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 <linux/swap.h>
88 #include <asm/tlbflush.h>
89 #include <asm/uaccess.h>
91 /* Internal MPOL_MF_xxx flags */
92 #define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
94 static kmem_cache_t *policy_cache;
95 static kmem_cache_t *sn_cache;
97 #define PDprintk(fmt...)
99 /* Highest zone. An specific allocation for a zone below that is not
101 int policy_zone = ZONE_DMA;
103 struct mempolicy default_policy = {
104 .refcnt = ATOMIC_INIT(1), /* never free it */
105 .policy = MPOL_DEFAULT,
108 /* Do sanity checking on a policy */
109 static int mpol_check_policy(int mode, nodemask_t *nodes)
111 int empty = nodes_empty(*nodes);
119 case MPOL_INTERLEAVE:
120 /* Preferred will only use the first bit, but allow
126 return nodes_subset(*nodes, node_online_map) ? 0 : -EINVAL;
128 /* Generate a custom zonelist for the BIND policy. */
129 static struct zonelist *bind_zonelist(nodemask_t *nodes)
134 max = 1 + MAX_NR_ZONES * nodes_weight(*nodes);
135 zl = kmalloc(sizeof(void *) * max, GFP_KERNEL);
139 for_each_node_mask(nd, *nodes)
140 zl->zones[num++] = &NODE_DATA(nd)->node_zones[policy_zone];
141 zl->zones[num] = NULL;
145 /* Create a new policy */
146 static struct mempolicy *mpol_new(int mode, nodemask_t *nodes)
148 struct mempolicy *policy;
150 PDprintk("setting mode %d nodes[0] %lx\n", mode, nodes_addr(*nodes)[0]);
151 if (mode == MPOL_DEFAULT)
153 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
155 return ERR_PTR(-ENOMEM);
156 atomic_set(&policy->refcnt, 1);
158 case MPOL_INTERLEAVE:
159 policy->v.nodes = *nodes;
160 if (nodes_weight(*nodes) == 0) {
161 kmem_cache_free(policy_cache, policy);
162 return ERR_PTR(-EINVAL);
166 policy->v.preferred_node = first_node(*nodes);
167 if (policy->v.preferred_node >= MAX_NUMNODES)
168 policy->v.preferred_node = -1;
171 policy->v.zonelist = bind_zonelist(nodes);
172 if (policy->v.zonelist == NULL) {
173 kmem_cache_free(policy_cache, policy);
174 return ERR_PTR(-ENOMEM);
178 policy->policy = mode;
182 /* Check if we are the only process mapping the page in question */
183 static inline int single_mm_mapping(struct mm_struct *mm,
184 struct address_space *mapping)
186 struct vm_area_struct *vma;
187 struct prio_tree_iter iter;
190 spin_lock(&mapping->i_mmap_lock);
191 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
192 if (mm != vma->vm_mm) {
196 list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
197 if (mm != vma->vm_mm) {
202 spin_unlock(&mapping->i_mmap_lock);
207 * Add a page to be migrated to the pagelist
209 static void migrate_page_add(struct vm_area_struct *vma,
210 struct page *page, struct list_head *pagelist, unsigned long flags)
213 * Avoid migrating a page that is shared by others and not writable.
215 if ((flags & MPOL_MF_MOVE_ALL) || !page->mapping || PageAnon(page) ||
216 mapping_writably_mapped(page->mapping) ||
217 single_mm_mapping(vma->vm_mm, page->mapping)) {
218 int rc = isolate_lru_page(page);
221 list_add(&page->lru, pagelist);
223 * If the isolate attempt was not successful then we just
224 * encountered an unswappable page. Something must be wrong.
