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 int 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 four possible return values:
433 * 0 - pages are placed on the right node or queued successfully.
434 * 1 - there is unmovable page, and MPOL_MF_MOVE* & MPOL_MF_STRICT were
437 * -EIO - is migration entry or only MPOL_MF_STRICT was specified and an
438 * existing page was already on a node that does not follow the
441 static int queue_pages_pmd(pmd_t
*pmd
, spinlock_t
*ptl
, unsigned long addr
,
442 unsigned long end
, struct mm_walk
*walk
)
446 struct queue_pages
*qp
= walk
->private;
449 if (unlikely(is_pmd_migration_entry(*pmd
))) {
453 page
= pmd_page(*pmd
);
454 if (is_huge_zero_page(page
)) {
456 __split_huge_pmd(walk
->vma
, pmd
, addr
, false, NULL
);
460 if (!thp_migration_supported()) {
464 ret
= split_huge_page(page
);
469 if (!queue_pages_required(page
, qp
))
473 /* go to thp migration */
474 if (flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
)) {
475 if (!vma_migratable(walk
->vma
) ||
476 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 * queue_pages_pte_range() has three possible return values:
493 * 0 - pages are placed on the right node or queued successfully.
494 * 1 - there is unmovable page, and MPOL_MF_MOVE* & MPOL_MF_STRICT were
496 * -EIO - only MPOL_MF_STRICT was specified and an existing page was already
497 * on a node that does not follow the policy.
499 static int queue_pages_pte_range(pmd_t
*pmd
, unsigned long addr
,
500 unsigned long end
, struct mm_walk
*walk
)
502 struct vm_area_struct
*vma
= walk
->vma
;
504 struct queue_pages
*qp
= walk
->private;
505 unsigned long flags
= qp
->flags
;
507 bool has_unmovable
= false;
511 ptl
= pmd_trans_huge_lock(pmd
, vma
);
513 ret
= queue_pages_pmd(pmd
, ptl
, addr
, end
, walk
);
517 /* THP was split, fall through to pte walk */
519 if (pmd_trans_unstable(pmd
))
522 pte
= pte_offset_map_lock(walk
->mm
, pmd
, addr
, &ptl
);
523 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
) {
524 if (!pte_present(*pte
))
526 page
= vm_normal_page(vma
, addr
, *pte
);
530 * vm_normal_page() filters out zero pages, but there might
531 * still be PageReserved pages to skip, perhaps in a VDSO.
533 if (PageReserved(page
))
535 if (!queue_pages_required(page
, qp
))
537 if (PageTransCompound(page
) && !thp_migration_supported()) {
539 pte_unmap_unlock(pte
, ptl
);
541 ret
= split_huge_page(page
);
544 /* Failed to split -- skip. */
546 pte
= pte_offset_map_lock(walk
->mm
, pmd
,
553 if (flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
)) {
554 /* MPOL_MF_STRICT must be specified if we get here */
555 if (!vma_migratable(vma
)) {
556 has_unmovable
= true;
561 * Do not abort immediately since there may be
562 * temporary off LRU pages in the range. Still
563 * need migrate other LRU pages.
565 if (migrate_page_add(page
, qp
->pagelist
, flags
))
566 has_unmovable
= true;
570 pte_unmap_unlock(pte
- 1, ptl
);
576 return addr
!= end
? -EIO
: 0;
579 static int queue_pages_hugetlb(pte_t
*pte
, unsigned long hmask
,
580 unsigned long addr
, unsigned long end
,
581 struct mm_walk
*walk
)
583 #ifdef CONFIG_HUGETLB_PAGE
584 struct queue_pages
*qp
= walk
->private;
585 unsigned long flags
= qp
->flags
;
590 ptl
= huge_pte_lock(hstate_vma(walk
->vma
), walk
->mm
, pte
);
591 entry
= huge_ptep_get(pte
);
592 if (!pte_present(entry
))
594 page
= pte_page(entry
);
595 if (!queue_pages_required(page
, qp
))
597 /* With MPOL_MF_MOVE, we migrate only unshared hugepage. */
598 if (flags
& (MPOL_MF_MOVE_ALL
) ||
599 (flags
& MPOL_MF_MOVE
&& page_mapcount(page
) == 1))
600 isolate_huge_page(page
, qp
->pagelist
);
609 #ifdef CONFIG_NUMA_BALANCING
611 * This is used to mark a range of virtual addresses to be inaccessible.
612 * These are later cleared by a NUMA hinting fault. Depending on these
613 * faults, pages may be migrated for better NUMA placement.
615 * This is assuming that NUMA faults are handled using PROT_NONE. If
616 * an architecture makes a different choice, it will need further
617 * changes to the core.
619 unsigned long change_prot_numa(struct vm_area_struct
*vma
,
620 unsigned long addr
, unsigned long end
)
624 nr_updated
= change_protection(vma
, addr
, end
, PAGE_NONE
, 0, 1);
626 count_vm_numa_events(NUMA_PTE_UPDATES
, nr_updated
);
631 static unsigned long change_prot_numa(struct vm_area_struct
*vma
,
632 unsigned long addr
, unsigned long end
)
636 #endif /* CONFIG_NUMA_BALANCING */
638 static int queue_pages_test_walk(unsigned long start
, unsigned long end
,
639 struct mm_walk
*walk
)
641 struct vm_area_struct
*vma
= walk
->vma
;
642 struct queue_pages
*qp
= walk
->private;
643 unsigned long endvma
= vma
->vm_end
;
644 unsigned long flags
= qp
->flags
;
647 * Need check MPOL_MF_STRICT to return -EIO if possible
648 * regardless of vma_migratable
650 if (!vma_migratable(vma
) &&
651 !(flags
& MPOL_MF_STRICT
))
656 if (vma
->vm_start
> start
)
657 start
= vma
->vm_start
;
659 if (!(flags
& MPOL_MF_DISCONTIG_OK
)) {
660 if (!vma
->vm_next
&& vma
->vm_end
< end
)
662 if (qp
->prev
&& qp
->prev
->vm_end
< vma
->vm_start
)
668 if (flags
& MPOL_MF_LAZY
) {
669 /* Similar to task_numa_work, skip inaccessible VMAs */
670 if (!is_vm_hugetlb_page(vma
) &&
671 (vma
->vm_flags
& (VM_READ
| VM_EXEC
| VM_WRITE
)) &&
672 !(vma
->vm_flags
& VM_MIXEDMAP
))
673 change_prot_numa(vma
, start
, endvma
);
677 /* queue pages from current vma */
678 if (flags
& MPOL_MF_VALID
)
684 * Walk through page tables and collect pages to be migrated.
686 * If pages found in a given range are on a set of nodes (determined by
687 * @nodes and @flags,) it's isolated and queued to the pagelist which is
688 * passed via @private.
690 * queue_pages_range() has three possible return values:
691 * 1 - there is unmovable page, but MPOL_MF_MOVE* & MPOL_MF_STRICT were
693 * 0 - queue pages successfully or no misplaced page.
694 * errno - i.e. misplaced pages with MPOL_MF_STRICT specified (-EIO) or
695 * memory range specified by nodemask and maxnode points outside
696 * your accessible address space (-EFAULT)
699 queue_pages_range(struct mm_struct
*mm
, unsigned long start
, unsigned long end
,
700 nodemask_t
*nodes
, unsigned long flags
,
701 struct list_head
*pagelist
)
703 struct queue_pages qp
= {
704 .pagelist
= pagelist
,
709 struct mm_walk queue_pages_walk
= {
710 .hugetlb_entry
= queue_pages_hugetlb
,
711 .pmd_entry
= queue_pages_pte_range
,
712 .test_walk
= queue_pages_test_walk
,
717 return walk_page_range(start
, end
, &queue_pages_walk
);
721 * Apply policy to a single VMA
722 * This must be called with the mmap_sem held for writing.
