1 // SPDX-License-Identifier: GPL-2.0-only
3 * Simple NUMA memory policy for the Linux kernel.
5 * Copyright 2003,2004 Andi Kleen, SuSE Labs.
6 * (C) Copyright 2005 Christoph Lameter, Silicon Graphics, Inc.
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>
71 #include <linux/pagewalk.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 */
127 static struct mempolicy preferred_node_policy
[MAX_NUMNODES
];
130 * numa_map_to_online_node - Find closest online node
131 * @node: Node id to start the search
133 * Lookup the next closest node by distance if @nid is not online.
135 int numa_map_to_online_node(int node
)
137 int min_dist
= INT_MAX
, dist
, n
, min_node
;
139 if (node
== NUMA_NO_NODE
|| node_online(node
))
143 for_each_online_node(n
) {
144 dist
= node_distance(node
, n
);
145 if (dist
< min_dist
) {
153 EXPORT_SYMBOL_GPL(numa_map_to_online_node
);
155 struct mempolicy
*get_task_policy(struct task_struct
*p
)
157 struct mempolicy
*pol
= p
->mempolicy
;
163 node
= numa_node_id();
164 if (node
!= NUMA_NO_NODE
) {
165 pol
= &preferred_node_policy
[node
];
166 /* preferred_node_policy is not initialised early in boot */
171 return &default_policy
;
174 static const struct mempolicy_operations
{
175 int (*create
)(struct mempolicy
*pol
, const nodemask_t
*nodes
);
176 void (*rebind
)(struct mempolicy
*pol
, const nodemask_t
*nodes
);
177 } mpol_ops
[MPOL_MAX
];
179 static inline int mpol_store_user_nodemask(const struct mempolicy
*pol
)
181 return pol
->flags
& MPOL_MODE_FLAGS
;
184 static void mpol_relative_nodemask(nodemask_t
*ret
, const nodemask_t
*orig
,
185 const nodemask_t
*rel
)
188 nodes_fold(tmp
, *orig
, nodes_weight(*rel
));
189 nodes_onto(*ret
, tmp
, *rel
);
192 static int mpol_new_interleave(struct mempolicy
*pol
, const nodemask_t
*nodes
)
194 if (nodes_empty(*nodes
))
200 static int mpol_new_preferred(struct mempolicy
*pol
, const nodemask_t
*nodes
)
202 if (nodes_empty(*nodes
))
205 nodes_clear(pol
->nodes
);
206 node_set(first_node(*nodes
), pol
->nodes
);
210 static int mpol_new_bind(struct mempolicy
*pol
, const nodemask_t
*nodes
)
212 if (nodes_empty(*nodes
))
219 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
220 * any, for the new policy. mpol_new() has already validated the nodes
221 * parameter with respect to the policy mode and flags.
223 * Must be called holding task's alloc_lock to protect task's mems_allowed
224 * and mempolicy. May also be called holding the mmap_lock for write.
226 static int mpol_set_nodemask(struct mempolicy
*pol
,
227 const nodemask_t
*nodes
, struct nodemask_scratch
*nsc
)
232 * Default (pol==NULL) resp. local memory policies are not a
233 * subject of any remapping. They also do not need any special
236 if (!pol
|| pol
->mode
== MPOL_LOCAL
)
240 nodes_and(nsc
->mask1
,
241 cpuset_current_mems_allowed
, node_states
[N_MEMORY
]);
245 if (pol
->flags
& MPOL_F_RELATIVE_NODES
)
246 mpol_relative_nodemask(&nsc
->mask2
, nodes
, &nsc
->mask1
);
248 nodes_and(nsc
->mask2
, *nodes
, nsc
->mask1
);
250 if (mpol_store_user_nodemask(pol
))
251 pol
->w
.user_nodemask
= *nodes
;
253 pol
->w
.cpuset_mems_allowed
= cpuset_current_mems_allowed
;
255 ret
= mpol_ops
[pol
->mode
].create(pol
, &nsc
->mask2
);
260 * This function just creates a new policy, does some check and simple
261 * initialization. You must invoke mpol_set_nodemask() to set nodes.
263 static struct mempolicy
*mpol_new(unsigned short mode
, unsigned short flags
,
266 struct mempolicy
*policy
;
268 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
269 mode
, flags
, nodes
? nodes_addr(*nodes
)[0] : NUMA_NO_NODE
);
271 if (mode
== MPOL_DEFAULT
) {
272 if (nodes
&& !nodes_empty(*nodes
))
273 return ERR_PTR(-EINVAL
);
279 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
280 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
281 * All other modes require a valid pointer to a non-empty nodemask.
283 if (mode
== MPOL_PREFERRED
) {
284 if (nodes_empty(*nodes
)) {
285 if (((flags
& MPOL_F_STATIC_NODES
) ||
286 (flags
& MPOL_F_RELATIVE_NODES
)))
287 return ERR_PTR(-EINVAL
);
291 } else if (mode
== MPOL_LOCAL
) {
292 if (!nodes_empty(*nodes
) ||
293 (flags
& MPOL_F_STATIC_NODES
) ||
294 (flags
& MPOL_F_RELATIVE_NODES
))
295 return ERR_PTR(-EINVAL
);
296 } else if (nodes_empty(*nodes
))
297 return ERR_PTR(-EINVAL
);
298 policy
= kmem_cache_alloc(policy_cache
, GFP_KERNEL
);
300 return ERR_PTR(-ENOMEM
);
301 atomic_set(&policy
->refcnt
, 1);
303 policy
->flags
= flags
;
308 /* Slow path of a mpol destructor. */
309 void __mpol_put(struct mempolicy
*p
)
311 if (!atomic_dec_and_test(&p
->refcnt
))
313 kmem_cache_free(policy_cache
, p
);
316 static void mpol_rebind_default(struct mempolicy
*pol
, const nodemask_t
*nodes
)
320 static void mpol_rebind_nodemask(struct mempolicy
*pol
, const nodemask_t
*nodes
)
324 if (pol
->flags
& MPOL_F_STATIC_NODES
)
325 nodes_and(tmp
, pol
->w
.user_nodemask
, *nodes
);
326 else if (pol
->flags
& MPOL_F_RELATIVE_NODES
)
327 mpol_relative_nodemask(&tmp
, &pol
->w
.user_nodemask
, nodes
);
329 nodes_remap(tmp
, pol
->nodes
, pol
->w
.cpuset_mems_allowed
,
331 pol
->w
.cpuset_mems_allowed
= *nodes
;
334 if (nodes_empty(tmp
))
340 static void mpol_rebind_preferred(struct mempolicy
*pol
,
341 const nodemask_t
*nodes
)
343 pol
->w
.cpuset_mems_allowed
= *nodes
;
347 * mpol_rebind_policy - Migrate a policy to a different set of nodes
349 * Per-vma policies are protected by mmap_lock. Allocations using per-task
350 * policies are protected by task->mems_allowed_seq to prevent a premature
351 * OOM/allocation failure due to parallel nodemask modification.
353 static void mpol_rebind_policy(struct mempolicy
*pol
, const nodemask_t
*newmask
)
357 if (!mpol_store_user_nodemask(pol
) &&
358 nodes_equal(pol
->w
.cpuset_mems_allowed
, *newmask
))
361 mpol_ops
[pol
->mode
].rebind(pol
, newmask
);
365 * Wrapper for mpol_rebind_policy() that just requires task
366 * pointer, and updates task mempolicy.
368 * Called with task's alloc_lock held.
371 void mpol_rebind_task(struct task_struct
*tsk
, const nodemask_t
*new)
373 mpol_rebind_policy(tsk
->mempolicy
, new);
377 * Rebind each vma in mm to new nodemask.
379 * Call holding a reference to mm. Takes mm->mmap_lock during call.
382 void mpol_rebind_mm(struct mm_struct
*mm
, nodemask_t
*new)
384 struct vm_area_struct
*vma
;
387 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
)
388 mpol_rebind_policy(vma
->vm_policy
, new);
389 mmap_write_unlock(mm
);
392 static const struct mempolicy_operations mpol_ops
[MPOL_MAX
] = {
394 .rebind
= mpol_rebind_default
,
396 [MPOL_INTERLEAVE
] = {
397 .create
= mpol_new_interleave
,
398 .rebind
= mpol_rebind_nodemask
,
401 .create
= mpol_new_preferred
,
402 .rebind
= mpol_rebind_preferred
,
405 .create
= mpol_new_bind
,
406 .rebind
= mpol_rebind_nodemask
,
409 .rebind
= mpol_rebind_default
,
413 static int migrate_page_add(struct page
*page
, struct list_head
*pagelist
,
414 unsigned long flags
);
417 struct list_head
*pagelist
;
422 struct vm_area_struct
*first
;
426 * Check if the page's nid is in qp->nmask.
428 * If MPOL_MF_INVERT is set in qp->flags, check if the nid is
429 * in the invert of qp->nmask.
431 static inline bool queue_pages_required(struct page
*page
,
432 struct queue_pages
*qp
)
434 int nid
= page_to_nid(page
);
435 unsigned long flags
= qp
->flags
;
437 return node_isset(nid
, *qp
->nmask
) == !(flags
& MPOL_MF_INVERT
);
441 * queue_pages_pmd() has four possible return values:
442 * 0 - pages are placed on the right node or queued successfully, or
443 * special page is met, i.e. huge zero page.
444 * 1 - there is unmovable page, and MPOL_MF_MOVE* & MPOL_MF_STRICT were
447 * -EIO - is migration entry or only MPOL_MF_STRICT was specified and an
448 * existing page was already on a node that does not follow the
451 static int queue_pages_pmd(pmd_t
*pmd
, spinlock_t
*ptl
, unsigned long addr
,
452 unsigned long end
, struct mm_walk
*walk
)
457 struct queue_pages
*qp
= walk
->private;
460 if (unlikely(is_pmd_migration_entry(*pmd
))) {
464 page
= pmd_page(*pmd
);
465 if (is_huge_zero_page(page
)) {
467 walk
->action
= ACTION_CONTINUE
;
470 if (!queue_pages_required(page
, qp
))
474 /* go to thp migration */
475 if (flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
)) {
476 if (!vma_migratable(walk
->vma
) ||
477 migrate_page_add(page
, qp
->pagelist
, flags
)) {
490 * Scan through pages checking if pages follow certain conditions,
491 * and move them to the pagelist if they do.
