blob: a237295f81901522146e83c7a6900a71c6acf29a [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * Simple NUMA memory policy for the Linux kernel.
3 *
4 * Copyright 2003,2004 Andi Kleen, SuSE Labs.
Christoph Lameter8bccd852005-10-29 18:16:59 -07005 * (C) Copyright 2005 Christoph Lameter, Silicon Graphics, Inc.
Linus Torvalds1da177e2005-04-16 15:20:36 -07006 * Subject to the GNU Public License, version 2.
7 *
8 * NUMA policy allows the user to give hints in which node(s) memory should
9 * be allocated.
10 *
11 * Support four policies per VMA and per process:
12 *
13 * The VMA policy has priority over the process policy for a page fault.
14 *
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
20 * is used.
Christoph Lameter8bccd852005-10-29 18:16:59 -070021 *
Linus Torvalds1da177e2005-04-16 15:20:36 -070022 * bind Only allocate memory on a specific set of nodes,
23 * no fallback.
Christoph Lameter8bccd852005-10-29 18:16:59 -070024 * 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
27 *
Linus Torvalds1da177e2005-04-16 15:20:36 -070028 * 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
32 * process policy.
Christoph Lameter8bccd852005-10-29 18:16:59 -070033 *
Linus Torvalds1da177e2005-04-16 15:20:36 -070034 * 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.
37 *
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.
42 *
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.
46 *
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.
51 *
52 * For shmfs/tmpfs/hugetlbfs shared memory the policy is shared between
53 * all users and remembered even when nobody has memory mapped.
54 */
55
56/* Notebook:
57 fix mmap readahead to honour policy and enable policy for any page cache
58 object
59 statistics for bigpages
60 global policy for page cache? currently it uses process policy. Requires
61 first item above.
62 handle mremap for shared memory (currently ignored for the policy)
63 grows down?
64 make bind policy root only? It can trigger oom much faster and the
65 kernel is not always grateful with that.
Linus Torvalds1da177e2005-04-16 15:20:36 -070066*/
67
68#include <linux/mempolicy.h>
69#include <linux/mm.h>
70#include <linux/highmem.h>
71#include <linux/hugetlb.h>
72#include <linux/kernel.h>
73#include <linux/sched.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070074#include <linux/nodemask.h>
75#include <linux/cpuset.h>
76#include <linux/gfp.h>
77#include <linux/slab.h>
78#include <linux/string.h>
79#include <linux/module.h>
Pavel Emelyanovb4888932007-10-18 23:40:14 -070080#include <linux/nsproxy.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070081#include <linux/interrupt.h>
82#include <linux/init.h>
83#include <linux/compat.h>
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -080084#include <linux/swap.h>
Christoph Lameter1a75a6c2006-01-08 01:01:02 -080085#include <linux/seq_file.h>
86#include <linux/proc_fs.h>
Christoph Lameterb20a3502006-03-22 00:09:12 -080087#include <linux/migrate.h>
Christoph Lameter95a402c2006-06-23 02:03:53 -070088#include <linux/rmap.h>
David Quigley86c3a762006-06-23 02:04:02 -070089#include <linux/security.h>
Adrian Bunkdbcb0f12007-10-16 01:26:26 -070090#include <linux/syscalls.h>
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -080091
Linus Torvalds1da177e2005-04-16 15:20:36 -070092#include <asm/tlbflush.h>
93#include <asm/uaccess.h>
94
Christoph Lameter38e35862006-01-08 01:01:01 -080095/* Internal flags */
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -080096#define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
Christoph Lameter38e35862006-01-08 01:01:01 -080097#define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
Christoph Lameter1a75a6c2006-01-08 01:01:02 -080098#define MPOL_MF_STATS (MPOL_MF_INTERNAL << 2) /* Gather statistics */
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -080099
Pekka Enbergfcc234f2006-03-22 00:08:13 -0800100static struct kmem_cache *policy_cache;
101static struct kmem_cache *sn_cache;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700102
Linus Torvalds1da177e2005-04-16 15:20:36 -0700103/* Highest zone. An specific allocation for a zone below that is not
104 policied. */
Christoph Lameter62672762007-02-10 01:43:07 -0800105enum zone_type policy_zone = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700106
Andi Kleend42c6992005-07-06 19:56:03 +0200107struct mempolicy default_policy = {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700108 .refcnt = ATOMIC_INIT(1), /* never free it */
Lee Schermerhorn45c47452008-04-28 02:13:12 -0700109 .mode = MPOL_DEFAULT,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700110};
111
David Rientjes37012942008-04-28 02:12:33 -0700112static const struct mempolicy_operations {
113 int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
114 void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes);
115} mpol_ops[MPOL_MAX];
116
Mel Gorman19770b32008-04-28 02:12:18 -0700117/* Check that the nodemask contains at least one populated zone */
David Rientjes37012942008-04-28 02:12:33 -0700118static int is_valid_nodemask(const nodemask_t *nodemask)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700119{
Mel Gorman19770b32008-04-28 02:12:18 -0700120 int nd, k;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700121
Mel Gorman19770b32008-04-28 02:12:18 -0700122 /* Check that there is something useful in this mask */
123 k = policy_zone;
124
125 for_each_node_mask(nd, *nodemask) {
126 struct zone *z;
127
128 for (k = 0; k <= policy_zone; k++) {
129 z = &NODE_DATA(nd)->node_zones[k];
130 if (z->present_pages > 0)
131 return 1;
Andi Kleendd942ae2006-02-17 01:39:16 +0100132 }
133 }
Mel Gorman19770b32008-04-28 02:12:18 -0700134
135 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700136}
137
David Rientjesf5b087b2008-04-28 02:12:27 -0700138static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
139{
David Rientjes4c50bc02008-04-28 02:12:30 -0700140 return pol->flags & (MPOL_F_STATIC_NODES | MPOL_F_RELATIVE_NODES);
141}
142
143static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
144 const nodemask_t *rel)
145{
146 nodemask_t tmp;
147 nodes_fold(tmp, *orig, nodes_weight(*rel));
148 nodes_onto(*ret, tmp, *rel);
David Rientjesf5b087b2008-04-28 02:12:27 -0700149}
150
David Rientjes37012942008-04-28 02:12:33 -0700151static int mpol_new_interleave(struct mempolicy *pol, const nodemask_t *nodes)
152{
153 if (nodes_empty(*nodes))
154 return -EINVAL;
155 pol->v.nodes = *nodes;
156 return 0;
157}
158
159static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
160{
161 if (!nodes)
162 pol->v.preferred_node = -1; /* local allocation */
163 else if (nodes_empty(*nodes))
164 return -EINVAL; /* no allowed nodes */
165 else
166 pol->v.preferred_node = first_node(*nodes);
167 return 0;
168}
169
170static int mpol_new_bind(struct mempolicy *pol, const nodemask_t *nodes)
171{
172 if (!is_valid_nodemask(nodes))
173 return -EINVAL;
174 pol->v.nodes = *nodes;
175 return 0;
176}
177
Linus Torvalds1da177e2005-04-16 15:20:36 -0700178/* Create a new policy */
David Rientjes028fec42008-04-28 02:12:25 -0700179static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
180 nodemask_t *nodes)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700181{
182 struct mempolicy *policy;
David Rientjesf5b087b2008-04-28 02:12:27 -0700183 nodemask_t cpuset_context_nmask;
David Rientjes37012942008-04-28 02:12:33 -0700184 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700185
David Rientjes028fec42008-04-28 02:12:25 -0700186 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
187 mode, flags, nodes ? nodes_addr(*nodes)[0] : -1);
Paul Mundt140d5a42007-07-15 23:38:16 -0700188
David Rientjes3e1f0642008-04-28 02:12:34 -0700189 if (mode == MPOL_DEFAULT) {
190 if (nodes && !nodes_empty(*nodes))
David Rientjes37012942008-04-28 02:12:33 -0700191 return ERR_PTR(-EINVAL);
David Rientjes3e1f0642008-04-28 02:12:34 -0700192 return NULL;
David Rientjes37012942008-04-28 02:12:33 -0700193 }
David Rientjes3e1f0642008-04-28 02:12:34 -0700194 VM_BUG_ON(!nodes);
195
196 /*
197 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
198 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
199 * All other modes require a valid pointer to a non-empty nodemask.
200 */
201 if (mode == MPOL_PREFERRED) {
202 if (nodes_empty(*nodes)) {
203 if (((flags & MPOL_F_STATIC_NODES) ||
204 (flags & MPOL_F_RELATIVE_NODES)))
205 return ERR_PTR(-EINVAL);
206 nodes = NULL; /* flag local alloc */
207 }
208 } else if (nodes_empty(*nodes))
209 return ERR_PTR(-EINVAL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700210 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
211 if (!policy)
212 return ERR_PTR(-ENOMEM);
213 atomic_set(&policy->refcnt, 1);
Lee Schermerhorn45c47452008-04-28 02:13:12 -0700214 policy->mode = mode;
David Rientjes3e1f0642008-04-28 02:12:34 -0700215 policy->flags = flags;
David Rientjesf5b087b2008-04-28 02:12:27 -0700216
David Rientjes3e1f0642008-04-28 02:12:34 -0700217 if (nodes) {
218 /*
219 * cpuset related setup doesn't apply to local allocation
220 */
David Rientjes37012942008-04-28 02:12:33 -0700221 cpuset_update_task_memory_state();
222 if (flags & MPOL_F_RELATIVE_NODES)
223 mpol_relative_nodemask(&cpuset_context_nmask, nodes,
224 &cpuset_current_mems_allowed);
225 else
226 nodes_and(cpuset_context_nmask, *nodes,
227 cpuset_current_mems_allowed);
228 if (mpol_store_user_nodemask(policy))
229 policy->w.user_nodemask = *nodes;
230 else
231 policy->w.cpuset_mems_allowed =
232 cpuset_mems_allowed(current);
233 }
234
235 ret = mpol_ops[mode].create(policy,
David Rientjes3e1f0642008-04-28 02:12:34 -0700236 nodes ? &cpuset_context_nmask : NULL);
David Rientjes37012942008-04-28 02:12:33 -0700237 if (ret < 0) {
238 kmem_cache_free(policy_cache, policy);
239 return ERR_PTR(ret);
240 }
241 return policy;
242}
243
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -0700244/* Slow path of a mpol destructor. */
245void __mpol_put(struct mempolicy *p)
246{
247 if (!atomic_dec_and_test(&p->refcnt))
248 return;
249 p->mode = MPOL_DEFAULT;
250 kmem_cache_free(policy_cache, p);
251}
252
David Rientjes37012942008-04-28 02:12:33 -0700253static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes)
254{
255}
256
257static void mpol_rebind_nodemask(struct mempolicy *pol,
258 const nodemask_t *nodes)
259{
260 nodemask_t tmp;
261
262 if (pol->flags & MPOL_F_STATIC_NODES)
263 nodes_and(tmp, pol->w.user_nodemask, *nodes);
264 else if (pol->flags & MPOL_F_RELATIVE_NODES)
265 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
266 else {
267 nodes_remap(tmp, pol->v.nodes, pol->w.cpuset_mems_allowed,
268 *nodes);
269 pol->w.cpuset_mems_allowed = *nodes;
270 }
271
272 pol->v.nodes = tmp;
273 if (!node_isset(current->il_next, tmp)) {
274 current->il_next = next_node(current->il_next, tmp);
275 if (current->il_next >= MAX_NUMNODES)
276 current->il_next = first_node(tmp);
277 if (current->il_next >= MAX_NUMNODES)
278 current->il_next = numa_node_id();
279 }
280}
281
282static void mpol_rebind_preferred(struct mempolicy *pol,
283 const nodemask_t *nodes)
284{
285 nodemask_t tmp;
286
David Rientjes37012942008-04-28 02:12:33 -0700287 if (pol->flags & MPOL_F_STATIC_NODES) {
288 int node = first_node(pol->w.user_nodemask);
289
290 if (node_isset(node, *nodes))
291 pol->v.preferred_node = node;
292 else
293 pol->v.preferred_node = -1;
David Rientjes37012942008-04-28 02:12:33 -0700294 } else if (pol->flags & MPOL_F_RELATIVE_NODES) {
295 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
296 pol->v.preferred_node = first_node(tmp);
David Rientjes3e1f0642008-04-28 02:12:34 -0700297 } else if (pol->v.preferred_node != -1) {
David Rientjes37012942008-04-28 02:12:33 -0700298 pol->v.preferred_node = node_remap(pol->v.preferred_node,
299 pol->w.cpuset_mems_allowed,
300 *nodes);
301 pol->w.cpuset_mems_allowed = *nodes;
302 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700303}
304
David Rientjes1d0d2682008-04-28 02:12:32 -0700305/* Migrate a policy to a different set of nodes */
306static void mpol_rebind_policy(struct mempolicy *pol,
307 const nodemask_t *newmask)
308{
David Rientjes1d0d2682008-04-28 02:12:32 -0700309 if (!pol)
310 return;
David Rientjes1d0d2682008-04-28 02:12:32 -0700311 if (!mpol_store_user_nodemask(pol) &&
312 nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
313 return;
Lee Schermerhorn45c47452008-04-28 02:13:12 -0700314 mpol_ops[pol->mode].rebind(pol, newmask);
David Rientjes1d0d2682008-04-28 02:12:32 -0700315}
316
317/*
318 * Wrapper for mpol_rebind_policy() that just requires task
319 * pointer, and updates task mempolicy.