724 static int vma_replace_policy(struct vm_area_struct
*vma
,
725 struct mempolicy
*pol
)
728 struct mempolicy
*old
;
729 struct mempolicy
*new;
731 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
732 vma
->vm_start
, vma
->vm_end
, vma
->vm_pgoff
,
733 vma
->vm_ops
, vma
->vm_file
,
734 vma
->vm_ops
? vma
->vm_ops
->set_policy
: NULL
);
740 if (vma
->vm_ops
&& vma
->vm_ops
->set_policy
) {
741 err
= vma
->vm_ops
->set_policy(vma
, new);
746 old
= vma
->vm_policy
;
747 vma
->vm_policy
= new; /* protected by mmap_sem */
756 /* Step 2: apply policy to a range and do splits. */
757 static int mbind_range(struct mm_struct
*mm
, unsigned long start
,
758 unsigned long end
, struct mempolicy
*new_pol
)
760 struct vm_area_struct
*next
;
761 struct vm_area_struct
*prev
;
762 struct vm_area_struct
*vma
;
765 unsigned long vmstart
;
768 vma
= find_vma(mm
, start
);
769 if (!vma
|| vma
->vm_start
> start
)
773 if (start
> vma
->vm_start
)
776 for (; vma
&& vma
->vm_start
< end
; prev
= vma
, vma
= next
) {
778 vmstart
= max(start
, vma
->vm_start
);
779 vmend
= min(end
, vma
->vm_end
);
781 if (mpol_equal(vma_policy(vma
), new_pol
))
784 pgoff
= vma
->vm_pgoff
+
785 ((vmstart
- vma
->vm_start
) >> PAGE_SHIFT
);
786 prev
= vma_merge(mm
, prev
, vmstart
, vmend
, vma
->vm_flags
,
787 vma
->anon_vma
, vma
->vm_file
, pgoff
,
788 new_pol
, vma
->vm_userfaultfd_ctx
);
792 if (mpol_equal(vma_policy(vma
), new_pol
))
794 /* vma_merge() joined vma && vma->next, case 8 */
797 if (vma
->vm_start
!= vmstart
) {
798 err
= split_vma(vma
->vm_mm
, vma
, vmstart
, 1);
802 if (vma
->vm_end
!= vmend
) {
803 err
= split_vma(vma
->vm_mm
, vma
, vmend
, 0);
808 err
= vma_replace_policy(vma
, new_pol
);
817 /* Set the process memory policy */
818 static long do_set_mempolicy(unsigned short mode
, unsigned short flags
,
821 struct mempolicy
*new, *old
;
822 NODEMASK_SCRATCH(scratch
);
828 new = mpol_new(mode
, flags
, nodes
);
835 ret
= mpol_set_nodemask(new, nodes
, scratch
);
837 task_unlock(current
);
841 old
= current
->mempolicy
;
842 current
->mempolicy
= new;
843 if (new && new->mode
== MPOL_INTERLEAVE
)
844 current
->il_prev
= MAX_NUMNODES
-1;
845 task_unlock(current
);
849 NODEMASK_SCRATCH_FREE(scratch
);
854 * Return nodemask for policy for get_mempolicy() query
856 * Called with task's alloc_lock held
858 static void get_policy_nodemask(struct mempolicy
*p
, nodemask_t
*nodes
)
861 if (p
== &default_policy
)
867 case MPOL_INTERLEAVE
:
871 if (!(p
->flags
& MPOL_F_LOCAL
))
872 node_set(p
->v
.preferred_node
, *nodes
);
873 /* else return empty node mask for local allocation */
880 static int lookup_node(unsigned long addr
)
885 err
= get_user_pages(addr
& PAGE_MASK
, 1, 0, &p
, NULL
);
887 err
= page_to_nid(p
);
893 /* Retrieve NUMA policy */
894 static long do_get_mempolicy(int *policy
, nodemask_t
*nmask
,
895 unsigned long addr
, unsigned long flags
)
898 struct mm_struct
*mm
= current
->mm
;
899 struct vm_area_struct
*vma
= NULL
;
900 struct mempolicy
*pol
= current
->mempolicy
;
903 ~(unsigned long)(MPOL_F_NODE
|MPOL_F_ADDR
|MPOL_F_MEMS_ALLOWED
))
906 if (flags
& MPOL_F_MEMS_ALLOWED
) {
907 if (flags
& (MPOL_F_NODE
|MPOL_F_ADDR
))
909 *policy
= 0; /* just so it's initialized */
911 *nmask
= cpuset_current_mems_allowed
;
912 task_unlock(current
);
916 if (flags
& MPOL_F_ADDR
) {
918 * Do NOT fall back to task policy if the
919 * vma/shared policy at addr is NULL. We
920 * want to return MPOL_DEFAULT in this case.
922 down_read(&mm
->mmap_sem
);
923 vma
= find_vma_intersection(mm
, addr
, addr
+1);
925 up_read(&mm
->mmap_sem
);
928 if (vma
->vm_ops
&& vma
->vm_ops
->get_policy
)
929 pol
= vma
->vm_ops
->get_policy(vma
, addr
);
931 pol
= vma
->vm_policy
;
936 pol
= &default_policy
; /* indicates default behavior */
938 if (flags
& MPOL_F_NODE
) {
939 if (flags
& MPOL_F_ADDR
) {
940 err
= lookup_node(addr
);
944 } else if (pol
== current
->mempolicy
&&
945 pol
->mode
== MPOL_INTERLEAVE
) {
946 *policy
= next_node_in(current
->il_prev
, pol
->v
.nodes
);
952 *policy
= pol
== &default_policy
? MPOL_DEFAULT
:
955 * Internal mempolicy flags must be masked off before exposing
956 * the policy to userspace.
958 *policy
|= (pol
->flags
& MPOL_MODE_FLAGS
);
963 if (mpol_store_user_nodemask(pol
)) {
964 *nmask
= pol
->w
.user_nodemask
;
967 get_policy_nodemask(pol
, nmask
);
968 task_unlock(current
);
975 up_read(¤t
->mm
->mmap_sem
);
979 #ifdef CONFIG_MIGRATION
981 * page migration, thp tail pages can be passed.
983 static int migrate_page_add(struct page
*page
, struct list_head
*pagelist
,
986 struct page
*head
= compound_head(page
);
988 * Avoid migrating a page that is shared with others.
990 if ((flags
& MPOL_MF_MOVE_ALL
) || page_mapcount(head
) == 1) {
991 if (!isolate_lru_page(head
)) {
992 list_add_tail(&head
->lru
, pagelist
);
993 mod_node_page_state(page_pgdat(head
),
994 NR_ISOLATED_ANON
+ page_is_file_cache(head
),
995 hpage_nr_pages(head
));
996 } else if (flags
& MPOL_MF_STRICT
) {
998 * Non-movable page may reach here. And, there may be
999 * temporary off LRU pages or non-LRU movable pages.
1000 * Treat them as unmovable pages since they can't be
1001 * isolated, so they can't be moved at the moment. It
1002 * should return -EIO for this case too.
1011 static struct page
*new_node_page(struct page
*page
, unsigned long node
, int **x
)
1014 return alloc_huge_page_node(page_hstate(compound_head(page
)),
1016 else if (thp_migration_supported() && PageTransHuge(page
)) {
1019 thp
= alloc_pages_node(node
,
1020 (GFP_TRANSHUGE
| __GFP_THISNODE
),
1024 prep_transhuge_page(thp
);
1027 return __alloc_pages_node(node
, GFP_HIGHUSER_MOVABLE
|
1032 * Migrate pages from one node to a target node.
1033 * Returns error or the number of pages not migrated.
1035 static int migrate_to_node(struct mm_struct
*mm
, int source
, int dest
,
1039 LIST_HEAD(pagelist
);
1043 node_set(source
, nmask
);
1046 * This does not "check" the range but isolates all pages that
1047 * need migration. Between passing in the full user address
1048 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1050 VM_BUG_ON(!(flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
)));
1051 queue_pages_range(mm
, mm
->mmap
->vm_start
, mm
->task_size
, &nmask
,
1052 flags
| MPOL_MF_DISCONTIG_OK
, &pagelist
);
1054 if (!list_empty(&pagelist
)) {
1055 err
= migrate_pages(&pagelist
, new_node_page
, NULL
, dest
,
1056 MIGRATE_SYNC
, MR_SYSCALL
);
1058 putback_movable_pages(&pagelist
);
1065 * Move pages between the two nodesets so as to preserve the physical
1066 * layout as much as possible.
1068 * Returns the number of page that could not be moved.
1070 int do_migrate_pages(struct mm_struct
*mm
, const nodemask_t
*from
,
1071 const nodemask_t
*to
, int flags
)
1077 err
= migrate_prep();
1081 down_read(&mm
->mmap_sem
);
1084 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1085 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1086 * bit in 'tmp', and return that <source, dest> pair for migration.
1087 * The pair of nodemasks 'to' and 'from' define the map.
1089 * If no pair of bits is found that way, fallback to picking some
1090 * pair of 'source' and 'dest' bits that are not the same. If the
1091 * 'source' and 'dest' bits are the same, this represents a node
1092 * that will be migrating to itself, so no pages need move.
1094 * If no bits are left in 'tmp', or if all remaining bits left
1095 * in 'tmp' correspond to the same bit in 'to', return false
1096 * (nothing left to migrate).