493 * queue_pages_pte_range() has three possible return values:
494 * 0 - pages are placed on the right node or queued successfully, or
495 * special page is met, i.e. zero page.
496 * 1 - there is unmovable page, and MPOL_MF_MOVE* & MPOL_MF_STRICT were
498 * -EIO - only MPOL_MF_STRICT was specified and an existing page was already
499 * on a node that does not follow the policy.
501 static int queue_pages_pte_range(pmd_t
*pmd
, unsigned long addr
,
502 unsigned long end
, struct mm_walk
*walk
)
504 struct vm_area_struct
*vma
= walk
->vma
;
506 struct queue_pages
*qp
= walk
->private;
507 unsigned long flags
= qp
->flags
;
509 bool has_unmovable
= false;
510 pte_t
*pte
, *mapped_pte
;
513 ptl
= pmd_trans_huge_lock(pmd
, vma
);
515 ret
= queue_pages_pmd(pmd
, ptl
, addr
, end
, walk
);
519 /* THP was split, fall through to pte walk */
521 if (pmd_trans_unstable(pmd
))
524 mapped_pte
= pte
= pte_offset_map_lock(walk
->mm
, pmd
, addr
, &ptl
);
525 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
) {
526 if (!pte_present(*pte
))
528 page
= vm_normal_page(vma
, addr
, *pte
);
532 * vm_normal_page() filters out zero pages, but there might
533 * still be PageReserved pages to skip, perhaps in a VDSO.
535 if (PageReserved(page
))
537 if (!queue_pages_required(page
, qp
))
539 if (flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
)) {
540 /* MPOL_MF_STRICT must be specified if we get here */
541 if (!vma_migratable(vma
)) {
542 has_unmovable
= true;
547 * Do not abort immediately since there may be
548 * temporary off LRU pages in the range. Still
549 * need migrate other LRU pages.
551 if (migrate_page_add(page
, qp
->pagelist
, flags
))
552 has_unmovable
= true;
556 pte_unmap_unlock(mapped_pte
, ptl
);
562 return addr
!= end
? -EIO
: 0;
565 static int queue_pages_hugetlb(pte_t
*pte
, unsigned long hmask
,
566 unsigned long addr
, unsigned long end
,
567 struct mm_walk
*walk
)
570 #ifdef CONFIG_HUGETLB_PAGE
571 struct queue_pages
*qp
= walk
->private;
572 unsigned long flags
= (qp
->flags
& MPOL_MF_VALID
);
577 ptl
= huge_pte_lock(hstate_vma(walk
->vma
), walk
->mm
, pte
);
578 entry
= huge_ptep_get(pte
);
579 if (!pte_present(entry
))
581 page
= pte_page(entry
);
582 if (!queue_pages_required(page
, qp
))
585 if (flags
== MPOL_MF_STRICT
) {
587 * STRICT alone means only detecting misplaced page and no
588 * need to further check other vma.
594 if (!vma_migratable(walk
->vma
)) {
596 * Must be STRICT with MOVE*, otherwise .test_walk() have
597 * stopped walking current vma.
598 * Detecting misplaced page but allow migrating pages which
605 /* With MPOL_MF_MOVE, we migrate only unshared hugepage. */
606 if (flags
& (MPOL_MF_MOVE_ALL
) ||
607 (flags
& MPOL_MF_MOVE
&& page_mapcount(page
) == 1)) {
608 if (!isolate_huge_page(page
, qp
->pagelist
) &&
609 (flags
& MPOL_MF_STRICT
))
611 * Failed to isolate page but allow migrating pages
612 * which have been queued.
624 #ifdef CONFIG_NUMA_BALANCING
626 * This is used to mark a range of virtual addresses to be inaccessible.
627 * These are later cleared by a NUMA hinting fault. Depending on these
628 * faults, pages may be migrated for better NUMA placement.
630 * This is assuming that NUMA faults are handled using PROT_NONE. If
631 * an architecture makes a different choice, it will need further
632 * changes to the core.
634 unsigned long change_prot_numa(struct vm_area_struct
*vma
,
635 unsigned long addr
, unsigned long end
)
639 nr_updated
= change_protection(vma
, addr
, end
, PAGE_NONE
, MM_CP_PROT_NUMA
);
641 count_vm_numa_events(NUMA_PTE_UPDATES
, nr_updated
);
646 static unsigned long change_prot_numa(struct vm_area_struct
*vma
,
647 unsigned long addr
, unsigned long end
)
651 #endif /* CONFIG_NUMA_BALANCING */
653 static int queue_pages_test_walk(unsigned long start
, unsigned long end
,
654 struct mm_walk
*walk
)
656 struct vm_area_struct
*vma
= walk
->vma
;
657 struct queue_pages
*qp
= walk
->private;
658 unsigned long endvma
= vma
->vm_end
;
659 unsigned long flags
= qp
->flags
;
661 /* range check first */
662 VM_BUG_ON_VMA(!range_in_vma(vma
, start
, end
), vma
);
666 if (!(flags
& MPOL_MF_DISCONTIG_OK
) &&
667 (qp
->start
< vma
->vm_start
))
668 /* hole at head side of range */
671 if (!(flags
& MPOL_MF_DISCONTIG_OK
) &&
672 ((vma
->vm_end
< qp
->end
) &&
673 (!vma
->vm_next
|| vma
->vm_end
< vma
->vm_next
->vm_start
)))
674 /* hole at middle or tail of range */
678 * Need check MPOL_MF_STRICT to return -EIO if possible
679 * regardless of vma_migratable
681 if (!vma_migratable(vma
) &&
682 !(flags
& MPOL_MF_STRICT
))
688 if (flags
& MPOL_MF_LAZY
) {
689 /* Similar to task_numa_work, skip inaccessible VMAs */
690 if (!is_vm_hugetlb_page(vma
) && vma_is_accessible(vma
) &&
691 !(vma
->vm_flags
& VM_MIXEDMAP
))
692 change_prot_numa(vma
, start
, endvma
);
696 /* queue pages from current vma */
697 if (flags
& MPOL_MF_VALID
)
702 static const struct mm_walk_ops queue_pages_walk_ops
= {
703 .hugetlb_entry
= queue_pages_hugetlb
,
704 .pmd_entry
= queue_pages_pte_range
,
705 .test_walk
= queue_pages_test_walk
,
709 * Walk through page tables and collect pages to be migrated.
711 * If pages found in a given range are on a set of nodes (determined by
712 * @nodes and @flags,) it's isolated and queued to the pagelist which is
713 * passed via @private.
715 * queue_pages_range() has three possible return values:
716 * 1 - there is unmovable page, but MPOL_MF_MOVE* & MPOL_MF_STRICT were
718 * 0 - queue pages successfully or no misplaced page.
719 * errno - i.e. misplaced pages with MPOL_MF_STRICT specified (-EIO) or
720 * memory range specified by nodemask and maxnode points outside
721 * your accessible address space (-EFAULT)
724 queue_pages_range(struct mm_struct
*mm
, unsigned long start
, unsigned long end
,
725 nodemask_t
*nodes
, unsigned long flags
,
726 struct list_head
*pagelist
)
729 struct queue_pages qp
= {
730 .pagelist
= pagelist
,
738 err
= walk_page_range(mm
, start
, end
, &queue_pages_walk_ops
, &qp
);
741 /* whole range in hole */
748 * Apply policy to a single VMA
749 * This must be called with the mmap_lock held for writing.
