]> git.proxmox.com Git - mirror_ubuntu-bionic-kernel.git/blame - mm/mempolicy.c
vfs: pull btrfs clone API to vfs layer
[mirror_ubuntu-bionic-kernel.git] / mm / mempolicy.c
CommitLineData
1da177e4
LT
1/*
2 * Simple NUMA memory policy for the Linux kernel.
3 *
4 * Copyright 2003,2004 Andi Kleen, SuSE Labs.
8bccd85f 5 * (C) Copyright 2005 Christoph Lameter, Silicon Graphics, Inc.
1da177e4
LT
6 * 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.
8bccd85f 21 *
1da177e4
LT
22 * bind Only allocate memory on a specific set of nodes,
23 * no fallback.
8bccd85f
CL
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
27 *
1da177e4 28 * preferred Try a specific node first before normal fallback.
00ef2d2f 29 * As a special case NUMA_NO_NODE here means do the allocation
1da177e4
LT
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.
8bccd85f 33 *
1da177e4
LT
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.
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.
1da177e4
LT
66*/
67
b1de0d13
MH
68#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
69
1da177e4
LT
70#include <linux/mempolicy.h>
71#include <linux/mm.h>
72#include <linux/highmem.h>
73#include <linux/hugetlb.h>
74#include <linux/kernel.h>
75#include <linux/sched.h>
1da177e4
LT
76#include <linux/nodemask.h>
77#include <linux/cpuset.h>
1da177e4
LT
78#include <linux/slab.h>
79#include <linux/string.h>
b95f1b31 80#include <linux/export.h>
b488893a 81#include <linux/nsproxy.h>
1da177e4
LT
82#include <linux/interrupt.h>
83#include <linux/init.h>
84#include <linux/compat.h>
dc9aa5b9 85#include <linux/swap.h>
1a75a6c8
CL
86#include <linux/seq_file.h>
87#include <linux/proc_fs.h>
b20a3503 88#include <linux/migrate.h>
62b61f61 89#include <linux/ksm.h>
95a402c3 90#include <linux/rmap.h>
86c3a764 91#include <linux/security.h>
dbcb0f19 92#include <linux/syscalls.h>
095f1fc4 93#include <linux/ctype.h>
6d9c285a 94#include <linux/mm_inline.h>
b24f53a0 95#include <linux/mmu_notifier.h>
b1de0d13 96#include <linux/printk.h>
dc9aa5b9 97
1da177e4
LT
98#include <asm/tlbflush.h>
99#include <asm/uaccess.h>
778d3b0f 100#include <linux/random.h>
1da177e4 101
62695a84
NP
102#include "internal.h"
103
38e35860 104/* Internal flags */
dc9aa5b9 105#define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
38e35860 106#define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
dc9aa5b9 107
fcc234f8
PE
108static struct kmem_cache *policy_cache;
109static struct kmem_cache *sn_cache;
1da177e4 110
1da177e4
LT
111/* Highest zone. An specific allocation for a zone below that is not
112 policied. */
6267276f 113enum zone_type policy_zone = 0;
1da177e4 114
bea904d5
LS
115/*
116 * run-time system-wide default policy => local allocation
117 */
e754d79d 118static struct mempolicy default_policy = {
1da177e4 119 .refcnt = ATOMIC_INIT(1), /* never free it */
bea904d5 120 .mode = MPOL_PREFERRED,
fc36b8d3 121 .flags = MPOL_F_LOCAL,
1da177e4
LT
122};
123
5606e387
MG
124static struct mempolicy preferred_node_policy[MAX_NUMNODES];
125
74d2c3a0 126struct mempolicy *get_task_policy(struct task_struct *p)
5606e387
MG
127{
128 struct mempolicy *pol = p->mempolicy;
f15ca78e 129 int node;
5606e387 130
f15ca78e
ON
131 if (pol)
132 return pol;
5606e387 133
f15ca78e
ON
134 node = numa_node_id();
135 if (node != NUMA_NO_NODE) {
136 pol = &preferred_node_policy[node];
137 /* preferred_node_policy is not initialised early in boot */
138 if (pol->mode)
139 return pol;
5606e387
MG
140 }
141
f15ca78e 142 return &default_policy;
5606e387
MG
143}
144
37012946
DR
145static const struct mempolicy_operations {
146 int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
708c1bbc
MX
147 /*
148 * If read-side task has no lock to protect task->mempolicy, write-side
149 * task will rebind the task->mempolicy by two step. The first step is
150 * setting all the newly nodes, and the second step is cleaning all the
151 * disallowed nodes. In this way, we can avoid finding no node to alloc
152 * page.
153 * If we have a lock to protect task->mempolicy in read-side, we do
154 * rebind directly.
155 *
156 * step:
157 * MPOL_REBIND_ONCE - do rebind work at once
158 * MPOL_REBIND_STEP1 - set all the newly nodes
159 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
160 */
161 void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes,
162 enum mpol_rebind_step step);
37012946
DR
163} mpol_ops[MPOL_MAX];
164
f5b087b5
DR
165static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
166{
6d556294 167 return pol->flags & MPOL_MODE_FLAGS;
4c50bc01
DR
168}
169
170static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
171 const nodemask_t *rel)
172{
173 nodemask_t tmp;
174 nodes_fold(tmp, *orig, nodes_weight(*rel));
175 nodes_onto(*ret, tmp, *rel);
f5b087b5
DR
176}
177
37012946
DR
178static int mpol_new_interleave(struct mempolicy *pol, const nodemask_t *nodes)
179{
180 if (nodes_empty(*nodes))
181 return -EINVAL;
182 pol->v.nodes = *nodes;
183 return 0;
184}
185
186static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
187{
188 if (!nodes)
fc36b8d3 189 pol->flags |= MPOL_F_LOCAL; /* local allocation */
37012946
DR
190 else if (nodes_empty(*nodes))
191 return -EINVAL; /* no allowed nodes */
192 else
193 pol->v.preferred_node = first_node(*nodes);
194 return 0;
195}
196
197static int mpol_new_bind(struct mempolicy *pol, const nodemask_t *nodes)
198{
859f7ef1 199 if (nodes_empty(*nodes))
37012946
DR
200 return -EINVAL;
201 pol->v.nodes = *nodes;
202 return 0;
203}
204
58568d2a
MX
205/*
206 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
207 * any, for the new policy. mpol_new() has already validated the nodes
208 * parameter with respect to the policy mode and flags. But, we need to
209 * handle an empty nodemask with MPOL_PREFERRED here.
210 *
211 * Must be called holding task's alloc_lock to protect task's mems_allowed
212 * and mempolicy. May also be called holding the mmap_semaphore for write.
213 */
4bfc4495
KH
214static int mpol_set_nodemask(struct mempolicy *pol,
215 const nodemask_t *nodes, struct nodemask_scratch *nsc)
58568d2a 216{
58568d2a
MX
217 int ret;
218
219 /* if mode is MPOL_DEFAULT, pol is NULL. This is right. */
220 if (pol == NULL)
221 return 0;
01f13bd6 222 /* Check N_MEMORY */
4bfc4495 223 nodes_and(nsc->mask1,
01f13bd6 224 cpuset_current_mems_allowed, node_states[N_MEMORY]);
58568d2a
MX
225
226 VM_BUG_ON(!nodes);
227 if (pol->mode == MPOL_PREFERRED && nodes_empty(*nodes))
228 nodes = NULL; /* explicit local allocation */
229 else {
230 if (pol->flags & MPOL_F_RELATIVE_NODES)
859f7ef1 231 mpol_relative_nodemask(&nsc->mask2, nodes, &nsc->mask1);
58568d2a 232 else
4bfc4495
KH
233 nodes_and(nsc->mask2, *nodes, nsc->mask1);
234
58568d2a
MX
235 if (mpol_store_user_nodemask(pol))
236 pol->w.user_nodemask = *nodes;
237 else
238 pol->w.cpuset_mems_allowed =
239 cpuset_current_mems_allowed;
240 }
241
4bfc4495
KH
242 if (nodes)
243 ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
244 else
245 ret = mpol_ops[pol->mode].create(pol, NULL);
58568d2a
MX
246 return ret;
247}
248
249/*
250 * This function just creates a new policy, does some check and simple
251 * initialization. You must invoke mpol_set_nodemask() to set nodes.
252 */
028fec41
DR
253static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
254 nodemask_t *nodes)
1da177e4
LT
255{
256 struct mempolicy *policy;
257
028fec41 258 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
00ef2d2f 259 mode, flags, nodes ? nodes_addr(*nodes)[0] : NUMA_NO_NODE);
140d5a49 260
3e1f0645
DR
261 if (mode == MPOL_DEFAULT) {
262 if (nodes && !nodes_empty(*nodes))
37012946 263 return ERR_PTR(-EINVAL);
d3a71033 264 return NULL;
37012946 265 }
3e1f0645
DR
266 VM_BUG_ON(!nodes);
267
268 /*
269 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
270 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
271 * All other modes require a valid pointer to a non-empty nodemask.
272 */
273 if (mode == MPOL_PREFERRED) {
274 if (nodes_empty(*nodes)) {
275 if (((flags & MPOL_F_STATIC_NODES) ||
276 (flags & MPOL_F_RELATIVE_NODES)))
277 return ERR_PTR(-EINVAL);
3e1f0645 278 }
479e2802
PZ
279 } else if (mode == MPOL_LOCAL) {
280 if (!nodes_empty(*nodes))
281 return ERR_PTR(-EINVAL);
282 mode = MPOL_PREFERRED;
3e1f0645
DR
283 } else if (nodes_empty(*nodes))
284 return ERR_PTR(-EINVAL);
1da177e4
LT
285 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
286 if (!policy)
287 return ERR_PTR(-ENOMEM);
288 atomic_set(&policy->refcnt, 1);
45c4745a 289 policy->mode = mode;
3e1f0645 290 policy->flags = flags;
37012946 291
1da177e4 292 return policy;
37012946
DR
293}
294
52cd3b07
LS
295/* Slow path of a mpol destructor. */
296void __mpol_put(struct mempolicy *p)
297{
298 if (!atomic_dec_and_test(&p->refcnt))
299 return;
52cd3b07
LS
300 kmem_cache_free(policy_cache, p);
301}
302
708c1bbc
MX
303static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes,
304 enum mpol_rebind_step step)
37012946
DR
305{
306}
307
708c1bbc
MX
308/*
309 * step:
310 * MPOL_REBIND_ONCE - do rebind work at once
311 * MPOL_REBIND_STEP1 - set all the newly nodes
312 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
313 */
314static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes,
315 enum mpol_rebind_step step)
37012946
DR
316{
317 nodemask_t tmp;
318
319 if (pol->flags & MPOL_F_STATIC_NODES)
320 nodes_and(tmp, pol->w.user_nodemask, *nodes);
321 else if (pol->flags & MPOL_F_RELATIVE_NODES)
322 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
323 else {
708c1bbc
MX
324 /*
325 * if step == 1, we use ->w.cpuset_mems_allowed to cache the
326 * result
327 */
328 if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP1) {
329 nodes_remap(tmp, pol->v.nodes,
330 pol->w.cpuset_mems_allowed, *nodes);
331 pol->w.cpuset_mems_allowed = step ? tmp : *nodes;
332 } else if (step == MPOL_REBIND_STEP2) {
333 tmp = pol->w.cpuset_mems_allowed;
334 pol->w.cpuset_mems_allowed = *nodes;
335 } else
336 BUG();
37012946 337 }
f5b087b5 338
708c1bbc
MX
339 if (nodes_empty(tmp))
340 tmp = *nodes;
341
342 if (step == MPOL_REBIND_STEP1)
343 nodes_or(pol->v.nodes, pol->v.nodes, tmp);
344 else if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP2)
345 pol->v.nodes = tmp;
346 else
347 BUG();
348
37012946
DR
349 if (!node_isset(current->il_next, tmp)) {
350 current->il_next = next_node(current->il_next, tmp);
351 if (current->il_next >= MAX_NUMNODES)
352 current->il_next = first_node(tmp);
353 if (current->il_next >= MAX_NUMNODES)
354 current->il_next = numa_node_id();
355 }
356}
357
358static void mpol_rebind_preferred(struct mempolicy *pol,
708c1bbc
MX
359 const nodemask_t *nodes,
360 enum mpol_rebind_step step)
37012946
DR
361{
362 nodemask_t tmp;
363
37012946
DR
364 if (pol->flags & MPOL_F_STATIC_NODES) {
365 int node = first_node(pol->w.user_nodemask);
366
fc36b8d3 367 if (node_isset(node, *nodes)) {
37012946 368 pol->v.preferred_node = node;
fc36b8d3
LS
369 pol->flags &= ~MPOL_F_LOCAL;
370 } else
371 pol->flags |= MPOL_F_LOCAL;
37012946
DR
372 } else if (pol->flags & MPOL_F_RELATIVE_NODES) {
373 mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
374 pol->v.preferred_node = first_node(tmp);
fc36b8d3 375 } else if (!(pol->flags & MPOL_F_LOCAL)) {
37012946
DR
376 pol->v.preferred_node = node_remap(pol->v.preferred_node,
377 pol->w.cpuset_mems_allowed,
378 *nodes);
379 pol->w.cpuset_mems_allowed = *nodes;
380 }
1da177e4
LT
381}
382
708c1bbc
MX
383/*
384 * mpol_rebind_policy - Migrate a policy to a different set of nodes
385 *
386 * If read-side task has no lock to protect task->mempolicy, write-side
387 * task will rebind the task->mempolicy by two step. The first step is
388 * setting all the newly nodes, and the second step is cleaning all the
389 * disallowed nodes. In this way, we can avoid finding no node to alloc
390 * page.
