1 /*
2 * Simple NUMA memory policy for the Linux kernel.
3 *
4 * Copyright 2003,2004 Andi Kleen, SuSE Labs.
5 * (C) Copyright 2005 Christoph Lameter, Silicon Graphics, Inc.
6 * Subject to the GNU Public License, version 2.
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.
21 *
22 * bind Only allocate memory on a specific set of nodes,
23 * no fallback.
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 *
28 * preferred Try a specific node first before normal fallback.
29 * As a special case NUMA_NO_NODE here means do the allocation
30 * on the local CPU. This is normally identical to default,
31 * but useful to set in a VMA when you have a non default
32 * process policy.
33 *
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.
66 */
67
68 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
69
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>
76 #include <linux/nodemask.h>
77 #include <linux/cpuset.h>
78 #include <linux/slab.h>
79 #include <linux/string.h>
80 #include <linux/export.h>
81 #include <linux/nsproxy.h>
82 #include <linux/interrupt.h>
83 #include <linux/init.h>
84 #include <linux/compat.h>
85 #include <linux/swap.h>
86 #include <linux/seq_file.h>
87 #include <linux/proc_fs.h>
88 #include <linux/migrate.h>
89 #include <linux/ksm.h>
90 #include <linux/rmap.h>
91 #include <linux/security.h>
92 #include <linux/syscalls.h>
93 #include <linux/ctype.h>
94 #include <linux/mm_inline.h>
95 #include <linux/mmu_notifier.h>
96 #include <linux/printk.h>
97
98 #include <asm/tlbflush.h>
99 #include <asm/uaccess.h>
100 #include <linux/random.h>
101
102 #include "internal.h"
103
104 /* Internal flags */
105 #define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */
106 #define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */
107
108 static struct kmem_cache *policy_cache;
109 static struct kmem_cache *sn_cache;
110
111 /* Highest zone. An specific allocation for a zone below that is not
112 policied. */
113 enum zone_type policy_zone = 0;
114
115 /*
116 * run-time system-wide default policy => local allocation
117 */
118 static struct mempolicy default_policy = {
119 .refcnt = ATOMIC_INIT(1), /* never free it */
120 .mode = MPOL_PREFERRED,
121 .flags = MPOL_F_LOCAL,
122 };
123
124 static struct mempolicy preferred_node_policy[MAX_NUMNODES];
125
get_task_policy(struct task_struct * p)126 struct mempolicy *get_task_policy(struct task_struct *p)
127 {
128 struct mempolicy *pol = p->mempolicy;
129 int node;
130
131 if (pol)
132 return pol;
133
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;
140 }
141
142 return &default_policy;
143 }
144
145 static const struct mempolicy_operations {
146 int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
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);
163 } mpol_ops[MPOL_MAX];
164
mpol_store_user_nodemask(const struct mempolicy * pol)165 static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
166 {
167 return pol->flags & MPOL_MODE_FLAGS;
168 }
169
mpol_relative_nodemask(nodemask_t * ret,const nodemask_t * orig,const nodemask_t * rel)170 static 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);
176 }
177
mpol_new_interleave(struct mempolicy * pol,const nodemask_t * nodes)178 static 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
mpol_new_preferred(struct mempolicy * pol,const nodemask_t * nodes)186 static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
187 {
188 if (!nodes)
189 pol->flags |= MPOL_F_LOCAL; /* local allocation */
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
mpol_new_bind(struct mempolicy * pol,const nodemask_t * nodes)197 static int mpol_new_bind(struct mempolicy *pol, const nodemask_t *nodes)
198 {
199 if (nodes_empty(*nodes))
200 return -EINVAL;
201 pol->v.nodes = *nodes;
202 return 0;
203 }
204
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 */
mpol_set_nodemask(struct mempolicy * pol,const nodemask_t * nodes,struct nodemask_scratch * nsc)214 static int mpol_set_nodemask(struct mempolicy *pol,
215 const nodemask_t *nodes, struct nodemask_scratch *nsc)
216 {
217 int ret;
218
219 /* if mode is MPOL_DEFAULT, pol is NULL. This is right. */
220 if (pol == NULL)
221 return 0;
222 /* Check N_MEMORY */
223 nodes_and(nsc->mask1,
224 cpuset_current_mems_allowed, node_states[N_MEMORY]);
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)
231 mpol_relative_nodemask(&nsc->mask2, nodes, &nsc->mask1);
232 else
233 nodes_and(nsc->mask2, *nodes, nsc->mask1);
234
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
242 if (nodes)
243 ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
244 else
245 ret = mpol_ops[pol->mode].create(pol, NULL);
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 */
mpol_new(unsigned short mode,unsigned short flags,nodemask_t * nodes)253 static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
254 nodemask_t *nodes)
255 {
256 struct mempolicy *policy;
257
258 pr_debug("setting mode %d flags %d nodes[0] %lx\n",
259 mode, flags, nodes ? nodes_addr(*nodes)[0] : NUMA_NO_NODE);
260
261 if (mode == MPOL_DEFAULT) {
262 if (nodes && !nodes_empty(*nodes))
263 return ERR_PTR(-EINVAL);
264 return NULL;
265 }
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);
278 }
279 } else if (mode == MPOL_LOCAL) {
280 if (!nodes_empty(*nodes))
281 return ERR_PTR(-EINVAL);
282 mode = MPOL_PREFERRED;
283 } else if (nodes_empty(*nodes))
284 return ERR_PTR(-EINVAL);
285 policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
286 if (!policy)
287 return ERR_PTR(-ENOMEM);
288 atomic_set(&policy->refcnt, 1);
289 policy->mode = mode;
290 policy->flags = flags;
291
292 return policy;
293 }
294
295 /* Slow path of a mpol destructor. */
__mpol_put(struct mempolicy * p)296 void __mpol_put(struct mempolicy *p)
297 {
298 if (!atomic_dec_and_test(&p->refcnt))
299 return;
300 kmem_cache_free(policy_cache, p);
301 }
302
mpol_rebind_default(struct mempolicy * pol,const nodemask_t * nodes,enum mpol_rebind_step step)303 static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes,
304 enum mpol_rebind_step step)
305 {
306 }
307
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 */
mpol_rebind_nodemask(struct mempolicy * pol,const nodemask_t * nodes,enum mpol_rebind_step step)314 static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes,
315 enum mpol_rebind_step step)
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 {
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();
337 }
338
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
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
mpol_rebind_preferred(struct mempolicy * pol,const nodemask_t * nodes,enum mpol_rebind_step step)358 static void mpol_rebind_preferred(struct mempolicy *pol,
359 const nodemask_t *nodes,
360 enum mpol_rebind_step step)
361 {
362 nodemask_t tmp;
363
364 if (pol->flags & MPOL_F_STATIC_NODES) {
365 int node = first_node(pol->w.user_nodemask);
366
367 if (node_isset(node, *nodes)) {
368 pol->v.preferred_node = node;
369 pol->flags &= ~MPOL_F_LOCAL;
370 } else
371 pol->flags |= MPOL_F_LOCAL;
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);
375 } else if (!(pol->flags & MPOL_F_LOCAL)) {
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 }
381 }
382
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 */
mpol_rebind_policy(struct mempolicy * pol,const nodemask_t * newmask,enum mpol_rebind_step step)399 static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask,
400 enum mpol_rebind_step step)
401 {
402 if (!pol)
403 return;
404 if (!mpol_store_user_nodemask(pol) && step == MPOL_REBIND_ONCE &&
405 nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
406 return;
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);
422 }
423
424 /*
425 * Wrapper for mpol_rebind_policy() that just requires task
426 * pointer, and updates task mempolicy.
427 *
428 * Called with task's alloc_lock held.
429 */
430
mpol_rebind_task(struct task_struct * tsk,const nodemask_t * new,enum mpol_rebind_step step)431 void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new,
432 enum mpol_rebind_step step)
433 {
434 mpol_rebind_policy(tsk->mempolicy, new, step);
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
mpol_rebind_mm(struct mm_struct * mm,nodemask_t * new)443 void 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)
449 mpol_rebind_policy(vma->vm_policy, new, MPOL_REBIND_ONCE);
450 up_write(&mm->mmap_sem);
451 }
452
453 static 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
471 static void migrate_page_add(struct page *page, struct list_head *pagelist,
472 unsigned long flags);
473
474 struct queue_pages {
475 struct list_head *pagelist;
476 unsigned long flags;
477 nodemask_t *nmask;
478 struct vm_area_struct *prev;
479 };
480
481 /*
482 * Scan through pages checking if pages follow certain conditions,
483 * and move them to the pagelist if they do.
484 */
queue_pages_pte_range(pmd_t * pmd,unsigned long addr,unsigned long end,struct mm_walk * walk)485 static int queue_pages_pte_range(pmd_t *pmd, unsigned long addr,
486 unsigned long end, struct mm_walk *walk)
487 {
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;
493 pte_t *pte, *mapped_pte;
494 spinlock_t *ptl;
495
496 split_huge_page_pmd(vma, addr, pmd);
497 if (pmd_trans_unstable(pmd))
498 return 0;
499
500 mapped_pte = pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
501 for (; addr != end; pte++, addr += PAGE_SIZE) {
502 if (!pte_present(*pte))
503 continue;
504 page = vm_normal_page(vma, addr, *pte);
505 if (!page)
506 continue;
507 /*
508 * vm_normal_page() filters out zero pages, but there might
509 * still be PageReserved pages to skip, perhaps in a VDSO.
510 */
511 if (PageReserved(page))
512 continue;
513 nid = page_to_nid(page);
514 if (node_isset(nid, *qp->nmask) == !!(flags & MPOL_MF_INVERT))
515 continue;
516
517 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
518 if (!vma_migratable(vma))
519 break;
520 migrate_page_add(page, qp->pagelist, flags);
521 } else
522 break;
523 }
524 pte_unmap_unlock(mapped_pte, ptl);
525 cond_resched();
526 return addr != end ? -EIO : 0;
527 }
528
queue_pages_hugetlb(pte_t * pte,unsigned long hmask,unsigned long addr,unsigned long end,struct mm_walk * walk)529 static int queue_pages_hugetlb(pte_t *pte, unsigned long hmask,
530 unsigned long addr, unsigned long end,
531 struct mm_walk *walk)
532 {
533 #ifdef CONFIG_HUGETLB_PAGE
534 struct queue_pages *qp = walk->private;
535 unsigned long flags = qp->flags;
536 int nid;
537 struct page *page;
538 spinlock_t *ptl;
539 pte_t entry;
540
541 ptl = huge_pte_lock(hstate_vma(walk->vma), walk->mm, pte);
542 entry = huge_ptep_get(pte);
543 if (!pte_present(entry))
544 goto unlock;
545 page = pte_page(entry);
546 nid = page_to_nid(page);
547 if (node_isset(nid, *qp->nmask) == !!(flags & MPOL_MF_INVERT))
548 goto unlock;
549 /* With MPOL_MF_MOVE, we migrate only unshared hugepage. */
550 if (flags & (MPOL_MF_MOVE_ALL) ||
551 (flags & MPOL_MF_MOVE && page_mapcount(page) == 1))
552 isolate_huge_page(page, qp->pagelist);
553 unlock:
554 spin_unlock(ptl);
555 #else
556 BUG();
557 #endif
558 return 0;
559 }
560
561 #ifdef CONFIG_NUMA_BALANCING
562 /*
563 * This is used to mark a range of virtual addresses to be inaccessible.
