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 */
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
;
415 struct vm_area_struct
*first
;
419 * Check if the page's nid is in qp->nmask.
421 * If MPOL_MF_INVERT is set in qp->flags, check if the nid is
422 * in the invert of qp->nmask.
424 static inline bool queue_pages_required(struct page
*page
,
425 struct queue_pages
*qp
)
427 int nid
= page_to_nid(page
);
428 unsigned long flags
= qp
->flags
;
430 return node_isset(nid
, *qp
->nmask
) == !(flags
& MPOL_MF_INVERT
);
434 * queue_pages_pmd() has four possible return values:
435 * 0 - pages are placed on the right node or queued successfully.
436 * 1 - there is unmovable page, and MPOL_MF_MOVE* & MPOL_MF_STRICT were
439 * -EIO - is migration entry or only MPOL_MF_STRICT was specified and an
440 * existing page was already on a node that does not follow the
443 static int queue_pages_pmd(pmd_t
*pmd
, spinlock_t
*ptl
, unsigned long addr
,
444 unsigned long end
, struct mm_walk
*walk
)
448 struct queue_pages
*qp
= walk
->private;
451 if (unlikely(is_pmd_migration_entry(*pmd
))) {
455 page
= pmd_page(*pmd
);
456 if (is_huge_zero_page(page
)) {
458 __split_huge_pmd(walk
->vma
, pmd
, addr
, false, NULL
);
462 if (!queue_pages_required(page
, qp
))
466 /* go to thp migration */
467 if (flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
)) {
468 if (!vma_migratable(walk
->vma
) ||
469 migrate_page_add(page
, qp
->pagelist
, flags
)) {
482 * Scan through pages checking if pages follow certain conditions,
483 * and move them to the pagelist if they do.
485 * queue_pages_pte_range() has three possible return values:
486 * 0 - pages are placed on the right node or queued successfully.
487 * 1 - there is unmovable page, and MPOL_MF_MOVE* & MPOL_MF_STRICT were
489 * -EIO - only MPOL_MF_STRICT was specified and an existing page was already
490 * on a node that does not follow the policy.
492 static int queue_pages_pte_range(pmd_t
*pmd
, unsigned long addr
,
493 unsigned long end
, struct mm_walk
*walk
)
495 struct vm_area_struct
*vma
= walk
->vma
;
497 struct queue_pages
*qp
= walk
->private;
498 unsigned long flags
= qp
->flags
;
500 bool has_unmovable
= false;
504 ptl
= pmd_trans_huge_lock(pmd
, vma
);
506 ret
= queue_pages_pmd(pmd
, ptl
, addr
, end
, walk
);
510 /* THP was split, fall through to pte walk */
512 if (pmd_trans_unstable(pmd
))
515 pte
= pte_offset_map_lock(walk
->mm
, pmd
, addr
, &ptl
);
516 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
) {
517 if (!pte_present(*pte
))
519 page
= vm_normal_page(vma
, addr
, *pte
);
523 * vm_normal_page() filters out zero pages, but there might
524 * still be PageReserved pages to skip, perhaps in a VDSO.
526 if (PageReserved(page
))
528 if (!queue_pages_required(page
, qp
))
530 if (flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
)) {
531 /* MPOL_MF_STRICT must be specified if we get here */
532 if (!vma_migratable(vma
)) {
533 has_unmovable
= true;
538 * Do not abort immediately since there may be
539 * temporary off LRU pages in the range. Still
540 * need migrate other LRU pages.
542 if (migrate_page_add(page
, qp
->pagelist
, flags
))
543 has_unmovable
= true;
547 pte_unmap_unlock(pte
- 1, ptl
);
553 return addr
!= end
? -EIO
: 0;
556 static int queue_pages_hugetlb(pte_t
*pte
, unsigned long hmask
,
557 unsigned long addr
, unsigned long end
,
558 struct mm_walk
*walk
)
560 #ifdef CONFIG_HUGETLB_PAGE
561 struct queue_pages
*qp
= walk
->private;
562 unsigned long flags
= qp
->flags
;
567 ptl
= huge_pte_lock(hstate_vma(walk
->vma
), walk
->mm
, pte
);
568 entry
= huge_ptep_get(pte
);
569 if (!pte_present(entry
))
571 page
= pte_page(entry
);
572 if (!queue_pages_required(page
, qp
))
574 /* With MPOL_MF_MOVE, we migrate only unshared hugepage. */
575 if (flags
& (MPOL_MF_MOVE_ALL
) ||
576 (flags
& MPOL_MF_MOVE
&& page_mapcount(page
) == 1))
577 isolate_huge_page(page
, qp
->pagelist
);
586 #ifdef CONFIG_NUMA_BALANCING
588 * This is used to mark a range of virtual addresses to be inaccessible.
589 * These are later cleared by a NUMA hinting fault. Depending on these
590 * faults, pages may be migrated for better NUMA placement.
592 * This is assuming that NUMA faults are handled using PROT_NONE. If
593 * an architecture makes a different choice, it will need further
594 * changes to the core.
596 unsigned long change_prot_numa(struct vm_area_struct
*vma
,
597 unsigned long addr
, unsigned long end
)
601 nr_updated
= change_protection(vma
, addr
, end
, PAGE_NONE
, 0, 1);
603 count_vm_numa_events(NUMA_PTE_UPDATES
, nr_updated
);
608 static unsigned long change_prot_numa(struct vm_area_struct
*vma
,
609 unsigned long addr
, unsigned long end
)
613 #endif /* CONFIG_NUMA_BALANCING */
615 static int queue_pages_test_walk(unsigned long start
, unsigned long end
,
616 struct mm_walk
*walk
)
618 struct vm_area_struct
*vma
= walk
->vma
;
619 struct queue_pages
*qp
= walk
->private;
620 unsigned long endvma
= vma
->vm_end
;
621 unsigned long flags
= qp
->flags
;
623 /* range check first */
624 VM_BUG_ON((vma
->vm_start
> start
) || (vma
->vm_end
< end
));
628 if (!(flags
& MPOL_MF_DISCONTIG_OK
) &&
629 (qp
->start
< vma
->vm_start
))
630 /* hole at head side of range */
633 if (!(flags
& MPOL_MF_DISCONTIG_OK
) &&
634 ((vma
->vm_end
< qp
->end
) &&
635 (!vma
->vm_next
|| vma
->vm_end
< vma
->vm_next
->vm_start
)))
636 /* hole at middle or tail of range */
640 * Need check MPOL_MF_STRICT to return -EIO if possible
641 * regardless of vma_migratable
643 if (!vma_migratable(vma
) &&
644 !(flags
& MPOL_MF_STRICT
))
650 if (flags
& MPOL_MF_LAZY
) {
651 /* Similar to task_numa_work, skip inaccessible VMAs */
652 if (!is_vm_hugetlb_page(vma
) &&
653 (vma
->vm_flags
& (VM_READ
| VM_EXEC
| VM_WRITE
)) &&
654 !(vma
->vm_flags
& VM_MIXEDMAP
))
655 change_prot_numa(vma
, start
, endvma
);
659 /* queue pages from current vma */
660 if (flags
& MPOL_MF_VALID
)
665 static const struct mm_walk_ops queue_pages_walk_ops
= {
666 .hugetlb_entry
= queue_pages_hugetlb
,
667 .pmd_entry
= queue_pages_pte_range
,
668 .test_walk
= queue_pages_test_walk
,
672 * Walk through page tables and collect pages to be migrated.
674 * If pages found in a given range are on a set of nodes (determined by
675 * @nodes and @flags,) it's isolated and queued to the pagelist which is
676 * passed via @private.
678 * queue_pages_range() has three possible return values:
679 * 1 - there is unmovable page, but MPOL_MF_MOVE* & MPOL_MF_STRICT were
681 * 0 - queue pages successfully or no misplaced page.
682 * errno - i.e. misplaced pages with MPOL_MF_STRICT specified (-EIO) or
683 * memory range specified by nodemask and maxnode points outside
684 * your accessible address space (-EFAULT)
687 queue_pages_range(struct mm_struct
*mm
, unsigned long start
, unsigned long end
,
688 nodemask_t
*nodes
, unsigned long flags
,
689 struct list_head
*pagelist
)
692 struct queue_pages qp
= {
693 .pagelist
= pagelist
,
701 err
= walk_page_range(mm
, start
, end
, &queue_pages_walk_ops
, &qp
);
704 /* whole range in hole */
711 * Apply policy to a single VMA
712 * This must be called with the mmap_sem held for writing.