391 * If we have a lock to protect task->mempolicy in read-side, we do
392 * rebind directly.
393 *
394 * step:
395 * MPOL_REBIND_ONCE - do rebind work at once
396 * MPOL_REBIND_STEP1 - set all the newly nodes
397 * MPOL_REBIND_STEP2 - clean all the disallowed nodes
398 */
399static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask,
400 enum mpol_rebind_step step)
1d0d2680 401{
1d0d2680
DR
402 if (!pol)
403 return;
89c522c7 404 if (!mpol_store_user_nodemask(pol) && step == MPOL_REBIND_ONCE &&
1d0d2680
DR
405 nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
406 return;
708c1bbc
MX
407
408 if (step == MPOL_REBIND_STEP1 && (pol->flags & MPOL_F_REBINDING))
409 return;
410
411 if (step == MPOL_REBIND_STEP2 && !(pol->flags & MPOL_F_REBINDING))
412 BUG();
413
414 if (step == MPOL_REBIND_STEP1)
415 pol->flags |= MPOL_F_REBINDING;
416 else if (step == MPOL_REBIND_STEP2)
417 pol->flags &= ~MPOL_F_REBINDING;
418 else if (step >= MPOL_REBIND_NSTEP)
419 BUG();
420
421 mpol_ops[pol->mode].rebind(pol, newmask, step);
1d0d2680
DR
422}
423
424/*
425 * Wrapper for mpol_rebind_policy() that just requires task
426 * pointer, and updates task mempolicy.
58568d2a
MX
427 *
428 * Called with task's alloc_lock held.
1d0d2680
DR
429 */
430
708c1bbc
MX
431void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new,
432 enum mpol_rebind_step step)
1d0d2680 433{
708c1bbc 434 mpol_rebind_policy(tsk->mempolicy, new, step);
1d0d2680
DR
435}
436
437/*
438 * Rebind each vma in mm to new nodemask.
439 *
440 * Call holding a reference to mm. Takes mm->mmap_sem during call.
441 */
442
443void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
444{
445 struct vm_area_struct *vma;
446
447 down_write(&mm->mmap_sem);
448 for (vma = mm->mmap; vma; vma = vma->vm_next)
708c1bbc 449 mpol_rebind_policy(vma->vm_policy, new, MPOL_REBIND_ONCE);
1d0d2680
DR
450 up_write(&mm->mmap_sem);
451}
452
37012946
DR
453static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
454 [MPOL_DEFAULT] = {
455 .rebind = mpol_rebind_default,
456 },
457 [MPOL_INTERLEAVE] = {
458 .create = mpol_new_interleave,
459 .rebind = mpol_rebind_nodemask,
460 },
461 [MPOL_PREFERRED] = {
462 .create = mpol_new_preferred,
463 .rebind = mpol_rebind_preferred,
464 },
465 [MPOL_BIND] = {
466 .create = mpol_new_bind,
467 .rebind = mpol_rebind_nodemask,
468 },
469};
470
fc301289
CL
471static void migrate_page_add(struct page *page, struct list_head *pagelist,
472 unsigned long flags);
1a75a6c8 473
6f4576e3
NH
474struct queue_pages {
475 struct list_head *pagelist;
476 unsigned long flags;
477 nodemask_t *nmask;
478 struct vm_area_struct *prev;
479};
480
98094945
NH
481/*
482 * Scan through pages checking if pages follow certain conditions,
483 * and move them to the pagelist if they do.
484 */
6f4576e3
NH
485static int queue_pages_pte_range(pmd_t *pmd, unsigned long addr,
486 unsigned long end, struct mm_walk *walk)
1da177e4 487{
6f4576e3
NH
488 struct vm_area_struct *vma = walk->vma;
489 struct page *page;
490 struct queue_pages *qp = walk->private;
491 unsigned long flags = qp->flags;
492 int nid;
91612e0d 493 pte_t *pte;
705e87c0 494 spinlock_t *ptl;
941150a3 495
6f4576e3
NH
496 split_huge_page_pmd(vma, addr, pmd);
497 if (pmd_trans_unstable(pmd))
498 return 0;
91612e0d 499
6f4576e3
NH
500 pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
501 for (; addr != end; pte++, addr += PAGE_SIZE) {
91612e0d 502 if (!pte_present(*pte))
1da177e4 503 continue;
6aab341e
LT
504 page = vm_normal_page(vma, addr, *pte);
505 if (!page)
1da177e4 506 continue;
053837fc 507 /*
62b61f61
HD
508 * vm_normal_page() filters out zero pages, but there might
509 * still be PageReserved pages to skip, perhaps in a VDSO.
053837fc 510 */
b79bc0a0 511 if (PageReserved(page))
f4598c8b 512 continue;
6aab341e 513 nid = page_to_nid(page);
6f4576e3 514 if (node_isset(nid, *qp->nmask) == !!(flags & MPOL_MF_INVERT))
38e35860
CL
515 continue;
516
b1f72d18 517 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
6f4576e3
NH
518 migrate_page_add(page, qp->pagelist, flags);
519 }
520 pte_unmap_unlock(pte - 1, ptl);
521 cond_resched();
522 return 0;
91612e0d
HD
523}
524
6f4576e3
NH
525static int queue_pages_hugetlb(pte_t *pte, unsigned long hmask,
526 unsigned long addr, unsigned long end,
527 struct mm_walk *walk)
e2d8cf40
NH
528{
529#ifdef CONFIG_HUGETLB_PAGE
6f4576e3
NH
530 struct queue_pages *qp = walk->private;
531 unsigned long flags = qp->flags;
e2d8cf40
NH
532 int nid;
533 struct page *page;
cb900f41 534 spinlock_t *ptl;
d4c54919 535 pte_t entry;
e2d8cf40 536
6f4576e3
NH
537 ptl = huge_pte_lock(hstate_vma(walk->vma), walk->mm, pte);
538 entry = huge_ptep_get(pte);
d4c54919
NH
539 if (!pte_present(entry))
540 goto unlock;
541 page = pte_page(entry);
e2d8cf40 542 nid = page_to_nid(page);
6f4576e3 543 if (node_isset(nid, *qp->nmask) == !!(flags & MPOL_MF_INVERT))
e2d8cf40
NH
544 goto unlock;
545 /* With MPOL_MF_MOVE, we migrate only unshared hugepage. */
546 if (flags & (MPOL_MF_MOVE_ALL) ||
547 (flags & MPOL_MF_MOVE && page_mapcount(page) == 1))
6f4576e3 548 isolate_huge_page(page, qp->pagelist);
e2d8cf40 549unlock:
cb900f41 550 spin_unlock(ptl);
e2d8cf40
NH
551#else
552 BUG();
553#endif
91612e0d 554 return 0;
1da177e4
LT
555}
556
5877231f 557#ifdef CONFIG_NUMA_BALANCING
b24f53a0 558/*
4b10e7d5
MG
559 * This is used to mark a range of virtual addresses to be inaccessible.
560 * These are later cleared by a NUMA hinting fault. Depending on these
561 * faults, pages may be migrated for better NUMA placement.
562 *
563 * This is assuming that NUMA faults are handled using PROT_NONE. If
564 * an architecture makes a different choice, it will need further
565 * changes to the core.
b24f53a0 566 */
4b10e7d5
MG
567unsigned long change_prot_numa(struct vm_area_struct *vma,
568 unsigned long addr, unsigned long end)
b24f53a0 569{
4b10e7d5 570 int nr_updated;
b24f53a0 571
4d942466 572 nr_updated = change_protection(vma, addr, end, PAGE_NONE, 0, 1);
03c5a6e1
MG
573 if (nr_updated)
574 count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
b24f53a0 575
4b10e7d5 576 return nr_updated;
b24f53a0
LS
577}
578#else
579static unsigned long change_prot_numa(struct vm_area_struct *vma,
580 unsigned long addr, unsigned long end)
581{
582 return 0;
583}
5877231f 584#endif /* CONFIG_NUMA_BALANCING */
b24f53a0 585
6f4576e3
NH
586static int queue_pages_test_walk(unsigned long start, unsigned long end,
587 struct mm_walk *walk)
588{
589 struct vm_area_struct *vma = walk->vma;
590 struct queue_pages *qp = walk->private;
591 unsigned long endvma = vma->vm_end;
592 unsigned long flags = qp->flags;
593
48684a65
NH
594 if (vma->vm_flags & VM_PFNMAP)
595 return 1;
596
6f4576e3
NH
597 if (endvma > end)
598 endvma = end;
599 if (vma->vm_start > start)
600 start = vma->vm_start;
601
602 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
603 if (!vma->vm_next && vma->vm_end < end)
604 return -EFAULT;
605 if (qp->prev && qp->prev->vm_end < vma->vm_start)
606 return -EFAULT;
607 }
608
609 qp->prev = vma;
610
6f4576e3
NH
611 if (flags & MPOL_MF_LAZY) {
612 /* Similar to task_numa_work, skip inaccessible VMAs */
613 if (vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))
614 change_prot_numa(vma, start, endvma);
615 return 1;
616 }
617
618 if ((flags & MPOL_MF_STRICT) ||
619 ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) &&
620 vma_migratable(vma)))
621 /* queue pages from current vma */
622 return 0;
623 return 1;
624}
625
dc9aa5b9 626/*
98094945
NH
627 * Walk through page tables and collect pages to be migrated.
628 *
629 * If pages found in a given range are on a set of nodes (determined by
630 * @nodes and @flags,) it's isolated and queued to the pagelist which is
631 * passed via @private.)
dc9aa5b9 632 */
d05f0cdc 633static int
98094945 634queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
6f4576e3
NH
635 nodemask_t *nodes, unsigned long flags,
636 struct list_head *pagelist)
1da177e4 637{
6f4576e3
NH
638 struct queue_pages qp = {
639 .pagelist = pagelist,
640 .flags = flags,
641 .nmask = nodes,
642 .prev = NULL,
643 };
644 struct mm_walk queue_pages_walk = {
645 .hugetlb_entry = queue_pages_hugetlb,
646 .pmd_entry = queue_pages_pte_range,
647 .test_walk = queue_pages_test_walk,
648 .mm = mm,
649 .private = &qp,
650 };
651
652 return walk_page_range(start, end, &queue_pages_walk);
1da177e4
LT
653}
654
869833f2
KM
655/*
656 * Apply policy to a single VMA
657 * This must be called with the mmap_sem held for writing.
