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
= tmp
;
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
) &&
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
);
431 static int queue_pages_pmd(pmd_t
*pmd
, spinlock_t
*ptl
, unsigned long addr
,
432 unsigned long end
, struct mm_walk
*walk
)
436 struct queue_pages
*qp
= walk
->private;
439 if (unlikely(is_pmd_migration_entry(*pmd
))) {
443 page
= pmd_page(*pmd
);
444 if (is_huge_zero_page(page
)) {
446 __split_huge_pmd(walk
->vma
, pmd
, addr
, false, NULL
);
449 if (!thp_migration_supported()) {
453 ret
= split_huge_page(page
);
458 if (!queue_pages_required(page
, qp
)) {
465 /* go to thp migration */
466 if (flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
))
467 migrate_page_add(page
, qp
->pagelist
, flags
);
475 * Scan through pages checking if pages follow certain conditions,
476 * and move them to the pagelist if they do.
478 static int queue_pages_pte_range(pmd_t
*pmd
, unsigned long addr
,
479 unsigned long end
, struct mm_walk
*walk
)
481 struct vm_area_struct
*vma
= walk
->vma
;
483 struct queue_pages
*qp
= walk
->private;
484 unsigned long flags
= qp
->flags
;
489 ptl
= pmd_trans_huge_lock(pmd
, vma
);
491 ret
= queue_pages_pmd(pmd
, ptl
, addr
, end
, walk
);
496 if (pmd_trans_unstable(pmd
))
499 pte
= pte_offset_map_lock(walk
->mm
, pmd
, addr
, &ptl
);
500 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
) {
501 if (!pte_present(*pte
))
503 page
= vm_normal_page(vma
, addr
, *pte
);
507 * vm_normal_page() filters out zero pages, but there might
508 * still be PageReserved pages to skip, perhaps in a VDSO.
510 if (PageReserved(page
))
512 if (!queue_pages_required(page
, qp
))
514 if (PageTransCompound(page
) && !thp_migration_supported()) {
516 pte_unmap_unlock(pte
, ptl
);
518 ret
= split_huge_page(page
);
521 /* Failed to split -- skip. */
523 pte
= pte_offset_map_lock(walk
->mm
, pmd
,
530 migrate_page_add(page
, qp
->pagelist
, flags
);
532 pte_unmap_unlock(pte
- 1, ptl
);
537 static int queue_pages_hugetlb(pte_t
*pte
, unsigned long hmask
,
538 unsigned long addr
, unsigned long end
,
539 struct mm_walk
*walk
)
541 #ifdef CONFIG_HUGETLB_PAGE
542 struct queue_pages
*qp
= walk
->private;
543 unsigned long flags
= qp
->flags
;
548 ptl
= huge_pte_lock(hstate_vma(walk
->vma
), walk
->mm
, pte
);
549 entry
= huge_ptep_get(pte
);
550 if (!pte_present(entry
))
552 page
= pte_page(entry
);
553 if (!queue_pages_required(page
, qp
))
555 /* With MPOL_MF_MOVE, we migrate only unshared hugepage. */
556 if (flags
& (MPOL_MF_MOVE_ALL
) ||
557 (flags
& MPOL_MF_MOVE
&& page_mapcount(page
) == 1))
558 isolate_huge_page(page
, qp
->pagelist
);
567 #ifdef CONFIG_NUMA_BALANCING
569 * This is used to mark a range of virtual addresses to be inaccessible.
570 * These are later cleared by a NUMA hinting fault. Depending on these
571 * faults, pages may be migrated for better NUMA placement.
573 * This is assuming that NUMA faults are handled using PROT_NONE. If
574 * an architecture makes a different choice, it will need further
575 * changes to the core.
577 unsigned long change_prot_numa(struct vm_area_struct
*vma
,
578 unsigned long addr
, unsigned long end
)
582 nr_updated
= change_protection(vma
, addr
, end
, PAGE_NONE
, 0, 1);
584 count_vm_numa_events(NUMA_PTE_UPDATES
, nr_updated
);
589 static unsigned long change_prot_numa(struct vm_area_struct
*vma
,
590 unsigned long addr
, unsigned long end
)
594 #endif /* CONFIG_NUMA_BALANCING */
596 static int queue_pages_test_walk(unsigned long start
, unsigned long end
,
597 struct mm_walk
*walk
)
599 struct vm_area_struct
*vma
= walk
->vma
;
600 struct queue_pages
*qp
= walk
->private;
601 unsigned long endvma
= vma
->vm_end
;
602 unsigned long flags
= qp
->flags
;
604 if (!vma_migratable(vma
))
609 if (vma
->vm_start
> start
)
610 start
= vma
->vm_start
;
612 if (!(flags
& MPOL_MF_DISCONTIG_OK
)) {
613 if (!vma
->vm_next
&& vma
->vm_end
< end
)
615 if (qp
->prev
&& qp
->prev
->vm_end
< vma
->vm_start
)
621 if (flags
& MPOL_MF_LAZY
) {
622 /* Similar to task_numa_work, skip inaccessible VMAs */
623 if (!is_vm_hugetlb_page(vma
) &&
624 (vma
->vm_flags
& (VM_READ
| VM_EXEC
| VM_WRITE
)) &&
625 !(vma
->vm_flags
& VM_MIXEDMAP
))
626 change_prot_numa(vma
, start
, endvma
);
630 /* queue pages from current vma */
631 if (flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
))
637 * Walk through page tables and collect pages to be migrated.
639 * If pages found in a given range are on a set of nodes (determined by
640 * @nodes and @flags,) it's isolated and queued to the pagelist which is
641 * passed via @private.)
644 queue_pages_range(struct mm_struct
*mm
, unsigned long start
, unsigned long end
,
645 nodemask_t
*nodes
, unsigned long flags
,
646 struct list_head
*pagelist
)
648 struct queue_pages qp
= {
649 .pagelist
= pagelist
,
654 struct mm_walk queue_pages_walk
= {
655 .hugetlb_entry
= queue_pages_hugetlb
,
656 .pmd_entry
= queue_pages_pte_range
,
657 .test_walk
= queue_pages_test_walk
,
662 return walk_page_range(start
, end
, &queue_pages_walk
);
666 * Apply policy to a single VMA
667 * This must be called with the mmap_sem held for writing.
