2 * Simple NUMA memory policy for the Linux kernel.
4 * Copyright 2003,2004 Andi Kleen, SuSE Labs.
5 * (C) Copyright 2005 Christoph Lameter, Silicon Graphics, Inc.
6 * Subject to the GNU Public License, version 2.
8 * NUMA policy allows the user to give hints in which node(s) memory should
11 * Support four policies per VMA and per process:
13 * The VMA policy has priority over the process policy for a page fault.
15 * interleave Allocate memory interleaved over a set of nodes,
16 * with normal fallback if it fails.
17 * For VMA based allocations this interleaves based on the
18 * offset into the backing object or offset into the mapping
19 * for anonymous memory. For process policy an process counter
22 * bind Only allocate memory on a specific set of nodes,
24 * FIXME: memory is allocated starting with the first node
25 * to the last. It would be better if bind would truly restrict
26 * the allocation to memory nodes instead
28 * preferred Try a specific node first before normal fallback.
29 * As a special case node -1 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.
66 could replace all the switch()es with a mempolicy_ops structure.
69 #include <linux/mempolicy.h>
71 #include <linux/highmem.h>
72 #include <linux/hugetlb.h>
73 #include <linux/kernel.h>
74 #include <linux/sched.h>
76 #include <linux/nodemask.h>
77 #include <linux/cpuset.h>
78 #include <linux/gfp.h>
79 #include <linux/slab.h>
80 #include <linux/string.h>
81 #include <linux/module.h>
82 #include <linux/interrupt.h>
83 #include <linux/init.h>
84 #include <linux/compat.h>
85 #include <linux/mempolicy.h>
86 #include <asm/tlbflush.h>
87 #include <asm/uaccess.h>
89 static kmem_cache_t *policy_cache;
90 static kmem_cache_t *sn_cache;
92 #define PDprintk(fmt...)
94 /* Highest zone. An specific allocation for a zone below that is not
96 static int policy_zone;
98 struct mempolicy default_policy = {
99 .refcnt = ATOMIC_INIT(1), /* never free it */
100 .policy = MPOL_DEFAULT,
103 /* Do sanity checking on a policy */
104 static int mpol_check_policy(int mode, nodemask_t *nodes)
106 int empty = nodes_empty(*nodes);
114 case MPOL_INTERLEAVE:
115 /* Preferred will only use the first bit, but allow
121 return nodes_subset(*nodes, node_online_map) ? 0 : -EINVAL;
123 /* Generate a custom zonelist for the BIND policy. */
124 static struct zonelist *bind_zonelist(nodemask_t *nodes)
129 max = 1 + MAX_NR_ZONES * nodes_weight(*nodes);
130 zl = kmalloc(sizeof(void *) * max, GFP_KERNEL);
134 for_each_node_mask(nd, *nodes) {
136 for (k = MAX_NR_ZONES-1; k >= 0; k--) {
137 struct zone *z = &NODE_DATA(nd)->node_zones[k];
138 if (!z->present_pages)
140 zl->zones[num++] = z;
145 zl->zones[num] = NULL;
149 /* Create a new policy */
150 static struct mempolicy *mpol_new(int mode, nodemask_t *nodes)
152 struct mempolicy *policy;
154 PDprintk("setting mode %d nodes[0] %lx\n", mode, nodes_addr(*nodes)[0]);
155 if (mode == MPOL_DEFAULT)
157 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
159 return ERR_PTR(-ENOMEM);
160 atomic_set(&policy->refcnt, 1);
162 case MPOL_INTERLEAVE:
163 policy->v.nodes = *nodes;
164 if (nodes_weight(*nodes) == 0) {
165 kmem_cache_free(policy_cache, policy);
166 return ERR_PTR(-EINVAL);
170 policy->v.preferred_node = first_node(*nodes);