320 */
321
322void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new)
323{
324 mpol_rebind_policy(tsk->mempolicy, new);
325}
326
327/*
328 * Rebind each vma in mm to new nodemask.
329 *
330 * Call holding a reference to mm. Takes mm->mmap_sem during call.
331 */
332
333void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
334{
335 struct vm_area_struct *vma;
336
337 down_write(&mm->mmap_sem);
338 for (vma = mm->mmap; vma; vma = vma->vm_next)
339 mpol_rebind_policy(vma->vm_policy, new);
340 up_write(&mm->mmap_sem);
341}
342
David Rientjes37012942008-04-28 02:12:33 -0700343static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
344 [MPOL_DEFAULT] = {
345 .rebind = mpol_rebind_default,
346 },
347 [MPOL_INTERLEAVE] = {
348 .create = mpol_new_interleave,
349 .rebind = mpol_rebind_nodemask,
350 },
351 [MPOL_PREFERRED] = {
352 .create = mpol_new_preferred,
353 .rebind = mpol_rebind_preferred,
354 },
355 [MPOL_BIND] = {
356 .create = mpol_new_bind,
357 .rebind = mpol_rebind_nodemask,
358 },
359};
360
Christoph Lameter397874d2006-03-06 15:42:53 -0800361static void gather_stats(struct page *, void *, int pte_dirty);
Christoph Lameterfc301282006-01-18 17:42:29 -0800362static void migrate_page_add(struct page *page, struct list_head *pagelist,
363 unsigned long flags);
Christoph Lameter1a75a6c2006-01-08 01:01:02 -0800364
Christoph Lameter38e35862006-01-08 01:01:01 -0800365/* Scan through pages checking if pages follow certain conditions. */
Nick Pigginb5810032005-10-29 18:16:12 -0700366static int check_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800367 unsigned long addr, unsigned long end,
368 const nodemask_t *nodes, unsigned long flags,
Christoph Lameter38e35862006-01-08 01:01:01 -0800369 void *private)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700370{
Hugh Dickins91612e02005-06-21 17:15:07 -0700371 pte_t *orig_pte;
372 pte_t *pte;
Hugh Dickins705e87c2005-10-29 18:16:27 -0700373 spinlock_t *ptl;
Hugh Dickins941150a2005-06-21 17:15:06 -0700374
Hugh Dickins705e87c2005-10-29 18:16:27 -0700375 orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
Hugh Dickins91612e02005-06-21 17:15:07 -0700376 do {
Linus Torvalds6aab3412005-11-28 14:34:23 -0800377 struct page *page;
Andy Whitcroft25ba77c2006-12-06 20:33:03 -0800378 int nid;
Hugh Dickins91612e02005-06-21 17:15:07 -0700379
380 if (!pte_present(*pte))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700381 continue;
Linus Torvalds6aab3412005-11-28 14:34:23 -0800382 page = vm_normal_page(vma, addr, *pte);
383 if (!page)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700384 continue;
Nick Piggin053837f2006-01-18 17:42:27 -0800385 /*
386 * The check for PageReserved here is important to avoid
387 * handling zero pages and other pages that may have been
388 * marked special by the system.
389 *
390 * If the PageReserved would not be checked here then f.e.
391 * the location of the zero page could have an influence
392 * on MPOL_MF_STRICT, zero pages would be counted for
393 * the per node stats, and there would be useless attempts
394 * to put zero pages on the migration list.
395 */
Christoph Lameterf4598c82006-01-12 01:05:20 -0800396 if (PageReserved(page))
397 continue;
Linus Torvalds6aab3412005-11-28 14:34:23 -0800398 nid = page_to_nid(page);
Christoph Lameter38e35862006-01-08 01:01:01 -0800399 if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
400 continue;
401
Christoph Lameter1a75a6c2006-01-08 01:01:02 -0800402 if (flags & MPOL_MF_STATS)
Christoph Lameter397874d2006-03-06 15:42:53 -0800403 gather_stats(page, private, pte_dirty(*pte));
Nick Piggin053837f2006-01-18 17:42:27 -0800404 else if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
Christoph Lameterfc301282006-01-18 17:42:29 -0800405 migrate_page_add(page, private, flags);
Christoph Lameter38e35862006-01-08 01:01:01 -0800406 else
407 break;
Hugh Dickins91612e02005-06-21 17:15:07 -0700408 } while (pte++, addr += PAGE_SIZE, addr != end);
Hugh Dickins705e87c2005-10-29 18:16:27 -0700409 pte_unmap_unlock(orig_pte, ptl);
Hugh Dickins91612e02005-06-21 17:15:07 -0700410 return addr != end;
411}
412
Nick Pigginb5810032005-10-29 18:16:12 -0700413static inline int check_pmd_range(struct vm_area_struct *vma, pud_t *pud,
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800414 unsigned long addr, unsigned long end,
415 const nodemask_t *nodes, unsigned long flags,
Christoph Lameter38e35862006-01-08 01:01:01 -0800416 void *private)
Hugh Dickins91612e02005-06-21 17:15:07 -0700417{
418 pmd_t *pmd;
419 unsigned long next;
420
421 pmd = pmd_offset(pud, addr);
422 do {
423 next = pmd_addr_end(addr, end);
424 if (pmd_none_or_clear_bad(pmd))
425 continue;
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800426 if (check_pte_range(vma, pmd, addr, next, nodes,
Christoph Lameter38e35862006-01-08 01:01:01 -0800427 flags, private))
Hugh Dickins91612e02005-06-21 17:15:07 -0700428 return -EIO;
429 } while (pmd++, addr = next, addr != end);
430 return 0;
431}
432
Nick Pigginb5810032005-10-29 18:16:12 -0700433static inline int check_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800434 unsigned long addr, unsigned long end,
435 const nodemask_t *nodes, unsigned long flags,
Christoph Lameter38e35862006-01-08 01:01:01 -0800436 void *private)
Hugh Dickins91612e02005-06-21 17:15:07 -0700437{
438 pud_t *pud;
439 unsigned long next;
440
441 pud = pud_offset(pgd, addr);
442 do {
443 next = pud_addr_end(addr, end);
444 if (pud_none_or_clear_bad(pud))
445 continue;
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800446 if (check_pmd_range(vma, pud, addr, next, nodes,
Christoph Lameter38e35862006-01-08 01:01:01 -0800447 flags, private))
Hugh Dickins91612e02005-06-21 17:15:07 -0700448 return -EIO;
449 } while (pud++, addr = next, addr != end);
450 return 0;
451}
452
Nick Pigginb5810032005-10-29 18:16:12 -0700453static inline int check_pgd_range(struct vm_area_struct *vma,
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800454 unsigned long addr, unsigned long end,
455 const nodemask_t *nodes, unsigned long flags,
Christoph Lameter38e35862006-01-08 01:01:01 -0800456 void *private)
Hugh Dickins91612e02005-06-21 17:15:07 -0700457{
458 pgd_t *pgd;
459 unsigned long next;
460
Nick Pigginb5810032005-10-29 18:16:12 -0700461 pgd = pgd_offset(vma->vm_mm, addr);
Hugh Dickins91612e02005-06-21 17:15:07 -0700462 do {
463 next = pgd_addr_end(addr, end);
464 if (pgd_none_or_clear_bad(pgd))
465 continue;
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800466 if (check_pud_range(vma, pgd, addr, next, nodes,
Christoph Lameter38e35862006-01-08 01:01:01 -0800467 flags, private))
Hugh Dickins91612e02005-06-21 17:15:07 -0700468 return -EIO;
469 } while (pgd++, addr = next, addr != end);
470 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700471}
472
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800473/*
474 * Check if all pages in a range are on a set of nodes.
475 * If pagelist != NULL then isolate pages from the LRU and
476 * put them on the pagelist.
477 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700478static struct vm_area_struct *
479check_range(struct mm_struct *mm, unsigned long start, unsigned long end,
Christoph Lameter38e35862006-01-08 01:01:01 -0800480 const nodemask_t *nodes, unsigned long flags, void *private)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700481{
482 int err;
483 struct vm_area_struct *first, *vma, *prev;
484
Christoph Lameter90036ee2006-03-16 23:03:59 -0800485 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
Christoph Lameter90036ee2006-03-16 23:03:59 -0800486
Christoph Lameterb20a3502006-03-22 00:09:12 -0800487 err = migrate_prep();
488 if (err)
489 return ERR_PTR(err);
Christoph Lameter90036ee2006-03-16 23:03:59 -0800490 }
Nick Piggin053837f2006-01-18 17:42:27 -0800491
Linus Torvalds1da177e2005-04-16 15:20:36 -0700492 first = find_vma(mm, start);
493 if (!first)
494 return ERR_PTR(-EFAULT);
495 prev = NULL;
496 for (vma = first; vma && vma->vm_start < end; vma = vma->vm_next) {
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800497 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
498 if (!vma->vm_next && vma->vm_end < end)
499 return ERR_PTR(-EFAULT);
500 if (prev && prev->vm_end < vma->vm_start)
501 return ERR_PTR(-EFAULT);
502 }
503 if (!is_vm_hugetlb_page(vma) &&
504 ((flags & MPOL_MF_STRICT) ||
505 ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) &&
506 vma_migratable(vma)))) {
Andi Kleen5b952b32005-09-13 01:25:08 -0700507 unsigned long endvma = vma->vm_end;
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800508
Andi Kleen5b952b32005-09-13 01:25:08 -0700509 if (endvma > end)
510 endvma = end;
511 if (vma->vm_start > start)
512 start = vma->vm_start;
Christoph Lameterdc9aa5b2006-01-08 01:00:50 -0800513 err = check_pgd_range(vma, start, endvma, nodes,
Christoph Lameter38e35862006-01-08 01:01:01 -0800514 flags, private);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700515 if (err) {
516 first = ERR_PTR(err);
517 break;
518 }
519 }
520 prev = vma;
521 }
522 return first;
523}
524
525/* Apply policy to a single VMA */
526static int policy_vma(struct vm_area_struct *vma, struct mempolicy *new)
527{
528 int err = 0;
529 struct mempolicy *old = vma->vm_policy;
530
Paul Mundt140d5a42007-07-15 23:38:16 -0700531 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700532 vma->vm_start, vma->vm_end, vma->vm_pgoff,
533 vma->vm_ops, vma->vm_file,
534 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
535
536 if (vma->vm_ops && vma->vm_ops->set_policy)
537 err = vma->vm_ops->set_policy(vma, new);
538 if (!err) {
539 mpol_get(new);
540 vma->vm_policy = new;
Lee Schermerhornf0be3d32008-04-28 02:13:08 -0700541 mpol_put(old);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700542 }
543 return err;
544}
545
546/* Step 2: apply policy to a range and do splits. */
547static int mbind_range(struct vm_area_struct *vma, unsigned long start,
548 unsigned long end, struct mempolicy *new)
549{
550 struct vm_area_struct *next;
551 int err;
552
553 err = 0;
554 for (; vma && vma->vm_start < end; vma = next) {
555 next = vma->vm_next;
556 if (vma->vm_start < start)
557 err = split_vma(vma->vm_mm, vma, start, 1);
558 if (!err && vma->vm_end > end)
559 err = split_vma(vma->vm_mm, vma, end, 0);
560 if (!err)
561 err = policy_vma(vma, new);
562 if (err)
563 break;
564 }
565 return err;
566}
567
Paul Jacksonc61afb12006-03-24 03:16:08 -0800568/*
569 * Update task->flags PF_MEMPOLICY bit: set iff non-default
570 * mempolicy. Allows more rapid checking of this (combined perhaps
571 * with other PF_* flag bits) on memory allocation hot code paths.