564 * These are later cleared by a NUMA hinting fault. Depending on these
565 * faults, pages may be migrated for better NUMA placement.
566 *
567 * This is assuming that NUMA faults are handled using PROT_NONE. If
568 * an architecture makes a different choice, it will need further
569 * changes to the core.
570 */
change_prot_numa(struct vm_area_struct * vma,unsigned long addr,unsigned long end)571 unsigned long change_prot_numa(struct vm_area_struct *vma,
572 unsigned long addr, unsigned long end)
573 {
574 int nr_updated;
575
576 nr_updated = change_protection(vma, addr, end, PAGE_NONE, 0, 1);
577 if (nr_updated)
578 count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
579
580 return nr_updated;
581 }
582 #else
change_prot_numa(struct vm_area_struct * vma,unsigned long addr,unsigned long end)583 static unsigned long change_prot_numa(struct vm_area_struct *vma,
584 unsigned long addr, unsigned long end)
585 {
586 return 0;
587 }
588 #endif /* CONFIG_NUMA_BALANCING */
589
queue_pages_test_walk(unsigned long start,unsigned long end,struct mm_walk * walk)590 static int queue_pages_test_walk(unsigned long start, unsigned long end,
591 struct mm_walk *walk)
592 {
593 struct vm_area_struct *vma = walk->vma;
594 struct queue_pages *qp = walk->private;
595 unsigned long endvma = vma->vm_end;
596 unsigned long flags = qp->flags;
597
598 if (vma->vm_flags & VM_PFNMAP)
599 return 1;
600
601 if (endvma > end)
602 endvma = end;
603 if (vma->vm_start > start)
604 start = vma->vm_start;
605
606 if (!(flags & MPOL_MF_DISCONTIG_OK)) {
607 if (!vma->vm_next && vma->vm_end < end)
608 return -EFAULT;
609 if (qp->prev && qp->prev->vm_end < vma->vm_start)
610 return -EFAULT;
611 }
612
613 qp->prev = vma;
614
615 if (flags & MPOL_MF_LAZY) {
616 /* Similar to task_numa_work, skip inaccessible VMAs */
617 if (vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))
618 change_prot_numa(vma, start, endvma);
619 return 1;
620 }
621
622 if ((flags & MPOL_MF_STRICT) ||
623 ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) &&
624 vma_migratable(vma)))
625 /* queue pages from current vma */
626 return 0;
627 return 1;
628 }
629
630 /*
631 * Walk through page tables and collect pages to be migrated.
632 *
633 * If pages found in a given range are on a set of nodes (determined by
634 * @nodes and @flags,) it's isolated and queued to the pagelist which is
635 * passed via @private.)
636 */
637 static int
queue_pages_range(struct mm_struct * mm,unsigned long start,unsigned long end,nodemask_t * nodes,unsigned long flags,struct list_head * pagelist)638 queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
639 nodemask_t *nodes, unsigned long flags,
640 struct list_head *pagelist)
641 {
642 struct queue_pages qp = {
643 .pagelist = pagelist,
644 .flags = flags,
645 .nmask = nodes,
646 .prev = NULL,
647 };
648 struct mm_walk queue_pages_walk = {
649 .hugetlb_entry = queue_pages_hugetlb,
650 .pmd_entry = queue_pages_pte_range,
651 .test_walk = queue_pages_test_walk,
652 .mm = mm,
653 .private = &qp,
654 };
655
656 return walk_page_range(start, end, &queue_pages_walk);
657 }
658
659 /*
660 * Apply policy to a single VMA
661 * This must be called with the mmap_sem held for writing.
662 */
vma_replace_policy(struct vm_area_struct * vma,struct mempolicy * pol)663 static int vma_replace_policy(struct vm_area_struct *vma,
664 struct mempolicy *pol)
665 {
666 int err;
667 struct mempolicy *old;
668 struct mempolicy *new;
669
670 pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
671 vma->vm_start, vma->vm_end, vma->vm_pgoff,
672 vma->vm_ops, vma->vm_file,
673 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
674
675 new = mpol_dup(pol);
676 if (IS_ERR(new))
677 return PTR_ERR(new);
678
679 if (vma->vm_ops && vma->vm_ops->set_policy) {
680 err = vma->vm_ops->set_policy(vma, new);
681 if (err)
682 goto err_out;
683 }
684
685 old = vma->vm_policy;
686 vma->vm_policy = new; /* protected by mmap_sem */
687 mpol_put(old);
688
689 return 0;
690 err_out:
691 mpol_put(new);
692 return err;
693 }
694
695 /* Step 2: apply policy to a range and do splits. */
mbind_range(struct mm_struct * mm,unsigned long start,unsigned long end,struct mempolicy * new_pol)696 static int mbind_range(struct mm_struct *mm, unsigned long start,
697 unsigned long end, struct mempolicy *new_pol)
698 {
699 struct vm_area_struct *next;
700 struct vm_area_struct *prev;
701 struct vm_area_struct *vma;
702 int err = 0;
703 pgoff_t pgoff;
704 unsigned long vmstart;
705 unsigned long vmend;
706
707 vma = find_vma(mm, start);
708 if (!vma || vma->vm_start > start)
709 return -EFAULT;
710
711 prev = vma->vm_prev;
712 if (start > vma->vm_start)
713 prev = vma;
714
715 for (; vma && vma->vm_start < end; prev = vma, vma = next) {
716 next = vma->vm_next;
717 vmstart = max(start, vma->vm_start);
718 vmend = min(end, vma->vm_end);
719
720 if (mpol_equal(vma_policy(vma), new_pol))
721 continue;
722
723 pgoff = vma->vm_pgoff +
724 ((vmstart - vma->vm_start) >> PAGE_SHIFT);
725 prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
726 vma->anon_vma, vma->vm_file, pgoff,
727 new_pol, vma->vm_userfaultfd_ctx,
728 vma_get_anon_name(vma));
729 if (prev) {
730 vma = prev;
731 next = vma->vm_next;
732 if (mpol_equal(vma_policy(vma), new_pol))
733 continue;
734 /* vma_merge() joined vma && vma->next, case 8 */
735 goto replace;
736 }
737 if (vma->vm_start != vmstart) {
738 err = split_vma(vma->vm_mm, vma, vmstart, 1);
739 if (err)
740 goto out;
741 }
742 if (vma->vm_end != vmend) {
743 err = split_vma(vma->vm_mm, vma, vmend, 0);
744 if (err)
745 goto out;
746 }
747 replace:
748 err = vma_replace_policy(vma, new_pol);
749 if (err)
750 goto out;
751 }
752
753 out:
754 return err;
755 }
756
757 /* Set the process memory policy */
do_set_mempolicy(unsigned short mode,unsigned short flags,nodemask_t * nodes)758 static long do_set_mempolicy(unsigned short mode, unsigned short flags,
759 nodemask_t *nodes)
760 {
761 struct mempolicy *new, *old;
762 NODEMASK_SCRATCH(scratch);
763 int ret;
764
765 if (!scratch)
766 return -ENOMEM;
767
768 new = mpol_new(mode, flags, nodes);
769 if (IS_ERR(new)) {
770 ret = PTR_ERR(new);
771 goto out;
772 }
773
774 task_lock(current);
775 ret = mpol_set_nodemask(new, nodes, scratch);
776 if (ret) {
777 task_unlock(current);
778 mpol_put(new);
779 goto out;
780 }
781 old = current->mempolicy;
782 current->mempolicy = new;
783 if (new && new->mode == MPOL_INTERLEAVE &&
784 nodes_weight(new->v.nodes))
785 current->il_next = first_node(new->v.nodes);
786 task_unlock(current);
787 mpol_put(old);
788 ret = 0;
789 out:
790 NODEMASK_SCRATCH_FREE(scratch);
791 return ret;
792 }
793
794 /*
795 * Return nodemask for policy for get_mempolicy() query
796 *
797 * Called with task's alloc_lock held
798 */
get_policy_nodemask(struct mempolicy * p,nodemask_t * nodes)799 static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
800 {
801 nodes_clear(*nodes);
802 if (p == &default_policy)
803 return;
804
805 switch (p->mode) {
806 case MPOL_BIND:
807 /* Fall through */
808 case MPOL_INTERLEAVE:
809 *nodes = p->v.nodes;
810 break;
811 case MPOL_PREFERRED:
812 if (!(p->flags & MPOL_F_LOCAL))
813 node_set(p->v.preferred_node, *nodes);
814 /* else return empty node mask for local allocation */
815 break;
816 default:
817 BUG();
818 }
819 }
820
lookup_node(struct mm_struct * mm,unsigned long addr)821 static int lookup_node(struct mm_struct *mm, unsigned long addr)
822 {
823 struct page *p;
824 int err;
825
826 err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, &p, NULL);
827 if (err >= 0) {
828 err = page_to_nid(p);
829 put_page(p);
830 }
831 return err;
832 }
833
834 /* Retrieve NUMA policy */
do_get_mempolicy(int * policy,nodemask_t * nmask,unsigned long addr,unsigned long flags)835 static long do_get_mempolicy(int *policy, nodemask_t *nmask,
836 unsigned long addr, unsigned long flags)
837 {
838 int err;
839 struct mm_struct *mm = current->mm;
840 struct vm_area_struct *vma = NULL;
841 struct mempolicy *pol = current->mempolicy;
842
843 if (flags &
844 ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
845 return -EINVAL;
846
847 if (flags & MPOL_F_MEMS_ALLOWED) {
848 if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
849 return -EINVAL;
850 *policy = 0; /* just so it's initialized */
851 task_lock(current);
852 *nmask = cpuset_current_mems_allowed;
853 task_unlock(current);
854 return 0;
855 }
856
857 if (flags & MPOL_F_ADDR) {
858 /*
859 * Do NOT fall back to task policy if the
860 * vma/shared policy at addr is NULL. We
861 * want to return MPOL_DEFAULT in this case.