171 if (policy->v.preferred_node >= MAX_NUMNODES)
172 policy->v.preferred_node = -1;
175 policy->v.zonelist = bind_zonelist(nodes);
176 if (policy->v.zonelist == NULL) {
177 kmem_cache_free(policy_cache, policy);
178 return ERR_PTR(-ENOMEM);
182 policy->policy = mode;
186 /* Ensure all existing pages follow the policy. */
187 static int check_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
188 unsigned long addr, unsigned long end, nodemask_t *nodes)
194 orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
199 if (!pte_present(*pte))
201 page = vm_normal_page(vma, addr, *pte);
204 nid = page_to_nid(page);
205 if (!node_isset(nid, *nodes))
207 } while (pte++, addr += PAGE_SIZE, addr != end);
208 pte_unmap_unlock(orig_pte, ptl);
212 static inline int check_pmd_range(struct vm_area_struct *vma, pud_t *pud,
213 unsigned long addr, unsigned long end, nodemask_t *nodes)
218 pmd = pmd_offset(pud, addr);
220 next = pmd_addr_end(addr, end);
221 if (pmd_none_or_clear_bad(pmd))
223 if (check_pte_range(vma, pmd, addr, next, nodes))
225 } while (pmd++, addr = next, addr != end);
229 static inline int check_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
230 unsigned long addr, unsigned long end, nodemask_t *nodes)
235 pud = pud_offset(pgd, addr);
237 next = pud_addr_end(addr, end);
238 if (pud_none_or_clear_bad(pud))
240 if (check_pmd_range(vma, pud, addr, next, nodes))
242 } while (pud++, addr = next, addr != end);
246 static inline int check_pgd_range(struct vm_area_struct *vma,
247 unsigned long addr, unsigned long end, nodemask_t *nodes)
252 pgd = pgd_offset(vma->vm_mm, addr);
254 next = pgd_addr_end(addr, end);
255 if (pgd_none_or_clear_bad(pgd))
257 if (check_pud_range(vma, pgd, addr, next, nodes))
259 } while (pgd++, addr = next, addr != end);
263 /* Step 1: check the range */
264 static struct vm_area_struct *
265 check_range(struct mm_struct *mm, unsigned long start, unsigned long end,
266 nodemask_t *nodes, unsigned long flags)
269 struct vm_area_struct *first, *vma, *prev;
271 first = find_vma(mm, start);
273 return ERR_PTR(-EFAULT);
275 for (vma = first; vma && vma->vm_start < end; vma = vma->vm_next) {
276 if (!vma->vm_next && vma->vm_end < end)
277 return ERR_PTR(-EFAULT);
278 if (prev && prev->vm_end < vma->vm_start)
279 return ERR_PTR(-EFAULT);
280 if ((flags & MPOL_MF_STRICT) && !is_vm_hugetlb_page(vma)) {
281 unsigned long endvma = vma->vm_end;
284 if (vma->vm_start > start)
285 start = vma->vm_start;
286 err = check_pgd_range(vma, start, endvma, nodes);
288 first = ERR_PTR(err);
297 /* Apply policy to a single VMA */
298 static int policy_vma(struct vm_area_struct *vma, struct mempolicy *new)
301 struct mempolicy *old = vma->vm_policy;
303 PDprintk("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
304 vma->vm_start, vma->vm_end, vma->vm_pgoff,
305 vma->vm_ops, vma->vm_file,
306 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
308 if (vma->vm_ops && vma->vm_ops->set_policy)
309 err = vma->vm_ops->set_policy(vma, new);
312 vma->vm_policy = new;
318 /* Step 2: apply policy to a range and do splits. */
319 static int mbind_range(struct vm_area_struct *vma, unsigned long start,
320 unsigned long end, struct mempolicy *new)
322 struct vm_area_struct *next;
