Commit | Line | Data |
---|---|---|
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> | |
6e84f315 | 76 | #include <linux/sched/mm.h> |
6a3827d7 | 77 | #include <linux/sched/numa_balancing.h> |
f719ff9b | 78 | #include <linux/sched/task.h> |
1da177e4 LT |
79 | #include <linux/nodemask.h> |
80 | #include <linux/cpuset.h> | |
1da177e4 LT |
81 | #include <linux/slab.h> |
82 | #include <linux/string.h> | |
b95f1b31 | 83 | #include <linux/export.h> |
b488893a | 84 | #include <linux/nsproxy.h> |
1da177e4 LT |
85 | #include <linux/interrupt.h> |
86 | #include <linux/init.h> | |
87 | #include <linux/compat.h> | |
dc9aa5b9 | 88 | #include <linux/swap.h> |
1a75a6c8 CL |
89 | #include <linux/seq_file.h> |
90 | #include <linux/proc_fs.h> | |
b20a3503 | 91 | #include <linux/migrate.h> |
62b61f61 | 92 | #include <linux/ksm.h> |
95a402c3 | 93 | #include <linux/rmap.h> |
86c3a764 | 94 | #include <linux/security.h> |
dbcb0f19 | 95 | #include <linux/syscalls.h> |
095f1fc4 | 96 | #include <linux/ctype.h> |
6d9c285a | 97 | #include <linux/mm_inline.h> |
b24f53a0 | 98 | #include <linux/mmu_notifier.h> |
b1de0d13 | 99 | #include <linux/printk.h> |
dc9aa5b9 | 100 | |
1da177e4 | 101 | #include <asm/tlbflush.h> |
7c0f6ba6 | 102 | #include <linux/uaccess.h> |
1da177e4 | 103 | |
62695a84 NP |
104 | #include "internal.h" |
105 | ||
38e35860 | 106 | /* Internal flags */ |
dc9aa5b9 | 107 | #define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */ |
38e35860 | 108 | #define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */ |
dc9aa5b9 | 109 | |
fcc234f8 PE |
110 | static struct kmem_cache *policy_cache; |
111 | static struct kmem_cache *sn_cache; | |
1da177e4 | 112 | |
1da177e4 LT |
113 | /* Highest zone. An specific allocation for a zone below that is not |
114 | policied. */ | |
6267276f | 115 | enum zone_type policy_zone = 0; |
1da177e4 | 116 | |
bea904d5 LS |
117 | /* |
118 | * run-time system-wide default policy => local allocation | |
119 | */ | |
e754d79d | 120 | static struct mempolicy default_policy = { |
1da177e4 | 121 | .refcnt = ATOMIC_INIT(1), /* never free it */ |
bea904d5 | 122 | .mode = MPOL_PREFERRED, |
fc36b8d3 | 123 | .flags = MPOL_F_LOCAL, |
1da177e4 LT |
124 | }; |
125 | ||
5606e387 MG |
126 | static struct mempolicy preferred_node_policy[MAX_NUMNODES]; |
127 | ||
74d2c3a0 | 128 | struct mempolicy *get_task_policy(struct task_struct *p) |
5606e387 MG |
129 | { |
130 | struct mempolicy *pol = p->mempolicy; | |
f15ca78e | 131 | int node; |
5606e387 | 132 | |
f15ca78e ON |
133 | if (pol) |
134 | return pol; | |
5606e387 | 135 | |
f15ca78e ON |
136 | node = numa_node_id(); |
137 | if (node != NUMA_NO_NODE) { | |
138 | pol = &preferred_node_policy[node]; | |
139 | /* preferred_node_policy is not initialised early in boot */ | |
140 | if (pol->mode) | |
141 | return pol; | |
5606e387 MG |
142 | } |
143 | ||
f15ca78e | 144 | return &default_policy; |
5606e387 MG |
145 | } |
146 | ||
37012946 DR |
147 | static const struct mempolicy_operations { |
148 | int (*create)(struct mempolicy *pol, const nodemask_t *nodes); | |
213980c0 | 149 | void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes); |
37012946 DR |
150 | } mpol_ops[MPOL_MAX]; |
151 | ||
f5b087b5 DR |
152 | static inline int mpol_store_user_nodemask(const struct mempolicy *pol) |
153 | { | |
6d556294 | 154 | return pol->flags & MPOL_MODE_FLAGS; |
4c50bc01 DR |
155 | } |
156 | ||
157 | static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig, | |
158 | const nodemask_t *rel) | |
159 | { | |
160 | nodemask_t tmp; | |
161 | nodes_fold(tmp, *orig, nodes_weight(*rel)); | |
162 | nodes_onto(*ret, tmp, *rel); | |
f5b087b5 DR |
163 | } |
164 | ||
37012946 DR |
165 | static int mpol_new_interleave(struct mempolicy *pol, const nodemask_t *nodes) |
166 | { | |
167 | if (nodes_empty(*nodes)) | |
168 | return -EINVAL; | |
169 | pol->v.nodes = *nodes; | |
170 | return 0; | |
171 | } | |
172 | ||
173 | static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes) | |
174 | { | |
175 | if (!nodes) | |
fc36b8d3 | 176 | pol->flags |= MPOL_F_LOCAL; /* local allocation */ |
37012946 DR |
177 | else if (nodes_empty(*nodes)) |
178 | return -EINVAL; /* no allowed nodes */ | |
179 | else | |
180 | pol->v.preferred_node = first_node(*nodes); | |
181 | return 0; | |
182 | } | |
183 | ||
184 | static int mpol_new_bind(struct mempolicy *pol, const nodemask_t *nodes) | |
185 | { | |
859f7ef1 | 186 | if (nodes_empty(*nodes)) |
37012946 DR |
187 | return -EINVAL; |
188 | pol->v.nodes = *nodes; | |
189 | return 0; | |
190 | } | |
191 | ||
58568d2a MX |
192 | /* |
193 | * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if | |
194 | * any, for the new policy. mpol_new() has already validated the nodes | |
195 | * parameter with respect to the policy mode and flags. But, we need to | |
196 | * handle an empty nodemask with MPOL_PREFERRED here. | |
197 | * | |
198 | * Must be called holding task's alloc_lock to protect task's mems_allowed | |
199 | * and mempolicy. May also be called holding the mmap_semaphore for write. | |
200 | */ | |
4bfc4495 KH |
201 | static int mpol_set_nodemask(struct mempolicy *pol, |
202 | const nodemask_t *nodes, struct nodemask_scratch *nsc) | |
58568d2a | 203 | { |
58568d2a MX |
204 | int ret; |
205 | ||
206 | /* if mode is MPOL_DEFAULT, pol is NULL. This is right. */ | |
207 | if (pol == NULL) | |
208 | return 0; | |
01f13bd6 | 209 | /* Check N_MEMORY */ |
4bfc4495 | 210 | nodes_and(nsc->mask1, |
01f13bd6 | 211 | cpuset_current_mems_allowed, node_states[N_MEMORY]); |
58568d2a MX |
212 | |
213 | VM_BUG_ON(!nodes); | |
214 | if (pol->mode == MPOL_PREFERRED && nodes_empty(*nodes)) | |
215 | nodes = NULL; /* explicit local allocation */ | |
216 | else { | |
217 | if (pol->flags & MPOL_F_RELATIVE_NODES) | |
859f7ef1 | 218 | mpol_relative_nodemask(&nsc->mask2, nodes, &nsc->mask1); |
58568d2a | 219 | else |
4bfc4495 KH |
220 | nodes_and(nsc->mask2, *nodes, nsc->mask1); |
221 | ||
58568d2a MX |
222 | if (mpol_store_user_nodemask(pol)) |
223 | pol->w.user_nodemask = *nodes; | |
224 | else | |
225 | pol->w.cpuset_mems_allowed = | |
226 | cpuset_current_mems_allowed; | |
227 | } | |
228 | ||
4bfc4495 KH |
229 | if (nodes) |
230 | ret = mpol_ops[pol->mode].create(pol, &nsc->mask2); | |
231 | else | |
232 | ret = mpol_ops[pol->mode].create(pol, NULL); | |
58568d2a MX |
233 | return ret; |
234 | } | |
235 | ||
236 | /* | |
237 | * This function just creates a new policy, does some check and simple | |
238 | * initialization. You must invoke mpol_set_nodemask() to set nodes. | |
239 | */ | |
028fec41 DR |
240 | static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags, |
241 | nodemask_t *nodes) | |
1da177e4 LT |
242 | { |
243 | struct mempolicy *policy; | |
244 | ||
028fec41 | 245 | pr_debug("setting mode %d flags %d nodes[0] %lx\n", |
00ef2d2f | 246 | mode, flags, nodes ? nodes_addr(*nodes)[0] : NUMA_NO_NODE); |
140d5a49 | 247 | |
3e1f0645 DR |
248 | if (mode == MPOL_DEFAULT) { |
249 | if (nodes && !nodes_empty(*nodes)) | |
37012946 | 250 | return ERR_PTR(-EINVAL); |
d3a71033 | 251 | return NULL; |
37012946 | 252 | } |
3e1f0645 DR |
253 | VM_BUG_ON(!nodes); |
254 | ||
255 | /* | |
256 | * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or | |
257 | * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation). | |
258 | * All other modes require a valid pointer to a non-empty nodemask. | |
259 | */ | |
260 | if (mode == MPOL_PREFERRED) { | |
261 | if (nodes_empty(*nodes)) { | |