862 */
863 down_read(&mm->mmap_sem);
864 vma = find_vma_intersection(mm, addr, addr+1);
865 if (!vma) {
866 up_read(&mm->mmap_sem);
867 return -EFAULT;
868 }
869 if (vma->vm_ops && vma->vm_ops->get_policy)
870 pol = vma->vm_ops->get_policy(vma, addr);
871 else
872 pol = vma->vm_policy;
873 } else if (addr)
874 return -EINVAL;
875
876 if (!pol)
877 pol = &default_policy; /* indicates default behavior */
878
879 if (flags & MPOL_F_NODE) {
880 if (flags & MPOL_F_ADDR) {
881 err = lookup_node(mm, addr);
882 if (err < 0)
883 goto out;
884 *policy = err;
885 } else if (pol == current->mempolicy &&
886 pol->mode == MPOL_INTERLEAVE) {
887 *policy = current->il_next;
888 } else {
889 err = -EINVAL;
890 goto out;
891 }
892 } else {
893 *policy = pol == &default_policy ? MPOL_DEFAULT :
894 pol->mode;
895 /*
896 * Internal mempolicy flags must be masked off before exposing
897 * the policy to userspace.
898 */
899 *policy |= (pol->flags & MPOL_MODE_FLAGS);
900 }
901
902 err = 0;
903 if (nmask) {
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 }
911 }
912
913 out:
914 mpol_cond_put(pol);
915 if (vma)
916 up_read(¤t->mm->mmap_sem);
917 return err;
918 }
919
920 #ifdef CONFIG_MIGRATION
921 /*
922 * page migration
923 */
migrate_page_add(struct page * page,struct list_head * pagelist,unsigned long flags)924 static void migrate_page_add(struct page *page, struct list_head *pagelist,
925 unsigned long flags)
926 {
927 /*
928 * Avoid migrating a page that is shared with others.
929 */
930 if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) {
931 if (!isolate_lru_page(page)) {
932 list_add_tail(&page->lru, pagelist);
933 inc_zone_page_state(page, NR_ISOLATED_ANON +
934 page_is_file_cache(page));
935 }
936 }
937 }
938
new_node_page(struct page * page,unsigned long node,int ** x)939 static struct page *new_node_page(struct page *page, unsigned long node, int **x)
940 {
941 if (PageHuge(page))
942 return alloc_huge_page_node(page_hstate(compound_head(page)),
943 node);
944 else
945 return __alloc_pages_node(node, GFP_HIGHUSER_MOVABLE |
946 __GFP_THISNODE, 0);
947 }
948
949 /*
950 * Migrate pages from one node to a target node.
951 * Returns error or the number of pages not migrated.
952 */
migrate_to_node(struct mm_struct * mm,int source,int dest,int flags)953 static int migrate_to_node(struct mm_struct *mm, int source, int dest,
954 int flags)
955 {
956 nodemask_t nmask;
957 LIST_HEAD(pagelist);
958 int err = 0;
959
960 nodes_clear(nmask);
961 node_set(source, nmask);
962
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)));
969 queue_pages_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
970 flags | MPOL_MF_DISCONTIG_OK, &pagelist);
971
972 if (!list_empty(&pagelist)) {
973 err = migrate_pages(&pagelist, new_node_page, NULL, dest,
974 MIGRATE_SYNC, MR_SYSCALL);
975 if (err)
976 putback_movable_pages(&pagelist);
977 }
978
979 return err;
980 }
981
982 /*
983 * Move pages between the two nodesets so as to preserve the physical
984 * layout as much as possible.
985 *
986 * Returns the number of page that could not be moved.
987 */
do_migrate_pages(struct mm_struct * mm,const nodemask_t * from,const nodemask_t * to,int flags)988 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
989 const nodemask_t *to, int flags)
990 {
991 int busy = 0;
992 int err;
993 nodemask_t tmp;
994
995 err = migrate_prep();
996 if (err)
997 return err;
998
999 down_read(&mm->mmap_sem);
1000
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.
1026 * Otherwise when we finish scanning from_tmp, we at least have the
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 */
1031
1032 tmp = *from;
1033 while (!nodes_empty(tmp)) {
1034 int s,d;
1035 int source = NUMA_NO_NODE;
1036 int dest = 0;
1037
1038 for_each_node_mask(s, tmp) {
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
1055 if ((nodes_weight(*from) != nodes_weight(*to)) &&
1056 (node_isset(s, *to)))
1057 continue;
1058
1059 d = node_remap(s, *from, *to);
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 }
1070 if (source == NUMA_NO_NODE)
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;
1079 }
1080 up_read(&mm->mmap_sem);
1081 if (err < 0)
1082 return err;
1083 return busy;
1084
1085 }
1086
1087 /*
1088 * Allocate a new page for page migration based on vma policy.
1089 * Start by assuming the page is mapped by the same vma as contains @start.
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 */
new_page(struct page * page,unsigned long start,int ** x)1094 static struct page *new_page(struct page *page, unsigned long start, int **x)
1095 {
1096 struct vm_area_struct *vma;
1097 unsigned long uninitialized_var(address);
1098
1099 vma = find_vma(current->mm, start);
1100 while (vma) {
1101 address = page_address_in_vma(page, vma);
1102 if (address != -EFAULT)
1103 break;
1104 vma = vma->vm_next;
1105 }
1106
1107 if (PageHuge(page)) {
1108 BUG_ON(!vma);
1109 return alloc_huge_page_noerr(vma, address, 1);
1110 }
1111 /*
1112 * if !vma, alloc_page_vma() will use task or system default policy
1113 */
1114 return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1115 }
1116 #else
1117
migrate_page_add(struct page * page,struct list_head * pagelist,unsigned long flags)1118 static void migrate_page_add(struct page *page, struct list_head *pagelist,
1119 unsigned long flags)
1120 {
1121 }
1122
do_migrate_pages(struct mm_struct * mm,const nodemask_t * from,const nodemask_t * to,int flags)1123 int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1124 const nodemask_t *to, int flags)
1125 {
1126 return -ENOSYS;
1127 }
1128
new_page(struct page * page,unsigned long start,int ** x)1129 static struct page *new_page(struct page *page, unsigned long start, int **x)
1130 {
1131 return NULL;
1132 }
1133 #endif
1134
do_mbind(unsigned long start,unsigned long len,unsigned short mode,unsigned short mode_flags,nodemask_t * nmask,unsigned long flags)1135 static long do_mbind(unsigned long start, unsigned long len,
1136 unsigned short mode, unsigned short mode_flags,
1137 nodemask_t *nmask, unsigned long flags)
1138 {
1139 struct mm_struct *mm = current->mm;
1140 struct mempolicy *new;
1141 unsigned long end;
1142 int err;
1143 LIST_HEAD(pagelist);
1144
1145 if (flags & ~(unsigned long)MPOL_MF_VALID)
1146 return -EINVAL;
1147 if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
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
1164 new = mpol_new(mode, mode_flags, nmask);
1165 if (IS_ERR(new))
1166 return PTR_ERR(new);
1167
1168 if (flags & MPOL_MF_LAZY)
1169 new->flags |= MPOL_F_MOF;
1170
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
1178 pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1179 start, start + len, mode, mode_flags,
1180 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
1181
1182 if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1183
1184 err = migrate_prep();
1185 if (err)
1186 goto mpol_out;
1187 }
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 }
1201 if (err)
1202 goto mpol_out;
1203
1204 err = queue_pages_range(mm, start, end, nmask,
1205 flags | MPOL_MF_INVERT, &pagelist);
1206 if (!err)
1207 err = mbind_range(mm, start, end, new);
1208
1209 if (!err) {
1210 int nr_failed = 0;
1211
1212 if (!list_empty(&pagelist)) {
1213 WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1214 nr_failed = migrate_pages(&pagelist, new_page, NULL,
1215 start, MIGRATE_SYNC, MR_MEMPOLICY_MBIND);
1216 if (nr_failed)
1217 putback_movable_pages(&pagelist);
1218 }
1219
1220 if (nr_failed && (flags & MPOL_MF_STRICT))
1221 err = -EIO;
1222 } else
1223 putback_movable_pages(&pagelist);
1224
1225 up_write(&mm->mmap_sem);
1226 mpol_out:
1227 mpol_put(new);
1228 return err;
1229 }
1230
1231 /*
1232 * User space interface with variable sized bitmaps for nodelists.
1233 */
1234
1235 /* Copy a node mask from user space. */
get_nodes(nodemask_t * nodes,const unsigned long __user * nmask,unsigned long maxnode)1236 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1237 unsigned long maxnode)
1238 {
1239 unsigned long k;
1240 unsigned long t;
1241 unsigned long nlongs;
1242 unsigned long endmask;
1243
1244 --maxnode;
1245 nodes_clear(*nodes);
1246 if (maxnode == 0 || !nmask)
1247 return 0;
1248 if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1249 return -EINVAL;
1250
1251 nlongs = BITS_TO_LONGS(maxnode);
1252 if ((maxnode % BITS_PER_LONG) == 0)
1253 endmask = ~0UL;
1254 else
1255 endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1256
1257 /*
1258 * When the user specified more nodes than supported just check
1259 * if the non supported part is all zero.
1260 *
1261 * If maxnode have more longs than MAX_NUMNODES, check
1262 * the bits in that area first. And then go through to
1263 * check the rest bits which equal or bigger than MAX_NUMNODES.
1264 * Otherwise, just check bits [MAX_NUMNODES, maxnode).