326 for (; vma && vma->vm_start < end; vma = next) {
328 if (vma->vm_start < start)
329 err = split_vma(vma->vm_mm, vma, start, 1);
330 if (!err && vma->vm_end > end)
331 err = split_vma(vma->vm_mm, vma, end, 0);
333 err = policy_vma(vma, new);
340 static int contextualize_policy(int mode, nodemask_t *nodes)
345 /* Update current mems_allowed */
346 cpuset_update_current_mems_allowed();
347 /* Ignore nodes not set in current->mems_allowed */
348 cpuset_restrict_to_mems_allowed(nodes->bits);
349 return mpol_check_policy(mode, nodes);
352 long do_mbind(unsigned long start, unsigned long len,
353 unsigned long mode, nodemask_t *nmask, unsigned long flags)
355 struct vm_area_struct *vma;
356 struct mm_struct *mm = current->mm;
357 struct mempolicy *new;
361 if ((flags & ~(unsigned long)(MPOL_MF_STRICT)) || mode > MPOL_MAX)
363 if (start & ~PAGE_MASK)
365 if (mode == MPOL_DEFAULT)
366 flags &= ~MPOL_MF_STRICT;
367 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
373 if (mpol_check_policy(mode, nmask))
375 new = mpol_new(mode, nmask);
379 PDprintk("mbind %lx-%lx mode:%ld nodes:%lx\n",start,start+len,
380 mode,nodes_addr(nodes)[0]);
382 down_write(&mm->mmap_sem);
383 vma = check_range(mm, start, end, nmask, flags);
386 err = mbind_range(vma, start, end, new);
387 up_write(&mm->mmap_sem);
392 /* Set the process memory policy */
393 long do_set_mempolicy(int mode, nodemask_t *nodes)
395 struct mempolicy *new;
397 if (contextualize_policy(mode, nodes))
399 new = mpol_new(mode, nodes);
402 mpol_free(current->mempolicy);
403 current->mempolicy = new;
404 if (new && new->policy == MPOL_INTERLEAVE)
405 current->il_next = first_node(new->v.nodes);
409 /* Fill a zone bitmap for a policy */
410 static void get_zonemask(struct mempolicy *p, nodemask_t *nodes)
417 for (i = 0; p->v.zonelist->zones[i]; i++)
418 node_set(p->v.zonelist->zones[i]->zone_pgdat->node_id,
423 case MPOL_INTERLEAVE:
427 /* or use current node instead of online map? */
428 if (p->v.preferred_node < 0)
429 *nodes = node_online_map;
431 node_set(p->v.preferred_node, *nodes);
438 static int lookup_node(struct mm_struct *mm, unsigned long addr)
443 err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL);
445 err = page_to_nid(p);
451 /* Retrieve NUMA policy */
452 long do_get_mempolicy(int *policy, nodemask_t *nmask,
453 unsigned long addr, unsigned long flags)
456 struct mm_struct *mm = current->mm;
457 struct vm_area_struct *vma = NULL;
458 struct mempolicy *pol = current->mempolicy;
460 cpuset_update_current_mems_allowed();
461 if (flags & ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR))
463 if (flags & MPOL_F_ADDR) {
464 down_read(&mm->mmap_sem);
465 vma = find_vma_intersection(mm, addr, addr+1);
467 up_read(&mm->mmap_sem);
470 if (vma->vm_ops && vma->vm_ops->get_policy)
471 pol = vma->vm_ops->get_policy(vma, addr);
473 pol = vma->vm_policy;
478 pol = &default_policy;
480 if (flags & MPOL_F_NODE) {
481 if (flags & MPOL_F_ADDR) {
482 err = lookup_node(mm, addr);
486 } else if (pol == current->mempolicy &&
487 pol->policy == MPOL_INTERLEAVE) {
488 *policy = current->il_next;
494 *policy = pol->policy;
497 up_read(¤t->mm->mmap_sem);
503 get_zonemask(pol, nmask);
507 up_read(¤t->mm->mmap_sem);
512 * User space interface with variable sized bitmaps for nodelists.