262 | if (((flags & MPOL_F_STATIC_NODES) || | |
263 | (flags & MPOL_F_RELATIVE_NODES))) | |
264 | return ERR_PTR(-EINVAL); | |
3e1f0645 | 265 | } |
479e2802 | 266 | } else if (mode == MPOL_LOCAL) { |
8d303e44 PK |
267 | if (!nodes_empty(*nodes) || |
268 | (flags & MPOL_F_STATIC_NODES) || | |
269 | (flags & MPOL_F_RELATIVE_NODES)) | |
479e2802 PZ |
270 | return ERR_PTR(-EINVAL); |
271 | mode = MPOL_PREFERRED; | |
3e1f0645 DR |
272 | } else if (nodes_empty(*nodes)) |
273 | return ERR_PTR(-EINVAL); | |
1da177e4 LT |
274 | policy = kmem_cache_alloc(policy_cache, GFP_KERNEL); |
275 | if (!policy) | |
276 | return ERR_PTR(-ENOMEM); | |
277 | atomic_set(&policy->refcnt, 1); | |
45c4745a | 278 | policy->mode = mode; |
3e1f0645 | 279 | policy->flags = flags; |
37012946 | 280 | |
1da177e4 | 281 | return policy; |
37012946 DR |
282 | } |
283 | ||
52cd3b07 LS |
284 | /* Slow path of a mpol destructor. */ |
285 | void __mpol_put(struct mempolicy *p) | |
286 | { | |
287 | if (!atomic_dec_and_test(&p->refcnt)) | |
288 | return; | |
52cd3b07 LS |
289 | kmem_cache_free(policy_cache, p); |
290 | } | |
291 | ||
213980c0 | 292 | static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes) |
37012946 DR |
293 | { |
294 | } | |
295 | ||
213980c0 | 296 | static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes) |
37012946 DR |
297 | { |
298 | nodemask_t tmp; | |
299 | ||
300 | if (pol->flags & MPOL_F_STATIC_NODES) | |
301 | nodes_and(tmp, pol->w.user_nodemask, *nodes); | |
302 | else if (pol->flags & MPOL_F_RELATIVE_NODES) | |
303 | mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes); | |
304 | else { | |
213980c0 VB |
305 | nodes_remap(tmp, pol->v.nodes,pol->w.cpuset_mems_allowed, |
306 | *nodes); | |
307 | pol->w.cpuset_mems_allowed = tmp; | |
37012946 | 308 | } |
f5b087b5 | 309 | |
708c1bbc MX |
310 | if (nodes_empty(tmp)) |
311 | tmp = *nodes; | |
312 | ||
213980c0 | 313 | pol->v.nodes = tmp; |
37012946 DR |
314 | } |
315 | ||
316 | static void mpol_rebind_preferred(struct mempolicy *pol, | |
213980c0 | 317 | const nodemask_t *nodes) |
37012946 DR |
318 | { |
319 | nodemask_t tmp; | |
320 | ||
37012946 DR |
321 | if (pol->flags & MPOL_F_STATIC_NODES) { |
322 | int node = first_node(pol->w.user_nodemask); | |
323 | ||
fc36b8d3 | 324 | if (node_isset(node, *nodes)) { |
37012946 | 325 | pol->v.preferred_node = node; |
fc36b8d3 LS |
326 | pol->flags &= ~MPOL_F_LOCAL; |
327 | } else | |
328 | pol->flags |= MPOL_F_LOCAL; | |
37012946 DR |
329 | } else if (pol->flags & MPOL_F_RELATIVE_NODES) { |
330 | mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes); | |
331 | pol->v.preferred_node = first_node(tmp); | |
fc36b8d3 | 332 | } else if (!(pol->flags & MPOL_F_LOCAL)) { |
37012946 DR |
333 | pol->v.preferred_node = node_remap(pol->v.preferred_node, |
334 | pol->w.cpuset_mems_allowed, | |
335 | *nodes); | |
336 | pol->w.cpuset_mems_allowed = *nodes; | |
337 | } | |
1da177e4 LT |
338 | } |
339 | ||
708c1bbc MX |
340 | /* |
341 | * mpol_rebind_policy - Migrate a policy to a different set of nodes | |
342 | * | |
213980c0 VB |
343 | * Per-vma policies are protected by mmap_sem. Allocations using per-task |
344 | * policies are protected by task->mems_allowed_seq to prevent a premature | |
345 | * OOM/allocation failure due to parallel nodemask modification. | |
708c1bbc | 346 | */ |
213980c0 | 347 | static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask) |
1d0d2680 | 348 | { |
1d0d2680 DR |
349 | if (!pol) |
350 | return; | |
213980c0 | 351 | if (!mpol_store_user_nodemask(pol) && |
1d0d2680 DR |
352 | nodes_equal(pol->w.cpuset_mems_allowed, *newmask)) |
353 | return; | |
708c1bbc | 354 | |
213980c0 | 355 | mpol_ops[pol->mode].rebind(pol, newmask); |
1d0d2680 DR |
356 | } |
357 | ||
358 | /* | |
359 | * Wrapper for mpol_rebind_policy() that just requires task | |
360 | * pointer, and updates task mempolicy. | |
58568d2a MX |
361 | * |
362 | * Called with task's alloc_lock held. | |
1d0d2680 DR |
363 | */ |
364 | ||
213980c0 | 365 | void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new) |
1d0d2680 | 366 | { |
213980c0 | 367 | mpol_rebind_policy(tsk->mempolicy, new); |
1d0d2680 DR |
368 | } |
369 | ||
370 | /* | |
371 | * Rebind each vma in mm to new nodemask. | |
372 | * | |
373 | * Call holding a reference to mm. Takes mm->mmap_sem during call. | |
374 | */ | |
375 | ||
376 | void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new) | |
377 | { | |
378 | struct vm_area_struct *vma; | |
379 | ||
380 | down_write(&mm->mmap_sem); | |
381 | for (vma = mm->mmap; vma; vma = vma->vm_next) | |
213980c0 | 382 | mpol_rebind_policy(vma->vm_policy, new); |
1d0d2680 DR |
383 | up_write(&mm->mmap_sem); |
384 | } | |
385 | ||
37012946 DR |
386 | static const struct mempolicy_operations mpol_ops[MPOL_MAX] = { |
387 | [MPOL_DEFAULT] = { | |
388 | .rebind = mpol_rebind_default, | |
389 | }, | |
390 | [MPOL_INTERLEAVE] = { | |
391 | .create = mpol_new_interleave, | |
392 | .rebind = mpol_rebind_nodemask, | |
393 | }, | |
394 | [MPOL_PREFERRED] = { | |
395 | .create = mpol_new_preferred, | |
396 | .rebind = mpol_rebind_preferred, | |
397 | }, | |
398 | [MPOL_BIND] = { | |
399 | .create = mpol_new_bind, | |
400 | .rebind = mpol_rebind_nodemask, | |
401 | }, | |
402 | }; | |
403 | ||
fc301289 CL |
404 | static void migrate_page_add(struct page *page, struct list_head *pagelist, |
405 | unsigned long flags); | |
1a75a6c8 | 406 | |
6f4576e3 NH |
407 | struct queue_pages { |
408 | struct list_head *pagelist; | |
409 | unsigned long flags; | |
410 | nodemask_t *nmask; | |
411 | struct vm_area_struct *prev; | |
412 | }; | |
413 | ||
88aaa2a1 NH |
414 | /* |
415 | * Check if the page's nid is in qp->nmask. | |
416 | * | |
417 | * If MPOL_MF_INVERT is set in qp->flags, check if the nid is | |
418 | * in the invert of qp->nmask. | |
419 | */ | |
420 | static inline bool queue_pages_required(struct page *page, | |
421 | struct queue_pages *qp) | |
422 | { | |
423 | int nid = page_to_nid(page); | |
424 | unsigned long flags = qp->flags; | |
425 | ||
426 | return node_isset(nid, *qp->nmask) == !(flags & MPOL_MF_INVERT); | |
427 | } | |
428 | ||
98094945 NH |
429 | /* |
430 | * Scan through pages checking if pages follow certain conditions, | |
431 | * and move them to the pagelist if they do. | |
432 | */ | |
6f4576e3 NH |
433 | static int queue_pages_pte_range(pmd_t *pmd, unsigned long addr, |
434 | unsigned long end, struct mm_walk *walk) | |
1da177e4 | 435 | { |
6f4576e3 NH |
436 | struct vm_area_struct *vma = walk->vma; |
437 | struct page *page; | |
438 | struct queue_pages *qp = walk->private; | |
439 | unsigned long flags = qp->flags; | |
248db92d | 440 | int nid, ret; |
91612e0d | 441 | pte_t *pte; |
705e87c0 | 442 | spinlock_t *ptl; |
941150a3 | 443 | |
248db92d KS |
444 | if (pmd_trans_huge(*pmd)) { |
445 | ptl = pmd_lock(walk->mm, pmd); | |
446 | if (pmd_trans_huge(*pmd)) { | |
447 | page = pmd_page(*pmd); | |
448 | if (is_huge_zero_page(page)) { | |
449 | spin_unlock(ptl); | |
fd60775a | 450 | __split_huge_pmd(vma, pmd, addr, false, NULL); |
248db92d KS |
451 | } else { |
452 | get_page(page); | |
453 | spin_unlock(ptl); | |
454 | lock_page(page); | |
455 | ret = split_huge_page(page); | |
456 | unlock_page(page); | |
457 | put_page(page); | |
458 | if (ret) | |
459 | return 0; | |
460 | } | |
461 | } else { | |
462 | spin_unlock(ptl); | |
463 | } | |
464 | } | |
91612e0d | 465 | |
337d9abf NH |
466 | if (pmd_trans_unstable(pmd)) |
467 | return 0; | |
248db92d | 468 | retry: |
6f4576e3 NH |