1098 * This lets us pick a pair of nodes to migrate between, such that
1099 * if possible the dest node is not already occupied by some other
1100 * source node, minimizing the risk of overloading the memory on a
1101 * node that would happen if we migrated incoming memory to a node
1102 * before migrating outgoing memory source that same node.
1104 * A single scan of tmp is sufficient. As we go, we remember the
1105 * most recent <s, d> pair that moved (s != d). If we find a pair
1106 * that not only moved, but what's better, moved to an empty slot
1107 * (d is not set in tmp), then we break out then, with that pair.
1108 * Otherwise when we finish scanning from_tmp, we at least have the
1109 * most recent <s, d> pair that moved. If we get all the way through
1110 * the scan of tmp without finding any node that moved, much less
1111 * moved to an empty node, then there is nothing left worth migrating.
1115 while (!nodes_empty(tmp
)) {
1117 int source
= NUMA_NO_NODE
;
1120 for_each_node_mask(s
, tmp
) {
1123 * do_migrate_pages() tries to maintain the relative
1124 * node relationship of the pages established between
1125 * threads and memory areas.
1127 * However if the number of source nodes is not equal to
1128 * the number of destination nodes we can not preserve
1129 * this node relative relationship. In that case, skip
1130 * copying memory from a node that is in the destination
1133 * Example: [2,3,4] -> [3,4,5] moves everything.
1134 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1137 if ((nodes_weight(*from
) != nodes_weight(*to
)) &&
1138 (node_isset(s
, *to
)))
1141 d
= node_remap(s
, *from
, *to
);
1145 source
= s
; /* Node moved. Memorize */
1148 /* dest not in remaining from nodes? */
1149 if (!node_isset(dest
, tmp
))
1152 if (source
== NUMA_NO_NODE
)
1155 node_clear(source
, tmp
);
1156 err
= migrate_to_node(mm
, source
, dest
, flags
);
1162 up_read(&mm
->mmap_sem
);
1170 * Allocate a new page for page migration based on vma policy.
1171 * Start by assuming the page is mapped by the same vma as contains @start.
1172 * Search forward from there, if not. N.B., this assumes that the
1173 * list of pages handed to migrate_pages()--which is how we get here--
1174 * is in virtual address order.
1176 static struct page
*new_page(struct page
*page
, unsigned long start
, int **x
)
1178 struct vm_area_struct
*vma
;
1179 unsigned long uninitialized_var(address
);
1181 vma
= find_vma(current
->mm
, start
);
1183 address
= page_address_in_vma(page
, vma
);
1184 if (address
!= -EFAULT
)
1189 if (PageHuge(page
)) {
1191 return alloc_huge_page_noerr(vma
, address
, 1);
1192 } else if (thp_migration_supported() && PageTransHuge(page
)) {
1195 thp
= alloc_hugepage_vma(GFP_TRANSHUGE
, vma
, address
,
1199 prep_transhuge_page(thp
);
1203 * if !vma, alloc_page_vma() will use task or system default policy
1205 return alloc_page_vma(GFP_HIGHUSER_MOVABLE
| __GFP_RETRY_MAYFAIL
,
1210 static int migrate_page_add(struct page
*page
, struct list_head
*pagelist
,
1211 unsigned long flags
)
1216 int do_migrate_pages(struct mm_struct
*mm
, const nodemask_t
*from
,
1217 const nodemask_t
*to
, int flags
)
1222 static struct page
*new_page(struct page
*page
, unsigned long start
, int **x
)
1228 static long do_mbind(unsigned long start
, unsigned long len
,
1229 unsigned short mode
, unsigned short mode_flags
,
1230 nodemask_t
*nmask
, unsigned long flags
)
1232 struct mm_struct
*mm
= current
->mm
;
1233 struct mempolicy
*new;
1237 LIST_HEAD(pagelist
);
1239 if (flags
& ~(unsigned long)MPOL_MF_VALID
)
1241 if ((flags
& MPOL_MF_MOVE_ALL
) && !capable(CAP_SYS_NICE
))
1244 if (start
& ~PAGE_MASK
)
1247 if (mode
== MPOL_DEFAULT
)
1248 flags
&= ~MPOL_MF_STRICT
;
1250 len
= (len
+ PAGE_SIZE
- 1) & PAGE_MASK
;
1258 new = mpol_new(mode
, mode_flags
, nmask
);
1260 return PTR_ERR(new);
1262 if (flags
& MPOL_MF_LAZY
)
1263 new->flags
|= MPOL_F_MOF
;
1266 * If we are using the default policy then operation
1267 * on discontinuous address spaces is okay after all
1270 flags
|= MPOL_MF_DISCONTIG_OK
;
1272 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1273 start
, start
+ len
, mode
, mode_flags
,
1274 nmask
? nodes_addr(*nmask
)[0] : NUMA_NO_NODE
);
1276 if (flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
)) {
1278 err
= migrate_prep();
1283 NODEMASK_SCRATCH(scratch
);
1285 down_write(&mm
->mmap_sem
);
1287 err
= mpol_set_nodemask(new, nmask
, scratch
);
1288 task_unlock(current
);
1290 up_write(&mm
->mmap_sem
);
1293 NODEMASK_SCRATCH_FREE(scratch
);
1298 ret
= queue_pages_range(mm
, start
, end
, nmask
,
1299 flags
| MPOL_MF_INVERT
, &pagelist
);
1306 err
= mbind_range(mm
, start
, end
, new);
1311 if (!list_empty(&pagelist
)) {
1312 WARN_ON_ONCE(flags
& MPOL_MF_LAZY
);
1313 nr_failed
= migrate_pages(&pagelist
, new_page
, NULL
,
1314 start
, MIGRATE_SYNC
, MR_MEMPOLICY_MBIND
);
1316 putback_movable_pages(&pagelist
);
1319 if ((ret
> 0) || (nr_failed
&& (flags
& MPOL_MF_STRICT
)))
1323 if (!list_empty(&pagelist
))
1324 putback_movable_pages(&pagelist
);
1327 up_write(&mm
->mmap_sem
);
1334 * User space interface with variable sized bitmaps for nodelists.
1337 /* Copy a node mask from user space. */
1338 static int get_nodes(nodemask_t
*nodes
, const unsigned long __user
*nmask
,
1339 unsigned long maxnode
)
1343 unsigned long nlongs
;
1344 unsigned long endmask
;
1347 nodes_clear(*nodes
);
1348 if (maxnode
== 0 || !nmask
)
1350 if (maxnode
> PAGE_SIZE
*BITS_PER_BYTE
)
1353 nlongs
= BITS_TO_LONGS(maxnode
);
1354 if ((maxnode
% BITS_PER_LONG
) == 0)
1357 endmask
= (1UL << (maxnode
% BITS_PER_LONG
)) - 1;
1360 * When the user specified more nodes than supported just check
1361 * if the non supported part is all zero.
1363 * If maxnode have more longs than MAX_NUMNODES, check
1364 * the bits in that area first. And then go through to
1365 * check the rest bits which equal or bigger than MAX_NUMNODES.
1366 * Otherwise, just check bits [MAX_NUMNODES, maxnode).