669 static int vma_replace_policy(struct vm_area_struct
*vma
,
670 struct mempolicy
*pol
)
673 struct mempolicy
*old
;
674 struct mempolicy
*new;
676 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
677 vma
->vm_start
, vma
->vm_end
, vma
->vm_pgoff
,
678 vma
->vm_ops
, vma
->vm_file
,
679 vma
->vm_ops
? vma
->vm_ops
->set_policy
: NULL
);
685 if (vma
->vm_ops
&& vma
->vm_ops
->set_policy
) {
686 err
= vma
->vm_ops
->set_policy(vma
, new);
691 old
= vma
->vm_policy
;
692 vma
->vm_policy
= new; /* protected by mmap_sem */
701 /* Step 2: apply policy to a range and do splits. */
702 static int mbind_range(struct mm_struct
*mm
, unsigned long start
,
703 unsigned long end
, struct mempolicy
*new_pol
)
705 struct vm_area_struct
*next
;
706 struct vm_area_struct
*prev
;
707 struct vm_area_struct
*vma
;
710 unsigned long vmstart
;
713 vma
= find_vma(mm
, start
);
714 if (!vma
|| vma
->vm_start
> start
)
718 if (start
> vma
->vm_start
)
721 for (; vma
&& vma
->vm_start
< end
; prev
= vma
, vma
= next
) {
723 vmstart
= max(start
, vma
->vm_start
);
724 vmend
= min(end
, vma
->vm_end
);
726 if (mpol_equal(vma_policy(vma
), new_pol
))
729 pgoff
= vma
->vm_pgoff
+
730 ((vmstart
- vma
->vm_start
) >> PAGE_SHIFT
);
731 prev
= vma_merge(mm
, prev
, vmstart
, vmend
, vma
->vm_flags
,
732 vma
->anon_vma
, vma
->vm_file
, pgoff
,
733 new_pol
, vma
->vm_userfaultfd_ctx
);
737 if (mpol_equal(vma_policy(vma
), new_pol
))
739 /* vma_merge() joined vma && vma->next, case 8 */
742 if (vma
->vm_start
!= vmstart
) {
743 err
= split_vma(vma
->vm_mm
, vma
, vmstart
, 1);
747 if (vma
->vm_end
!= vmend
) {
748 err
= split_vma(vma
->vm_mm
, vma
, vmend
, 0);
753 err
= vma_replace_policy(vma
, new_pol
);
762 /* Set the process memory policy */
763 static long do_set_mempolicy(unsigned short mode
, unsigned short flags
,
766 struct mempolicy
*new, *old
;
767 NODEMASK_SCRATCH(scratch
);
773 new = mpol_new(mode
, flags
, nodes
);
780 ret
= mpol_set_nodemask(new, nodes
, scratch
);
782 task_unlock(current
);
786 old
= current
->mempolicy
;
787 current
->mempolicy
= new;
788 if (new && new->mode
== MPOL_INTERLEAVE
)
789 current
->il_prev
= MAX_NUMNODES
-1;
790 task_unlock(current
);
794 NODEMASK_SCRATCH_FREE(scratch
);
799 * Return nodemask for policy for get_mempolicy() query
801 * Called with task's alloc_lock held
803 static void get_policy_nodemask(struct mempolicy
*p
, nodemask_t
*nodes
)
806 if (p
== &default_policy
)
812 case MPOL_INTERLEAVE
:
816 if (!(p
->flags
& MPOL_F_LOCAL
))
817 node_set(p
->v
.preferred_node
, *nodes
);
818 /* else return empty node mask for local allocation */
825 static int lookup_node(unsigned long addr
)
830 err
= get_user_pages(addr
& PAGE_MASK
, 1, 0, &p
, NULL
);
832 err
= page_to_nid(p
);
838 /* Retrieve NUMA policy */
839 static long do_get_mempolicy(int *policy
, nodemask_t
*nmask
,
840 unsigned long addr
, unsigned long flags
)
843 struct mm_struct
*mm
= current
->mm
;
844 struct vm_area_struct
*vma
= NULL
;
845 struct mempolicy
*pol
= current
->mempolicy
;
848 ~(unsigned long)(MPOL_F_NODE
|MPOL_F_ADDR
|MPOL_F_MEMS_ALLOWED
))
851 if (flags
& MPOL_F_MEMS_ALLOWED
) {
852 if (flags
& (MPOL_F_NODE
|MPOL_F_ADDR
))
854 *policy
= 0; /* just so it's initialized */
856 *nmask
= cpuset_current_mems_allowed
;
857 task_unlock(current
);
861 if (flags
& MPOL_F_ADDR
) {
863 * Do NOT fall back to task policy if the
864 * vma/shared policy at addr is NULL. We
865 * want to return MPOL_DEFAULT in this case.
867 down_read(&mm
->mmap_sem
);
868 vma
= find_vma_intersection(mm
, addr
, addr
+1);
870 up_read(&mm
->mmap_sem
);
873 if (vma
->vm_ops
&& vma
->vm_ops
->get_policy
)
874 pol
= vma
->vm_ops
->get_policy(vma
, addr
);
876 pol
= vma
->vm_policy
;
881 pol
= &default_policy
; /* indicates default behavior */
883 if (flags
& MPOL_F_NODE
) {
884 if (flags
& MPOL_F_ADDR
) {
885 err
= lookup_node(addr
);
889 } else if (pol
== current
->mempolicy
&&
890 pol
->mode
== MPOL_INTERLEAVE
) {
891 *policy
= next_node_in(current
->il_prev
, pol
->v
.nodes
);
897 *policy
= pol
== &default_policy
? MPOL_DEFAULT
:
900 * Internal mempolicy flags must be masked off before exposing
901 * the policy to userspace.
903 *policy
|= (pol
->flags
& MPOL_MODE_FLAGS
);
908 if (mpol_store_user_nodemask(pol
)) {
909 *nmask
= pol
->w
.user_nodemask
;
912 get_policy_nodemask(pol
, nmask
);
913 task_unlock(current
);
920 up_read(¤t
->mm
->mmap_sem
);
924 #ifdef CONFIG_MIGRATION
926 * page migration, thp tail pages can be passed.
928 static void migrate_page_add(struct page
*page
, struct list_head
*pagelist
,
931 struct page
*head
= compound_head(page
);
933 * Avoid migrating a page that is shared with others.
935 if ((flags
& MPOL_MF_MOVE_ALL
) || page_mapcount(head
) == 1) {
936 if (!isolate_lru_page(head
)) {
937 list_add_tail(&head
->lru
, pagelist
);
938 mod_node_page_state(page_pgdat(head
),
939 NR_ISOLATED_ANON
+ page_is_file_cache(head
),
940 hpage_nr_pages(head
));
945 static struct page
*new_node_page(struct page
*page
, unsigned long node
, int **x
)
948 return alloc_huge_page_node(page_hstate(compound_head(page
)),
950 else if (thp_migration_supported() && PageTransHuge(page
)) {
953 thp
= alloc_pages_node(node
,
954 (GFP_TRANSHUGE
| __GFP_THISNODE
),
958 prep_transhuge_page(thp
);
961 return __alloc_pages_node(node
, GFP_HIGHUSER_MOVABLE
|
966 * Migrate pages from one node to a target node.
967 * Returns error or the number of pages not migrated.
969 static int migrate_to_node(struct mm_struct
*mm
, int source
, int dest
,
977 node_set(source
, nmask
);
980 * This does not "check" the range but isolates all pages that
981 * need migration. Between passing in the full user address
982 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
984 VM_BUG_ON(!(flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
)));
985 queue_pages_range(mm
, mm
->mmap
->vm_start
, mm
->task_size
, &nmask
,
986 flags
| MPOL_MF_DISCONTIG_OK
, &pagelist
);
988 if (!list_empty(&pagelist
)) {
989 err
= migrate_pages(&pagelist
, new_node_page
, NULL
, dest
,
990 MIGRATE_SYNC
, MR_SYSCALL
);
992 putback_movable_pages(&pagelist
);
999 * Move pages between the two nodesets so as to preserve the physical
1000 * layout as much as possible.
1002 * Returns the number of page that could not be moved.
1004 int do_migrate_pages(struct mm_struct
*mm
, const nodemask_t
*from
,
1005 const nodemask_t
*to
, int flags
)
1011 err
= migrate_prep();
1015 down_read(&mm
->mmap_sem
);
1018 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1019 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1020 * bit in 'tmp', and return that <source, dest> pair for migration.
1021 * The pair of nodemasks 'to' and 'from' define the map.
1023 * If no pair of bits is found that way, fallback to picking some
1024 * pair of 'source' and 'dest' bits that are not the same. If the
1025 * 'source' and 'dest' bits are the same, this represents a node
1026 * that will be migrating to itself, so no pages need move.
1028 * If no bits are left in 'tmp', or if all remaining bits left
1029 * in 'tmp' correspond to the same bit in 'to', return false
1030 * (nothing left to migrate).
1032 * This lets us pick a pair of nodes to migrate between, such that
1033 * if possible the dest node is not already occupied by some other
1034 * source node, minimizing the risk of overloading the memory on a
1035 * node that would happen if we migrated incoming memory to a node
1036 * before migrating outgoing memory source that same node.
1038 * A single scan of tmp is sufficient. As we go, we remember the
1039 * most recent <s, d> pair that moved (s != d). If we find a pair
1040 * that not only moved, but what's better, moved to an empty slot
1041 * (d is not set in tmp), then we break out then, with that pair.
1042 * Otherwise when we finish scanning from_tmp, we at least have the
1043 * most recent <s, d> pair that moved. If we get all the way through
1044 * the scan of tmp without finding any node that moved, much less
1045 * moved to an empty node, then there is nothing left worth migrating.
1049 while (!nodes_empty(tmp
)) {
1051 int source
= NUMA_NO_NODE
;
1054 for_each_node_mask(s
, tmp
) {
1057 * do_migrate_pages() tries to maintain the relative
1058 * node relationship of the pages established between
1059 * threads and memory areas.
1061 * However if the number of source nodes is not equal to
1062 * the number of destination nodes we can not preserve
1063 * this node relative relationship. In that case, skip
1064 * copying memory from a node that is in the destination
1067 * Example: [2,3,4] -> [3,4,5] moves everything.