658 */
659static int vma_replace_policy(struct vm_area_struct *vma,
660 struct mempolicy *pol)
8d34694c 661{
869833f2
KM
662 int err;
663 struct mempolicy *old;
664 struct mempolicy *new;
8d34694c
KM
665
666 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
667 vma->vm_start, vma->vm_end, vma->vm_pgoff,
668 vma->vm_ops, vma->vm_file,
669 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
670
869833f2
KM
671 new = mpol_dup(pol);
672 if (IS_ERR(new))
673 return PTR_ERR(new);
674
675 if (vma->vm_ops && vma->vm_ops->set_policy) {
8d34694c 676 err = vma->vm_ops->set_policy(vma, new);
869833f2
KM
677 if (err)
678 goto err_out;
8d34694c 679 }
869833f2
KM
680
681 old = vma->vm_policy;
682 vma->vm_policy = new; /* protected by mmap_sem */
683 mpol_put(old);
684
685 return 0;
686 err_out:
687 mpol_put(new);
8d34694c
KM
688 return err;
689}
690
1da177e4 691/* Step 2: apply policy to a range and do splits. */
9d8cebd4
KM
692static int mbind_range(struct mm_struct *mm, unsigned long start,
693 unsigned long end, struct mempolicy *new_pol)
1da177e4
LT
694{
695 struct vm_area_struct *next;
9d8cebd4
KM
696 struct vm_area_struct *prev;
697 struct vm_area_struct *vma;
698 int err = 0;
e26a5114 699 pgoff_t pgoff;
9d8cebd4
KM
700 unsigned long vmstart;
701 unsigned long vmend;
1da177e4 702
097d5910 703 vma = find_vma(mm, start);
9d8cebd4
KM
704 if (!vma || vma->vm_start > start)
705 return -EFAULT;
706
097d5910 707 prev = vma->vm_prev;
e26a5114
KM
708 if (start > vma->vm_start)
709 prev = vma;
710
9d8cebd4 711 for (; vma && vma->vm_start < end; prev = vma, vma = next) {
1da177e4 712 next = vma->vm_next;
9d8cebd4
KM
713 vmstart = max(start, vma->vm_start);
714 vmend = min(end, vma->vm_end);
715
e26a5114
KM
716 if (mpol_equal(vma_policy(vma), new_pol))
717 continue;
718
719 pgoff = vma->vm_pgoff +
720 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
9d8cebd4 721 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
19a809af
AA
722 vma->anon_vma, vma->vm_file, pgoff,
723 new_pol, vma->vm_userfaultfd_ctx);
9d8cebd4
KM
724 if (prev) {
725 vma = prev;
726 next = vma->vm_next;
3964acd0
ON
727 if (mpol_equal(vma_policy(vma), new_pol))
728 continue;
729 /* vma_merge() joined vma && vma->next, case 8 */
730 goto replace;
9d8cebd4
KM
731 }
732 if (vma->vm_start != vmstart) {
733 err = split_vma(vma->vm_mm, vma, vmstart, 1);
734 if (err)
735 goto out;
736 }
737 if (vma->vm_end != vmend) {
738 err = split_vma(vma->vm_mm, vma, vmend, 0);
739 if (err)
740 goto out;
741 }
3964acd0 742 replace:
869833f2 743 err = vma_replace_policy(vma, new_pol);
8d34694c
KM
744 if (err)
745 goto out;
1da177e4 746 }
9d8cebd4
KM
747
748 out:
1da177e4
LT
749 return err;
750}
751
1da177e4 752/* Set the process memory policy */
028fec41
DR
753static long do_set_mempolicy(unsigned short mode, unsigned short flags,
754 nodemask_t *nodes)
1da177e4 755{
58568d2a 756 struct mempolicy *new, *old;
4bfc4495 757 NODEMASK_SCRATCH(scratch);
58568d2a 758 int ret;
1da177e4 759
4bfc4495
KH
760 if (!scratch)
761 return -ENOMEM;
f4e53d91 762
4bfc4495
KH
763 new = mpol_new(mode, flags, nodes);
764 if (IS_ERR(new)) {
765 ret = PTR_ERR(new);
766 goto out;
767 }
2c7c3a7d 768
58568d2a 769 task_lock(current);
4bfc4495 770 ret = mpol_set_nodemask(new, nodes, scratch);
58568d2a
MX
771 if (ret) {
772 task_unlock(current);
58568d2a 773 mpol_put(new);
4bfc4495 774 goto out;
58568d2a
MX
775 }
776 old = current->mempolicy;
1da177e4 777 current->mempolicy = new;
45c4745a 778 if (new && new->mode == MPOL_INTERLEAVE &&
f5b087b5 779 nodes_weight(new->v.nodes))
dfcd3c0d 780 current->il_next = first_node(new->v.nodes);
58568d2a 781 task_unlock(current);
58568d2a 782 mpol_put(old);
4bfc4495
KH
783 ret = 0;
784out:
785 NODEMASK_SCRATCH_FREE(scratch);
786 return ret;
1da177e4
LT
787}
788
bea904d5
LS
789/*
790 * Return nodemask for policy for get_mempolicy() query
58568d2a
MX
791 *
792 * Called with task's alloc_lock held
bea904d5
LS
793 */
794static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
1da177e4 795{
dfcd3c0d 796 nodes_clear(*nodes);
bea904d5
LS
797 if (p == &default_policy)
798 return;
799
45c4745a 800 switch (p->mode) {
19770b32
MG
801 case MPOL_BIND:
802 /* Fall through */
1da177e4 803 case MPOL_INTERLEAVE:
dfcd3c0d 804 *nodes = p->v.nodes;
1da177e4
LT
805 break;
806 case MPOL_PREFERRED:
fc36b8d3 807 if (!(p->flags & MPOL_F_LOCAL))
dfcd3c0d 808 node_set(p->v.preferred_node, *nodes);
53f2556b 809 /* else return empty node mask for local allocation */
1da177e4
LT
810 break;
811 default:
812 BUG();
813 }
814}
815
816static int lookup_node(struct mm_struct *mm, unsigned long addr)
817{
818 struct page *p;
819 int err;
820
821 err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL);
822 if (err >= 0) {
823 err = page_to_nid(p);
824 put_page(p);
825 }
826 return err;
827}
828
1da177e4 829/* Retrieve NUMA policy */
dbcb0f19
AB
830static long do_get_mempolicy(int *policy, nodemask_t *nmask,
831 unsigned long addr, unsigned long flags)
1da177e4 832{
8bccd85f 833 int err;
1da177e4
LT
834 struct mm_struct *mm = current->mm;
835 struct vm_area_struct *vma = NULL;
836 struct mempolicy *pol = current->mempolicy;
837
754af6f5
LS
838 if (flags &
839 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
1da177e4 840 return -EINVAL;
754af6f5
LS
841
842 if (flags & MPOL_F_MEMS_ALLOWED) {
843 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
844 return -EINVAL;
845 *policy = 0; /* just so it's initialized */
58568d2a 846 task_lock(current);
754af6f5 847 *nmask = cpuset_current_mems_allowed;
58568d2a 848 task_unlock(current);
754af6f5
LS
849 return 0;
850 }
851
1da177e4 852 if (flags & MPOL_F_ADDR) {
bea904d5
LS
853 /*
854 * Do NOT fall back to task policy if the
855 * vma/shared policy at addr is NULL. We
856 * want to return MPOL_DEFAULT in this case.
857 */
1da177e4
LT
858 down_read(&mm->mmap_sem);
859 vma = find_vma_intersection(mm, addr, addr+1);
860 if (!vma) {
861 up_read(&mm->mmap_sem);
862 return -EFAULT;
863 }
864 if (vma->vm_ops && vma->vm_ops->get_policy)
865 pol = vma->vm_ops->get_policy(vma, addr);
866 else
867 pol = vma->vm_policy;
868 } else if (addr)
869 return -EINVAL;
870
871 if (!pol)
bea904d5 872 pol = &default_policy; /* indicates default behavior */
1da177e4
LT
873
874 if (flags & MPOL_F_NODE) {
875 if (flags & MPOL_F_ADDR) {
876 err = lookup_node(mm, addr);
877 if (err < 0)
878 goto out;
8bccd85f 879 *policy = err;
1da177e4 880 } else if (pol == current->mempolicy &&
45c4745a 881 pol->mode == MPOL_INTERLEAVE) {
8bccd85f 882 *policy = current->il_next;
1da177e4
LT
883 } else {
884 err = -EINVAL;
885 goto out;
886 }
bea904d5
LS
887 } else {
888 *policy = pol == &default_policy ? MPOL_DEFAULT :
889 pol->mode;
d79df630
DR
890 /*
891 * Internal mempolicy flags must be masked off before exposing
892 * the policy to userspace.
893 */
894 *policy |= (pol->flags & MPOL_MODE_FLAGS);
bea904d5 895 }
1da177e4
LT
896
897 if (vma) {
898 up_read(&current->mm->mmap_sem);
899 vma = NULL;
900 }
901
1da177e4 902 err = 0;
58568d2a 903 if (nmask) {
c6b6ef8b
LS
904 if (mpol_store_user_nodemask(pol)) {
905 *nmask = pol->w.user_nodemask;
906 } else {
907 task_lock(current);
908 get_policy_nodemask(pol, nmask);
909 task_unlock(current);
910 }
58568d2a 911 }
1da177e4
LT
912
913 out:
52cd3b07 914 mpol_cond_put(pol);
1da177e4
LT
915 if (vma)
916 up_read(&current->mm->mmap_sem);
917 return err;
918}
919
b20a3503 920#ifdef CONFIG_MIGRATION
6ce3c4c0
CL
921/*
922 * page migration
923 */
fc301289
CL
924static void migrate_page_add(struct page *page, struct list_head *pagelist,
925 unsigned long flags)
6ce3c4c0
CL
926{
927 /*
fc301289 928 * Avoid migrating a page that is shared with others.
6ce3c4c0 929 */
62695a84
NP
930 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) {
931 if (!isolate_lru_page(page)) {
932 list_add_tail(&page->lru, pagelist);
6d9c285a
KM
933 inc_zone_page_state(page, NR_ISOLATED_ANON +
934 page_is_file_cache(page));
62695a84
NP
935 }
936 }
7e2ab150 937}
6ce3c4c0 938
742755a1 939static struct page *new_node_page(struct page *page, unsigned long node, int **x)
95a402c3 940{
e2d8cf40
NH
941 if (PageHuge(page))
942 return alloc_huge_page_node(page_hstate(compound_head(page)),
943 node);
944 else
96db800f 945 return __alloc_pages_node(node, GFP_HIGHUSER_MOVABLE |
b360edb4 946 __GFP_THISNODE, 0);
95a402c3
CL
947}
948
7e2ab150
CL
949/*
950 * Migrate pages from one node to a target node.
951 * Returns error or the number of pages not migrated.
952 */
dbcb0f19
AB
953static int migrate_to_node(struct mm_struct *mm, int source, int dest,
954 int flags)
7e2ab150
CL
955{
956 nodemask_t nmask;
957 LIST_HEAD(pagelist);
958 int err = 0;
959
960 nodes_clear(nmask);
961 node_set(source, nmask);
6ce3c4c0 962
08270807
MK
963 /*
964 * This does not "check" the range but isolates all pages that
965 * need migration. Between passing in the full user address
966 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
967 */
968 VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
98094945 969 queue_pages_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
7e2ab150
CL
970 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
971
cf608ac1 972 if (!list_empty(&pagelist)) {
68711a74 973 err = migrate_pages(&pagelist, new_node_page, NULL, dest,
9c620e2b 974 MIGRATE_SYNC, MR_SYSCALL);
cf608ac1 975 if (err)
e2d8cf40 976 putback_movable_pages(&pagelist);
cf608ac1 977 }
95a402c3 978
7e2ab150 979 return err;
6ce3c4c0
CL
980}
981
39743889 982/*
7e2ab150
CL
983 * Move pages between the two nodesets so as to preserve the physical
984 * layout as much as possible.
39743889
CL
985 *
986 * Returns the number of page that could not be moved.
987 */
0ce72d4f
AM
988int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
989 const nodemask_t *to, int flags)
39743889 990{
7e2ab150 991 int busy = 0;
0aedadf9 992 int err;
7e2ab150 993 nodemask_t tmp;
39743889 994
0aedadf9
CL
995 err = migrate_prep();
996 if (err)
997 return err;
998
53f2556b 999 down_read(&mm->mmap_sem);
39743889 1000
da0aa138
KM
1001 /*
1002 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1003 * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
1004 * bit in 'tmp', and return that <source, dest> pair for migration.
1005 * The pair of nodemasks 'to' and 'from' define the map.
1006 *
1007 * If no pair of bits is found that way, fallback to picking some
1008 * pair of 'source' and 'dest' bits that are not the same. If the
1009 * 'source' and 'dest' bits are the same, this represents a node
1010 * that will be migrating to itself, so no pages need move.
1011 *
1012 * If no bits are left in 'tmp', or if all remaining bits left
1013 * in 'tmp' correspond to the same bit in 'to', return false
1014 * (nothing left to migrate).
1015 *
1016 * This lets us pick a pair of nodes to migrate between, such that
1017 * if possible the dest node is not already occupied by some other
1018 * source node, minimizing the risk of overloading the memory on a
1019 * node that would happen if we migrated incoming memory to a node
1020 * before migrating outgoing memory source that same node.
1021 *
1022 * A single scan of tmp is sufficient. As we go, we remember the
1023 * most recent <s, d> pair that moved (s != d). If we find a pair
1024 * that not only moved, but what's better, moved to an empty slot
1025 * (d is not set in tmp), then we break out then, with that pair.
ae0e47f0 1026 * Otherwise when we finish scanning from_tmp, we at least have the
da0aa138
KM
1027 * most recent <s, d> pair that moved. If we get all the way through
1028 * the scan of tmp without finding any node that moved, much less
1029 * moved to an empty node, then there is nothing left worth migrating.
1030 */
d4984711 1031
0ce72d4f 1032 tmp = *from;
7e2ab150
CL
1033 while (!nodes_empty(tmp)) {
1034 int s,d;
b76ac7e7 1035 int source = NUMA_NO_NODE;
7e2ab150
CL
1036 int dest = 0;
1037
1038 for_each_node_mask(s, tmp) {
4a5b18cc
LW
1039
1040 /*
1041 * do_migrate_pages() tries to maintain the relative
1042 * node relationship of the pages established between
1043 * threads and memory areas.
1044 *
1045 * However if the number of source nodes is not equal to
1046 * the number of destination nodes we can not preserve
1047 * this node relative relationship. In that case, skip
1048 * copying memory from a node that is in the destination
1049 * mask.
1050 *
1051 * Example: [2,3,4] -> [3,4,5] moves everything.
1052 * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1053 */
1054
0ce72d4f
AM
1055 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1056 (node_isset(s, *to)))
4a5b18cc
LW
1057 continue;
1058
0ce72d4f 1059 d = node_remap(s, *from, *to);
7e2ab150
CL
1060 if (s == d)
1061 continue;
1062
1063 source = s; /* Node moved. Memorize */
1064 dest = d;
1065
1066 /* dest not in remaining from nodes? */
1067 if (!node_isset(dest, tmp))
1068 break;
1069 }
b76ac7e7 1070 if (source == NUMA_NO_NODE)
7e2ab150
CL
1071 break;
1072
1073 node_clear(source, tmp);
1074 err = migrate_to_node(mm, source, dest, flags);
1075 if (err > 0)
1076 busy += err;
1077 if (err < 0)
1078 break;
39743889
CL
1079 }
1080 up_read(&mm->mmap_sem);
7e2ab150
CL
1081 if (err < 0)
1082 return err;
1083 return busy;
b20a3503
CL
1084
1085}
1086
3ad33b24
LS
1087/*
1088 * Allocate a new page for page migration based on vma policy.
d05f0cdc 1089 * Start by assuming the page is mapped by the same vma as contains @start.