1368 if (nlongs
> BITS_TO_LONGS(MAX_NUMNODES
)) {
1369 if (nlongs
> PAGE_SIZE
/sizeof(long))
1371 for (k
= BITS_TO_LONGS(MAX_NUMNODES
); k
< nlongs
; k
++) {
1372 if (get_user(t
, nmask
+ k
))
1374 if (k
== nlongs
- 1) {
1380 nlongs
= BITS_TO_LONGS(MAX_NUMNODES
);
1384 if (maxnode
> MAX_NUMNODES
&& MAX_NUMNODES
% BITS_PER_LONG
!= 0) {
1385 unsigned long valid_mask
= endmask
;
1387 valid_mask
&= ~((1UL << (MAX_NUMNODES
% BITS_PER_LONG
)) - 1);
1388 if (get_user(t
, nmask
+ nlongs
- 1))
1394 if (copy_from_user(nodes_addr(*nodes
), nmask
, nlongs
*sizeof(unsigned long)))
1396 nodes_addr(*nodes
)[nlongs
-1] &= endmask
;
1400 /* Copy a kernel node mask to user space */
1401 static int copy_nodes_to_user(unsigned long __user
*mask
, unsigned long maxnode
,
1404 unsigned long copy
= ALIGN(maxnode
-1, 64) / 8;
1405 unsigned int nbytes
= BITS_TO_LONGS(nr_node_ids
) * sizeof(long);
1407 if (copy
> nbytes
) {
1408 if (copy
> PAGE_SIZE
)
1410 if (clear_user((char __user
*)mask
+ nbytes
, copy
- nbytes
))
1414 return copy_to_user(mask
, nodes_addr(*nodes
), copy
) ? -EFAULT
: 0;
1417 SYSCALL_DEFINE6(mbind
, unsigned long, start
, unsigned long, len
,
1418 unsigned long, mode
, const unsigned long __user
*, nmask
,
1419 unsigned long, maxnode
, unsigned, flags
)
1423 unsigned short mode_flags
;
1425 mode_flags
= mode
& MPOL_MODE_FLAGS
;
1426 mode
&= ~MPOL_MODE_FLAGS
;
1427 if (mode
>= MPOL_MAX
)
1429 if ((mode_flags
& MPOL_F_STATIC_NODES
) &&
1430 (mode_flags
& MPOL_F_RELATIVE_NODES
))
1432 err
= get_nodes(&nodes
, nmask
, maxnode
);
1435 return do_mbind(start
, len
, mode
, mode_flags
, &nodes
, flags
);
1438 /* Set the process memory policy */
1439 SYSCALL_DEFINE3(set_mempolicy
, int, mode
, const unsigned long __user
*, nmask
,
1440 unsigned long, maxnode
)
1444 unsigned short flags
;
1446 flags
= mode
& MPOL_MODE_FLAGS
;
1447 mode
&= ~MPOL_MODE_FLAGS
;
1448 if ((unsigned int)mode
>= MPOL_MAX
)
1450 if ((flags
& MPOL_F_STATIC_NODES
) && (flags
& MPOL_F_RELATIVE_NODES
))
1452 err
= get_nodes(&nodes
, nmask
, maxnode
);
1455 return do_set_mempolicy(mode
, flags
, &nodes
);
1458 SYSCALL_DEFINE4(migrate_pages
, pid_t
, pid
, unsigned long, maxnode
,
1459 const unsigned long __user
*, old_nodes
,
1460 const unsigned long __user
*, new_nodes
)
1462 struct mm_struct
*mm
= NULL
;
1463 struct task_struct
*task
;
1464 nodemask_t task_nodes
;
1468 NODEMASK_SCRATCH(scratch
);
1473 old
= &scratch
->mask1
;
1474 new = &scratch
->mask2
;
1476 err
= get_nodes(old
, old_nodes
, maxnode
);
1480 err
= get_nodes(new, new_nodes
, maxnode
);
1484 /* Find the mm_struct */
1486 task
= pid
? find_task_by_vpid(pid
) : current
;
1492 get_task_struct(task
);
1497 * Check if this process has the right to modify the specified process.
1498 * Use the regular "ptrace_may_access()" checks.
1500 if (!ptrace_may_access(task
, PTRACE_MODE_READ_REALCREDS
)) {
1507 task_nodes
= cpuset_mems_allowed(task
);
1508 /* Is the user allowed to access the target nodes? */
1509 if (!nodes_subset(*new, task_nodes
) && !capable(CAP_SYS_NICE
)) {
1514 task_nodes
= cpuset_mems_allowed(current
);
1515 nodes_and(*new, *new, task_nodes
);
1516 if (nodes_empty(*new))
1519 nodes_and(*new, *new, node_states
[N_MEMORY
]);
1520 if (nodes_empty(*new))
1523 err
= security_task_movememory(task
);
1527 mm
= get_task_mm(task
);
1528 put_task_struct(task
);
1535 err
= do_migrate_pages(mm
, old
, new,
1536 capable(CAP_SYS_NICE
) ? MPOL_MF_MOVE_ALL
: MPOL_MF_MOVE
);
1540 NODEMASK_SCRATCH_FREE(scratch
);
1545 put_task_struct(task
);
1551 /* Retrieve NUMA policy */
1552 SYSCALL_DEFINE5(get_mempolicy
, int __user
*, policy
,
1553 unsigned long __user
*, nmask
, unsigned long, maxnode
,
1554 unsigned long, addr
, unsigned long, flags
)
1557 int uninitialized_var(pval
);
1560 if (nmask
!= NULL
&& maxnode
< nr_node_ids
)
1563 err
= do_get_mempolicy(&pval
, &nodes
, addr
, flags
);
1568 if (policy
&& put_user(pval
, policy
))
1572 err
= copy_nodes_to_user(nmask
, maxnode
, &nodes
);
1577 #ifdef CONFIG_COMPAT
1579 COMPAT_SYSCALL_DEFINE5(get_mempolicy
, int __user
*, policy
,
1580 compat_ulong_t __user
*, nmask
,
1581 compat_ulong_t
, maxnode
,
1582 compat_ulong_t
, addr
, compat_ulong_t
, flags
)
1585 unsigned long __user
*nm
= NULL
;
1586 unsigned long nr_bits
, alloc_size
;
1587 DECLARE_BITMAP(bm
, MAX_NUMNODES
);
1589 nr_bits
= min_t(unsigned long, maxnode
-1, nr_node_ids
);
1590 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
1593 nm
= compat_alloc_user_space(alloc_size
);
1595 err
= sys_get_mempolicy(policy
, nm
, nr_bits
+1, addr
, flags
);
1597 if (!err
&& nmask
) {
1598 unsigned long copy_size
;
1599 copy_size
= min_t(unsigned long, sizeof(bm
), alloc_size
);
1600 err
= copy_from_user(bm
, nm
, copy_size
);
1601 /* ensure entire bitmap is zeroed */
1602 err
|= clear_user(nmask
, ALIGN(maxnode
-1, 8) / 8);
1603 err
|= compat_put_bitmap(nmask
, bm
, nr_bits
);
1609 COMPAT_SYSCALL_DEFINE3(set_mempolicy
, int, mode
, compat_ulong_t __user
*, nmask
,
1610 compat_ulong_t
, maxnode
)
1612 unsigned long __user
*nm
= NULL
;
1613 unsigned long nr_bits
, alloc_size
;
1614 DECLARE_BITMAP(bm
, MAX_NUMNODES
);
1616 nr_bits
= min_t(unsigned long, maxnode
-1, MAX_NUMNODES
);
1617 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
1620 if (compat_get_bitmap(bm
, nmask
, nr_bits
))
1622 nm
= compat_alloc_user_space(alloc_size
);
1623 if (copy_to_user(nm
, bm
, alloc_size
))
1627 return sys_set_mempolicy(mode
, nm
, nr_bits
+1);
1630 COMPAT_SYSCALL_DEFINE6(mbind
, compat_ulong_t
, start
, compat_ulong_t
, len
,
1631 compat_ulong_t
, mode
, compat_ulong_t __user
*, nmask
,
1632 compat_ulong_t
, maxnode
, compat_ulong_t
, flags
)
1634 unsigned long __user
*nm
= NULL
;
1635 unsigned long nr_bits
, alloc_size
;
1638 nr_bits
= min_t(unsigned long, maxnode
-1, MAX_NUMNODES
);
1639 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
1642 if (compat_get_bitmap(nodes_addr(bm
), nmask
, nr_bits
))
1644 nm
= compat_alloc_user_space(alloc_size
);
1645 if (copy_to_user(nm
, nodes_addr(bm
), alloc_size
))
1649 return sys_mbind(start
, len
, mode
, nm
, nr_bits
+1, flags
);
1654 struct mempolicy
*__get_vma_policy(struct vm_area_struct
*vma
,
1657 struct mempolicy
*pol
= NULL
;
1660 if (vma
->vm_ops
&& vma
->vm_ops
->get_policy
) {
1661 pol
= vma
->vm_ops
->get_policy(vma
, addr
);
1662 } else if (vma
->vm_policy
) {
1663 pol
= vma
->vm_policy
;
1666 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1667 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1668 * count on these policies which will be dropped by
1669 * mpol_cond_put() later
1671 if (mpol_needs_cond_ref(pol
))
1680 * get_vma_policy(@vma, @addr)
1681 * @vma: virtual memory area whose policy is sought
1682 * @addr: address in @vma for shared policy lookup
1684 * Returns effective policy for a VMA at specified address.
1685 * Falls back to current->mempolicy or system default policy, as necessary.
1686 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1687 * count--added by the get_policy() vm_op, as appropriate--to protect against
1688 * freeing by another task. It is the caller's responsibility to free the
1689 * extra reference for shared policies.