1265 */
1266 if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1267 if (nlongs > PAGE_SIZE/sizeof(long))
1268 return -EINVAL;
1269 for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1270 if (get_user(t, nmask + k))
1271 return -EFAULT;
1272 if (k == nlongs - 1) {
1273 if (t & endmask)
1274 return -EINVAL;
1275 } else if (t)
1276 return -EINVAL;
1277 }
1278 nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1279 endmask = ~0UL;
1280 }
1281
1282 if (maxnode > MAX_NUMNODES && MAX_NUMNODES % BITS_PER_LONG != 0) {
1283 unsigned long valid_mask = endmask;
1284
1285 valid_mask &= ~((1UL << (MAX_NUMNODES % BITS_PER_LONG)) - 1);
1286 if (get_user(t, nmask + nlongs - 1))
1287 return -EFAULT;
1288 if (t & valid_mask)
1289 return -EINVAL;
1290 }
1291
1292 if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1293 return -EFAULT;
1294 nodes_addr(*nodes)[nlongs-1] &= endmask;
1295 return 0;
1296 }
1297
1298 /* Copy a kernel node mask to user space */
copy_nodes_to_user(unsigned long __user * mask,unsigned long maxnode,nodemask_t * nodes)1299 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1300 nodemask_t *nodes)
1301 {
1302 unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1303 unsigned int nbytes = BITS_TO_LONGS(nr_node_ids) * sizeof(long);
1304
1305 if (copy > nbytes) {
1306 if (copy > PAGE_SIZE)
1307 return -EINVAL;
1308 if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1309 return -EFAULT;
1310 copy = nbytes;
1311 }
1312 return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1313 }
1314
SYSCALL_DEFINE6(mbind,unsigned long,start,unsigned long,len,unsigned long,mode,const unsigned long __user *,nmask,unsigned long,maxnode,unsigned,flags)1315 SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1316 unsigned long, mode, const unsigned long __user *, nmask,
1317 unsigned long, maxnode, unsigned, flags)
1318 {
1319 nodemask_t nodes;
1320 int err;
1321 unsigned short mode_flags;
1322
1323 mode_flags = mode & MPOL_MODE_FLAGS;
1324 mode &= ~MPOL_MODE_FLAGS;
1325 if (mode >= MPOL_MAX)
1326 return -EINVAL;
1327 if ((mode_flags & MPOL_F_STATIC_NODES) &&
1328 (mode_flags & MPOL_F_RELATIVE_NODES))
1329 return -EINVAL;
1330 err = get_nodes(&nodes, nmask, maxnode);
1331 if (err)
1332 return err;
1333 return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1334 }
1335
1336 /* Set the process memory policy */
SYSCALL_DEFINE3(set_mempolicy,int,mode,const unsigned long __user *,nmask,unsigned long,maxnode)1337 SYSCALL_DEFINE3(set_mempolicy, int, mode, const unsigned long __user *, nmask,
1338 unsigned long, maxnode)
1339 {
1340 int err;
1341 nodemask_t nodes;
1342 unsigned short flags;
1343
1344 flags = mode & MPOL_MODE_FLAGS;
1345 mode &= ~MPOL_MODE_FLAGS;
1346 if ((unsigned int)mode >= MPOL_MAX)
1347 return -EINVAL;
1348 if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1349 return -EINVAL;
1350 err = get_nodes(&nodes, nmask, maxnode);
1351 if (err)
1352 return err;
1353 return do_set_mempolicy(mode, flags, &nodes);
1354 }
1355
SYSCALL_DEFINE4(migrate_pages,pid_t,pid,unsigned long,maxnode,const unsigned long __user *,old_nodes,const unsigned long __user *,new_nodes)1356 SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1357 const unsigned long __user *, old_nodes,
1358 const unsigned long __user *, new_nodes)
1359 {
1360 const struct cred *cred = current_cred(), *tcred;
1361 struct mm_struct *mm = NULL;
1362 struct task_struct *task;
1363 nodemask_t task_nodes;
1364 int err;
1365 nodemask_t *old;
1366 nodemask_t *new;
1367 NODEMASK_SCRATCH(scratch);
1368
1369 if (!scratch)
1370 return -ENOMEM;
1371
1372 old = &scratch->mask1;
1373 new = &scratch->mask2;
1374
1375 err = get_nodes(old, old_nodes, maxnode);
1376 if (err)
1377 goto out;
1378
1379 err = get_nodes(new, new_nodes, maxnode);
1380 if (err)
1381 goto out;
1382
1383 /* Find the mm_struct */
1384 rcu_read_lock();
1385 task = pid ? find_task_by_vpid(pid) : current;
1386 if (!task) {
1387 rcu_read_unlock();
1388 err = -ESRCH;
1389 goto out;
1390 }
1391 get_task_struct(task);
1392
1393 err = -EINVAL;
1394
1395 /*
1396 * Check if this process has the right to modify the specified
1397 * process. The right exists if the process has administrative
1398 * capabilities, superuser privileges or the same
1399 * userid as the target process.
1400 */
1401 tcred = __task_cred(task);
1402 if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1403 !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) &&
1404 !capable(CAP_SYS_NICE)) {
1405 rcu_read_unlock();
1406 err = -EPERM;
1407 goto out_put;
1408 }
1409 rcu_read_unlock();
1410
1411 task_nodes = cpuset_mems_allowed(task);
1412 /* Is the user allowed to access the target nodes? */
1413 if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1414 err = -EPERM;
1415 goto out_put;
1416 }
1417
1418 task_nodes = cpuset_mems_allowed(current);
1419 nodes_and(*new, *new, task_nodes);
1420 if (nodes_empty(*new))
1421 goto out_put;
1422
1423 nodes_and(*new, *new, node_states[N_MEMORY]);
1424 if (nodes_empty(*new))
1425 goto out_put;
1426
1427 err = security_task_movememory(task);
1428 if (err)
1429 goto out_put;
1430
1431 mm = get_task_mm(task);
1432 put_task_struct(task);
1433
1434 if (!mm) {
1435 err = -EINVAL;
1436 goto out;
1437 }
1438
1439 err = do_migrate_pages(mm, old, new,
1440 capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1441
1442 mmput(mm);
1443 out:
1444 NODEMASK_SCRATCH_FREE(scratch);
1445
1446 return err;
1447
1448 out_put:
1449 put_task_struct(task);
1450 goto out;
1451
1452 }
1453
1454
1455 /* Retrieve NUMA policy */
SYSCALL_DEFINE5(get_mempolicy,int __user *,policy,unsigned long __user *,nmask,unsigned long,maxnode,unsigned long,addr,unsigned long,flags)1456 SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1457 unsigned long __user *, nmask, unsigned long, maxnode,
1458 unsigned long, addr, unsigned long, flags)
1459 {
1460 int err;
1461 int uninitialized_var(pval);
1462 nodemask_t nodes;
1463
1464 if (nmask != NULL && maxnode < nr_node_ids)
1465 return -EINVAL;
1466
1467 err = do_get_mempolicy(&pval, &nodes, addr, flags);
1468
1469 if (err)
1470 return err;
1471
1472 if (policy && put_user(pval, policy))
1473 return -EFAULT;
1474
1475 if (nmask)
1476 err = copy_nodes_to_user(nmask, maxnode, &nodes);
1477
1478 return err;
1479 }
1480
1481 #ifdef CONFIG_COMPAT
1482
COMPAT_SYSCALL_DEFINE5(get_mempolicy,int __user *,policy,compat_ulong_t __user *,nmask,compat_ulong_t,maxnode,compat_ulong_t,addr,compat_ulong_t,flags)1483 COMPAT_SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1484 compat_ulong_t __user *, nmask,
1485 compat_ulong_t, maxnode,
1486 compat_ulong_t, addr, compat_ulong_t, flags)
1487 {
1488 long err;
1489 unsigned long __user *nm = NULL;
1490 unsigned long nr_bits, alloc_size;
1491 DECLARE_BITMAP(bm, MAX_NUMNODES);
1492
1493 nr_bits = min_t(unsigned long, maxnode-1, nr_node_ids);
1494 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1495
1496 if (nmask)
1497 nm = compat_alloc_user_space(alloc_size);
1498
1499 err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1500
1501 if (!err && nmask) {
1502 unsigned long copy_size;
1503 copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1504 err = copy_from_user(bm, nm, copy_size);
1505 /* ensure entire bitmap is zeroed */
1506 err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1507 err |= compat_put_bitmap(nmask, bm, nr_bits);
1508 }
1509
1510 return err;
1511 }
1512
COMPAT_SYSCALL_DEFINE3(set_mempolicy,int,mode,compat_ulong_t __user *,nmask,compat_ulong_t,maxnode)1513 COMPAT_SYSCALL_DEFINE3(set_mempolicy, int, mode, compat_ulong_t __user *, nmask,
1514 compat_ulong_t, maxnode)
1515 {
1516 unsigned long __user *nm = NULL;
1517 unsigned long nr_bits, alloc_size;
1518 DECLARE_BITMAP(bm, MAX_NUMNODES);
1519
1520 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1521 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1522
1523 if (nmask) {
1524 if (compat_get_bitmap(bm, nmask, nr_bits))
1525 return -EFAULT;
1526 nm = compat_alloc_user_space(alloc_size);
1527 if (copy_to_user(nm, bm, alloc_size))
1528 return -EFAULT;
1529 }
1530
1531 return sys_set_mempolicy(mode, nm, nr_bits+1);
1532 }
1533
COMPAT_SYSCALL_DEFINE6(mbind,compat_ulong_t,start,compat_ulong_t,len,compat_ulong_t,mode,compat_ulong_t __user *,nmask,compat_ulong_t,maxnode,compat_ulong_t,flags)1534 COMPAT_SYSCALL_DEFINE6(mbind, compat_ulong_t, start, compat_ulong_t, len,
1535 compat_ulong_t, mode, compat_ulong_t __user *, nmask,
1536 compat_ulong_t, maxnode, compat_ulong_t, flags)
1537 {
1538 unsigned long __user *nm = NULL;
1539 unsigned long nr_bits, alloc_size;
1540 nodemask_t bm;
1541
1542 nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1543 alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1544
1545 if (nmask) {
1546 if (compat_get_bitmap(nodes_addr(bm), nmask, nr_bits))
1547 return -EFAULT;
1548 nm = compat_alloc_user_space(alloc_size);
1549 if (copy_to_user(nm, nodes_addr(bm), alloc_size))
1550 return -EFAULT;
1551 }
1552
1553 return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1554 }
1555
1556 #endif
1557
__get_vma_policy(struct vm_area_struct * vma,unsigned long addr)1558 struct mempolicy *__get_vma_policy(struct vm_area_struct *vma,
1559 unsigned long addr)
1560 {
1561 struct mempolicy *pol = NULL;
1562
1563 if (vma) {
1564 if (vma->vm_ops && vma->vm_ops->get_policy) {
1565 pol = vma->vm_ops->get_policy(vma, addr);
1566 } else if (vma->vm_policy) {
1567 pol = vma->vm_policy;
1568
1569 /*
1570 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1571 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1572 * count on these policies which will be dropped by
1573 * mpol_cond_put() later
1574 */
1575 if (mpol_needs_cond_ref(pol))
1576 mpol_get(pol);
1577 }
1578 }
1579
1580 return pol;
1581 }
1582
1583 /*
1584 * get_vma_policy(@vma, @addr)
1585 * @vma: virtual memory area whose policy is sought
1586 * @addr: address in @vma for shared policy lookup
1587 *
1588 * Returns effective policy for a VMA at specified address.
1589 * Falls back to current->mempolicy or system default policy, as necessary.
1590 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1591 * count--added by the get_policy() vm_op, as appropriate--to protect against
1592 * freeing by another task. It is the caller's responsibility to free the
1593 * extra reference for shared policies.
1594 */
get_vma_policy(struct vm_area_struct * vma,unsigned long addr)1595 static struct mempolicy *get_vma_policy(struct vm_area_struct *vma,
1596 unsigned long addr)
1597 {
1598 struct mempolicy *pol = __get_vma_policy(vma, addr);
1599
1600 if (!pol)
1601 pol = get_task_policy(current);
1602
1603 return pol;
1604 }
1605
vma_policy_mof(struct vm_area_struct * vma)1606 bool vma_policy_mof(struct vm_area_struct *vma)
1607 {
1608 struct mempolicy *pol;
1609
1610 if (vma->vm_ops && vma->vm_ops->get_policy) {
1611 bool ret = false;
1612
1613 pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1614 if (pol && (pol->flags & MPOL_F_MOF))
1615 ret = true;
1616 mpol_cond_put(pol);
1617
1618 return ret;
1619 }
1620
1621 pol = vma->vm_policy;
1622 if (!pol)
1623 pol = get_task_policy(current);
1624
1625 return pol->flags & MPOL_F_MOF;
1626 }
1627
apply_policy_zone(struct mempolicy * policy,enum zone_type zone)1628 static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1629 {
1630 enum zone_type dynamic_policy_zone = policy_zone;
1631
1632 BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1633
1634 /*
1635 * if policy->v.nodes has movable memory only,
1636 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1637 *
1638 * policy->v.nodes is intersect with node_states[N_MEMORY].