230 /* Ensure all existing pages follow the policy. */
231 static int check_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
232 unsigned long addr, unsigned long end,
233 const nodemask_t *nodes, unsigned long flags,
234 struct list_head *pagelist)
240 orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
245 if (!pte_present(*pte))
247 page = vm_normal_page(vma, addr, *pte);
250 nid = page_to_nid(page);
251 if (!node_isset(nid, *nodes)) {
253 migrate_page_add(vma, page, pagelist, flags);
257 } while (pte++, addr += PAGE_SIZE, addr != end);
258 pte_unmap_unlock(orig_pte, ptl);
262 static inline int check_pmd_range(struct vm_area_struct *vma, pud_t *pud,
263 unsigned long addr, unsigned long end,
264 const nodemask_t *nodes, unsigned long flags,
265 struct list_head *pagelist)
270 pmd = pmd_offset(pud, addr);
272 next = pmd_addr_end(addr, end);
273 if (pmd_none_or_clear_bad(pmd))
275 if (check_pte_range(vma, pmd, addr, next, nodes,
278 } while (pmd++, addr = next, addr != end);
282 static inline int check_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
283 unsigned long addr, unsigned long end,
284 const nodemask_t *nodes, unsigned long flags,
285 struct list_head *pagelist)
290 pud = pud_offset(pgd, addr);
292 next = pud_addr_end(addr, end);
293 if (pud_none_or_clear_bad(pud))
295 if (check_pmd_range(vma, pud, addr, next, nodes,
298 } while (pud++, addr = next, addr != end);
302 static inline int check_pgd_range(struct vm_area_struct *vma,
303 unsigned long addr, unsigned long end,
304 const nodemask_t *nodes, unsigned long flags,
305 struct list_head *pagelist)
310 pgd = pgd_offset(vma->vm_mm, addr);
312 next = pgd_addr_end(addr, end);
313 if (pgd_none_or_clear_bad(pgd))
315 if (check_pud_range(vma, pgd, addr, next, nodes,
318 } while (pgd++, addr = next, addr != end);
322 /* Check if a vma is migratable */
323 static inline int vma_migratable(struct vm_area_struct *vma)
325 if (vma->vm_flags & (
326 VM_LOCKED|VM_IO|VM_HUGETLB|VM_PFNMAP))
332 * Check if all pages in a range are on a set of nodes.
333 * If pagelist != NULL then isolate pages from the LRU and
334 * put them on the pagelist.
336 static struct vm_area_struct *
337 check_range(struct mm_struct *mm, unsigned long start, unsigned long end,
338 const nodemask_t *nodes, unsigned long flags,
339 struct list_head *pagelist)
342 struct vm_area_struct *first, *vma, *prev;
344 first = find_vma(mm, start);
346 return ERR_PTR(-EFAULT);
348 for (vma = first; vma && vma->vm_start < end; vma = vma->vm_next) {
349 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
350 if (!vma->vm_next && vma->vm_end < end)
351 return ERR_PTR(-EFAULT);
352 if (prev && prev->vm_end < vma->vm_start)
353 return ERR_PTR(-EFAULT);
355 if (!is_vm_hugetlb_page(vma) &&
356 ((flags & MPOL_MF_STRICT) ||
357 ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) &&
358 vma_migratable(vma)))) {
359 unsigned long endvma = vma->vm_end;
363 if (vma->vm_start > start)
364 start = vma->vm_start;
365 err = check_pgd_range(vma, start, endvma, nodes,
368 first = ERR_PTR(err);
377 /* Apply policy to a single VMA */
378 static int policy_vma(struct vm_area_struct *vma, struct mempolicy *new)
381 struct mempolicy *old = vma->vm_policy;
383 PDprintk("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
384 vma->vm_start, vma->vm_end, vma->vm_pgoff,
385 vma->vm_ops, vma->vm_file,
386 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
388 if (vma->vm_ops && vma->vm_ops->set_policy)
389 err = vma->vm_ops->set_policy(vma, new);
392 vma->vm_policy = new;
398 /* Step 2: apply policy to a range and do splits. */
399 static int mbind_range(struct vm_area_struct *vma, unsigned long start,
400 unsigned long end, struct mempolicy *new)
402 struct vm_area_struct *next;
406 for (; vma && vma->vm_start < end; vma = next) {
408 if (vma->vm_start < start)
409 err = split_vma(vma->vm_mm, vma, start, 1);
