for hugepages try only the preferred node if possible
1992 * This function allocates a page from the kernel page pool and applies
1993 * a NUMA policy associated with the VMA or the current process.
1994 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1995 * mm_struct of the VMA to prevent it from going away. Should be used for
1996 * all allocations for pages that will be mapped into user space. Returns
1997 * NULL when no page can be allocated.
2000 alloc_pages_vma(gfp_t gfp
, int order
, struct vm_area_struct
*vma
,
2001 unsigned long addr
, int node
, bool hugepage
)
2003 struct mempolicy
*pol
;
2008 pol
= get_vma_policy(vma
, addr
);
2010 if (pol
->mode
== MPOL_INTERLEAVE
) {
2013 nid
= interleave_nid(pol
, vma
, addr
, PAGE_SHIFT
+ order
);
2015 page
= alloc_page_interleave(gfp
, order
, nid
);
2019 if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE
) && hugepage
)) {
2020 int hpage_node
= node
;
2023 * For hugepage allocation and non-interleave policy which
2024 * allows the current node (or other explicitly preferred
2025 * node) we only try to allocate from the current/preferred
2026 * node and don't fall back to other nodes, as the cost of
2027 * remote accesses would likely offset THP benefits.
2029 * If the policy is interleave, or does not allow the current
2030 * node in its nodemask, we allocate the standard way.
2032 if (pol
->mode
== MPOL_PREFERRED
&&
2033 !(pol
->flags
& MPOL_F_LOCAL
))
2034 hpage_node
= pol
->v
.preferred_node
;
2036 nmask
= policy_nodemask(gfp
, pol
);
2037 if (!nmask
|| node_isset(hpage_node
, *nmask
)) {
2040 * We cannot invoke reclaim if __GFP_THISNODE
2041 * is set. Invoking reclaim with
2042 * __GFP_THISNODE set, would cause THP
2043 * allocations to trigger heavy swapping
2044 * despite there may be tons of free memory
2045 * (including potentially plenty of THP
2046 * already available in the buddy) on all the
2049 * At most we could invoke compaction when
2050 * __GFP_THISNODE is set (but we would need to
2051 * refrain from invoking reclaim even if
2052 * compaction returned COMPACT_SKIPPED because
2053 * there wasn't not enough memory to succeed
2054 * compaction). For now just avoid
2055 * __GFP_THISNODE instead of limiting the
2056 * allocation path to a strict and single
2057 * compaction invocation.
2059 * Supposedly if direct reclaim was enabled by
2060 * the caller, the app prefers THP regardless
2061 * of the node it comes from so this would be
2062 * more desiderable behavior than only
2063 * providing THP originated from the local
2064 * node in such case.
2066 if (!(gfp
& __GFP_DIRECT_RECLAIM
))
2067 gfp
|= __GFP_THISNODE
;
2068 page
= __alloc_pages_node(hpage_node
, gfp
, order
);
2073 nmask
= policy_nodemask(gfp
, pol
);
2074 preferred_nid
= policy_node(gfp
, pol
, node
);
2075 page
= __alloc_pages_nodemask(gfp
, order
, preferred_nid
, nmask
);
2082 * alloc_pages_current - Allocate pages.
2085 * %GFP_USER user allocation,
2086 * %GFP_KERNEL kernel allocation,
2087 * %GFP_HIGHMEM highmem allocation,
2088 * %GFP_FS don't call back into a file system.
2089 * %GFP_ATOMIC don't sleep.
2090 * @order: Power of two of allocation size in pages. 0 is a single page.
2092 * Allocate a page from the kernel page pool. When not in
2093 * interrupt context and apply the current process NUMA policy.
2094 * Returns NULL when no page can be allocated.