1068 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1071 if ((nodes_weight(*from
) != nodes_weight(*to
)) &&
1072 (node_isset(s
, *to
)))
1075 d
= node_remap(s
, *from
, *to
);
1079 source
= s
; /* Node moved. Memorize */
1082 /* dest not in remaining from nodes? */
1083 if (!node_isset(dest
, tmp
))
1086 if (source
== NUMA_NO_NODE
)
1089 node_clear(source
, tmp
);
1090 err
= migrate_to_node(mm
, source
, dest
, flags
);
1096 up_read(&mm
->mmap_sem
);
1104 * Allocate a new page for page migration based on vma policy.
1105 * Start by assuming the page is mapped by the same vma as contains @start.
1106 * Search forward from there, if not. N.B., this assumes that the
1107 * list of pages handed to migrate_pages()--which is how we get here--
1108 * is in virtual address order.
1110 static struct page
*new_page(struct page
*page
, unsigned long start
, int **x
)
1112 struct vm_area_struct
*vma
;
1113 unsigned long uninitialized_var(address
);
1115 vma
= find_vma(current
->mm
, start
);
1117 address
= page_address_in_vma(page
, vma
);
1118 if (address
!= -EFAULT
)
1123 if (PageHuge(page
)) {
1125 return alloc_huge_page_noerr(vma
, address
, 1);
1126 } else if (thp_migration_supported() && PageTransHuge(page
)) {
1129 thp
= alloc_hugepage_vma(GFP_TRANSHUGE
, vma
, address
,
1133 prep_transhuge_page(thp
);
1137 * if !vma, alloc_page_vma() will use task or system default policy
1139 return alloc_page_vma(GFP_HIGHUSER_MOVABLE
| __GFP_RETRY_MAYFAIL
,
1144 static void migrate_page_add(struct page
*page
, struct list_head
*pagelist
,
1145 unsigned long flags
)
1149 int do_migrate_pages(struct mm_struct
*mm
, const nodemask_t
*from
,
1150 const nodemask_t
*to
, int flags
)
1155 static struct page
*new_page(struct page
*page
, unsigned long start
, int **x
)
1161 static long do_mbind(unsigned long start
, unsigned long len
,
1162 unsigned short mode
, unsigned short mode_flags
,
1163 nodemask_t
*nmask
, unsigned long flags
)
1165 struct mm_struct
*mm
= current
->mm
;
1166 struct mempolicy
*new;
1169 LIST_HEAD(pagelist
);
1171 if (flags
& ~(unsigned long)MPOL_MF_VALID
)
1173 if ((flags
& MPOL_MF_MOVE_ALL
) && !capable(CAP_SYS_NICE
))
1176 if (start
& ~PAGE_MASK
)
1179 if (mode
== MPOL_DEFAULT
)
1180 flags
&= ~MPOL_MF_STRICT
;
1182 len
= (len
+ PAGE_SIZE
- 1) & PAGE_MASK
;
1190 new = mpol_new(mode
, mode_flags
, nmask
);
1192 return PTR_ERR(new);
1194 if (flags
& MPOL_MF_LAZY
)
1195 new->flags
|= MPOL_F_MOF
;
1198 * If we are using the default policy then operation
1199 * on discontinuous address spaces is okay after all
1202 flags
|= MPOL_MF_DISCONTIG_OK
;
1204 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1205 start
, start
+ len
, mode
, mode_flags
,
1206 nmask
? nodes_addr(*nmask
)[0] : NUMA_NO_NODE
);
1208 if (flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
)) {
1210 err
= migrate_prep();
1215 NODEMASK_SCRATCH(scratch
);
1217 down_write(&mm
->mmap_sem
);
1219 err
= mpol_set_nodemask(new, nmask
, scratch
);
1220 task_unlock(current
);
1222 up_write(&mm
->mmap_sem
);
1225 NODEMASK_SCRATCH_FREE(scratch
);
1230 err
= queue_pages_range(mm
, start
, end
, nmask
,
1231 flags
| MPOL_MF_INVERT
, &pagelist
);
1233 err
= mbind_range(mm
, start
, end
, new);
1238 if (!list_empty(&pagelist
)) {
1239 WARN_ON_ONCE(flags
& MPOL_MF_LAZY
);
1240 nr_failed
= migrate_pages(&pagelist
, new_page
, NULL
,
1241 start
, MIGRATE_SYNC
, MR_MEMPOLICY_MBIND
);
1243 putback_movable_pages(&pagelist
);
1246 if (nr_failed
&& (flags
& MPOL_MF_STRICT
))
1249 putback_movable_pages(&pagelist
);
1251 up_write(&mm
->mmap_sem
);
1258 * User space interface with variable sized bitmaps for nodelists.
1261 /* Copy a node mask from user space. */
1262 static int get_nodes(nodemask_t
*nodes
, const unsigned long __user
*nmask
,
1263 unsigned long maxnode
)
1267 unsigned long nlongs
;
1268 unsigned long endmask
;
1271 nodes_clear(*nodes
);
1272 if (maxnode
== 0 || !nmask
)
1274 if (maxnode
> PAGE_SIZE
*BITS_PER_BYTE
)
1277 nlongs
= BITS_TO_LONGS(maxnode
);
1278 if ((maxnode
% BITS_PER_LONG
) == 0)
1281 endmask
= (1UL << (maxnode
% BITS_PER_LONG
)) - 1;
1284 * When the user specified more nodes than supported just check
1285 * if the non supported part is all zero.
1287 * If maxnode have more longs than MAX_NUMNODES, check
1288 * the bits in that area first. And then go through to
1289 * check the rest bits which equal or bigger than MAX_NUMNODES.
1290 * Otherwise, just check bits [MAX_NUMNODES, maxnode).
1292 if (nlongs
> BITS_TO_LONGS(MAX_NUMNODES
)) {
1293 if (nlongs
> PAGE_SIZE
/sizeof(long))
1295 for (k
= BITS_TO_LONGS(MAX_NUMNODES
); k
< nlongs
; k
++) {
1296 if (get_user(t
, nmask
+ k
))
1298 if (k
== nlongs
- 1) {
1304 nlongs
= BITS_TO_LONGS(MAX_NUMNODES
);
1308 if (maxnode
> MAX_NUMNODES
&& MAX_NUMNODES
% BITS_PER_LONG
!= 0) {
1309 unsigned long valid_mask
= endmask
;
1311 valid_mask
&= ~((1UL << (MAX_NUMNODES
% BITS_PER_LONG
)) - 1);
1312 if (get_user(t
, nmask
+ nlongs
- 1))
1318 if (copy_from_user(nodes_addr(*nodes
), nmask
, nlongs
*sizeof(unsigned long)))
1320 nodes_addr(*nodes
)[nlongs
-1] &= endmask
;
1324 /* Copy a kernel node mask to user space */
1325 static int copy_nodes_to_user(unsigned long __user
*mask
, unsigned long maxnode
,
1328 unsigned long copy
= ALIGN(maxnode
-1, 64) / 8;
1329 const int nbytes
= BITS_TO_LONGS(MAX_NUMNODES
) * sizeof(long);
1331 if (copy
> nbytes
) {
1332 if (copy
> PAGE_SIZE
)
1334 if (clear_user((char __user
*)mask
+ nbytes
, copy
- nbytes
))
1338 return copy_to_user(mask
, nodes_addr(*nodes
), copy
) ? -EFAULT
: 0;
1341 SYSCALL_DEFINE6(mbind
, unsigned long, start
, unsigned long, len
,
1342 unsigned long, mode
, const unsigned long __user
*, nmask
,
1343 unsigned long, maxnode
, unsigned, flags
)
1347 unsigned short mode_flags
;
1349 mode_flags
= mode
& MPOL_MODE_FLAGS
;
1350 mode
&= ~MPOL_MODE_FLAGS
;
1351 if (mode
>= MPOL_MAX
)
1353 if ((mode_flags
& MPOL_F_STATIC_NODES
) &&
1354 (mode_flags
& MPOL_F_RELATIVE_NODES
))
1356 err
= get_nodes(&nodes
, nmask
, maxnode
);
1359 return do_mbind(start
, len
, mode
, mode_flags
, &nodes
, flags
);
1362 /* Set the process memory policy */
1363 SYSCALL_DEFINE3(set_mempolicy
, int, mode
, const unsigned long __user
*, nmask
,
1364 unsigned long, maxnode
)
1368 unsigned short flags
;
1370 flags
= mode
& MPOL_MODE_FLAGS
;
1371 mode
&= ~MPOL_MODE_FLAGS
;
1372 if ((unsigned int)mode
>= MPOL_MAX
)
1374 if ((flags
& MPOL_F_STATIC_NODES
) && (flags
& MPOL_F_RELATIVE_NODES
))
1376 err
= get_nodes(&nodes
, nmask
, maxnode
);
1379 return do_set_mempolicy(mode
, flags
, &nodes
);
1382 SYSCALL_DEFINE4(migrate_pages
, pid_t
, pid
, unsigned long, maxnode
,
1383 const unsigned long __user
*, old_nodes
,
1384 const unsigned long __user
*, new_nodes
)
1386 struct mm_struct
*mm
= NULL
;
1387 struct task_struct
*task
;
1388 nodemask_t task_nodes
;
1392 NODEMASK_SCRATCH(scratch
);
1397 old
= &scratch
->mask1
;
1398 new = &scratch
->mask2
;
1400 err
= get_nodes(old
, old_nodes
, maxnode
);
1404 err
= get_nodes(new, new_nodes
, maxnode
);
1408 /* Find the mm_struct */
1410 task
= pid
? find_task_by_vpid(pid
) : current
;
1416 get_task_struct(task
);
1421 * Check if this process has the right to modify the specified process.