3ad33b24
LS
1090 * Search forward from there, if not. N.B., this assumes that the
1091 * list of pages handed to migrate_pages()--which is how we get here--
1092 * is in virtual address order.
1093 */
d05f0cdc 1094static struct page *new_page(struct page *page, unsigned long start, int **x)
95a402c3 1095{
d05f0cdc 1096 struct vm_area_struct *vma;
3ad33b24 1097 unsigned long uninitialized_var(address);
95a402c3 1098
d05f0cdc 1099 vma = find_vma(current->mm, start);
3ad33b24
LS
1100 while (vma) {
1101 address = page_address_in_vma(page, vma);
1102 if (address != -EFAULT)
1103 break;
1104 vma = vma->vm_next;
1105 }
11c731e8
WL
1106
1107 if (PageHuge(page)) {
cc81717e
MH
1108 BUG_ON(!vma);
1109 return alloc_huge_page_noerr(vma, address, 1);
11c731e8 1110 }
0bf598d8 1111 /*
11c731e8 1112 * if !vma, alloc_page_vma() will use task or system default policy
0bf598d8 1113 */
3ad33b24 1114 return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
95a402c3 1115}
b20a3503
CL
1116#else
1117
1118static void migrate_page_add(struct page *page, struct list_head *pagelist,
1119 unsigned long flags)
1120{
39743889
CL
1121}
1122
0ce72d4f
AM
1123int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1124 const nodemask_t *to, int flags)
b20a3503
CL
1125{
1126 return -ENOSYS;
1127}
95a402c3 1128
d05f0cdc 1129static struct page *new_page(struct page *page, unsigned long start, int **x)
95a402c3
CL
1130{
1131 return NULL;
1132}
b20a3503
CL
1133#endif
1134
dbcb0f19 1135static long do_mbind(unsigned long start, unsigned long len,
028fec41
DR
1136 unsigned short mode, unsigned short mode_flags,
1137 nodemask_t *nmask, unsigned long flags)
6ce3c4c0 1138{
6ce3c4c0
CL
1139 struct mm_struct *mm = current->mm;
1140 struct mempolicy *new;
1141 unsigned long end;
1142 int err;
1143 LIST_HEAD(pagelist);
1144
b24f53a0 1145 if (flags & ~(unsigned long)MPOL_MF_VALID)
6ce3c4c0 1146 return -EINVAL;
74c00241 1147 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
6ce3c4c0
CL
1148 return -EPERM;
1149
1150 if (start & ~PAGE_MASK)
1151 return -EINVAL;
1152
1153 if (mode == MPOL_DEFAULT)
1154 flags &= ~MPOL_MF_STRICT;
1155
1156 len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1157 end = start + len;
1158
1159 if (end < start)
1160 return -EINVAL;
1161 if (end == start)
1162 return 0;
1163
028fec41 1164 new = mpol_new(mode, mode_flags, nmask);
6ce3c4c0
CL
1165 if (IS_ERR(new))
1166 return PTR_ERR(new);
1167
b24f53a0
LS
1168 if (flags & MPOL_MF_LAZY)
1169 new->flags |= MPOL_F_MOF;
1170
6ce3c4c0
CL
1171 /*
1172 * If we are using the default policy then operation
1173 * on discontinuous address spaces is okay after all
1174 */
1175 if (!new)
1176 flags |= MPOL_MF_DISCONTIG_OK;
1177
028fec41
DR
1178 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1179 start, start + len, mode, mode_flags,
00ef2d2f 1180 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
6ce3c4c0 1181
0aedadf9
CL
1182 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1183
1184 err = migrate_prep();
1185 if (err)
b05ca738 1186 goto mpol_out;
0aedadf9 1187 }
4bfc4495
KH
1188 {
1189 NODEMASK_SCRATCH(scratch);
1190 if (scratch) {
1191 down_write(&mm->mmap_sem);
1192 task_lock(current);
1193 err = mpol_set_nodemask(new, nmask, scratch);
1194 task_unlock(current);
1195 if (err)
1196 up_write(&mm->mmap_sem);
1197 } else
1198 err = -ENOMEM;
1199 NODEMASK_SCRATCH_FREE(scratch);
1200 }
b05ca738
KM
1201 if (err)
1202 goto mpol_out;
1203
d05f0cdc 1204 err = queue_pages_range(mm, start, end, nmask,
6ce3c4c0 1205 flags | MPOL_MF_INVERT, &pagelist);
d05f0cdc 1206 if (!err)
9d8cebd4 1207 err = mbind_range(mm, start, end, new);
7e2ab150 1208
b24f53a0
LS
1209 if (!err) {
1210 int nr_failed = 0;
1211
cf608ac1 1212 if (!list_empty(&pagelist)) {
b24f53a0 1213 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
d05f0cdc
HD
1214 nr_failed = migrate_pages(&pagelist, new_page, NULL,
1215 start, MIGRATE_SYNC, MR_MEMPOLICY_MBIND);
cf608ac1 1216 if (nr_failed)
74060e4d 1217 putback_movable_pages(&pagelist);
cf608ac1 1218 }
6ce3c4c0 1219
b24f53a0 1220 if (nr_failed && (flags & MPOL_MF_STRICT))
6ce3c4c0 1221 err = -EIO;
ab8a3e14 1222 } else
b0e5fd73 1223 putback_movable_pages(&pagelist);
b20a3503 1224
6ce3c4c0 1225 up_write(&mm->mmap_sem);
b05ca738 1226 mpol_out:
f0be3d32 1227 mpol_put(new);
6ce3c4c0
CL
1228 return err;
1229}
1230
8bccd85f
CL
1231/*
1232 * User space interface with variable sized bitmaps for nodelists.
1233 */
1234
1235/* Copy a node mask from user space. */
39743889 1236static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
8bccd85f
CL
1237 unsigned long maxnode)
1238{
1239 unsigned long k;
1240 unsigned long nlongs;
1241 unsigned long endmask;
1242
1243 --maxnode;
1244 nodes_clear(*nodes);
1245 if (maxnode == 0 || !nmask)
1246 return 0;
a9c930ba 1247 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
636f13c1 1248 return -EINVAL;
8bccd85f
CL
1249
1250 nlongs = BITS_TO_LONGS(maxnode);
1251 if ((maxnode % BITS_PER_LONG) == 0)
1252 endmask = ~0UL;
1253 else
1254 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1255
1256 /* When the user specified more nodes than supported just check
1257 if the non supported part is all zero. */
1258 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1259 if (nlongs > PAGE_SIZE/sizeof(long))
1260 return -EINVAL;
1261 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1262 unsigned long t;
1263 if (get_user(t, nmask + k))
1264 return -EFAULT;
1265 if (k == nlongs - 1) {
1266 if (t & endmask)
1267 return -EINVAL;
1268 } else if (t)
1269 return -EINVAL;
1270 }
1271 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1272 endmask = ~0UL;
1273 }
1274
1275 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1276 return -EFAULT;
1277 nodes_addr(*nodes)[nlongs-1] &= endmask;
1278 return 0;
1279}
1280
1281/* Copy a kernel node mask to user space */
1282static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1283 nodemask_t *nodes)
1284{
1285 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1286 const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
1287
1288 if (copy > nbytes) {
1289 if (copy > PAGE_SIZE)
1290 return -EINVAL;
1291 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1292 return -EFAULT;
1293 copy = nbytes;
1294 }
1295 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1296}
1297
938bb9f5 1298SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
f7f28ca9 1299 unsigned long, mode, const unsigned long __user *, nmask,
938bb9f5 1300 unsigned long, maxnode, unsigned, flags)
8bccd85f
CL
1301{
1302 nodemask_t nodes;
1303 int err;
028fec41 1304 unsigned short mode_flags;
8bccd85f 1305
028fec41
DR
1306 mode_flags = mode & MPOL_MODE_FLAGS;
1307 mode &= ~MPOL_MODE_FLAGS;
a3b51e01
DR
1308 if (mode >= MPOL_MAX)
1309 return -EINVAL;
4c50bc01
DR
1310 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1311 (mode_flags & MPOL_F_RELATIVE_NODES))
1312 return -EINVAL;
8bccd85f
CL
1313 err = get_nodes(&nodes, nmask, maxnode);
1314 if (err)
1315 return err;
028fec41 1316 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
8bccd85f
CL
1317}
1318
1319/* Set the process memory policy */
23c8902d 1320SYSCALL_DEFINE3(set_mempolicy, int, mode, const unsigned long __user *, nmask,
938bb9f5 1321 unsigned long, maxnode)
8bccd85f
CL
1322{
1323 int err;
1324 nodemask_t nodes;
028fec41 1325 unsigned short flags;
8bccd85f 1326
028fec41
DR
1327 flags = mode & MPOL_MODE_FLAGS;
1328 mode &= ~MPOL_MODE_FLAGS;
1329 if ((unsigned int)mode >= MPOL_MAX)
8bccd85f 1330 return -EINVAL;
4c50bc01
DR
1331 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1332 return -EINVAL;
8bccd85f
CL
1333 err = get_nodes(&nodes, nmask, maxnode);
1334 if (err)
1335 return err;
028fec41 1336 return do_set_mempolicy(mode, flags, &nodes);
8bccd85f
CL
1337}
1338
938bb9f5
HC
1339SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1340 const unsigned long __user *, old_nodes,
1341 const unsigned long __user *, new_nodes)
39743889 1342{
c69e8d9c 1343 const struct cred *cred = current_cred(), *tcred;
596d7cfa 1344 struct mm_struct *mm = NULL;
39743889 1345 struct task_struct *task;
39743889
CL
1346 nodemask_t task_nodes;
1347 int err;
596d7cfa
KM
1348 nodemask_t *old;
1349 nodemask_t *new;
1350 NODEMASK_SCRATCH(scratch);
1351
1352 if (!scratch)
1353 return -ENOMEM;
39743889 1354
596d7cfa
KM
1355 old = &scratch->mask1;
1356 new = &scratch->mask2;
1357
1358 err = get_nodes(old, old_nodes, maxnode);
39743889 1359 if (err)
596d7cfa 1360 goto out;
39743889 1361
596d7cfa 1362 err = get_nodes(new, new_nodes, maxnode);
39743889 1363 if (err)
596d7cfa 1364 goto out;
39743889
CL
1365
1366 /* Find the mm_struct */
55cfaa3c 1367 rcu_read_lock();
228ebcbe 1368 task = pid ? find_task_by_vpid(pid) : current;
39743889 1369 if (!task) {
55cfaa3c 1370 rcu_read_unlock();
596d7cfa
KM
1371 err = -ESRCH;
1372 goto out;
39743889 1373 }
3268c63e 1374 get_task_struct(task);
39743889 1375
596d7cfa 1376 err = -EINVAL;
39743889
CL
1377
1378 /*
1379 * Check if this process has the right to modify the specified
1380 * process. The right exists if the process has administrative
7f927fcc 1381 * capabilities, superuser privileges or the same
39743889
CL
1382 * userid as the target process.