1691 static struct mempolicy
*get_vma_policy(struct vm_area_struct
*vma
,
1694 struct mempolicy
*pol
= __get_vma_policy(vma
, addr
);
1697 pol
= get_task_policy(current
);
1702 bool vma_policy_mof(struct vm_area_struct
*vma
)
1704 struct mempolicy
*pol
;
1706 if (vma
->vm_ops
&& vma
->vm_ops
->get_policy
) {
1709 pol
= vma
->vm_ops
->get_policy(vma
, vma
->vm_start
);
1710 if (pol
&& (pol
->flags
& MPOL_F_MOF
))
1717 pol
= vma
->vm_policy
;
1719 pol
= get_task_policy(current
);
1721 return pol
->flags
& MPOL_F_MOF
;
1724 static int apply_policy_zone(struct mempolicy
*policy
, enum zone_type zone
)
1726 enum zone_type dynamic_policy_zone
= policy_zone
;
1728 BUG_ON(dynamic_policy_zone
== ZONE_MOVABLE
);
1731 * if policy->v.nodes has movable memory only,
1732 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1734 * policy->v.nodes is intersect with node_states[N_MEMORY].
1735 * so if the following test faile, it implies
1736 * policy->v.nodes has movable memory only.
1738 if (!nodes_intersects(policy
->v
.nodes
, node_states
[N_HIGH_MEMORY
]))
1739 dynamic_policy_zone
= ZONE_MOVABLE
;
1741 return zone
>= dynamic_policy_zone
;
1745 * Return a nodemask representing a mempolicy for filtering nodes for
1748 static nodemask_t
*policy_nodemask(gfp_t gfp
, struct mempolicy
*policy
)
1750 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1751 if (unlikely(policy
->mode
== MPOL_BIND
) &&
1752 apply_policy_zone(policy
, gfp_zone(gfp
)) &&
1753 cpuset_nodemask_valid_mems_allowed(&policy
->v
.nodes
))
1754 return &policy
->v
.nodes
;
1759 /* Return the node id preferred by the given mempolicy, or the given id */
1760 static int policy_node(gfp_t gfp
, struct mempolicy
*policy
,
1763 if (policy
->mode
== MPOL_PREFERRED
&& !(policy
->flags
& MPOL_F_LOCAL
))
1764 nd
= policy
->v
.preferred_node
;
1767 * __GFP_THISNODE shouldn't even be used with the bind policy
1768 * because we might easily break the expectation to stay on the
1769 * requested node and not break the policy.
1771 WARN_ON_ONCE(policy
->mode
== MPOL_BIND
&& (gfp
& __GFP_THISNODE
));
1777 /* Do dynamic interleaving for a process */
1778 static unsigned interleave_nodes(struct mempolicy
*policy
)
1781 struct task_struct
*me
= current
;
1783 next
= next_node_in(me
->il_prev
, policy
->v
.nodes
);
1784 if (next
< MAX_NUMNODES
)
1790 * Depending on the memory policy provide a node from which to allocate the
1793 unsigned int mempolicy_slab_node(void)
1795 struct mempolicy
*policy
;
1796 int node
= numa_mem_id();
1801 policy
= current
->mempolicy
;
1802 if (!policy
|| policy
->flags
& MPOL_F_LOCAL
)
1805 switch (policy
->mode
) {
1806 case MPOL_PREFERRED
:
1808 * handled MPOL_F_LOCAL above
1810 return policy
->v
.preferred_node
;
1812 case MPOL_INTERLEAVE
:
1813 return interleave_nodes(policy
);
1819 * Follow bind policy behavior and start allocation at the
1822 struct zonelist
*zonelist
;
1823 enum zone_type highest_zoneidx
= gfp_zone(GFP_KERNEL
);
1824 zonelist
= &NODE_DATA(node
)->node_zonelists
[ZONELIST_FALLBACK
];
1825 z
= first_zones_zonelist(zonelist
, highest_zoneidx
,
1827 return z
->zone
? z
->zone
->node
: node
;
1836 * Do static interleaving for a VMA with known offset @n. Returns the n'th
1837 * node in pol->v.nodes (starting from n=0), wrapping around if n exceeds the
1838 * number of present nodes.
1840 static unsigned offset_il_node(struct mempolicy
*pol
, unsigned long n
)
1842 unsigned nnodes
= nodes_weight(pol
->v
.nodes
);
1848 return numa_node_id();
1849 target
= (unsigned int)n
% nnodes
;
1850 nid
= first_node(pol
->v
.nodes
);
1851 for (i
= 0; i
< target
; i
++)
1852 nid
= next_node(nid
, pol
->v
.nodes
);
1856 /* Determine a node number for interleave */
1857 static inline unsigned interleave_nid(struct mempolicy
*pol
,
1858 struct vm_area_struct
*vma
, unsigned long addr
, int shift
)
1864 * for small pages, there is no difference between
1865 * shift and PAGE_SHIFT, so the bit-shift is safe.
1866 * for huge pages, since vm_pgoff is in units of small
1867 * pages, we need to shift off the always 0 bits to get
1870 BUG_ON(shift
< PAGE_SHIFT
);
1871 off
= vma
->vm_pgoff
>> (shift
- PAGE_SHIFT
);
1872 off
+= (addr
- vma
->vm_start
) >> shift
;
1873 return offset_il_node(pol
, off
);
1875 return interleave_nodes(pol
);
1878 #ifdef CONFIG_HUGETLBFS
1880 * huge_node(@vma, @addr, @gfp_flags, @mpol)
1881 * @vma: virtual memory area whose policy is sought
1882 * @addr: address in @vma for shared policy lookup and interleave policy
1883 * @gfp_flags: for requested zone
1884 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
1885 * @nodemask: pointer to nodemask pointer for MPOL_BIND nodemask
1887 * Returns a nid suitable for a huge page allocation and a pointer
1888 * to the struct mempolicy for conditional unref after allocation.
1889 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1890 * @nodemask for filtering the zonelist.
1892 * Must be protected by read_mems_allowed_begin()
1894 int huge_node(struct vm_area_struct
*vma
, unsigned long addr
, gfp_t gfp_flags
,
1895 struct mempolicy
**mpol
, nodemask_t
**nodemask
)
1899 *mpol
= get_vma_policy(vma
, addr
);
1900 *nodemask
= NULL
; /* assume !MPOL_BIND */
1902 if (unlikely((*mpol
)->mode
== MPOL_INTERLEAVE
)) {
1903 nid
= interleave_nid(*mpol
, vma
, addr
,
1904 huge_page_shift(hstate_vma(vma
)));
1906 nid
= policy_node(gfp_flags
, *mpol
, numa_node_id());
1907 if ((*mpol
)->mode
== MPOL_BIND
)
1908 *nodemask
= &(*mpol
)->v
.nodes
;
1914 * init_nodemask_of_mempolicy
1916 * If the current task's mempolicy is "default" [NULL], return 'false'
1917 * to indicate default policy. Otherwise, extract the policy nodemask
1918 * for 'bind' or 'interleave' policy into the argument nodemask, or
1919 * initialize the argument nodemask to contain the single node for
1920 * 'preferred' or 'local' policy and return 'true' to indicate presence
1921 * of non-default mempolicy.
1923 * We don't bother with reference counting the mempolicy [mpol_get/put]
1924 * because the current task is examining it's own mempolicy and a task's
1925 * mempolicy is only ever changed by the task itself.
1927 * N.B., it is the caller's responsibility to free a returned nodemask.
1929 bool init_nodemask_of_mempolicy(nodemask_t
*mask
)
1931 struct mempolicy
*mempolicy
;
1934 if (!(mask
&& current
->mempolicy
))
1938 mempolicy
= current
->mempolicy
;
1939 switch (mempolicy
->mode
) {
1940 case MPOL_PREFERRED
:
1941 if (mempolicy
->flags
& MPOL_F_LOCAL
)
1942 nid
= numa_node_id();
1944 nid
= mempolicy
->v
.preferred_node
;
1945 init_nodemask_of_node(mask
, nid
);
1950 case MPOL_INTERLEAVE
:
1951 *mask
= mempolicy
->v
.nodes
;
1957 task_unlock(current
);
1964 * mempolicy_nodemask_intersects
1966 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1967 * policy. Otherwise, check for intersection between mask and the policy
1968 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1969 * policy, always return true since it may allocate elsewhere on fallback.
1971 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1973 bool mempolicy_nodemask_intersects(struct task_struct
*tsk
,
1974 const nodemask_t
*mask
)
1976 struct mempolicy
*mempolicy
;
1982 mempolicy
= tsk
->mempolicy
;
1986 switch (mempolicy
->mode
) {
1987 case MPOL_PREFERRED
:
1989 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1990 * allocate from, they may fallback to other nodes when oom.
1991 * Thus, it's possible for tsk to have allocated memory from
1996 case MPOL_INTERLEAVE
:
1997 ret
= nodes_intersects(mempolicy
->v
.nodes
, *mask
);
2007 /* Allocate a page in interleaved policy.