1422 * Use the regular "ptrace_may_access()" checks.
1424 if (!ptrace_may_access(task
, PTRACE_MODE_READ_REALCREDS
)) {
1431 task_nodes
= cpuset_mems_allowed(task
);
1432 /* Is the user allowed to access the target nodes? */
1433 if (!nodes_subset(*new, task_nodes
) && !capable(CAP_SYS_NICE
)) {
1438 task_nodes
= cpuset_mems_allowed(current
);
1439 nodes_and(*new, *new, task_nodes
);
1440 if (nodes_empty(*new))
1443 nodes_and(*new, *new, node_states
[N_MEMORY
]);
1444 if (nodes_empty(*new))
1447 err
= security_task_movememory(task
);
1451 mm
= get_task_mm(task
);
1452 put_task_struct(task
);
1459 err
= do_migrate_pages(mm
, old
, new,
1460 capable(CAP_SYS_NICE
) ? MPOL_MF_MOVE_ALL
: MPOL_MF_MOVE
);
1464 NODEMASK_SCRATCH_FREE(scratch
);
1469 put_task_struct(task
);
1475 /* Retrieve NUMA policy */
1476 SYSCALL_DEFINE5(get_mempolicy
, int __user
*, policy
,
1477 unsigned long __user
*, nmask
, unsigned long, maxnode
,
1478 unsigned long, addr
, unsigned long, flags
)
1481 int uninitialized_var(pval
);
1484 if (nmask
!= NULL
&& maxnode
< MAX_NUMNODES
)
1487 err
= do_get_mempolicy(&pval
, &nodes
, addr
, flags
);
1492 if (policy
&& put_user(pval
, policy
))
1496 err
= copy_nodes_to_user(nmask
, maxnode
, &nodes
);
1501 #ifdef CONFIG_COMPAT
1503 COMPAT_SYSCALL_DEFINE5(get_mempolicy
, int __user
*, policy
,
1504 compat_ulong_t __user
*, nmask
,
1505 compat_ulong_t
, maxnode
,
1506 compat_ulong_t
, addr
, compat_ulong_t
, flags
)
1509 unsigned long __user
*nm
= NULL
;
1510 unsigned long nr_bits
, alloc_size
;
1511 DECLARE_BITMAP(bm
, MAX_NUMNODES
);
1513 nr_bits
= min_t(unsigned long, maxnode
-1, MAX_NUMNODES
);
1514 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
1517 nm
= compat_alloc_user_space(alloc_size
);
1519 err
= sys_get_mempolicy(policy
, nm
, nr_bits
+1, addr
, flags
);
1521 if (!err
&& nmask
) {
1522 unsigned long copy_size
;
1523 copy_size
= min_t(unsigned long, sizeof(bm
), alloc_size
);
1524 err
= copy_from_user(bm
, nm
, copy_size
);
1525 /* ensure entire bitmap is zeroed */
1526 err
|= clear_user(nmask
, ALIGN(maxnode
-1, 8) / 8);
1527 err
|= compat_put_bitmap(nmask
, bm
, nr_bits
);
1533 COMPAT_SYSCALL_DEFINE3(set_mempolicy
, int, mode
, compat_ulong_t __user
*, nmask
,
1534 compat_ulong_t
, maxnode
)
1536 unsigned long __user
*nm
= NULL
;
1537 unsigned long nr_bits
, alloc_size
;
1538 DECLARE_BITMAP(bm
, MAX_NUMNODES
);
1540 nr_bits
= min_t(unsigned long, maxnode
-1, MAX_NUMNODES
);
1541 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
1544 if (compat_get_bitmap(bm
, nmask
, nr_bits
))
1546 nm
= compat_alloc_user_space(alloc_size
);
1547 if (copy_to_user(nm
, bm
, alloc_size
))
1551 return sys_set_mempolicy(mode
, nm
, nr_bits
+1);
1554 COMPAT_SYSCALL_DEFINE6(mbind
, compat_ulong_t
, start
, compat_ulong_t
, len
,
1555 compat_ulong_t
, mode
, compat_ulong_t __user
*, nmask
,
1556 compat_ulong_t
, maxnode
, compat_ulong_t
, flags
)
1558 unsigned long __user
*nm
= NULL
;
1559 unsigned long nr_bits
, alloc_size
;
1562 nr_bits
= min_t(unsigned long, maxnode
-1, MAX_NUMNODES
);
1563 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
1566 if (compat_get_bitmap(nodes_addr(bm
), nmask
, nr_bits
))
1568 nm
= compat_alloc_user_space(alloc_size
);
1569 if (copy_to_user(nm
, nodes_addr(bm
), alloc_size
))
1573 return sys_mbind(start
, len
, mode
, nm
, nr_bits
+1, flags
);
1578 struct mempolicy
*__get_vma_policy(struct vm_area_struct
*vma
,
1581 struct mempolicy
*pol
= NULL
;
1584 if (vma
->vm_ops
&& vma
->vm_ops
->get_policy
) {
1585 pol
= vma
->vm_ops
->get_policy(vma
, addr
);
1586 } else if (vma
->vm_policy
) {
1587 pol
= vma
->vm_policy
;
1590 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1591 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1592 * count on these policies which will be dropped by
1593 * mpol_cond_put() later
1595 if (mpol_needs_cond_ref(pol
))
1604 * get_vma_policy(@vma, @addr)
1605 * @vma: virtual memory area whose policy is sought
1606 * @addr: address in @vma for shared policy lookup
1608 * Returns effective policy for a VMA at specified address.
1609 * Falls back to current->mempolicy or system default policy, as necessary.
1610 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1611 * count--added by the get_policy() vm_op, as appropriate--to protect against
1612 * freeing by another task. It is the caller's responsibility to free the
1613 * extra reference for shared policies.
1615 static struct mempolicy
*get_vma_policy(struct vm_area_struct
*vma
,
1618 struct mempolicy
*pol
= __get_vma_policy(vma
, addr
);
1621 pol
= get_task_policy(current
);
1626 bool vma_policy_mof(struct vm_area_struct
*vma
)
1628 struct mempolicy
*pol
;
1630 if (vma
->vm_ops
&& vma
->vm_ops
->get_policy
) {
1633 pol
= vma
->vm_ops
->get_policy(vma
, vma
->vm_start
);
1634 if (pol
&& (pol
->flags
& MPOL_F_MOF
))
1641 pol
= vma
->vm_policy
;
1643 pol
= get_task_policy(current
);
1645 return pol
->flags
& MPOL_F_MOF
;
1648 static int apply_policy_zone(struct mempolicy
*policy
, enum zone_type zone
)
1650 enum zone_type dynamic_policy_zone
= policy_zone
;
1652 BUG_ON(dynamic_policy_zone
== ZONE_MOVABLE
);
1655 * if policy->v.nodes has movable memory only,
1656 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1658 * policy->v.nodes is intersect with node_states[N_MEMORY].