572 *
573 * If called from outside this file, the task 'p' should -only- be
574 * a newly forked child not yet visible on the task list, because
575 * manipulating the task flags of a visible task is not safe.
576 *
577 * The above limitation is why this routine has the funny name
578 * mpol_fix_fork_child_flag().
579 *
580 * It is also safe to call this with a task pointer of current,
581 * which the static wrapper mpol_set_task_struct_flag() does,
582 * for use within this file.
583 */
584
585void mpol_fix_fork_child_flag(struct task_struct *p)
586{
587 if (p->mempolicy)
588 p->flags |= PF_MEMPOLICY;
589 else
590 p->flags &= ~PF_MEMPOLICY;
591}
592
593static void mpol_set_task_struct_flag(void)
594{
595 mpol_fix_fork_child_flag(current);
596}
597
Linus Torvalds1da177e2005-04-16 15:20:36 -0700598/* Set the process memory policy */
David Rientjes028fec42008-04-28 02:12:25 -0700599static long do_set_mempolicy(unsigned short mode, unsigned short flags,
600 nodemask_t *nodes)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700601{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700602 struct mempolicy *new;
Lee Schermerhornf4e53d92008-04-28 02:13:10 -0700603 struct mm_struct *mm = current->mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700604
David Rientjes028fec42008-04-28 02:12:25 -0700605 new = mpol_new(mode, flags, nodes);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700606 if (IS_ERR(new))
607 return PTR_ERR(new);
Lee Schermerhornf4e53d92008-04-28 02:13:10 -0700608
609 /*
610 * prevent changing our mempolicy while show_numa_maps()
611 * is using it.
612 * Note: do_set_mempolicy() can be called at init time
613 * with no 'mm'.
614 */
615 if (mm)
616 down_write(&mm->mmap_sem);
Lee Schermerhornf0be3d32008-04-28 02:13:08 -0700617 mpol_put(current->mempolicy);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700618 current->mempolicy = new;
Paul Jacksonc61afb12006-03-24 03:16:08 -0800619 mpol_set_task_struct_flag();
Lee Schermerhorn45c47452008-04-28 02:13:12 -0700620 if (new && new->mode == MPOL_INTERLEAVE &&
David Rientjesf5b087b2008-04-28 02:12:27 -0700621 nodes_weight(new->v.nodes))
Andi Kleendfcd3c02005-10-29 18:15:48 -0700622 current->il_next = first_node(new->v.nodes);
Lee Schermerhornf4e53d92008-04-28 02:13:10 -0700623 if (mm)
624 up_write(&mm->mmap_sem);
625
Linus Torvalds1da177e2005-04-16 15:20:36 -0700626 return 0;
627}
628
629/* Fill a zone bitmap for a policy */
Andi Kleendfcd3c02005-10-29 18:15:48 -0700630static void get_zonemask(struct mempolicy *p, nodemask_t *nodes)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700631{
Andi Kleendfcd3c02005-10-29 18:15:48 -0700632 nodes_clear(*nodes);
Lee Schermerhorn45c47452008-04-28 02:13:12 -0700633 switch (p->mode) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700634 case MPOL_DEFAULT:
635 break;
Mel Gorman19770b32008-04-28 02:12:18 -0700636 case MPOL_BIND:
637 /* Fall through */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700638 case MPOL_INTERLEAVE:
Andi Kleendfcd3c02005-10-29 18:15:48 -0700639 *nodes = p->v.nodes;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700640 break;
641 case MPOL_PREFERRED:
Christoph Lameter56bbd652007-10-16 01:25:35 -0700642 /* or use current node instead of memory_map? */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700643 if (p->v.preferred_node < 0)
Christoph Lameter56bbd652007-10-16 01:25:35 -0700644 *nodes = node_states[N_HIGH_MEMORY];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700645 else
Andi Kleendfcd3c02005-10-29 18:15:48 -0700646 node_set(p->v.preferred_node, *nodes);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700647 break;
648 default:
649 BUG();
650 }
651}
652
653static int lookup_node(struct mm_struct *mm, unsigned long addr)
654{
655 struct page *p;
656 int err;
657
658 err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL);
659 if (err >= 0) {
660 err = page_to_nid(p);
661 put_page(p);
662 }
663 return err;
664}
665
Linus Torvalds1da177e2005-04-16 15:20:36 -0700666/* Retrieve NUMA policy */
Adrian Bunkdbcb0f12007-10-16 01:26:26 -0700667static long do_get_mempolicy(int *policy, nodemask_t *nmask,
668 unsigned long addr, unsigned long flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700669{
Christoph Lameter8bccd852005-10-29 18:16:59 -0700670 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700671 struct mm_struct *mm = current->mm;
672 struct vm_area_struct *vma = NULL;
673 struct mempolicy *pol = current->mempolicy;
674
Paul Jacksoncf2a473c2006-01-08 01:01:54 -0800675 cpuset_update_task_memory_state();
Lee Schermerhorn754af6f2007-10-16 01:24:51 -0700676 if (flags &
677 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700678 return -EINVAL;
Lee Schermerhorn754af6f2007-10-16 01:24:51 -0700679
680 if (flags & MPOL_F_MEMS_ALLOWED) {
681 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
682 return -EINVAL;
683 *policy = 0; /* just so it's initialized */
684 *nmask = cpuset_current_mems_allowed;
685 return 0;
686 }
687
Linus Torvalds1da177e2005-04-16 15:20:36 -0700688 if (flags & MPOL_F_ADDR) {
689 down_read(&mm->mmap_sem);
690 vma = find_vma_intersection(mm, addr, addr+1);
691 if (!vma) {
692 up_read(&mm->mmap_sem);
693 return -EFAULT;
694 }
695 if (vma->vm_ops && vma->vm_ops->get_policy)
696 pol = vma->vm_ops->get_policy(vma, addr);
697 else
698 pol = vma->vm_policy;
699 } else if (addr)
700 return -EINVAL;
701
702 if (!pol)
703 pol = &default_policy;
704
705 if (flags & MPOL_F_NODE) {
706 if (flags & MPOL_F_ADDR) {
707 err = lookup_node(mm, addr);
708 if (err < 0)
709 goto out;
Christoph Lameter8bccd852005-10-29 18:16:59 -0700710 *policy = err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700711 } else if (pol == current->mempolicy &&
Lee Schermerhorn45c47452008-04-28 02:13:12 -0700712 pol->mode == MPOL_INTERLEAVE) {
Christoph Lameter8bccd852005-10-29 18:16:59 -0700713 *policy = current->il_next;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700714 } else {
715 err = -EINVAL;
716 goto out;
717 }
718 } else
Lee Schermerhorn45c47452008-04-28 02:13:12 -0700719 *policy = pol->mode | pol->flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700720
721 if (vma) {
722 up_read(&current->mm->mmap_sem);
723 vma = NULL;
724 }
725
Linus Torvalds1da177e2005-04-16 15:20:36 -0700726 err = 0;
Christoph Lameter8bccd852005-10-29 18:16:59 -0700727 if (nmask)
728 get_zonemask(pol, nmask);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700729
730 out:
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -0700731 mpol_cond_put(pol);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700732 if (vma)
733 up_read(&current->mm->mmap_sem);
734 return err;
735}
736
Christoph Lameterb20a3502006-03-22 00:09:12 -0800737#ifdef CONFIG_MIGRATION
Christoph Lameter8bccd852005-10-29 18:16:59 -0700738/*
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -0800739 * page migration
740 */
Christoph Lameterfc301282006-01-18 17:42:29 -0800741static void migrate_page_add(struct page *page, struct list_head *pagelist,
742 unsigned long flags)
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -0800743{
744 /*
Christoph Lameterfc301282006-01-18 17:42:29 -0800745 * Avoid migrating a page that is shared with others.
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -0800746 */
Christoph Lameterb20a3502006-03-22 00:09:12 -0800747 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1)
748 isolate_lru_page(page, pagelist);
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -0800749}
750
Christoph Lameter742755a2006-06-23 02:03:55 -0700751static struct page *new_node_page(struct page *page, unsigned long node, int **x)
Christoph Lameter95a402c2006-06-23 02:03:53 -0700752{
Mel Gorman769848c2007-07-17 04:03:05 -0700753 return alloc_pages_node(node, GFP_HIGHUSER_MOVABLE, 0);
Christoph Lameter95a402c2006-06-23 02:03:53 -0700754}
755
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -0800756/*
Christoph Lameter7e2ab152006-02-01 03:05:40 -0800757 * Migrate pages from one node to a target node.
758 * Returns error or the number of pages not migrated.
759 */
Adrian Bunkdbcb0f12007-10-16 01:26:26 -0700760static int migrate_to_node(struct mm_struct *mm, int source, int dest,
761 int flags)
Christoph Lameter7e2ab152006-02-01 03:05:40 -0800762{
763 nodemask_t nmask;
764 LIST_HEAD(pagelist);
765 int err = 0;
766
767 nodes_clear(nmask);
768 node_set(source, nmask);
769
770 check_range(mm, mm->mmap->vm_start, TASK_SIZE, &nmask,
771 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
772
Christoph Lameteraaa994b2006-06-23 02:03:52 -0700773 if (!list_empty(&pagelist))
Christoph Lameter95a402c2006-06-23 02:03:53 -0700774 err = migrate_pages(&pagelist, new_node_page, dest);
775
Christoph Lameter7e2ab152006-02-01 03:05:40 -0800776 return err;
777}
778
779/*
780 * Move pages between the two nodesets so as to preserve the physical
781 * layout as much as possible.
Christoph Lameter39743882006-01-08 01:00:51 -0800782 *
783 * Returns the number of page that could not be moved.