2008 Own path because it needs to do special accounting. */
2009 static struct page
*alloc_page_interleave(gfp_t gfp
, unsigned order
,
2014 page
= __alloc_pages(gfp
, order
, nid
);
2015 /* skip NUMA_INTERLEAVE_HIT counter update if numa stats is disabled */
2016 if (!static_branch_likely(&vm_numa_stat_key
))
2018 if (page
&& page_to_nid(page
) == nid
) {
2020 __inc_numa_state(page_zone(page
), NUMA_INTERLEAVE_HIT
);
2027 * alloc_pages_vma - Allocate a page for a VMA.
2030 * %GFP_USER user allocation.
2031 * %GFP_KERNEL kernel allocations,
2032 * %GFP_HIGHMEM highmem/user allocations,
2033 * %GFP_FS allocation should not call back into a file system.
2034 * %GFP_ATOMIC don't sleep.
2036 * @order:Order of the GFP allocation.
2037 * @vma: Pointer to VMA or NULL if not available.
2038 * @addr: Virtual Address of the allocation. Must be inside the VMA.
2039 * @node: Which node to prefer for allocation (modulo policy).
2040 * @hugepage: for hugepages try only the preferred node if possible
2042 * This function allocates a page from the kernel page pool and applies
2043 * a NUMA policy associated with the VMA or the current process.
2044 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
2045 * mm_struct of the VMA to prevent it from going away. Should be used for
2046 * all allocations for pages that will be mapped into user space. Returns
2047 * NULL when no page can be allocated.
2050 alloc_pages_vma(gfp_t gfp
, int order
, struct vm_area_struct
*vma
,
2051 unsigned long addr
, int node
, bool hugepage
)
2053 struct mempolicy
*pol
;
2058 pol
= get_vma_policy(vma
, addr
);
2060 if (pol
->mode
== MPOL_INTERLEAVE
) {
2063 nid
= interleave_nid(pol
, vma
, addr
, PAGE_SHIFT
+ order
);
2065 page
= alloc_page_interleave(gfp
, order
, nid
);
2069 if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE
) && hugepage
)) {
2070 int hpage_node
= node
;
2073 * For hugepage allocation and non-interleave policy which
2074 * allows the current node (or other explicitly preferred
2075 * node) we only try to allocate from the current/preferred
2076 * node and don't fall back to other nodes, as the cost of
2077 * remote accesses would likely offset THP benefits.
2079 * If the policy is interleave, or does not allow the current
2080 * node in its nodemask, we allocate the standard way.
2082 if (pol
->mode
== MPOL_PREFERRED
&&
2083 !(pol
->flags
& MPOL_F_LOCAL
))
2084 hpage_node
= pol
->v
.preferred_node
;
2086 nmask
= policy_nodemask(gfp
, pol
);
2087 if (!nmask
|| node_isset(hpage_node
, *nmask
)) {
2090 * We cannot invoke reclaim if __GFP_THISNODE
2091 * is set. Invoking reclaim with
2092 * __GFP_THISNODE set, would cause THP
2093 * allocations to trigger heavy swapping
2094 * despite there may be tons of free memory
2095 * (including potentially plenty of THP
2096 * already available in the buddy) on all the
2099 * At most we could invoke compaction when
2100 * __GFP_THISNODE is set (but we would need to
2101 * refrain from invoking reclaim even if
2102 * compaction returned COMPACT_SKIPPED because
2103 * there wasn't not enough memory to succeed
2104 * compaction). For now just avoid
2105 * __GFP_THISNODE instead of limiting the
2106 * allocation path to a strict and single
2107 * compaction invocation.
2109 * Supposedly if direct reclaim was enabled by
2110 * the caller, the app prefers THP regardless
2111 * of the node it comes from so this would be
2112 * more desiderable behavior than only
2113 * providing THP originated from the local
2114 * node in such case.
2116 if (!(gfp
& __GFP_DIRECT_RECLAIM
))
2117 gfp
|= __GFP_THISNODE
;
2118 page
= __alloc_pages_node(hpage_node
, gfp
, order
);
2123 nmask
= policy_nodemask(gfp
, pol
);
2124 preferred_nid
= policy_node(gfp
, pol
, node
);
2125 page
= __alloc_pages_nodemask(gfp
, order
, preferred_nid
, nmask
);
2132 * alloc_pages_current - Allocate pages.
2135 * %GFP_USER user allocation,
2136 * %GFP_KERNEL kernel allocation,
2137 * %GFP_HIGHMEM highmem allocation,
2138 * %GFP_FS don't call back into a file system.
2139 * %GFP_ATOMIC don't sleep.
2140 * @order: Power of two of allocation size in pages. 0 is a single page.
2142 * Allocate a page from the kernel page pool. When not in
2143 * interrupt context and apply the current process NUMA policy.
2144 * Returns NULL when no page can be allocated.
2146 struct page
*alloc_pages_current(gfp_t gfp
, unsigned order
)
2148 struct mempolicy
*pol
= &default_policy
;
2151 if (!in_interrupt() && !(gfp
& __GFP_THISNODE
))
2152 pol
= get_task_policy(current
);
2155 * No reference counting needed for current->mempolicy
2156 * nor system default_policy
2158 if (pol
->mode
== MPOL_INTERLEAVE
)
2159 page
= alloc_page_interleave(gfp
, order
, interleave_nodes(pol
));
2161 page
= __alloc_pages_nodemask(gfp
, order
,
2162 policy_node(gfp
, pol
, numa_node_id()),
2163 policy_nodemask(gfp
, pol
));
2167 EXPORT_SYMBOL(alloc_pages_current
);
2169 int vma_dup_policy(struct vm_area_struct
*src
, struct vm_area_struct
*dst
)
2171 struct mempolicy
*pol
= mpol_dup(vma_policy(src
));
2174 return PTR_ERR(pol
);
2175 dst
->vm_policy
= pol
;
2180 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2181 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2182 * with the mems_allowed returned by cpuset_mems_allowed(). This
2183 * keeps mempolicies cpuset relative after its cpuset moves. See
2184 * further kernel/cpuset.c update_nodemask().
2186 * current's mempolicy may be rebinded by the other task(the task that changes
2187 * cpuset's mems), so we needn't do rebind work for current task.
2190 /* Slow path of a mempolicy duplicate */
2191 struct mempolicy
*__mpol_dup(struct mempolicy
*old
)
2193 struct mempolicy
*new = kmem_cache_alloc(policy_cache
, GFP_KERNEL
);
2196 return ERR_PTR(-ENOMEM
);
2198 /* task's mempolicy is protected by alloc_lock */
2199 if (old
== current
->mempolicy
) {
2202 task_unlock(current
);
2206 if (current_cpuset_is_being_rebound()) {
2207 nodemask_t mems
= cpuset_mems_allowed(current
);
2208 mpol_rebind_policy(new, &mems
);
2210 atomic_set(&new->refcnt
, 1);
2214 /* Slow path of a mempolicy comparison */
2215 bool __mpol_equal(struct mempolicy
*a
, struct mempolicy
*b
)
2219 if (a
->mode
!= b
->mode
)
2221 if (a
->flags
!= b
->flags
)
2223 if (mpol_store_user_nodemask(a
))
2224 if (!nodes_equal(a
->w
.user_nodemask
, b
->w
.user_nodemask
))
2230 case MPOL_INTERLEAVE
:
2231 return !!nodes_equal(a
->v
.nodes
, b
->v
.nodes
);
2232 case MPOL_PREFERRED
:
2233 /* a's ->flags is the same as b's */
2234 if (a
->flags
& MPOL_F_LOCAL
)
2236 return a
->v
.preferred_node
== b
->v
.preferred_node
;
2244 * Shared memory backing store policy support.
2246 * Remember policies even when nobody has shared memory mapped.
2247 * The policies are kept in Red-Black tree linked from the inode.
2248 * They are protected by the sp->lock rwlock, which should be held
2249 * for any accesses to the tree.