469 | pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl); |
470 | for (; addr != end; pte++, addr += PAGE_SIZE) { | |
91612e0d | 471 | if (!pte_present(*pte)) |
1da177e4 | 472 | continue; |
6aab341e LT |
473 | page = vm_normal_page(vma, addr, *pte); |
474 | if (!page) | |
1da177e4 | 475 | continue; |
053837fc | 476 | /* |
62b61f61 HD |
477 | * vm_normal_page() filters out zero pages, but there might |
478 | * still be PageReserved pages to skip, perhaps in a VDSO. | |
053837fc | 479 | */ |
b79bc0a0 | 480 | if (PageReserved(page)) |
f4598c8b | 481 | continue; |
88aaa2a1 | 482 | if (!queue_pages_required(page, qp)) |
38e35860 | 483 | continue; |
800d8c63 | 484 | if (PageTransCompound(page)) { |
248db92d KS |
485 | get_page(page); |
486 | pte_unmap_unlock(pte, ptl); | |
487 | lock_page(page); | |
488 | ret = split_huge_page(page); | |
489 | unlock_page(page); | |
490 | put_page(page); | |
491 | /* Failed to split -- skip. */ | |
492 | if (ret) { | |
493 | pte = pte_offset_map_lock(walk->mm, pmd, | |
494 | addr, &ptl); | |
495 | continue; | |
496 | } | |
497 | goto retry; | |
498 | } | |
38e35860 | 499 | |
77bf45e7 | 500 | migrate_page_add(page, qp->pagelist, flags); |
6f4576e3 NH |
501 | } |
502 | pte_unmap_unlock(pte - 1, ptl); | |
503 | cond_resched(); | |
504 | return 0; | |
91612e0d HD |
505 | } |
506 | ||
6f4576e3 NH |
507 | static int queue_pages_hugetlb(pte_t *pte, unsigned long hmask, |
508 | unsigned long addr, unsigned long end, | |
509 | struct mm_walk *walk) | |
e2d8cf40 NH |
510 | { |
511 | #ifdef CONFIG_HUGETLB_PAGE | |
6f4576e3 NH |
512 | struct queue_pages *qp = walk->private; |
513 | unsigned long flags = qp->flags; | |
e2d8cf40 | 514 | struct page *page; |
cb900f41 | 515 | spinlock_t *ptl; |
d4c54919 | 516 | pte_t entry; |
e2d8cf40 | 517 | |
6f4576e3 NH |
518 | ptl = huge_pte_lock(hstate_vma(walk->vma), walk->mm, pte); |
519 | entry = huge_ptep_get(pte); | |
d4c54919 NH |
520 | if (!pte_present(entry)) |
521 | goto unlock; | |
522 | page = pte_page(entry); | |
88aaa2a1 | 523 | if (!queue_pages_required(page, qp)) |
e2d8cf40 NH |
524 | goto unlock; |
525 | /* With MPOL_MF_MOVE, we migrate only unshared hugepage. */ | |
526 | if (flags & (MPOL_MF_MOVE_ALL) || | |
527 | (flags & MPOL_MF_MOVE && page_mapcount(page) == 1)) | |
6f4576e3 | 528 | isolate_huge_page(page, qp->pagelist); |
e2d8cf40 | 529 | unlock: |
cb900f41 | 530 | spin_unlock(ptl); |
e2d8cf40 NH |
531 | #else |
532 | BUG(); | |
533 | #endif | |
91612e0d | 534 | return 0; |
1da177e4 LT |
535 | } |
536 | ||
5877231f | 537 | #ifdef CONFIG_NUMA_BALANCING |
b24f53a0 | 538 | /* |
4b10e7d5 MG |
539 | * This is used to mark a range of virtual addresses to be inaccessible. |
540 | * These are later cleared by a NUMA hinting fault. Depending on these | |
541 | * faults, pages may be migrated for better NUMA placement. | |
542 | * | |
543 | * This is assuming that NUMA faults are handled using PROT_NONE. If | |
544 | * an architecture makes a different choice, it will need further | |
545 | * changes to the core. | |
b24f53a0 | 546 | */ |
4b10e7d5 MG |
547 | unsigned long change_prot_numa(struct vm_area_struct *vma, |
548 | unsigned long addr, unsigned long end) | |
b24f53a0 | 549 | { |
4b10e7d5 | 550 | int nr_updated; |
b24f53a0 | 551 | |
4d942466 | 552 | nr_updated = change_protection(vma, addr, end, PAGE_NONE, 0, 1); |
03c5a6e1 MG |
553 | if (nr_updated) |
554 | count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated); | |
b24f53a0 | 555 | |
4b10e7d5 | 556 | return nr_updated; |
b24f53a0 LS |
557 | } |
558 | #else | |
559 | static unsigned long change_prot_numa(struct vm_area_struct *vma, | |
560 | unsigned long addr, unsigned long end) | |
561 | { | |
562 | return 0; | |
563 | } | |
5877231f | 564 | #endif /* CONFIG_NUMA_BALANCING */ |
b24f53a0 | 565 | |
6f4576e3 NH |
566 | static int queue_pages_test_walk(unsigned long start, unsigned long end, |
567 | struct mm_walk *walk) | |
568 | { | |
569 | struct vm_area_struct *vma = walk->vma; | |
570 | struct queue_pages *qp = walk->private; | |
571 | unsigned long endvma = vma->vm_end; | |
572 | unsigned long flags = qp->flags; | |
573 | ||
77bf45e7 | 574 | if (!vma_migratable(vma)) |
48684a65 NH |
575 | return 1; |
576 | ||
6f4576e3 NH |
577 | if (endvma > end) |
578 | endvma = end; | |
579 | if (vma->vm_start > start) | |
580 | start = vma->vm_start; | |
581 | ||
582 | if (!(flags & MPOL_MF_DISCONTIG_OK)) { | |
583 | if (!vma->vm_next && vma->vm_end < end) | |
584 | return -EFAULT; | |
585 | if (qp->prev && qp->prev->vm_end < vma->vm_start) | |
586 | return -EFAULT; | |
587 | } | |
588 | ||
589 | qp->prev = vma; | |
590 | ||
6f4576e3 NH |
591 | if (flags & MPOL_MF_LAZY) { |
592 | /* Similar to task_numa_work, skip inaccessible VMAs */ | |
4355c018 LC |
593 | if (!is_vm_hugetlb_page(vma) && |
594 | (vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)) && | |
595 | !(vma->vm_flags & VM_MIXEDMAP)) | |
6f4576e3 NH |
596 | change_prot_numa(vma, start, endvma); |
597 | return 1; | |
598 | } | |
599 | ||
77bf45e7 KS |
600 | /* queue pages from current vma */ |
601 | if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) | |
6f4576e3 NH |
602 | return 0; |
603 | return 1; | |
604 | } | |
605 | ||
dc9aa5b9 | 606 | /* |
98094945 NH |
607 | * Walk through page tables and collect pages to be migrated. |
608 | * | |
609 | * If pages found in a given range are on a set of nodes (determined by | |
610 | * @nodes and @flags,) it's isolated and queued to the pagelist which is | |
611 | * passed via @private.) | |
dc9aa5b9 | 612 | */ |
d05f0cdc | 613 | static int |
98094945 | 614 | queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end, |
6f4576e3 NH |
615 | nodemask_t *nodes, unsigned long flags, |
616 | struct list_head *pagelist) | |
1da177e4 | 617 | { |
6f4576e3 NH |
618 | struct queue_pages qp = { |
619 | .pagelist = pagelist, | |
620 | .flags = flags, | |
621 | .nmask = nodes, | |
622 | .prev = NULL, | |
623 | }; | |
624 | struct mm_walk queue_pages_walk = { | |
625 | .hugetlb_entry = queue_pages_hugetlb, | |
626 | .pmd_entry = queue_pages_pte_range, | |
627 | .test_walk = queue_pages_test_walk, | |
628 | .mm = mm, | |
629 | .private = &qp, | |
630 | }; | |
631 | ||
632 | return walk_page_range(start, end, &queue_pages_walk); | |
1da177e4 LT |
633 | } |
634 | ||
869833f2 KM |
635 | /* |
636 | * Apply policy to a single VMA | |
637 | * This must be called with the mmap_sem held for writing. | |
638 | */ | |
639 | static int vma_replace_policy(struct vm_area_struct *vma, | |
640 | struct mempolicy *pol) | |
8d34694c | 641 | { |
869833f2 KM |
642 | int err; |
643 | struct mempolicy *old; | |
644 | struct mempolicy *new; | |
8d34694c KM |
645 | |
646 | pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n", | |
647 | vma->vm_start, vma->vm_end, vma->vm_pgoff, | |
648 | vma->vm_ops, vma->vm_file, | |
649 | vma->vm_ops ? vma->vm_ops->set_policy : NULL); | |
650 | ||
869833f2 KM |
651 | new = mpol_dup(pol); |
652 | if (IS_ERR(new)) | |
653 | return PTR_ERR(new); | |
654 | ||
655 | if (vma->vm_ops && vma->vm_ops->set_policy) { | |
8d34694c | 656 | err = vma->vm_ops->set_policy(vma, new); |
869833f2 KM |
657 | if (err) |
658 | goto err_out; | |
8d34694c | 659 | } |
869833f2 KM |
660 | |
661 | old = vma->vm_policy; | |
662 | vma->vm_policy = new; /* protected by mmap_sem */ | |
663 | mpol_put(old); | |
664 | ||
665 | return 0; | |
666 | err_out: | |
667 | mpol_put(new); | |
8d34694c KM |
668 | return err; |
669 | } | |
670 | ||