714 static int vma_replace_policy(struct vm_area_struct
*vma
,
715 struct mempolicy
*pol
)
718 struct mempolicy
*old
;
719 struct mempolicy
*new;
721 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
722 vma
->vm_start
, vma
->vm_end
, vma
->vm_pgoff
,
723 vma
->vm_ops
, vma
->vm_file
,
724 vma
->vm_ops
? vma
->vm_ops
->set_policy
: NULL
);
730 if (vma
->vm_ops
&& vma
->vm_ops
->set_policy
) {
731 err
= vma
->vm_ops
->set_policy(vma
, new);
736 old
= vma
->vm_policy
;
737 vma
->vm_policy
= new; /* protected by mmap_sem */
746 /* Step 2: apply policy to a range and do splits. */
747 static int mbind_range(struct mm_struct
*mm
, unsigned long start
,
748 unsigned long end
, struct mempolicy
*new_pol
)
750 struct vm_area_struct
*next
;
751 struct vm_area_struct
*prev
;
752 struct vm_area_struct
*vma
;
755 unsigned long vmstart
;
758 vma
= find_vma(mm
, start
);
762 if (start
> vma
->vm_start
)
765 for (; vma
&& vma
->vm_start
< end
; prev
= vma
, vma
= next
) {
767 vmstart
= max(start
, vma
->vm_start
);
768 vmend
= min(end
, vma
->vm_end
);
770 if (mpol_equal(vma_policy(vma
), new_pol
))
773 pgoff
= vma
->vm_pgoff
+
774 ((vmstart
- vma
->vm_start
) >> PAGE_SHIFT
);
775 prev
= vma_merge(mm
, prev
, vmstart
, vmend
, vma
->vm_flags
,
776 vma
->anon_vma
, vma
->vm_file
, pgoff
,
777 new_pol
, vma
->vm_userfaultfd_ctx
);
781 if (mpol_equal(vma_policy(vma
), new_pol
))
783 /* vma_merge() joined vma && vma->next, case 8 */
786 if (vma
->vm_start
!= vmstart
) {
787 err
= split_vma(vma
->vm_mm
, vma
, vmstart
, 1);
791 if (vma
->vm_end
!= vmend
) {
792 err
= split_vma(vma
->vm_mm
, vma
, vmend
, 0);
797 err
= vma_replace_policy(vma
, new_pol
);
806 /* Set the process memory policy */
807 static long do_set_mempolicy(unsigned short mode
, unsigned short flags
,
810 struct mempolicy
*new, *old
;
811 NODEMASK_SCRATCH(scratch
);
817 new = mpol_new(mode
, flags
, nodes
);
824 ret
= mpol_set_nodemask(new, nodes
, scratch
);
826 task_unlock(current
);
830 old
= current
->mempolicy
;
831 current
->mempolicy
= new;
832 if (new && new->mode
== MPOL_INTERLEAVE
)
833 current
->il_prev
= MAX_NUMNODES
-1;
834 task_unlock(current
);
838 NODEMASK_SCRATCH_FREE(scratch
);
843 * Return nodemask for policy for get_mempolicy() query
845 * Called with task's alloc_lock held
847 static void get_policy_nodemask(struct mempolicy
*p
, nodemask_t
*nodes
)
850 if (p
== &default_policy
)
856 case MPOL_INTERLEAVE
:
860 if (!(p
->flags
& MPOL_F_LOCAL
))
861 node_set(p
->v
.preferred_node
, *nodes
);
862 /* else return empty node mask for local allocation */
869 static int lookup_node(struct mm_struct
*mm
, unsigned long addr
)
875 err
= get_user_pages_locked(addr
& PAGE_MASK
, 1, 0, &p
, &locked
);
877 err
= page_to_nid(p
);
881 up_read(&mm
->mmap_sem
);
885 /* Retrieve NUMA policy */
886 static long do_get_mempolicy(int *policy
, nodemask_t
*nmask
,
887 unsigned long addr
, unsigned long flags
)
890 struct mm_struct
*mm
= current
->mm
;
891 struct vm_area_struct
*vma
= NULL
;
892 struct mempolicy
*pol
= current
->mempolicy
, *pol_refcount
= NULL
;
895 ~(unsigned long)(MPOL_F_NODE
|MPOL_F_ADDR
|MPOL_F_MEMS_ALLOWED
))
898 if (flags
& MPOL_F_MEMS_ALLOWED
) {
899 if (flags
& (MPOL_F_NODE
|MPOL_F_ADDR
))
901 *policy
= 0; /* just so it's initialized */
903 *nmask
= cpuset_current_mems_allowed
;
904 task_unlock(current
);
908 if (flags
& MPOL_F_ADDR
) {
910 * Do NOT fall back to task policy if the
911 * vma/shared policy at addr is NULL. We
912 * want to return MPOL_DEFAULT in this case.
914 down_read(&mm
->mmap_sem
);
915 vma
= find_vma_intersection(mm
, addr
, addr
+1);
917 up_read(&mm
->mmap_sem
);
920 if (vma
->vm_ops
&& vma
->vm_ops
->get_policy
)
921 pol
= vma
->vm_ops
->get_policy(vma
, addr
);
923 pol
= vma
->vm_policy
;
928 pol
= &default_policy
; /* indicates default behavior */
930 if (flags
& MPOL_F_NODE
) {
931 if (flags
& MPOL_F_ADDR
) {
933 * Take a refcount on the mpol, lookup_node()
934 * wil drop the mmap_sem, so after calling
935 * lookup_node() only "pol" remains valid, "vma"
941 err
= lookup_node(mm
, addr
);
945 } else if (pol
== current
->mempolicy
&&
946 pol
->mode
== MPOL_INTERLEAVE
) {
947 *policy
= next_node_in(current
->il_prev
, pol
->v
.nodes
);
953 *policy
= pol
== &default_policy
? MPOL_DEFAULT
:
956 * Internal mempolicy flags must be masked off before exposing
957 * the policy to userspace.
959 *policy
|= (pol
->flags
& MPOL_MODE_FLAGS
);
964 if (mpol_store_user_nodemask(pol
)) {
965 *nmask
= pol
->w
.user_nodemask
;
968 get_policy_nodemask(pol
, nmask
);
969 task_unlock(current
);
976 up_read(&mm
->mmap_sem
);
978 mpol_put(pol_refcount
);
982 #ifdef CONFIG_MIGRATION
984 * page migration, thp tail pages can be passed.
986 static int migrate_page_add(struct page
*page
, struct list_head
*pagelist
,
989 struct page
*head
= compound_head(page
);
991 * Avoid migrating a page that is shared with others.
993 if ((flags
& MPOL_MF_MOVE_ALL
) || page_mapcount(head
) == 1) {
994 if (!isolate_lru_page(head
)) {
995 list_add_tail(&head
->lru
, pagelist
);
996 mod_node_page_state(page_pgdat(head
),
997 NR_ISOLATED_ANON
+ page_is_file_cache(head
),
998 hpage_nr_pages(head
));
999 } else if (flags
& MPOL_MF_STRICT
) {
1001 * Non-movable page may reach here. And, there may be
1002 * temporary off LRU pages or non-LRU movable pages.
1003 * Treat them as unmovable pages since they can't be
1004 * isolated, so they can't be moved at the moment. It
1005 * should return -EIO for this case too.
1014 /* page allocation callback for NUMA node migration */
1015 struct page
*alloc_new_node_page(struct page
*page
, unsigned long node
)
1018 return alloc_huge_page_node(page_hstate(compound_head(page
)),
1020 else if (PageTransHuge(page
)) {
1023 thp
= alloc_pages_node(node
,
1024 (GFP_TRANSHUGE
| __GFP_THISNODE
),
1028 prep_transhuge_page(thp
);
1031 return __alloc_pages_node(node
, GFP_HIGHUSER_MOVABLE
|
1036 * Migrate pages from one node to a target node.
1037 * Returns error or the number of pages not migrated.
1039 static int migrate_to_node(struct mm_struct
*mm
, int source
, int dest
,
1043 LIST_HEAD(pagelist
);
1047 node_set(source
, nmask
);
1050 * This does not "check" the range but isolates all pages that
1051 * need migration. Between passing in the full user address
1052 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1054 VM_BUG_ON(!(flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
)));
1055 queue_pages_range(mm
, mm
->mmap
->vm_start
, mm
->task_size
, &nmask
,
1056 flags
| MPOL_MF_DISCONTIG_OK
, &pagelist
);
1058 if (!list_empty(&pagelist
)) {
1059 err
= migrate_pages(&pagelist
, alloc_new_node_page
, NULL
, dest
,
1060 MIGRATE_SYNC
, MR_SYSCALL
);
1062 putback_movable_pages(&pagelist
);
1069 * Move pages between the two nodesets so as to preserve the physical
1070 * layout as much as possible.
1072 * Returns the number of page that could not be moved.
1074 int do_migrate_pages(struct mm_struct
*mm
, const nodemask_t
*from
,
1075 const nodemask_t
*to
, int flags
)
1081 err
= migrate_prep();
1085 down_read(&mm
->mmap_sem
);
1088 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1089 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1090 * bit in 'tmp', and return that <source, dest> pair for migration.
1091 * The pair of nodemasks 'to' and 'from' define the map.
1093 * If no pair of bits is found that way, fallback to picking some
1094 * pair of 'source' and 'dest' bits that are not the same. If the
1095 * 'source' and 'dest' bits are the same, this represents a node
1096 * that will be migrating to itself, so no pages need move.