1383 */
c69e8d9c 1384 tcred = __task_cred(task);
b38a86eb
EB
1385 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1386 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
74c00241 1387 !capable(CAP_SYS_NICE)) {
c69e8d9c 1388 rcu_read_unlock();
39743889 1389 err = -EPERM;
3268c63e 1390 goto out_put;
39743889 1391 }
c69e8d9c 1392 rcu_read_unlock();
39743889
CL
1393
1394 task_nodes = cpuset_mems_allowed(task);
1395 /* Is the user allowed to access the target nodes? */
596d7cfa 1396 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
39743889 1397 err = -EPERM;
3268c63e 1398 goto out_put;
39743889
CL
1399 }
1400
01f13bd6 1401 if (!nodes_subset(*new, node_states[N_MEMORY])) {
3b42d28b 1402 err = -EINVAL;
3268c63e 1403 goto out_put;
3b42d28b
CL
1404 }
1405
86c3a764
DQ
1406 err = security_task_movememory(task);
1407 if (err)
3268c63e 1408 goto out_put;
86c3a764 1409
3268c63e
CL
1410 mm = get_task_mm(task);
1411 put_task_struct(task);
f2a9ef88
SL
1412
1413 if (!mm) {
3268c63e 1414 err = -EINVAL;
f2a9ef88
SL
1415 goto out;
1416 }
1417
1418 err = do_migrate_pages(mm, old, new,
1419 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
3268c63e
CL
1420
1421 mmput(mm);
1422out:
596d7cfa
KM
1423 NODEMASK_SCRATCH_FREE(scratch);
1424
39743889 1425 return err;
3268c63e
CL
1426
1427out_put:
1428 put_task_struct(task);
1429 goto out;
1430
39743889
CL
1431}
1432
1433
8bccd85f 1434/* Retrieve NUMA policy */
938bb9f5
HC
1435SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1436 unsigned long __user *, nmask, unsigned long, maxnode,
1437 unsigned long, addr, unsigned long, flags)
8bccd85f 1438{
dbcb0f19
AB
1439 int err;
1440 int uninitialized_var(pval);
8bccd85f
CL
1441 nodemask_t nodes;
1442
1443 if (nmask != NULL && maxnode < MAX_NUMNODES)
1444 return -EINVAL;
1445
1446 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1447
1448 if (err)
1449 return err;
1450
1451 if (policy && put_user(pval, policy))
1452 return -EFAULT;
1453
1454 if (nmask)
1455 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1456
1457 return err;
1458}
1459
1da177e4
LT
1460#ifdef CONFIG_COMPAT
1461
c93e0f6c
HC
1462COMPAT_SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1463 compat_ulong_t __user *, nmask,
1464 compat_ulong_t, maxnode,
1465 compat_ulong_t, addr, compat_ulong_t, flags)
1da177e4
LT
1466{
1467 long err;
1468 unsigned long __user *nm = NULL;
1469 unsigned long nr_bits, alloc_size;
1470 DECLARE_BITMAP(bm, MAX_NUMNODES);
1471
1472 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1473 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1474
1475 if (nmask)
1476 nm = compat_alloc_user_space(alloc_size);
1477
1478 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1479
1480 if (!err && nmask) {
2bbff6c7
KH
1481 unsigned long copy_size;
1482 copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1483 err = copy_from_user(bm, nm, copy_size);
1da177e4
LT
1484 /* ensure entire bitmap is zeroed */
1485 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1486 err |= compat_put_bitmap(nmask, bm, nr_bits);
1487 }
1488
1489 return err;
1490}
1491
c93e0f6c
HC
1492COMPAT_SYSCALL_DEFINE3(set_mempolicy, int, mode, compat_ulong_t __user *, nmask,
1493 compat_ulong_t, maxnode)
1da177e4
LT
1494{
1495 long err = 0;
1496 unsigned long __user *nm = NULL;
1497 unsigned long nr_bits, alloc_size;
1498 DECLARE_BITMAP(bm, MAX_NUMNODES);
1499
1500 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1501 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1502
1503 if (nmask) {
1504 err = compat_get_bitmap(bm, nmask, nr_bits);
1505 nm = compat_alloc_user_space(alloc_size);
1506 err |= copy_to_user(nm, bm, alloc_size);
1507 }
1508
1509 if (err)
1510 return -EFAULT;
1511
1512 return sys_set_mempolicy(mode, nm, nr_bits+1);
1513}
1514
c93e0f6c
HC
1515COMPAT_SYSCALL_DEFINE6(mbind, compat_ulong_t, start, compat_ulong_t, len,
1516 compat_ulong_t, mode, compat_ulong_t __user *, nmask,
1517 compat_ulong_t, maxnode, compat_ulong_t, flags)
1da177e4
LT
1518{
1519 long err = 0;
1520 unsigned long __user *nm = NULL;
1521 unsigned long nr_bits, alloc_size;
dfcd3c0d 1522 nodemask_t bm;
1da177e4
LT
1523
1524 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1525 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1526
1527 if (nmask) {
dfcd3c0d 1528 err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
1da177e4 1529 nm = compat_alloc_user_space(alloc_size);
dfcd3c0d 1530 err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
1da177e4
LT
1531 }
1532
1533 if (err)
1534 return -EFAULT;
1535
1536 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1537}
1538
1539#endif
1540
74d2c3a0
ON
1541struct mempolicy *__get_vma_policy(struct vm_area_struct *vma,
1542 unsigned long addr)
1da177e4 1543{
8d90274b 1544 struct mempolicy *pol = NULL;
1da177e4
LT
1545
1546 if (vma) {
480eccf9 1547 if (vma->vm_ops && vma->vm_ops->get_policy) {
8d90274b 1548 pol = vma->vm_ops->get_policy(vma, addr);
00442ad0 1549 } else if (vma->vm_policy) {
1da177e4 1550 pol = vma->vm_policy;
00442ad0
MG
1551
1552 /*
1553 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1554 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1555 * count on these policies which will be dropped by
1556 * mpol_cond_put() later
1557 */
1558 if (mpol_needs_cond_ref(pol))
1559 mpol_get(pol);
1560 }
1da177e4 1561 }
f15ca78e 1562
74d2c3a0
ON
1563 return pol;
1564}
1565
1566/*
dd6eecb9 1567 * get_vma_policy(@vma, @addr)
74d2c3a0
ON
1568 * @vma: virtual memory area whose policy is sought
1569 * @addr: address in @vma for shared policy lookup
1570 *
1571 * Returns effective policy for a VMA at specified address.
dd6eecb9 1572 * Falls back to current->mempolicy or system default policy, as necessary.
74d2c3a0
ON
1573 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1574 * count--added by the get_policy() vm_op, as appropriate--to protect against
1575 * freeing by another task. It is the caller's responsibility to free the
1576 * extra reference for shared policies.
1577 */
dd6eecb9
ON
1578static struct mempolicy *get_vma_policy(struct vm_area_struct *vma,
1579 unsigned long addr)
74d2c3a0
ON
1580{
1581 struct mempolicy *pol = __get_vma_policy(vma, addr);
1582
8d90274b 1583 if (!pol)
dd6eecb9 1584 pol = get_task_policy(current);
8d90274b 1585
1da177e4
LT
1586 return pol;
1587}
1588
6b6482bb 1589bool vma_policy_mof(struct vm_area_struct *vma)
fc314724 1590{
6b6482bb 1591 struct mempolicy *pol;
fc314724 1592
6b6482bb
ON
1593 if (vma->vm_ops && vma->vm_ops->get_policy) {
1594 bool ret = false;
fc314724 1595
6b6482bb
ON
1596 pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1597 if (pol && (pol->flags & MPOL_F_MOF))
1598 ret = true;
1599 mpol_cond_put(pol);
8d90274b 1600
6b6482bb 1601 return ret;
fc314724
MG
1602 }
1603
6b6482bb 1604 pol = vma->vm_policy;
8d90274b 1605 if (!pol)
6b6482bb 1606 pol = get_task_policy(current);
8d90274b 1607
fc314724
MG
1608 return pol->flags & MPOL_F_MOF;
1609}
1610
d3eb1570
LJ
1611static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1612{
1613 enum zone_type dynamic_policy_zone = policy_zone;
1614
1615 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1616
1617 /*
1618 * if policy->v.nodes has movable memory only,
1619 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1620 *
1621 * policy->v.nodes is intersect with node_states[N_MEMORY].
1622 * so if the following test faile, it implies
1623 * policy->v.nodes has movable memory only.
1624 */
1625 if (!nodes_intersects(policy->v.nodes, node_states[N_HIGH_MEMORY]))
1626 dynamic_policy_zone = ZONE_MOVABLE;
1627
1628 return zone >= dynamic_policy_zone;
1629}
1630
52cd3b07
LS
1631/*
1632 * Return a nodemask representing a mempolicy for filtering nodes for
1633 * page allocation
1634 */
1635static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
19770b32
MG
1636{
1637 /* Lower zones don't get a nodemask applied for MPOL_BIND */
45c4745a 1638 if (unlikely(policy->mode == MPOL_BIND) &&
d3eb1570 1639 apply_policy_zone(policy, gfp_zone(gfp)) &&
19770b32
MG
1640 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1641 return &policy->v.nodes;
1642
1643 return NULL;
1644}
1645
52cd3b07 1646/* Return a zonelist indicated by gfp for node representing a mempolicy */
2f5f9486
AK
1647static struct zonelist *policy_zonelist(gfp_t gfp, struct mempolicy *policy,
1648 int nd)
1da177e4 1649{
45c4745a 1650 switch (policy->mode) {
1da177e4 1651 case MPOL_PREFERRED:
fc36b8d3
LS
1652 if (!(policy->flags & MPOL_F_LOCAL))
1653 nd = policy->v.preferred_node;
1da177e4
LT
1654 break;
1655 case MPOL_BIND:
19770b32 1656 /*
52cd3b07
LS
1657 * Normally, MPOL_BIND allocations are node-local within the
1658 * allowed nodemask. However, if __GFP_THISNODE is set and the
6eb27e1f 1659 * current node isn't part of the mask, we use the zonelist for
52cd3b07 1660 * the first node in the mask instead.
19770b32 1661 */
19770b32
MG
1662 if (unlikely(gfp & __GFP_THISNODE) &&
1663 unlikely(!node_isset(nd, policy->v.nodes)))
1664 nd = first_node(policy->v.nodes);
1665 break;
1da177e4 1666 default:
1da177e4
LT
1667 BUG();
1668 }
0e88460d 1669 return node_zonelist(nd, gfp);
1da177e4
LT
1670}
1671
1672/* Do dynamic interleaving for a process */
1673static unsigned interleave_nodes(struct mempolicy *policy)
1674{
1675 unsigned nid, next;
1676 struct task_struct *me = current;
1677
1678 nid = me->il_next;
dfcd3c0d 1679 next = next_node(nid, policy->v.nodes);
1da177e4 1680 if (next >= MAX_NUMNODES)
dfcd3c0d 1681 next = first_node(policy->v.nodes);
f5b087b5
DR
1682 if (next < MAX_NUMNODES)
1683 me->il_next = next;
1da177e4
LT
1684 return nid;
1685}
1686
dc85da15
CL
1687/*
1688 * Depending on the memory policy provide a node from which to allocate the
1689 * next slab entry.
1690 */
2a389610 1691unsigned int mempolicy_slab_node(void)
dc85da15 1692{
e7b691b0 1693 struct mempolicy *policy;
2a389610 1694 int node = numa_mem_id();
e7b691b0
AK
1695
1696 if (in_interrupt())
2a389610 1697 return node;
e7b691b0
AK
1698
1699 policy = current->mempolicy;
fc36b8d3 1700 if (!policy || policy->flags & MPOL_F_LOCAL)
2a389610 1701 return node;
bea904d5
LS
1702
1703 switch (policy->mode) {
1704 case MPOL_PREFERRED:
fc36b8d3
LS
1705 /*
1706 * handled MPOL_F_LOCAL above
1707 */
1708 return policy->v.preferred_node;
765c4507 1709
dc85da15
CL
1710 case MPOL_INTERLEAVE:
1711 return interleave_nodes(policy);
1712
dd1a239f 1713 case MPOL_BIND: {
dc85da15
CL
1714 /*
1715 * Follow bind policy behavior and start allocation at the
1716 * first node.
1717 */
19770b32
MG
1718 struct zonelist *zonelist;
1719 struct zone *zone;
1720 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
2a389610 1721 zonelist = &NODE_DATA(node)->node_zonelists[0];
19770b32
MG
1722 (void)first_zones_zonelist(zonelist, highest_zoneidx,
1723 &policy->v.nodes,
1724 &zone);
2a389610 1725 return zone ? zone->node : node;
dd1a239f 1726 }
dc85da15 1727
dc85da15 1728 default:
bea904d5 1729 BUG();
dc85da15
CL
1730 }
1731}
1732
1da177e4
LT
1733/* Do static interleaving for a VMA with known offset. */
1734static unsigned offset_il_node(struct mempolicy *pol,
1735 struct vm_area_struct *vma, unsigned long off)
1736{
dfcd3c0d 1737 unsigned nnodes = nodes_weight(pol->v.nodes);
f5b087b5 1738 unsigned target;
1da177e4 1739 int c;
b76ac7e7 1740 int nid = NUMA_NO_NODE;
1da177e4 1741
f5b087b5
DR
1742 if (!nnodes)
1743 return numa_node_id();
1744 target = (unsigned int)off % nnodes;
1da177e4
LT
1745 c = 0;
1746 do {
dfcd3c0d 1747 nid = next_node(nid, pol->v.nodes);
1da177e4
LT
1748 c++;
1749 } while (c <= target);
1da177e4
LT
1750 return nid;
1751}
1752
5da7ca86
CL
1753/* Determine a node number for interleave */
1754static inline unsigned interleave_nid(struct mempolicy *pol,
1755 struct vm_area_struct *vma, unsigned long addr, int shift)
1756{
1757 if (vma) {
1758 unsigned long off;
1759
3b98b087
NA
1760 /*
1761 * for small pages, there is no difference between
1762 * shift and PAGE_SHIFT, so the bit-shift is safe.
1763 * for huge pages, since vm_pgoff is in units of small
1764 * pages, we need to shift off the always 0 bits to get
1765 * a useful offset.
1766 */
1767 BUG_ON(shift < PAGE_SHIFT);
1768 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
5da7ca86
CL
1769 off += (addr - vma->vm_start) >> shift;
1770 return offset_il_node(pol, vma, off);
1771 } else
1772 return interleave_nodes(pol);
1773}
1774
778d3b0f
MH
1775/*
1776 * Return the bit number of a random bit set in the nodemask.
b76ac7e7 1777 * (returns NUMA_NO_NODE if nodemask is empty)
778d3b0f
MH
1778 */
1779int node_random(const nodemask_t *maskp)
1780{
b76ac7e7 1781 int w, bit = NUMA_NO_NODE;
778d3b0f
MH
1782
1783 w = nodes_weight(*maskp);
1784 if (w)
1785 bit = bitmap_ord_to_pos(maskp->bits,
1786 get_random_int() % w, MAX_NUMNODES);
1787 return bit;
1788}
1789
00ac59ad 1790#ifdef CONFIG_HUGETLBFS
480eccf9
LS
1791/*
1792 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
b46e14ac
FF
1793 * @vma: virtual memory area whose policy is sought
1794 * @addr: address in @vma for shared policy lookup and interleave policy
1795 * @gfp_flags: for requested zone
1796 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
1797 * @nodemask: pointer to nodemask pointer for MPOL_BIND nodemask
480eccf9 1798 *
52cd3b07
LS
1799 * Returns a zonelist suitable for a huge page allocation and a pointer
1800 * to the struct mempolicy for conditional unref after allocation.