784 */
785int do_migrate_pages(struct mm_struct *mm,
786 const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
787{
788 LIST_HEAD(pagelist);
Christoph Lameter7e2ab152006-02-01 03:05:40 -0800789 int busy = 0;
790 int err = 0;
791 nodemask_t tmp;
Christoph Lameter39743882006-01-08 01:00:51 -0800792
Christoph Lameter7e2ab152006-02-01 03:05:40 -0800793 down_read(&mm->mmap_sem);
Christoph Lameter39743882006-01-08 01:00:51 -0800794
Christoph Lameter7b2259b2006-06-25 05:46:48 -0700795 err = migrate_vmas(mm, from_nodes, to_nodes, flags);
796 if (err)
797 goto out;
798
Christoph Lameter7e2ab152006-02-01 03:05:40 -0800799/*
800 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
801 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
802 * bit in 'tmp', and return that <source, dest> pair for migration.
803 * The pair of nodemasks 'to' and 'from' define the map.
804 *
805 * If no pair of bits is found that way, fallback to picking some
806 * pair of 'source' and 'dest' bits that are not the same. If the
807 * 'source' and 'dest' bits are the same, this represents a node
808 * that will be migrating to itself, so no pages need move.
809 *
810 * If no bits are left in 'tmp', or if all remaining bits left
811 * in 'tmp' correspond to the same bit in 'to', return false
812 * (nothing left to migrate).
813 *
814 * This lets us pick a pair of nodes to migrate between, such that
815 * if possible the dest node is not already occupied by some other
816 * source node, minimizing the risk of overloading the memory on a
817 * node that would happen if we migrated incoming memory to a node
818 * before migrating outgoing memory source that same node.
819 *
820 * A single scan of tmp is sufficient. As we go, we remember the
821 * most recent <s, d> pair that moved (s != d). If we find a pair
822 * that not only moved, but what's better, moved to an empty slot
823 * (d is not set in tmp), then we break out then, with that pair.
824 * Otherwise when we finish scannng from_tmp, we at least have the
825 * most recent <s, d> pair that moved. If we get all the way through
826 * the scan of tmp without finding any node that moved, much less
827 * moved to an empty node, then there is nothing left worth migrating.
828 */
Christoph Lameterd4984712006-01-08 01:00:55 -0800829
Christoph Lameter7e2ab152006-02-01 03:05:40 -0800830 tmp = *from_nodes;
831 while (!nodes_empty(tmp)) {
832 int s,d;
833 int source = -1;
834 int dest = 0;
835
836 for_each_node_mask(s, tmp) {
837 d = node_remap(s, *from_nodes, *to_nodes);
838 if (s == d)
839 continue;
840
841 source = s; /* Node moved. Memorize */
842 dest = d;
843
844 /* dest not in remaining from nodes? */
845 if (!node_isset(dest, tmp))
846 break;
847 }
848 if (source == -1)
849 break;
850
851 node_clear(source, tmp);
852 err = migrate_to_node(mm, source, dest, flags);
853 if (err > 0)
854 busy += err;
855 if (err < 0)
856 break;
Christoph Lameter39743882006-01-08 01:00:51 -0800857 }
Christoph Lameter7b2259b2006-06-25 05:46:48 -0700858out:
Christoph Lameter39743882006-01-08 01:00:51 -0800859 up_read(&mm->mmap_sem);
Christoph Lameter7e2ab152006-02-01 03:05:40 -0800860 if (err < 0)
861 return err;
862 return busy;
Christoph Lameterb20a3502006-03-22 00:09:12 -0800863
Christoph Lameter39743882006-01-08 01:00:51 -0800864}
865
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -0800866/*
867 * Allocate a new page for page migration based on vma policy.
868 * Start assuming that page is mapped by vma pointed to by @private.
869 * Search forward from there, if not. N.B., this assumes that the
870 * list of pages handed to migrate_pages()--which is how we get here--
871 * is in virtual address order.
872 */
Christoph Lameter742755a2006-06-23 02:03:55 -0700873static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
Christoph Lameter95a402c2006-06-23 02:03:53 -0700874{
875 struct vm_area_struct *vma = (struct vm_area_struct *)private;
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -0800876 unsigned long uninitialized_var(address);
Christoph Lameter95a402c2006-06-23 02:03:53 -0700877
Lee Schermerhorn3ad33b22007-11-14 16:59:10 -0800878 while (vma) {
879 address = page_address_in_vma(page, vma);
880 if (address != -EFAULT)
881 break;
882 vma = vma->vm_next;
883 }
884
885 /*
886 * if !vma, alloc_page_vma() will use task or system default policy
887 */
888 return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
Christoph Lameter95a402c2006-06-23 02:03:53 -0700889}
Christoph Lameterb20a3502006-03-22 00:09:12 -0800890#else
891
892static void migrate_page_add(struct page *page, struct list_head *pagelist,
893 unsigned long flags)
894{
895}
896
897int do_migrate_pages(struct mm_struct *mm,
898 const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
899{
900 return -ENOSYS;
901}
Christoph Lameter95a402c2006-06-23 02:03:53 -0700902
Keith Owens69939742006-10-11 01:21:28 -0700903static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
Christoph Lameter95a402c2006-06-23 02:03:53 -0700904{
905 return NULL;
906}
Christoph Lameterb20a3502006-03-22 00:09:12 -0800907#endif
908
Adrian Bunkdbcb0f12007-10-16 01:26:26 -0700909static long do_mbind(unsigned long start, unsigned long len,
David Rientjes028fec42008-04-28 02:12:25 -0700910 unsigned short mode, unsigned short mode_flags,
911 nodemask_t *nmask, unsigned long flags)
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -0800912{
913 struct vm_area_struct *vma;
914 struct mm_struct *mm = current->mm;
915 struct mempolicy *new;
916 unsigned long end;
917 int err;
918 LIST_HEAD(pagelist);
919
David Rientjesa3b51e02008-04-28 02:12:23 -0700920 if (flags & ~(unsigned long)(MPOL_MF_STRICT |
921 MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -0800922 return -EINVAL;
Christoph Lameter74c00242006-03-14 19:50:21 -0800923 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -0800924 return -EPERM;
925
926 if (start & ~PAGE_MASK)
927 return -EINVAL;
928
929 if (mode == MPOL_DEFAULT)
930 flags &= ~MPOL_MF_STRICT;
931
932 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
933 end = start + len;
934
935 if (end < start)
936 return -EINVAL;
937 if (end == start)
938 return 0;
939
David Rientjes028fec42008-04-28 02:12:25 -0700940 new = mpol_new(mode, mode_flags, nmask);
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -0800941 if (IS_ERR(new))
942 return PTR_ERR(new);
943
944 /*
945 * If we are using the default policy then operation
946 * on discontinuous address spaces is okay after all
947 */
948 if (!new)
949 flags |= MPOL_MF_DISCONTIG_OK;
950
David Rientjes028fec42008-04-28 02:12:25 -0700951 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
952 start, start + len, mode, mode_flags,
953 nmask ? nodes_addr(*nmask)[0] : -1);
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -0800954
955 down_write(&mm->mmap_sem);
956 vma = check_range(mm, start, end, nmask,
957 flags | MPOL_MF_INVERT, &pagelist);
958
959 err = PTR_ERR(vma);
960 if (!IS_ERR(vma)) {
961 int nr_failed = 0;
962
963 err = mbind_range(vma, start, end, new);
Christoph Lameter7e2ab152006-02-01 03:05:40 -0800964
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -0800965 if (!list_empty(&pagelist))
Christoph Lameter95a402c2006-06-23 02:03:53 -0700966 nr_failed = migrate_pages(&pagelist, new_vma_page,
967 (unsigned long)vma);
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -0800968
969 if (!err && nr_failed && (flags & MPOL_MF_STRICT))
970 err = -EIO;
971 }
Christoph Lameterb20a3502006-03-22 00:09:12 -0800972
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -0800973 up_write(&mm->mmap_sem);
Lee Schermerhornf0be3d32008-04-28 02:13:08 -0700974 mpol_put(new);
Christoph Lameter6ce3c4c2006-01-08 01:01:04 -0800975 return err;
976}
977
Christoph Lameter39743882006-01-08 01:00:51 -0800978/*
Christoph Lameter8bccd852005-10-29 18:16:59 -0700979 * User space interface with variable sized bitmaps for nodelists.
980 */
981
982/* Copy a node mask from user space. */
Christoph Lameter39743882006-01-08 01:00:51 -0800983static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
Christoph Lameter8bccd852005-10-29 18:16:59 -0700984 unsigned long maxnode)
985{
986 unsigned long k;
987 unsigned long nlongs;
988 unsigned long endmask;
989
990 --maxnode;
991 nodes_clear(*nodes);
992 if (maxnode == 0 || !nmask)
993 return 0;
Andi Kleena9c930b2006-02-20 18:27:59 -0800994 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
Chris Wright636f13c2006-02-17 13:59:36 -0800995 return -EINVAL;
Christoph Lameter8bccd852005-10-29 18:16:59 -0700996
997 nlongs = BITS_TO_LONGS(maxnode);
998 if ((maxnode % BITS_PER_LONG) == 0)
999 endmask = ~0UL;
1000 else
1001 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1002
1003 /* When the user specified more nodes than supported just check
1004 if the non supported part is all zero. */
1005 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1006 if (nlongs > PAGE_SIZE/sizeof(long))
1007 return -EINVAL;
1008 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1009 unsigned long t;
1010 if (get_user(t, nmask + k))
1011 return -EFAULT;
1012 if (k == nlongs - 1) {
1013 if (t & endmask)
1014 return -EINVAL;
1015 } else if (t)
1016 return -EINVAL;
1017 }
1018 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1019 endmask = ~0UL;
1020 }
1021
1022 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1023 return -EFAULT;
1024 nodes_addr(*nodes)[nlongs-1] &= endmask;
1025 return 0;
1026}
1027
1028/* Copy a kernel node mask to user space */
1029static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1030 nodemask_t *nodes)
1031{
1032 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1033 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
1034
1035 if (copy > nbytes) {
1036 if (copy > PAGE_SIZE)
1037 return -EINVAL;
1038 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1039 return -EFAULT;
1040 copy = nbytes;
1041 }
1042 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1043}
1044
1045asmlinkage long sys_mbind(unsigned long start, unsigned long len,
1046 unsigned long mode,
1047 unsigned long __user *nmask, unsigned long maxnode,
1048 unsigned flags)
1049{
1050 nodemask_t nodes;
1051 int err;
David Rientjes028fec42008-04-28 02:12:25 -07001052 unsigned short mode_flags;
Christoph Lameter8bccd852005-10-29 18:16:59 -07001053
David Rientjes028fec42008-04-28 02:12:25 -07001054 mode_flags = mode & MPOL_MODE_FLAGS;
1055 mode &= ~MPOL_MODE_FLAGS;
David Rientjesa3b51e02008-04-28 02:12:23 -07001056 if (mode >= MPOL_MAX)
1057 return -EINVAL;
David Rientjes4c50bc02008-04-28 02:12:30 -07001058 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1059 (mode_flags & MPOL_F_RELATIVE_NODES))
1060 return -EINVAL;
Christoph Lameter8bccd852005-10-29 18:16:59 -07001061 err = get_nodes(&nodes, nmask, maxnode);
1062 if (err)
1063 return err;
David Rientjes028fec42008-04-28 02:12:25 -07001064 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
Christoph Lameter8bccd852005-10-29 18:16:59 -07001065}
1066
1067/* Set the process memory policy */
1068asmlinkage long sys_set_mempolicy(int mode, unsigned long __user *nmask,
1069 unsigned long maxnode)
1070{
1071 int err;
1072 nodemask_t nodes;
David Rientjes028fec42008-04-28 02:12:25 -07001073 unsigned short flags;
Christoph Lameter8bccd852005-10-29 18:16:59 -07001074
David Rientjes028fec42008-04-28 02:12:25 -07001075 flags = mode & MPOL_MODE_FLAGS;
1076 mode &= ~MPOL_MODE_FLAGS;
1077 if ((unsigned int)mode >= MPOL_MAX)
Christoph Lameter8bccd852005-10-29 18:16:59 -07001078 return -EINVAL;
David Rientjes4c50bc02008-04-28 02:12:30 -07001079 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1080 return -EINVAL;
Christoph Lameter8bccd852005-10-29 18:16:59 -07001081 err = get_nodes(&nodes, nmask, maxnode);
1082 if (err)
1083 return err;
David Rientjes028fec42008-04-28 02:12:25 -07001084 return do_set_mempolicy(mode, flags, &nodes);
Christoph Lameter8bccd852005-10-29 18:16:59 -07001085}
1086
Christoph Lameter39743882006-01-08 01:00:51 -08001087asmlinkage long sys_migrate_pages(pid_t pid, unsigned long maxnode,
1088 const unsigned long __user *old_nodes,
1089 const unsigned long __user *new_nodes)
1090{
1091 struct mm_struct *mm;
1092 struct task_struct *task;
1093 nodemask_t old;
1094 nodemask_t new;
1095 nodemask_t task_nodes;
1096 int err;
1097
1098 err = get_nodes(&old, old_nodes, maxnode);
1099 if (err)
1100 return err;
1101
1102 err = get_nodes(&new, new_nodes, maxnode);
1103 if (err)
1104 return err;
1105
1106 /* Find the mm_struct */
1107 read_lock(&tasklist_lock);
Pavel Emelyanov228ebcb2007-10-18 23:40:16 -07001108 task = pid ? find_task_by_vpid(pid) : current;
Christoph Lameter39743882006-01-08 01:00:51 -08001109 if (!task) {
1110 read_unlock(&tasklist_lock);
1111 return -ESRCH;
1112 }
1113 mm = get_task_mm(task);
1114 read_unlock(&tasklist_lock);
1115
1116 if (!mm)
1117 return -EINVAL;
1118
1119 /*
1120 * Check if this process has the right to modify the specified
1121 * process. The right exists if the process has administrative
Alexey Dobriyan7f927fc2006-03-28 01:56:53 -08001122 * capabilities, superuser privileges or the same
Christoph Lameter39743882006-01-08 01:00:51 -08001123 * userid as the target process.