2253 * lookup first element intersecting start-end. Caller holds sp->lock for
2254 * reading or for writing
2256 static struct sp_node
*
2257 sp_lookup(struct shared_policy
*sp
, unsigned long start
, unsigned long end
)
2259 struct rb_node
*n
= sp
->root
.rb_node
;
2262 struct sp_node
*p
= rb_entry(n
, struct sp_node
, nd
);
2264 if (start
>= p
->end
)
2266 else if (end
<= p
->start
)
2274 struct sp_node
*w
= NULL
;
2275 struct rb_node
*prev
= rb_prev(n
);
2278 w
= rb_entry(prev
, struct sp_node
, nd
);
2279 if (w
->end
<= start
)
2283 return rb_entry(n
, struct sp_node
, nd
);
2287 * Insert a new shared policy into the list. Caller holds sp->lock for
2290 static void sp_insert(struct shared_policy
*sp
, struct sp_node
*new)
2292 struct rb_node
**p
= &sp
->root
.rb_node
;
2293 struct rb_node
*parent
= NULL
;
2298 nd
= rb_entry(parent
, struct sp_node
, nd
);
2299 if (new->start
< nd
->start
)
2301 else if (new->end
> nd
->end
)
2302 p
= &(*p
)->rb_right
;
2306 rb_link_node(&new->nd
, parent
, p
);
2307 rb_insert_color(&new->nd
, &sp
->root
);
2308 pr_debug("inserting %lx-%lx: %d\n", new->start
, new->end
,
2309 new->policy
? new->policy
->mode
: 0);
2312 /* Find shared policy intersecting idx */
2314 mpol_shared_policy_lookup(struct shared_policy
*sp
, unsigned long idx
)
2316 struct mempolicy
*pol
= NULL
;
2319 if (!sp
->root
.rb_node
)
2321 read_lock(&sp
->lock
);
2322 sn
= sp_lookup(sp
, idx
, idx
+1);
2324 mpol_get(sn
->policy
);
2327 read_unlock(&sp
->lock
);
2331 static void sp_free(struct sp_node
*n
)
2333 mpol_put(n
->policy
);
2334 kmem_cache_free(sn_cache
, n
);
2338 * mpol_misplaced - check whether current page node is valid in policy
2340 * @page: page to be checked
2341 * @vma: vm area where page mapped
2342 * @addr: virtual address where page mapped
2344 * Lookup current policy node id for vma,addr and "compare to" page's
2348 * -1 - not misplaced, page is in the right node
2349 * node - node id where the page should be
2351 * Policy determination "mimics" alloc_page_vma().
2352 * Called from fault path where we know the vma and faulting address.
2354 int mpol_misplaced(struct page
*page
, struct vm_area_struct
*vma
, unsigned long addr
)
2356 struct mempolicy
*pol
;
2358 int curnid
= page_to_nid(page
);
2359 unsigned long pgoff
;
2360 int thiscpu
= raw_smp_processor_id();
2361 int thisnid
= cpu_to_node(thiscpu
);
2365 pol
= get_vma_policy(vma
, addr
);
2366 if (!(pol
->flags
& MPOL_F_MOF
))
2369 switch (pol
->mode
) {
2370 case MPOL_INTERLEAVE
:
2371 pgoff
= vma
->vm_pgoff
;
2372 pgoff
+= (addr
- vma
->vm_start
) >> PAGE_SHIFT
;
2373 polnid
= offset_il_node(pol
, pgoff
);
2376 case MPOL_PREFERRED
:
2377 if (pol
->flags
& MPOL_F_LOCAL
)
2378 polnid
= numa_node_id();
2380 polnid
= pol
->v
.preferred_node
;
2386 * allows binding to multiple nodes.
2387 * use current page if in policy nodemask,
2388 * else select nearest allowed node, if any.
2389 * If no allowed nodes, use current [!misplaced].
2391 if (node_isset(curnid
, pol
->v
.nodes
))
2393 z
= first_zones_zonelist(
2394 node_zonelist(numa_node_id(), GFP_HIGHUSER
),
2395 gfp_zone(GFP_HIGHUSER
),
2397 polnid
= z
->zone
->node
;
2404 /* Migrate the page towards the node whose CPU is referencing it */
2405 if (pol
->flags
& MPOL_F_MORON
) {
2408 if (!should_numa_migrate_memory(current
, page
, curnid
, thiscpu
))
2412 if (curnid
!= polnid
)
2421 * Drop the (possibly final) reference to task->mempolicy. It needs to be
2422 * dropped after task->mempolicy is set to NULL so that any allocation done as
2423 * part of its kmem_cache_free(), such as by KASAN, doesn't reference a freed
2426 void mpol_put_task_policy(struct task_struct
*task
)
2428 struct mempolicy
*pol
;
2431 pol
= task
->mempolicy
;
2432 task
->mempolicy
= NULL
;
2437 static void sp_delete(struct shared_policy
*sp
, struct sp_node
*n
)
2439 pr_debug("deleting %lx-l%lx\n", n
->start
, n
->end
);
2440 rb_erase(&n
->nd
, &sp
->root
);
2444 static void sp_node_init(struct sp_node
*node
, unsigned long start
,
2445 unsigned long end
, struct mempolicy
*pol
)
2447 node
->start
= start
;
2452 static struct sp_node
*sp_alloc(unsigned long start
, unsigned long end
,
2453 struct mempolicy
*pol
)
2456 struct mempolicy
*newpol
;
2458 n
= kmem_cache_alloc(sn_cache
, GFP_KERNEL
);
2462 newpol
= mpol_dup(pol
);
2463 if (IS_ERR(newpol
)) {
2464 kmem_cache_free(sn_cache
, n
);
2467 newpol
->flags
|= MPOL_F_SHARED
;
2468 sp_node_init(n
, start
, end
, newpol
);
2473 /* Replace a policy range. */
2474 static int shared_policy_replace(struct shared_policy
*sp
, unsigned long start
,
2475 unsigned long end
, struct sp_node
*new)
2478 struct sp_node
*n_new
= NULL
;
2479 struct mempolicy
*mpol_new
= NULL
;
2483 write_lock(&sp
->lock
);
2484 n
= sp_lookup(sp
, start
, end
);
2485 /* Take care of old policies in the same range. */
2486 while (n
&& n
->start
< end
) {
2487 struct rb_node
*next
= rb_next(&n
->nd
);
2488 if (n
->start
>= start
) {
2494 /* Old policy spanning whole new range. */
2499 *mpol_new
= *n
->policy
;
2500 atomic_set(&mpol_new
->refcnt
, 1);
2501 sp_node_init(n_new
, end
, n
->end
, mpol_new
);
2503 sp_insert(sp
, n_new
);
2512 n
= rb_entry(next
, struct sp_node
, nd
);
2516 write_unlock(&sp
->lock
);
2523 kmem_cache_free(sn_cache
, n_new
);
2528 write_unlock(&sp
->lock
);
2530 n_new
= kmem_cache_alloc(sn_cache
, GFP_KERNEL
);
2533 mpol_new
= kmem_cache_alloc(policy_cache
, GFP_KERNEL
);
2540 * mpol_shared_policy_init - initialize shared policy for inode
2541 * @sp: pointer to inode shared policy
2542 * @mpol: struct mempolicy to install
2544 * Install non-NULL @mpol in inode's shared policy rb-tree.
2545 * On entry, the current task has a reference on a non-NULL @mpol.
2546 * This must be released on exit.
2547 * This is called at get_inode() calls and we can use GFP_KERNEL.