1da177e4 | 671 | /* Step 2: apply policy to a range and do splits. */ |
9d8cebd4 KM |
672 | static int mbind_range(struct mm_struct *mm, unsigned long start, |
673 | unsigned long end, struct mempolicy *new_pol) | |
1da177e4 LT |
674 | { |
675 | struct vm_area_struct *next; | |
9d8cebd4 KM |
676 | struct vm_area_struct *prev; |
677 | struct vm_area_struct *vma; | |
678 | int err = 0; | |
e26a5114 | 679 | pgoff_t pgoff; |
9d8cebd4 KM |
680 | unsigned long vmstart; |
681 | unsigned long vmend; | |
1da177e4 | 682 | |
097d5910 | 683 | vma = find_vma(mm, start); |
9d8cebd4 KM |
684 | if (!vma || vma->vm_start > start) |
685 | return -EFAULT; | |
686 | ||
097d5910 | 687 | prev = vma->vm_prev; |
e26a5114 KM |
688 | if (start > vma->vm_start) |
689 | prev = vma; | |
690 | ||
9d8cebd4 | 691 | for (; vma && vma->vm_start < end; prev = vma, vma = next) { |
1da177e4 | 692 | next = vma->vm_next; |
9d8cebd4 KM |
693 | vmstart = max(start, vma->vm_start); |
694 | vmend = min(end, vma->vm_end); | |
695 | ||
e26a5114 KM |
696 | if (mpol_equal(vma_policy(vma), new_pol)) |
697 | continue; | |
698 | ||
699 | pgoff = vma->vm_pgoff + | |
700 | ((vmstart - vma->vm_start) >> PAGE_SHIFT); | |
9d8cebd4 | 701 | prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags, |
19a809af AA |
702 | vma->anon_vma, vma->vm_file, pgoff, |
703 | new_pol, vma->vm_userfaultfd_ctx); | |
9d8cebd4 KM |
704 | if (prev) { |
705 | vma = prev; | |
706 | next = vma->vm_next; | |
3964acd0 ON |
707 | if (mpol_equal(vma_policy(vma), new_pol)) |
708 | continue; | |
709 | /* vma_merge() joined vma && vma->next, case 8 */ | |
710 | goto replace; | |
9d8cebd4 KM |
711 | } |
712 | if (vma->vm_start != vmstart) { | |
713 | err = split_vma(vma->vm_mm, vma, vmstart, 1); | |
714 | if (err) | |
715 | goto out; | |
716 | } | |
717 | if (vma->vm_end != vmend) { | |
718 | err = split_vma(vma->vm_mm, vma, vmend, 0); | |
719 | if (err) | |
720 | goto out; | |
721 | } | |
3964acd0 | 722 | replace: |
869833f2 | 723 | err = vma_replace_policy(vma, new_pol); |
8d34694c KM |
724 | if (err) |
725 | goto out; | |
1da177e4 | 726 | } |
9d8cebd4 KM |
727 | |
728 | out: | |
1da177e4 LT |
729 | return err; |
730 | } | |
731 | ||
1da177e4 | 732 | /* Set the process memory policy */ |
028fec41 DR |
733 | static long do_set_mempolicy(unsigned short mode, unsigned short flags, |
734 | nodemask_t *nodes) | |
1da177e4 | 735 | { |
58568d2a | 736 | struct mempolicy *new, *old; |
4bfc4495 | 737 | NODEMASK_SCRATCH(scratch); |
58568d2a | 738 | int ret; |
1da177e4 | 739 | |
4bfc4495 KH |
740 | if (!scratch) |
741 | return -ENOMEM; | |
f4e53d91 | 742 | |
4bfc4495 KH |
743 | new = mpol_new(mode, flags, nodes); |
744 | if (IS_ERR(new)) { | |
745 | ret = PTR_ERR(new); | |
746 | goto out; | |
747 | } | |
2c7c3a7d | 748 | |
58568d2a | 749 | task_lock(current); |
4bfc4495 | 750 | ret = mpol_set_nodemask(new, nodes, scratch); |
58568d2a MX |
751 | if (ret) { |
752 | task_unlock(current); | |
58568d2a | 753 | mpol_put(new); |
4bfc4495 | 754 | goto out; |
58568d2a MX |
755 | } |
756 | old = current->mempolicy; | |
1da177e4 | 757 | current->mempolicy = new; |
45816682 VB |
758 | if (new && new->mode == MPOL_INTERLEAVE) |
759 | current->il_prev = MAX_NUMNODES-1; | |
58568d2a | 760 | task_unlock(current); |
58568d2a | 761 | mpol_put(old); |
4bfc4495 KH |
762 | ret = 0; |
763 | out: | |
764 | NODEMASK_SCRATCH_FREE(scratch); | |
765 | return ret; | |
1da177e4 LT |
766 | } |
767 | ||
bea904d5 LS |
768 | /* |
769 | * Return nodemask for policy for get_mempolicy() query | |
58568d2a MX |
770 | * |
771 | * Called with task's alloc_lock held | |
bea904d5 LS |
772 | */ |
773 | static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes) | |
1da177e4 | 774 | { |
dfcd3c0d | 775 | nodes_clear(*nodes); |
bea904d5 LS |
776 | if (p == &default_policy) |
777 | return; | |
778 | ||
45c4745a | 779 | switch (p->mode) { |
19770b32 MG |
780 | case MPOL_BIND: |
781 | /* Fall through */ | |
1da177e4 | 782 | case MPOL_INTERLEAVE: |
dfcd3c0d | 783 | *nodes = p->v.nodes; |
1da177e4 LT |
784 | break; |
785 | case MPOL_PREFERRED: | |
fc36b8d3 | 786 | if (!(p->flags & MPOL_F_LOCAL)) |
dfcd3c0d | 787 | node_set(p->v.preferred_node, *nodes); |
53f2556b | 788 | /* else return empty node mask for local allocation */ |
1da177e4 LT |
789 | break; |
790 | default: | |
791 | BUG(); | |
792 | } | |
793 | } | |
794 | ||
d4edcf0d | 795 | static int lookup_node(unsigned long addr) |
1da177e4 LT |
796 | { |
797 | struct page *p; | |
798 | int err; | |
799 | ||
768ae309 | 800 | err = get_user_pages(addr & PAGE_MASK, 1, 0, &p, NULL); |
1da177e4 LT |
801 | if (err >= 0) { |
802 | err = page_to_nid(p); | |
803 | put_page(p); | |
804 | } | |
805 | return err; | |
806 | } | |
807 | ||
1da177e4 | 808 | /* Retrieve NUMA policy */ |
dbcb0f19 AB |
809 | static long do_get_mempolicy(int *policy, nodemask_t *nmask, |
810 | unsigned long addr, unsigned long flags) | |
1da177e4 | 811 | { |
8bccd85f | 812 | int err; |
1da177e4 LT |
813 | struct mm_struct *mm = current->mm; |
814 | struct vm_area_struct *vma = NULL; | |
815 | struct mempolicy *pol = current->mempolicy; | |
816 | ||
754af6f5 LS |
817 | if (flags & |
818 | ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED)) | |
1da177e4 | 819 | return -EINVAL; |
754af6f5 LS |
820 | |
821 | if (flags & MPOL_F_MEMS_ALLOWED) { | |
822 | if (flags & (MPOL_F_NODE|MPOL_F_ADDR)) | |
823 | return -EINVAL; | |
824 | *policy = 0; /* just so it's initialized */ | |
58568d2a | 825 | task_lock(current); |
754af6f5 | 826 | *nmask = cpuset_current_mems_allowed; |
58568d2a | 827 | task_unlock(current); |
754af6f5 LS |
828 | return 0; |
829 | } | |
830 | ||
1da177e4 | 831 | if (flags & MPOL_F_ADDR) { |
bea904d5 LS |
832 | /* |
833 | * Do NOT fall back to task policy if the | |
834 | * vma/shared policy at addr is NULL. We | |
835 | * want to return MPOL_DEFAULT in this case. | |
836 | */ | |
1da177e4 LT |
837 | down_read(&mm->mmap_sem); |
838 | vma = find_vma_intersection(mm, addr, addr+1); | |
839 | if (!vma) { | |
840 | up_read(&mm->mmap_sem); | |
841 | return -EFAULT; | |
842 | } | |
843 | if (vma->vm_ops && vma->vm_ops->get_policy) | |
844 | pol = vma->vm_ops->get_policy(vma, addr); | |
845 | else | |
846 | pol = vma->vm_policy; | |
847 | } else if (addr) | |
848 | return -EINVAL; | |
849 | ||
850 | if (!pol) | |
bea904d5 | 851 | pol = &default_policy; /* indicates default behavior */ |
1da177e4 LT |
852 | |
853 | if (flags & MPOL_F_NODE) { | |
854 | if (flags & MPOL_F_ADDR) { | |
d4edcf0d | 855 | err = lookup_node(addr); |
1da177e4 LT |
856 | if (err < 0) |
857 | goto out; | |
8bccd85f | 858 | *policy = err; |
1da177e4 | 859 | } else if (pol == current->mempolicy && |
45c4745a | 860 | pol->mode == MPOL_INTERLEAVE) { |
45816682 | 861 | *policy = next_node_in(current->il_prev, pol->v.nodes); |
1da177e4 LT |
862 | } else { |
863 | err = -EINVAL; | |
864 | goto out; | |
865 | } | |
bea904d5 LS |
866 | } else { |
867 | *policy = pol == &default_policy ? MPOL_DEFAULT : | |
868 | pol->mode; | |
d79df630 DR |
869 | /* |
870 | * Internal mempolicy flags must be masked off before exposing | |
871 | * the policy to userspace. | |
872 | */ | |
873 | *policy |= (pol->flags & MPOL_MODE_FLAGS); | |
bea904d5 | 874 | } |
1da177e4 | 875 | |
1da177e4 | 876 | err = 0; |
58568d2a | 877 | if (nmask) { |
c6b6ef8b LS |
878 | if (mpol_store_user_nodemask(pol)) { |
879 | *nmask = pol->w.user_nodemask; | |
880 | } else { | |
881 | task_lock(current); | |
882 | get_policy_nodemask(pol, nmask); | |
883 | task_unlock(current); | |