1124 */
1125 if ((current->euid != task->suid) && (current->euid != task->uid) &&
1126 (current->uid != task->suid) && (current->uid != task->uid) &&
Christoph Lameter74c00242006-03-14 19:50:21 -08001127 !capable(CAP_SYS_NICE)) {
Christoph Lameter39743882006-01-08 01:00:51 -08001128 err = -EPERM;
1129 goto out;
1130 }
1131
1132 task_nodes = cpuset_mems_allowed(task);
1133 /* Is the user allowed to access the target nodes? */
Christoph Lameter74c00242006-03-14 19:50:21 -08001134 if (!nodes_subset(new, task_nodes) && !capable(CAP_SYS_NICE)) {
Christoph Lameter39743882006-01-08 01:00:51 -08001135 err = -EPERM;
1136 goto out;
1137 }
1138
Lee Schermerhorn37b07e42007-10-16 01:25:39 -07001139 if (!nodes_subset(new, node_states[N_HIGH_MEMORY])) {
Christoph Lameter3b42d282007-08-31 00:12:08 -07001140 err = -EINVAL;
1141 goto out;
1142 }
1143
David Quigley86c3a762006-06-23 02:04:02 -07001144 err = security_task_movememory(task);
1145 if (err)
1146 goto out;
1147
Christoph Lameter511030b2006-02-28 16:58:57 -08001148 err = do_migrate_pages(mm, &old, &new,
Christoph Lameter74c00242006-03-14 19:50:21 -08001149 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
Christoph Lameter39743882006-01-08 01:00:51 -08001150out:
1151 mmput(mm);
1152 return err;
1153}
1154
1155
Christoph Lameter8bccd852005-10-29 18:16:59 -07001156/* Retrieve NUMA policy */
1157asmlinkage long sys_get_mempolicy(int __user *policy,
1158 unsigned long __user *nmask,
1159 unsigned long maxnode,
1160 unsigned long addr, unsigned long flags)
1161{
Adrian Bunkdbcb0f12007-10-16 01:26:26 -07001162 int err;
1163 int uninitialized_var(pval);
Christoph Lameter8bccd852005-10-29 18:16:59 -07001164 nodemask_t nodes;
1165
1166 if (nmask != NULL && maxnode < MAX_NUMNODES)
1167 return -EINVAL;
1168
1169 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1170
1171 if (err)
1172 return err;
1173
1174 if (policy && put_user(pval, policy))
1175 return -EFAULT;
1176
1177 if (nmask)
1178 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1179
1180 return err;
1181}
1182
Linus Torvalds1da177e2005-04-16 15:20:36 -07001183#ifdef CONFIG_COMPAT
1184
1185asmlinkage long compat_sys_get_mempolicy(int __user *policy,
1186 compat_ulong_t __user *nmask,
1187 compat_ulong_t maxnode,
1188 compat_ulong_t addr, compat_ulong_t flags)
1189{
1190 long err;
1191 unsigned long __user *nm = NULL;
1192 unsigned long nr_bits, alloc_size;
1193 DECLARE_BITMAP(bm, MAX_NUMNODES);
1194
1195 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1196 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1197
1198 if (nmask)
1199 nm = compat_alloc_user_space(alloc_size);
1200
1201 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1202
1203 if (!err && nmask) {
1204 err = copy_from_user(bm, nm, alloc_size);
1205 /* ensure entire bitmap is zeroed */
1206 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1207 err |= compat_put_bitmap(nmask, bm, nr_bits);
1208 }
1209
1210 return err;
1211}
1212
1213asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask,
1214 compat_ulong_t maxnode)
1215{
1216 long err = 0;
1217 unsigned long __user *nm = NULL;
1218 unsigned long nr_bits, alloc_size;
1219 DECLARE_BITMAP(bm, MAX_NUMNODES);
1220
1221 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1222 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1223
1224 if (nmask) {
1225 err = compat_get_bitmap(bm, nmask, nr_bits);
1226 nm = compat_alloc_user_space(alloc_size);
1227 err |= copy_to_user(nm, bm, alloc_size);
1228 }
1229
1230 if (err)
1231 return -EFAULT;
1232
1233 return sys_set_mempolicy(mode, nm, nr_bits+1);
1234}
1235
1236asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len,
1237 compat_ulong_t mode, compat_ulong_t __user *nmask,
1238 compat_ulong_t maxnode, compat_ulong_t flags)
1239{
1240 long err = 0;
1241 unsigned long __user *nm = NULL;
1242 unsigned long nr_bits, alloc_size;
Andi Kleendfcd3c02005-10-29 18:15:48 -07001243 nodemask_t bm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001244
1245 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1246 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1247
1248 if (nmask) {
Andi Kleendfcd3c02005-10-29 18:15:48 -07001249 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001250 nm = compat_alloc_user_space(alloc_size);
Andi Kleendfcd3c02005-10-29 18:15:48 -07001251 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001252 }
1253
1254 if (err)
1255 return -EFAULT;
1256
1257 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1258}
1259
1260#endif
1261
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001262/*
1263 * get_vma_policy(@task, @vma, @addr)
1264 * @task - task for fallback if vma policy == default
1265 * @vma - virtual memory area whose policy is sought
1266 * @addr - address in @vma for shared policy lookup
1267 *
1268 * Returns effective policy for a VMA at specified address.
1269 * Falls back to @task or system default policy, as necessary.
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001270 * Current or other task's task mempolicy and non-shared vma policies
1271 * are protected by the task's mmap_sem, which must be held for read by
1272 * the caller.
1273 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1274 * count--added by the get_policy() vm_op, as appropriate--to protect against
1275 * freeing by another task. It is the caller's responsibility to free the
1276 * extra reference for shared policies.
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001277 */
Lee Schermerhornae4d8c12008-04-28 02:13:11 -07001278static struct mempolicy *get_vma_policy(struct task_struct *task,
Christoph Lameter48fce342006-01-08 01:01:03 -08001279 struct vm_area_struct *vma, unsigned long addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001280{
Christoph Lameter6e21c8f2005-09-03 15:54:45 -07001281 struct mempolicy *pol = task->mempolicy;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001282
1283 if (vma) {
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001284 if (vma->vm_ops && vma->vm_ops->get_policy) {
Lee Schermerhornae4d8c12008-04-28 02:13:11 -07001285 struct mempolicy *vpol = vma->vm_ops->get_policy(vma,
1286 addr);
1287 if (vpol)
1288 pol = vpol;
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001289 } else if (vma->vm_policy &&
Lee Schermerhorn45c47452008-04-28 02:13:12 -07001290 vma->vm_policy->mode != MPOL_DEFAULT)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001291 pol = vma->vm_policy;
1292 }
1293 if (!pol)
1294 pol = &default_policy;
1295 return pol;
1296}
1297
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001298/*
1299 * Return a nodemask representing a mempolicy for filtering nodes for
1300 * page allocation
1301 */
1302static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
Mel Gorman19770b32008-04-28 02:12:18 -07001303{
1304 /* Lower zones don't get a nodemask applied for MPOL_BIND */
Lee Schermerhorn45c47452008-04-28 02:13:12 -07001305 if (unlikely(policy->mode == MPOL_BIND) &&
Mel Gorman19770b32008-04-28 02:12:18 -07001306 gfp_zone(gfp) >= policy_zone &&
1307 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1308 return &policy->v.nodes;
1309
1310 return NULL;
1311}
1312
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001313/* Return a zonelist indicated by gfp for node representing a mempolicy */
1314static struct zonelist *policy_zonelist(gfp_t gfp, struct mempolicy *policy)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001315{
1316 int nd;
1317
Lee Schermerhorn45c47452008-04-28 02:13:12 -07001318 switch (policy->mode) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001319 case MPOL_PREFERRED:
1320 nd = policy->v.preferred_node;
1321 if (nd < 0)
1322 nd = numa_node_id();
1323 break;
1324 case MPOL_BIND:
Mel Gorman19770b32008-04-28 02:12:18 -07001325 /*
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001326 * Normally, MPOL_BIND allocations are node-local within the
1327 * allowed nodemask. However, if __GFP_THISNODE is set and the
1328 * current node is part of the mask, we use the zonelist for
1329 * the first node in the mask instead.
Mel Gorman19770b32008-04-28 02:12:18 -07001330 */
1331 nd = numa_node_id();
1332 if (unlikely(gfp & __GFP_THISNODE) &&
1333 unlikely(!node_isset(nd, policy->v.nodes)))
1334 nd = first_node(policy->v.nodes);
1335 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001336 case MPOL_INTERLEAVE: /* should not happen */
1337 case MPOL_DEFAULT:
1338 nd = numa_node_id();
1339 break;
1340 default:
1341 nd = 0;
1342 BUG();
1343 }
Mel Gorman0e884602008-04-28 02:12:14 -07001344 return node_zonelist(nd, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001345}
1346
1347/* Do dynamic interleaving for a process */
1348static unsigned interleave_nodes(struct mempolicy *policy)
1349{
1350 unsigned nid, next;
1351 struct task_struct *me = current;
1352
1353 nid = me->il_next;
Andi Kleendfcd3c02005-10-29 18:15:48 -07001354 next = next_node(nid, policy->v.nodes);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001355 if (next >= MAX_NUMNODES)
Andi Kleendfcd3c02005-10-29 18:15:48 -07001356 next = first_node(policy->v.nodes);
David Rientjesf5b087b2008-04-28 02:12:27 -07001357 if (next < MAX_NUMNODES)
1358 me->il_next = next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001359 return nid;
1360}
1361
Christoph Lameterdc85da12006-01-18 17:42:36 -08001362/*
1363 * Depending on the memory policy provide a node from which to allocate the
1364 * next slab entry.