1801 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1802 * @nodemask for filtering the zonelist.
c0ff7453 1803 *
d26914d1 1804 * Must be protected by read_mems_allowed_begin()
480eccf9 1805 */
396faf03 1806struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
19770b32
MG
1807 gfp_t gfp_flags, struct mempolicy **mpol,
1808 nodemask_t **nodemask)
5da7ca86 1809{
480eccf9 1810 struct zonelist *zl;
5da7ca86 1811
dd6eecb9 1812 *mpol = get_vma_policy(vma, addr);
19770b32 1813 *nodemask = NULL; /* assume !MPOL_BIND */
5da7ca86 1814
52cd3b07
LS
1815 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1816 zl = node_zonelist(interleave_nid(*mpol, vma, addr,
a5516438 1817 huge_page_shift(hstate_vma(vma))), gfp_flags);
52cd3b07 1818 } else {
2f5f9486 1819 zl = policy_zonelist(gfp_flags, *mpol, numa_node_id());
52cd3b07
LS
1820 if ((*mpol)->mode == MPOL_BIND)
1821 *nodemask = &(*mpol)->v.nodes;
480eccf9
LS
1822 }
1823 return zl;
5da7ca86 1824}
06808b08
LS
1825
1826/*
1827 * init_nodemask_of_mempolicy
1828 *
1829 * If the current task's mempolicy is "default" [NULL], return 'false'
1830 * to indicate default policy. Otherwise, extract the policy nodemask
1831 * for 'bind' or 'interleave' policy into the argument nodemask, or
1832 * initialize the argument nodemask to contain the single node for
1833 * 'preferred' or 'local' policy and return 'true' to indicate presence
1834 * of non-default mempolicy.
1835 *
1836 * We don't bother with reference counting the mempolicy [mpol_get/put]
1837 * because the current task is examining it's own mempolicy and a task's
1838 * mempolicy is only ever changed by the task itself.
1839 *
1840 * N.B., it is the caller's responsibility to free a returned nodemask.
1841 */
1842bool init_nodemask_of_mempolicy(nodemask_t *mask)
1843{
1844 struct mempolicy *mempolicy;
1845 int nid;
1846
1847 if (!(mask && current->mempolicy))
1848 return false;
1849
c0ff7453 1850 task_lock(current);
06808b08
LS
1851 mempolicy = current->mempolicy;
1852 switch (mempolicy->mode) {
1853 case MPOL_PREFERRED:
1854 if (mempolicy->flags & MPOL_F_LOCAL)
1855 nid = numa_node_id();
1856 else
1857 nid = mempolicy->v.preferred_node;
1858 init_nodemask_of_node(mask, nid);
1859 break;
1860
1861 case MPOL_BIND:
1862 /* Fall through */
1863 case MPOL_INTERLEAVE:
1864 *mask = mempolicy->v.nodes;
1865 break;
1866
1867 default:
1868 BUG();
1869 }
c0ff7453 1870 task_unlock(current);
06808b08
LS
1871
1872 return true;
1873}
00ac59ad 1874#endif
5da7ca86 1875
6f48d0eb
DR
1876/*
1877 * mempolicy_nodemask_intersects
1878 *
1879 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1880 * policy. Otherwise, check for intersection between mask and the policy
1881 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1882 * policy, always return true since it may allocate elsewhere on fallback.
1883 *
1884 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1885 */
1886bool mempolicy_nodemask_intersects(struct task_struct *tsk,
1887 const nodemask_t *mask)
1888{
1889 struct mempolicy *mempolicy;
1890 bool ret = true;
1891
1892 if (!mask)
1893 return ret;
1894 task_lock(tsk);
1895 mempolicy = tsk->mempolicy;
1896 if (!mempolicy)
1897 goto out;
1898
1899 switch (mempolicy->mode) {
1900 case MPOL_PREFERRED:
1901 /*
1902 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1903 * allocate from, they may fallback to other nodes when oom.
1904 * Thus, it's possible for tsk to have allocated memory from
1905 * nodes in mask.
1906 */
1907 break;
1908 case MPOL_BIND:
1909 case MPOL_INTERLEAVE:
1910 ret = nodes_intersects(mempolicy->v.nodes, *mask);
1911 break;
1912 default:
1913 BUG();
1914 }
1915out:
1916 task_unlock(tsk);
1917 return ret;
1918}
1919
1da177e4
LT
1920/* Allocate a page in interleaved policy.
1921 Own path because it needs to do special accounting. */
662f3a0b
AK
1922static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1923 unsigned nid)
1da177e4
LT
1924{
1925 struct zonelist *zl;
1926 struct page *page;
1927
0e88460d 1928 zl = node_zonelist(nid, gfp);
1da177e4 1929 page = __alloc_pages(gfp, order, zl);
dd1a239f 1930 if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0]))
ca889e6c 1931 inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
1da177e4
LT
1932 return page;
1933}
1934
1935/**
0bbbc0b3 1936 * alloc_pages_vma - Allocate a page for a VMA.
1da177e4
LT
1937 *
1938 * @gfp:
1939 * %GFP_USER user allocation.
1940 * %GFP_KERNEL kernel allocations,
1941 * %GFP_HIGHMEM highmem/user allocations,
1942 * %GFP_FS allocation should not call back into a file system.
1943 * %GFP_ATOMIC don't sleep.
1944 *
0bbbc0b3 1945 * @order:Order of the GFP allocation.
1da177e4
LT
1946 * @vma: Pointer to VMA or NULL if not available.
1947 * @addr: Virtual Address of the allocation. Must be inside the VMA.
be97a41b
VB
1948 * @node: Which node to prefer for allocation (modulo policy).
1949 * @hugepage: for hugepages try only the preferred node if possible
1da177e4
LT
1950 *
1951 * This function allocates a page from the kernel page pool and applies
1952 * a NUMA policy associated with the VMA or the current process.
1953 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1954 * mm_struct of the VMA to prevent it from going away. Should be used for
be97a41b
VB
1955 * all allocations for pages that will be mapped into user space. Returns
1956 * NULL when no page can be allocated.
1da177e4
LT
1957 */
1958struct page *
0bbbc0b3 1959alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
be97a41b 1960 unsigned long addr, int node, bool hugepage)
1da177e4 1961{
cc9a6c87 1962 struct mempolicy *pol;
c0ff7453 1963 struct page *page;
cc9a6c87 1964 unsigned int cpuset_mems_cookie;
be97a41b
VB
1965 struct zonelist *zl;
1966 nodemask_t *nmask;
cc9a6c87
MG
1967
1968retry_cpuset:
dd6eecb9 1969 pol = get_vma_policy(vma, addr);
d26914d1 1970 cpuset_mems_cookie = read_mems_allowed_begin();
1da177e4 1971
0867a57c
VB
1972 if (pol->mode == MPOL_INTERLEAVE) {
1973 unsigned nid;
1974
1975 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
1976 mpol_cond_put(pol);
1977 page = alloc_page_interleave(gfp, order, nid);
1978 goto out;
1979 }
1980
1981 if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && hugepage)) {
1982 int hpage_node = node;
1983
be97a41b
VB
1984 /*
1985 * For hugepage allocation and non-interleave policy which
0867a57c
VB
1986 * allows the current node (or other explicitly preferred
1987 * node) we only try to allocate from the current/preferred
1988 * node and don't fall back to other nodes, as the cost of
1989 * remote accesses would likely offset THP benefits.
be97a41b
VB
1990 *
1991 * If the policy is interleave, or does not allow the current
1992 * node in its nodemask, we allocate the standard way.
1993 */
0867a57c
VB
1994 if (pol->mode == MPOL_PREFERRED &&
1995 !(pol->flags & MPOL_F_LOCAL))
1996 hpage_node = pol->v.preferred_node;
1997
be97a41b 1998 nmask = policy_nodemask(gfp, pol);
0867a57c 1999 if (!nmask || node_isset(hpage_node, *nmask)) {
be97a41b 2000 mpol_cond_put(pol);
96db800f 2001 page = __alloc_pages_node(hpage_node,
5265047a 2002 gfp | __GFP_THISNODE, order);
be97a41b
VB
2003 goto out;
2004 }
2005 }
2006
be97a41b
VB
2007 nmask = policy_nodemask(gfp, pol);
2008 zl = policy_zonelist(gfp, pol, node);
2386740d 2009 mpol_cond_put(pol);
be97a41b
VB
2010 page = __alloc_pages_nodemask(gfp, order, zl, nmask);
2011out:
d26914d1 2012 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
cc9a6c87 2013 goto retry_cpuset;
c0ff7453 2014 return page;
1da177e4
LT
2015}
2016
2017/**
2018 * alloc_pages_current - Allocate pages.
2019 *
2020 * @gfp:
2021 * %GFP_USER user allocation,
2022 * %GFP_KERNEL kernel allocation,
2023 * %GFP_HIGHMEM highmem allocation,
2024 * %GFP_FS don't call back into a file system.
2025 * %GFP_ATOMIC don't sleep.
2026 * @order: Power of two of allocation size in pages. 0 is a single page.
2027 *
2028 * Allocate a page from the kernel page pool. When not in
2029 * interrupt context and apply the current process NUMA policy.
2030 * Returns NULL when no page can be allocated.
2031 *
cf2a473c 2032 * Don't call cpuset_update_task_memory_state() unless
1da177e4
LT
2033 * 1) it's ok to take cpuset_sem (can WAIT), and
2034 * 2) allocating for current task (not interrupt).
2035 */
dd0fc66f 2036struct page *alloc_pages_current(gfp_t gfp, unsigned order)
1da177e4 2037{
8d90274b 2038 struct mempolicy *pol = &default_policy;
c0ff7453 2039 struct page *page;
cc9a6c87 2040 unsigned int cpuset_mems_cookie;
1da177e4 2041
8d90274b
ON
2042 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2043 pol = get_task_policy(current);
52cd3b07 2044
cc9a6c87 2045retry_cpuset:
d26914d1 2046 cpuset_mems_cookie = read_mems_allowed_begin();
cc9a6c87 2047
52cd3b07
LS
2048 /*
2049 * No reference counting needed for current->mempolicy
2050 * nor system default_policy
2051 */
45c4745a 2052 if (pol->mode == MPOL_INTERLEAVE)
c0ff7453
MX
2053 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2054 else
2055 page = __alloc_pages_nodemask(gfp, order,
5c4b4be3
AK
2056 policy_zonelist(gfp, pol, numa_node_id()),
2057 policy_nodemask(gfp, pol));
cc9a6c87 2058
d26914d1 2059 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
cc9a6c87
MG
2060 goto retry_cpuset;
2061
c0ff7453 2062 return page;
1da177e4
LT
2063}
2064EXPORT_SYMBOL(alloc_pages_current);
2065
ef0855d3
ON
2066int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2067{
2068 struct mempolicy *pol = mpol_dup(vma_policy(src));
2069
2070 if (IS_ERR(pol))
2071 return PTR_ERR(pol);
2072 dst->vm_policy = pol;
2073 return 0;
2074}
2075
4225399a 2076/*
846a16bf 2077 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
4225399a
PJ
2078 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2079 * with the mems_allowed returned by cpuset_mems_allowed(). This
2080 * keeps mempolicies cpuset relative after its cpuset moves. See
2081 * further kernel/cpuset.c update_nodemask().
708c1bbc
MX
2082 *
2083 * current's mempolicy may be rebinded by the other task(the task that changes
2084 * cpuset's mems), so we needn't do rebind work for current task.
4225399a 2085 */
4225399a 2086
846a16bf
LS
2087/* Slow path of a mempolicy duplicate */
2088struct mempolicy *__mpol_dup(struct mempolicy *old)
1da177e4
LT
2089{
2090 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2091
2092 if (!new)
2093 return ERR_PTR(-ENOMEM);
708c1bbc
MX
2094
2095 /* task's mempolicy is protected by alloc_lock */
2096 if (old == current->mempolicy) {
2097 task_lock(current);
2098 *new = *old;
2099 task_unlock(current);
2100 } else
2101 *new = *old;
2102
4225399a
PJ
2103 if (current_cpuset_is_being_rebound()) {
2104 nodemask_t mems = cpuset_mems_allowed(current);
708c1bbc
MX
2105 if (new->flags & MPOL_F_REBINDING)
2106 mpol_rebind_policy(new, &mems, MPOL_REBIND_STEP2);
2107 else
2108 mpol_rebind_policy(new, &mems, MPOL_REBIND_ONCE);
4225399a 2109 }
1da177e4 2110 atomic_set(&new->refcnt, 1);
1da177e4
LT
2111 return new;
2112}
2113
2114/* Slow path of a mempolicy comparison */
fcfb4dcc 2115bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
1da177e4
LT
2116{
2117 if (!a || !b)
fcfb4dcc 2118 return false;
45c4745a 2119 if (a->mode != b->mode)
fcfb4dcc 2120 return false;
19800502 2121 if (a->flags != b->flags)
fcfb4dcc 2122 return false;
19800502
BL
2123 if (mpol_store_user_nodemask(a))
2124 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
fcfb4dcc 2125 return false;
19800502 2126
45c4745a 2127 switch (a->mode) {
19770b32
MG
2128 case MPOL_BIND:
2129 /* Fall through */
1da177e4 2130 case MPOL_INTERLEAVE:
fcfb4dcc 2131 return !!nodes_equal(a->v.nodes, b->v.nodes);
1da177e4 2132 case MPOL_PREFERRED:
75719661 2133 return a->v.preferred_node == b->v.preferred_node;
1da177e4
LT
2134 default:
2135 BUG();
fcfb4dcc 2136 return false;
1da177e4
LT
2137 }
2138}
2139
1da177e4
LT
2140/*
2141 * Shared memory backing store policy support.