884 | } | |
58568d2a | 885 | } |
1da177e4 LT |
886 | |
887 | out: | |
52cd3b07 | 888 | mpol_cond_put(pol); |
1da177e4 LT |
889 | if (vma) |
890 | up_read(¤t->mm->mmap_sem); | |
891 | return err; | |
892 | } | |
893 | ||
b20a3503 | 894 | #ifdef CONFIG_MIGRATION |
6ce3c4c0 CL |
895 | /* |
896 | * page migration | |
897 | */ | |
fc301289 CL |
898 | static void migrate_page_add(struct page *page, struct list_head *pagelist, |
899 | unsigned long flags) | |
6ce3c4c0 CL |
900 | { |
901 | /* | |
fc301289 | 902 | * Avoid migrating a page that is shared with others. |
6ce3c4c0 | 903 | */ |
62695a84 NP |
904 | if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) { |
905 | if (!isolate_lru_page(page)) { | |
906 | list_add_tail(&page->lru, pagelist); | |
599d0c95 | 907 | inc_node_page_state(page, NR_ISOLATED_ANON + |
6d9c285a | 908 | page_is_file_cache(page)); |
62695a84 NP |
909 | } |
910 | } | |
7e2ab150 | 911 | } |
6ce3c4c0 | 912 | |
742755a1 | 913 | static struct page *new_node_page(struct page *page, unsigned long node, int **x) |
95a402c3 | 914 | { |
e2d8cf40 NH |
915 | if (PageHuge(page)) |
916 | return alloc_huge_page_node(page_hstate(compound_head(page)), | |
917 | node); | |
918 | else | |
96db800f | 919 | return __alloc_pages_node(node, GFP_HIGHUSER_MOVABLE | |
b360edb4 | 920 | __GFP_THISNODE, 0); |
95a402c3 CL |
921 | } |
922 | ||
7e2ab150 CL |
923 | /* |
924 | * Migrate pages from one node to a target node. | |
925 | * Returns error or the number of pages not migrated. | |
926 | */ | |
dbcb0f19 AB |
927 | static int migrate_to_node(struct mm_struct *mm, int source, int dest, |
928 | int flags) | |
7e2ab150 CL |
929 | { |
930 | nodemask_t nmask; | |
931 | LIST_HEAD(pagelist); | |
932 | int err = 0; | |
933 | ||
934 | nodes_clear(nmask); | |
935 | node_set(source, nmask); | |
6ce3c4c0 | 936 | |
08270807 MK |
937 | /* |
938 | * This does not "check" the range but isolates all pages that | |
939 | * need migration. Between passing in the full user address | |
940 | * space range and MPOL_MF_DISCONTIG_OK, this call can not fail. | |
941 | */ | |
942 | VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))); | |
98094945 | 943 | queue_pages_range(mm, mm->mmap->vm_start, mm->task_size, &nmask, |
7e2ab150 CL |
944 | flags | MPOL_MF_DISCONTIG_OK, &pagelist); |
945 | ||
cf608ac1 | 946 | if (!list_empty(&pagelist)) { |
68711a74 | 947 | err = migrate_pages(&pagelist, new_node_page, NULL, dest, |
9c620e2b | 948 | MIGRATE_SYNC, MR_SYSCALL); |
cf608ac1 | 949 | if (err) |
e2d8cf40 | 950 | putback_movable_pages(&pagelist); |
cf608ac1 | 951 | } |
95a402c3 | 952 | |
7e2ab150 | 953 | return err; |
6ce3c4c0 CL |
954 | } |
955 | ||
39743889 | 956 | /* |
7e2ab150 CL |
957 | * Move pages between the two nodesets so as to preserve the physical |
958 | * layout as much as possible. | |
39743889 CL |
959 | * |
960 | * Returns the number of page that could not be moved. | |
961 | */ | |
0ce72d4f AM |
962 | int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from, |
963 | const nodemask_t *to, int flags) | |
39743889 | 964 | { |
7e2ab150 | 965 | int busy = 0; |
0aedadf9 | 966 | int err; |
7e2ab150 | 967 | nodemask_t tmp; |
39743889 | 968 | |
0aedadf9 CL |
969 | err = migrate_prep(); |
970 | if (err) | |
971 | return err; | |
972 | ||
53f2556b | 973 | down_read(&mm->mmap_sem); |
39743889 | 974 | |
da0aa138 KM |
975 | /* |
976 | * Find a 'source' bit set in 'tmp' whose corresponding 'dest' | |
977 | * bit in 'to' is not also set in 'tmp'. Clear the found 'source' | |
978 | * bit in 'tmp', and return that <source, dest> pair for migration. | |
979 | * The pair of nodemasks 'to' and 'from' define the map. | |
980 | * | |
981 | * If no pair of bits is found that way, fallback to picking some | |
982 | * pair of 'source' and 'dest' bits that are not the same. If the | |
983 | * 'source' and 'dest' bits are the same, this represents a node | |
984 | * that will be migrating to itself, so no pages need move. | |
985 | * | |
986 | * If no bits are left in 'tmp', or if all remaining bits left | |
987 | * in 'tmp' correspond to the same bit in 'to', return false | |
988 | * (nothing left to migrate). | |
989 | * | |
990 | * This lets us pick a pair of nodes to migrate between, such that | |
991 | * if possible the dest node is not already occupied by some other | |
992 | * source node, minimizing the risk of overloading the memory on a | |
993 | * node that would happen if we migrated incoming memory to a node | |
994 | * before migrating outgoing memory source that same node. | |
995 | * | |
996 | * A single scan of tmp is sufficient. As we go, we remember the | |
997 | * most recent <s, d> pair that moved (s != d). If we find a pair | |
998 | * that not only moved, but what's better, moved to an empty slot | |
999 | * (d is not set in tmp), then we break out then, with that pair. | |
ae0e47f0 | 1000 | * Otherwise when we finish scanning from_tmp, we at least have the |
da0aa138 KM |
1001 | * most recent <s, d> pair that moved. If we get all the way through |
1002 | * the scan of tmp without finding any node that moved, much less | |
1003 | * moved to an empty node, then there is nothing left worth migrating. | |
1004 | */ | |
d4984711 | 1005 | |
0ce72d4f | 1006 | tmp = *from; |
7e2ab150 CL |
1007 | while (!nodes_empty(tmp)) { |
1008 | int s,d; | |
b76ac7e7 | 1009 | int source = NUMA_NO_NODE; |
7e2ab150 CL |
1010 | int dest = 0; |
1011 | ||
1012 | for_each_node_mask(s, tmp) { | |
4a5b18cc LW |
1013 | |
1014 | /* | |
1015 | * do_migrate_pages() tries to maintain the relative | |
1016 | * node relationship of the pages established between | |
1017 | * threads and memory areas. | |
1018 | * | |
1019 | * However if the number of source nodes is not equal to | |
1020 | * the number of destination nodes we can not preserve | |
1021 | * this node relative relationship. In that case, skip | |
1022 | * copying memory from a node that is in the destination | |
1023 | * mask. | |
1024 | * | |
1025 | * Example: [2,3,4] -> [3,4,5] moves everything. | |
1026 | * [0-7] - > [3,4,5] moves only 0,1,2,6,7. | |
1027 | */ | |
1028 | ||
0ce72d4f AM |
1029 | if ((nodes_weight(*from) != nodes_weight(*to)) && |
1030 | (node_isset(s, *to))) | |
4a5b18cc LW |
1031 | continue; |
1032 | ||
0ce72d4f | 1033 | d = node_remap(s, *from, *to); |
7e2ab150 CL |
1034 | if (s == d) |
1035 | continue; | |
1036 | ||
1037 | source = s; /* Node moved. Memorize */ | |
1038 | dest = d; | |
1039 | ||
1040 | /* dest not in remaining from nodes? */ | |
1041 | if (!node_isset(dest, tmp)) | |
1042 | break; | |
1043 | } | |
b76ac7e7 | 1044 | if (source == NUMA_NO_NODE) |
7e2ab150 CL |
1045 | break; |
1046 | ||
1047 | node_clear(source, tmp); | |
1048 | err = migrate_to_node(mm, source, dest, flags); | |
1049 | if (err > 0) | |
1050 | busy += err; | |
1051 | if (err < 0) | |
1052 | break; | |
39743889 CL |
1053 | } |
1054 | up_read(&mm->mmap_sem); | |
7e2ab150 CL |
1055 | if (err < 0) |
1056 | return err; | |
1057 | return busy; | |
b20a3503 CL |
1058 | |
1059 | } | |
1060 | ||
3ad33b24 LS |
1061 | /* |
1062 | * Allocate a new page for page migration based on vma policy. | |
d05f0cdc | 1063 | * Start by assuming the page is mapped by the same vma as contains @start. |
3ad33b24 LS |
1064 | * Search forward from there, if not. N.B., this assumes that the |
1065 | * list of pages handed to migrate_pages()--which is how we get here-- | |
1066 | * is in virtual address order. | |
1067 | */ | |