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001365 * @policy must be protected by freeing by the caller. If @policy is
1366 * the current task's mempolicy, this protection is implicit, as only the
1367 * task can change it's policy. The system default policy requires no
1368 * such protection.
Christoph Lameterdc85da12006-01-18 17:42:36 -08001369 */
1370unsigned slab_node(struct mempolicy *policy)
1371{
Lee Schermerhorn45c47452008-04-28 02:13:12 -07001372 unsigned short pol = policy ? policy->mode : MPOL_DEFAULT;
Christoph Lameter765c4502006-09-27 01:50:08 -07001373
1374 switch (pol) {
Christoph Lameterdc85da12006-01-18 17:42:36 -08001375 case MPOL_INTERLEAVE:
1376 return interleave_nodes(policy);
1377
Mel Gormandd1a2392008-04-28 02:12:17 -07001378 case MPOL_BIND: {
Christoph Lameterdc85da12006-01-18 17:42:36 -08001379 /*
1380 * Follow bind policy behavior and start allocation at the
1381 * first node.
1382 */
Mel Gorman19770b32008-04-28 02:12:18 -07001383 struct zonelist *zonelist;
1384 struct zone *zone;
1385 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1386 zonelist = &NODE_DATA(numa_node_id())->node_zonelists[0];
1387 (void)first_zones_zonelist(zonelist, highest_zoneidx,
1388 &policy->v.nodes,
1389 &zone);
1390 return zone->node;
Mel Gormandd1a2392008-04-28 02:12:17 -07001391 }
Christoph Lameterdc85da12006-01-18 17:42:36 -08001392
1393 case MPOL_PREFERRED:
1394 if (policy->v.preferred_node >= 0)
1395 return policy->v.preferred_node;
1396 /* Fall through */
1397
1398 default:
1399 return numa_node_id();
1400 }
1401}
1402
Linus Torvalds1da177e2005-04-16 15:20:36 -07001403/* Do static interleaving for a VMA with known offset. */
1404static unsigned offset_il_node(struct mempolicy *pol,
1405 struct vm_area_struct *vma, unsigned long off)
1406{
Andi Kleendfcd3c02005-10-29 18:15:48 -07001407 unsigned nnodes = nodes_weight(pol->v.nodes);
David Rientjesf5b087b2008-04-28 02:12:27 -07001408 unsigned target;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001409 int c;
1410 int nid = -1;
1411
David Rientjesf5b087b2008-04-28 02:12:27 -07001412 if (!nnodes)
1413 return numa_node_id();
1414 target = (unsigned int)off % nnodes;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001415 c = 0;
1416 do {
Andi Kleendfcd3c02005-10-29 18:15:48 -07001417 nid = next_node(nid, pol->v.nodes);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001418 c++;
1419 } while (c <= target);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001420 return nid;
1421}
1422
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001423/* Determine a node number for interleave */
1424static inline unsigned interleave_nid(struct mempolicy *pol,
1425 struct vm_area_struct *vma, unsigned long addr, int shift)
1426{
1427 if (vma) {
1428 unsigned long off;
1429
Nishanth Aravamudan3b98b082006-08-31 21:27:53 -07001430 /*
1431 * for small pages, there is no difference between
1432 * shift and PAGE_SHIFT, so the bit-shift is safe.
1433 * for huge pages, since vm_pgoff is in units of small
1434 * pages, we need to shift off the always 0 bits to get
1435 * a useful offset.
1436 */
1437 BUG_ON(shift < PAGE_SHIFT);
1438 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001439 off += (addr - vma->vm_start) >> shift;
1440 return offset_il_node(pol, vma, off);
1441 } else
1442 return interleave_nodes(pol);
1443}
1444
Chen, Kenneth W00ac59a2006-02-03 21:51:14 +01001445#ifdef CONFIG_HUGETLBFS
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001446/*
1447 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1448 * @vma = virtual memory area whose policy is sought
1449 * @addr = address in @vma for shared policy lookup and interleave policy
1450 * @gfp_flags = for requested zone
Mel Gorman19770b32008-04-28 02:12:18 -07001451 * @mpol = pointer to mempolicy pointer for reference counted mempolicy
1452 * @nodemask = pointer to nodemask pointer for MPOL_BIND nodemask
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001453 *
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001454 * Returns a zonelist suitable for a huge page allocation and a pointer
1455 * to the struct mempolicy for conditional unref after allocation.
1456 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1457 * @nodemask for filtering the zonelist.
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001458 */
Mel Gorman396faf02007-07-17 04:03:13 -07001459struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
Mel Gorman19770b32008-04-28 02:12:18 -07001460 gfp_t gfp_flags, struct mempolicy **mpol,
1461 nodemask_t **nodemask)
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001462{
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001463 struct zonelist *zl;
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001464
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001465 *mpol = get_vma_policy(current, vma, addr);
Mel Gorman19770b32008-04-28 02:12:18 -07001466 *nodemask = NULL; /* assume !MPOL_BIND */
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001467
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001468 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1469 zl = node_zonelist(interleave_nid(*mpol, vma, addr,
1470 HPAGE_SHIFT), gfp_flags);
1471 } else {
1472 zl = policy_zonelist(gfp_flags, *mpol);
1473 if ((*mpol)->mode == MPOL_BIND)
1474 *nodemask = &(*mpol)->v.nodes;
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001475 }
1476 return zl;
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001477}
Chen, Kenneth W00ac59a2006-02-03 21:51:14 +01001478#endif
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001479
Linus Torvalds1da177e2005-04-16 15:20:36 -07001480/* Allocate a page in interleaved policy.
1481 Own path because it needs to do special accounting. */
Andi Kleen662f3a02005-10-29 18:15:49 -07001482static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1483 unsigned nid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001484{
1485 struct zonelist *zl;
1486 struct page *page;
1487
Mel Gorman0e884602008-04-28 02:12:14 -07001488 zl = node_zonelist(nid, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001489 page = __alloc_pages(gfp, order, zl);
Mel Gormandd1a2392008-04-28 02:12:17 -07001490 if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0]))
Christoph Lameterca889e62006-06-30 01:55:44 -07001491 inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001492 return page;
1493}
1494
1495/**
1496 * alloc_page_vma - Allocate a page for a VMA.
1497 *
1498 * @gfp:
1499 * %GFP_USER user allocation.
1500 * %GFP_KERNEL kernel allocations,
1501 * %GFP_HIGHMEM highmem/user allocations,
1502 * %GFP_FS allocation should not call back into a file system.
1503 * %GFP_ATOMIC don't sleep.
1504 *
1505 * @vma: Pointer to VMA or NULL if not available.
1506 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1507 *
1508 * This function allocates a page from the kernel page pool and applies
1509 * a NUMA policy associated with the VMA or the current process.
1510 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1511 * mm_struct of the VMA to prevent it from going away. Should be used for
1512 * all allocations for pages that will be mapped into
1513 * user space. Returns NULL when no page can be allocated.
1514 *
1515 * Should be called with the mm_sem of the vma hold.
1516 */
1517struct page *
Al Virodd0fc662005-10-07 07:46:04 +01001518alloc_page_vma(gfp_t gfp, struct vm_area_struct *vma, unsigned long addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001519{
Christoph Lameter6e21c8f2005-09-03 15:54:45 -07001520 struct mempolicy *pol = get_vma_policy(current, vma, addr);
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001521 struct zonelist *zl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001522
Paul Jacksoncf2a473c2006-01-08 01:01:54 -08001523 cpuset_update_task_memory_state();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001524
Lee Schermerhorn45c47452008-04-28 02:13:12 -07001525 if (unlikely(pol->mode == MPOL_INTERLEAVE)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001526 unsigned nid;
Christoph Lameter5da7ca82006-01-06 00:10:46 -08001527
1528 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT);
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001529 mpol_cond_put(pol);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001530 return alloc_page_interleave(gfp, 0, nid);
1531 }
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001532 zl = policy_zonelist(gfp, pol);
1533 if (unlikely(mpol_needs_cond_ref(pol))) {
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001534 /*
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001535 * slow path: ref counted shared policy
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001536 */
Mel Gorman19770b32008-04-28 02:12:18 -07001537 struct page *page = __alloc_pages_nodemask(gfp, 0,
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001538 zl, policy_nodemask(gfp, pol));
Lee Schermerhornf0be3d32008-04-28 02:13:08 -07001539 __mpol_put(pol);
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001540 return page;
1541 }
1542 /*
1543 * fast path: default or task policy
1544 */
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001545 return __alloc_pages_nodemask(gfp, 0, zl, policy_nodemask(gfp, pol));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001546}
1547
1548/**
1549 * alloc_pages_current - Allocate pages.
1550 *
1551 * @gfp:
1552 * %GFP_USER user allocation,
1553 * %GFP_KERNEL kernel allocation,
1554 * %GFP_HIGHMEM highmem allocation,
1555 * %GFP_FS don't call back into a file system.
1556 * %GFP_ATOMIC don't sleep.
1557 * @order: Power of two of allocation size in pages. 0 is a single page.
1558 *
1559 * Allocate a page from the kernel page pool. When not in
1560 * interrupt context and apply the current process NUMA policy.
1561 * Returns NULL when no page can be allocated.
1562 *
Paul Jacksoncf2a473c2006-01-08 01:01:54 -08001563 * Don't call cpuset_update_task_memory_state() unless
Linus Torvalds1da177e2005-04-16 15:20:36 -07001564 * 1) it's ok to take cpuset_sem (can WAIT), and
1565 * 2) allocating for current task (not interrupt).
1566 */
Al Virodd0fc662005-10-07 07:46:04 +01001567struct page *alloc_pages_current(gfp_t gfp, unsigned order)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001568{
1569 struct mempolicy *pol = current->mempolicy;
1570
1571 if ((gfp & __GFP_WAIT) && !in_interrupt())
Paul Jacksoncf2a473c2006-01-08 01:01:54 -08001572 cpuset_update_task_memory_state();
Christoph Lameter9b819d22006-09-25 23:31:40 -07001573 if (!pol || in_interrupt() || (gfp & __GFP_THISNODE))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001574 pol = &default_policy;
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001575
1576 /*
1577 * No reference counting needed for current->mempolicy
1578 * nor system default_policy
1579 */
Lee Schermerhorn45c47452008-04-28 02:13:12 -07001580 if (pol->mode == MPOL_INTERLEAVE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001581 return alloc_page_interleave(gfp, order, interleave_nodes(pol));
Mel Gorman19770b32008-04-28 02:12:18 -07001582 return __alloc_pages_nodemask(gfp, order,
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001583 policy_zonelist(gfp, pol), policy_nodemask(gfp, pol));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001584}
1585EXPORT_SYMBOL(alloc_pages_current);
1586
Paul Jackson42253992006-01-08 01:01:59 -08001587/*
Lee Schermerhorn846a16b2008-04-28 02:13:09 -07001588 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
Paul Jackson42253992006-01-08 01:01:59 -08001589 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
1590 * with the mems_allowed returned by cpuset_mems_allowed(). This
1591 * keeps mempolicies cpuset relative after its cpuset moves. See
1592 * further kernel/cpuset.c update_nodemask().