751 static int vma_replace_policy(struct vm_area_struct
*vma
,
752 struct mempolicy
*pol
)
755 struct mempolicy
*old
;
756 struct mempolicy
*new;
758 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
759 vma
->vm_start
, vma
->vm_end
, vma
->vm_pgoff
,
760 vma
->vm_ops
, vma
->vm_file
,
761 vma
->vm_ops
? vma
->vm_ops
->set_policy
: NULL
);
767 if (vma
->vm_ops
&& vma
->vm_ops
->set_policy
) {
768 err
= vma
->vm_ops
->set_policy(vma
, new);
773 old
= vma
->vm_policy
;
774 vma
->vm_policy
= new; /* protected by mmap_lock */
783 /* Step 2: apply policy to a range and do splits. */
784 static int mbind_range(struct mm_struct
*mm
, unsigned long start
,
785 unsigned long end
, struct mempolicy
*new_pol
)
787 struct vm_area_struct
*next
;
788 struct vm_area_struct
*prev
;
789 struct vm_area_struct
*vma
;
792 unsigned long vmstart
;
795 vma
= find_vma(mm
, start
);
799 if (start
> vma
->vm_start
)
802 for (; vma
&& vma
->vm_start
< end
; prev
= vma
, vma
= next
) {
804 vmstart
= max(start
, vma
->vm_start
);
805 vmend
= min(end
, vma
->vm_end
);
807 if (mpol_equal(vma_policy(vma
), new_pol
))
810 pgoff
= vma
->vm_pgoff
+
811 ((vmstart
- vma
->vm_start
) >> PAGE_SHIFT
);
812 prev
= vma_merge(mm
, prev
, vmstart
, vmend
, vma
->vm_flags
,
813 vma
->anon_vma
, vma
->vm_file
, pgoff
,
814 new_pol
, vma
->vm_userfaultfd_ctx
);
818 if (mpol_equal(vma_policy(vma
), new_pol
))
820 /* vma_merge() joined vma && vma->next, case 8 */
823 if (vma
->vm_start
!= vmstart
) {
824 err
= split_vma(vma
->vm_mm
, vma
, vmstart
, 1);
828 if (vma
->vm_end
!= vmend
) {
829 err
= split_vma(vma
->vm_mm
, vma
, vmend
, 0);
834 err
= vma_replace_policy(vma
, new_pol
);
843 /* Set the process memory policy */
844 static long do_set_mempolicy(unsigned short mode
, unsigned short flags
,
847 struct mempolicy
*new, *old
;
848 NODEMASK_SCRATCH(scratch
);
854 new = mpol_new(mode
, flags
, nodes
);
860 if (flags
& MPOL_F_NUMA_BALANCING
) {
861 if (new && new->mode
== MPOL_BIND
) {
862 new->flags
|= (MPOL_F_MOF
| MPOL_F_MORON
);
870 ret
= mpol_set_nodemask(new, nodes
, scratch
);
876 old
= current
->mempolicy
;
877 current
->mempolicy
= new;
878 if (new && new->mode
== MPOL_INTERLEAVE
)
879 current
->il_prev
= MAX_NUMNODES
-1;
880 task_unlock(current
);
884 NODEMASK_SCRATCH_FREE(scratch
);
889 * Return nodemask for policy for get_mempolicy() query
891 * Called with task's alloc_lock held
893 static void get_policy_nodemask(struct mempolicy
*p
, nodemask_t
*nodes
)
896 if (p
== &default_policy
)
901 case MPOL_INTERLEAVE
:
906 /* return empty node mask for local allocation */
913 static int lookup_node(struct mm_struct
*mm
, unsigned long addr
)
915 struct page
*p
= NULL
;
919 err
= get_user_pages_locked(addr
& PAGE_MASK
, 1, 0, &p
, &locked
);
921 err
= page_to_nid(p
);
925 mmap_read_unlock(mm
);
929 /* Retrieve NUMA policy */
930 static long do_get_mempolicy(int *policy
, nodemask_t
*nmask
,
931 unsigned long addr
, unsigned long flags
)
934 struct mm_struct
*mm
= current
->mm
;
935 struct vm_area_struct
*vma
= NULL
;
936 struct mempolicy
*pol
= current
->mempolicy
, *pol_refcount
= NULL
;
939 ~(unsigned long)(MPOL_F_NODE
|MPOL_F_ADDR
|MPOL_F_MEMS_ALLOWED
))
942 if (flags
& MPOL_F_MEMS_ALLOWED
) {
943 if (flags
& (MPOL_F_NODE
|MPOL_F_ADDR
))
945 *policy
= 0; /* just so it's initialized */
947 *nmask
= cpuset_current_mems_allowed
;
948 task_unlock(current
);
952 if (flags
& MPOL_F_ADDR
) {
954 * Do NOT fall back to task policy if the
955 * vma/shared policy at addr is NULL. We
956 * want to return MPOL_DEFAULT in this case.
959 vma
= vma_lookup(mm
, addr
);
961 mmap_read_unlock(mm
);
964 if (vma
->vm_ops
&& vma
->vm_ops
->get_policy
)
965 pol
= vma
->vm_ops
->get_policy(vma
, addr
);
967 pol
= vma
->vm_policy
;
972 pol
= &default_policy
; /* indicates default behavior */
974 if (flags
& MPOL_F_NODE
) {
975 if (flags
& MPOL_F_ADDR
) {
977 * Take a refcount on the mpol, lookup_node()
978 * will drop the mmap_lock, so after calling
979 * lookup_node() only "pol" remains valid, "vma"
985 err
= lookup_node(mm
, addr
);
989 } else if (pol
== current
->mempolicy
&&
990 pol
->mode
== MPOL_INTERLEAVE
) {
991 *policy
= next_node_in(current
->il_prev
, pol
->nodes
);
997 *policy
= pol
== &default_policy
? MPOL_DEFAULT
:
1000 * Internal mempolicy flags must be masked off before exposing
1001 * the policy to userspace.
1003 *policy
|= (pol
->flags
& MPOL_MODE_FLAGS
);
1008 if (mpol_store_user_nodemask(pol
)) {
1009 *nmask
= pol
->w
.user_nodemask
;
1012 get_policy_nodemask(pol
, nmask
);
1013 task_unlock(current
);
1020 mmap_read_unlock(mm
);
1022 mpol_put(pol_refcount
);
1026 #ifdef CONFIG_MIGRATION
1028 * page migration, thp tail pages can be passed.
1030 static int migrate_page_add(struct page
*page
, struct list_head
*pagelist
,
1031 unsigned long flags
)
1033 struct page
*head
= compound_head(page
);
1035 * Avoid migrating a page that is shared with others.
1037 if ((flags
& MPOL_MF_MOVE_ALL
) || page_mapcount(head
) == 1) {
1038 if (!isolate_lru_page(head
)) {
1039 list_add_tail(&head
->lru
, pagelist
);
1040 mod_node_page_state(page_pgdat(head
),
1041 NR_ISOLATED_ANON
+ page_is_file_lru(head
),
1042 thp_nr_pages(head
));
1043 } else if (flags
& MPOL_MF_STRICT
) {
1045 * Non-movable page may reach here. And, there may be
1046 * temporary off LRU pages or non-LRU movable pages.
1047 * Treat them as unmovable pages since they can't be
1048 * isolated, so they can't be moved at the moment. It
1049 * should return -EIO for this case too.
1059 * Migrate pages from one node to a target node.
1060 * Returns error or the number of pages not migrated.
1062 static int migrate_to_node(struct mm_struct
*mm
, int source
, int dest
,
1066 LIST_HEAD(pagelist
);
1068 struct migration_target_control mtc
= {
1070 .gfp_mask
= GFP_HIGHUSER_MOVABLE
| __GFP_THISNODE
,
1074 node_set(source
, nmask
);
1077 * This does not "check" the range but isolates all pages that
1078 * need migration. Between passing in the full user address
1079 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1081 VM_BUG_ON(!(flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
)));
1082 queue_pages_range(mm
, mm
->mmap
->vm_start
, mm
->task_size
, &nmask
,
1083 flags
| MPOL_MF_DISCONTIG_OK
, &pagelist
);
1085 if (!list_empty(&pagelist
)) {
1086 err
= migrate_pages(&pagelist
, alloc_migration_target
, NULL
,
1087 (unsigned long)&mtc
, MIGRATE_SYNC
, MR_SYSCALL
);
1089 putback_movable_pages(&pagelist
);
1096 * Move pages between the two nodesets so as to preserve the physical
1097 * layout as much as possible.
1099 * Returns the number of page that could not be moved.
1101 int do_migrate_pages(struct mm_struct
*mm
, const nodemask_t
*from
,
1102 const nodemask_t
*to
, int flags
)
1108 lru_cache_disable();
1113 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1114 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1115 * bit in 'tmp', and return that <source, dest> pair for migration.
1116 * The pair of nodemasks 'to' and 'from' define the map.
1118 * If no pair of bits is found that way, fallback to picking some
1119 * pair of 'source' and 'dest' bits that are not the same. If the
1120 * 'source' and 'dest' bits are the same, this represents a node
1121 * that will be migrating to itself, so no pages need move.
1123 * If no bits are left in 'tmp', or if all remaining bits left
1124 * in 'tmp' correspond to the same bit in 'to', return false
1125 * (nothing left to migrate).
1127 * This lets us pick a pair of nodes to migrate between, such that
1128 * if possible the dest node is not already occupied by some other
1129 * source node, minimizing the risk of overloading the memory on a
1130 * node that would happen if we migrated incoming memory to a node
1131 * before migrating outgoing memory source that same node.
1133 * A single scan of tmp is sufficient. As we go, we remember the
1134 * most recent <s, d> pair that moved (s != d). If we find a pair
1135 * that not only moved, but what's better, moved to an empty slot
1136 * (d is not set in tmp), then we break out then, with that pair.
1137 * Otherwise when we finish scanning from_tmp, we at least have the
1138 * most recent <s, d> pair that moved. If we get all the way through
1139 * the scan of tmp without finding any node that moved, much less
1140 * moved to an empty node, then there is nothing left worth migrating.
1144 while (!nodes_empty(tmp
)) {
1146 int source
= NUMA_NO_NODE
;
1149 for_each_node_mask(s
, tmp
) {
1152 * do_migrate_pages() tries to maintain the relative
1153 * node relationship of the pages established between
1154 * threads and memory areas.
1156 * However if the number of source nodes is not equal to
1157 * the number of destination nodes we can not preserve
1158 * this node relative relationship. In that case, skip
1159 * copying memory from a node that is in the destination
1162 * Example: [2,3,4] -> [3,4,5] moves everything.
1163 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1166 if ((nodes_weight(*from
) != nodes_weight(*to
)) &&
1167 (node_isset(s
, *to
)))
1170 d
= node_remap(s
, *from
, *to
);
1174 source
= s
; /* Node moved. Memorize */
1177 /* dest not in remaining from nodes? */
1178 if (!node_isset(dest
, tmp
))
1181 if (source
== NUMA_NO_NODE
)
1184 node_clear(source
, tmp
);
1185 err
= migrate_to_node(mm
, source
, dest
, flags
);
1191 mmap_read_unlock(mm
);
1201 * Allocate a new page for page migration based on vma policy.
1202 * Start by assuming the page is mapped by the same vma as contains @start.
1203 * Search forward from there, if not. N.B., this assumes that the
1204 * list of pages handed to migrate_pages()--which is how we get here--
1205 * is in virtual address order.