1098 * If no bits are left in 'tmp', or if all remaining bits left
1099 * in 'tmp' correspond to the same bit in 'to', return false
1100 * (nothing left to migrate).
1102 * This lets us pick a pair of nodes to migrate between, such that
1103 * if possible the dest node is not already occupied by some other
1104 * source node, minimizing the risk of overloading the memory on a
1105 * node that would happen if we migrated incoming memory to a node
1106 * before migrating outgoing memory source that same node.
1108 * A single scan of tmp is sufficient. As we go, we remember the
1109 * most recent <s, d> pair that moved (s != d). If we find a pair
1110 * that not only moved, but what's better, moved to an empty slot
1111 * (d is not set in tmp), then we break out then, with that pair.
1112 * Otherwise when we finish scanning from_tmp, we at least have the
1113 * most recent <s, d> pair that moved. If we get all the way through
1114 * the scan of tmp without finding any node that moved, much less
1115 * moved to an empty node, then there is nothing left worth migrating.
1119 while (!nodes_empty(tmp
)) {
1121 int source
= NUMA_NO_NODE
;
1124 for_each_node_mask(s
, tmp
) {
1127 * do_migrate_pages() tries to maintain the relative
1128 * node relationship of the pages established between
1129 * threads and memory areas.
1131 * However if the number of source nodes is not equal to
1132 * the number of destination nodes we can not preserve
1133 * this node relative relationship. In that case, skip
1134 * copying memory from a node that is in the destination
1137 * Example: [2,3,4] -> [3,4,5] moves everything.
1138 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1141 if ((nodes_weight(*from
) != nodes_weight(*to
)) &&
1142 (node_isset(s
, *to
)))
1145 d
= node_remap(s
, *from
, *to
);
1149 source
= s
; /* Node moved. Memorize */
1152 /* dest not in remaining from nodes? */
1153 if (!node_isset(dest
, tmp
))
1156 if (source
== NUMA_NO_NODE
)
1159 node_clear(source
, tmp
);
1160 err
= migrate_to_node(mm
, source
, dest
, flags
);
1166 up_read(&mm
->mmap_sem
);
1174 * Allocate a new page for page migration based on vma policy.
1175 * Start by assuming the page is mapped by the same vma as contains @start.
1176 * Search forward from there, if not. N.B., this assumes that the
1177 * list of pages handed to migrate_pages()--which is how we get here--
1178 * is in virtual address order.
1180 static struct page
*new_page(struct page
*page
, unsigned long start
)
1182 struct vm_area_struct
*vma
;
1183 unsigned long uninitialized_var(address
);
1185 vma
= find_vma(current
->mm
, start
);
1187 address
= page_address_in_vma(page
, vma
);
1188 if (address
!= -EFAULT
)
1193 if (PageHuge(page
)) {
1194 return alloc_huge_page_vma(page_hstate(compound_head(page
)),
1196 } else if (PageTransHuge(page
)) {
1199 thp
= alloc_hugepage_vma(GFP_TRANSHUGE
, vma
, address
,
1203 prep_transhuge_page(thp
);
1207 * if !vma, alloc_page_vma() will use task or system default policy
1209 return alloc_page_vma(GFP_HIGHUSER_MOVABLE
| __GFP_RETRY_MAYFAIL
,
1214 static int migrate_page_add(struct page
*page
, struct list_head
*pagelist
,
1215 unsigned long flags
)
1220 int do_migrate_pages(struct mm_struct
*mm
, const nodemask_t
*from
,
1221 const nodemask_t
*to
, int flags
)
1226 static struct page
*new_page(struct page
*page
, unsigned long start
)
1232 static long do_mbind(unsigned long start
, unsigned long len
,
1233 unsigned short mode
, unsigned short mode_flags
,
1234 nodemask_t
*nmask
, unsigned long flags
)
1236 struct mm_struct
*mm
= current
->mm
;
1237 struct mempolicy
*new;
1241 LIST_HEAD(pagelist
);
1243 if (flags
& ~(unsigned long)MPOL_MF_VALID
)
1245 if ((flags
& MPOL_MF_MOVE_ALL
) && !capable(CAP_SYS_NICE
))
1248 if (start
& ~PAGE_MASK
)
1251 if (mode
== MPOL_DEFAULT
)
1252 flags
&= ~MPOL_MF_STRICT
;
1254 len
= (len
+ PAGE_SIZE
- 1) & PAGE_MASK
;
1262 new = mpol_new(mode
, mode_flags
, nmask
);
1264 return PTR_ERR(new);
1266 if (flags
& MPOL_MF_LAZY
)
1267 new->flags
|= MPOL_F_MOF
;
1270 * If we are using the default policy then operation
1271 * on discontinuous address spaces is okay after all
1274 flags
|= MPOL_MF_DISCONTIG_OK
;
1276 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1277 start
, start
+ len
, mode
, mode_flags
,
1278 nmask
? nodes_addr(*nmask
)[0] : NUMA_NO_NODE
);
1280 if (flags
& (MPOL_MF_MOVE
| MPOL_MF_MOVE_ALL
)) {
1282 err
= migrate_prep();
1287 NODEMASK_SCRATCH(scratch
);
1289 down_write(&mm
->mmap_sem
);
1291 err
= mpol_set_nodemask(new, nmask
, scratch
);
1292 task_unlock(current
);
1294 up_write(&mm
->mmap_sem
);
1297 NODEMASK_SCRATCH_FREE(scratch
);
1302 ret
= queue_pages_range(mm
, start
, end
, nmask
,
1303 flags
| MPOL_MF_INVERT
, &pagelist
);
1310 err
= mbind_range(mm
, start
, end
, new);
1315 if (!list_empty(&pagelist
)) {
1316 WARN_ON_ONCE(flags
& MPOL_MF_LAZY
);
1317 nr_failed
= migrate_pages(&pagelist
, new_page
, NULL
,
1318 start
, MIGRATE_SYNC
, MR_MEMPOLICY_MBIND
);
1320 putback_movable_pages(&pagelist
);
1323 if ((ret
> 0) || (nr_failed
&& (flags
& MPOL_MF_STRICT
)))
1327 if (!list_empty(&pagelist
))
1328 putback_movable_pages(&pagelist
);
1331 up_write(&mm
->mmap_sem
);
1338 * User space interface with variable sized bitmaps for nodelists.
1341 /* Copy a node mask from user space. */
1342 static int get_nodes(nodemask_t
*nodes
, const unsigned long __user
*nmask
,
1343 unsigned long maxnode
)
1347 unsigned long nlongs
;
1348 unsigned long endmask
;
1351 nodes_clear(*nodes
);
1352 if (maxnode
== 0 || !nmask
)
1354 if (maxnode
> PAGE_SIZE
*BITS_PER_BYTE
)
1357 nlongs
= BITS_TO_LONGS(maxnode
);
1358 if ((maxnode
% BITS_PER_LONG
) == 0)
1361 endmask
= (1UL << (maxnode
% BITS_PER_LONG
)) - 1;
1364 * When the user specified more nodes than supported just check
1365 * if the non supported part is all zero.
1367 * If maxnode have more longs than MAX_NUMNODES, check
1368 * the bits in that area first. And then go through to
1369 * check the rest bits which equal or bigger than MAX_NUMNODES.
1370 * Otherwise, just check bits [MAX_NUMNODES, maxnode).