1593 */
Paul Jackson42253992006-01-08 01:01:59 -08001594
Lee Schermerhorn846a16b2008-04-28 02:13:09 -07001595/* Slow path of a mempolicy duplicate */
1596struct mempolicy *__mpol_dup(struct mempolicy *old)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001597{
1598 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
1599
1600 if (!new)
1601 return ERR_PTR(-ENOMEM);
Paul Jackson42253992006-01-08 01:01:59 -08001602 if (current_cpuset_is_being_rebound()) {
1603 nodemask_t mems = cpuset_mems_allowed(current);
1604 mpol_rebind_policy(old, &mems);
1605 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001606 *new = *old;
1607 atomic_set(&new->refcnt, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001608 return new;
1609}
1610
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07001611/*
1612 * If *frompol needs [has] an extra ref, copy *frompol to *tompol ,
1613 * eliminate the * MPOL_F_* flags that require conditional ref and
1614 * [NOTE!!!] drop the extra ref. Not safe to reference *frompol directly
1615 * after return. Use the returned value.
1616 *
1617 * Allows use of a mempolicy for, e.g., multiple allocations with a single
1618 * policy lookup, even if the policy needs/has extra ref on lookup.
1619 * shmem_readahead needs this.
1620 */
1621struct mempolicy *__mpol_cond_copy(struct mempolicy *tompol,
1622 struct mempolicy *frompol)
1623{
1624 if (!mpol_needs_cond_ref(frompol))
1625 return frompol;
1626
1627 *tompol = *frompol;
1628 tompol->flags &= ~MPOL_F_SHARED; /* copy doesn't need unref */
1629 __mpol_put(frompol);
1630 return tompol;
1631}
1632
David Rientjesf5b087b2008-04-28 02:12:27 -07001633static int mpol_match_intent(const struct mempolicy *a,
1634 const struct mempolicy *b)
1635{
1636 if (a->flags != b->flags)
1637 return 0;
1638 if (!mpol_store_user_nodemask(a))
1639 return 1;
1640 return nodes_equal(a->w.user_nodemask, b->w.user_nodemask);
1641}
1642
Linus Torvalds1da177e2005-04-16 15:20:36 -07001643/* Slow path of a mempolicy comparison */
1644int __mpol_equal(struct mempolicy *a, struct mempolicy *b)
1645{
1646 if (!a || !b)
1647 return 0;
Lee Schermerhorn45c47452008-04-28 02:13:12 -07001648 if (a->mode != b->mode)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001649 return 0;
Lee Schermerhorn45c47452008-04-28 02:13:12 -07001650 if (a->mode != MPOL_DEFAULT && !mpol_match_intent(a, b))
David Rientjesf5b087b2008-04-28 02:12:27 -07001651 return 0;
Lee Schermerhorn45c47452008-04-28 02:13:12 -07001652 switch (a->mode) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001653 case MPOL_DEFAULT:
1654 return 1;
Mel Gorman19770b32008-04-28 02:12:18 -07001655 case MPOL_BIND:
1656 /* Fall through */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001657 case MPOL_INTERLEAVE:
Andi Kleendfcd3c02005-10-29 18:15:48 -07001658 return nodes_equal(a->v.nodes, b->v.nodes);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001659 case MPOL_PREFERRED:
1660 return a->v.preferred_node == b->v.preferred_node;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001661 default:
1662 BUG();
1663 return 0;
1664 }
1665}
1666
Linus Torvalds1da177e2005-04-16 15:20:36 -07001667/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001668 * Shared memory backing store policy support.
1669 *
1670 * Remember policies even when nobody has shared memory mapped.
1671 * The policies are kept in Red-Black tree linked from the inode.
1672 * They are protected by the sp->lock spinlock, which should be held
1673 * for any accesses to the tree.
1674 */
1675
1676/* lookup first element intersecting start-end */
1677/* Caller holds sp->lock */
1678static struct sp_node *
1679sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
1680{
1681 struct rb_node *n = sp->root.rb_node;
1682
1683 while (n) {
1684 struct sp_node *p = rb_entry(n, struct sp_node, nd);
1685
1686 if (start >= p->end)
1687 n = n->rb_right;
1688 else if (end <= p->start)
1689 n = n->rb_left;
1690 else
1691 break;
1692 }
1693 if (!n)
1694 return NULL;
1695 for (;;) {
1696 struct sp_node *w = NULL;
1697 struct rb_node *prev = rb_prev(n);
1698 if (!prev)
1699 break;
1700 w = rb_entry(prev, struct sp_node, nd);
1701 if (w->end <= start)
1702 break;
1703 n = prev;
1704 }
1705 return rb_entry(n, struct sp_node, nd);
1706}
1707
1708/* Insert a new shared policy into the list. */
1709/* Caller holds sp->lock */
1710static void sp_insert(struct shared_policy *sp, struct sp_node *new)
1711{
1712 struct rb_node **p = &sp->root.rb_node;
1713 struct rb_node *parent = NULL;
1714 struct sp_node *nd;
1715
1716 while (*p) {
1717 parent = *p;
1718 nd = rb_entry(parent, struct sp_node, nd);
1719 if (new->start < nd->start)
1720 p = &(*p)->rb_left;
1721 else if (new->end > nd->end)
1722 p = &(*p)->rb_right;
1723 else
1724 BUG();
1725 }
1726 rb_link_node(&new->nd, parent, p);
1727 rb_insert_color(&new->nd, &sp->root);
Paul Mundt140d5a42007-07-15 23:38:16 -07001728 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
Lee Schermerhorn45c47452008-04-28 02:13:12 -07001729 new->policy ? new->policy->mode : 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001730}
1731
1732/* Find shared policy intersecting idx */
1733struct mempolicy *
1734mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
1735{
1736 struct mempolicy *pol = NULL;
1737 struct sp_node *sn;
1738
1739 if (!sp->root.rb_node)
1740 return NULL;
1741 spin_lock(&sp->lock);
1742 sn = sp_lookup(sp, idx, idx+1);
1743 if (sn) {
1744 mpol_get(sn->policy);
1745 pol = sn->policy;
1746 }
1747 spin_unlock(&sp->lock);
1748 return pol;
1749}
1750
1751static void sp_delete(struct shared_policy *sp, struct sp_node *n)
1752{
Paul Mundt140d5a42007-07-15 23:38:16 -07001753 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001754 rb_erase(&n->nd, &sp->root);
Lee Schermerhornf0be3d32008-04-28 02:13:08 -07001755 mpol_put(n->policy);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001756 kmem_cache_free(sn_cache, n);
1757}
1758
Adrian Bunkdbcb0f12007-10-16 01:26:26 -07001759static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
1760 struct mempolicy *pol)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001761{
1762 struct sp_node *n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
1763
1764 if (!n)
1765 return NULL;
1766 n->start = start;
1767 n->end = end;
1768 mpol_get(pol);
Lee Schermerhornaab0b102008-04-28 02:13:13 -07001769 pol->flags |= MPOL_F_SHARED; /* for unref */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001770 n->policy = pol;
1771 return n;
1772}
1773
1774/* Replace a policy range. */
1775static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
1776 unsigned long end, struct sp_node *new)
1777{
1778 struct sp_node *n, *new2 = NULL;
1779
1780restart:
1781 spin_lock(&sp->lock);
1782 n = sp_lookup(sp, start, end);
1783 /* Take care of old policies in the same range. */
1784 while (n && n->start < end) {
1785 struct rb_node *next = rb_next(&n->nd);
1786 if (n->start >= start) {
1787 if (n->end <= end)
1788 sp_delete(sp, n);
1789 else
1790 n->start = end;
1791 } else {
1792 /* Old policy spanning whole new range. */
1793 if (n->end > end) {
1794 if (!new2) {
1795 spin_unlock(&sp->lock);
1796 new2 = sp_alloc(end, n->end, n->policy);
1797 if (!new2)
1798 return -ENOMEM;
1799 goto restart;
1800 }
1801 n->end = start;
1802 sp_insert(sp, new2);
1803 new2 = NULL;
1804 break;
1805 } else
1806 n->end = start;
1807 }
1808 if (!next)
1809 break;
1810 n = rb_entry(next, struct sp_node, nd);
1811 }
1812 if (new)
1813 sp_insert(sp, new);
1814 spin_unlock(&sp->lock);
1815 if (new2) {
Lee Schermerhornf0be3d32008-04-28 02:13:08 -07001816 mpol_put(new2->policy);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001817 kmem_cache_free(sn_cache, new2);
1818 }
1819 return 0;
1820}
1821
David Rientjesa3b51e02008-04-28 02:12:23 -07001822void mpol_shared_policy_init(struct shared_policy *info, unsigned short policy,
David Rientjes028fec42008-04-28 02:12:25 -07001823 unsigned short flags, nodemask_t *policy_nodes)
Robin Holt7339ff82006-01-14 13:20:48 -08001824{
1825 info->root = RB_ROOT;
1826 spin_lock_init(&info->lock);
1827
1828 if (policy != MPOL_DEFAULT) {
1829 struct mempolicy *newpol;
1830
1831 /* Falls back to MPOL_DEFAULT on any error */
David Rientjes028fec42008-04-28 02:12:25 -07001832 newpol = mpol_new(policy, flags, policy_nodes);
Robin Holt7339ff82006-01-14 13:20:48 -08001833 if (!IS_ERR(newpol)) {
1834 /* Create pseudo-vma that contains just the policy */
1835 struct vm_area_struct pvma;
1836
1837 memset(&pvma, 0, sizeof(struct vm_area_struct));
1838 /* Policy covers entire file */
1839 pvma.vm_end = TASK_SIZE;
1840 mpol_set_shared_policy(info, &pvma, newpol);
Lee Schermerhornf0be3d32008-04-28 02:13:08 -07001841 mpol_put(newpol);
Robin Holt7339ff82006-01-14 13:20:48 -08001842 }
1843 }
1844}
1845
Linus Torvalds1da177e2005-04-16 15:20:36 -07001846int mpol_set_shared_policy(struct shared_policy *info,
1847 struct vm_area_struct *vma, struct mempolicy *npol)
1848{
1849 int err;
1850 struct sp_node *new = NULL;
1851 unsigned long sz = vma_pages(vma);
1852
David Rientjes028fec42008-04-28 02:12:25 -07001853 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -07001854 vma->vm_pgoff,
Lee Schermerhorn45c47452008-04-28 02:13:12 -07001855 sz, npol ? npol->mode : -1,
David Rientjes028fec42008-04-28 02:12:25 -07001856 npol ? npol->flags : -1,
Paul Mundt140d5a42007-07-15 23:38:16 -07001857 npol ? nodes_addr(npol->v.nodes)[0] : -1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001858
1859 if (npol) {
1860 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
1861 if (!new)
1862 return -ENOMEM;
1863 }
1864 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
1865 if (err && new)
1866 kmem_cache_free(sn_cache, new);
1867 return err;
1868}
1869
1870/* Free a backing policy store on inode delete. */
1871void mpol_free_shared_policy(struct shared_policy *p)
1872{
1873 struct sp_node *n;
1874 struct rb_node *next;
1875
1876 if (!p->root.rb_node)
1877 return;
1878 spin_lock(&p->lock);
1879 next = rb_first(&p->root);
1880 while (next) {
1881 n = rb_entry(next, struct sp_node, nd);
1882 next = rb_next(&n->nd);
Andi Kleen90c50292005-07-27 11:43:50 -07001883 rb_erase(&n->nd, &p->root);
Lee Schermerhornf0be3d32008-04-28 02:13:08 -07001884 mpol_put(n->policy);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001885 kmem_cache_free(sn_cache, n);
1886 }
1887 spin_unlock(&p->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001888}
1889
1890/* assumes fs == KERNEL_DS */
1891void __init numa_policy_init(void)
1892{
Paul Mundtb71636e2007-07-15 23:38:15 -07001893 nodemask_t interleave_nodes;
1894 unsigned long largest = 0;
1895 int nid, prefer = 0;
1896
Linus Torvalds1da177e2005-04-16 15:20:36 -07001897 policy_cache = kmem_cache_create("numa_policy",
1898 sizeof(struct mempolicy),
Paul Mundt20c2df82007-07-20 10:11:58 +09001899 0, SLAB_PANIC, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001900
1901 sn_cache = kmem_cache_create("shared_policy_node",
1902 sizeof(struct sp_node),
Paul Mundt20c2df82007-07-20 10:11:58 +09001903 0, SLAB_PANIC, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001904
Paul Mundtb71636e2007-07-15 23:38:15 -07001905 /*
1906 * Set interleaving policy for system init. Interleaving is only
1907 * enabled across suitably sized nodes (default is >= 16MB), or
1908 * fall back to the largest node if they're all smaller.