1639 * so if the following test faile, it implies
1640 * policy->v.nodes has movable memory only.
1641 */
1642 if (!nodes_intersects(policy->v.nodes, node_states[N_HIGH_MEMORY]))
1643 dynamic_policy_zone = ZONE_MOVABLE;
1644
1645 return zone >= dynamic_policy_zone;
1646 }
1647
1648 /*
1649 * Return a nodemask representing a mempolicy for filtering nodes for
1650 * page allocation
1651 */
policy_nodemask(gfp_t gfp,struct mempolicy * policy)1652 static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1653 {
1654 /* Lower zones don't get a nodemask applied for MPOL_BIND */
1655 if (unlikely(policy->mode == MPOL_BIND) &&
1656 apply_policy_zone(policy, gfp_zone(gfp)) &&
1657 cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1658 return &policy->v.nodes;
1659
1660 return NULL;
1661 }
1662
1663 /* Return a zonelist indicated by gfp for node representing a mempolicy */
policy_zonelist(gfp_t gfp,struct mempolicy * policy,int nd)1664 static struct zonelist *policy_zonelist(gfp_t gfp, struct mempolicy *policy,
1665 int nd)
1666 {
1667 switch (policy->mode) {
1668 case MPOL_PREFERRED:
1669 if (!(policy->flags & MPOL_F_LOCAL))
1670 nd = policy->v.preferred_node;
1671 break;
1672 case MPOL_BIND:
1673 /*
1674 * Normally, MPOL_BIND allocations are node-local within the
1675 * allowed nodemask. However, if __GFP_THISNODE is set and the
1676 * current node isn't part of the mask, we use the zonelist for
1677 * the first node in the mask instead.
1678 */
1679 if (unlikely(gfp & __GFP_THISNODE) &&
1680 unlikely(!node_isset(nd, policy->v.nodes)))
1681 nd = first_node(policy->v.nodes);
1682 break;
1683 default:
1684 BUG();
1685 }
1686 return node_zonelist(nd, gfp);
1687 }
1688
1689 /* Do dynamic interleaving for a process */
interleave_nodes(struct mempolicy * policy)1690 static unsigned interleave_nodes(struct mempolicy *policy)
1691 {
1692 unsigned nid, next;
1693 struct task_struct *me = current;
1694
1695 nid = me->il_next;
1696 next = next_node(nid, policy->v.nodes);
1697 if (next >= MAX_NUMNODES)
1698 next = first_node(policy->v.nodes);
1699 if (next < MAX_NUMNODES)
1700 me->il_next = next;
1701 return nid;
1702 }
1703
1704 /*
1705 * Depending on the memory policy provide a node from which to allocate the
1706 * next slab entry.
1707 */
mempolicy_slab_node(void)1708 unsigned int mempolicy_slab_node(void)
1709 {
1710 struct mempolicy *policy;
1711 int node = numa_mem_id();
1712
1713 if (in_interrupt())
1714 return node;
1715
1716 policy = current->mempolicy;
1717 if (!policy || policy->flags & MPOL_F_LOCAL)
1718 return node;
1719
1720 switch (policy->mode) {
1721 case MPOL_PREFERRED:
1722 /*
1723 * handled MPOL_F_LOCAL above
1724 */
1725 return policy->v.preferred_node;
1726
1727 case MPOL_INTERLEAVE:
1728 return interleave_nodes(policy);
1729
1730 case MPOL_BIND: {
1731 /*
1732 * Follow bind policy behavior and start allocation at the
1733 * first node.
1734 */
1735 struct zonelist *zonelist;
1736 struct zone *zone;
1737 enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1738 zonelist = &NODE_DATA(node)->node_zonelists[0];
1739 (void)first_zones_zonelist(zonelist, highest_zoneidx,
1740 &policy->v.nodes,
1741 &zone);
1742 return zone ? zone->node : node;
1743 }
1744
1745 default:
1746 BUG();
1747 }
1748 }
1749
1750 /* Do static interleaving for a VMA with known offset. */
offset_il_node(struct mempolicy * pol,struct vm_area_struct * vma,unsigned long off)1751 static unsigned offset_il_node(struct mempolicy *pol,
1752 struct vm_area_struct *vma, unsigned long off)
1753 {
1754 unsigned nnodes = nodes_weight(pol->v.nodes);
1755 unsigned target;
1756 int c;
1757 int nid = NUMA_NO_NODE;
1758
1759 if (!nnodes)
1760 return numa_node_id();
1761 target = (unsigned int)off % nnodes;
1762 c = 0;
1763 do {
1764 nid = next_node(nid, pol->v.nodes);
1765 c++;
1766 } while (c <= target);
1767 return nid;
1768 }
1769
1770 /* Determine a node number for interleave */
interleave_nid(struct mempolicy * pol,struct vm_area_struct * vma,unsigned long addr,int shift)1771 static inline unsigned interleave_nid(struct mempolicy *pol,
1772 struct vm_area_struct *vma, unsigned long addr, int shift)
1773 {
1774 if (vma) {
1775 unsigned long off;
1776
1777 /*
1778 * for small pages, there is no difference between
1779 * shift and PAGE_SHIFT, so the bit-shift is safe.
1780 * for huge pages, since vm_pgoff is in units of small
1781 * pages, we need to shift off the always 0 bits to get
1782 * a useful offset.
1783 */
1784 BUG_ON(shift < PAGE_SHIFT);
1785 off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1786 off += (addr - vma->vm_start) >> shift;
1787 return offset_il_node(pol, vma, off);
1788 } else
1789 return interleave_nodes(pol);
1790 }
1791
1792 /*
1793 * Return the bit number of a random bit set in the nodemask.
1794 * (returns NUMA_NO_NODE if nodemask is empty)
1795 */
node_random(const nodemask_t * maskp)1796 int node_random(const nodemask_t *maskp)
1797 {
1798 int w, bit = NUMA_NO_NODE;
1799
1800 w = nodes_weight(*maskp);
1801 if (w)
1802 bit = bitmap_ord_to_pos(maskp->bits,
1803 get_random_int() % w, MAX_NUMNODES);
1804 return bit;
1805 }
1806
1807 #ifdef CONFIG_HUGETLBFS
1808 /*
1809 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1810 * @vma: virtual memory area whose policy is sought
1811 * @addr: address in @vma for shared policy lookup and interleave policy
1812 * @gfp_flags: for requested zone
1813 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
1814 * @nodemask: pointer to nodemask pointer for MPOL_BIND nodemask
1815 *
1816 * Returns a zonelist suitable for a huge page allocation and a pointer
1817 * to the struct mempolicy for conditional unref after allocation.
1818 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1819 * @nodemask for filtering the zonelist.
1820 *
1821 * Must be protected by read_mems_allowed_begin()
1822 */
huge_zonelist(struct vm_area_struct * vma,unsigned long addr,gfp_t gfp_flags,struct mempolicy ** mpol,nodemask_t ** nodemask)1823 struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
1824 gfp_t gfp_flags, struct mempolicy **mpol,
1825 nodemask_t **nodemask)
1826 {
1827 struct zonelist *zl;
1828
1829 *mpol = get_vma_policy(vma, addr);
1830 *nodemask = NULL; /* assume !MPOL_BIND */
1831
1832 if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1833 zl = node_zonelist(interleave_nid(*mpol, vma, addr,
1834 huge_page_shift(hstate_vma(vma))), gfp_flags);
1835 } else {
1836 zl = policy_zonelist(gfp_flags, *mpol, numa_node_id());
1837 if ((*mpol)->mode == MPOL_BIND)
1838 *nodemask = &(*mpol)->v.nodes;
1839 }
1840 return zl;
1841 }
1842
1843 /*
1844 * init_nodemask_of_mempolicy
1845 *
1846 * If the current task's mempolicy is "default" [NULL], return 'false'
1847 * to indicate default policy. Otherwise, extract the policy nodemask
1848 * for 'bind' or 'interleave' policy into the argument nodemask, or
1849 * initialize the argument nodemask to contain the single node for
1850 * 'preferred' or 'local' policy and return 'true' to indicate presence
1851 * of non-default mempolicy.
1852 *
1853 * We don't bother with reference counting the mempolicy [mpol_get/put]
1854 * because the current task is examining it's own mempolicy and a task's
1855 * mempolicy is only ever changed by the task itself.
1856 *
1857 * N.B., it is the caller's responsibility to free a returned nodemask.
1858 */
init_nodemask_of_mempolicy(nodemask_t * mask)1859 bool init_nodemask_of_mempolicy(nodemask_t *mask)
1860 {
1861 struct mempolicy *mempolicy;
1862 int nid;
1863
1864 if (!(mask && current->mempolicy))
1865 return false;
1866
1867 task_lock(current);
1868 mempolicy = current->mempolicy;
1869 switch (mempolicy->mode) {
1870 case MPOL_PREFERRED:
1871 if (mempolicy->flags & MPOL_F_LOCAL)
1872 nid = numa_node_id();
1873 else
1874 nid = mempolicy->v.preferred_node;
1875 init_nodemask_of_node(mask, nid);
1876 break;
1877
1878 case MPOL_BIND:
1879 /* Fall through */
1880 case MPOL_INTERLEAVE:
1881 *mask = mempolicy->v.nodes;
1882 break;
1883
1884 default:
1885 BUG();
1886 }
1887 task_unlock(current);
1888
1889 return true;
1890 }
1891 #endif
1892
1893 /*
1894 * mempolicy_nodemask_intersects
1895 *
1896 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1897 * policy. Otherwise, check for intersection between mask and the policy
1898 * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local'
1899 * policy, always return true since it may allocate elsewhere on fallback.
1900 *
1901 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1902 */
mempolicy_nodemask_intersects(struct task_struct * tsk,const nodemask_t * mask)1903 bool mempolicy_nodemask_intersects(struct task_struct *tsk,
1904 const nodemask_t *mask)
1905 {
1906 struct mempolicy *mempolicy;
1907 bool ret = true;
1908
1909 if (!mask)
1910 return ret;
1911 task_lock(tsk);
1912 mempolicy = tsk->mempolicy;
1913 if (!mempolicy)
1914 goto out;
1915
1916 switch (mempolicy->mode) {
1917 case MPOL_PREFERRED:
1918 /*
1919 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1920 * allocate from, they may fallback to other nodes when oom.