1372 if (nlongs
> BITS_TO_LONGS(MAX_NUMNODES
)) {
1373 for (k
= BITS_TO_LONGS(MAX_NUMNODES
); k
< nlongs
; k
++) {
1374 if (get_user(t
, nmask
+ k
))
1376 if (k
== nlongs
- 1) {
1382 nlongs
= BITS_TO_LONGS(MAX_NUMNODES
);
1386 if (maxnode
> MAX_NUMNODES
&& MAX_NUMNODES
% BITS_PER_LONG
!= 0) {
1387 unsigned long valid_mask
= endmask
;
1389 valid_mask
&= ~((1UL << (MAX_NUMNODES
% BITS_PER_LONG
)) - 1);
1390 if (get_user(t
, nmask
+ nlongs
- 1))
1396 if (copy_from_user(nodes_addr(*nodes
), nmask
, nlongs
*sizeof(unsigned long)))
1398 nodes_addr(*nodes
)[nlongs
-1] &= endmask
;
1402 /* Copy a kernel node mask to user space */
1403 static int copy_nodes_to_user(unsigned long __user
*mask
, unsigned long maxnode
,
1406 unsigned long copy
= ALIGN(maxnode
-1, 64) / 8;
1407 unsigned int nbytes
= BITS_TO_LONGS(nr_node_ids
) * sizeof(long);
1409 if (copy
> nbytes
) {
1410 if (copy
> PAGE_SIZE
)
1412 if (clear_user((char __user
*)mask
+ nbytes
, copy
- nbytes
))
1416 return copy_to_user(mask
, nodes_addr(*nodes
), copy
) ? -EFAULT
: 0;
1419 static long kernel_mbind(unsigned long start
, unsigned long len
,
1420 unsigned long mode
, const unsigned long __user
*nmask
,
1421 unsigned long maxnode
, unsigned int flags
)
1425 unsigned short mode_flags
;
1427 start
= untagged_addr(start
);
1428 mode_flags
= mode
& MPOL_MODE_FLAGS
;
1429 mode
&= ~MPOL_MODE_FLAGS
;
1430 if (mode
>= MPOL_MAX
)
1432 if ((mode_flags
& MPOL_F_STATIC_NODES
) &&
1433 (mode_flags
& MPOL_F_RELATIVE_NODES
))
1435 err
= get_nodes(&nodes
, nmask
, maxnode
);
1438 return do_mbind(start
, len
, mode
, mode_flags
, &nodes
, flags
);
1441 SYSCALL_DEFINE6(mbind
, unsigned long, start
, unsigned long, len
,
1442 unsigned long, mode
, const unsigned long __user
*, nmask
,
1443 unsigned long, maxnode
, unsigned int, flags
)
1445 return kernel_mbind(start
, len
, mode
, nmask
, maxnode
, flags
);
1448 /* Set the process memory policy */
1449 static long kernel_set_mempolicy(int mode
, const unsigned long __user
*nmask
,
1450 unsigned long maxnode
)
1454 unsigned short flags
;
1456 flags
= mode
& MPOL_MODE_FLAGS
;
1457 mode
&= ~MPOL_MODE_FLAGS
;
1458 if ((unsigned int)mode
>= MPOL_MAX
)
1460 if ((flags
& MPOL_F_STATIC_NODES
) && (flags
& MPOL_F_RELATIVE_NODES
))
1462 err
= get_nodes(&nodes
, nmask
, maxnode
);
1465 return do_set_mempolicy(mode
, flags
, &nodes
);
1468 SYSCALL_DEFINE3(set_mempolicy
, int, mode
, const unsigned long __user
*, nmask
,
1469 unsigned long, maxnode
)
1471 return kernel_set_mempolicy(mode
, nmask
, maxnode
);
1474 static int kernel_migrate_pages(pid_t pid
, unsigned long maxnode
,
1475 const unsigned long __user
*old_nodes
,
1476 const unsigned long __user
*new_nodes
)
1478 struct mm_struct
*mm
= NULL
;
1479 struct task_struct
*task
;
1480 nodemask_t task_nodes
;
1484 NODEMASK_SCRATCH(scratch
);
1489 old
= &scratch
->mask1
;
1490 new = &scratch
->mask2
;
1492 err
= get_nodes(old
, old_nodes
, maxnode
);
1496 err
= get_nodes(new, new_nodes
, maxnode
);
1500 /* Find the mm_struct */
1502 task
= pid
? find_task_by_vpid(pid
) : current
;
1508 get_task_struct(task
);
1513 * Check if this process has the right to modify the specified process.
1514 * Use the regular "ptrace_may_access()" checks.
1516 if (!ptrace_may_access(task
, PTRACE_MODE_READ_REALCREDS
)) {
1523 task_nodes
= cpuset_mems_allowed(task
);
1524 /* Is the user allowed to access the target nodes? */
1525 if (!nodes_subset(*new, task_nodes
) && !capable(CAP_SYS_NICE
)) {
1530 task_nodes
= cpuset_mems_allowed(current
);
1531 nodes_and(*new, *new, task_nodes
);
1532 if (nodes_empty(*new))
1535 err
= security_task_movememory(task
);
1539 mm
= get_task_mm(task
);
1540 put_task_struct(task
);
1547 err
= do_migrate_pages(mm
, old
, new,
1548 capable(CAP_SYS_NICE
) ? MPOL_MF_MOVE_ALL
: MPOL_MF_MOVE
);
1552 NODEMASK_SCRATCH_FREE(scratch
);
1557 put_task_struct(task
);
1562 SYSCALL_DEFINE4(migrate_pages
, pid_t
, pid
, unsigned long, maxnode
,
1563 const unsigned long __user
*, old_nodes
,
1564 const unsigned long __user
*, new_nodes
)
1566 return kernel_migrate_pages(pid
, maxnode
, old_nodes
, new_nodes
);
1570 /* Retrieve NUMA policy */
1571 static int kernel_get_mempolicy(int __user
*policy
,
1572 unsigned long __user
*nmask
,
1573 unsigned long maxnode
,
1575 unsigned long flags
)
1578 int uninitialized_var(pval
);
1581 addr
= untagged_addr(addr
);
1583 if (nmask
!= NULL
&& maxnode
< nr_node_ids
)
1586 err
= do_get_mempolicy(&pval
, &nodes
, addr
, flags
);
1591 if (policy
&& put_user(pval
, policy
))
1595 err
= copy_nodes_to_user(nmask
, maxnode
, &nodes
);
1600 SYSCALL_DEFINE5(get_mempolicy
, int __user
*, policy
,
1601 unsigned long __user
*, nmask
, unsigned long, maxnode
,
1602 unsigned long, addr
, unsigned long, flags
)
1604 return kernel_get_mempolicy(policy
, nmask
, maxnode
, addr
, flags
);
1607 #ifdef CONFIG_COMPAT
1609 COMPAT_SYSCALL_DEFINE5(get_mempolicy
, int __user
*, policy
,
1610 compat_ulong_t __user
*, nmask
,
1611 compat_ulong_t
, maxnode
,
1612 compat_ulong_t
, addr
, compat_ulong_t
, flags
)
1615 unsigned long __user
*nm
= NULL
;
1616 unsigned long nr_bits
, alloc_size
;
1617 DECLARE_BITMAP(bm
, MAX_NUMNODES
);
1619 nr_bits
= min_t(unsigned long, maxnode
-1, nr_node_ids
);
1620 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
1623 nm
= compat_alloc_user_space(alloc_size
);
1625 err
= kernel_get_mempolicy(policy
, nm
, nr_bits
+1, addr
, flags
);
1627 if (!err
&& nmask
) {
1628 unsigned long copy_size
;
1629 copy_size
= min_t(unsigned long, sizeof(bm
), alloc_size
);
1630 err
= copy_from_user(bm
, nm
, copy_size
);
1631 /* ensure entire bitmap is zeroed */
1632 err
|= clear_user(nmask
, ALIGN(maxnode
-1, 8) / 8);
1633 err
|= compat_put_bitmap(nmask
, bm
, nr_bits
);
1639 COMPAT_SYSCALL_DEFINE3(set_mempolicy
, int, mode
, compat_ulong_t __user
*, nmask
,
1640 compat_ulong_t
, maxnode
)
1642 unsigned long __user
*nm
= NULL
;
1643 unsigned long nr_bits
, alloc_size
;
1644 DECLARE_BITMAP(bm
, MAX_NUMNODES
);
1646 nr_bits
= min_t(unsigned long, maxnode
-1, MAX_NUMNODES
);
1647 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
1650 if (compat_get_bitmap(bm
, nmask
, nr_bits
))
1652 nm
= compat_alloc_user_space(alloc_size
);
1653 if (copy_to_user(nm
, bm
, alloc_size
))
1657 return kernel_set_mempolicy(mode
, nm
, nr_bits
+1);
1660 COMPAT_SYSCALL_DEFINE6(mbind
, compat_ulong_t
, start
, compat_ulong_t
, len
,
1661 compat_ulong_t
, mode
, compat_ulong_t __user
*, nmask
,
1662 compat_ulong_t
, maxnode
, compat_ulong_t
, flags
)
1664 unsigned long __user
*nm
= NULL
;
1665 unsigned long nr_bits
, alloc_size
;
1668 nr_bits
= min_t(unsigned long, maxnode
-1, MAX_NUMNODES
);
1669 alloc_size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
1672 if (compat_get_bitmap(nodes_addr(bm
), nmask
, nr_bits
))
1674 nm
= compat_alloc_user_space(alloc_size
);
1675 if (copy_to_user(nm
, nodes_addr(bm
), alloc_size
))
1679 return kernel_mbind(start
, len
, mode
, nm
, nr_bits
+1, flags
);
1682 COMPAT_SYSCALL_DEFINE4(migrate_pages
, compat_pid_t
, pid
,
1683 compat_ulong_t
, maxnode
,
1684 const compat_ulong_t __user
*, old_nodes
,
1685 const compat_ulong_t __user
*, new_nodes
)
1687 unsigned long __user
*old
= NULL
;
1688 unsigned long __user
*new = NULL
;
1689 nodemask_t tmp_mask
;
1690 unsigned long nr_bits
;
1693 nr_bits
= min_t(unsigned long, maxnode
- 1, MAX_NUMNODES
);
1694 size
= ALIGN(nr_bits
, BITS_PER_LONG
) / 8;
1696 if (compat_get_bitmap(nodes_addr(tmp_mask
), old_nodes
, nr_bits
))
1698 old
= compat_alloc_user_space(new_nodes
? size
* 2 : size
);
1700 new = old
+ size
/ sizeof(unsigned long);
1701 if (copy_to_user(old
, nodes_addr(tmp_mask
), size
))
1705 if (compat_get_bitmap(nodes_addr(tmp_mask
), new_nodes
, nr_bits
))
1708 new = compat_alloc_user_space(size
);
1709 if (copy_to_user(new, nodes_addr(tmp_mask
), size
))
1712 return kernel_migrate_pages(pid
, nr_bits
+ 1, old
, new);
1715 #endif /* CONFIG_COMPAT */
1717 struct mempolicy
*__get_vma_policy(struct vm_area_struct
*vma
,
1720 struct mempolicy
*pol
= NULL
;
1723 if (vma
->vm_ops
&& vma
->vm_ops
->get_policy
) {
1724 pol
= vma
->vm_ops
->get_policy(vma
, addr
);
1725 } else if (vma
->vm_policy
) {
1726 pol
= vma
->vm_policy
;
1729 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1730 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1731 * count on these policies which will be dropped by
1732 * mpol_cond_put() later
1734 if (mpol_needs_cond_ref(pol
))
1743 * get_vma_policy(@vma, @addr)
1744 * @vma: virtual memory area whose policy is sought
1745 * @addr: address in @vma for shared policy lookup
1747 * Returns effective policy for a VMA at specified address.