515 /* Copy a node mask from user space. */
516 static int get_nodes(nodemask_t *nodes, unsigned long __user *nmask,
517 unsigned long maxnode)
520 unsigned long nlongs;
521 unsigned long endmask;
525 if (maxnode == 0 || !nmask)
528 nlongs = BITS_TO_LONGS(maxnode);
529 if ((maxnode % BITS_PER_LONG) == 0)
532 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
534 /* When the user specified more nodes than supported just check
535 if the non supported part is all zero. */
536 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
537 if (nlongs > PAGE_SIZE/sizeof(long))
539 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
541 if (get_user(t, nmask + k))
543 if (k == nlongs - 1) {
549 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
553 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
555 nodes_addr(*nodes)[nlongs-1] &= endmask;
559 /* Copy a kernel node mask to user space */
560 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
563 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
564 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
567 if (copy > PAGE_SIZE)
569 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
573 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
576 asmlinkage long sys_mbind(unsigned long start, unsigned long len,
578 unsigned long __user *nmask, unsigned long maxnode,
584 err = get_nodes(&nodes, nmask, maxnode);
587 return do_mbind(start, len, mode, &nodes, flags);
590 /* Set the process memory policy */
591 asmlinkage long sys_set_mempolicy(int mode, unsigned long __user *nmask,
592 unsigned long maxnode)
597 if (mode < 0 || mode > MPOL_MAX)
599 err = get_nodes(&nodes, nmask, maxnode);
602 return do_set_mempolicy(mode, &nodes);
605 /* Retrieve NUMA policy */
606 asmlinkage long sys_get_mempolicy(int __user *policy,
607 unsigned long __user *nmask,
608 unsigned long maxnode,
609 unsigned long addr, unsigned long flags)
614 if (nmask != NULL && maxnode < MAX_NUMNODES)
617 err = do_get_mempolicy(&pval, &nodes, addr, flags);
622 if (policy && put_user(pval, policy))
626 err = copy_nodes_to_user(nmask, maxnode, &nodes);
633 asmlinkage long compat_sys_get_mempolicy(int __user *policy,
634 compat_ulong_t __user *nmask,
635 compat_ulong_t maxnode,
636 compat_ulong_t addr, compat_ulong_t flags)
639 unsigned long __user *nm = NULL;
640 unsigned long nr_bits, alloc_size;
641 DECLARE_BITMAP(bm, MAX_NUMNODES);
643 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
644 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
647 nm = compat_alloc_user_space(alloc_size);
649 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
652 err = copy_from_user(bm, nm, alloc_size);
653 /* ensure entire bitmap is zeroed */
654 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
655 err |= compat_put_bitmap(nmask, bm, nr_bits);
661 asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask,
662 compat_ulong_t maxnode)
665 unsigned long __user *nm = NULL;
666 unsigned long nr_bits, alloc_size;
667 DECLARE_BITMAP(bm, MAX_NUMNODES);
669 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
670 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
673 err = compat_get_bitmap(bm, nmask, nr_bits);
674 nm = compat_alloc_user_space(alloc_size);
675 err |= copy_to_user(nm, bm, alloc_size);
681 return sys_set_mempolicy(mode, nm, nr_bits+1);
684 asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len,
685 compat_ulong_t mode, compat_ulong_t __user *nmask,
686 compat_ulong_t maxnode, compat_ulong_t flags)
689 unsigned long __user *nm = NULL;
690 unsigned long nr_bits, alloc_size;
693 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
694 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
697 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
698 nm = compat_alloc_user_space(alloc_size);
699 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
705 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
710 /* Return effective policy for a VMA */
712 get_vma_policy(struct task_struct *task, struct vm_area_struct *vma, unsigned long addr)
714 struct mempolicy *pol = task->mempolicy;
717 if (vma->vm_ops && vma->vm_ops->get_policy)
718 pol = vma->vm_ops->get_policy(vma, addr);
719 else if (vma->vm_policy &&
720 vma->vm_policy->policy != MPOL_DEFAULT)
721 pol = vma->vm_policy;
724 pol = &default_policy;
728 /* Return a zonelist representing a mempolicy */
729 static struct zonelist *zonelist_policy(gfp_t gfp, struct mempolicy *policy)
733 switch (policy->policy) {
735 nd = policy->v.preferred_node;
740 /* Lower zones don't get a policy applied */
741 /* Careful: current->mems_allowed might have moved */
742 if (gfp_zone(gfp) >= policy_zone)
743 if (cpuset_zonelist_valid_mems_allowed(policy->v.zonelist))
744 return policy->v.zonelist;
746 case MPOL_INTERLEAVE: /* should not happen */
754 return NODE_DATA(nd)->node_zonelists + gfp_zone(gfp);
757 /* Do dynamic interleaving for a process */
758 static unsigned interleave_nodes(struct mempolicy *policy)
761 struct task_struct *me = current;
764 next = next_node(nid, policy->v.nodes);
765 if (next >= MAX_NUMNODES)
766 next = first_node(policy->v.nodes);
771 /* Do static interleaving for a VMA with known offset. */
772 static unsigned offset_il_node(struct mempolicy *pol,
773 struct vm_area_struct *vma, unsigned long off)
775 unsigned nnodes = nodes_weight(pol->v.nodes);
776 unsigned target = (unsigned)off % nnodes;
782 nid = next_node(nid, pol->v.nodes);
784 } while (c <= target);
788 /* Allocate a page in interleaved policy.