2142 *
2143 * Remember policies even when nobody has shared memory mapped.
2144 * The policies are kept in Red-Black tree linked from the inode.
2145 * They are protected by the sp->lock spinlock, which should be held
2146 * for any accesses to the tree.
2147 */
2148
2149/* lookup first element intersecting start-end */
42288fe3 2150/* Caller holds sp->lock */
1da177e4
LT
2151static struct sp_node *
2152sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2153{
2154 struct rb_node *n = sp->root.rb_node;
2155
2156 while (n) {
2157 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2158
2159 if (start >= p->end)
2160 n = n->rb_right;
2161 else if (end <= p->start)
2162 n = n->rb_left;
2163 else
2164 break;
2165 }
2166 if (!n)
2167 return NULL;
2168 for (;;) {
2169 struct sp_node *w = NULL;
2170 struct rb_node *prev = rb_prev(n);
2171 if (!prev)
2172 break;
2173 w = rb_entry(prev, struct sp_node, nd);
2174 if (w->end <= start)
2175 break;
2176 n = prev;
2177 }
2178 return rb_entry(n, struct sp_node, nd);
2179}
2180
2181/* Insert a new shared policy into the list. */
2182/* Caller holds sp->lock */
2183static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2184{
2185 struct rb_node **p = &sp->root.rb_node;
2186 struct rb_node *parent = NULL;
2187 struct sp_node *nd;
2188
2189 while (*p) {
2190 parent = *p;
2191 nd = rb_entry(parent, struct sp_node, nd);
2192 if (new->start < nd->start)
2193 p = &(*p)->rb_left;
2194 else if (new->end > nd->end)
2195 p = &(*p)->rb_right;
2196 else
2197 BUG();
2198 }
2199 rb_link_node(&new->nd, parent, p);
2200 rb_insert_color(&new->nd, &sp->root);
140d5a49 2201 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
45c4745a 2202 new->policy ? new->policy->mode : 0);
1da177e4
LT
2203}
2204
2205/* Find shared policy intersecting idx */
2206struct mempolicy *
2207mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2208{
2209 struct mempolicy *pol = NULL;
2210 struct sp_node *sn;
2211
2212 if (!sp->root.rb_node)
2213 return NULL;
42288fe3 2214 spin_lock(&sp->lock);
1da177e4
LT
2215 sn = sp_lookup(sp, idx, idx+1);
2216 if (sn) {
2217 mpol_get(sn->policy);
2218 pol = sn->policy;
2219 }
42288fe3 2220 spin_unlock(&sp->lock);
1da177e4
LT
2221 return pol;
2222}
2223
63f74ca2
KM
2224static void sp_free(struct sp_node *n)
2225{
2226 mpol_put(n->policy);
2227 kmem_cache_free(sn_cache, n);
2228}
2229
771fb4d8
LS
2230/**
2231 * mpol_misplaced - check whether current page node is valid in policy
2232 *
b46e14ac
FF
2233 * @page: page to be checked
2234 * @vma: vm area where page mapped
2235 * @addr: virtual address where page mapped
771fb4d8
LS
2236 *
2237 * Lookup current policy node id for vma,addr and "compare to" page's
2238 * node id.
2239 *
2240 * Returns:
2241 * -1 - not misplaced, page is in the right node
2242 * node - node id where the page should be
2243 *
2244 * Policy determination "mimics" alloc_page_vma().
2245 * Called from fault path where we know the vma and faulting address.
2246 */
2247int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2248{
2249 struct mempolicy *pol;
2250 struct zone *zone;
2251 int curnid = page_to_nid(page);
2252 unsigned long pgoff;
90572890
PZ
2253 int thiscpu = raw_smp_processor_id();
2254 int thisnid = cpu_to_node(thiscpu);
771fb4d8
LS
2255 int polnid = -1;
2256 int ret = -1;
2257
2258 BUG_ON(!vma);
2259
dd6eecb9 2260 pol = get_vma_policy(vma, addr);
771fb4d8
LS
2261 if (!(pol->flags & MPOL_F_MOF))
2262 goto out;
2263
2264 switch (pol->mode) {
2265 case MPOL_INTERLEAVE:
2266 BUG_ON(addr >= vma->vm_end);
2267 BUG_ON(addr < vma->vm_start);
2268
2269 pgoff = vma->vm_pgoff;
2270 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2271 polnid = offset_il_node(pol, vma, pgoff);
2272 break;
2273
2274 case MPOL_PREFERRED:
2275 if (pol->flags & MPOL_F_LOCAL)
2276 polnid = numa_node_id();
2277 else
2278 polnid = pol->v.preferred_node;
2279 break;
2280
2281 case MPOL_BIND:
2282 /*
2283 * allows binding to multiple nodes.
2284 * use current page if in policy nodemask,
2285 * else select nearest allowed node, if any.
2286 * If no allowed nodes, use current [!misplaced].
2287 */
2288 if (node_isset(curnid, pol->v.nodes))
2289 goto out;
2290 (void)first_zones_zonelist(
2291 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2292 gfp_zone(GFP_HIGHUSER),
2293 &pol->v.nodes, &zone);
2294 polnid = zone->node;
2295 break;
2296
2297 default:
2298 BUG();
2299 }
5606e387
MG
2300
2301 /* Migrate the page towards the node whose CPU is referencing it */
e42c8ff2 2302 if (pol->flags & MPOL_F_MORON) {
90572890 2303 polnid = thisnid;
5606e387 2304
10f39042 2305 if (!should_numa_migrate_memory(current, page, curnid, thiscpu))
de1c9ce6 2306 goto out;
e42c8ff2
MG
2307 }
2308
771fb4d8
LS
2309 if (curnid != polnid)
2310 ret = polnid;
2311out:
2312 mpol_cond_put(pol);
2313
2314 return ret;
2315}
2316
1da177e4
LT
2317static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2318{
140d5a49 2319 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
1da177e4 2320 rb_erase(&n->nd, &sp->root);
63f74ca2 2321 sp_free(n);
1da177e4
LT
2322}
2323
42288fe3
MG
2324static void sp_node_init(struct sp_node *node, unsigned long start,
2325 unsigned long end, struct mempolicy *pol)
2326{
2327 node->start = start;
2328 node->end = end;
2329 node->policy = pol;
2330}
2331
dbcb0f19
AB
2332static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2333 struct mempolicy *pol)
1da177e4 2334{
869833f2
KM
2335 struct sp_node *n;
2336 struct mempolicy *newpol;
1da177e4 2337
869833f2 2338 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
1da177e4
LT
2339 if (!n)
2340 return NULL;
869833f2
KM
2341
2342 newpol = mpol_dup(pol);
2343 if (IS_ERR(newpol)) {
2344 kmem_cache_free(sn_cache, n);
2345 return NULL;
2346 }
2347 newpol->flags |= MPOL_F_SHARED;
42288fe3 2348 sp_node_init(n, start, end, newpol);
869833f2 2349
1da177e4
LT
2350 return n;
2351}
2352
2353/* Replace a policy range. */
2354static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2355 unsigned long end, struct sp_node *new)
2356{
b22d127a 2357 struct sp_node *n;
42288fe3
MG
2358 struct sp_node *n_new = NULL;
2359 struct mempolicy *mpol_new = NULL;
b22d127a 2360 int ret = 0;
1da177e4 2361
42288fe3
MG
2362restart:
2363 spin_lock(&sp->lock);
1da177e4
LT
2364 n = sp_lookup(sp, start, end);
2365 /* Take care of old policies in the same range. */
2366 while (n && n->start < end) {
2367 struct rb_node *next = rb_next(&n->nd);
2368 if (n->start >= start) {
2369 if (n->end <= end)
2370 sp_delete(sp, n);
2371 else
2372 n->start = end;
2373 } else {
2374 /* Old policy spanning whole new range. */
2375 if (n->end > end) {
42288fe3
MG
2376 if (!n_new)
2377 goto alloc_new;
2378
2379 *mpol_new = *n->policy;
2380 atomic_set(&mpol_new->refcnt, 1);
7880639c 2381 sp_node_init(n_new, end, n->end, mpol_new);
1da177e4 2382 n->end = start;
5ca39575 2383 sp_insert(sp, n_new);
42288fe3
MG
2384 n_new = NULL;
2385 mpol_new = NULL;
1da177e4
LT
2386 break;
2387 } else
2388 n->end = start;
2389 }
2390 if (!next)
2391 break;
2392 n = rb_entry(next, struct sp_node, nd);
2393 }
2394 if (new)
2395 sp_insert(sp, new);
42288fe3
MG
2396 spin_unlock(&sp->lock);
2397 ret = 0;
2398
2399err_out:
2400 if (mpol_new)
2401 mpol_put(mpol_new);
2402 if (n_new)
2403 kmem_cache_free(sn_cache, n_new);
2404
b22d127a 2405 return ret;
42288fe3
MG
2406
2407alloc_new:
2408 spin_unlock(&sp->lock);
2409 ret = -ENOMEM;
2410 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2411 if (!n_new)
2412 goto err_out;
2413 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2414 if (!mpol_new)
2415 goto err_out;
2416 goto restart;
1da177e4
LT
2417}
2418
71fe804b
LS
2419/**
2420 * mpol_shared_policy_init - initialize shared policy for inode
2421 * @sp: pointer to inode shared policy
2422 * @mpol: struct mempolicy to install
2423 *
2424 * Install non-NULL @mpol in inode's shared policy rb-tree.
2425 * On entry, the current task has a reference on a non-NULL @mpol.
2426 * This must be released on exit.
4bfc4495 2427 * This is called at get_inode() calls and we can use GFP_KERNEL.