2549 void mpol_shared_policy_init(struct shared_policy
*sp
, struct mempolicy
*mpol
)
2553 sp
->root
= RB_ROOT
; /* empty tree == default mempolicy */
2554 rwlock_init(&sp
->lock
);
2557 struct vm_area_struct pvma
;
2558 struct mempolicy
*new;
2559 NODEMASK_SCRATCH(scratch
);
2563 /* contextualize the tmpfs mount point mempolicy */
2564 new = mpol_new(mpol
->mode
, mpol
->flags
, &mpol
->w
.user_nodemask
);
2566 goto free_scratch
; /* no valid nodemask intersection */
2569 ret
= mpol_set_nodemask(new, &mpol
->w
.user_nodemask
, scratch
);
2570 task_unlock(current
);
2574 /* Create pseudo-vma that contains just the policy */
2575 memset(&pvma
, 0, sizeof(struct vm_area_struct
));
2576 pvma
.vm_end
= TASK_SIZE
; /* policy covers entire file */
2577 mpol_set_shared_policy(sp
, &pvma
, new); /* adds ref */
2580 mpol_put(new); /* drop initial ref */
2582 NODEMASK_SCRATCH_FREE(scratch
);
2584 mpol_put(mpol
); /* drop our incoming ref on sb mpol */
2588 int mpol_set_shared_policy(struct shared_policy
*info
,
2589 struct vm_area_struct
*vma
, struct mempolicy
*npol
)
2592 struct sp_node
*new = NULL
;
2593 unsigned long sz
= vma_pages(vma
);
2595 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2597 sz
, npol
? npol
->mode
: -1,
2598 npol
? npol
->flags
: -1,
2599 npol
? nodes_addr(npol
->v
.nodes
)[0] : NUMA_NO_NODE
);
2602 new = sp_alloc(vma
->vm_pgoff
, vma
->vm_pgoff
+ sz
, npol
);
2606 err
= shared_policy_replace(info
, vma
->vm_pgoff
, vma
->vm_pgoff
+sz
, new);
2612 /* Free a backing policy store on inode delete. */
2613 void mpol_free_shared_policy(struct shared_policy
*p
)
2616 struct rb_node
*next
;
2618 if (!p
->root
.rb_node
)
2620 write_lock(&p
->lock
);
2621 next
= rb_first(&p
->root
);
2623 n
= rb_entry(next
, struct sp_node
, nd
);
2624 next
= rb_next(&n
->nd
);
2627 write_unlock(&p
->lock
);
2630 #ifdef CONFIG_NUMA_BALANCING
2631 static int __initdata numabalancing_override
;
2633 static void __init
check_numabalancing_enable(void)
2635 bool numabalancing_default
= false;
2637 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED
))
2638 numabalancing_default
= true;
2640 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2641 if (numabalancing_override
)
2642 set_numabalancing_state(numabalancing_override
== 1);
2644 if (num_online_nodes() > 1 && !numabalancing_override
) {
2645 pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
2646 numabalancing_default
? "Enabling" : "Disabling");
2647 set_numabalancing_state(numabalancing_default
);
2651 static int __init
setup_numabalancing(char *str
)
2657 if (!strcmp(str
, "enable")) {
2658 numabalancing_override
= 1;
2660 } else if (!strcmp(str
, "disable")) {
2661 numabalancing_override
= -1;
2666 pr_warn("Unable to parse numa_balancing=\n");
2670 __setup("numa_balancing=", setup_numabalancing
);
2672 static inline void __init
check_numabalancing_enable(void)
2675 #endif /* CONFIG_NUMA_BALANCING */
2677 /* assumes fs == KERNEL_DS */
2678 void __init
numa_policy_init(void)
2680 nodemask_t interleave_nodes
;
2681 unsigned long largest
= 0;
2682 int nid
, prefer
= 0;
2684 policy_cache
= kmem_cache_create("numa_policy",
2685 sizeof(struct mempolicy
),
2686 0, SLAB_PANIC
, NULL
);
2688 sn_cache
= kmem_cache_create("shared_policy_node",
2689 sizeof(struct sp_node
),
2690 0, SLAB_PANIC
, NULL
);
2692 for_each_node(nid
) {
2693 preferred_node_policy
[nid
] = (struct mempolicy
) {
2694 .refcnt
= ATOMIC_INIT(1),
2695 .mode
= MPOL_PREFERRED
,
2696 .flags
= MPOL_F_MOF
| MPOL_F_MORON
,
2697 .v
= { .preferred_node
= nid
, },
2702 * Set interleaving policy for system init. Interleaving is only
2703 * enabled across suitably sized nodes (default is >= 16MB), or
2704 * fall back to the largest node if they're all smaller.
2706 nodes_clear(interleave_nodes
);
2707 for_each_node_state(nid
, N_MEMORY
) {
2708 unsigned long total_pages
= node_present_pages(nid
);
2710 /* Preserve the largest node */
2711 if (largest
< total_pages
) {
2712 largest
= total_pages
;
2716 /* Interleave this node? */
2717 if ((total_pages
<< PAGE_SHIFT
) >= (16 << 20))
2718 node_set(nid
, interleave_nodes
);
2721 /* All too small, use the largest */
2722 if (unlikely(nodes_empty(interleave_nodes
)))
2723 node_set(prefer
, interleave_nodes
);
2725 if (do_set_mempolicy(MPOL_INTERLEAVE
, 0, &interleave_nodes
))
2726 pr_err("%s: interleaving failed\n", __func__
);
2728 check_numabalancing_enable();
2731 /* Reset policy of current process to default */
2732 void numa_default_policy(void)
2734 do_set_mempolicy(MPOL_DEFAULT
, 0, NULL
);
2738 * Parse and format mempolicy from/to strings
2742 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
2744 static const char * const policy_modes
[] =
2746 [MPOL_DEFAULT
] = "default",
2747 [MPOL_PREFERRED
] = "prefer",
2748 [MPOL_BIND
] = "bind",
2749 [MPOL_INTERLEAVE
] = "interleave",
2750 [MPOL_LOCAL
] = "local",
2756 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2757 * @str: string containing mempolicy to parse
2758 * @mpol: pointer to struct mempolicy pointer, returned on success.
2761 * <mode>[=<flags>][:<nodelist>]
2763 * On success, returns 0, else 1
2765 int mpol_parse_str(char *str
, struct mempolicy
**mpol
)
2767 struct mempolicy
*new = NULL
;
2768 unsigned short mode
;
2769 unsigned short mode_flags
;
2771 char *nodelist
= strchr(str
, ':');
2772 char *flags
= strchr(str
, '=');
2776 *flags
++ = '\0'; /* terminate mode string */
2779 /* NUL-terminate mode or flags string */
2781 if (nodelist_parse(nodelist
, nodes
))
2783 if (!nodes_subset(nodes
, node_states
[N_MEMORY
]))
2788 for (mode
= 0; mode
< MPOL_MAX
; mode
++) {
2789 if (!strcmp(str
, policy_modes
[mode
])) {
2793 if (mode
>= MPOL_MAX
)
2797 case MPOL_PREFERRED
:
2799 * Insist on a nodelist of one node only
2802 char *rest
= nodelist
;
2803 while (isdigit(*rest
))
2809 case MPOL_INTERLEAVE
:
2811 * Default to online nodes with memory if no nodelist
2814 nodes
= node_states
[N_MEMORY
];
2818 * Don't allow a nodelist; mpol_new() checks flags
2822 mode
= MPOL_PREFERRED
;
2826 * Insist on a empty nodelist
2833 * Insist on a nodelist
2842 * Currently, we only support two mutually exclusive
2845 if (!strcmp(flags
, "static"))
2846 mode_flags
|= MPOL_F_STATIC_NODES
;
2847 else if (!strcmp(flags
, "relative"))
2848 mode_flags
|= MPOL_F_RELATIVE_NODES
;
2853 new = mpol_new(mode
, mode_flags
, &nodes
);
2858 * Save nodes for mpol_to_str() to show the tmpfs mount options
2859 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2861 if (mode
!= MPOL_PREFERRED
)
2862 new->v
.nodes
= nodes
;
2864 new->v
.preferred_node
= first_node(nodes
);
2866 new->flags
|= MPOL_F_LOCAL
;
2869 * Save nodes for contextualization: this will be used to "clone"
2870 * the mempolicy in a specific context [cpuset] at a later time.
2872 new->w
.user_nodemask
= nodes
;
2877 /* Restore string for error message */
2886 #endif /* CONFIG_TMPFS */
2889 * mpol_to_str - format a mempolicy structure for printing
2890 * @buffer: to contain formatted mempolicy string
2891 * @maxlen: length of @buffer
2892 * @pol: pointer to mempolicy to be formatted
2894 * Convert @pol into a string. If @buffer is too short, truncate the string.
2895 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
2896 * longest flag, "relative", and to display at least a few node ids.
2898 void mpol_to_str(char *buffer
, int maxlen
, struct mempolicy
*pol
)
2901 nodemask_t nodes
= NODE_MASK_NONE
;
2902 unsigned short mode
= MPOL_DEFAULT
;
2903 unsigned short flags
= 0;
2905 if (pol
&& pol
!= &default_policy
&& !(pol
->flags
& MPOL_F_MORON
)) {
2913 case MPOL_PREFERRED
:
2914 if (flags
& MPOL_F_LOCAL
)
2917 node_set(pol
->v
.preferred_node
, nodes
);
2920 case MPOL_INTERLEAVE
:
2921 nodes
= pol
->v
.nodes
;
2925 snprintf(p
, maxlen
, "unknown");
2929 p
+= snprintf(p
, maxlen
, "%s", policy_modes
[mode
]);
2931 if (flags
& MPOL_MODE_FLAGS
) {
2932 p
+= snprintf(p
, buffer
+ maxlen
- p
, "=");
2935 * Currently, the only defined flags are mutually exclusive
2937 if (flags
& MPOL_F_STATIC_NODES
)
2938 p
+= snprintf(p
, buffer
+ maxlen
- p
, "static");
2939 else if (flags
& MPOL_F_RELATIVE_NODES
)
2940 p
+= snprintf(p
, buffer
+ maxlen
- p
, "relative");
2943 if (!nodes_empty(nodes
))
2944 p
+= scnprintf(p
, buffer
+ maxlen
- p
, ":%*pbl",
2945 nodemask_pr_args(&nodes
));