1748 * Falls back to current->mempolicy or system default policy, as necessary.
1749 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1750 * count--added by the get_policy() vm_op, as appropriate--to protect against
1751 * freeing by another task. It is the caller's responsibility to free the
1752 * extra reference for shared policies.
1754 static struct mempolicy
*get_vma_policy(struct vm_area_struct
*vma
,
1757 struct mempolicy
*pol
= __get_vma_policy(vma
, addr
);
1760 pol
= get_task_policy(current
);
1765 bool vma_policy_mof(struct vm_area_struct
*vma
)
1767 struct mempolicy
*pol
;
1769 if (vma
->vm_ops
&& vma
->vm_ops
->get_policy
) {
1772 pol
= vma
->vm_ops
->get_policy(vma
, vma
->vm_start
);
1773 if (pol
&& (pol
->flags
& MPOL_F_MOF
))
1780 pol
= vma
->vm_policy
;
1782 pol
= get_task_policy(current
);
1784 return pol
->flags
& MPOL_F_MOF
;
1787 static int apply_policy_zone(struct mempolicy
*policy
, enum zone_type zone
)
1789 enum zone_type dynamic_policy_zone
= policy_zone
;
1791 BUG_ON(dynamic_policy_zone
== ZONE_MOVABLE
);
1794 * if policy->v.nodes has movable memory only,
1795 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1797 * policy->v.nodes is intersect with node_states[N_MEMORY].
1798 * so if the following test faile, it implies
1799 * policy->v.nodes has movable memory only.
1801 if (!nodes_intersects(policy
->v
.nodes
, node_states
[N_HIGH_MEMORY
]))
1802 dynamic_policy_zone
= ZONE_MOVABLE
;
1804 return zone
>= dynamic_policy_zone
;
1808 * Return a nodemask representing a mempolicy for filtering nodes for
1811 static nodemask_t
*policy_nodemask(gfp_t gfp
, struct mempolicy
*policy
)
1813 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1814 if (unlikely(policy
->mode
== MPOL_BIND
) &&
1815 apply_policy_zone(policy
, gfp_zone(gfp
)) &&
1816 cpuset_nodemask_valid_mems_allowed(&policy
->v
.nodes
))
1817 return &policy
->v
.nodes
;
1822 /* Return the node id preferred by the given mempolicy, or the given id */
1823 static int policy_node(gfp_t gfp
, struct mempolicy
*policy
,
1826 if (policy
->mode
== MPOL_PREFERRED
&& !(policy
->flags
& MPOL_F_LOCAL
))
1827 nd
= policy
->v
.preferred_node
;
1830 * __GFP_THISNODE shouldn't even be used with the bind policy
1831 * because we might easily break the expectation to stay on the
1832 * requested node and not break the policy.
1834 WARN_ON_ONCE(policy
->mode
== MPOL_BIND
&& (gfp
& __GFP_THISNODE
));
1840 /* Do dynamic interleaving for a process */
1841 static unsigned interleave_nodes(struct mempolicy
*policy
)
1844 struct task_struct
*me
= current
;
1846 next
= next_node_in(me
->il_prev
, policy
->v
.nodes
);
1847 if (next
< MAX_NUMNODES
)
1853 * Depending on the memory policy provide a node from which to allocate the
1856 unsigned int mempolicy_slab_node(void)
1858 struct mempolicy
*policy
;
1859 int node
= numa_mem_id();
1864 policy
= current
->mempolicy
;
1865 if (!policy
|| policy
->flags
& MPOL_F_LOCAL
)
1868 switch (policy
->mode
) {
1869 case MPOL_PREFERRED
:
1871 * handled MPOL_F_LOCAL above
1873 return policy
->v
.preferred_node
;
1875 case MPOL_INTERLEAVE
:
1876 return interleave_nodes(policy
);
1882 * Follow bind policy behavior and start allocation at the
1885 struct zonelist
*zonelist
;
1886 enum zone_type highest_zoneidx
= gfp_zone(GFP_KERNEL
);
1887 zonelist
= &NODE_DATA(node
)->node_zonelists
[ZONELIST_FALLBACK
];
1888 z
= first_zones_zonelist(zonelist
, highest_zoneidx
,
1890 return z
->zone
? zone_to_nid(z
->zone
) : node
;
1899 * Do static interleaving for a VMA with known offset @n. Returns the n'th
1900 * node in pol->v.nodes (starting from n=0), wrapping around if n exceeds the
1901 * number of present nodes.
1903 static unsigned offset_il_node(struct mempolicy
*pol
, unsigned long n
)
1905 unsigned nnodes
= nodes_weight(pol
->v
.nodes
);
1911 return numa_node_id();
1912 target
= (unsigned int)n
% nnodes
;
1913 nid
= first_node(pol
->v
.nodes
);
1914 for (i
= 0; i
< target
; i
++)
1915 nid
= next_node(nid
, pol
->v
.nodes
);
1919 /* Determine a node number for interleave */
1920 static inline unsigned interleave_nid(struct mempolicy
*pol
,
1921 struct vm_area_struct
*vma
, unsigned long addr
, int shift
)
1927 * for small pages, there is no difference between
1928 * shift and PAGE_SHIFT, so the bit-shift is safe.
1929 * for huge pages, since vm_pgoff is in units of small
1930 * pages, we need to shift off the always 0 bits to get
1933 BUG_ON(shift
< PAGE_SHIFT
);
1934 off
= vma
->vm_pgoff
>> (shift
- PAGE_SHIFT
);
1935 off
+= (addr
- vma
->vm_start
) >> shift
;
1936 return offset_il_node(pol
, off
);
1938 return interleave_nodes(pol
);
1941 #ifdef CONFIG_HUGETLBFS
1943 * huge_node(@vma, @addr, @gfp_flags, @mpol)
1944 * @vma: virtual memory area whose policy is sought
1945 * @addr: address in @vma for shared policy lookup and interleave policy
1946 * @gfp_flags: for requested zone
1947 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
1948 * @nodemask: pointer to nodemask pointer for MPOL_BIND nodemask
1950 * Returns a nid suitable for a huge page allocation and a pointer
1951 * to the struct mempolicy for conditional unref after allocation.
1952 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1953 * @nodemask for filtering the zonelist.
1955 * Must be protected by read_mems_allowed_begin()
1957 int huge_node(struct vm_area_struct
*vma
, unsigned long addr
, gfp_t gfp_flags
,
1958 struct mempolicy
**mpol
, nodemask_t
**nodemask
)
1962 *mpol
= get_vma_policy(vma
, addr
);
1963 *nodemask
= NULL
; /* assume !MPOL_BIND */
1965 if (unlikely((*mpol
)->mode
== MPOL_INTERLEAVE
)) {
1966 nid
= interleave_nid(*mpol
, vma
, addr
,
1967 huge_page_shift(hstate_vma(vma
)));
1969 nid
= policy_node(gfp_flags
, *mpol
, numa_node_id());
1970 if ((*mpol
)->mode
== MPOL_BIND
)
1971 *nodemask
= &(*mpol
)->v
.nodes
;
1977 * init_nodemask_of_mempolicy
1979 * If the current task's mempolicy is "default" [NULL], return 'false'
1980 * to indicate default policy. Otherwise, extract the policy nodemask
1981 * for 'bind' or 'interleave' policy into the argument nodemask, or
1982 * initialize the argument nodemask to contain the single node for
1983 * 'preferred' or 'local' policy and return 'true' to indicate presence
1984 * of non-default mempolicy.
1986 * We don't bother with reference counting the mempolicy [mpol_get/put]
1987 * because the current task is examining it's own mempolicy and a task's
1988 * mempolicy is only ever changed by the task itself.