1207 static struct page
*new_page(struct page
*page
, unsigned long start
)
1209 struct vm_area_struct
*vma
;
1210 unsigned long address
;
1212 vma
= find_vma(current
->mm
, start
);
1214 address
= page_address_in_vma(page
, vma
);
1215 if (address
!= -EFAULT
)
1220 if (PageHuge(page
)) {
1221 return alloc_huge_page_vma(page_hstate(compound_head(page
)),
1223 } else if (PageTransHuge(page
)) {
1226 thp
= alloc_hugepage_vma(GFP_TRANSHUGE
, vma
, address
,
1230 prep_transhuge_page(thp
);
1234 * if !vma, alloc_page_vma() will use task or system default policy
1236 return alloc_page_vma(GFP_HIGHUSER_MOVABLE
| __GFP_RETRY_MAYFAIL
,
1241 static int migrate_page_add(struct page
*page
, struct list_head
*pagelist
,
1242 unsigned long flags
)
1247 int do_migrate_pages(struct mm_struct
*mm
, const nodemask_t
*from
,
1248 const nodemask_t
*to
, int flags
)
1253 static struct page
*new_page(struct page
*page
, unsigned long start
)
1259 static long do_mbind(unsigned long start
, unsigned long len
,
1260 unsigned short mode
, unsigned short mode_flags
,
1261 nodemask_t
*nmask
, unsigned long flags
)
1263 struct mm_struct
*mm
= current
->mm
;
1264 struct mempolicy
*new;
1268 LIST_HEAD(pagelist
);
1270 if (flags
& ~(unsigned long)MPOL_MF_VALID
)
1272 if ((flags
& MPOL_MF_MOVE_ALL
) && !capable(CAP_SYS_NICE
))
1275 if (start
& ~PAGE_MASK
)
1278 if (mode
== MPOL_DEFAULT
)
1279 flags
&= ~MPOL_MF_STRICT
;
1281 len
= (len
+ PAGE_SIZE
- 1) & PAGE_MASK
;
1289 new = mpol_new(mode
, mode_flags
, nmask
);
1291 return PTR_ERR(new);
1293 if (flags
& MPOL_MF_LAZY
)
1294 new->flags
|= MPOL_F_MOF
;
1297 * If we are using the default policy then operation
1298 * on discontinuous address spaces is okay after all
1301 flags
|= MPOL_MF_DISCONTIG_OK
;
1303 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1304 start
, start
+ len
, mode
, mode_flags
,
1305 nmask
? nodes_addr(*nmask
)[0] : NUMA_NO_NODE
);
1307 if (flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
)) {
1309 lru_cache_disable();
1312 NODEMASK_SCRATCH(scratch
);
1314 mmap_write_lock(mm
);
1315 err
= mpol_set_nodemask(new, nmask
, scratch
);
1317 mmap_write_unlock(mm
);
1320 NODEMASK_SCRATCH_FREE(scratch
);
1325 ret
= queue_pages_range(mm
, start
, end
, nmask
,
1326 flags
| MPOL_MF_INVERT
, &pagelist
);
1333 err
= mbind_range(mm
, start
, end
, new);
1338 if (!list_empty(&pagelist
)) {
1339 WARN_ON_ONCE(flags
& MPOL_MF_LAZY
);
1340 nr_failed
= migrate_pages(&pagelist
, new_page
, NULL
,
1341 start
, MIGRATE_SYNC
, MR_MEMPOLICY_MBIND
);
1343 putback_movable_pages(&pagelist
);
1346 if ((ret
> 0) || (nr_failed
&& (flags
& MPOL_MF_STRICT
)))
1350 if (!list_empty(&pagelist
))
1351 putback_movable_pages(&pagelist
);
1354 mmap_write_unlock(mm
);
1357 if (flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
))
1363 * User space interface with variable sized bitmaps for nodelists.
1366 /* Copy a node mask from user space. */
1367 static int get_nodes(nodemask_t
*nodes
, const unsigned long __user
*nmask
,
1368 unsigned long maxnode
)
1372 unsigned long nlongs
;
1373 unsigned long endmask
;
1376 nodes_clear(*nodes
);
1377 if (maxnode
== 0 || !nmask
)
1379 if (maxnode
> PAGE_SIZE
*BITS_PER_BYTE
)
1382 nlongs
= BITS_TO_LONGS(maxnode
);
1383 if ((maxnode
% BITS_PER_LONG
) == 0)
1386 endmask
= (1UL << (maxnode
% BITS_PER_LONG
)) - 1;
1389 * When the user specified more nodes than supported just check
1390 * if the non supported part is all zero.
1392 * If maxnode have more longs than MAX_NUMNODES, check
1393 * the bits in that area first. And then go through to
1394 * check the rest bits which equal or bigger than MAX_NUMNODES.
1395 * Otherwise, just check bits [MAX_NUMNODES, maxnode).
1397 if (nlongs
> BITS_TO_LONGS(MAX_NUMNODES
)) {
1398 for (k
= BITS_TO_LONGS(MAX_NUMNODES
); k
< nlongs
; k
++) {
1399 if (get_user(t
, nmask
+ k
))
1401 if (k
== nlongs
- 1) {
1407 nlongs
= BITS_TO_LONGS(MAX_NUMNODES
);
1411 if (maxnode
> MAX_NUMNODES
&& MAX_NUMNODES
% BITS_PER_LONG
!= 0) {
1412 unsigned long valid_mask
= endmask
;
1414 valid_mask
&= ~((1UL << (MAX_NUMNODES
% BITS_PER_LONG
)) - 1);
1415 if (get_user(t
, nmask
+ nlongs
- 1))
1421 if (copy_from_user(nodes_addr(*nodes
), nmask
, nlongs
*sizeof(unsigned long)))
1423 nodes_addr(*nodes
)[nlongs
-1] &= endmask
;
1427 /* Copy a kernel node mask to user space */
1428 static int copy_nodes_to_user(unsigned long __user
*mask
, unsigned long maxnode
,
1431 unsigned long copy
= ALIGN(maxnode
-1, 64) / 8;
1432 unsigned int nbytes
= BITS_TO_LONGS(nr_node_ids
) * sizeof(long);
1434 if (copy
> nbytes
) {
1435 if (copy
> PAGE_SIZE
)
1437 if (clear_user((char __user
*)mask
+ nbytes
, copy
- nbytes
))
1441 return copy_to_user(mask
, nodes_addr(*nodes
), copy
) ? -EFAULT
: 0;
1444 /* Basic parameter sanity check used by both mbind() and set_mempolicy() */
1445 static inline int sanitize_mpol_flags(int *mode
, unsigned short *flags
)
1447 *flags
= *mode
& MPOL_MODE_FLAGS
;
1448 *mode
&= ~MPOL_MODE_FLAGS
;
1449 if ((unsigned int)(*mode
) >= MPOL_MAX
)
1451 if ((*flags
& MPOL_F_STATIC_NODES
) && (*flags
& MPOL_F_RELATIVE_NODES
))
1457 static long kernel_mbind(unsigned long start
, unsigned long len
,
1458 unsigned long mode
, const unsigned long __user
*nmask
,
1459 unsigned long maxnode
, unsigned int flags
)
1461 unsigned short mode_flags
;
1466 start
= untagged_addr(start
);
1467 err
= sanitize_mpol_flags(&lmode
, &mode_flags
);
1471 err
= get_nodes(&nodes
, nmask
, maxnode
);
1475 return do_mbind(start
, len
, lmode
, mode_flags
, &nodes
, flags
);
1478 SYSCALL_DEFINE6(mbind
, unsigned long, start
, unsigned long, len
,
1479 unsigned long, mode
, const unsigned long __user
*, nmask
,
1480 unsigned long, maxnode
, unsigned int, flags
)
1482 return kernel_mbind(start
, len
, mode
, nmask
, maxnode
, flags
);
1485 /* Set the process memory policy */
1486 static long kernel_set_mempolicy(int mode
, const unsigned long __user
*nmask
,
1487 unsigned long maxnode
)
1489 unsigned short mode_flags
;
1494 err
= sanitize_mpol_flags(&lmode
, &mode_flags
);
1498 err
= get_nodes(&nodes
, nmask
, maxnode
);
1502 return do_set_mempolicy(lmode
, mode_flags
, &nodes
);
1505 SYSCALL_DEFINE3(set_mempolicy
, int, mode
, const unsigned long __user
*, nmask
,
1506 unsigned long, maxnode
)
1508 return kernel_set_mempolicy(mode
, nmask
, maxnode
);
1511 static int kernel_migrate_pages(pid_t pid
, unsigned long maxnode
,
1512 const unsigned long __user
*old_nodes
,
1513 const unsigned long __user
*new_nodes
)
1515 struct mm_struct
*mm
= NULL
;
1516 struct task_struct
*task
;
1517 nodemask_t task_nodes
;
1521 NODEMASK_SCRATCH(scratch
);
1526 old
= &scratch
->mask1
;
1527 new = &scratch
->mask2
;
1529 err
= get_nodes(old
, old_nodes
, maxnode
);
1533 err
= get_nodes(new, new_nodes
, maxnode
);
1537 /* Find the mm_struct */
1539 task
= pid
? find_task_by_vpid(pid
) : current
;
1545 get_task_struct(task
);
1550 * Check if this process has the right to modify the specified process.
1551 * Use the regular "ptrace_may_access()" checks.