2096 struct page
*alloc_pages_current(gfp_t gfp
, unsigned order
)
2098 struct mempolicy
*pol
= &default_policy
;
2101 if (!in_interrupt() && !(gfp
& __GFP_THISNODE
))
2102 pol
= get_task_policy(current
);
2105 * No reference counting needed for current->mempolicy
2106 * nor system default_policy
2108 if (pol
->mode
== MPOL_INTERLEAVE
)
2109 page
= alloc_page_interleave(gfp
, order
, interleave_nodes(pol
));
2111 page
= __alloc_pages_nodemask(gfp
, order
,
2112 policy_node(gfp
, pol
, numa_node_id()),
2113 policy_nodemask(gfp
, pol
));
2117 EXPORT_SYMBOL(alloc_pages_current
);
2119 int vma_dup_policy(struct vm_area_struct
*src
, struct vm_area_struct
*dst
)
2121 struct mempolicy
*pol
= mpol_dup(vma_policy(src
));
2124 return PTR_ERR(pol
);
2125 dst
->vm_policy
= pol
;
2130 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2131 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2132 * with the mems_allowed returned by cpuset_mems_allowed(). This
2133 * keeps mempolicies cpuset relative after its cpuset moves. See
2134 * further kernel/cpuset.c update_nodemask().
2136 * current's mempolicy may be rebinded by the other task(the task that changes
2137 * cpuset's mems), so we needn't do rebind work for current task.
2140 /* Slow path of a mempolicy duplicate */
2141 struct mempolicy
*__mpol_dup(struct mempolicy
*old
)
2143 struct mempolicy
*new = kmem_cache_alloc(policy_cache
, GFP_KERNEL
);
2146 return ERR_PTR(-ENOMEM
);
2148 /* task's mempolicy is protected by alloc_lock */
2149 if (old
== current
->mempolicy
) {
2152 task_unlock(current
);
2156 if (current_cpuset_is_being_rebound()) {
2157 nodemask_t mems
= cpuset_mems_allowed(current
);
2158 mpol_rebind_policy(new, &mems
);
2160 atomic_set(&new->refcnt
, 1);
2164 /* Slow path of a mempolicy comparison */
2165 bool __mpol_equal(struct mempolicy
*a
, struct mempolicy
*b
)
2169 if (a
->mode
!= b
->mode
)
2171 if (a
->flags
!= b
->flags
)
2173 if (mpol_store_user_nodemask(a
))
2174 if (!nodes_equal(a
->w
.user_nodemask
, b
->w
.user_nodemask
))
2180 case MPOL_INTERLEAVE
:
2181 return !!nodes_equal(a
->v
.nodes
, b
->v
.nodes
);
2182 case MPOL_PREFERRED
:
2183 /* a's ->flags is the same as b's */
2184 if (a
->flags
& MPOL_F_LOCAL
)
2186 return a
->v
.preferred_node
== b
->v
.preferred_node
;
2194 * Shared memory backing store policy support.
2196 * Remember policies even when nobody has shared memory mapped.
2197 * The policies are kept in Red-Black tree linked from the inode.
2198 * They are protected by the sp->lock rwlock, which should be held
2199 * for any accesses to the tree.
2203 * lookup first element intersecting start-end. Caller holds sp->lock for
2204 * reading or for writing
2206 static struct sp_node
*
2207 sp_lookup(struct shared_policy
*sp
, unsigned long start
, unsigned long end
)
2209 struct rb_node
*n
= sp
->root
.rb_node
;
2212 struct sp_node
*p
= rb_entry(n
, struct sp_node
, nd
);
2214 if (start
>= p
->end
)
2216 else if (end
<= p
->start
)
2224 struct sp_node
*w
= NULL
;
2225 struct rb_node
*prev
= rb_prev(n
);
2228 w
= rb_entry(prev
, struct sp_node
, nd
);
2229 if (w
->end
<= start
)
2233 return rb_entry(n
, struct sp_node
, nd
);
2237 * Insert a new shared policy into the list. Caller holds sp->lock for
2240 static void sp_insert(struct shared_policy
*sp
, struct sp_node
*new)
2242 struct rb_node
**p
= &sp
->root
.rb_node
;
2243 struct rb_node
*parent
= NULL
;
2248 nd
= rb_entry(parent
, struct sp_node
, nd
);
2249 if (new->start
< nd
->start
)
2251 else if (new->end
> nd
->end
)
2252 p
= &(*p
)->rb_right
;
2256 rb_link_node(&new->nd
, parent
, p
);
2257 rb_insert_color(&new->nd
, &sp
->root
);
2258 pr_debug("inserting %lx-%lx: %d\n", new->start
, new->end
,
2259 new->policy
? new->policy
->mode
: 0);
2262 /* Find shared policy intersecting idx */
2264 mpol_shared_policy_lookup(struct shared_policy
*sp
, unsigned long idx
)
2266 struct mempolicy
*pol
= NULL
;
2269 if (!sp
->root
.rb_node
)
2271 read_lock(&sp
->lock
);
2272 sn
= sp_lookup(sp
, idx
, idx
+1);
2274 mpol_get(sn
->policy
);
2277 read_unlock(&sp
->lock
);
2281 static void sp_free(struct sp_node
*n
)
2283 mpol_put(n
->policy
);
2284 kmem_cache_free(sn_cache
, n
);
2288 * mpol_misplaced - check whether current page node is valid in policy
2290 * @page: page to be checked
2291 * @vma: vm area where page mapped
2292 * @addr: virtual address where page mapped
2294 * Lookup current policy node id for vma,addr and "compare to" page's
2298 * -1 - not misplaced, page is in the right node
2299 * node - node id where the page should be
2301 * Policy determination "mimics" alloc_page_vma().