71fe804b
LS
2428 */
2429void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2430{
58568d2a
MX
2431 int ret;
2432
71fe804b 2433 sp->root = RB_ROOT; /* empty tree == default mempolicy */
42288fe3 2434 spin_lock_init(&sp->lock);
71fe804b
LS
2435
2436 if (mpol) {
2437 struct vm_area_struct pvma;
2438 struct mempolicy *new;
4bfc4495 2439 NODEMASK_SCRATCH(scratch);
71fe804b 2440
4bfc4495 2441 if (!scratch)
5c0c1654 2442 goto put_mpol;
71fe804b
LS
2443 /* contextualize the tmpfs mount point mempolicy */
2444 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
15d77835 2445 if (IS_ERR(new))
0cae3457 2446 goto free_scratch; /* no valid nodemask intersection */
58568d2a
MX
2447
2448 task_lock(current);
4bfc4495 2449 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
58568d2a 2450 task_unlock(current);
15d77835 2451 if (ret)
5c0c1654 2452 goto put_new;
71fe804b
LS
2453
2454 /* Create pseudo-vma that contains just the policy */
2455 memset(&pvma, 0, sizeof(struct vm_area_struct));
2456 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2457 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
15d77835 2458
5c0c1654 2459put_new:
71fe804b 2460 mpol_put(new); /* drop initial ref */
0cae3457 2461free_scratch:
4bfc4495 2462 NODEMASK_SCRATCH_FREE(scratch);
5c0c1654
LS
2463put_mpol:
2464 mpol_put(mpol); /* drop our incoming ref on sb mpol */
7339ff83
RH
2465 }
2466}
2467
1da177e4
LT
2468int mpol_set_shared_policy(struct shared_policy *info,
2469 struct vm_area_struct *vma, struct mempolicy *npol)
2470{
2471 int err;
2472 struct sp_node *new = NULL;
2473 unsigned long sz = vma_pages(vma);
2474
028fec41 2475 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
1da177e4 2476 vma->vm_pgoff,
45c4745a 2477 sz, npol ? npol->mode : -1,
028fec41 2478 npol ? npol->flags : -1,
00ef2d2f 2479 npol ? nodes_addr(npol->v.nodes)[0] : NUMA_NO_NODE);
1da177e4
LT
2480
2481 if (npol) {
2482 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2483 if (!new)
2484 return -ENOMEM;
2485 }
2486 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2487 if (err && new)
63f74ca2 2488 sp_free(new);
1da177e4
LT
2489 return err;
2490}
2491
2492/* Free a backing policy store on inode delete. */
2493void mpol_free_shared_policy(struct shared_policy *p)
2494{
2495 struct sp_node *n;
2496 struct rb_node *next;
2497
2498 if (!p->root.rb_node)
2499 return;
42288fe3 2500 spin_lock(&p->lock);
1da177e4
LT
2501 next = rb_first(&p->root);
2502 while (next) {
2503 n = rb_entry(next, struct sp_node, nd);
2504 next = rb_next(&n->nd);
63f74ca2 2505 sp_delete(p, n);
1da177e4 2506 }
42288fe3 2507 spin_unlock(&p->lock);
1da177e4
LT
2508}
2509
1a687c2e 2510#ifdef CONFIG_NUMA_BALANCING
c297663c 2511static int __initdata numabalancing_override;
1a687c2e
MG
2512
2513static void __init check_numabalancing_enable(void)
2514{
2515 bool numabalancing_default = false;
2516
2517 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2518 numabalancing_default = true;
2519
c297663c
MG
2520 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2521 if (numabalancing_override)
2522 set_numabalancing_state(numabalancing_override == 1);
2523
b0dc2b9b 2524 if (num_online_nodes() > 1 && !numabalancing_override) {
4a404bea 2525 pr_info("%s automatic NUMA balancing. "
c297663c
MG
2526 "Configure with numa_balancing= or the "
2527 "kernel.numa_balancing sysctl",
2528 numabalancing_default ? "Enabling" : "Disabling");
1a687c2e
MG
2529 set_numabalancing_state(numabalancing_default);
2530 }
2531}
2532
2533static int __init setup_numabalancing(char *str)
2534{
2535 int ret = 0;
2536 if (!str)
2537 goto out;
1a687c2e
MG
2538
2539 if (!strcmp(str, "enable")) {
c297663c 2540 numabalancing_override = 1;
1a687c2e
MG
2541 ret = 1;
2542 } else if (!strcmp(str, "disable")) {
c297663c 2543 numabalancing_override = -1;
1a687c2e
MG
2544 ret = 1;
2545 }
2546out:
2547 if (!ret)
4a404bea 2548 pr_warn("Unable to parse numa_balancing=\n");
1a687c2e
MG
2549
2550 return ret;
2551}
2552__setup("numa_balancing=", setup_numabalancing);
2553#else
2554static inline void __init check_numabalancing_enable(void)
2555{
2556}
2557#endif /* CONFIG_NUMA_BALANCING */
2558
1da177e4
LT
2559/* assumes fs == KERNEL_DS */
2560void __init numa_policy_init(void)
2561{
b71636e2
PM
2562 nodemask_t interleave_nodes;
2563 unsigned long largest = 0;
2564 int nid, prefer = 0;
2565
1da177e4
LT
2566 policy_cache = kmem_cache_create("numa_policy",
2567 sizeof(struct mempolicy),
20c2df83 2568 0, SLAB_PANIC, NULL);
1da177e4
LT
2569
2570 sn_cache = kmem_cache_create("shared_policy_node",
2571 sizeof(struct sp_node),
20c2df83 2572 0, SLAB_PANIC, NULL);
1da177e4 2573
5606e387
MG
2574 for_each_node(nid) {
2575 preferred_node_policy[nid] = (struct mempolicy) {
2576 .refcnt = ATOMIC_INIT(1),
2577 .mode = MPOL_PREFERRED,
2578 .flags = MPOL_F_MOF | MPOL_F_MORON,
2579 .v = { .preferred_node = nid, },
2580 };
2581 }
2582
b71636e2
PM
2583 /*
2584 * Set interleaving policy for system init. Interleaving is only
2585 * enabled across suitably sized nodes (default is >= 16MB), or
2586 * fall back to the largest node if they're all smaller.
2587 */
2588 nodes_clear(interleave_nodes);
01f13bd6 2589 for_each_node_state(nid, N_MEMORY) {
b71636e2
PM
2590 unsigned long total_pages = node_present_pages(nid);
2591
2592 /* Preserve the largest node */
2593 if (largest < total_pages) {
2594 largest = total_pages;
2595 prefer = nid;
2596 }
2597
2598 /* Interleave this node? */
2599 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2600 node_set(nid, interleave_nodes);
2601 }
2602
2603 /* All too small, use the largest */
2604 if (unlikely(nodes_empty(interleave_nodes)))
2605 node_set(prefer, interleave_nodes);
1da177e4 2606
028fec41 2607 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
b1de0d13 2608 pr_err("%s: interleaving failed\n", __func__);
1a687c2e
MG
2609
2610 check_numabalancing_enable();
1da177e4
LT
2611}
2612
8bccd85f 2613/* Reset policy of current process to default */
1da177e4
LT
2614void numa_default_policy(void)
2615{
028fec41 2616 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
1da177e4 2617}
68860ec1 2618
095f1fc4
LS
2619/*
2620 * Parse and format mempolicy from/to strings
2621 */
2622
1a75a6c8 2623/*
f2a07f40 2624 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
1a75a6c8 2625 */
345ace9c
LS
2626static const char * const policy_modes[] =
2627{
2628 [MPOL_DEFAULT] = "default",
2629 [MPOL_PREFERRED] = "prefer",
2630 [MPOL_BIND] = "bind",
2631 [MPOL_INTERLEAVE] = "interleave",
d3a71033 2632 [MPOL_LOCAL] = "local",
345ace9c 2633};
1a75a6c8 2634
095f1fc4
LS
2635
2636#ifdef CONFIG_TMPFS
2637/**
f2a07f40 2638 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
095f1fc4 2639 * @str: string containing mempolicy to parse
71fe804b 2640 * @mpol: pointer to struct mempolicy pointer, returned on success.
095f1fc4
LS
2641 *
2642 * Format of input:
2643 * <mode>[=<flags>][:<nodelist>]
2644 *
71fe804b 2645 * On success, returns 0, else 1
095f1fc4 2646 */
a7a88b23 2647int mpol_parse_str(char *str, struct mempolicy **mpol)
095f1fc4 2648{
71fe804b 2649 struct mempolicy *new = NULL;
b4652e84 2650 unsigned short mode;
f2a07f40 2651 unsigned short mode_flags;
71fe804b 2652 nodemask_t nodes;
095f1fc4
LS
2653 char *nodelist = strchr(str, ':');
2654 char *flags = strchr(str, '=');
095f1fc4
LS
2655 int err = 1;
2656
2657 if (nodelist) {
2658 /* NUL-terminate mode or flags string */
2659 *nodelist++ = '\0';
71fe804b 2660 if (nodelist_parse(nodelist, nodes))
095f1fc4 2661 goto out;
01f13bd6 2662 if (!nodes_subset(nodes, node_states[N_MEMORY]))
095f1fc4 2663 goto out;
71fe804b
LS
2664 } else
2665 nodes_clear(nodes);
2666
095f1fc4
LS
2667 if (flags)
2668 *flags++ = '\0'; /* terminate mode string */
2669
479e2802 2670 for (mode = 0; mode < MPOL_MAX; mode++) {
345ace9c 2671 if (!strcmp(str, policy_modes[mode])) {
095f1fc4
LS
2672 break;
2673 }
2674 }
a720094d 2675 if (mode >= MPOL_MAX)
095f1fc4
LS
2676 goto out;
2677
71fe804b 2678 switch (mode) {
095f1fc4 2679 case MPOL_PREFERRED:
71fe804b
LS
2680 /*
2681 * Insist on a nodelist of one node only
2682 */
095f1fc4
LS
2683 if (nodelist) {
2684 char *rest = nodelist;
2685 while (isdigit(*rest))
2686 rest++;
926f2ae0
KM
2687 if (*rest)
2688 goto out;
095f1fc4
LS
2689 }
2690 break;
095f1fc4
LS
2691 case MPOL_INTERLEAVE:
2692 /*
2693 * Default to online nodes with memory if no nodelist
2694 */
2695 if (!nodelist)
01f13bd6 2696 nodes = node_states[N_MEMORY];
3f226aa1 2697 break;
71fe804b 2698 case MPOL_LOCAL:
3f226aa1 2699 /*
71fe804b 2700 * Don't allow a nodelist; mpol_new() checks flags
3f226aa1 2701 */
71fe804b 2702 if (nodelist)
3f226aa1 2703 goto out;
71fe804b 2704 mode = MPOL_PREFERRED;
3f226aa1 2705 break;
413b43de
RT
2706 case MPOL_DEFAULT:
2707 /*
2708 * Insist on a empty nodelist
2709 */
2710 if (!nodelist)
2711 err = 0;
2712 goto out;
d69b2e63
KM
2713 case MPOL_BIND:
2714 /*
2715 * Insist on a nodelist
2716 */
2717 if (!nodelist)
2718 goto out;
095f1fc4
LS
2719 }
2720
71fe804b 2721 mode_flags = 0;
095f1fc4
LS
2722 if (flags) {
2723 /*
2724 * Currently, we only support two mutually exclusive
2725 * mode flags.
2726 */
2727 if (!strcmp(flags, "static"))
71fe804b 2728 mode_flags |= MPOL_F_STATIC_NODES;
095f1fc4 2729 else if (!strcmp(flags, "relative"))
71fe804b 2730 mode_flags |= MPOL_F_RELATIVE_NODES;
095f1fc4 2731 else
926f2ae0 2732 goto out;
095f1fc4 2733 }
71fe804b
LS
2734
2735 new = mpol_new(mode, mode_flags, &nodes);
2736 if (IS_ERR(new))
926f2ae0
KM
2737 goto out;
2738
f2a07f40
HD
2739 /*
2740 * Save nodes for mpol_to_str() to show the tmpfs mount options
2741 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2742 */
2743 if (mode != MPOL_PREFERRED)
2744 new->v.nodes = nodes;
2745 else if (nodelist)
2746 new->v.preferred_node = first_node(nodes);
2747 else
2748 new->flags |= MPOL_F_LOCAL;
2749
2750 /*
2751 * Save nodes for contextualization: this will be used to "clone"
2752 * the mempolicy in a specific context [cpuset] at a later time.
2753 */
2754 new->w.user_nodemask = nodes;
2755
926f2ae0 2756 err = 0;
71fe804b 2757
095f1fc4
LS
2758out:
2759 /* Restore string for error message */
2760 if (nodelist)
2761 *--nodelist = ':';
2762 if (flags)
2763 *--flags = '=';
71fe804b
LS
2764 if (!err)
2765 *mpol = new;
095f1fc4
LS
2766 return err;
2767}
2768#endif /* CONFIG_TMPFS */
2769
71fe804b
LS
2770/**
2771 * mpol_to_str - format a mempolicy structure for printing
2772 * @buffer: to contain formatted mempolicy string
2773 * @maxlen: length of @buffer
2774 * @pol: pointer to mempolicy to be formatted
71fe804b 2775 *
948927ee
DR
2776 * Convert @pol into a string. If @buffer is too short, truncate the string.
2777 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
2778 * longest flag, "relative", and to display at least a few node ids.
1a75a6c8 2779 */
948927ee 2780void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
1a75a6c8
CL
2781{
2782 char *p = buffer;
948927ee
DR
2783 nodemask_t nodes = NODE_MASK_NONE;
2784 unsigned short mode = MPOL_DEFAULT;
2785 unsigned short flags = 0;
2291990a 2786
8790c71a 2787 if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
bea904d5 2788 mode = pol->mode;
948927ee
DR
2789 flags = pol->flags;
2790 }
bea904d5 2791
1a75a6c8
CL
2792 switch (mode) {
2793 case MPOL_DEFAULT:
1a75a6c8 2794 break;
1a75a6c8 2795 case MPOL_PREFERRED:
fc36b8d3 2796 if (flags & MPOL_F_LOCAL)
f2a07f40 2797 mode = MPOL_LOCAL;
53f2556b 2798 else
fc36b8d3 2799 node_set(pol->v.preferred_node, nodes);
1a75a6c8 2800 break;
1a75a6c8 2801 case MPOL_BIND:
1a75a6c8 2802 case MPOL_INTERLEAVE:
f2a07f40 2803 nodes = pol->v.nodes;
1a75a6c8 2804 break;
1a75a6c8 2805 default:
948927ee
DR
2806 WARN_ON_ONCE(1);
2807 snprintf(p, maxlen, "unknown");
2808 return;
1a75a6c8
CL
2809 }
2810
b7a9f420 2811 p += snprintf(p, maxlen, "%s", policy_modes[mode]);
1a75a6c8 2812
fc36b8d3 2813 if (flags & MPOL_MODE_FLAGS) {
948927ee 2814 p += snprintf(p, buffer + maxlen - p, "=");
f5b087b5 2815
2291990a
LS
2816 /*
2817 * Currently, the only defined flags are mutually exclusive
2818 */
f5b087b5 2819 if (flags & MPOL_F_STATIC_NODES)
2291990a
LS
2820 p += snprintf(p, buffer + maxlen - p, "static");
2821 else if (flags & MPOL_F_RELATIVE_NODES)
2822 p += snprintf(p, buffer + maxlen - p, "relative");
f5b087b5
DR
2823 }
2824
9e763e0f
TH
2825 if (!nodes_empty(nodes))
2826 p += scnprintf(p, buffer + maxlen - p, ":%*pbl",
2827 nodemask_pr_args(&nodes));
1a75a6c8 2828}