1553 if (!ptrace_may_access(task
, PTRACE_MODE_READ_REALCREDS
)) {
1560 task_nodes
= cpuset_mems_allowed(task
);
1561 /* Is the user allowed to access the target nodes? */
1562 if (!nodes_subset(*new, task_nodes
) && !capable(CAP_SYS_NICE
)) {
1567 task_nodes
= cpuset_mems_allowed(current
);
1568 nodes_and(*new, *new, task_nodes
);
1569 if (nodes_empty(*new))
1572 err
= security_task_movememory(task
);
1576 mm
= get_task_mm(task
);
1577 put_task_struct(task
);
1584 err
= do_migrate_pages(mm
, old
, new,
1585 capable(CAP_SYS_NICE
) ? MPOL_MF_MOVE_ALL
: MPOL_MF_MOVE
);
1589 NODEMASK_SCRATCH_FREE(scratch
);
1594 put_task_struct(task
);
1599 SYSCALL_DEFINE4(migrate_pages
, pid_t
, pid
, unsigned long, maxnode
,
1600 const unsigned long __user
*, old_nodes
,
1601 const unsigned long __user
*, new_nodes
)
1603 return kernel_migrate_pages(pid
, maxnode
, old_nodes
, new_nodes
);
1607 /* Retrieve NUMA policy */
1608 static int kernel_get_mempolicy(int __user
*policy
,
1609 unsigned long __user
*nmask
,
1610 unsigned long maxnode
,
1612 unsigned long flags
)
1618 if (nmask
!= NULL
&& maxnode
< nr_node_ids
)
1621 addr
= untagged_addr(addr
);
1623 err
= do_get_mempolicy(&pval
, &nodes
, addr
, flags
);
1628 if (policy
&& put_user(pval
, policy
))
1632 err
= copy_nodes_to_user(nmask
, maxnode
, &nodes
);
1637 SYSCALL_DEFINE5(get_mempolicy
, int __user
*, policy
,
1638 unsigned long __user
*, nmask
, unsigned long, maxnode
,
1639 unsigned long, addr
, unsigned long, flags
)
1641 return kernel_get_mempolicy(policy
, nmask
, maxnode
, addr
, flags
);
1644 #ifdef CONFIG_COMPAT
1646 COMPAT_SYSCALL_DEFINE5(get_mempolicy
, int __user
*, policy
,
1647 compat_ulong_t __user
*, nmask
,
1648 compat_ulong_t
, maxnode
,
1649 compat_ulong_t
, addr
, compat_ulong_t
, flags
)
1652 unsigned long __user
*nm
= NULL
;
1653 unsigned long nr_bits
, alloc_size
;
1654 DECLARE_BITMAP(bm
, MAX_NUMNODES
);
1656 nr_bits
= min_t(unsigned long, maxnode
-1, nr_node_ids
);
1657 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
1660 nm
= compat_alloc_user_space(alloc_size
);
1662 err
= kernel_get_mempolicy(policy
, nm
, nr_bits
+1, addr
, flags
);
1664 if (!err
&& nmask
) {
1665 unsigned long copy_size
;
1666 copy_size
= min_t(unsigned long, sizeof(bm
), alloc_size
);
1667 err
= copy_from_user(bm
, nm
, copy_size
);
1668 /* ensure entire bitmap is zeroed */
1669 err
|= clear_user(nmask
, ALIGN(maxnode
-1, 8) / 8);
1670 err
|= compat_put_bitmap(nmask
, bm
, nr_bits
);
1676 COMPAT_SYSCALL_DEFINE3(set_mempolicy
, int, mode
, compat_ulong_t __user
*, nmask
,
1677 compat_ulong_t
, maxnode
)
1679 unsigned long __user
*nm
= NULL
;
1680 unsigned long nr_bits
, alloc_size
;
1681 DECLARE_BITMAP(bm
, MAX_NUMNODES
);
1683 nr_bits
= min_t(unsigned long, maxnode
-1, MAX_NUMNODES
);
1684 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
1687 if (compat_get_bitmap(bm
, nmask
, nr_bits
))
1689 nm
= compat_alloc_user_space(alloc_size
);
1690 if (copy_to_user(nm
, bm
, alloc_size
))
1694 return kernel_set_mempolicy(mode
, nm
, nr_bits
+1);
1697 COMPAT_SYSCALL_DEFINE6(mbind
, compat_ulong_t
, start
, compat_ulong_t
, len
,
1698 compat_ulong_t
, mode
, compat_ulong_t __user
*, nmask
,
1699 compat_ulong_t
, maxnode
, compat_ulong_t
, flags
)
1701 unsigned long __user
*nm
= NULL
;
1702 unsigned long nr_bits
, alloc_size
;
1705 nr_bits
= min_t(unsigned long, maxnode
-1, MAX_NUMNODES
);
1706 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
1709 if (compat_get_bitmap(nodes_addr(bm
), nmask
, nr_bits
))
1711 nm
= compat_alloc_user_space(alloc_size
);
1712 if (copy_to_user(nm
, nodes_addr(bm
), alloc_size
))
1716 return kernel_mbind(start
, len
, mode
, nm
, nr_bits
+1, flags
);
1719 COMPAT_SYSCALL_DEFINE4(migrate_pages
, compat_pid_t
, pid
,
1720 compat_ulong_t
, maxnode
,
1721 const compat_ulong_t __user
*, old_nodes
,
1722 const compat_ulong_t __user
*, new_nodes
)
1724 unsigned long __user
*old
= NULL
;
1725 unsigned long __user
*new = NULL
;
1726 nodemask_t tmp_mask
;
1727 unsigned long nr_bits
;
1730 nr_bits
= min_t(unsigned long, maxnode
- 1, MAX_NUMNODES
);
1731 size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
1733 if (compat_get_bitmap(nodes_addr(tmp_mask
), old_nodes
, nr_bits
))
1735 old
= compat_alloc_user_space(new_nodes
? size
* 2 : size
);
1737 new = old
+ size
/ sizeof(unsigned long);
1738 if (copy_to_user(old
, nodes_addr(tmp_mask
), size
))
1742 if (compat_get_bitmap(nodes_addr(tmp_mask
), new_nodes
, nr_bits
))
1745 new = compat_alloc_user_space(size
);
1746 if (copy_to_user(new, nodes_addr(tmp_mask
), size
))
1749 return kernel_migrate_pages(pid
, nr_bits
+ 1, old
, new);
1752 #endif /* CONFIG_COMPAT */
1754 bool vma_migratable(struct vm_area_struct
*vma
)
1756 if (vma
->vm_flags
& (VM_IO
| VM_PFNMAP
))
1760 * DAX device mappings require predictable access latency, so avoid
1761 * incurring periodic faults.
1763 if (vma_is_dax(vma
))
1766 if (is_vm_hugetlb_page(vma
) &&
1767 !hugepage_migration_supported(hstate_vma(vma
)))
1771 * Migration allocates pages in the highest zone. If we cannot
1772 * do so then migration (at least from node to node) is not
1776 gfp_zone(mapping_gfp_mask(vma
->vm_file
->f_mapping
))
1782 struct mempolicy
*__get_vma_policy(struct vm_area_struct
*vma
,
1785 struct mempolicy
*pol
= NULL
;
1788 if (vma
->vm_ops
&& vma
->vm_ops
->get_policy
) {
1789 pol
= vma
->vm_ops
->get_policy(vma
, addr
);
1790 } else if (vma
->vm_policy
) {
1791 pol
= vma
->vm_policy
;
1794 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1795 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1796 * count on these policies which will be dropped by
1797 * mpol_cond_put() later
1799 if (mpol_needs_cond_ref(pol
))
1808 * get_vma_policy(@vma, @addr)
1809 * @vma: virtual memory area whose policy is sought
1810 * @addr: address in @vma for shared policy lookup
1812 * Returns effective policy for a VMA at specified address.
1813 * Falls back to current->mempolicy or system default policy, as necessary.
1814 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1815 * count--added by the get_policy() vm_op, as appropriate--to protect against
1816 * freeing by another task. It is the caller's responsibility to free the
1817 * extra reference for shared policies.
1819 static struct mempolicy
*get_vma_policy(struct vm_area_struct
*vma
,
1822 struct mempolicy
*pol
= __get_vma_policy(vma
, addr
);
1825 pol
= get_task_policy(current
);
1830 bool vma_policy_mof(struct vm_area_struct
*vma
)
1832 struct mempolicy
*pol
;
1834 if (vma
->vm_ops
&& vma
->vm_ops
->get_policy
) {
1837 pol
= vma
->vm_ops
->get_policy(vma
, vma
->vm_start
);
1838 if (pol
&& (pol
->flags
& MPOL_F_MOF
))
1845 pol
= vma
->vm_policy
;
1847 pol
= get_task_policy(current
);
1849 return pol
->flags
& MPOL_F_MOF
;
1852 static int apply_policy_zone(struct mempolicy
*policy
, enum zone_type zone
)
1854 enum zone_type dynamic_policy_zone
= policy_zone
;
1856 BUG_ON(dynamic_policy_zone
== ZONE_MOVABLE
);
1859 * if policy->nodes has movable memory only,
1860 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1862 * policy->nodes is intersect with node_states[N_MEMORY].
1863 * so if the following test fails, it implies
1864 * policy->nodes has movable memory only.
1866 if (!nodes_intersects(policy
->nodes
, node_states
[N_HIGH_MEMORY
]))
1867 dynamic_policy_zone
= ZONE_MOVABLE
;
1869 return zone
>= dynamic_policy_zone
;
1873 * Return a nodemask representing a mempolicy for filtering nodes for
1876 nodemask_t
*policy_nodemask(gfp_t gfp
, struct mempolicy
*policy
)
1878 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1879 if (unlikely(policy
->mode
== MPOL_BIND
) &&
1880 apply_policy_zone(policy
, gfp_zone(gfp
)) &&
1881 cpuset_nodemask_valid_mems_allowed(&policy
->nodes
))
1882 return &policy
->nodes
;
1887 /* Return the node id preferred by the given mempolicy, or the given id */
1888 static int policy_node(gfp_t gfp
, struct mempolicy
*policy
, int nd
)
1890 if (policy
->mode
== MPOL_PREFERRED
) {
1891 nd
= first_node(policy
->nodes
);
1894 * __GFP_THISNODE shouldn't even be used with the bind policy
1895 * because we might easily break the expectation to stay on the
1896 * requested node and not break the policy.
1898 WARN_ON_ONCE(policy
->mode
== MPOL_BIND
&& (gfp
& __GFP_THISNODE
));
1904 /* Do dynamic interleaving for a process */
1905 static unsigned interleave_nodes(struct mempolicy
*policy
)
1908 struct task_struct
*me
= current
;
1910 next
= next_node_in(me
->il_prev
, policy
->nodes
);
1911 if (next
< MAX_NUMNODES
)
1917 * Depending on the memory policy provide a node from which to allocate the
1920 unsigned int mempolicy_slab_node(void)
1922 struct mempolicy
*policy
;
1923 int node
= numa_mem_id();
1928 policy
= current
->mempolicy
;
1932 switch (policy
->mode
) {
1933 case MPOL_PREFERRED
:
1934 return first_node(policy
->nodes
);
1936 case MPOL_INTERLEAVE
:
1937 return interleave_nodes(policy
);
1943 * Follow bind policy behavior and start allocation at the
1946 struct zonelist
*zonelist
;
1947 enum zone_type highest_zoneidx
= gfp_zone(GFP_KERNEL
);
1948 zonelist
= &NODE_DATA(node
)->node_zonelists
[ZONELIST_FALLBACK
];
1949 z
= first_zones_zonelist(zonelist
, highest_zoneidx
,
1951 return z
->zone
? zone_to_nid(z
->zone
) : node
;
1962 * Do static interleaving for a VMA with known offset @n. Returns the n'th
1963 * node in pol->nodes (starting from n=0), wrapping around if n exceeds the
1964 * number of present nodes.