1990 * N.B., it is the caller's responsibility to free a returned nodemask.
1992 bool init_nodemask_of_mempolicy(nodemask_t
*mask
)
1994 struct mempolicy
*mempolicy
;
1997 if (!(mask
&& current
->mempolicy
))
2001 mempolicy
= current
->mempolicy
;
2002 switch (mempolicy
->mode
) {
2003 case MPOL_PREFERRED
:
2004 if (mempolicy
->flags
& MPOL_F_LOCAL
)
2005 nid
= numa_node_id();
2007 nid
= mempolicy
->v
.preferred_node
;
2008 init_nodemask_of_node(mask
, nid
);
2013 case MPOL_INTERLEAVE
:
2014 *mask
= mempolicy
->v
.nodes
;
2020 task_unlock(current
);
2027 * mempolicy_nodemask_intersects
2029 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
2030 * policy. Otherwise, check for intersection between mask and the policy
2031 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
2032 * policy, always return true since it may allocate elsewhere on fallback.
2034 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
2036 bool mempolicy_nodemask_intersects(struct task_struct
*tsk
,
2037 const nodemask_t
*mask
)
2039 struct mempolicy
*mempolicy
;
2045 mempolicy
= tsk
->mempolicy
;
2049 switch (mempolicy
->mode
) {
2050 case MPOL_PREFERRED
:
2052 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
2053 * allocate from, they may fallback to other nodes when oom.
2054 * Thus, it's possible for tsk to have allocated memory from
2059 case MPOL_INTERLEAVE
:
2060 ret
= nodes_intersects(mempolicy
->v
.nodes
, *mask
);
2070 /* Allocate a page in interleaved policy.
2071 Own path because it needs to do special accounting. */
2072 static struct page
*alloc_page_interleave(gfp_t gfp
, unsigned order
,
2077 page
= __alloc_pages(gfp
, order
, nid
);
2078 /* skip NUMA_INTERLEAVE_HIT counter update if numa stats is disabled */
2079 if (!static_branch_likely(&vm_numa_stat_key
))
2081 if (page
&& page_to_nid(page
) == nid
) {
2083 __inc_numa_state(page_zone(page
), NUMA_INTERLEAVE_HIT
);
2090 * alloc_pages_vma - Allocate a page for a VMA.
2093 * %GFP_USER user allocation.
2094 * %GFP_KERNEL kernel allocations,
2095 * %GFP_HIGHMEM highmem/user allocations,
2096 * %GFP_FS allocation should not call back into a file system.
2097 * %GFP_ATOMIC don't sleep.
2099 * @order:Order of the GFP allocation.
2100 * @vma: Pointer to VMA or NULL if not available.
2101 * @addr: Virtual Address of the allocation. Must be inside the VMA.
2102 * @node: Which node to prefer for allocation (modulo policy).
2103 * @hugepage: for hugepages try only the preferred node if possible
2105 * This function allocates a page from the kernel page pool and applies
2106 * a NUMA policy associated with the VMA or the current process.
2107 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
2108 * mm_struct of the VMA to prevent it from going away. Should be used for
2109 * all allocations for pages that will be mapped into user space. Returns
2110 * NULL when no page can be allocated.
2113 alloc_pages_vma(gfp_t gfp
, int order
, struct vm_area_struct
*vma
,
2114 unsigned long addr
, int node
, bool hugepage
)
2116 struct mempolicy
*pol
;
2121 pol
= get_vma_policy(vma
, addr
);
2123 if (pol
->mode
== MPOL_INTERLEAVE
) {
2126 nid
= interleave_nid(pol
, vma
, addr
, PAGE_SHIFT
+ order
);
2128 page
= alloc_page_interleave(gfp
, order
, nid
);
2132 if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE
) && hugepage
)) {
2133 int hpage_node
= node
;
2136 * For hugepage allocation and non-interleave policy which
2137 * allows the current node (or other explicitly preferred
2138 * node) we only try to allocate from the current/preferred
2139 * node and don't fall back to other nodes, as the cost of
2140 * remote accesses would likely offset THP benefits.
2142 * If the policy is interleave, or does not allow the current
2143 * node in its nodemask, we allocate the standard way.
2145 if (pol
->mode
== MPOL_PREFERRED
&& !(pol
->flags
& MPOL_F_LOCAL
))
2146 hpage_node
= pol
->v
.preferred_node
;
2148 nmask
= policy_nodemask(gfp
, pol
);
2149 if (!nmask
|| node_isset(hpage_node
, *nmask
)) {
2152 * First, try to allocate THP only on local node, but
2153 * don't reclaim unnecessarily, just compact.
2155 page
= __alloc_pages_node(hpage_node
,
2156 gfp
| __GFP_THISNODE
| __GFP_NORETRY
, order
);
2159 * If hugepage allocations are configured to always
2160 * synchronous compact or the vma has been madvised
2161 * to prefer hugepage backing, retry allowing remote
2162 * memory with both reclaim and compact as well.
2164 if (!page
&& (gfp
& __GFP_DIRECT_RECLAIM
))
2165 page
= __alloc_pages_node(hpage_node
,
2172 nmask
= policy_nodemask(gfp
, pol
);
2173 preferred_nid
= policy_node(gfp
, pol
, node
);
2174 page
= __alloc_pages_nodemask(gfp
, order
, preferred_nid
, nmask
);
2179 EXPORT_SYMBOL(alloc_pages_vma
);
2182 * alloc_pages_current - Allocate pages.
2185 * %GFP_USER user allocation,
2186 * %GFP_KERNEL kernel allocation,
2187 * %GFP_HIGHMEM highmem allocation,
2188 * %GFP_FS don't call back into a file system.
2189 * %GFP_ATOMIC don't sleep.
2190 * @order: Power of two of allocation size in pages. 0 is a single page.
2192 * Allocate a page from the kernel page pool. When not in
2193 * interrupt context and apply the current process NUMA policy.
2194 * Returns NULL when no page can be allocated.
2196 struct page
*alloc_pages_current(gfp_t gfp
, unsigned order
)
2198 struct mempolicy
*pol
= &default_policy
;
2201 if (!in_interrupt() && !(gfp
& __GFP_THISNODE
))
2202 pol
= get_task_policy(current
);
2205 * No reference counting needed for current->mempolicy
2206 * nor system default_policy
2208 if (pol
->mode
== MPOL_INTERLEAVE
)
2209 page
= alloc_page_interleave(gfp
, order
, interleave_nodes(pol
));
2211 page
= __alloc_pages_nodemask(gfp
, order
,
2212 policy_node(gfp
, pol
, numa_node_id()),
2213 policy_nodemask(gfp
, pol
));
2217 EXPORT_SYMBOL(alloc_pages_current
);
2219 int vma_dup_policy(struct vm_area_struct
*src
, struct vm_area_struct
*dst
)
2221 struct mempolicy
*pol
= mpol_dup(vma_policy(src
));
2224 return PTR_ERR(pol
);
2225 dst
->vm_policy
= pol
;
2230 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2231 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2232 * with the mems_allowed returned by cpuset_mems_allowed(). This
2233 * keeps mempolicies cpuset relative after its cpuset moves. See
2234 * further kernel/cpuset.c update_nodemask().
2236 * current's mempolicy may be rebinded by the other task(the task that changes
2237 * cpuset's mems), so we needn't do rebind work for current task.
2240 /* Slow path of a mempolicy duplicate */
2241 struct mempolicy
*__mpol_dup(struct mempolicy
*old
)
2243 struct mempolicy
*new = kmem_cache_alloc(policy_cache
, GFP_KERNEL
);
2246 return ERR_PTR(-ENOMEM
);
2248 /* task's mempolicy is protected by alloc_lock */
2249 if (old
== current
->mempolicy
) {
2252 task_unlock(current
);
2256 if (current_cpuset_is_being_rebound()) {
2257 nodemask_t mems
= cpuset_mems_allowed(current
);
2258 mpol_rebind_policy(new, &mems
);
2260 atomic_set(&new->refcnt
, 1);
2264 /* Slow path of a mempolicy comparison */
2265 bool __mpol_equal(struct mempolicy
*a
, struct mempolicy
*b
)
2269 if (a
->mode
!= b
->mode
)
2271 if (a
->flags
!= b
->flags
)
2273 if (mpol_store_user_nodemask(a
))
2274 if (!nodes_equal(a
->w
.user_nodemask
, b
->w
.user_nodemask
))
2280 case MPOL_INTERLEAVE
:
2281 return !!nodes_equal(a
->v
.nodes
, b
->v
.nodes
);
2282 case MPOL_PREFERRED
:
2283 /* a's ->flags is the same as b's */
2284 if (a
->flags
& MPOL_F_LOCAL
)
2286 return a
->v
.preferred_node
== b
->v
.preferred_node
;
2294 * Shared memory backing store policy support.
2296 * Remember policies even when nobody has shared memory mapped.