410 if (!err && vma->vm_end > end)
411 err = split_vma(vma->vm_mm, vma, end, 0);
413 err = policy_vma(vma, new);
420 static int contextualize_policy(int mode, nodemask_t *nodes)
425 /* Update current mems_allowed */
426 cpuset_update_current_mems_allowed();
427 /* Ignore nodes not set in current->mems_allowed */
428 cpuset_restrict_to_mems_allowed(nodes->bits);
429 return mpol_check_policy(mode, nodes);
432 long do_mbind(unsigned long start, unsigned long len,
433 unsigned long mode, nodemask_t *nmask, unsigned long flags)
435 struct vm_area_struct *vma;
436 struct mm_struct *mm = current->mm;
437 struct mempolicy *new;
442 if ((flags & ~(unsigned long)(MPOL_MF_STRICT|MPOL_MF_MOVE|MPOL_MF_MOVE_ALL))
445 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_RESOURCE))
448 if (start & ~PAGE_MASK)
451 if (mode == MPOL_DEFAULT)
452 flags &= ~MPOL_MF_STRICT;
454 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
462 if (mpol_check_policy(mode, nmask))
465 new = mpol_new(mode, nmask);
470 * If we are using the default policy then operation
471 * on discontinuous address spaces is okay after all
474 flags |= MPOL_MF_DISCONTIG_OK;
476 PDprintk("mbind %lx-%lx mode:%ld nodes:%lx\n",start,start+len,
477 mode,nodes_addr(nodes)[0]);
479 down_write(&mm->mmap_sem);
480 vma = check_range(mm, start, end, nmask, flags,
481 (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) ? &pagelist : NULL);
484 err = mbind_range(vma, start, end, new);
485 if (!list_empty(&pagelist))
486 migrate_pages(&pagelist, NULL);
487 if (!err && !list_empty(&pagelist) && (flags & MPOL_MF_STRICT))
490 if (!list_empty(&pagelist))
491 putback_lru_pages(&pagelist);
493 up_write(&mm->mmap_sem);
498 /* Set the process memory policy */
499 long do_set_mempolicy(int mode, nodemask_t *nodes)
501 struct mempolicy *new;
503 if (contextualize_policy(mode, nodes))
505 new = mpol_new(mode, nodes);
508 mpol_free(current->mempolicy);
509 current->mempolicy = new;
510 if (new && new->policy == MPOL_INTERLEAVE)
511 current->il_next = first_node(new->v.nodes);
515 /* Fill a zone bitmap for a policy */
516 static void get_zonemask(struct mempolicy *p, nodemask_t *nodes)
523 for (i = 0; p->v.zonelist->zones[i]; i++)
524 node_set(p->v.zonelist->zones[i]->zone_pgdat->node_id,
529 case MPOL_INTERLEAVE:
533 /* or use current node instead of online map? */
534 if (p->v.preferred_node < 0)
535 *nodes = node_online_map;
537 node_set(p->v.preferred_node, *nodes);
544 static int lookup_node(struct mm_struct *mm, unsigned long addr)
549 err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL);
551 err = page_to_nid(p);
557 /* Retrieve NUMA policy */
558 long do_get_mempolicy(int *policy, nodemask_t *nmask,
559 unsigned long addr, unsigned long flags)
562 struct mm_struct *mm = current->mm;
563 struct vm_area_struct *vma = NULL;
564 struct mempolicy *pol = current->mempolicy;
566 cpuset_update_current_mems_allowed();
567 if (flags & ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR))
569 if (flags & MPOL_F_ADDR) {
570 down_read(&mm->mmap_sem);
571 vma = find_vma_intersection(mm, addr, addr+1);
573 up_read(&mm->mmap_sem);
576 if (vma->vm_ops && vma->vm_ops->get_policy)
577 pol = vma->vm_ops->get_policy(vma, addr);
579 pol = vma->vm_policy;
584 pol = &default_policy;
586 if (flags & MPOL_F_NODE) {
587 if (flags & MPOL_F_ADDR) {
588 err = lookup_node(mm, addr);
592 } else if (pol == current->mempolicy &&
593 pol->policy == MPOL_INTERLEAVE) {
594 *policy = current->il_next;
600 *policy = pol->policy;
603 up_read(¤t->mm->mmap_sem);
609 get_zonemask(pol, nmask);
613 up_read(¤t->mm->mmap_sem);
618 * For now migrate_pages simply swaps out the pages from nodes that are in
619 * the source set but not in the target set. In the future, we would
620 * want a function that moves pages between the two nodesets in such
621 * a way as to preserve the physical layout as much as possible.