1966 static unsigned offset_il_node(struct mempolicy
*pol
, unsigned long n
)
1968 unsigned nnodes
= nodes_weight(pol
->nodes
);
1974 return numa_node_id();
1975 target
= (unsigned int)n
% nnodes
;
1976 nid
= first_node(pol
->nodes
);
1977 for (i
= 0; i
< target
; i
++)
1978 nid
= next_node(nid
, pol
->nodes
);
1982 /* Determine a node number for interleave */
1983 static inline unsigned interleave_nid(struct mempolicy
*pol
,
1984 struct vm_area_struct
*vma
, unsigned long addr
, int shift
)
1990 * for small pages, there is no difference between
1991 * shift and PAGE_SHIFT, so the bit-shift is safe.
1992 * for huge pages, since vm_pgoff is in units of small
1993 * pages, we need to shift off the always 0 bits to get
1996 BUG_ON(shift
< PAGE_SHIFT
);
1997 off
= vma
->vm_pgoff
>> (shift
- PAGE_SHIFT
);
1998 off
+= (addr
- vma
->vm_start
) >> shift
;
1999 return offset_il_node(pol
, off
);
2001 return interleave_nodes(pol
);
2004 #ifdef CONFIG_HUGETLBFS
2006 * huge_node(@vma, @addr, @gfp_flags, @mpol)
2007 * @vma: virtual memory area whose policy is sought
2008 * @addr: address in @vma for shared policy lookup and interleave policy
2009 * @gfp_flags: for requested zone
2010 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
2011 * @nodemask: pointer to nodemask pointer for MPOL_BIND nodemask
2013 * Returns a nid suitable for a huge page allocation and a pointer
2014 * to the struct mempolicy for conditional unref after allocation.
2015 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
2016 * @nodemask for filtering the zonelist.
2018 * Must be protected by read_mems_allowed_begin()
2020 int huge_node(struct vm_area_struct
*vma
, unsigned long addr
, gfp_t gfp_flags
,
2021 struct mempolicy
**mpol
, nodemask_t
**nodemask
)
2025 *mpol
= get_vma_policy(vma
, addr
);
2026 *nodemask
= NULL
; /* assume !MPOL_BIND */
2028 if (unlikely((*mpol
)->mode
== MPOL_INTERLEAVE
)) {
2029 nid
= interleave_nid(*mpol
, vma
, addr
,
2030 huge_page_shift(hstate_vma(vma
)));
2032 nid
= policy_node(gfp_flags
, *mpol
, numa_node_id());
2033 if ((*mpol
)->mode
== MPOL_BIND
)
2034 *nodemask
= &(*mpol
)->nodes
;
2040 * init_nodemask_of_mempolicy
2042 * If the current task's mempolicy is "default" [NULL], return 'false'
2043 * to indicate default policy. Otherwise, extract the policy nodemask
2044 * for 'bind' or 'interleave' policy into the argument nodemask, or
2045 * initialize the argument nodemask to contain the single node for
2046 * 'preferred' or 'local' policy and return 'true' to indicate presence
2047 * of non-default mempolicy.
2049 * We don't bother with reference counting the mempolicy [mpol_get/put]
2050 * because the current task is examining it's own mempolicy and a task's
2051 * mempolicy is only ever changed by the task itself.
2053 * N.B., it is the caller's responsibility to free a returned nodemask.
2055 bool init_nodemask_of_mempolicy(nodemask_t
*mask
)
2057 struct mempolicy
*mempolicy
;
2059 if (!(mask
&& current
->mempolicy
))
2063 mempolicy
= current
->mempolicy
;
2064 switch (mempolicy
->mode
) {
2065 case MPOL_PREFERRED
:
2067 case MPOL_INTERLEAVE
:
2068 *mask
= mempolicy
->nodes
;
2072 init_nodemask_of_node(mask
, numa_node_id());
2078 task_unlock(current
);
2085 * mempolicy_in_oom_domain
2087 * If tsk's mempolicy is "bind", check for intersection between mask and
2088 * the policy nodemask. Otherwise, return true for all other policies
2089 * including "interleave", as a tsk with "interleave" policy may have
2090 * memory allocated from all nodes in system.
2092 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
2094 bool mempolicy_in_oom_domain(struct task_struct
*tsk
,
2095 const nodemask_t
*mask
)
2097 struct mempolicy
*mempolicy
;
2104 mempolicy
= tsk
->mempolicy
;
2105 if (mempolicy
&& mempolicy
->mode
== MPOL_BIND
)
2106 ret
= nodes_intersects(mempolicy
->nodes
, *mask
);
2112 /* Allocate a page in interleaved policy.
2113 Own path because it needs to do special accounting. */
2114 static struct page
*alloc_page_interleave(gfp_t gfp
, unsigned order
,
2119 page
= __alloc_pages(gfp
, order
, nid
, NULL
);
2120 /* skip NUMA_INTERLEAVE_HIT counter update if numa stats is disabled */
2121 if (!static_branch_likely(&vm_numa_stat_key
))
2123 if (page
&& page_to_nid(page
) == nid
) {
2125 __count_numa_event(page_zone(page
), NUMA_INTERLEAVE_HIT
);
2132 * alloc_pages_vma - Allocate a page for a VMA.
2134 * @order: Order of the GFP allocation.
2135 * @vma: Pointer to VMA or NULL if not available.
2136 * @addr: Virtual address of the allocation. Must be inside @vma.
2137 * @node: Which node to prefer for allocation (modulo policy).
2138 * @hugepage: For hugepages try only the preferred node if possible.
2140 * Allocate a page for a specific address in @vma, using the appropriate
2141 * NUMA policy. When @vma is not NULL the caller must hold the mmap_lock
2142 * of the mm_struct of the VMA to prevent it from going away. Should be
2143 * used for all allocations for pages that will be mapped into user space.
2145 * Return: The page on success or NULL if allocation fails.
2147 struct page
*alloc_pages_vma(gfp_t gfp
, int order
, struct vm_area_struct
*vma
,
2148 unsigned long addr
, int node
, bool hugepage
)
2150 struct mempolicy
*pol
;
2155 pol
= get_vma_policy(vma
, addr
);
2157 if (pol
->mode
== MPOL_INTERLEAVE
) {
2160 nid
= interleave_nid(pol
, vma
, addr
, PAGE_SHIFT
+ order
);
2162 page
= alloc_page_interleave(gfp
, order
, nid
);
2166 if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE
) && hugepage
)) {
2167 int hpage_node
= node
;
2170 * For hugepage allocation and non-interleave policy which
2171 * allows the current node (or other explicitly preferred
2172 * node) we only try to allocate from the current/preferred
2173 * node and don't fall back to other nodes, as the cost of
2174 * remote accesses would likely offset THP benefits.
2176 * If the policy is interleave, or does not allow the current
2177 * node in its nodemask, we allocate the standard way.
2179 if (pol
->mode
== MPOL_PREFERRED
)
2180 hpage_node
= first_node(pol
->nodes
);
2182 nmask
= policy_nodemask(gfp
, pol
);
2183 if (!nmask
|| node_isset(hpage_node
, *nmask
)) {
2186 * First, try to allocate THP only on local node, but
2187 * don't reclaim unnecessarily, just compact.
2189 page
= __alloc_pages_node(hpage_node
,
2190 gfp
| __GFP_THISNODE
| __GFP_NORETRY
, order
);
2193 * If hugepage allocations are configured to always
2194 * synchronous compact or the vma has been madvised
2195 * to prefer hugepage backing, retry allowing remote
2196 * memory with both reclaim and compact as well.
2198 if (!page
&& (gfp
& __GFP_DIRECT_RECLAIM
))
2199 page
= __alloc_pages_node(hpage_node
,
2206 nmask
= policy_nodemask(gfp
, pol
);
2207 preferred_nid
= policy_node(gfp
, pol
, node
);
2208 page
= __alloc_pages(gfp
, order
, preferred_nid
, nmask
);
2213 EXPORT_SYMBOL(alloc_pages_vma
);
2216 * alloc_pages - Allocate pages.
2218 * @order: Power of two of number of pages to allocate.
2220 * Allocate 1 << @order contiguous pages. The physical address of the
2221 * first page is naturally aligned (eg an order-3 allocation will be aligned
2222 * to a multiple of 8 * PAGE_SIZE bytes). The NUMA policy of the current
2223 * process is honoured when in process context.
2225 * Context: Can be called from any context, providing the appropriate GFP
2227 * Return: The page on success or NULL if allocation fails.
2229 struct page
*alloc_pages(gfp_t gfp
, unsigned order
)
2231 struct mempolicy
*pol
= &default_policy
;
2234 if (!in_interrupt() && !(gfp
& __GFP_THISNODE
))
2235 pol
= get_task_policy(current
);
2238 * No reference counting needed for current->mempolicy
2239 * nor system default_policy
2241 if (pol
->mode
== MPOL_INTERLEAVE
)
2242 page
= alloc_page_interleave(gfp
, order
, interleave_nodes(pol
));
2244 page
= __alloc_pages(gfp
, order
,
2245 policy_node(gfp
, pol
, numa_node_id()),
2246 policy_nodemask(gfp
, pol
));
2250 EXPORT_SYMBOL(alloc_pages
);
2252 int vma_dup_policy(struct vm_area_struct
*src
, struct vm_area_struct
*dst
)
2254 struct mempolicy
*pol
= mpol_dup(vma_policy(src
));
2257 return PTR_ERR(pol
);
2258 dst
->vm_policy
= pol
;
2263 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2264 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2265 * with the mems_allowed returned by cpuset_mems_allowed(). This
2266 * keeps mempolicies cpuset relative after its cpuset moves. See
2267 * further kernel/cpuset.c update_nodemask().