2302 * Called from fault path where we know the vma and faulting address.
2304 int mpol_misplaced(struct page
*page
, struct vm_area_struct
*vma
, unsigned long addr
)
2306 struct mempolicy
*pol
;
2308 int curnid
= page_to_nid(page
);
2309 unsigned long pgoff
;
2310 int thiscpu
= raw_smp_processor_id();
2311 int thisnid
= cpu_to_node(thiscpu
);
2315 pol
= get_vma_policy(vma
, addr
);
2316 if (!(pol
->flags
& MPOL_F_MOF
))
2319 switch (pol
->mode
) {
2320 case MPOL_INTERLEAVE
:
2321 pgoff
= vma
->vm_pgoff
;
2322 pgoff
+= (addr
- vma
->vm_start
) >> PAGE_SHIFT
;
2323 polnid
= offset_il_node(pol
, pgoff
);
2326 case MPOL_PREFERRED
:
2327 if (pol
->flags
& MPOL_F_LOCAL
)
2328 polnid
= numa_node_id();
2330 polnid
= pol
->v
.preferred_node
;
2336 * allows binding to multiple nodes.
2337 * use current page if in policy nodemask,
2338 * else select nearest allowed node, if any.
2339 * If no allowed nodes, use current [!misplaced].
2341 if (node_isset(curnid
, pol
->v
.nodes
))
2343 z
= first_zones_zonelist(
2344 node_zonelist(numa_node_id(), GFP_HIGHUSER
),
2345 gfp_zone(GFP_HIGHUSER
),
2347 polnid
= z
->zone
->node
;
2354 /* Migrate the page towards the node whose CPU is referencing it */
2355 if (pol
->flags
& MPOL_F_MORON
) {
2358 if (!should_numa_migrate_memory(current
, page
, curnid
, thiscpu
))
2362 if (curnid
!= polnid
)
2371 * Drop the (possibly final) reference to task->mempolicy. It needs to be
2372 * dropped after task->mempolicy is set to NULL so that any allocation done as
2373 * part of its kmem_cache_free(), such as by KASAN, doesn't reference a freed
2376 void mpol_put_task_policy(struct task_struct
*task
)
2378 struct mempolicy
*pol
;
2381 pol
= task
->mempolicy
;
2382 task
->mempolicy
= NULL
;
2387 static void sp_delete(struct shared_policy
*sp
, struct sp_node
*n
)
2389 pr_debug("deleting %lx-l%lx\n", n
->start
, n
->end
);
2390 rb_erase(&n
->nd
, &sp
->root
);
2394 static void sp_node_init(struct sp_node
*node
, unsigned long start
,
2395 unsigned long end
, struct mempolicy
*pol
)
2397 node
->start
= start
;
2402 static struct sp_node
*sp_alloc(unsigned long start
, unsigned long end
,
2403 struct mempolicy
*pol
)
2406 struct mempolicy
*newpol
;
2408 n
= kmem_cache_alloc(sn_cache
, GFP_KERNEL
);
2412 newpol
= mpol_dup(pol
);
2413 if (IS_ERR(newpol
)) {
2414 kmem_cache_free(sn_cache
, n
);
2417 newpol
->flags
|= MPOL_F_SHARED
;
2418 sp_node_init(n
, start
, end
, newpol
);
2423 /* Replace a policy range. */
2424 static int shared_policy_replace(struct shared_policy
*sp
, unsigned long start
,
2425 unsigned long end
, struct sp_node
*new)
2428 struct sp_node
*n_new
= NULL
;
2429 struct mempolicy
*mpol_new
= NULL
;
2433 write_lock(&sp
->lock
);
2434 n
= sp_lookup(sp
, start
, end
);
2435 /* Take care of old policies in the same range. */
2436 while (n
&& n
->start
< end
) {
2437 struct rb_node
*next
= rb_next(&n
->nd
);
2438 if (n
->start
>= start
) {
2444 /* Old policy spanning whole new range. */