d05f0cdc | 1068 | static struct page *new_page(struct page *page, unsigned long start, int **x) |
95a402c3 | 1069 | { |
d05f0cdc | 1070 | struct vm_area_struct *vma; |
3ad33b24 | 1071 | unsigned long uninitialized_var(address); |
95a402c3 | 1072 | |
d05f0cdc | 1073 | vma = find_vma(current->mm, start); |
3ad33b24 LS |
1074 | while (vma) { |
1075 | address = page_address_in_vma(page, vma); | |
1076 | if (address != -EFAULT) | |
1077 | break; | |
1078 | vma = vma->vm_next; | |
1079 | } | |
11c731e8 WL |
1080 | |
1081 | if (PageHuge(page)) { | |
cc81717e MH |
1082 | BUG_ON(!vma); |
1083 | return alloc_huge_page_noerr(vma, address, 1); | |
11c731e8 | 1084 | } |
0bf598d8 | 1085 | /* |
11c731e8 | 1086 | * if !vma, alloc_page_vma() will use task or system default policy |
0bf598d8 | 1087 | */ |
0f556856 MH |
1088 | return alloc_page_vma(GFP_HIGHUSER_MOVABLE | __GFP_RETRY_MAYFAIL, |
1089 | vma, address); | |
95a402c3 | 1090 | } |
b20a3503 CL |
1091 | #else |
1092 | ||
1093 | static void migrate_page_add(struct page *page, struct list_head *pagelist, | |
1094 | unsigned long flags) | |
1095 | { | |
39743889 CL |
1096 | } |
1097 | ||
0ce72d4f AM |
1098 | int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from, |
1099 | const nodemask_t *to, int flags) | |
b20a3503 CL |
1100 | { |
1101 | return -ENOSYS; | |
1102 | } | |
95a402c3 | 1103 | |
d05f0cdc | 1104 | static struct page *new_page(struct page *page, unsigned long start, int **x) |
95a402c3 CL |
1105 | { |
1106 | return NULL; | |
1107 | } | |
b20a3503 CL |
1108 | #endif |
1109 | ||
dbcb0f19 | 1110 | static long do_mbind(unsigned long start, unsigned long len, |
028fec41 DR |
1111 | unsigned short mode, unsigned short mode_flags, |
1112 | nodemask_t *nmask, unsigned long flags) | |
6ce3c4c0 | 1113 | { |
6ce3c4c0 CL |
1114 | struct mm_struct *mm = current->mm; |
1115 | struct mempolicy *new; | |
1116 | unsigned long end; | |
1117 | int err; | |
1118 | LIST_HEAD(pagelist); | |
1119 | ||
b24f53a0 | 1120 | if (flags & ~(unsigned long)MPOL_MF_VALID) |
6ce3c4c0 | 1121 | return -EINVAL; |
74c00241 | 1122 | if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE)) |
6ce3c4c0 CL |
1123 | return -EPERM; |
1124 | ||
1125 | if (start & ~PAGE_MASK) | |
1126 | return -EINVAL; | |
1127 | ||
1128 | if (mode == MPOL_DEFAULT) | |
1129 | flags &= ~MPOL_MF_STRICT; | |
1130 | ||
1131 | len = (len + PAGE_SIZE - 1) & PAGE_MASK; | |
1132 | end = start + len; | |
1133 | ||
1134 | if (end < start) | |
1135 | return -EINVAL; | |
1136 | if (end == start) | |
1137 | return 0; | |
1138 | ||
028fec41 | 1139 | new = mpol_new(mode, mode_flags, nmask); |
6ce3c4c0 CL |
1140 | if (IS_ERR(new)) |
1141 | return PTR_ERR(new); | |
1142 | ||
b24f53a0 LS |
1143 | if (flags & MPOL_MF_LAZY) |
1144 | new->flags |= MPOL_F_MOF; | |
1145 | ||
6ce3c4c0 CL |
1146 | /* |
1147 | * If we are using the default policy then operation | |
1148 | * on discontinuous address spaces is okay after all | |
1149 | */ | |
1150 | if (!new) | |
1151 | flags |= MPOL_MF_DISCONTIG_OK; | |
1152 | ||
028fec41 DR |
1153 | pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n", |
1154 | start, start + len, mode, mode_flags, | |
00ef2d2f | 1155 | nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE); |
6ce3c4c0 | 1156 | |
0aedadf9 CL |
1157 | if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) { |
1158 | ||
1159 | err = migrate_prep(); | |
1160 | if (err) | |
b05ca738 | 1161 | goto mpol_out; |
0aedadf9 | 1162 | } |
4bfc4495 KH |
1163 | { |
1164 | NODEMASK_SCRATCH(scratch); | |
1165 | if (scratch) { | |
1166 | down_write(&mm->mmap_sem); | |
1167 | task_lock(current); | |
1168 | err = mpol_set_nodemask(new, nmask, scratch); | |
1169 | task_unlock(current); | |
1170 | if (err) | |
1171 | up_write(&mm->mmap_sem); | |
1172 | } else | |
1173 | err = -ENOMEM; | |
1174 | NODEMASK_SCRATCH_FREE(scratch); | |
1175 | } | |
b05ca738 KM |
1176 | if (err) |
1177 | goto mpol_out; | |
1178 | ||
d05f0cdc | 1179 | err = queue_pages_range(mm, start, end, nmask, |
6ce3c4c0 | 1180 | flags | MPOL_MF_INVERT, &pagelist); |
d05f0cdc | 1181 | if (!err) |
9d8cebd4 | 1182 | err = mbind_range(mm, start, end, new); |
7e2ab150 | 1183 | |
b24f53a0 LS |
1184 | if (!err) { |
1185 | int nr_failed = 0; | |
1186 | ||
cf608ac1 | 1187 | if (!list_empty(&pagelist)) { |
b24f53a0 | 1188 | WARN_ON_ONCE(flags & MPOL_MF_LAZY); |
d05f0cdc HD |
1189 | nr_failed = migrate_pages(&pagelist, new_page, NULL, |
1190 | start, MIGRATE_SYNC, MR_MEMPOLICY_MBIND); | |
cf608ac1 | 1191 | if (nr_failed) |
74060e4d | 1192 | putback_movable_pages(&pagelist); |
cf608ac1 | 1193 | } |
6ce3c4c0 | 1194 | |
b24f53a0 | 1195 | if (nr_failed && (flags & MPOL_MF_STRICT)) |
6ce3c4c0 | 1196 | err = -EIO; |
ab8a3e14 | 1197 | } else |
b0e5fd73 | 1198 | putback_movable_pages(&pagelist); |
b20a3503 | 1199 | |
6ce3c4c0 | 1200 | up_write(&mm->mmap_sem); |
b05ca738 | 1201 | mpol_out: |
f0be3d32 | 1202 | mpol_put(new); |
6ce3c4c0 CL |
1203 | return err; |
1204 | } | |
1205 | ||
8bccd85f CL |
1206 | /* |
1207 | * User space interface with variable sized bitmaps for nodelists. | |
1208 | */ | |
1209 | ||
1210 | /* Copy a node mask from user space. */ | |
39743889 | 1211 | static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask, |
8bccd85f CL |
1212 | unsigned long maxnode) |
1213 | { | |
1214 | unsigned long k; | |
1215 | unsigned long nlongs; | |
1216 | unsigned long endmask; | |
1217 | ||
1218 | --maxnode; | |
1219 | nodes_clear(*nodes); | |
1220 | if (maxnode == 0 || !nmask) | |
1221 | return 0; | |
a9c930ba | 1222 | if (maxnode > PAGE_SIZE*BITS_PER_BYTE) |
636f13c1 | 1223 | return -EINVAL; |
8bccd85f CL |
1224 | |
1225 | nlongs = BITS_TO_LONGS(maxnode); | |
1226 | if ((maxnode % BITS_PER_LONG) == 0) | |
1227 | endmask = ~0UL; | |
1228 | else | |
1229 | endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1; | |
1230 | ||
1231 | /* When the user specified more nodes than supported just check | |
1232 | if the non supported part is all zero. */ | |
1233 | if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) { | |
1234 | if (nlongs > PAGE_SIZE/sizeof(long)) | |
1235 | return -EINVAL; | |
1236 | for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) { | |
1237 | unsigned long t; | |
1238 | if (get_user(t, nmask + k)) | |
1239 | return -EFAULT; | |
1240 | if (k == nlongs - 1) { | |
1241 | if (t & endmask) | |
1242 | return -EINVAL; | |
1243 | } else if (t) | |
1244 | return -EINVAL; | |
1245 | } | |
1246 | nlongs = BITS_TO_LONGS(MAX_NUMNODES); | |
1247 | endmask = ~0UL; | |
1248 | } | |
1249 | ||
1250 | if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long))) | |
1251 | return -EFAULT; | |
1252 | nodes_addr(*nodes)[nlongs-1] &= endmask; | |
1253 | return 0; | |
1254 | } | |
1255 | ||
1256 | /* Copy a kernel node mask to user space */ | |
1257 | static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode, | |
1258 | nodemask_t *nodes) | |
1259 | { | |
1260 | unsigned long copy = ALIGN(maxnode-1, 64) / 8; | |
1261 | const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long); | |
1262 | ||
1263 | if (copy > nbytes) { | |
1264 | if (copy > PAGE_SIZE) | |
1265 | return -EINVAL; | |
1266 | if (clear_user((char __user *)mask + nbytes, copy - nbytes)) | |
1267 | return -EFAULT; | |
1268 | copy = nbytes; | |
1269 | } | |
1270 | return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0; | |
1271 | } | |
1272 | ||
938bb9f5 | 1273 | SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len, |
f7f28ca9 | 1274 | unsigned long, mode, const unsigned long __user *, nmask, |