2269 * current's mempolicy may be rebinded by the other task(the task that changes
2270 * cpuset's mems), so we needn't do rebind work for current task.
2273 /* Slow path of a mempolicy duplicate */
2274 struct mempolicy
*__mpol_dup(struct mempolicy
*old
)
2276 struct mempolicy
*new = kmem_cache_alloc(policy_cache
, GFP_KERNEL
);
2279 return ERR_PTR(-ENOMEM
);
2281 /* task's mempolicy is protected by alloc_lock */
2282 if (old
== current
->mempolicy
) {
2285 task_unlock(current
);
2289 if (current_cpuset_is_being_rebound()) {
2290 nodemask_t mems
= cpuset_mems_allowed(current
);
2291 mpol_rebind_policy(new, &mems
);
2293 atomic_set(&new->refcnt
, 1);
2297 /* Slow path of a mempolicy comparison */
2298 bool __mpol_equal(struct mempolicy
*a
, struct mempolicy
*b
)
2302 if (a
->mode
!= b
->mode
)
2304 if (a
->flags
!= b
->flags
)
2306 if (mpol_store_user_nodemask(a
))
2307 if (!nodes_equal(a
->w
.user_nodemask
, b
->w
.user_nodemask
))
2312 case MPOL_INTERLEAVE
:
2313 case MPOL_PREFERRED
:
2314 return !!nodes_equal(a
->nodes
, b
->nodes
);
2324 * Shared memory backing store policy support.
2326 * Remember policies even when nobody has shared memory mapped.
2327 * The policies are kept in Red-Black tree linked from the inode.
2328 * They are protected by the sp->lock rwlock, which should be held
2329 * for any accesses to the tree.
2333 * lookup first element intersecting start-end. Caller holds sp->lock for
2334 * reading or for writing
2336 static struct sp_node
*
2337 sp_lookup(struct shared_policy
*sp
, unsigned long start
, unsigned long end
)
2339 struct rb_node
*n
= sp
->root
.rb_node
;
2342 struct sp_node
*p
= rb_entry(n
, struct sp_node
, nd
);
2344 if (start
>= p
->end
)
2346 else if (end
<= p
->start
)
2354 struct sp_node
*w
= NULL
;
2355 struct rb_node
*prev
= rb_prev(n
);
2358 w
= rb_entry(prev
, struct sp_node
, nd
);
2359 if (w
->end
<= start
)
2363 return rb_entry(n
, struct sp_node
, nd
);
2367 * Insert a new shared policy into the list. Caller holds sp->lock for
2370 static void sp_insert(struct shared_policy
*sp
, struct sp_node
*new)
2372 struct rb_node
**p
= &sp
->root
.rb_node
;
2373 struct rb_node
*parent
= NULL
;
2378 nd
= rb_entry(parent
, struct sp_node
, nd
);
2379 if (new->start
< nd
->start
)
2381 else if (new->end
> nd
->end
)
2382 p
= &(*p
)->rb_right
;
2386 rb_link_node(&new->nd
, parent
, p
);
2387 rb_insert_color(&new->nd
, &sp
->root
);
2388 pr_debug("inserting %lx-%lx: %d\n", new->start
, new->end
,
2389 new->policy
? new->policy
->mode
: 0);
2392 /* Find shared policy intersecting idx */
2394 mpol_shared_policy_lookup(struct shared_policy
*sp
, unsigned long idx
)
2396 struct mempolicy
*pol
= NULL
;
2399 if (!sp
->root
.rb_node
)
2401 read_lock(&sp
->lock
);
2402 sn
= sp_lookup(sp
, idx
, idx
+1);
2404 mpol_get(sn
->policy
);
2407 read_unlock(&sp
->lock
);
2411 static void sp_free(struct sp_node
*n
)
2413 mpol_put(n
->policy
);
2414 kmem_cache_free(sn_cache
, n
);
2418 * mpol_misplaced - check whether current page node is valid in policy
2420 * @page: page to be checked
2421 * @vma: vm area where page mapped
2422 * @addr: virtual address where page mapped
2424 * Lookup current policy node id for vma,addr and "compare to" page's
2425 * node id. Policy determination "mimics" alloc_page_vma().
2426 * Called from fault path where we know the vma and faulting address.
2428 * Return: -1 if the page is in a node that is valid for this policy, or a
2429 * suitable node ID to allocate a replacement page from.
2431 int mpol_misplaced(struct page
*page
, struct vm_area_struct
*vma
, unsigned long addr
)
2433 struct mempolicy
*pol
;
2435 int curnid
= page_to_nid(page
);
2436 unsigned long pgoff
;
2437 int thiscpu
= raw_smp_processor_id();
2438 int thisnid
= cpu_to_node(thiscpu
);
2439 int polnid
= NUMA_NO_NODE
;
2442 pol
= get_vma_policy(vma
, addr
);
2443 if (!(pol
->flags
& MPOL_F_MOF
))
2446 switch (pol
->mode
) {
2447 case MPOL_INTERLEAVE
:
2448 pgoff
= vma
->vm_pgoff
;
2449 pgoff
+= (addr
- vma
->vm_start
) >> PAGE_SHIFT
;
2450 polnid
= offset_il_node(pol
, pgoff
);
2453 case MPOL_PREFERRED
:
2454 polnid
= first_node(pol
->nodes
);
2458 polnid
= numa_node_id();
2462 /* Optimize placement among multiple nodes via NUMA balancing */
2463 if (pol
->flags
& MPOL_F_MORON
) {
2464 if (node_isset(thisnid
, pol
->nodes
))
2470 * allows binding to multiple nodes.
2471 * use current page if in policy nodemask,
2472 * else select nearest allowed node, if any.
2473 * If no allowed nodes, use current [!misplaced].
2475 if (node_isset(curnid
, pol
->nodes
))
2477 z
= first_zones_zonelist(
2478 node_zonelist(numa_node_id(), GFP_HIGHUSER
),
2479 gfp_zone(GFP_HIGHUSER
),
2481 polnid
= zone_to_nid(z
->zone
);
2488 /* Migrate the page towards the node whose CPU is referencing it */
2489 if (pol
->flags
& MPOL_F_MORON
) {
2492 if (!should_numa_migrate_memory(current
, page
, curnid
, thiscpu
))
2496 if (curnid
!= polnid
)
2505 * Drop the (possibly final) reference to task->mempolicy. It needs to be
2506 * dropped after task->mempolicy is set to NULL so that any allocation done as
2507 * part of its kmem_cache_free(), such as by KASAN, doesn't reference a freed
2510 void mpol_put_task_policy(struct task_struct
*task
)
2512 struct mempolicy
*pol
;
2515 pol
= task
->mempolicy
;
2516 task
->mempolicy
= NULL
;
2521 static void sp_delete(struct shared_policy
*sp
, struct sp_node
*n
)
2523 pr_debug("deleting %lx-l%lx\n", n
->start
, n
->end
);
2524 rb_erase(&n
->nd
, &sp
->root
);
2528 static void sp_node_init(struct sp_node
*node
, unsigned long start
,
2529 unsigned long end
, struct mempolicy
*pol
)
2531 node
->start
= start
;
2536 static struct sp_node
*sp_alloc(unsigned long start
, unsigned long end
,
2537 struct mempolicy
*pol
)
2540 struct mempolicy
*newpol
;
2542 n
= kmem_cache_alloc(sn_cache
, GFP_KERNEL
);
2546 newpol
= mpol_dup(pol
);
2547 if (IS_ERR(newpol
)) {
2548 kmem_cache_free(sn_cache
, n
);
2551 newpol
->flags
|= MPOL_F_SHARED
;
2552 sp_node_init(n
, start
, end
, newpol
);
2557 /* Replace a policy range. */
2558 static int shared_policy_replace(struct shared_policy
*sp
, unsigned long start
,
2559 unsigned long end
, struct sp_node
*new)
2562 struct sp_node
*n_new
= NULL
;
2563 struct mempolicy
*mpol_new
= NULL
;
2567 write_lock(&sp
->lock
);
2568 n
= sp_lookup(sp
, start
, end
);
2569 /* Take care of old policies in the same range. */
2570 while (n
&& n
->start
< end
) {
2571 struct rb_node
*next
= rb_next(&n
->nd
);
2572 if (n
->start
>= start
) {
2578 /* Old policy spanning whole new range. */
2583 *mpol_new
= *n
->policy
;
2584 atomic_set(&mpol_new
->refcnt
, 1);
2585 sp_node_init(n_new
, end
, n
->end
, mpol_new
);
2587 sp_insert(sp
, n_new
);
2596 n
= rb_entry(next
, struct sp_node
, nd
);
2600 write_unlock(&sp
->lock
);
2607 kmem_cache_free(sn_cache
, n_new
);
2612 write_unlock(&sp
->lock
);
2614 n_new
= kmem_cache_alloc(sn_cache
, GFP_KERNEL
);
2617 mpol_new
= kmem_cache_alloc(policy_cache
, GFP_KERNEL
);
2624 * mpol_shared_policy_init - initialize shared policy for inode
2625 * @sp: pointer to inode shared policy
2626 * @mpol: struct mempolicy to install
2628 * Install non-NULL @mpol in inode's shared policy rb-tree.
2629 * On entry, the current task has a reference on a non-NULL @mpol.
2630 * This must be released on exit.
2631 * This is called at get_inode() calls and we can use GFP_KERNEL.