623 * Returns the number of page that could not be moved.
625 int do_migrate_pages(struct mm_struct *mm,
626 const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
632 nodes_andnot(nodes, *from_nodes, *to_nodes);
633 nodes_complement(nodes, nodes);
635 down_read(&mm->mmap_sem);
636 check_range(mm, mm->mmap->vm_start, TASK_SIZE, &nodes,
637 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
638 if (!list_empty(&pagelist)) {
639 migrate_pages(&pagelist, NULL);
640 if (!list_empty(&pagelist))
641 count = putback_lru_pages(&pagelist);
643 up_read(&mm->mmap_sem);
648 * User space interface with variable sized bitmaps for nodelists.
651 /* Copy a node mask from user space. */
652 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
653 unsigned long maxnode)
656 unsigned long nlongs;
657 unsigned long endmask;
661 if (maxnode == 0 || !nmask)
664 nlongs = BITS_TO_LONGS(maxnode);
665 if ((maxnode % BITS_PER_LONG) == 0)
668 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
670 /* When the user specified more nodes than supported just check
671 if the non supported part is all zero. */
672 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
673 if (nlongs > PAGE_SIZE/sizeof(long))
675 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
677 if (get_user(t, nmask + k))
679 if (k == nlongs - 1) {
685 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
689 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
691 nodes_addr(*nodes)[nlongs-1] &= endmask;
695 /* Copy a kernel node mask to user space */
696 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
699 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
700 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
703 if (copy > PAGE_SIZE)
705 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
709 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
712 asmlinkage long sys_mbind(unsigned long start, unsigned long len,
714 unsigned long __user *nmask, unsigned long maxnode,
720 err = get_nodes(&nodes, nmask, maxnode);
723 return do_mbind(start, len, mode, &nodes, flags);
726 /* Set the process memory policy */
727 asmlinkage long sys_set_mempolicy(int mode, unsigned long __user *nmask,
728 unsigned long maxnode)
733 if (mode < 0 || mode > MPOL_MAX)
735 err = get_nodes(&nodes, nmask, maxnode);
738 return do_set_mempolicy(mode, &nodes);
741 /* Macro needed until Paul implements this function in kernel/cpusets.c */
742 #define cpuset_mems_allowed(task) node_online_map
744 asmlinkage long sys_migrate_pages(pid_t pid, unsigned long maxnode,
745 const unsigned long __user *old_nodes,
746 const unsigned long __user *new_nodes)
748 struct mm_struct *mm;
749 struct task_struct *task;
752 nodemask_t task_nodes;
755 err = get_nodes(&old, old_nodes, maxnode);
759 err = get_nodes(&new, new_nodes, maxnode);
763 /* Find the mm_struct */
764 read_lock(&tasklist_lock);
765 task = pid ? find_task_by_pid(pid) : current;
767 read_unlock(&tasklist_lock);
770 mm = get_task_mm(task);
771 read_unlock(&tasklist_lock);
777 * Check if this process has the right to modify the specified
778 * process. The right exists if the process has administrative
779 * capabilities, superuser priviledges or the same
780 * userid as the target process.