1659 * so if the following test faile, it implies
1660 * policy->v.nodes has movable memory only.
1662 if (!nodes_intersects(policy
->v
.nodes
, node_states
[N_HIGH_MEMORY
]))
1663 dynamic_policy_zone
= ZONE_MOVABLE
;
1665 return zone
>= dynamic_policy_zone
;
1669 * Return a nodemask representing a mempolicy for filtering nodes for
1672 static nodemask_t
*policy_nodemask(gfp_t gfp
, struct mempolicy
*policy
)
1674 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1675 if (unlikely(policy
->mode
== MPOL_BIND
) &&
1676 apply_policy_zone(policy
, gfp_zone(gfp
)) &&
1677 cpuset_nodemask_valid_mems_allowed(&policy
->v
.nodes
))
1678 return &policy
->v
.nodes
;
1683 /* Return the node id preferred by the given mempolicy, or the given id */
1684 static int policy_node(gfp_t gfp
, struct mempolicy
*policy
,
1687 if (policy
->mode
== MPOL_PREFERRED
&& !(policy
->flags
& MPOL_F_LOCAL
))
1688 nd
= policy
->v
.preferred_node
;
1691 * __GFP_THISNODE shouldn't even be used with the bind policy
1692 * because we might easily break the expectation to stay on the
1693 * requested node and not break the policy.
1695 WARN_ON_ONCE(policy
->mode
== MPOL_BIND
&& (gfp
& __GFP_THISNODE
));
1701 /* Do dynamic interleaving for a process */
1702 static unsigned interleave_nodes(struct mempolicy
*policy
)
1705 struct task_struct
*me
= current
;
1707 next
= next_node_in(me
->il_prev
, policy
->v
.nodes
);
1708 if (next
< MAX_NUMNODES
)
1714 * Depending on the memory policy provide a node from which to allocate the
1717 unsigned int mempolicy_slab_node(void)
1719 struct mempolicy
*policy
;
1720 int node
= numa_mem_id();
1725 policy
= current
->mempolicy
;
1726 if (!policy
|| policy
->flags
& MPOL_F_LOCAL
)
1729 switch (policy
->mode
) {
1730 case MPOL_PREFERRED
:
1732 * handled MPOL_F_LOCAL above
1734 return policy
->v
.preferred_node
;
1736 case MPOL_INTERLEAVE
:
1737 return interleave_nodes(policy
);
1743 * Follow bind policy behavior and start allocation at the
1746 struct zonelist
*zonelist
;
1747 enum zone_type highest_zoneidx
= gfp_zone(GFP_KERNEL
);
1748 zonelist
= &NODE_DATA(node
)->node_zonelists
[ZONELIST_FALLBACK
];
1749 z
= first_zones_zonelist(zonelist
, highest_zoneidx
,
1751 return z
->zone
? z
->zone
->node
: node
;
1760 * Do static interleaving for a VMA with known offset @n. Returns the n'th
1761 * node in pol->v.nodes (starting from n=0), wrapping around if n exceeds the
1762 * number of present nodes.
1764 static unsigned offset_il_node(struct mempolicy
*pol
, unsigned long n
)
1766 unsigned nnodes
= nodes_weight(pol
->v
.nodes
);
1772 return numa_node_id();
1773 target
= (unsigned int)n
% nnodes
;
1774 nid
= first_node(pol
->v
.nodes
);
1775 for (i
= 0; i
< target
; i
++)
1776 nid
= next_node(nid
, pol
->v
.nodes
);
1780 /* Determine a node number for interleave */
1781 static inline unsigned interleave_nid(struct mempolicy
*pol
,
1782 struct vm_area_struct
*vma
, unsigned long addr
, int shift
)
1788 * for small pages, there is no difference between
1789 * shift and PAGE_SHIFT, so the bit-shift is safe.
1790 * for huge pages, since vm_pgoff is in units of small
1791 * pages, we need to shift off the always 0 bits to get
1794 BUG_ON(shift
< PAGE_SHIFT
);
1795 off
= vma
->vm_pgoff
>> (shift
- PAGE_SHIFT
);
1796 off
+= (addr
- vma
->vm_start
) >> shift
;
1797 return offset_il_node(pol
, off
);
1799 return interleave_nodes(pol
);
1802 #ifdef CONFIG_HUGETLBFS
1804 * huge_node(@vma, @addr, @gfp_flags, @mpol)
1805 * @vma: virtual memory area whose policy is sought
1806 * @addr: address in @vma for shared policy lookup and interleave policy
1807 * @gfp_flags: for requested zone
1808 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
1809 * @nodemask: pointer to nodemask pointer for MPOL_BIND nodemask
1811 * Returns a nid suitable for a huge page allocation and a pointer
1812 * to the struct mempolicy for conditional unref after allocation.
1813 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1814 * @nodemask for filtering the zonelist.
1816 * Must be protected by read_mems_allowed_begin()
1818 int huge_node(struct vm_area_struct
*vma
, unsigned long addr
, gfp_t gfp_flags
,
1819 struct mempolicy
**mpol
, nodemask_t
**nodemask
)
1823 *mpol
= get_vma_policy(vma
, addr
);
1824 *nodemask
= NULL
; /* assume !MPOL_BIND */
1826 if (unlikely((*mpol
)->mode
== MPOL_INTERLEAVE
)) {
1827 nid
= interleave_nid(*mpol
, vma
, addr
,
1828 huge_page_shift(hstate_vma(vma
)));
1830 nid
= policy_node(gfp_flags
, *mpol
, numa_node_id());
1831 if ((*mpol
)->mode
== MPOL_BIND
)
1832 *nodemask
= &(*mpol
)->v
.nodes
;
1838 * init_nodemask_of_mempolicy
1840 * If the current task's mempolicy is "default" [NULL], return 'false'
1841 * to indicate default policy. Otherwise, extract the policy nodemask
1842 * for 'bind' or 'interleave' policy into the argument nodemask, or
1843 * initialize the argument nodemask to contain the single node for
1844 * 'preferred' or 'local' policy and return 'true' to indicate presence
1845 * of non-default mempolicy.
1847 * We don't bother with reference counting the mempolicy [mpol_get/put]
1848 * because the current task is examining it's own mempolicy and a task's
1849 * mempolicy is only ever changed by the task itself.
1851 * N.B., it is the caller's responsibility to free a returned nodemask.
1853 bool init_nodemask_of_mempolicy(nodemask_t
*mask
)
1855 struct mempolicy
*mempolicy
;
1858 if (!(mask
&& current
->mempolicy
))
1862 mempolicy
= current
->mempolicy
;
1863 switch (mempolicy
->mode
) {
1864 case MPOL_PREFERRED
:
1865 if (mempolicy
->flags
& MPOL_F_LOCAL
)
1866 nid
= numa_node_id();
1868 nid
= mempolicy
->v
.preferred_node
;
1869 init_nodemask_of_node(mask
, nid
);
1874 case MPOL_INTERLEAVE
:
1875 *mask
= mempolicy
->v
.nodes
;
1881 task_unlock(current
);
1888 * mempolicy_nodemask_intersects
1890 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1891 * policy. Otherwise, check for intersection between mask and the policy
1892 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1893 * policy, always return true since it may allocate elsewhere on fallback.
1895 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1897 bool mempolicy_nodemask_intersects(struct task_struct
*tsk
,
1898 const nodemask_t
*mask
)
1900 struct mempolicy
*mempolicy
;
1906 mempolicy
= tsk
->mempolicy
;
1910 switch (mempolicy
->mode
) {
1911 case MPOL_PREFERRED
:
1913 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1914 * allocate from, they may fallback to other nodes when oom.
1915 * Thus, it's possible for tsk to have allocated memory from
1920 case MPOL_INTERLEAVE
:
1921 ret
= nodes_intersects(mempolicy
->v
.nodes
, *mask
);
1931 /* Allocate a page in interleaved policy.