789 Own path because it needs to do special accounting. */
790 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
796 zl = NODE_DATA(nid)->node_zonelists + gfp_zone(gfp);
797 page = __alloc_pages(gfp, order, zl);
798 if (page && page_zone(page) == zl->zones[0]) {
799 zone_pcp(zl->zones[0],get_cpu())->interleave_hit++;
806 * alloc_page_vma - Allocate a page for a VMA.
809 * %GFP_USER user allocation.
810 * %GFP_KERNEL kernel allocations,
811 * %GFP_HIGHMEM highmem/user allocations,
812 * %GFP_FS allocation should not call back into a file system.
813 * %GFP_ATOMIC don't sleep.
815 * @vma: Pointer to VMA or NULL if not available.
816 * @addr: Virtual Address of the allocation. Must be inside the VMA.
818 * This function allocates a page from the kernel page pool and applies
819 * a NUMA policy associated with the VMA or the current process.
820 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
821 * mm_struct of the VMA to prevent it from going away. Should be used for
822 * all allocations for pages that will be mapped into
823 * user space. Returns NULL when no page can be allocated.
825 * Should be called with the mm_sem of the vma hold.
828 alloc_page_vma(gfp_t gfp, struct vm_area_struct *vma, unsigned long addr)
830 struct mempolicy *pol = get_vma_policy(current, vma, addr);
832 cpuset_update_current_mems_allowed();
834 if (unlikely(pol->policy == MPOL_INTERLEAVE)) {
839 off += (addr - vma->vm_start) >> PAGE_SHIFT;
840 nid = offset_il_node(pol, vma, off);
842 /* fall back to process interleaving */
843 nid = interleave_nodes(pol);
845 return alloc_page_interleave(gfp, 0, nid);
847 return __alloc_pages(gfp, 0, zonelist_policy(gfp, pol));
851 * alloc_pages_current - Allocate pages.
854 * %GFP_USER user allocation,
855 * %GFP_KERNEL kernel allocation,
856 * %GFP_HIGHMEM highmem allocation,
857 * %GFP_FS don't call back into a file system.
858 * %GFP_ATOMIC don't sleep.
859 * @order: Power of two of allocation size in pages. 0 is a single page.
861 * Allocate a page from the kernel page pool. When not in
862 * interrupt context and apply the current process NUMA policy.
863 * Returns NULL when no page can be allocated.
865 * Don't call cpuset_update_current_mems_allowed() unless
866 * 1) it's ok to take cpuset_sem (can WAIT), and
867 * 2) allocating for current task (not interrupt).
869 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
871 struct mempolicy *pol = current->mempolicy;
873 if ((gfp & __GFP_WAIT) && !in_interrupt())
874 cpuset_update_current_mems_allowed();
875 if (!pol || in_interrupt())
876 pol = &default_policy;
877 if (pol->policy == MPOL_INTERLEAVE)
878 return alloc_page_interleave(gfp, order, interleave_nodes(pol));
879 return __alloc_pages(gfp, order, zonelist_policy(gfp, pol));
881 EXPORT_SYMBOL(alloc_pages_current);
883 /* Slow path of a mempolicy copy */
884 struct mempolicy *__mpol_copy(struct mempolicy *old)
886 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
889 return ERR_PTR(-ENOMEM);
891 atomic_set(&new->refcnt, 1);
892 if (new->policy == MPOL_BIND) {
893 int sz = ksize(old->v.zonelist);
894 new->v.zonelist = kmalloc(sz, SLAB_KERNEL);
895 if (!new->v.zonelist) {
896 kmem_cache_free(policy_cache, new);
897 return ERR_PTR(-ENOMEM);
899 memcpy(new->v.zonelist, old->v.zonelist, sz);
904 /* Slow path of a mempolicy comparison */
905 int __mpol_equal(struct mempolicy *a, struct mempolicy *b)
909 if (a->policy != b->policy)
914 case MPOL_INTERLEAVE:
915 return nodes_equal(a->v.nodes, b->v.nodes);