782 if ((current->euid != task->suid) && (current->euid != task->uid) &&
783 (current->uid != task->suid) && (current->uid != task->uid) &&
784 !capable(CAP_SYS_ADMIN)) {
789 task_nodes = cpuset_mems_allowed(task);
790 /* Is the user allowed to access the target nodes? */
791 if (!nodes_subset(new, task_nodes) && !capable(CAP_SYS_ADMIN)) {
796 err = do_migrate_pages(mm, &old, &new, MPOL_MF_MOVE);
803 /* Retrieve NUMA policy */
804 asmlinkage long sys_get_mempolicy(int __user *policy,
805 unsigned long __user *nmask,
806 unsigned long maxnode,
807 unsigned long addr, unsigned long flags)
812 if (nmask != NULL && maxnode < MAX_NUMNODES)
815 err = do_get_mempolicy(&pval, &nodes, addr, flags);
820 if (policy && put_user(pval, policy))
824 err = copy_nodes_to_user(nmask, maxnode, &nodes);
831 asmlinkage long compat_sys_get_mempolicy(int __user *policy,
832 compat_ulong_t __user *nmask,
833 compat_ulong_t maxnode,
834 compat_ulong_t addr, compat_ulong_t flags)
837 unsigned long __user *nm = NULL;
838 unsigned long nr_bits, alloc_size;
839 DECLARE_BITMAP(bm, MAX_NUMNODES);
841 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
842 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
845 nm = compat_alloc_user_space(alloc_size);
847 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
850 err = copy_from_user(bm, nm, alloc_size);
851 /* ensure entire bitmap is zeroed */
852 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
853 err |= compat_put_bitmap(nmask, bm, nr_bits);
859 asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask,
860 compat_ulong_t maxnode)
863 unsigned long __user *nm = NULL;
864 unsigned long nr_bits, alloc_size;
865 DECLARE_BITMAP(bm, MAX_NUMNODES);
867 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
868 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
871 err = compat_get_bitmap(bm, nmask, nr_bits);
872 nm = compat_alloc_user_space(alloc_size);
873 err |= copy_to_user(nm, bm, alloc_size);
879 return sys_set_mempolicy(mode, nm, nr_bits+1);
882 asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len,
883 compat_ulong_t mode, compat_ulong_t __user *nmask,
884 compat_ulong_t maxnode, compat_ulong_t flags)
887 unsigned long __user *nm = NULL;
888 unsigned long nr_bits, alloc_size;
891 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
892 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
895 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
896 nm = compat_alloc_user_space(alloc_size);
897 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
903 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
908 /* Return effective policy for a VMA */
910 get_vma_policy(struct task_struct *task, struct vm_area_struct *vma, unsigned long addr)
912 struct mempolicy *pol = task->mempolicy;
915 if (vma->vm_ops && vma->vm_ops->get_policy)
916 pol = vma->vm_ops->get_policy(vma, addr);
917 else if (vma->vm_policy &&
918 vma->vm_policy->policy != MPOL_DEFAULT)
919 pol = vma->vm_policy;
922 pol = &default_policy;
926 /* Return a zonelist representing a mempolicy */
927 static struct zonelist *zonelist_policy(gfp_t gfp, struct mempolicy *policy)
931 switch (policy->policy) {
933 nd = policy->v.preferred_node;
938 /* Lower zones don't get a policy applied */
939 /* Careful: current->mems_allowed might have moved */
940 if (gfp_zone(gfp) >= policy_zone)
941 if (cpuset_zonelist_valid_mems_allowed(policy->v.zonelist))
942 return policy->v.zonelist;
944 case MPOL_INTERLEAVE: /* should not happen */
952 return NODE_DATA(nd)->node_zonelists + gfp_zone(gfp);
955 /* Do dynamic interleaving for a process */
956 static unsigned interleave_nodes(struct mempolicy *policy)
959 struct task_struct *me = current;
962 next = next_node(nid, policy->v.nodes);
963 if (next >= MAX_NUMNODES)
964 next = first_node(policy->v.nodes);
969 /* Do static interleaving for a VMA with known offset. */
970 static unsigned offset_il_node(struct mempolicy *pol,
971 struct vm_area_struct *vma, unsigned long off)
973 unsigned nnodes = nodes_weight(pol->v.nodes);
974 unsigned target = (unsigned)off % nnodes;
980 nid = next_node(nid, pol->v.nodes);
982 } while (c <= target);
986 /* Determine a node number for interleave */
987 static inline unsigned interleave_nid(struct mempolicy *pol,
988 struct vm_area_struct *vma, unsigned long addr, int shift)
994 off += (addr - vma->vm_start) >> shift;
995 return offset_il_node(pol, vma, off);
997 return interleave_nodes(pol);
1000 /* Return a zonelist suitable for a huge page allocation. */
1001 struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr)
1003 struct mempolicy *pol = get_vma_policy(current, vma, addr);
1005 if (pol->policy == MPOL_INTERLEAVE) {
1008 nid = interleave_nid(pol, vma, addr, HPAGE_SHIFT);
1009 return NODE_DATA(nid)->node_zonelists + gfp_zone(GFP_HIGHUSER);
1011 return zonelist_policy(GFP_HIGHUSER, pol);
1014 /* Allocate a page in interleaved policy.