2449 *mpol_new
= *n
->policy
;
2450 atomic_set(&mpol_new
->refcnt
, 1);
2451 sp_node_init(n_new
, end
, n
->end
, mpol_new
);
2453 sp_insert(sp
, n_new
);
2462 n
= rb_entry(next
, struct sp_node
, nd
);
2466 write_unlock(&sp
->lock
);
2473 kmem_cache_free(sn_cache
, n_new
);
2478 write_unlock(&sp
->lock
);
2480 n_new
= kmem_cache_alloc(sn_cache
, GFP_KERNEL
);
2483 mpol_new
= kmem_cache_alloc(policy_cache
, GFP_KERNEL
);
2490 * mpol_shared_policy_init - initialize shared policy for inode
2491 * @sp: pointer to inode shared policy
2492 * @mpol: struct mempolicy to install
2494 * Install non-NULL @mpol in inode's shared policy rb-tree.
2495 * On entry, the current task has a reference on a non-NULL @mpol.
2496 * This must be released on exit.
2497 * This is called at get_inode() calls and we can use GFP_KERNEL.
2499 void mpol_shared_policy_init(struct shared_policy
*sp
, struct mempolicy
*mpol
)
2503 sp
->root
= RB_ROOT
; /* empty tree == default mempolicy */
2504 rwlock_init(&sp
->lock
);
2507 struct vm_area_struct pvma
;
2508 struct mempolicy
*new;
2509 NODEMASK_SCRATCH(scratch
);
2513 /* contextualize the tmpfs mount point mempolicy */
2514 new = mpol_new(mpol
->mode
, mpol
->flags
, &mpol
->w
.user_nodemask
);
2516 goto free_scratch
; /* no valid nodemask intersection */
2519 ret
= mpol_set_nodemask(new, &mpol
->w
.user_nodemask
, scratch
);
2520 task_unlock(current
);
2524 /* Create pseudo-vma that contains just the policy */
2525 memset(&pvma
, 0, sizeof(struct vm_area_struct
));
2526 pvma
.vm_end
= TASK_SIZE
; /* policy covers entire file */
2527 mpol_set_shared_policy(sp
, &pvma
, new); /* adds ref */
2530 mpol_put(new); /* drop initial ref */
2532 NODEMASK_SCRATCH_FREE(scratch
);
2534 mpol_put(mpol
); /* drop our incoming ref on sb mpol */
2538 int mpol_set_shared_policy(struct shared_policy
*info
,
2539 struct vm_area_struct
*vma
, struct mempolicy
*npol
)
2542 struct sp_node
*new = NULL
;
2543 unsigned long sz
= vma_pages(vma
);
2545 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2547 sz
, npol
? npol
->mode
: -1,
2548 npol
? npol
->flags
: -1,
2549 npol
? nodes_addr(npol
->v
.nodes
)[0] : NUMA_NO_NODE
);
2552 new = sp_alloc(vma
->vm_pgoff
, vma
->vm_pgoff
+ sz
, npol
);
2556 err
= shared_policy_replace(info
, vma
->vm_pgoff
, vma
->vm_pgoff
+sz
, new);
2562 /* Free a backing policy store on inode delete. */
2563 void mpol_free_shared_policy(struct shared_policy
*p
)
2566 struct rb_node
*next
;
2568 if (!p
->root
.rb_node
)
2570 write_lock(&p
->lock
);
2571 next
= rb_first(&p
->root
);
2573 n
= rb_entry(next
, struct sp_node
, nd
);
2574 next
= rb_next(&n
->nd
);
2577 write_unlock(&p
->lock
);
2580 #ifdef CONFIG_NUMA_BALANCING
2581 static int __initdata numabalancing_override
;
2583 static void __init
check_numabalancing_enable(void)
2585 bool numabalancing_default
= false;
2587 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED
))
2588 numabalancing_default