917 return a->v.preferred_node == b->v.preferred_node;
920 for (i = 0; a->v.zonelist->zones[i]; i++)
921 if (a->v.zonelist->zones[i] != b->v.zonelist->zones[i])
923 return b->v.zonelist->zones[i] == NULL;
931 /* Slow path of a mpol destructor. */
932 void __mpol_free(struct mempolicy *p)
934 if (!atomic_dec_and_test(&p->refcnt))
936 if (p->policy == MPOL_BIND)
937 kfree(p->v.zonelist);
938 p->policy = MPOL_DEFAULT;
939 kmem_cache_free(policy_cache, p);
943 * Hugetlb policy. Same as above, just works with node numbers instead of
947 /* Find first node suitable for an allocation */
948 int mpol_first_node(struct vm_area_struct *vma, unsigned long addr)
950 struct mempolicy *pol = get_vma_policy(current, vma, addr);
952 switch (pol->policy) {
954 return numa_node_id();
956 return pol->v.zonelist->zones[0]->zone_pgdat->node_id;
957 case MPOL_INTERLEAVE:
958 return interleave_nodes(pol);
960 return pol->v.preferred_node >= 0 ?
961 pol->v.preferred_node : numa_node_id();
967 /* Find secondary valid nodes for an allocation */
968 int mpol_node_valid(int nid, struct vm_area_struct *vma, unsigned long addr)
970 struct mempolicy *pol = get_vma_policy(current, vma, addr);
972 switch (pol->policy) {
975 case MPOL_INTERLEAVE:
979 for (z = pol->v.zonelist->zones; *z; z++)
980 if ((*z)->zone_pgdat->node_id == nid)
991 * Shared memory backing store policy support.
993 * Remember policies even when nobody has shared memory mapped.
994 * The policies are kept in Red-Black tree linked from the inode.
995 * They are protected by the sp->lock spinlock, which should be held
996 * for any accesses to the tree.
999 /* lookup first element intersecting start-end */
1000 /* Caller holds sp->lock */
1001 static struct sp_node *
1002 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
1004 struct rb_node *n = sp->root.rb_node;
1007 struct sp_node *p = rb_entry(n, struct sp_node, nd);
1009 if (start >= p->end)
1011 else if (end <= p->start)
1019 struct sp_node *w = NULL;
1020 struct rb_node *prev = rb_prev(n);
1023 w = rb_entry(prev, struct sp_node, nd);
1024 if (w->end <= start)
1028 return rb_entry(n, struct sp_node, nd);
1031 /* Insert a new shared policy into the list. */
1032 /* Caller holds sp->lock */
1033 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
1035 struct rb_node **p = &sp->root.rb_node;
1036 struct rb_node *parent = NULL;
1041 nd = rb_entry(parent, struct sp_node, nd);
1042 if (new->start < nd->start)
1044 else if (new->end > nd->end)
1045 p = &(*p)->rb_right;
1049 rb_link_node(&new->nd, parent, p);
1050 rb_insert_color(&new->nd, &sp->root);
1051 PDprintk("inserting %lx-%lx: %d\n", new->start, new->end,
1052 new->policy ? new->policy->policy : 0);
1055 /* Find shared policy intersecting idx */
1057 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
1059 struct mempolicy *pol = NULL;
1062 if (!sp->root.rb_node)
1064 spin_lock(&sp->lock);
1065 sn = sp_lookup(sp, idx, idx+1);
1067 mpol_get(sn->policy);
1070 spin_unlock(&sp->lock);
1074 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
1076 PDprintk("deleting %lx-l%x\n", n->start, n->end);
1077 rb_erase(&n->nd, &sp->root);
1078 mpol_free(n->policy);
1079 kmem_cache_free(sn_cache, n);
1083 sp_alloc(unsigned long start, unsigned long end, struct mempolicy *pol)
1085 struct sp_node *n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
1096 /* Replace a policy range. */
1097 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
1098 unsigned long end, struct sp_node *new)