1015 Own path because it needs to do special accounting. */
1016 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1019 struct zonelist *zl;
1022 zl = NODE_DATA(nid)->node_zonelists + gfp_zone(gfp);
1023 page = __alloc_pages(gfp, order, zl);
1024 if (page && page_zone(page) == zl->zones[0]) {
1025 zone_pcp(zl->zones[0],get_cpu())->interleave_hit++;
1032 * alloc_page_vma - Allocate a page for a VMA.
1035 * %GFP_USER user allocation.
1036 * %GFP_KERNEL kernel allocations,
1037 * %GFP_HIGHMEM highmem/user allocations,
1038 * %GFP_FS allocation should not call back into a file system.
1039 * %GFP_ATOMIC don't sleep.
1041 * @vma: Pointer to VMA or NULL if not available.
1042 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1044 * This function allocates a page from the kernel page pool and applies
1045 * a NUMA policy associated with the VMA or the current process.
1046 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1047 * mm_struct of the VMA to prevent it from going away. Should be used for
1048 * all allocations for pages that will be mapped into
1049 * user space. Returns NULL when no page can be allocated.
1051 * Should be called with the mm_sem of the vma hold.
1054 alloc_page_vma(gfp_t gfp, struct vm_area_struct *vma, unsigned long addr)
1056 struct mempolicy *pol = get_vma_policy(current, vma, addr);
1058 cpuset_update_current_mems_allowed();
1060 if (unlikely(pol->policy == MPOL_INTERLEAVE)) {
1063 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT);
1064 return alloc_page_interleave(gfp, 0, nid);
1066 return __alloc_pages(gfp, 0, zonelist_policy(gfp, pol));
1070 * alloc_pages_current - Allocate pages.
1073 * %GFP_USER user allocation,
1074 * %GFP_KERNEL kernel allocation,
1075 * %GFP_HIGHMEM highmem allocation,
1076 * %GFP_FS don't call back into a file system.
1077 * %GFP_ATOMIC don't sleep.
1078 * @order: Power of two of allocation size in pages. 0 is a single page.
1080 * Allocate a page from the kernel page pool. When not in
1081 * interrupt context and apply the current process NUMA policy.
1082 * Returns NULL when no page can be allocated.
1084 * Don't call cpuset_update_current_mems_allowed() unless
1085 * 1) it's ok to take cpuset_sem (can WAIT), and
1086 * 2) allocating for current task (not interrupt).