2633 void mpol_shared_policy_init(struct shared_policy
*sp
, struct mempolicy
*mpol
)
2637 sp
->root
= RB_ROOT
; /* empty tree == default mempolicy */
2638 rwlock_init(&sp
->lock
);
2641 struct vm_area_struct pvma
;
2642 struct mempolicy
*new;
2643 NODEMASK_SCRATCH(scratch
);
2647 /* contextualize the tmpfs mount point mempolicy */
2648 new = mpol_new(mpol
->mode
, mpol
->flags
, &mpol
->w
.user_nodemask
);
2650 goto free_scratch
; /* no valid nodemask intersection */
2653 ret
= mpol_set_nodemask(new, &mpol
->w
.user_nodemask
, scratch
);
2654 task_unlock(current
);
2658 /* Create pseudo-vma that contains just the policy */
2659 vma_init(&pvma
, NULL
);
2660 pvma
.vm_end
= TASK_SIZE
; /* policy covers entire file */
2661 mpol_set_shared_policy(sp
, &pvma
, new); /* adds ref */
2664 mpol_put(new); /* drop initial ref */
2666 NODEMASK_SCRATCH_FREE(scratch
);
2668 mpol_put(mpol
); /* drop our incoming ref on sb mpol */
2672 int mpol_set_shared_policy(struct shared_policy
*info
,
2673 struct vm_area_struct
*vma
, struct mempolicy
*npol
)
2676 struct sp_node
*new = NULL
;
2677 unsigned long sz
= vma_pages(vma
);
2679 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2681 sz
, npol
? npol
->mode
: -1,
2682 npol
? npol
->flags
: -1,
2683 npol
? nodes_addr(npol
->nodes
)[0] : NUMA_NO_NODE
);
2686 new = sp_alloc(vma
->vm_pgoff
, vma
->vm_pgoff
+ sz
, npol
);
2690 err
= shared_policy_replace(info
, vma
->vm_pgoff
, vma
->vm_pgoff
+sz
, new);
2696 /* Free a backing policy store on inode delete. */
2697 void mpol_free_shared_policy(struct shared_policy
*p
)
2700 struct rb_node
*next
;
2702 if (!p
->root
.rb_node
)
2704 write_lock(&p
->lock
);
2705 next
= rb_first(&p
->root
);
2707 n
= rb_entry(next
, struct sp_node
, nd
);
2708 next
= rb_next(&n
->nd
);
2711 write_unlock(&p
->lock
);
2714 #ifdef CONFIG_NUMA_BALANCING
2715 static int __initdata numabalancing_override
;
2717 static void __init
check_numabalancing_enable(void)
2719 bool numabalancing_default
= false;
2721 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED
))
2722 numabalancing_default
= true;
2724 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2725 if (numabalancing_override
)
2726 set_numabalancing_state(numabalancing_override
== 1);
2728 if (num_online_nodes() > 1 && !numabalancing_override
) {
2729 pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
2730 numabalancing_default
? "Enabling" : "Disabling");
2731 set_numabalancing_state(numabalancing_default
);
2735 static int __init
setup_numabalancing(char *str
)
2741 if (!strcmp(str
, "enable")) {
2742 numabalancing_override
= 1;
2744 } else if (!strcmp(str
, "disable")) {
2745 numabalancing_override
= -1;
2750 pr_warn("Unable to parse numa_balancing=\n");
2754 __setup("numa_balancing=", setup_numabalancing
);
2756 static inline void __init
check_numabalancing_enable(void)
2759 #endif /* CONFIG_NUMA_BALANCING */
2761 /* assumes fs == KERNEL_DS */
2762 void __init
numa_policy_init(void)
2764 nodemask_t interleave_nodes
;
2765 unsigned long largest
= 0;
2766 int nid
, prefer
= 0;
2768 policy_cache
= kmem_cache_create("numa_policy",
2769 sizeof(struct mempolicy
),
2770 0, SLAB_PANIC
, NULL
);
2772 sn_cache
= kmem_cache_create("shared_policy_node",
2773 sizeof(struct sp_node
),
2774 0, SLAB_PANIC
, NULL
);
2776 for_each_node(nid
) {
2777 preferred_node_policy
[nid
] = (struct mempolicy
) {
2778 .refcnt
= ATOMIC_INIT(1),
2779 .mode
= MPOL_PREFERRED
,
2780 .flags
= MPOL_F_MOF
| MPOL_F_MORON
,
2781 .nodes
= nodemask_of_node(nid
),
2786 * Set interleaving policy for system init. Interleaving is only
2787 * enabled across suitably sized nodes (default is >= 16MB), or
2788 * fall back to the largest node if they're all smaller.
2790 nodes_clear(interleave_nodes
);
2791 for_each_node_state(nid
, N_MEMORY
) {
2792 unsigned long total_pages
= node_present_pages(nid
);
2794 /* Preserve the largest node */
2795 if (largest
< total_pages
) {
2796 largest
= total_pages
;
2800 /* Interleave this node? */
2801 if ((total_pages
<< PAGE_SHIFT
) >= (16 << 20))
2802 node_set(nid
, interleave_nodes
);
2805 /* All too small, use the largest */
2806 if (unlikely(nodes_empty(interleave_nodes
)))
2807 node_set(prefer
, interleave_nodes
);
2809 if (do_set_mempolicy(MPOL_INTERLEAVE
, 0, &interleave_nodes
))
2810 pr_err("%s: interleaving failed\n", __func__
);
2812 check_numabalancing_enable();
2815 /* Reset policy of current process to default */
2816 void numa_default_policy(void)
2818 do_set_mempolicy(MPOL_DEFAULT
, 0, NULL
);
2822 * Parse and format mempolicy from/to strings
2825 static const char * const policy_modes
[] =
2827 [MPOL_DEFAULT
] = "default",
2828 [MPOL_PREFERRED
] = "prefer",
2829 [MPOL_BIND
] = "bind",
2830 [MPOL_INTERLEAVE
] = "interleave",
2831 [MPOL_LOCAL
] = "local",
2837 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2838 * @str: string containing mempolicy to parse
2839 * @mpol: pointer to struct mempolicy pointer, returned on success.
2842 * <mode>[=<flags>][:<nodelist>]
2844 * On success, returns 0, else 1
2846 int mpol_parse_str(char *str
, struct mempolicy
**mpol
)
2848 struct mempolicy
*new = NULL
;
2849 unsigned short mode_flags
;
2851 char *nodelist
= strchr(str
, ':');
2852 char *flags
= strchr(str
, '=');
2856 *flags
++ = '\0'; /* terminate mode string */
2859 /* NUL-terminate mode or flags string */
2861 if (nodelist_parse(nodelist
, nodes
))
2863 if (!nodes_subset(nodes
, node_states
[N_MEMORY
]))
2868 mode
= match_string(policy_modes
, MPOL_MAX
, str
);
2873 case MPOL_PREFERRED
:
2875 * Insist on a nodelist of one node only, although later
2876 * we use first_node(nodes) to grab a single node, so here
2877 * nodelist (or nodes) cannot be empty.
2880 char *rest
= nodelist
;
2881 while (isdigit(*rest
))
2885 if (nodes_empty(nodes
))
2889 case MPOL_INTERLEAVE
:
2891 * Default to online nodes with memory if no nodelist
2894 nodes
= node_states
[N_MEMORY
];
2898 * Don't allow a nodelist; mpol_new() checks flags
2905 * Insist on a empty nodelist
2912 * Insist on a nodelist
2921 * Currently, we only support two mutually exclusive
2924 if (!strcmp(flags
, "static"))
2925 mode_flags
|= MPOL_F_STATIC_NODES
;
2926 else if (!strcmp(flags
, "relative"))
2927 mode_flags
|= MPOL_F_RELATIVE_NODES
;
2932 new = mpol_new(mode
, mode_flags
, &nodes
);
2937 * Save nodes for mpol_to_str() to show the tmpfs mount options
2938 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2940 if (mode
!= MPOL_PREFERRED
) {
2942 } else if (nodelist
) {
2943 nodes_clear(new->nodes
);
2944 node_set(first_node(nodes
), new->nodes
);
2946 new->mode
= MPOL_LOCAL
;
2950 * Save nodes for contextualization: this will be used to "clone"
2951 * the mempolicy in a specific context [cpuset] at a later time.
2953 new->w
.user_nodemask
= nodes
;
2958 /* Restore string for error message */
2967 #endif /* CONFIG_TMPFS */
2970 * mpol_to_str - format a mempolicy structure for printing
2971 * @buffer: to contain formatted mempolicy string
2972 * @maxlen: length of @buffer
2973 * @pol: pointer to mempolicy to be formatted
2975 * Convert @pol into a string. If @buffer is too short, truncate the string.
2976 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
2977 * longest flag, "relative", and to display at least a few node ids.
2979 void mpol_to_str(char *buffer
, int maxlen
, struct mempolicy
*pol
)
2982 nodemask_t nodes
= NODE_MASK_NONE
;
2983 unsigned short mode
= MPOL_DEFAULT
;
2984 unsigned short flags
= 0;
2986 if (pol
&& pol
!= &default_policy
&& !(pol
->flags
& MPOL_F_MORON
)) {
2995 case MPOL_PREFERRED
:
2997 case MPOL_INTERLEAVE
:
3002 snprintf(p
, maxlen
, "unknown");
3006 p
+= snprintf(p
, maxlen
, "%s", policy_modes
[mode
]);
3008 if (flags
& MPOL_MODE_FLAGS
) {
3009 p
+= snprintf(p
, buffer
+ maxlen
- p
, "=");
3012 * Currently, the only defined flags are mutually exclusive
3014 if (flags
& MPOL_F_STATIC_NODES
)
3015 p
+= snprintf(p
, buffer
+ maxlen
- p
, "static");
3016 else if (flags
& MPOL_F_RELATIVE_NODES
)
3017 p
+= snprintf(p
, buffer
+ maxlen
- p
, "relative");
3020 if (!nodes_empty(nodes
))
3021 p
+= scnprintf(p
, buffer
+ maxlen
- p
, ":%*pbl",
3022 nodemask_pr_args(&nodes
));