938bb9f5 | 1275 | unsigned long, maxnode, unsigned, flags) |
8bccd85f CL |
1276 | { |
1277 | nodemask_t nodes; | |
1278 | int err; | |
028fec41 | 1279 | unsigned short mode_flags; |
8bccd85f | 1280 | |
028fec41 DR |
1281 | mode_flags = mode & MPOL_MODE_FLAGS; |
1282 | mode &= ~MPOL_MODE_FLAGS; | |
a3b51e01 DR |
1283 | if (mode >= MPOL_MAX) |
1284 | return -EINVAL; | |
4c50bc01 DR |
1285 | if ((mode_flags & MPOL_F_STATIC_NODES) && |
1286 | (mode_flags & MPOL_F_RELATIVE_NODES)) | |
1287 | return -EINVAL; | |
8bccd85f CL |
1288 | err = get_nodes(&nodes, nmask, maxnode); |
1289 | if (err) | |
1290 | return err; | |
028fec41 | 1291 | return do_mbind(start, len, mode, mode_flags, &nodes, flags); |
8bccd85f CL |
1292 | } |
1293 | ||
1294 | /* Set the process memory policy */ | |
23c8902d | 1295 | SYSCALL_DEFINE3(set_mempolicy, int, mode, const unsigned long __user *, nmask, |
938bb9f5 | 1296 | unsigned long, maxnode) |
8bccd85f CL |
1297 | { |
1298 | int err; | |
1299 | nodemask_t nodes; | |
028fec41 | 1300 | unsigned short flags; |
8bccd85f | 1301 | |
028fec41 DR |
1302 | flags = mode & MPOL_MODE_FLAGS; |
1303 | mode &= ~MPOL_MODE_FLAGS; | |
1304 | if ((unsigned int)mode >= MPOL_MAX) | |
8bccd85f | 1305 | return -EINVAL; |
4c50bc01 DR |
1306 | if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES)) |
1307 | return -EINVAL; | |
8bccd85f CL |
1308 | err = get_nodes(&nodes, nmask, maxnode); |
1309 | if (err) | |
1310 | return err; | |
028fec41 | 1311 | return do_set_mempolicy(mode, flags, &nodes); |
8bccd85f CL |
1312 | } |
1313 | ||
938bb9f5 HC |
1314 | SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode, |
1315 | const unsigned long __user *, old_nodes, | |
1316 | const unsigned long __user *, new_nodes) | |
39743889 | 1317 | { |
c69e8d9c | 1318 | const struct cred *cred = current_cred(), *tcred; |
596d7cfa | 1319 | struct mm_struct *mm = NULL; |
39743889 | 1320 | struct task_struct *task; |
39743889 CL |
1321 | nodemask_t task_nodes; |
1322 | int err; | |
596d7cfa KM |
1323 | nodemask_t *old; |
1324 | nodemask_t *new; | |
1325 | NODEMASK_SCRATCH(scratch); | |
1326 | ||
1327 | if (!scratch) | |
1328 | return -ENOMEM; | |
39743889 | 1329 | |
596d7cfa KM |
1330 | old = &scratch->mask1; |
1331 | new = &scratch->mask2; | |
1332 | ||
1333 | err = get_nodes(old, old_nodes, maxnode); | |
39743889 | 1334 | if (err) |
596d7cfa | 1335 | goto out; |
39743889 | 1336 | |
596d7cfa | 1337 | err = get_nodes(new, new_nodes, maxnode); |
39743889 | 1338 | if (err) |
596d7cfa | 1339 | goto out; |
39743889 CL |
1340 | |
1341 | /* Find the mm_struct */ | |
55cfaa3c | 1342 | rcu_read_lock(); |
228ebcbe | 1343 | task = pid ? find_task_by_vpid(pid) : current; |
39743889 | 1344 | if (!task) { |
55cfaa3c | 1345 | rcu_read_unlock(); |
596d7cfa KM |
1346 | err = -ESRCH; |
1347 | goto out; | |
39743889 | 1348 | } |
3268c63e | 1349 | get_task_struct(task); |
39743889 | 1350 | |
596d7cfa | 1351 | err = -EINVAL; |
39743889 CL |
1352 | |
1353 | /* | |
1354 | * Check if this process has the right to modify the specified | |
1355 | * process. The right exists if the process has administrative | |
7f927fcc | 1356 | * capabilities, superuser privileges or the same |
39743889 CL |
1357 | * userid as the target process. |
1358 | */ | |
c69e8d9c | 1359 | tcred = __task_cred(task); |
b38a86eb EB |
1360 | if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) && |
1361 | !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) && | |
74c00241 | 1362 | !capable(CAP_SYS_NICE)) { |
c69e8d9c | 1363 | rcu_read_unlock(); |
39743889 | 1364 | err = -EPERM; |
3268c63e | 1365 | goto out_put; |
39743889 | 1366 | } |
c69e8d9c | 1367 | rcu_read_unlock(); |
39743889 CL |
1368 | |
1369 | task_nodes = cpuset_mems_allowed(task); | |
1370 | /* Is the user allowed to access the target nodes? */ | |
596d7cfa | 1371 | if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) { |
39743889 | 1372 | err = -EPERM; |
3268c63e | 1373 | goto out_put; |
39743889 CL |
1374 | } |
1375 | ||
01f13bd6 | 1376 | if (!nodes_subset(*new, node_states[N_MEMORY])) { |
3b42d28b | 1377 | err = -EINVAL; |
3268c63e | 1378 | goto out_put; |
3b42d28b CL |
1379 | } |
1380 | ||
86c3a764 DQ |
1381 | err = security_task_movememory(task); |
1382 | if (err) | |
3268c63e | 1383 | goto out_put; |
86c3a764 | 1384 | |
3268c63e CL |
1385 | mm = get_task_mm(task); |
1386 | put_task_struct(task); | |
f2a9ef88 SL |
1387 | |
1388 | if (!mm) { | |
3268c63e | 1389 | err = -EINVAL; |
f2a9ef88 SL |
1390 | goto out; |
1391 | } | |
1392 | ||
1393 | err = do_migrate_pages(mm, old, new, | |
1394 | capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE); | |
3268c63e CL |
1395 | |
1396 | mmput(mm); | |
1397 | out: | |
596d7cfa KM |
1398 | NODEMASK_SCRATCH_FREE(scratch); |
1399 | ||
39743889 | 1400 | return err; |
3268c63e CL |
1401 | |
1402 | out_put: | |
1403 | put_task_struct(task); | |
1404 | goto out; | |
1405 | ||
39743889 CL |
1406 | } |
1407 | ||
1408 | ||
8bccd85f | 1409 | /* Retrieve NUMA policy */ |
938bb9f5 HC |
1410 | SYSCALL_DEFINE5(get_mempolicy, int __user *, policy, |
1411 | unsigned long __user *, nmask, unsigned long, maxnode, | |
1412 | unsigned long, addr, unsigned long, flags) | |
8bccd85f | 1413 | { |
dbcb0f19 AB |
1414 | int err; |
1415 | int uninitialized_var(pval); | |
8bccd85f CL |
1416 | nodemask_t nodes; |
1417 | ||
1418 | if (nmask != NULL && maxnode < MAX_NUMNODES) | |
1419 | return -EINVAL; | |
1420 | ||
1421 | err = do_get_mempolicy(&pval, &nodes, addr, flags); | |
1422 | ||
1423 | if (err) | |
1424 | return err; | |
1425 | ||
1426 | if (policy && put_user(pval, policy)) | |
1427 | return -EFAULT; | |
1428 | ||
1429 | if (nmask) | |
1430 | err = copy_nodes_to_user(nmask, maxnode, &nodes); | |
1431 | ||
1432 | return err; | |
1433 | } | |
1434 | ||
1da177e4 LT |
1435 | #ifdef CONFIG_COMPAT |
1436 | ||
c93e0f6c HC |
1437 | COMPAT_SYSCALL_DEFINE5(get_mempolicy, int __user *, policy, |
1438 | compat_ulong_t __user *, nmask, | |
1439 | compat_ulong_t, maxnode, | |
1440 | compat_ulong_t, addr, compat_ulong_t, flags) | |
1da177e4 LT |
1441 | { |
1442 | long err; | |
1443 | unsigned long __user *nm = NULL; | |
1444 | unsigned long nr_bits, alloc_size; | |
1445 | DECLARE_BITMAP(bm, MAX_NUMNODES); | |
1446 | ||
1447 | nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES); | |
1448 | alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8; | |
1449 | ||
1450 | if (nmask) | |
1451 | nm = compat_alloc_user_space(alloc_size); | |
1452 | ||
1453 | err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags); | |
1454 | ||
1455 | if (!err && nmask) { | |
2bbff6c7 KH |
1456 | unsigned long copy_size; |
1457 | copy_size = min_t(unsigned long, sizeof(bm), alloc_size); | |
1458 | err = copy_from_user(bm, nm, copy_size); | |
1da177e4 LT |
1459 | /* ensure entire bitmap is zeroed */ |
1460 | err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8); | |
1461 | err |= compat_put_bitmap(nmask, bm, nr_bits); | |
1462 | } | |
1463 | ||
1464 | return err; | |
1465 | } | |
1466 | ||
c93e0f6c HC |
1467 | COMPAT_SYSCALL_DEFINE3(set_mempolicy, int, mode, compat_ulong_t __user *, nmask, |
1468 | compat_ulong_t, maxnode) | |
1da177e4 | 1469 | { |
1da177e4 LT |
1470 | unsigned long __user *nm = NULL; |
1471 | unsigned long nr_bits, alloc_size; | |
1472 | DECLARE_BITMAP(bm, MAX_NUMNODES); | |
1473 | ||
1474 | nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES); | |
1475 | alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8; | |
1476 | ||
1477 | if (nmask) { | |