= true;
2590 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2591 if (numabalancing_override
)
2592 set_numabalancing_state(numabalancing_override
== 1);
2594 if (num_online_nodes() > 1 && !numabalancing_override
) {
2595 pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
2596 numabalancing_default
? "Enabling" : "Disabling");
2597 set_numabalancing_state(numabalancing_default
);
2601 static int __init
setup_numabalancing(char *str
)
2607 if (!strcmp(str
, "enable")) {
2608 numabalancing_override
= 1;
2610 } else if (!strcmp(str
, "disable")) {
2611 numabalancing_override
= -1;
2616 pr_warn("Unable to parse numa_balancing=\n");
2620 __setup("numa_balancing=", setup_numabalancing
);
2622 static inline void __init
check_numabalancing_enable(void)
2625 #endif /* CONFIG_NUMA_BALANCING */
2627 /* assumes fs == KERNEL_DS */
2628 void __init
numa_policy_init(void)
2630 nodemask_t interleave_nodes
;
2631 unsigned long largest
= 0;
2632 int nid
, prefer
= 0;
2634 policy_cache
= kmem_cache_create("numa_policy",
2635 sizeof(struct mempolicy
),
2636 0, SLAB_PANIC
, NULL
);
2638 sn_cache
= kmem_cache_create("shared_policy_node",
2639 sizeof(struct sp_node
),
2640 0, SLAB_PANIC
, NULL
);
2642 for_each_node(nid
) {
2643 preferred_node_policy
[nid
] = (struct mempolicy
) {
2644 .refcnt
= ATOMIC_INIT(1),
2645 .mode
= MPOL_PREFERRED
,
2646 .flags
= MPOL_F_MOF
| MPOL_F_MORON
,
2647 .v
= { .preferred_node
= nid
, },
2652 * Set interleaving policy for system init. Interleaving is only
2653 * enabled across suitably sized nodes (default is >= 16MB), or
2654 * fall back to the largest node if they're all smaller.
2656 nodes_clear(interleave_nodes
);
2657 for_each_node_state(nid
, N_MEMORY
) {
2658 unsigned long total_pages
= node_present_pages(nid
);
2660 /* Preserve the largest node */
2661 if (largest
< total_pages
) {
2662 largest
= total_pages
;
2666 /* Interleave this node? */
2667 if ((total_pages
<< PAGE_SHIFT
) >= (16 << 20))
2668 node_set(nid
, interleave_nodes
);
2671 /* All too small, use the largest */
2672 if (unlikely(nodes_empty(interleave_nodes
)))
2673 node_set(prefer
, interleave_nodes
);
2675 if (do_set_mempolicy(MPOL_INTERLEAVE
, 0, &interleave_nodes
))
2676 pr_err("%s: interleaving failed\n", __func__
);
2678 check_numabalancing_enable();
2681 /* Reset policy of current process to default */
2682 void numa_default_policy(void)
2684 do_set_mempolicy(MPOL_DEFAULT
, 0, NULL
);
2688 * Parse and format mempolicy from/to strings
2692 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
2694 static const char * const policy_modes
[] =
2696 [MPOL_DEFAULT
] = "default",
2697 [MPOL_PREFERRED
] = "prefer",
2698 [MPOL_BIND
] = "bind",
2699 [MPOL_INTERLEAVE
] = "interleave",
2700 [MPOL_LOCAL
] = "local",
2706 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2707 * @str: string containing mempolicy to parse
2708 * @mpol: pointer to struct mempolicy pointer, returned on success.