2297 * The policies are kept in Red-Black tree linked from the inode.
2298 * They are protected by the sp->lock rwlock, which should be held
2299 * for any accesses to the tree.
2303 * lookup first element intersecting start-end. Caller holds sp->lock for
2304 * reading or for writing
2306 static struct sp_node
*
2307 sp_lookup(struct shared_policy
*sp
, unsigned long start
, unsigned long end
)
2309 struct rb_node
*n
= sp
->root
.rb_node
;
2312 struct sp_node
*p
= rb_entry(n
, struct sp_node
, nd
);
2314 if (start
>= p
->end
)
2316 else if (end
<= p
->start
)
2324 struct sp_node
*w
= NULL
;
2325 struct rb_node
*prev
= rb_prev(n
);
2328 w
= rb_entry(prev
, struct sp_node
, nd
);
2329 if (w
->end
<= start
)
2333 return rb_entry(n
, struct sp_node
, nd
);
2337 * Insert a new shared policy into the list. Caller holds sp->lock for
2340 static void sp_insert(struct shared_policy
*sp
, struct sp_node
*new)
2342 struct rb_node
**p
= &sp
->root
.rb_node
;
2343 struct rb_node
*parent
= NULL
;
2348 nd
= rb_entry(parent
, struct sp_node
, nd
);
2349 if (new->start
< nd
->start
)
2351 else if (new->end
> nd
->end
)
2352 p
= &(*p
)->rb_right
;
2356 rb_link_node(&new->nd
, parent
, p
);
2357 rb_insert_color(&new->nd
, &sp
->root
);
2358 pr_debug("inserting %lx-%lx: %d\n", new->start
, new->end
,
2359 new->policy
? new->policy
->mode
: 0);
2362 /* Find shared policy intersecting idx */
2364 mpol_shared_policy_lookup(struct shared_policy
*sp
, unsigned long idx
)
2366 struct mempolicy
*pol
= NULL
;
2369 if (!sp
->root
.rb_node
)
2371 read_lock(&sp
->lock
);
2372 sn
= sp_lookup(sp
, idx
, idx
+1);
2374 mpol_get(sn
->policy
);
2377 read_unlock(&sp
->lock
);
2381 static void sp_free(struct sp_node
*n
)
2383 mpol_put(n
->policy
);
2384 kmem_cache_free(sn_cache
, n
);
2388 * mpol_misplaced - check whether current page node is valid in policy
2390 * @page: page to be checked
2391 * @vma: vm area where page mapped
2392 * @addr: virtual address where page mapped
2394 * Lookup current policy node id for vma,addr and "compare to" page's
2398 * -1 - not misplaced, page is in the right node
2399 * node - node id where the page should be
2401 * Policy determination "mimics" alloc_page_vma().
2402 * Called from fault path where we know the vma and faulting address.
2404 int mpol_misplaced(struct page
*page
, struct vm_area_struct
*vma
, unsigned long addr
)
2406 struct mempolicy
*pol
;
2408 int curnid
= page_to_nid(page
);
2409 unsigned long pgoff
;
2410 int thiscpu
= raw_smp_processor_id();
2411 int thisnid
= cpu_to_node(thiscpu
);
2412 int polnid
= NUMA_NO_NODE
;
2415 pol
= get_vma_policy(vma
, addr
);
2416 if (!(pol
->flags
& MPOL_F_MOF
))
2419 switch (pol
->mode
) {
2420 case MPOL_INTERLEAVE
:
2421 pgoff
= vma
->vm_pgoff
;
2422 pgoff
+= (addr
- vma
->vm_start
) >> PAGE_SHIFT
;
2423 polnid
= offset_il_node(pol
, pgoff
);
2426 case MPOL_PREFERRED
:
2427 if (pol
->flags
& MPOL_F_LOCAL
)
2428 polnid
= numa_node_id();
2430 polnid
= pol
->v
.preferred_node
;
2436 * allows binding to multiple nodes.
2437 * use current page if in policy nodemask,
2438 * else select nearest allowed node, if any.
2439 * If no allowed nodes, use current [!misplaced].
2441 if (node_isset(curnid
, pol
->v
.nodes
))
2443 z
= first_zones_zonelist(
2444 node_zonelist(numa_node_id(), GFP_HIGHUSER
),
2445 gfp_zone(GFP_HIGHUSER
),
2447 polnid
= zone_to_nid(z
->zone
);
2454 /* Migrate the page towards the node whose CPU is referencing it */
2455 if (pol
->flags
& MPOL_F_MORON
) {
2458 if (!should_numa_migrate_memory(current
, page
, curnid
, thiscpu
))
2462 if (curnid
!= polnid
)
2471 * Drop the (possibly final) reference to task->mempolicy. It needs to be
2472 * dropped after task->mempolicy is set to NULL so that any allocation done as
2473 * part of its kmem_cache_free(), such as by KASAN, doesn't reference a freed
2476 void mpol_put_task_policy(struct task_struct
*task
)
2478 struct mempolicy
*pol
;
2481 pol
= task
->mempolicy
;
2482 task
->mempolicy
= NULL
;
2487 static void sp_delete(struct shared_policy
*sp
, struct sp_node
*n
)
2489 pr_debug("deleting %lx-l%lx\n", n
->start
, n
->end
);
2490 rb_erase(&n
->nd
, &sp
->root
);
2494 static void sp_node_init(struct sp_node
*node
, unsigned long start
,
2495 unsigned long end
, struct mempolicy
*pol
)
2497 node
->start
= start
;
2502 static struct sp_node
*sp_alloc(unsigned long start
, unsigned long end
,
2503 struct mempolicy
*pol
)
2506 struct mempolicy
*newpol
;
2508 n
= kmem_cache_alloc(sn_cache
, GFP_KERNEL
);
2512 newpol
= mpol_dup(pol
);
2513 if (IS_ERR(newpol
)) {
2514 kmem_cache_free(sn_cache
, n
);
2517 newpol
->flags
|= MPOL_F_SHARED
;
2518 sp_node_init(n
, start
, end
, newpol
);
2523 /* Replace a policy range. */
2524 static int shared_policy_replace(struct shared_policy
*sp
, unsigned long start
,
2525 unsigned long end
, struct sp_node
*new)
2528 struct sp_node
*n_new
= NULL
;
2529 struct mempolicy
*mpol_new
= NULL
;
2533 write_lock(&sp
->lock
);
2534 n
= sp_lookup(sp
, start
, end
);
2535 /* Take care of old policies in the same range. */
2536 while (n
&& n
->start
< end
) {
2537 struct rb_node
*next
= rb_next(&n
->nd
);
2538 if (n
->start
>= start
) {
2544 /* Old policy spanning whole new range. */
2549 *mpol_new
= *n
->policy
;
2550 atomic_set(&mpol_new
->refcnt
, 1);
2551 sp_node_init(n_new
, end
, n
->end
, mpol_new
);
2553 sp_insert(sp
, n_new
);
2562 n
= rb_entry(next
, struct sp_node
, nd
);
2566 write_unlock(&sp
->lock
);
2573 kmem_cache_free(sn_cache
, n_new
);
2578 write_unlock(&sp
->lock
);
2580 n_new
= kmem_cache_alloc(sn_cache
, GFP_KERNEL
);
2583 mpol_new
= kmem_cache_alloc(policy_cache
, GFP_KERNEL
);
2590 * mpol_shared_policy_init - initialize shared policy for inode
2591 * @sp: pointer to inode shared policy
2592 * @mpol: struct mempolicy to install
2594 * Install non-NULL @mpol in inode's shared policy rb-tree.
2595 * On entry, the current task has a reference on a non-NULL @mpol.
2596 * This must be released on exit.
2597 * This is called at get_inode() calls and we can use GFP_KERNEL.