1088 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
1090 struct mempolicy *pol = current->mempolicy;
1092 if ((gfp & __GFP_WAIT) && !in_interrupt())
1093 cpuset_update_current_mems_allowed();
1094 if (!pol || in_interrupt())
1095 pol = &default_policy;
1096 if (pol->policy == MPOL_INTERLEAVE)
1097 return alloc_page_interleave(gfp, order, interleave_nodes(pol));
1098 return __alloc_pages(gfp, order, zonelist_policy(gfp, pol));
1100 EXPORT_SYMBOL(alloc_pages_current);
1102 /* Slow path of a mempolicy copy */
1103 struct mempolicy *__mpol_copy(struct mempolicy *old)
1105 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
1108 return ERR_PTR(-ENOMEM);
1110 atomic_set(&new->refcnt, 1);
1111 if (new->policy == MPOL_BIND) {
1112 int sz = ksize(old->v.zonelist);
1113 new->v.zonelist = kmalloc(sz, SLAB_KERNEL);
1114 if (!new->v.zonelist) {
1115 kmem_cache_free(policy_cache, new);
1116 return ERR_PTR(-ENOMEM);
1118 memcpy(new->v.zonelist, old->v.zonelist, sz);
1123 /* Slow path of a mempolicy comparison */
1124 int __mpol_equal(struct mempolicy *a, struct mempolicy *b)
1128 if (a->policy != b->policy)
1130 switch (a->policy) {
1133 case MPOL_INTERLEAVE:
1134 return nodes_equal(a->v.nodes, b->v.nodes);
1135 case MPOL_PREFERRED:
1136 return a->v.preferred_node == b->v.preferred_node;
1139 for (i = 0; a->v.zonelist->zones[i]; i++)
1140 if (a->v.zonelist->zones[i] != b->v.zonelist->zones[i])
1142 return b->v.zonelist->zones[i] == NULL;
1150 /* Slow path of a mpol destructor. */
1151 void __mpol_free(struct mempolicy *p)
1153 if (!atomic_dec_and_test(&p->refcnt))
1155 if (p->policy == MPOL_BIND)
1156 kfree(p->v.zonelist);
1157 p->policy = MPOL_DEFAULT;
1158 kmem_cache_free(policy_cache, p);
1162 * Shared memory backing store policy support.
1164 * Remember policies even when nobody has shared memory mapped.
1165 * The policies are kept in Red-Black tree linked from the inode.
1166 * They are protected by the sp->lock spinlock, which should be held
1167 * for any accesses to the tree.
1170 /* lookup first element intersecting start-end */
1171 /* Caller holds sp->lock */
1172 static struct sp_node *
1173 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
1175 struct rb_node *n = sp->root.rb_node;
1178 struct sp_node *p = rb_entry(n, struct sp_node, nd);
1180 if (start >= p->end)
1182 else if (end <= p->start)
1190 struct sp_node *w = NULL;
1191 struct rb_node *prev = rb_prev(n);
1194 w = rb_entry(prev, struct sp_node, nd);
1195 if (w->end <= start)
1199 return rb_entry(n, struct sp_node, nd);
1202 /* Insert a new shared policy into the list. */
1203 /* Caller holds sp->lock */
1204 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
1206 struct rb_node **p = &sp->root.rb_node;
1207 struct rb_node *parent = NULL;
1212 nd = rb_entry(parent, struct sp_node, nd);
1213 if (new->start < nd->start)
1215 else if (new->end > nd->end)
1216 p = &(*p)->rb_right;
1220 rb_link_node(&new->nd, parent, p);
1221 rb_insert_color(&new->nd, &sp->root);
1222 PDprintk("inserting %lx-%lx: %d\n", new->start, new->end,
1223 new->policy ? new->policy->policy : 0);
1226 /* Find shared policy intersecting idx */
1228 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
1230 struct mempolicy *pol = NULL;
1233 if (!sp->root.rb_node)
1235 spin_lock(&sp->lock);
1236 sn = sp_lookup(sp, idx, idx+1);
1238 mpol_get(sn->policy);
1241 spin_unlock(&sp->lock);
1245 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
1247 PDprintk("deleting %lx-l%x\n", n->start, n->end);
1248 rb_erase(&n->nd, &sp->root);
1249 mpol_free(n->policy);
1250 kmem_cache_free(sn_cache, n);
1254 sp_alloc(unsigned long start, unsigned long end, struct mempolicy *pol)