1921 * Thus, it's possible for tsk to have allocated memory from
1922 * nodes in mask.
1923 */
1924 break;
1925 case MPOL_BIND:
1926 case MPOL_INTERLEAVE:
1927 ret = nodes_intersects(mempolicy->v.nodes, *mask);
1928 break;
1929 default:
1930 BUG();
1931 }
1932 out:
1933 task_unlock(tsk);
1934 return ret;
1935 }
1936
1937 /* Allocate a page in interleaved policy.
1938 Own path because it needs to do special accounting. */
alloc_page_interleave(gfp_t gfp,unsigned order,unsigned nid)1939 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1940 unsigned nid)
1941 {
1942 struct zonelist *zl;
1943 struct page *page;
1944
1945 zl = node_zonelist(nid, gfp);
1946 page = __alloc_pages(gfp, order, zl);
1947 if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0]))
1948 inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
1949 return page;
1950 }
1951
1952 /**
1953 * alloc_pages_vma - Allocate a page for a VMA.
1954 *
1955 * @gfp:
1956 * %GFP_USER user allocation.
1957 * %GFP_KERNEL kernel allocations,
1958 * %GFP_HIGHMEM highmem/user allocations,
1959 * %GFP_FS allocation should not call back into a file system.
1960 * %GFP_ATOMIC don't sleep.
1961 *
1962 * @order:Order of the GFP allocation.
1963 * @vma: Pointer to VMA or NULL if not available.
1964 * @addr: Virtual Address of the allocation. Must be inside the VMA.
1965 * @node: Which node to prefer for allocation (modulo policy).
1966 * @hugepage: for hugepages try only the preferred node if possible
1967 *
1968 * This function allocates a page from the kernel page pool and applies
1969 * a NUMA policy associated with the VMA or the current process.
1970 * When VMA is not NULL caller must hold down_read on the mmap_sem of the
1971 * mm_struct of the VMA to prevent it from going away. Should be used for
1972 * all allocations for pages that will be mapped into user space. Returns
1973 * NULL when no page can be allocated.
1974 */
1975 struct page *
alloc_pages_vma(gfp_t gfp,int order,struct vm_area_struct * vma,unsigned long addr,int node,bool hugepage)1976 alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
1977 unsigned long addr, int node, bool hugepage)
1978 {
1979 struct mempolicy *pol;
1980 struct page *page;
1981 unsigned int cpuset_mems_cookie;
1982 struct zonelist *zl;
1983 nodemask_t *nmask;
1984
1985 retry_cpuset:
1986 pol = get_vma_policy(vma, addr);
1987 cpuset_mems_cookie = read_mems_allowed_begin();
1988
1989 if (pol->mode == MPOL_INTERLEAVE) {
1990 unsigned nid;
1991
1992 nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
1993 mpol_cond_put(pol);
1994 page = alloc_page_interleave(gfp, order, nid);
1995 goto out;
1996 }
1997
1998 if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && hugepage)) {
1999 int hpage_node = node;
2000
2001 /*
2002 * For hugepage allocation and non-interleave policy which
2003 * allows the current node (or other explicitly preferred
2004 * node) we only try to allocate from the current/preferred
2005 * node and don't fall back to other nodes, as the cost of
2006 * remote accesses would likely offset THP benefits.
2007 *
2008 * If the policy is interleave, or does not allow the current
2009 * node in its nodemask, we allocate the standard way.
2010 */
2011 if (pol->mode == MPOL_PREFERRED &&
2012 !(pol->flags & MPOL_F_LOCAL))
2013 hpage_node = pol->v.preferred_node;
2014
2015 nmask = policy_nodemask(gfp, pol);
2016 if (!nmask || node_isset(hpage_node, *nmask)) {
2017 mpol_cond_put(pol);
2018 /*
2019 * We cannot invoke reclaim if __GFP_THISNODE
2020 * is set. Invoking reclaim with
2021 * __GFP_THISNODE set, would cause THP
2022 * allocations to trigger heavy swapping
2023 * despite there may be tons of free memory
2024 * (including potentially plenty of THP
2025 * already available in the buddy) on all the
2026 * other NUMA nodes.
2027 *
2028 * At most we could invoke compaction when
2029 * __GFP_THISNODE is set (but we would need to
2030 * refrain from invoking reclaim even if
2031 * compaction returned COMPACT_SKIPPED because
2032 * there wasn't not enough memory to succeed
2033 * compaction). For now just avoid
2034 * __GFP_THISNODE instead of limiting the
2035 * allocation path to a strict and single
2036 * compaction invocation.
2037 *
2038 * Supposedly if direct reclaim was enabled by
2039 * the caller, the app prefers THP regardless
2040 * of the node it comes from so this would be
2041 * more desiderable behavior than only
2042 * providing THP originated from the local
2043 * node in such case.
2044 */
2045 if (!(gfp & __GFP_DIRECT_RECLAIM))
2046 gfp |= __GFP_THISNODE;
2047 page = __alloc_pages_node(hpage_node, gfp, order);
2048 goto out;
2049 }
2050 }
2051
2052 nmask = policy_nodemask(gfp, pol);
2053 zl = policy_zonelist(gfp, pol, node);
2054 page = __alloc_pages_nodemask(gfp, order, zl, nmask);
2055 mpol_cond_put(pol);
2056 out:
2057 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
2058 goto retry_cpuset;
2059 return page;
2060 }
2061
2062 /**
2063 * alloc_pages_current - Allocate pages.
2064 *
2065 * @gfp:
2066 * %GFP_USER user allocation,
2067 * %GFP_KERNEL kernel allocation,
2068 * %GFP_HIGHMEM highmem allocation,
2069 * %GFP_FS don't call back into a file system.
2070 * %GFP_ATOMIC don't sleep.
2071 * @order: Power of two of allocation size in pages. 0 is a single page.
2072 *
2073 * Allocate a page from the kernel page pool. When not in
2074 * interrupt context and apply the current process NUMA policy.
2075 * Returns NULL when no page can be allocated.
2076 *
2077 * Don't call cpuset_update_task_memory_state() unless
2078 * 1) it's ok to take cpuset_sem (can WAIT), and
2079 * 2) allocating for current task (not interrupt).
2080 */
alloc_pages_current(gfp_t gfp,unsigned order)2081 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
2082 {
2083 struct mempolicy *pol = &default_policy;
2084 struct page *page;
2085 unsigned int cpuset_mems_cookie;
2086
2087 if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2088 pol = get_task_policy(current);
2089
2090 retry_cpuset:
2091 cpuset_mems_cookie = read_mems_allowed_begin();
2092
2093 /*
2094 * No reference counting needed for current->mempolicy
2095 * nor system default_policy
2096 */
2097 if (pol->mode == MPOL_INTERLEAVE)
2098 page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2099 else
2100 page = __alloc_pages_nodemask(gfp, order,
2101 policy_zonelist(gfp, pol, numa_node_id()),
2102 policy_nodemask(gfp, pol));
2103
2104 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
2105 goto retry_cpuset;
2106
2107 return page;
2108 }
2109 EXPORT_SYMBOL(alloc_pages_current);
2110
vma_dup_policy(struct vm_area_struct * src,struct vm_area_struct * dst)2111 int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2112 {
2113 struct mempolicy *pol = mpol_dup(vma_policy(src));
2114
2115 if (IS_ERR(pol))
2116 return PTR_ERR(pol);
2117 dst->vm_policy = pol;
2118 return 0;
2119 }
2120
2121 /*
2122 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2123 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2124 * with the mems_allowed returned by cpuset_mems_allowed(). This
2125 * keeps mempolicies cpuset relative after its cpuset moves. See
2126 * further kernel/cpuset.c update_nodemask().
2127 *
2128 * current's mempolicy may be rebinded by the other task(the task that changes
2129 * cpuset's mems), so we needn't do rebind work for current task.
2130 */
2131
2132 /* Slow path of a mempolicy duplicate */
__mpol_dup(struct mempolicy * old)2133 struct mempolicy *__mpol_dup(struct mempolicy *old)
2134 {
2135 struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2136
2137 if (!new)
2138 return ERR_PTR(-ENOMEM);
2139
2140 /* task's mempolicy is protected by alloc_lock */
2141 if (old == current->mempolicy) {
2142 task_lock(current);
2143 *new = *old;
2144 task_unlock(current);
2145 } else
2146 *new = *old;
2147
2148 if (current_cpuset_is_being_rebound()) {
2149 nodemask_t mems = cpuset_mems_allowed(current);
2150 if (new->flags & MPOL_F_REBINDING)
2151 mpol_rebind_policy(new, &mems, MPOL_REBIND_STEP2);
2152 else
2153 mpol_rebind_policy(new, &mems, MPOL_REBIND_ONCE);
2154 }
2155 atomic_set(&new->refcnt, 1);
2156 return new;
2157 }
2158
2159 /* Slow path of a mempolicy comparison */
__mpol_equal(struct mempolicy * a,struct mempolicy * b)2160 bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2161 {
2162 if (!a || !b)
2163 return false;
2164 if (a->mode != b->mode)
2165 return false;
2166 if (a->flags != b->flags)
2167 return false;
2168 if (mpol_store_user_nodemask(a))
2169 if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2170 return false;
2171
2172 switch (a->mode) {
2173 case MPOL_BIND:
2174 /* Fall through */
2175 case MPOL_INTERLEAVE:
2176 return !!nodes_equal(a->v.nodes, b->v.nodes);
2177 case MPOL_PREFERRED:
2178 /* a's ->flags is the same as b's */
2179 if (a->flags & MPOL_F_LOCAL)
2180 return true;
2181 return a->v.preferred_node == b->v.preferred_node;
2182 default:
2183 BUG();
2184 return false;
2185 }
2186 }
2187
2188 /*
2189 * Shared memory backing store policy support.
2190 *
2191 * Remember policies even when nobody has shared memory mapped.
2192 * The policies are kept in Red-Black tree linked from the inode.
2193 * They are protected by the sp->lock spinlock, which should be held
2194 * for any accesses to the tree.