1932 Own path because it needs to do special accounting. */
1933 static struct page
*alloc_page_interleave(gfp_t gfp
, unsigned order
,
1938 page
= __alloc_pages(gfp
, order
, nid
);
1939 /* skip NUMA_INTERLEAVE_HIT counter update if numa stats is disabled */
1940 if (!static_branch_likely(&vm_numa_stat_key
))
1942 if (page
&& page_to_nid(page
) == nid
) {
1944 __inc_numa_state(page_zone(page
), NUMA_INTERLEAVE_HIT
);
1951 * alloc_pages_vma - Allocate a page for a VMA.
1954 * %GFP_USER user allocation.
1955 * %GFP_KERNEL kernel allocations,
1956 * %GFP_HIGHMEM highmem/user allocations,
1957 * %GFP_FS allocation should not call back into a file system.
1958 * %GFP_ATOMIC don't sleep.
1960 * @order:Order of the GFP allocation.
1961 * @vma: Pointer to VMA or NULL if not available.
1962 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1963 * @node: Which node to prefer for allocation (modulo policy).
1964 * @hugepage: for hugepages try only the preferred node if possible
1966 * This function allocates a page from the kernel page pool and applies
1967 * a NUMA policy associated with the VMA or the current process.
1968 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1969 * mm_struct of the VMA to prevent it from going away. Should be used for
1970 * all allocations for pages that will be mapped into user space. Returns
1971 * NULL when no page can be allocated.
1974 alloc_pages_vma(gfp_t gfp
, int order
, struct vm_area_struct
*vma
,
1975 unsigned long addr
, int node
, bool hugepage
)
1977 struct mempolicy
*pol
;
1982 pol
= get_vma_policy(vma
, addr
);
1984 if (pol
->mode
== MPOL_INTERLEAVE
) {
1987 nid
= interleave_nid(pol
, vma
, addr
, PAGE_SHIFT
+ order
);
1989 page
= alloc_page_interleave(gfp
, order
, nid
);
1993 if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE
) && hugepage
)) {
1994 int hpage_node
= node
;
1997 * For hugepage allocation and non-interleave policy which
1998 * allows the current node (or other explicitly preferred
1999 * node) we only try to allocate from the current/preferred
2000 * node and don't fall back to other nodes, as the cost of
2001 * remote accesses would likely offset THP benefits.
2003 * If the policy is interleave, or does not allow the current
2004 * node in its nodemask, we allocate the standard way.
2006 if (pol
->mode
== MPOL_PREFERRED
&&
2007 !(pol
->flags
& MPOL_F_LOCAL
))
2008 hpage_node
= pol
->v
.preferred_node
;
2010 nmask
= policy_nodemask(gfp
, pol
);
2011 if (!nmask
|| node_isset(hpage_node
, *nmask
)) {
2013 page
= __alloc_pages_node(hpage_node
,
2014 gfp
| __GFP_THISNODE
, order
);
2019 nmask
= policy_nodemask(gfp
, pol
);
2020 preferred_nid
= policy_node(gfp
, pol
, node
);
2021 page
= __alloc_pages_nodemask(gfp
, order
, preferred_nid
, nmask
);
2028 * alloc_pages_current - Allocate pages.
2031 * %GFP_USER user allocation,
2032 * %GFP_KERNEL kernel allocation,
2033 * %GFP_HIGHMEM highmem allocation,
2034 * %GFP_FS don't call back into a file system.
2035 * %GFP_ATOMIC don't sleep.
2036 * @order: Power of two of allocation size in pages. 0 is a single page.
2038 * Allocate a page from the kernel page pool. When not in
2039 * interrupt context and apply the current process NUMA policy.
2040 * Returns NULL when no page can be allocated.
2042 struct page
*alloc_pages_current(gfp_t gfp
, unsigned order
)
2044 struct mempolicy
*pol
= &default_policy
;
2047 if (!in_interrupt() && !(gfp
& __GFP_THISNODE
))
2048 pol
= get_task_policy(current
);
2051 * No reference counting needed for current->mempolicy
2052 * nor system default_policy
2054 if (pol
->mode
== MPOL_INTERLEAVE
)
2055 page
= alloc_page_interleave(gfp
, order
, interleave_nodes(pol
));
2057 page
= __alloc_pages_nodemask(gfp
, order
,
2058 policy_node(gfp
, pol
, numa_node_id()),
2059 policy_nodemask(gfp
, pol
));
2063 EXPORT_SYMBOL(alloc_pages_current
);
2065 int vma_dup_policy(struct vm_area_struct
*src
, struct vm_area_struct
*dst
)
2067 struct mempolicy
*pol
= mpol_dup(vma_policy(src
));
2070 return PTR_ERR(pol
);
2071 dst
->vm_policy
= pol
;
2076 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2077 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2078 * with the mems_allowed returned by cpuset_mems_allowed(). This
2079 * keeps mempolicies cpuset relative after its cpuset moves. See
2080 * further kernel/cpuset.c update_nodemask().
2082 * current's mempolicy may be rebinded by the other task(the task that changes
2083 * cpuset's mems), so we needn't do rebind work for current task.
2086 /* Slow path of a mempolicy duplicate */
2087 struct mempolicy
*__mpol_dup(struct mempolicy
*old
)
2089 struct mempolicy
*new = kmem_cache_alloc(policy_cache
, GFP_KERNEL
);
2092 return ERR_PTR(-ENOMEM
);
2094 /* task's mempolicy is protected by alloc_lock */
2095 if (old
== current
->mempolicy
) {
2098 task_unlock(current
);
2102 if (current_cpuset_is_being_rebound()) {
2103 nodemask_t mems
= cpuset_mems_allowed(current
);
2104 mpol_rebind_policy(new, &mems
);
2106 atomic_set(&new->refcnt
, 1);
2110 /* Slow path of a mempolicy comparison */
2111 bool __mpol_equal(struct mempolicy
*a
, struct mempolicy
*b
)
2115 if (a
->mode
!= b
->mode
)
2117 if (a
->flags
!= b
->flags
)
2119 if (mpol_store_user_nodemask(a
))
2120 if (!nodes_equal(a
->w
.user_nodemask
, b
->w
.user_nodemask
))
2126 case MPOL_INTERLEAVE
:
2127 return !!nodes_equal(a
->v
.nodes
, b
->v
.nodes
);
2128 case MPOL_PREFERRED
:
2129 /* a's ->flags is the same as b's */
2130 if (a
->flags
& MPOL_F_LOCAL
)
2132 return a
->v
.preferred_node
== b
->v
.preferred_node
;
2140 * Shared memory backing store policy support.
2142 * Remember policies even when nobody has shared memory mapped.
2143 * The policies are kept in Red-Black tree linked from the inode.
2144 * They are protected by the sp->lock rwlock, which should be held
2145 * for any accesses to the tree.
2149 * lookup first element intersecting start-end. Caller holds sp->lock for
2150 * reading or for writing
2152 static struct sp_node
*
2153 sp_lookup(struct shared_policy
*sp
, unsigned long start
, unsigned long end
)
2155 struct rb_node
*n
= sp
->root
.rb_node
;
2158 struct sp_node
*p
= rb_entry(n
, struct sp_node
, nd
);
2160 if (start
>= p
->end
)
2162 else if (end
<= p
->start
)
2170 struct sp_node
*w
= NULL
;
2171 struct rb_node
*prev
= rb_prev(n
);
2174 w
= rb_entry(prev
, struct sp_node
, nd
);
2175 if (w
->end
<= start
)
2179 return rb_entry(n
, struct sp_node
, nd
);
2183 * Insert a new shared policy into the list. Caller holds sp->lock for
2186 static void sp_insert(struct shared_policy
*sp
, struct sp_node
*new)
2188 struct rb_node
**p
= &sp
->root
.rb_node
;
2189 struct rb_node
*parent
= NULL
;
2194 nd
= rb_entry(parent
, struct sp_node
, nd
);
2195 if (new->start
< nd
->start
)
2197 else if (new->end
> nd
->end
)
2198 p
= &(*p
)->rb_right
;
2202 rb_link_node(&new->nd
, parent
, p
);
2203 rb_insert_color(&new->nd
, &sp
->root
);
2204 pr_debug("inserting %lx-%lx: %d\n", new->start
, new->end
,
2205 new->policy
? new->policy
->mode
: 0);
2208 /* Find shared policy intersecting idx */
2210 mpol_shared_policy_lookup(struct shared_policy
*sp
, unsigned long idx
)
2212 struct mempolicy
*pol
= NULL
;
2215 if (!sp
->root
.rb_node
)
2217 read_lock(&sp
->lock
);
2218 sn
= sp_lookup(sp
, idx
, idx
+1);
2220 mpol_get(sn
->policy
);
2223 read_unlock(&sp
->lock
);
2227 static void sp_free(struct sp_node
*n
)
2229 mpol_put(n
->policy
);
2230 kmem_cache_free(sn_cache
, n
);
2234 * mpol_misplaced - check whether current page node is valid in policy
2236 * @page: page to be checked
2237 * @vma: vm area where page mapped
2238 * @addr: virtual address where page mapped
2240 * Lookup current policy node id for vma,addr and "compare to" page's
2244 * -1 - not misplaced, page is in the right node
2245 * node - node id where the page should be
2247 * Policy determination "mimics" alloc_page_vma().