1909 */
1910 nodes_clear(interleave_nodes);
Christoph Lameter56bbd652007-10-16 01:25:35 -07001911 for_each_node_state(nid, N_HIGH_MEMORY) {
Paul Mundtb71636e2007-07-15 23:38:15 -07001912 unsigned long total_pages = node_present_pages(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001913
Paul Mundtb71636e2007-07-15 23:38:15 -07001914 /* Preserve the largest node */
1915 if (largest < total_pages) {
1916 largest = total_pages;
1917 prefer = nid;
1918 }
1919
1920 /* Interleave this node? */
1921 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
1922 node_set(nid, interleave_nodes);
1923 }
1924
1925 /* All too small, use the largest */
1926 if (unlikely(nodes_empty(interleave_nodes)))
1927 node_set(prefer, interleave_nodes);
1928
David Rientjes028fec42008-04-28 02:12:25 -07001929 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001930 printk("numa_policy_init: interleaving failed\n");
1931}
1932
Christoph Lameter8bccd852005-10-29 18:16:59 -07001933/* Reset policy of current process to default */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001934void numa_default_policy(void)
1935{
David Rientjes028fec42008-04-28 02:12:25 -07001936 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001937}
Paul Jackson68860ec2005-10-30 15:02:36 -08001938
Paul Jackson42253992006-01-08 01:01:59 -08001939/*
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08001940 * Display pages allocated per node and memory policy via /proc.
1941 */
Helge Deller15ad7cd2006-12-06 20:40:36 -08001942static const char * const policy_types[] =
1943 { "default", "prefer", "bind", "interleave" };
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08001944
1945/*
1946 * Convert a mempolicy into a string.
1947 * Returns the number of characters in buffer (if positive)
1948 * or an error (negative)
1949 */
1950static inline int mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
1951{
1952 char *p = buffer;
1953 int l;
1954 nodemask_t nodes;
Lee Schermerhorn45c47452008-04-28 02:13:12 -07001955 unsigned short mode = pol ? pol->mode : MPOL_DEFAULT;
David Rientjesf5b087b2008-04-28 02:12:27 -07001956 unsigned short flags = pol ? pol->flags : 0;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08001957
1958 switch (mode) {
1959 case MPOL_DEFAULT:
1960 nodes_clear(nodes);
1961 break;
1962
1963 case MPOL_PREFERRED:
1964 nodes_clear(nodes);
1965 node_set(pol->v.preferred_node, nodes);
1966 break;
1967
1968 case MPOL_BIND:
Mel Gorman19770b32008-04-28 02:12:18 -07001969 /* Fall through */
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08001970 case MPOL_INTERLEAVE:
1971 nodes = pol->v.nodes;
1972 break;
1973
1974 default:
1975 BUG();
1976 return -EFAULT;
1977 }
1978
1979 l = strlen(policy_types[mode]);
1980 if (buffer + maxlen < p + l + 1)
1981 return -ENOSPC;
1982
1983 strcpy(p, policy_types[mode]);
1984 p += l;
1985
David Rientjesf5b087b2008-04-28 02:12:27 -07001986 if (flags) {
1987 int need_bar = 0;
1988
1989 if (buffer + maxlen < p + 2)
1990 return -ENOSPC;
1991 *p++ = '=';
1992
1993 if (flags & MPOL_F_STATIC_NODES)
1994 p += sprintf(p, "%sstatic", need_bar++ ? "|" : "");
David Rientjes4c50bc02008-04-28 02:12:30 -07001995 if (flags & MPOL_F_RELATIVE_NODES)
1996 p += sprintf(p, "%srelative", need_bar++ ? "|" : "");
David Rientjesf5b087b2008-04-28 02:12:27 -07001997 }
1998
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08001999 if (!nodes_empty(nodes)) {
2000 if (buffer + maxlen < p + 2)
2001 return -ENOSPC;
2002 *p++ = '=';
2003 p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);
2004 }
2005 return p - buffer;
2006}
2007
2008struct numa_maps {
2009 unsigned long pages;
2010 unsigned long anon;
Christoph Lameter397874d2006-03-06 15:42:53 -08002011 unsigned long active;
2012 unsigned long writeback;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002013 unsigned long mapcount_max;
Christoph Lameter397874d2006-03-06 15:42:53 -08002014 unsigned long dirty;
2015 unsigned long swapcache;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002016 unsigned long node[MAX_NUMNODES];
2017};
2018
Christoph Lameter397874d2006-03-06 15:42:53 -08002019static void gather_stats(struct page *page, void *private, int pte_dirty)
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002020{
2021 struct numa_maps *md = private;
2022 int count = page_mapcount(page);
2023
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002024 md->pages++;
Christoph Lameter397874d2006-03-06 15:42:53 -08002025 if (pte_dirty || PageDirty(page))
2026 md->dirty++;
2027
2028 if (PageSwapCache(page))
2029 md->swapcache++;
2030
2031 if (PageActive(page))
2032 md->active++;
2033
2034 if (PageWriteback(page))
2035 md->writeback++;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002036
2037 if (PageAnon(page))
2038 md->anon++;
2039
Christoph Lameter397874d2006-03-06 15:42:53 -08002040 if (count > md->mapcount_max)
2041 md->mapcount_max = count;
2042
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002043 md->node[page_to_nid(page)]++;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002044}
2045
Andrew Morton7f709ed2006-03-07 21:55:22 -08002046#ifdef CONFIG_HUGETLB_PAGE
Christoph Lameter397874d2006-03-06 15:42:53 -08002047static void check_huge_range(struct vm_area_struct *vma,
2048 unsigned long start, unsigned long end,
2049 struct numa_maps *md)
2050{
2051 unsigned long addr;
2052 struct page *page;
2053
2054 for (addr = start; addr < end; addr += HPAGE_SIZE) {
2055 pte_t *ptep = huge_pte_offset(vma->vm_mm, addr & HPAGE_MASK);
2056 pte_t pte;
2057
2058 if (!ptep)
2059 continue;
2060
2061 pte = *ptep;
2062 if (pte_none(pte))
2063 continue;
2064
2065 page = pte_page(pte);
2066 if (!page)
2067 continue;
2068
2069 gather_stats(page, md, pte_dirty(*ptep));
2070 }
2071}
Andrew Morton7f709ed2006-03-07 21:55:22 -08002072#else
2073static inline void check_huge_range(struct vm_area_struct *vma,
2074 unsigned long start, unsigned long end,
2075 struct numa_maps *md)
2076{
2077}
2078#endif
Christoph Lameter397874d2006-03-06 15:42:53 -08002079
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002080int show_numa_map(struct seq_file *m, void *v)
2081{
Eric W. Biederman99f89552006-06-26 00:25:55 -07002082 struct proc_maps_private *priv = m->private;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002083 struct vm_area_struct *vma = v;
2084 struct numa_maps *md;
Christoph Lameter397874d2006-03-06 15:42:53 -08002085 struct file *file = vma->vm_file;
2086 struct mm_struct *mm = vma->vm_mm;
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07002087 struct mempolicy *pol;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002088 int n;
2089 char buffer[50];
2090
Christoph Lameter397874d2006-03-06 15:42:53 -08002091 if (!mm)
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002092 return 0;
2093
2094 md = kzalloc(sizeof(struct numa_maps), GFP_KERNEL);
2095 if (!md)
2096 return 0;
2097
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07002098 pol = get_vma_policy(priv->task, vma, vma->vm_start);
2099 mpol_to_str(buffer, sizeof(buffer), pol);
Lee Schermerhorn52cd3b02008-04-28 02:13:16 -07002100 mpol_cond_put(pol);
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002101
Christoph Lameter397874d2006-03-06 15:42:53 -08002102 seq_printf(m, "%08lx %s", vma->vm_start, buffer);
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002103
Christoph Lameter397874d2006-03-06 15:42:53 -08002104 if (file) {
2105 seq_printf(m, " file=");
Jan Blunckc32c2f62008-02-14 19:38:43 -08002106 seq_path(m, &file->f_path, "\n\t= ");
Christoph Lameter397874d2006-03-06 15:42:53 -08002107 } else if (vma->vm_start <= mm->brk && vma->vm_end >= mm->start_brk) {
2108 seq_printf(m, " heap");
2109 } else if (vma->vm_start <= mm->start_stack &&
2110 vma->vm_end >= mm->start_stack) {
2111 seq_printf(m, " stack");
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002112 }
Christoph Lameter397874d2006-03-06 15:42:53 -08002113
2114 if (is_vm_hugetlb_page(vma)) {
2115 check_huge_range(vma, vma->vm_start, vma->vm_end, md);
2116 seq_printf(m, " huge");
2117 } else {
2118 check_pgd_range(vma, vma->vm_start, vma->vm_end,
Christoph Lameter56bbd652007-10-16 01:25:35 -07002119 &node_states[N_HIGH_MEMORY], MPOL_MF_STATS, md);
Christoph Lameter397874d2006-03-06 15:42:53 -08002120 }
2121
2122 if (!md->pages)
2123 goto out;
2124
2125 if (md->anon)
2126 seq_printf(m," anon=%lu",md->anon);
2127
2128 if (md->dirty)
2129 seq_printf(m," dirty=%lu",md->dirty);
2130
2131 if (md->pages != md->anon && md->pages != md->dirty)
2132 seq_printf(m, " mapped=%lu", md->pages);
2133
2134 if (md->mapcount_max > 1)
2135 seq_printf(m, " mapmax=%lu", md->mapcount_max);
2136
2137 if (md->swapcache)
2138 seq_printf(m," swapcache=%lu", md->swapcache);
2139
2140 if (md->active < md->pages && !is_vm_hugetlb_page(vma))
2141 seq_printf(m," active=%lu", md->active);
2142
2143 if (md->writeback)
2144 seq_printf(m," writeback=%lu", md->writeback);
2145
Christoph Lameter56bbd652007-10-16 01:25:35 -07002146 for_each_node_state(n, N_HIGH_MEMORY)
Christoph Lameter397874d2006-03-06 15:42:53 -08002147 if (md->node[n])
2148 seq_printf(m, " N%d=%lu", n, md->node[n]);
2149out:
2150 seq_putc(m, '\n');
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002151 kfree(md);
2152
2153 if (m->count < m->size)
Eric W. Biederman99f89552006-06-26 00:25:55 -07002154 m->version = (vma != priv->tail_vma) ? vma->vm_start : 0;
Christoph Lameter1a75a6c2006-01-08 01:01:02 -08002155 return 0;
2156}