2599 void mpol_shared_policy_init(struct shared_policy
*sp
, struct mempolicy
*mpol
)
2603 sp
->root
= RB_ROOT
; /* empty tree == default mempolicy */
2604 rwlock_init(&sp
->lock
);
2607 struct vm_area_struct pvma
;
2608 struct mempolicy
*new;
2609 NODEMASK_SCRATCH(scratch
);
2613 /* contextualize the tmpfs mount point mempolicy */
2614 new = mpol_new(mpol
->mode
, mpol
->flags
, &mpol
->w
.user_nodemask
);
2616 goto free_scratch
; /* no valid nodemask intersection */
2619 ret
= mpol_set_nodemask(new, &mpol
->w
.user_nodemask
, scratch
);
2620 task_unlock(current
);
2624 /* Create pseudo-vma that contains just the policy */
2625 vma_init(&pvma
, NULL
);
2626 pvma
.vm_end
= TASK_SIZE
; /* policy covers entire file */
2627 mpol_set_shared_policy(sp
, &pvma
, new); /* adds ref */
2630 mpol_put(new); /* drop initial ref */
2632 NODEMASK_SCRATCH_FREE(scratch
);
2634 mpol_put(mpol
); /* drop our incoming ref on sb mpol */
2638 int mpol_set_shared_policy(struct shared_policy
*info
,
2639 struct vm_area_struct
*vma
, struct mempolicy
*npol
)
2642 struct sp_node
*new = NULL
;
2643 unsigned long sz
= vma_pages(vma
);
2645 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2647 sz
, npol
? npol
->mode
: -1,
2648 npol
? npol
->flags
: -1,
2649 npol
? nodes_addr(npol
->v
.nodes
)[0] : NUMA_NO_NODE
);
2652 new = sp_alloc(vma
->vm_pgoff
, vma
->vm_pgoff
+ sz
, npol
);
2656 err
= shared_policy_replace(info
, vma
->vm_pgoff
, vma
->vm_pgoff
+sz
, new);
2662 /* Free a backing policy store on inode delete. */
2663 void mpol_free_shared_policy(struct shared_policy
*p
)
2666 struct rb_node
*next
;
2668 if (!p
->root
.rb_node
)
2670 write_lock(&p
->lock
);
2671 next
= rb_first(&p
->root
);
2673 n
= rb_entry(next
, struct sp_node
, nd
);
2674 next
= rb_next(&n
->nd
);
2677 write_unlock(&p
->lock
);
2680 #ifdef CONFIG_NUMA_BALANCING
2681 static int __initdata numabalancing_override
;
2683 static void __init
check_numabalancing_enable(void)
2685 bool numabalancing_default
= false;
2687 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED
))
2688 numabalancing_default
= true;
2690 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2691 if (numabalancing_override
)
2692 set_numabalancing_state(numabalancing_override
== 1);
2694 if (num_online_nodes() > 1 && !numabalancing_override
) {
2695 pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
2696 numabalancing_default
? "Enabling" : "Disabling");
2697 set_numabalancing_state(numabalancing_default
);
2701 static int __init
setup_numabalancing(char *str
)
2707 if (!strcmp(str
, "enable")) {
2708 numabalancing_override
= 1;
2710 } else if (!strcmp(str
, "disable")) {
2711 numabalancing_override
= -1;
2716 pr_warn("Unable to parse numa_balancing=\n");
2720 __setup("numa_balancing=", setup_numabalancing
);
2722 static inline void __init
check_numabalancing_enable(void)
2725 #endif /* CONFIG_NUMA_BALANCING */
2727 /* assumes fs == KERNEL_DS */
2728 void __init
numa_policy_init(void)
2730 nodemask_t interleave_nodes
;
2731 unsigned long largest
= 0;
2732 int nid
, prefer
= 0;
2734 policy_cache
= kmem_cache_create("numa_policy",
2735 sizeof(struct mempolicy
),
2736 0, SLAB_PANIC
, NULL
);
2738 sn_cache
= kmem_cache_create("shared_policy_node",
2739 sizeof(struct sp_node
),
2740 0, SLAB_PANIC
, NULL
);
2742 for_each_node(nid
) {
2743 preferred_node_policy
[nid
] = (struct mempolicy
) {
2744 .refcnt
= ATOMIC_INIT(1),
2745 .mode
= MPOL_PREFERRED
,
2746 .flags
= MPOL_F_MOF
| MPOL_F_MORON
,
2747 .v
= { .preferred_node
= nid
, },
2752 * Set interleaving policy for system init. Interleaving is only
2753 * enabled across suitably sized nodes (default is >= 16MB), or
2754 * fall back to the largest node if they're all smaller.
2756 nodes_clear(interleave_nodes
);
2757 for_each_node_state(nid
, N_MEMORY
) {
2758 unsigned long total_pages
= node_present_pages(nid
);
2760 /* Preserve the largest node */
2761 if (largest
< total_pages
) {
2762 largest
= total_pages
;
2766 /* Interleave this node? */
2767 if ((total_pages
<< PAGE_SHIFT
) >= (16 << 20))
2768 node_set(nid
, interleave_nodes
);
2771 /* All too small, use the largest */
2772 if (unlikely(nodes_empty(interleave_nodes
)))
2773 node_set(prefer
, interleave_nodes
);
2775 if (do_set_mempolicy(MPOL_INTERLEAVE
, 0, &interleave_nodes
))
2776 pr_err("%s: interleaving failed\n", __func__
);
2778 check_numabalancing_enable();
2781 /* Reset policy of current process to default */
2782 void numa_default_policy(void)
2784 do_set_mempolicy(MPOL_DEFAULT
, 0, NULL
);
2788 * Parse and format mempolicy from/to strings
2792 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
2794 static const char * const policy_modes
[] =
2796 [MPOL_DEFAULT
] = "default",
2797 [MPOL_PREFERRED
] = "prefer",
2798 [MPOL_BIND
] = "bind",
2799 [MPOL_INTERLEAVE
] = "interleave",
2800 [MPOL_LOCAL
] = "local",
2806 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2807 * @str: string containing mempolicy to parse
2808 * @mpol: pointer to struct mempolicy pointer, returned on success.
2811 * <mode>[=<flags>][:<nodelist>]
2813 * On success, returns 0, else 1
2815 int mpol_parse_str(char *str
, struct mempolicy
**mpol
)
2817 struct mempolicy
*new = NULL
;
2818 unsigned short mode_flags
;
2820 char *nodelist
= strchr(str
, ':');
2821 char *flags
= strchr(str
, '=');
2825 /* NUL-terminate mode or flags string */
2827 if (nodelist_parse(nodelist
, nodes
))
2829 if (!nodes_subset(nodes
, node_states
[N_MEMORY
]))
2835 *flags
++ = '\0'; /* terminate mode string */
2837 mode
= match_string(policy_modes
, MPOL_MAX
, str
);
2842 case MPOL_PREFERRED
:
2844 * Insist on a nodelist of one node only
2847 char *rest
= nodelist
;
2848 while (isdigit(*rest
))
2854 case MPOL_INTERLEAVE
:
2856 * Default to online nodes with memory if no nodelist
2859 nodes
= node_states
[N_MEMORY
];
2863 * Don't allow a nodelist; mpol_new() checks flags
2867 mode
= MPOL_PREFERRED
;
2871 * Insist on a empty nodelist
2878 * Insist on a nodelist
2887 * Currently, we only support two mutually exclusive
2890 if (!strcmp(flags
, "static"))
2891 mode_flags
|= MPOL_F_STATIC_NODES
;
2892 else if (!strcmp(flags
, "relative"))
2893 mode_flags
|= MPOL_F_RELATIVE_NODES
;
2898 new = mpol_new(mode
, mode_flags
, &nodes
);
2903 * Save nodes for mpol_to_str() to show the tmpfs mount options
2904 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2906 if (mode
!= MPOL_PREFERRED
)
2907 new->v
.nodes
= nodes
;
2909 new->v
.preferred_node
= first_node(nodes
);
2911 new->flags
|= MPOL_F_LOCAL
;
2914 * Save nodes for contextualization: this will be used to "clone"
2915 * the mempolicy in a specific context [cpuset] at a later time.
2917 new->w
.user_nodemask
= nodes
;
2922 /* Restore string for error message */
2931 #endif /* CONFIG_TMPFS */
2934 * mpol_to_str - format a mempolicy structure for printing
2935 * @buffer: to contain formatted mempolicy string
2936 * @maxlen: length of @buffer
2937 * @pol: pointer to mempolicy to be formatted
2939 * Convert @pol into a string. If @buffer is too short, truncate the string.
2940 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
2941 * longest flag, "relative", and to display at least a few node ids.
2943 void mpol_to_str(char *buffer
, int maxlen
, struct mempolicy
*pol
)
2946 nodemask_t nodes
= NODE_MASK_NONE
;
2947 unsigned short mode
= MPOL_DEFAULT
;
2948 unsigned short flags
= 0;
2950 if (pol
&& pol
!= &default_policy
&& !(pol
->flags
& MPOL_F_MORON
)) {
2958 case MPOL_PREFERRED
:
2959 if (flags
& MPOL_F_LOCAL
)
2962 node_set(pol
->v
.preferred_node
, nodes
);
2965 case MPOL_INTERLEAVE
:
2966 nodes
= pol
->v
.nodes
;
2970 snprintf(p
, maxlen
, "unknown");
2974 p
+= snprintf(p
, maxlen
, "%s", policy_modes
[mode
]);
2976 if (flags
& MPOL_MODE_FLAGS
) {
2977 p
+= snprintf(p
, buffer
+ maxlen
- p
, "=");
2980 * Currently, the only defined flags are mutually exclusive
2982 if (flags
& MPOL_F_STATIC_NODES
)
2983 p
+= snprintf(p
, buffer
+ maxlen
- p
, "static");
2984 else if (flags
& MPOL_F_RELATIVE_NODES
)
2985 p
+= snprintf(p
, buffer
+ maxlen
- p
, "relative");
2988 if (!nodes_empty(nodes
))
2989 p
+= scnprintf(p
, buffer
+ maxlen
- p
, ":%*pbl",
2990 nodemask_pr_args(&nodes
));