2248 * Called from fault path where we know the vma and faulting address.
2250 int mpol_misplaced(struct page
*page
, struct vm_area_struct
*vma
, unsigned long addr
)
2252 struct mempolicy
*pol
;
2254 int curnid
= page_to_nid(page
);
2255 unsigned long pgoff
;
2256 int thiscpu
= raw_smp_processor_id();
2257 int thisnid
= cpu_to_node(thiscpu
);
2261 pol
= get_vma_policy(vma
, addr
);
2262 if (!(pol
->flags
& MPOL_F_MOF
))
2265 switch (pol
->mode
) {
2266 case MPOL_INTERLEAVE
:
2267 pgoff
= vma
->vm_pgoff
;
2268 pgoff
+= (addr
- vma
->vm_start
) >> PAGE_SHIFT
;
2269 polnid
= offset_il_node(pol
, pgoff
);
2272 case MPOL_PREFERRED
:
2273 if (pol
->flags
& MPOL_F_LOCAL
)
2274 polnid
= numa_node_id();
2276 polnid
= pol
->v
.preferred_node
;
2282 * allows binding to multiple nodes.
2283 * use current page if in policy nodemask,
2284 * else select nearest allowed node, if any.
2285 * If no allowed nodes, use current [!misplaced].
2287 if (node_isset(curnid
, pol
->v
.nodes
))
2289 z
= first_zones_zonelist(
2290 node_zonelist(numa_node_id(), GFP_HIGHUSER
),
2291 gfp_zone(GFP_HIGHUSER
),
2293 polnid
= z
->zone
->node
;
2300 /* Migrate the page towards the node whose CPU is referencing it */
2301 if (pol
->flags
& MPOL_F_MORON
) {
2304 if (!should_numa_migrate_memory(current
, page
, curnid
, thiscpu
))
2308 if (curnid
!= polnid
)
2317 * Drop the (possibly final) reference to task->mempolicy. It needs to be
2318 * dropped after task->mempolicy is set to NULL so that any allocation done as
2319 * part of its kmem_cache_free(), such as by KASAN, doesn't reference a freed
2322 void mpol_put_task_policy(struct task_struct
*task
)
2324 struct mempolicy
*pol
;
2327 pol
= task
->mempolicy
;
2328 task
->mempolicy
= NULL
;
2333 static void sp_delete(struct shared_policy
*sp
, struct sp_node
*n
)
2335 pr_debug("deleting %lx-l%lx\n", n
->start
, n
->end
);
2336 rb_erase(&n
->nd
, &sp
->root
);
2340 static void sp_node_init(struct sp_node
*node
, unsigned long start
,
2341 unsigned long end
, struct mempolicy
*pol
)
2343 node
->start
= start
;
2348 static struct sp_node
*sp_alloc(unsigned long start
, unsigned long end
,
2349 struct mempolicy
*pol
)
2352 struct mempolicy
*newpol
;
2354 n
= kmem_cache_alloc(sn_cache
, GFP_KERNEL
);
2358 newpol
= mpol_dup(pol
);
2359 if (IS_ERR(newpol
)) {
2360 kmem_cache_free(sn_cache
, n
);
2363 newpol
->flags
|= MPOL_F_SHARED
;
2364 sp_node_init(n
, start
, end
, newpol
);
2369 /* Replace a policy range. */
2370 static int shared_policy_replace(struct shared_policy
*sp
, unsigned long start
,
2371 unsigned long end
, struct sp_node
*new)
2374 struct sp_node
*n_new
= NULL
;
2375 struct mempolicy
*mpol_new
= NULL
;
2379 write_lock(&sp
->lock
);
2380 n
= sp_lookup(sp
, start
, end
);
2381 /* Take care of old policies in the same range. */
2382 while (n
&& n
->start
< end
) {
2383 struct rb_node
*next
= rb_next(&n
->nd
);
2384 if (n
->start
>= start
) {
2390 /* Old policy spanning whole new range. */
2395 *mpol_new
= *n
->policy
;
2396 atomic_set(&mpol_new
->refcnt
, 1);
2397 sp_node_init(n_new
, end
, n
->end
, mpol_new
);
2399 sp_insert(sp
, n_new
);
2408 n
= rb_entry(next
, struct sp_node
, nd
);
2412 write_unlock(&sp
->lock
);
2419 kmem_cache_free(sn_cache
, n_new
);
2424 write_unlock(&sp
->lock
);
2426 n_new
= kmem_cache_alloc(sn_cache
, GFP_KERNEL
);
2429 mpol_new
= kmem_cache_alloc(policy_cache
, GFP_KERNEL
);
2436 * mpol_shared_policy_init - initialize shared policy for inode
2437 * @sp: pointer to inode shared policy
2438 * @mpol: struct mempolicy to install
2440 * Install non-NULL @mpol in inode's shared policy rb-tree.
2441 * On entry, the current task has a reference on a non-NULL @mpol.
2442 * This must be released on exit.
2443 * This is called at get_inode() calls and we can use GFP_KERNEL.