2711 * <mode>[=<flags>][:<nodelist>]
2713 * On success, returns 0, else 1
2715 int mpol_parse_str(char *str
, struct mempolicy
**mpol
)
2717 struct mempolicy
*new = NULL
;
2718 unsigned short mode
;
2719 unsigned short mode_flags
;
2721 char *nodelist
= strchr(str
, ':');
2722 char *flags
= strchr(str
, '=');
2726 /* NUL-terminate mode or flags string */
2728 if (nodelist_parse(nodelist
, nodes
))
2730 if (!nodes_subset(nodes
, node_states
[N_MEMORY
]))
2736 *flags
++ = '\0'; /* terminate mode string */
2738 for (mode
= 0; mode
< MPOL_MAX
; mode
++) {
2739 if (!strcmp(str
, policy_modes
[mode
])) {
2743 if (mode
>= MPOL_MAX
)
2747 case MPOL_PREFERRED
:
2749 * Insist on a nodelist of one node only
2752 char *rest
= nodelist
;
2753 while (isdigit(*rest
))
2759 case MPOL_INTERLEAVE
:
2761 * Default to online nodes with memory if no nodelist
2764 nodes
= node_states
[N_MEMORY
];
2768 * Don't allow a nodelist; mpol_new() checks flags
2772 mode
= MPOL_PREFERRED
;
2776 * Insist on a empty nodelist
2783 * Insist on a nodelist
2792 * Currently, we only support two mutually exclusive
2795 if (!strcmp(flags
, "static"))
2796 mode_flags
|= MPOL_F_STATIC_NODES
;
2797 else if (!strcmp(flags
, "relative"))
2798 mode_flags
|= MPOL_F_RELATIVE_NODES
;
2803 new = mpol_new(mode
, mode_flags
, &nodes
);
2808 * Save nodes for mpol_to_str() to show the tmpfs mount options
2809 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2811 if (mode
!= MPOL_PREFERRED
)
2812 new->v
.nodes
= nodes
;
2814 new->v
.preferred_node
= first_node(nodes
);
2816 new->flags
|= MPOL_F_LOCAL
;
2819 * Save nodes for contextualization: this will be used to "clone"
2820 * the mempolicy in a specific context [cpuset] at a later time.
2822 new->w
.user_nodemask
= nodes
;
2827 /* Restore string for error message */
2836 #endif /* CONFIG_TMPFS */
2839 * mpol_to_str - format a mempolicy structure for printing
2840 * @buffer: to contain formatted mempolicy string
2841 * @maxlen: length of @buffer
2842 * @pol: pointer to mempolicy to be formatted
2844 * Convert @pol into a string. If @buffer is too short, truncate the string.
2845 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
2846 * longest flag, "relative", and to display at least a few node ids.
2848 void mpol_to_str(char *buffer
, int maxlen
, struct mempolicy
*pol
)
2851 nodemask_t nodes
= NODE_MASK_NONE
;
2852 unsigned short mode
= MPOL_DEFAULT
;
2853 unsigned short flags
= 0;
2855 if (pol
&& pol
!= &default_policy
&& !(pol
->flags
& MPOL_F_MORON
)) {
2863 case MPOL_PREFERRED
:
2864 if (flags
& MPOL_F_LOCAL
)
2867 node_set(pol
->v
.preferred_node
, nodes
);
2870 case MPOL_INTERLEAVE
:
2871 nodes
= pol
->v
.nodes
;
2875 snprintf(p
, maxlen
, "unknown");
2879 p
+= snprintf(p
, maxlen
, "%s", policy_modes
[mode
]);
2881 if (flags
& MPOL_MODE_FLAGS
) {
2882 p
+= snprintf(p
, buffer
+ maxlen
- p
, "=");
2885 * Currently, the only defined flags are mutually exclusive
2887 if (flags
& MPOL_F_STATIC_NODES
)
2888 p
+= snprintf(p
, buffer
+ maxlen
- p
, "static");
2889 else if (flags
& MPOL_F_RELATIVE_NODES
)
2890 p
+= snprintf(p
, buffer
+ maxlen
- p
, "relative");
2893 if (!nodes_empty(nodes
))
2894 p
+= scnprintf(p
, buffer
+ maxlen
- p
, ":%*pbl",
2895 nodemask_pr_args(&nodes
));