2195 */
2196
2197 /* lookup first element intersecting start-end */
2198 /* Caller holds sp->lock */
2199 static struct sp_node *
sp_lookup(struct shared_policy * sp,unsigned long start,unsigned long end)2200 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2201 {
2202 struct rb_node *n = sp->root.rb_node;
2203
2204 while (n) {
2205 struct sp_node *p = rb_entry(n, struct sp_node, nd);
2206
2207 if (start >= p->end)
2208 n = n->rb_right;
2209 else if (end <= p->start)
2210 n = n->rb_left;
2211 else
2212 break;
2213 }
2214 if (!n)
2215 return NULL;
2216 for (;;) {
2217 struct sp_node *w = NULL;
2218 struct rb_node *prev = rb_prev(n);
2219 if (!prev)
2220 break;
2221 w = rb_entry(prev, struct sp_node, nd);
2222 if (w->end <= start)
2223 break;
2224 n = prev;
2225 }
2226 return rb_entry(n, struct sp_node, nd);
2227 }
2228
2229 /* Insert a new shared policy into the list. */
2230 /* Caller holds sp->lock */
sp_insert(struct shared_policy * sp,struct sp_node * new)2231 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2232 {
2233 struct rb_node **p = &sp->root.rb_node;
2234 struct rb_node *parent = NULL;
2235 struct sp_node *nd;
2236
2237 while (*p) {
2238 parent = *p;
2239 nd = rb_entry(parent, struct sp_node, nd);
2240 if (new->start < nd->start)
2241 p = &(*p)->rb_left;
2242 else if (new->end > nd->end)
2243 p = &(*p)->rb_right;
2244 else
2245 BUG();
2246 }
2247 rb_link_node(&new->nd, parent, p);
2248 rb_insert_color(&new->nd, &sp->root);
2249 pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2250 new->policy ? new->policy->mode : 0);
2251 }
2252
2253 /* Find shared policy intersecting idx */
2254 struct mempolicy *
mpol_shared_policy_lookup(struct shared_policy * sp,unsigned long idx)2255 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2256 {
2257 struct mempolicy *pol = NULL;
2258 struct sp_node *sn;
2259
2260 if (!sp->root.rb_node)
2261 return NULL;
2262 spin_lock(&sp->lock);
2263 sn = sp_lookup(sp, idx, idx+1);
2264 if (sn) {
2265 mpol_get(sn->policy);
2266 pol = sn->policy;
2267 }
2268 spin_unlock(&sp->lock);
2269 return pol;
2270 }
2271
sp_free(struct sp_node * n)2272 static void sp_free(struct sp_node *n)
2273 {
2274 mpol_put(n->policy);
2275 kmem_cache_free(sn_cache, n);
2276 }
2277
2278 /**
2279 * mpol_misplaced - check whether current page node is valid in policy
2280 *
2281 * @page: page to be checked
2282 * @vma: vm area where page mapped
2283 * @addr: virtual address where page mapped
2284 *
2285 * Lookup current policy node id for vma,addr and "compare to" page's
2286 * node id.
2287 *
2288 * Returns:
2289 * -1 - not misplaced, page is in the right node
2290 * node - node id where the page should be
2291 *
2292 * Policy determination "mimics" alloc_page_vma().
2293 * Called from fault path where we know the vma and faulting address.
2294 */
mpol_misplaced(struct page * page,struct vm_area_struct * vma,unsigned long addr)2295 int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2296 {
2297 struct mempolicy *pol;
2298 struct zone *zone;
2299 int curnid = page_to_nid(page);
2300 unsigned long pgoff;
2301 int thiscpu = raw_smp_processor_id();
2302 int thisnid = cpu_to_node(thiscpu);
2303 int polnid = -1;
2304 int ret = -1;
2305
2306 BUG_ON(!vma);
2307
2308 pol = get_vma_policy(vma, addr);
2309 if (!(pol->flags & MPOL_F_MOF))
2310 goto out;
2311
2312 switch (pol->mode) {
2313 case MPOL_INTERLEAVE:
2314 BUG_ON(addr >= vma->vm_end);
2315 BUG_ON(addr < vma->vm_start);
2316
2317 pgoff = vma->vm_pgoff;
2318 pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2319 polnid = offset_il_node(pol, vma, pgoff);
2320 break;
2321
2322 case MPOL_PREFERRED:
2323 if (pol->flags & MPOL_F_LOCAL)
2324 polnid = numa_node_id();
2325 else
2326 polnid = pol->v.preferred_node;
2327 break;
2328
2329 case MPOL_BIND:
2330 /*
2331 * allows binding to multiple nodes.
2332 * use current page if in policy nodemask,
2333 * else select nearest allowed node, if any.
2334 * If no allowed nodes, use current [!misplaced].
2335 */
2336 if (node_isset(curnid, pol->v.nodes))
2337 goto out;
2338 (void)first_zones_zonelist(
2339 node_zonelist(numa_node_id(), GFP_HIGHUSER),
2340 gfp_zone(GFP_HIGHUSER),
2341 &pol->v.nodes, &zone);
2342 polnid = zone->node;
2343 break;
2344
2345 default:
2346 BUG();
2347 }
2348
2349 /* Migrate the page towards the node whose CPU is referencing it */
2350 if (pol->flags & MPOL_F_MORON) {
2351 polnid = thisnid;
2352
2353 if (!should_numa_migrate_memory(current, page, curnid, thiscpu))
2354 goto out;
2355 }
2356
2357 if (curnid != polnid)
2358 ret = polnid;
2359 out:
2360 mpol_cond_put(pol);
2361
2362 return ret;
2363 }
2364
sp_delete(struct shared_policy * sp,struct sp_node * n)2365 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2366 {
2367 pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2368 rb_erase(&n->nd, &sp->root);
2369 sp_free(n);
2370 }
2371
sp_node_init(struct sp_node * node,unsigned long start,unsigned long end,struct mempolicy * pol)2372 static void sp_node_init(struct sp_node *node, unsigned long start,
2373 unsigned long end, struct mempolicy *pol)
2374 {
2375 node->start = start;
2376 node->end = end;
2377 node->policy = pol;
2378 }
2379
sp_alloc(unsigned long start,unsigned long end,struct mempolicy * pol)2380 static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2381 struct mempolicy *pol)
2382 {
2383 struct sp_node *n;
2384 struct mempolicy *newpol;
2385
2386 n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2387 if (!n)
2388 return NULL;
2389
2390 newpol = mpol_dup(pol);
2391 if (IS_ERR(newpol)) {
2392 kmem_cache_free(sn_cache, n);
2393 return NULL;
2394 }
2395 newpol->flags |= MPOL_F_SHARED;
2396 sp_node_init(n, start, end, newpol);
2397
2398 return n;
2399 }
2400
2401 /* Replace a policy range. */
shared_policy_replace(struct shared_policy * sp,unsigned long start,unsigned long end,struct sp_node * new)2402 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2403 unsigned long end, struct sp_node *new)
2404 {
2405 struct sp_node *n;
2406 struct sp_node *n_new = NULL;
2407 struct mempolicy *mpol_new = NULL;
2408 int ret = 0;
2409
2410 restart:
2411 spin_lock(&sp->lock);
2412 n = sp_lookup(sp, start, end);
2413 /* Take care of old policies in the same range. */
2414 while (n && n->start < end) {
2415 struct rb_node *next = rb_next(&n->nd);
2416 if (n->start >= start) {
2417 if (n->end <= end)
2418 sp_delete(sp, n);
2419 else
2420 n->start = end;
2421 } else {
2422 /* Old policy spanning whole new range. */
2423 if (n->end > end) {
2424 if (!n_new)
2425 goto alloc_new;
2426
2427 *mpol_new = *n->policy;
2428 atomic_set(&mpol_new->refcnt, 1);
2429 sp_node_init(n_new, end, n->end, mpol_new);
2430 n->end = start;
2431 sp_insert(sp, n_new);
2432 n_new = NULL;
2433 mpol_new = NULL;
2434 break;
2435 } else
2436 n->end = start;
2437 }
2438 if (!next)
2439 break;
2440 n = rb_entry(next, struct sp_node, nd);
2441 }
2442 if (new)
2443 sp_insert(sp, new);
2444 spin_unlock(&sp->lock);
2445 ret = 0;
2446
2447 err_out:
2448 if (mpol_new)
2449 mpol_put(mpol_new);
2450 if (n_new)
2451 kmem_cache_free(sn_cache, n_new);
2452
2453 return ret;
2454
2455 alloc_new:
2456 spin_unlock(&sp->lock);
2457 ret = -ENOMEM;
2458 n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2459 if (!n_new)
2460 goto err_out;
2461 mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2462 if (!mpol_new)
2463 goto err_out;
2464 goto restart;
2465 }
2466
2467 /**
2468 * mpol_shared_policy_init - initialize shared policy for inode
2469 * @sp: pointer to inode shared policy
2470 * @mpol: struct mempolicy to install
2471 *
2472 * Install non-NULL @mpol in inode's shared policy rb-tree.
2473 * On entry, the current task has a reference on a non-NULL @mpol.
2474 * This must be released on exit.
2475 * This is called at get_inode() calls and we can use GFP_KERNEL.