1256 struct sp_node *n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
1267 /* Replace a policy range. */
1268 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
1269 unsigned long end, struct sp_node *new)
1271 struct sp_node *n, *new2 = NULL;
1274 spin_lock(&sp->lock);
1275 n = sp_lookup(sp, start, end);
1276 /* Take care of old policies in the same range. */
1277 while (n && n->start < end) {
1278 struct rb_node *next = rb_next(&n->nd);
1279 if (n->start >= start) {
1285 /* Old policy spanning whole new range. */
1288 spin_unlock(&sp->lock);
1289 new2 = sp_alloc(end, n->end, n->policy);
1295 sp_insert(sp, new2);
1303 n = rb_entry(next, struct sp_node, nd);
1307 spin_unlock(&sp->lock);
1309 mpol_free(new2->policy);
1310 kmem_cache_free(sn_cache, new2);
1315 int mpol_set_shared_policy(struct shared_policy *info,
1316 struct vm_area_struct *vma, struct mempolicy *npol)
1319 struct sp_node *new = NULL;
1320 unsigned long sz = vma_pages(vma);
1322 PDprintk("set_shared_policy %lx sz %lu %d %lx\n",
1324 sz, npol? npol->policy : -1,
1325 npol ? nodes_addr(npol->v.nodes)[0] : -1);
1328 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
1332 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
1334 kmem_cache_free(sn_cache, new);
1338 /* Free a backing policy store on inode delete. */
1339 void mpol_free_shared_policy(struct shared_policy *p)
1342 struct rb_node *next;
1344 if (!p->root.rb_node)
1346 spin_lock(&p->lock);
1347 next = rb_first(&p->root);
1349 n = rb_entry(next, struct sp_node, nd);
1350 next = rb_next(&n->nd);
1351 rb_erase(&n->nd, &p->root);
1352 mpol_free(n->policy);
1353 kmem_cache_free(sn_cache, n);
1355 spin_unlock(&p->lock);
1358 /* assumes fs == KERNEL_DS */
1359 void __init numa_policy_init(void)
1361 policy_cache = kmem_cache_create("numa_policy",
1362 sizeof(struct mempolicy),
1363 0, SLAB_PANIC, NULL, NULL);
1365 sn_cache = kmem_cache_create("shared_policy_node",
1366 sizeof(struct sp_node),
1367 0, SLAB_PANIC, NULL, NULL);
1369 /* Set interleaving policy for system init. This way not all
1370 the data structures allocated at system boot end up in node zero. */
1372 if (do_set_mempolicy(MPOL_INTERLEAVE, &node_online_map))
1373 printk("numa_policy_init: interleaving failed\n");
1376 /* Reset policy of current process to default */
1377 void numa_default_policy(void)
1379 do_set_mempolicy(MPOL_DEFAULT, NULL);
1382 /* Migrate a policy to a different set of nodes */
1383 static void rebind_policy(struct mempolicy *pol, const nodemask_t *old,
1384 const nodemask_t *new)
1391 switch (pol->policy) {
1394 case MPOL_INTERLEAVE:
1395 nodes_remap(tmp, pol->v.nodes, *old, *new);
1397 current->il_next = node_remap(current->il_next, *old, *new);
1399 case MPOL_PREFERRED:
1400 pol->v.preferred_node = node_remap(pol->v.preferred_node,
1406 struct zonelist *zonelist;
1409 for (z = pol->v.zonelist->zones; *z; z++)
1410 node_set((*z)->zone_pgdat->node_id, nodes);
1411 nodes_remap(tmp, nodes, *old, *new);
1414 zonelist = bind_zonelist(&nodes);
1416 /* If no mem, then zonelist is NULL and we keep old zonelist.
1417 * If that old zonelist has no remaining mems_allowed nodes,
1418 * then zonelist_policy() will "FALL THROUGH" to MPOL_DEFAULT.
1422 /* Good - got mem - substitute new zonelist */
1423 kfree(pol->v.zonelist);
1424 pol->v.zonelist = zonelist;
1435 * Someone moved this task to different nodes. Fixup mempolicies.
1437 * TODO - fixup current->mm->vma and shmfs/tmpfs/hugetlbfs policies as well,
1438 * once we have a cpuset mechanism to mark which cpuset subtree is migrating.
1440 void numa_policy_rebind(const nodemask_t *old, const nodemask_t *new)
1442 rebind_policy(current->mempolicy, old, new);