2445 void mpol_shared_policy_init(struct shared_policy
*sp
, struct mempolicy
*mpol
)
2449 sp
->root
= RB_ROOT
; /* empty tree == default mempolicy */
2450 rwlock_init(&sp
->lock
);
2453 struct vm_area_struct pvma
;
2454 struct mempolicy
*new;
2455 NODEMASK_SCRATCH(scratch
);
2459 /* contextualize the tmpfs mount point mempolicy */
2460 new = mpol_new(mpol
->mode
, mpol
->flags
, &mpol
->w
.user_nodemask
);
2462 goto free_scratch
; /* no valid nodemask intersection */
2465 ret
= mpol_set_nodemask(new, &mpol
->w
.user_nodemask
, scratch
);
2466 task_unlock(current
);
2470 /* Create pseudo-vma that contains just the policy */
2471 memset(&pvma
, 0, sizeof(struct vm_area_struct
));
2472 pvma
.vm_end
= TASK_SIZE
; /* policy covers entire file */
2473 mpol_set_shared_policy(sp
, &pvma
, new); /* adds ref */
2476 mpol_put(new); /* drop initial ref */
2478 NODEMASK_SCRATCH_FREE(scratch
);
2480 mpol_put(mpol
); /* drop our incoming ref on sb mpol */
2484 int mpol_set_shared_policy(struct shared_policy
*info
,
2485 struct vm_area_struct
*vma
, struct mempolicy
*npol
)
2488 struct sp_node
*new = NULL
;
2489 unsigned long sz
= vma_pages(vma
);
2491 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2493 sz
, npol
? npol
->mode
: -1,
2494 npol
? npol
->flags
: -1,
2495 npol
? nodes_addr(npol
->v
.nodes
)[0] : NUMA_NO_NODE
);
2498 new = sp_alloc(vma
->vm_pgoff
, vma
->vm_pgoff
+ sz
, npol
);
2502 err
= shared_policy_replace(info
, vma
->vm_pgoff
, vma
->vm_pgoff
+sz
, new);
2508 /* Free a backing policy store on inode delete. */
2509 void mpol_free_shared_policy(struct shared_policy
*p
)
2512 struct rb_node
*next
;
2514 if (!p
->root
.rb_node
)
2516 write_lock(&p
->lock
);
2517 next
= rb_first(&p
->root
);
2519 n
= rb_entry(next
, struct sp_node
, nd
);
2520 next
= rb_next(&n
->nd
);
2523 write_unlock(&p
->lock
);
2526 #ifdef CONFIG_NUMA_BALANCING
2527 static int __initdata numabalancing_override
;
2529 static void __init
check_numabalancing_enable(void)
2531 bool numabalancing_default
= false;
2533 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED
))
2534 numabalancing_default
= true;
2536 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2537 if (numabalancing_override
)
2538 set_numabalancing_state(numabalancing_override
== 1);
2540 if (num_online_nodes() > 1 && !numabalancing_override
) {
2541 pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
2542 numabalancing_default
? "Enabling" : "Disabling");
2543 set_numabalancing_state(numabalancing_default
);
2547 static int __init
setup_numabalancing(char *str
)
2553 if (!strcmp(str
, "enable")) {
2554 numabalancing_override
= 1;
2556 } else if (!strcmp(str
, "disable")) {
2557 numabalancing_override
= -1;
2562 pr_warn("Unable to parse numa_balancing=\n");
2566 __setup("numa_balancing=", setup_numabalancing
);
2568 static inline void __init
check_numabalancing_enable(void)
2571 #endif /* CONFIG_NUMA_BALANCING */
2573 /* assumes fs == KERNEL_DS */
2574 void __init
numa_policy_init(void)
2576 nodemask_t interleave_nodes
;
2577 unsigned long largest
= 0;
2578 int nid
, prefer
= 0;
2580 policy_cache
= kmem_cache_create("numa_policy",
2581 sizeof(struct mempolicy
),
2582 0, SLAB_PANIC
, NULL
);
2584 sn_cache
= kmem_cache_create("shared_policy_node",
2585 sizeof(struct sp_node
),
2586 0, SLAB_PANIC
, NULL
);
2588 for_each_node(nid
) {
2589 preferred_node_policy
[nid
] = (struct mempolicy
) {
2590 .refcnt
= ATOMIC_INIT(1),
2591 .mode
= MPOL_PREFERRED
,
2592 .flags
= MPOL_F_MOF
| MPOL_F_MORON
,
2593 .v
= { .preferred_node
= nid
, },
2598 * Set interleaving policy for system init. Interleaving is only
2599 * enabled across suitably sized nodes (default is >= 16MB), or
2600 * fall back to the largest node if they're all smaller.
2602 nodes_clear(interleave_nodes
);
2603 for_each_node_state(nid
, N_MEMORY
) {
2604 unsigned long total_pages
= node_present_pages(nid
);
2606 /* Preserve the largest node */
2607 if (largest
< total_pages
) {
2608 largest
= total_pages
;
2612 /* Interleave this node? */
2613 if ((total_pages
<< PAGE_SHIFT
) >= (16 << 20))
2614 node_set(nid
, interleave_nodes
);
2617 /* All too small, use the largest */
2618 if (unlikely(nodes_empty(interleave_nodes
)))
2619 node_set(prefer
, interleave_nodes
);
2621 if (do_set_mempolicy(MPOL_INTERLEAVE
, 0, &interleave_nodes
))
2622 pr_err("%s: interleaving failed\n", __func__
);
2624 check_numabalancing_enable();
2627 /* Reset policy of current process to default */
2628 void numa_default_policy(void)
2630 do_set_mempolicy(MPOL_DEFAULT
, 0, NULL
);
2634 * Parse and format mempolicy from/to strings
2638 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
2640 static const char * const policy_modes
[] =
2642 [MPOL_DEFAULT
] = "default",
2643 [MPOL_PREFERRED
] = "prefer",
2644 [MPOL_BIND
] = "bind",
2645 [MPOL_INTERLEAVE
] = "interleave",
2646 [MPOL_LOCAL
] = "local",
2652 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2653 * @str: string containing mempolicy to parse
2654 * @mpol: pointer to struct mempolicy pointer, returned on success.
2657 * <mode>[=<flags>][:<nodelist>]
2659 * On success, returns 0, else 1
2661 int mpol_parse_str(char *str
, struct mempolicy
**mpol
)
2663 struct mempolicy
*new = NULL
;
2664 unsigned short mode
;
2665 unsigned short mode_flags
;
2667 char *nodelist
= strchr(str
, ':');
2668 char *flags
= strchr(str
, '=');
2672 /* NUL-terminate mode or flags string */
2674 if (nodelist_parse(nodelist
, nodes
))
2676 if (!nodes_subset(nodes
, node_states
[N_MEMORY
]))
2682 *flags
++ = '\0'; /* terminate mode string */
2684 for (mode
= 0; mode
< MPOL_MAX
; mode
++) {
2685 if (!strcmp(str
, policy_modes
[mode
])) {
2689 if (mode
>= MPOL_MAX
)
2693 case MPOL_PREFERRED
:
2695 * Insist on a nodelist of one node only
2698 char *rest
= nodelist
;
2699 while (isdigit(*rest
))
2705 case MPOL_INTERLEAVE
:
2707 * Default to online nodes with memory if no nodelist
2710 nodes
= node_states
[N_MEMORY
];
2714 * Don't allow a nodelist; mpol_new() checks flags
2718 mode
= MPOL_PREFERRED
;
2722 * Insist on a empty nodelist
2729 * Insist on a nodelist
2738 * Currently, we only support two mutually exclusive
2741 if (!strcmp(flags
, "static"))
2742 mode_flags
|= MPOL_F_STATIC_NODES
;
2743 else if (!strcmp(flags
, "relative"))
2744 mode_flags
|= MPOL_F_RELATIVE_NODES
;
2749 new = mpol_new(mode
, mode_flags
, &nodes
);
2754 * Save nodes for mpol_to_str() to show the tmpfs mount options
2755 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2757 if (mode
!= MPOL_PREFERRED
)
2758 new->v
.nodes
= nodes
;
2760 new->v
.preferred_node
= first_node(nodes
);
2762 new->flags
|= MPOL_F_LOCAL
;
2765 * Save nodes for contextualization: this will be used to "clone"
2766 * the mempolicy in a specific context [cpuset] at a later time.
2768 new->w
.user_nodemask
= nodes
;
2773 /* Restore string for error message */
2782 #endif /* CONFIG_TMPFS */
2785 * mpol_to_str - format a mempolicy structure for printing
2786 * @buffer: to contain formatted mempolicy string
2787 * @maxlen: length of @buffer
2788 * @pol: pointer to mempolicy to be formatted
2790 * Convert @pol into a string. If @buffer is too short, truncate the string.
2791 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
2792 * longest flag, "relative", and to display at least a few node ids.
2794 void mpol_to_str(char *buffer
, int maxlen
, struct mempolicy
*pol
)
2797 nodemask_t nodes
= NODE_MASK_NONE
;
2798 unsigned short mode
= MPOL_DEFAULT
;
2799 unsigned short flags
= 0;
2801 if (pol
&& pol
!= &default_policy
&& !(pol
->flags
& MPOL_F_MORON
)) {
2809 case MPOL_PREFERRED
:
2810 if (flags
& MPOL_F_LOCAL
)
2813 node_set(pol
->v
.preferred_node
, nodes
);
2816 case MPOL_INTERLEAVE
:
2817 nodes
= pol
->v
.nodes
;
2821 snprintf(p
, maxlen
, "unknown");
2825 p
+= snprintf(p
, maxlen
, "%s", policy_modes
[mode
]);
2827 if (flags
& MPOL_MODE_FLAGS
) {
2828 p
+= snprintf(p
, buffer
+ maxlen
- p
, "=");
2831 * Currently, the only defined flags are mutually exclusive
2833 if (flags
& MPOL_F_STATIC_NODES
)
2834 p
+= snprintf(p
, buffer
+ maxlen
- p
, "static");
2835 else if (flags
& MPOL_F_RELATIVE_NODES
)
2836 p
+= snprintf(p
, buffer
+ maxlen
- p
, "relative");
2839 if (!nodes_empty(nodes
))
2840 p
+= scnprintf(p
, buffer
+ maxlen
- p
, ":%*pbl",
2841 nodemask_pr_args(&nodes
));