2476 */
mpol_shared_policy_init(struct shared_policy * sp,struct mempolicy * mpol)2477 void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2478 {
2479 int ret;
2480
2481 sp->root = RB_ROOT; /* empty tree == default mempolicy */
2482 spin_lock_init(&sp->lock);
2483
2484 if (mpol) {
2485 struct vm_area_struct pvma;
2486 struct mempolicy *new;
2487 NODEMASK_SCRATCH(scratch);
2488
2489 if (!scratch)
2490 goto put_mpol;
2491 /* contextualize the tmpfs mount point mempolicy */
2492 new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2493 if (IS_ERR(new))
2494 goto free_scratch; /* no valid nodemask intersection */
2495
2496 task_lock(current);
2497 ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2498 task_unlock(current);
2499 if (ret)
2500 goto put_new;
2501
2502 /* Create pseudo-vma that contains just the policy */
2503 memset(&pvma, 0, sizeof(struct vm_area_struct));
2504 pvma.vm_end = TASK_SIZE; /* policy covers entire file */
2505 mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2506
2507 put_new:
2508 mpol_put(new); /* drop initial ref */
2509 free_scratch:
2510 NODEMASK_SCRATCH_FREE(scratch);
2511 put_mpol:
2512 mpol_put(mpol); /* drop our incoming ref on sb mpol */
2513 }
2514 }
2515
mpol_set_shared_policy(struct shared_policy * info,struct vm_area_struct * vma,struct mempolicy * npol)2516 int mpol_set_shared_policy(struct shared_policy *info,
2517 struct vm_area_struct *vma, struct mempolicy *npol)
2518 {
2519 int err;
2520 struct sp_node *new = NULL;
2521 unsigned long sz = vma_pages(vma);
2522
2523 pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2524 vma->vm_pgoff,
2525 sz, npol ? npol->mode : -1,
2526 npol ? npol->flags : -1,
2527 npol ? nodes_addr(npol->v.nodes)[0] : NUMA_NO_NODE);
2528
2529 if (npol) {
2530 new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2531 if (!new)
2532 return -ENOMEM;
2533 }
2534 err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2535 if (err && new)
2536 sp_free(new);
2537 return err;
2538 }
2539
2540 /* Free a backing policy store on inode delete. */
mpol_free_shared_policy(struct shared_policy * p)2541 void mpol_free_shared_policy(struct shared_policy *p)
2542 {
2543 struct sp_node *n;
2544 struct rb_node *next;
2545
2546 if (!p->root.rb_node)
2547 return;
2548 spin_lock(&p->lock);
2549 next = rb_first(&p->root);
2550 while (next) {
2551 n = rb_entry(next, struct sp_node, nd);
2552 next = rb_next(&n->nd);
2553 sp_delete(p, n);
2554 }
2555 spin_unlock(&p->lock);
2556 }
2557
2558 #ifdef CONFIG_NUMA_BALANCING
2559 static int __initdata numabalancing_override;
2560
check_numabalancing_enable(void)2561 static void __init check_numabalancing_enable(void)
2562 {
2563 bool numabalancing_default = false;
2564
2565 if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2566 numabalancing_default = true;
2567
2568 /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2569 if (numabalancing_override)
2570 set_numabalancing_state(numabalancing_override == 1);
2571
2572 if (num_online_nodes() > 1 && !numabalancing_override) {
2573 pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
2574 numabalancing_default ? "Enabling" : "Disabling");
2575 set_numabalancing_state(numabalancing_default);
2576 }
2577 }
2578
setup_numabalancing(char * str)2579 static int __init setup_numabalancing(char *str)
2580 {
2581 int ret = 0;
2582 if (!str)
2583 goto out;
2584
2585 if (!strcmp(str, "enable")) {
2586 numabalancing_override = 1;
2587 ret = 1;
2588 } else if (!strcmp(str, "disable")) {
2589 numabalancing_override = -1;
2590 ret = 1;
2591 }
2592 out:
2593 if (!ret)
2594 pr_warn("Unable to parse numa_balancing=\n");
2595
2596 return ret;
2597 }
2598 __setup("numa_balancing=", setup_numabalancing);
2599 #else
check_numabalancing_enable(void)2600 static inline void __init check_numabalancing_enable(void)
2601 {
2602 }
2603 #endif /* CONFIG_NUMA_BALANCING */
2604
2605 /* assumes fs == KERNEL_DS */
numa_policy_init(void)2606 void __init numa_policy_init(void)
2607 {
2608 nodemask_t interleave_nodes;
2609 unsigned long largest = 0;
2610 int nid, prefer = 0;
2611
2612 policy_cache = kmem_cache_create("numa_policy",
2613 sizeof(struct mempolicy),
2614 0, SLAB_PANIC, NULL);
2615
2616 sn_cache = kmem_cache_create("shared_policy_node",
2617 sizeof(struct sp_node),
2618 0, SLAB_PANIC, NULL);
2619
2620 for_each_node(nid) {
2621 preferred_node_policy[nid] = (struct mempolicy) {
2622 .refcnt = ATOMIC_INIT(1),
2623 .mode = MPOL_PREFERRED,
2624 .flags = MPOL_F_MOF | MPOL_F_MORON,
2625 .v = { .preferred_node = nid, },
2626 };
2627 }
2628
2629 /*
2630 * Set interleaving policy for system init. Interleaving is only
2631 * enabled across suitably sized nodes (default is >= 16MB), or
2632 * fall back to the largest node if they're all smaller.
2633 */
2634 nodes_clear(interleave_nodes);
2635 for_each_node_state(nid, N_MEMORY) {
2636 unsigned long total_pages = node_present_pages(nid);
2637
2638 /* Preserve the largest node */
2639 if (largest < total_pages) {
2640 largest = total_pages;
2641 prefer = nid;
2642 }
2643
2644 /* Interleave this node? */
2645 if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2646 node_set(nid, interleave_nodes);
2647 }
2648
2649 /* All too small, use the largest */
2650 if (unlikely(nodes_empty(interleave_nodes)))
2651 node_set(prefer, interleave_nodes);
2652
2653 if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2654 pr_err("%s: interleaving failed\n", __func__);
2655
2656 check_numabalancing_enable();
2657 }
2658
2659 /* Reset policy of current process to default */
numa_default_policy(void)2660 void numa_default_policy(void)
2661 {
2662 do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2663 }
2664
2665 /*
2666 * Parse and format mempolicy from/to strings
2667 */
2668
2669 /*
2670 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
2671 */
2672 static const char * const policy_modes[] =
2673 {
2674 [MPOL_DEFAULT] = "default",
2675 [MPOL_PREFERRED] = "prefer",
2676 [MPOL_BIND] = "bind",
2677 [MPOL_INTERLEAVE] = "interleave",
2678 [MPOL_LOCAL] = "local",
2679 };
2680
2681
2682 #ifdef CONFIG_TMPFS
2683 /**
2684 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2685 * @str: string containing mempolicy to parse
2686 * @mpol: pointer to struct mempolicy pointer, returned on success.
2687 *
2688 * Format of input:
2689 * <mode>[=<flags>][:<nodelist>]
2690 *
2691 * On success, returns 0, else 1
2692 */
mpol_parse_str(char * str,struct mempolicy ** mpol)2693 int mpol_parse_str(char *str, struct mempolicy **mpol)
2694 {
2695 struct mempolicy *new = NULL;
2696 unsigned short mode;
2697 unsigned short mode_flags;
2698 nodemask_t nodes;
2699 char *nodelist = strchr(str, ':');
2700 char *flags = strchr(str, '=');
2701 int err = 1;
2702
2703 if (flags)
2704 *flags++ = '\0'; /* terminate mode string */
2705
2706 if (nodelist) {
2707 /* NUL-terminate mode or flags string */
2708 *nodelist++ = '\0';
2709 if (nodelist_parse(nodelist, nodes))
2710 goto out;
2711 if (!nodes_subset(nodes, node_states[N_MEMORY]))
2712 goto out;
2713 } else
2714 nodes_clear(nodes);
2715
2716 for (mode = 0; mode < MPOL_MAX; mode++) {
2717 if (!strcmp(str, policy_modes[mode])) {
2718 break;
2719 }
2720 }
2721 if (mode >= MPOL_MAX)
2722 goto out;
2723
2724 switch (mode) {
2725 case MPOL_PREFERRED:
2726 /*
2727 * Insist on a nodelist of one node only, although later
2728 * we use first_node(nodes) to grab a single node, so here
2729 * nodelist (or nodes) cannot be empty.
2730 */
2731 if (nodelist) {
2732 char *rest = nodelist;
2733 while (isdigit(*rest))
2734 rest++;
2735 if (*rest)
2736 goto out;
2737 if (nodes_empty(nodes))
2738 goto out;
2739 }
2740 break;
2741 case MPOL_INTERLEAVE:
2742 /*
2743 * Default to online nodes with memory if no nodelist
2744 */
2745 if (!nodelist)
2746 nodes = node_states[N_MEMORY];
2747 break;
2748 case MPOL_LOCAL:
2749 /*
2750 * Don't allow a nodelist; mpol_new() checks flags
2751 */
2752 if (nodelist)
2753 goto out;
2754 mode = MPOL_PREFERRED;
2755 break;
2756 case MPOL_DEFAULT:
2757 /*
2758 * Insist on a empty nodelist
2759 */
2760 if (!nodelist)
2761 err = 0;
2762 goto out;
2763 case MPOL_BIND:
2764 /*
2765 * Insist on a nodelist
2766 */
2767 if (!nodelist)
2768 goto out;
2769 }
2770
2771 mode_flags = 0;
2772 if (flags) {
2773 /*
2774 * Currently, we only support two mutually exclusive
2775 * mode flags.
2776 */
2777 if (!strcmp(flags, "static"))
2778 mode_flags |= MPOL_F_STATIC_NODES;
2779 else if (!strcmp(flags, "relative"))
2780 mode_flags |= MPOL_F_RELATIVE_NODES;
2781 else
2782 goto out;
2783 }
2784
2785 new = mpol_new(mode, mode_flags, &nodes);
2786 if (IS_ERR(new))
2787 goto out;
2788
2789 /*
2790 * Save nodes for mpol_to_str() to show the tmpfs mount options
2791 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2792 */
2793 if (mode != MPOL_PREFERRED)
2794 new->v.nodes = nodes;
2795 else if (nodelist)
2796 new->v.preferred_node = first_node(nodes);
2797 else
2798 new->flags |= MPOL_F_LOCAL;
2799
2800 /*
2801 * Save nodes for contextualization: this will be used to "clone"
2802 * the mempolicy in a specific context [cpuset] at a later time.
2803 */
2804 new->w.user_nodemask = nodes;
2805
2806 err = 0;
2807
2808 out:
2809 /* Restore string for error message */
2810 if (nodelist)
2811 *--nodelist = ':';
2812 if (flags)
2813 *--flags = '=';
2814 if (!err)
2815 *mpol = new;
2816 return err;
2817 }
2818 #endif /* CONFIG_TMPFS */
2819
2820 /**
2821 * mpol_to_str - format a mempolicy structure for printing
2822 * @buffer: to contain formatted mempolicy string
2823 * @maxlen: length of @buffer
2824 * @pol: pointer to mempolicy to be formatted
2825 *
2826 * Convert @pol into a string. If @buffer is too short, truncate the string.
2827 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
2828 * longest flag, "relative", and to display at least a few node ids.
2829 */
mpol_to_str(char * buffer,int maxlen,struct mempolicy * pol)2830 void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
2831 {
2832 char *p = buffer;
2833 nodemask_t nodes = NODE_MASK_NONE;
2834 unsigned short mode = MPOL_DEFAULT;
2835 unsigned short flags = 0;
2836
2837 if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
2838 mode = pol->mode;
2839 flags = pol->flags;
2840 }
2841
2842 switch (mode) {
2843 case MPOL_DEFAULT:
2844 break;
2845 case MPOL_PREFERRED:
2846 if (flags & MPOL_F_LOCAL)
2847 mode = MPOL_LOCAL;
2848 else
2849 node_set(pol->v.preferred_node, nodes);
2850 break;
2851 case MPOL_BIND:
2852 case MPOL_INTERLEAVE:
2853 nodes = pol->v.nodes;
2854 break;
2855 default:
2856 WARN_ON_ONCE(1);
2857 snprintf(p, maxlen, "unknown");
2858 return;
2859 }
2860
2861 p += snprintf(p, maxlen, "%s", policy_modes[mode]);
2862
2863 if (flags & MPOL_MODE_FLAGS) {
2864 p += snprintf(p, buffer + maxlen - p, "=");
2865
2866 /*
2867 * Currently, the only defined flags are mutually exclusive
2868 */
2869 if (flags & MPOL_F_STATIC_NODES)
2870 p += snprintf(p, buffer + maxlen - p, "static");
2871 else if (flags & MPOL_F_RELATIVE_NODES)
2872 p += snprintf(p, buffer + maxlen - p, "relative");
2873 }
2874
2875 if (!nodes_empty(nodes))
2876 p += scnprintf(p, buffer + maxlen - p, ":%*pbl",
2877 nodemask_pr_args(&nodes));
2878 }
2879