1100 struct sp_node *n, *new2 = NULL;
1103 spin_lock(&sp->lock);
1104 n = sp_lookup(sp, start, end);
1105 /* Take care of old policies in the same range. */
1106 while (n && n->start < end) {
1107 struct rb_node *next = rb_next(&n->nd);
1108 if (n->start >= start) {
1114 /* Old policy spanning whole new range. */
1117 spin_unlock(&sp->lock);
1118 new2 = sp_alloc(end, n->end, n->policy);
1124 sp_insert(sp, new2);
1132 n = rb_entry(next, struct sp_node, nd);
1136 spin_unlock(&sp->lock);
1138 mpol_free(new2->policy);
1139 kmem_cache_free(sn_cache, new2);
1144 int mpol_set_shared_policy(struct shared_policy *info,
1145 struct vm_area_struct *vma, struct mempolicy *npol)
1148 struct sp_node *new = NULL;
1149 unsigned long sz = vma_pages(vma);
1151 PDprintk("set_shared_policy %lx sz %lu %d %lx\n",
1153 sz, npol? npol->policy : -1,
1154 npol ? nodes_addr(npol->v.nodes)[0] : -1);
1157 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
1161 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
1163 kmem_cache_free(sn_cache, new);
1167 /* Free a backing policy store on inode delete. */
1168 void mpol_free_shared_policy(struct shared_policy *p)
1171 struct rb_node *next;
1173 if (!p->root.rb_node)
1175 spin_lock(&p->lock);
1176 next = rb_first(&p->root);
1178 n = rb_entry(next, struct sp_node, nd);
1179 next = rb_next(&n->nd);
1180 rb_erase(&n->nd, &p->root);
1181 mpol_free(n->policy);
1182 kmem_cache_free(sn_cache, n);
1184 spin_unlock(&p->lock);
1187 /* assumes fs == KERNEL_DS */
1188 void __init numa_policy_init(void)
1190 policy_cache = kmem_cache_create("numa_policy",
1191 sizeof(struct mempolicy),
1192 0, SLAB_PANIC, NULL, NULL);
1194 sn_cache = kmem_cache_create("shared_policy_node",
1195 sizeof(struct sp_node),
1196 0, SLAB_PANIC, NULL, NULL);
1198 /* Set interleaving policy for system init. This way not all
1199 the data structures allocated at system boot end up in node zero. */
1201 if (do_set_mempolicy(MPOL_INTERLEAVE, &node_online_map))
1202 printk("numa_policy_init: interleaving failed\n");
1205 /* Reset policy of current process to default */
1206 void numa_default_policy(void)
1208 do_set_mempolicy(MPOL_DEFAULT, NULL);
1211 /* Migrate a policy to a different set of nodes */
1212 static void rebind_policy(struct mempolicy *pol, const nodemask_t *old,
1213 const nodemask_t *new)
1220 switch (pol->policy) {
1223 case MPOL_INTERLEAVE:
1224 nodes_remap(tmp, pol->v.nodes, *old, *new);
1226 current->il_next = node_remap(current->il_next, *old, *new);
1228 case MPOL_PREFERRED:
1229 pol->v.preferred_node = node_remap(pol->v.preferred_node,
1235 struct zonelist *zonelist;
1238 for (z = pol->v.zonelist->zones; *z; z++)
1239 node_set((*z)->zone_pgdat->node_id, nodes);
1240 nodes_remap(tmp, nodes, *old, *new);
1243 zonelist = bind_zonelist(&nodes);
1245 /* If no mem, then zonelist is NULL and we keep old zonelist.
1246 * If that old zonelist has no remaining mems_allowed nodes,
1247 * then zonelist_policy() will "FALL THROUGH" to MPOL_DEFAULT.
1251 /* Good - got mem - substitute new zonelist */
1252 kfree(pol->v.zonelist);
1253 pol->v.zonelist = zonelist;
1264 * Someone moved this task to different nodes. Fixup mempolicies.
1266 * TODO - fixup current->mm->vma and shmfs/tmpfs/hugetlbfs policies as well,
1267 * once we have a cpuset mechanism to mark which cpuset subtree is migrating.
1269 void numa_policy_rebind(const nodemask_t *old, const nodemask_t *new)
1271 rebind_policy(current->mempolicy, old, new);