cf01fb99 CS |
1478 | if (compat_get_bitmap(bm, nmask, nr_bits)) |
1479 | return -EFAULT; | |
1da177e4 | 1480 | nm = compat_alloc_user_space(alloc_size); |
cf01fb99 CS |
1481 | if (copy_to_user(nm, bm, alloc_size)) |
1482 | return -EFAULT; | |
1da177e4 LT |
1483 | } |
1484 | ||
1da177e4 LT |
1485 | return sys_set_mempolicy(mode, nm, nr_bits+1); |
1486 | } | |
1487 | ||
c93e0f6c HC |
1488 | COMPAT_SYSCALL_DEFINE6(mbind, compat_ulong_t, start, compat_ulong_t, len, |
1489 | compat_ulong_t, mode, compat_ulong_t __user *, nmask, | |
1490 | compat_ulong_t, maxnode, compat_ulong_t, flags) | |
1da177e4 | 1491 | { |
1da177e4 LT |
1492 | unsigned long __user *nm = NULL; |
1493 | unsigned long nr_bits, alloc_size; | |
dfcd3c0d | 1494 | nodemask_t bm; |
1da177e4 LT |
1495 | |
1496 | nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES); | |
1497 | alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8; | |
1498 | ||
1499 | if (nmask) { | |
cf01fb99 CS |
1500 | if (compat_get_bitmap(nodes_addr(bm), nmask, nr_bits)) |
1501 | return -EFAULT; | |
1da177e4 | 1502 | nm = compat_alloc_user_space(alloc_size); |
cf01fb99 CS |
1503 | if (copy_to_user(nm, nodes_addr(bm), alloc_size)) |
1504 | return -EFAULT; | |
1da177e4 LT |
1505 | } |
1506 | ||
1da177e4 LT |
1507 | return sys_mbind(start, len, mode, nm, nr_bits+1, flags); |
1508 | } | |
1509 | ||
1510 | #endif | |
1511 | ||
74d2c3a0 ON |
1512 | struct mempolicy *__get_vma_policy(struct vm_area_struct *vma, |
1513 | unsigned long addr) | |
1da177e4 | 1514 | { |
8d90274b | 1515 | struct mempolicy *pol = NULL; |
1da177e4 LT |
1516 | |
1517 | if (vma) { | |
480eccf9 | 1518 | if (vma->vm_ops && vma->vm_ops->get_policy) { |
8d90274b | 1519 | pol = vma->vm_ops->get_policy(vma, addr); |
00442ad0 | 1520 | } else if (vma->vm_policy) { |
1da177e4 | 1521 | pol = vma->vm_policy; |
00442ad0 MG |
1522 | |
1523 | /* | |
1524 | * shmem_alloc_page() passes MPOL_F_SHARED policy with | |
1525 | * a pseudo vma whose vma->vm_ops=NULL. Take a reference | |
1526 | * count on these policies which will be dropped by | |
1527 | * mpol_cond_put() later | |
1528 | */ | |
1529 | if (mpol_needs_cond_ref(pol)) | |
1530 | mpol_get(pol); | |
1531 | } | |
1da177e4 | 1532 | } |
f15ca78e | 1533 | |
74d2c3a0 ON |
1534 | return pol; |
1535 | } | |
1536 | ||
1537 | /* | |
dd6eecb9 | 1538 | * get_vma_policy(@vma, @addr) |
74d2c3a0 ON |
1539 | * @vma: virtual memory area whose policy is sought |
1540 | * @addr: address in @vma for shared policy lookup | |
1541 | * | |
1542 | * Returns effective policy for a VMA at specified address. | |
dd6eecb9 | 1543 | * Falls back to current->mempolicy or system default policy, as necessary. |
74d2c3a0 ON |
1544 | * Shared policies [those marked as MPOL_F_SHARED] require an extra reference |
1545 | * count--added by the get_policy() vm_op, as appropriate--to protect against | |
1546 | * freeing by another task. It is the caller's responsibility to free the | |
1547 | * extra reference for shared policies. | |
1548 | */ | |
dd6eecb9 ON |
1549 | static struct mempolicy *get_vma_policy(struct vm_area_struct *vma, |
1550 | unsigned long addr) | |
74d2c3a0 ON |
1551 | { |
1552 | struct mempolicy *pol = __get_vma_policy(vma, addr); | |
1553 | ||
8d90274b | 1554 | if (!pol) |
dd6eecb9 | 1555 | pol = get_task_policy(current); |
8d90274b | 1556 | |
1da177e4 LT |
1557 | return pol; |
1558 | } | |
1559 | ||
6b6482bb | 1560 | bool vma_policy_mof(struct vm_area_struct *vma) |
fc314724 | 1561 | { |
6b6482bb | 1562 | struct mempolicy *pol; |
fc314724 | 1563 | |
6b6482bb ON |
1564 | if (vma->vm_ops && vma->vm_ops->get_policy) { |
1565 | bool ret = false; | |
fc314724 | 1566 | |
6b6482bb ON |
1567 | pol = vma->vm_ops->get_policy(vma, vma->vm_start); |
1568 | if (pol && (pol->flags & MPOL_F_MOF)) | |
1569 | ret = true; | |
1570 | mpol_cond_put(pol); | |
8d90274b | 1571 | |
6b6482bb | 1572 | return ret; |
fc314724 MG |
1573 | } |
1574 | ||
6b6482bb | 1575 | pol = vma->vm_policy; |
8d90274b | 1576 | if (!pol) |
6b6482bb | 1577 | pol = get_task_policy(current); |
8d90274b | 1578 | |
fc314724 MG |
1579 | return pol->flags & MPOL_F_MOF; |
1580 | } | |
1581 | ||
d3eb1570 LJ |
1582 | static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone) |
1583 | { | |
1584 | enum zone_type dynamic_policy_zone = policy_zone; | |
1585 | ||
1586 | BUG_ON(dynamic_policy_zone == ZONE_MOVABLE); | |
1587 | ||
1588 | /* | |
1589 | * if policy->v.nodes has movable memory only, | |
1590 | * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only. | |
1591 | * | |
1592 | * policy->v.nodes is intersect with node_states[N_MEMORY]. | |
1593 | * so if the following test faile, it implies | |
1594 | * policy->v.nodes has movable memory only. | |
1595 | */ | |
1596 | if (!nodes_intersects(policy->v.nodes, node_states[N_HIGH_MEMORY])) | |
1597 | dynamic_policy_zone = ZONE_MOVABLE; | |
1598 | ||
1599 | return zone >= dynamic_policy_zone; | |
1600 | } | |
1601 | ||
52cd3b07 LS |
1602 | /* |
1603 | * Return a nodemask representing a mempolicy for filtering nodes for | |
1604 | * page allocation | |
1605 | */ | |
1606 | static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy) | |
19770b32 MG |
1607 | { |
1608 | /* Lower zones don't get a nodemask applied for MPOL_BIND */ | |
45c4745a | 1609 | if (unlikely(policy->mode == MPOL_BIND) && |
d3eb1570 | 1610 | apply_policy_zone(policy, gfp_zone(gfp)) && |
19770b32 MG |
1611 | cpuset_nodemask_valid_mems_allowed(&policy->v.nodes)) |
1612 | return &policy->v.nodes; | |
1613 | ||
1614 | return NULL; | |
1615 | } | |
1616 | ||
04ec6264 VB |
1617 | /* Return the node id preferred by the given mempolicy, or the given id */ |
1618 | static int policy_node(gfp_t gfp, struct mempolicy *policy, | |
1619 | int nd) | |
1da177e4 | 1620 | { |
6d840958 MH |
1621 | if (policy->mode == MPOL_PREFERRED && !(policy->flags & MPOL_F_LOCAL)) |
1622 | nd = policy->v.preferred_node; | |
1623 | else { | |
19770b32 | 1624 | /* |
6d840958 MH |
1625 | * __GFP_THISNODE shouldn't even be used with the bind policy |
1626 | * because we might easily break the expectation to stay on the | |
1627 | * requested node and not break the policy. | |
19770b32 | 1628 | */ |
6d840958 | 1629 | WARN_ON_ONCE(policy->mode == MPOL_BIND && (gfp & __GFP_THISNODE)); |
1da177e4 | 1630 | } |
6d840958 | 1631 | |
04ec6264 | 1632 | return nd; |
1da177e4 LT |
1633 | } |
1634 | ||
1635 | /* Do dynamic interleaving for a process */ | |
1636 | static unsigned interleave_nodes(struct mempolicy *policy) | |
1637 | { | |
45816682 | 1638 | unsigned next; |
1da177e4 LT |
1639 | struct task_struct *me = current; |
1640 | ||
45816682 | 1641 | next = next_node_in(me->il_prev, policy->v.nodes); |
f5b087b5 | 1642 | if (next < MAX_NUMNODES) |
45816682 VB |
1643 | me->il_prev = next; |
1644 | return next; | |
1da177e4 LT |
1645 | } |
1646 | ||
dc85da15 CL |
1647 | /* |
1648 | * Depending on the memory policy provide a node from which to allocate the | |
1649 | * next slab entry. | |
1650 | */ | |
2a389610 | 1651 | unsigned int mempolicy_slab_node(void) |
dc85da15 | 1652 | { |
e7b691b0 | 1653 | struct mempolicy *policy; |
2a389610 | 1654 | int node = numa_mem_id(); |
e7b691b0 AK |
1655 | |
1656 | if (in_interrupt()) | |
2a389610 | 1657 | return node; |
e7b691b0 AK |
1658 | |
1659 | policy = current->mempolicy; | |
fc36b8d3 | 1660 | if (!policy || policy->flags & MPOL_F_LOCAL) |
2a389610 | 1661 | return node; |
bea904d5 LS |
1662 | |
1663 | switch (policy->mode) { | |
1664 | case MPOL_PREFERRED: | |
fc36b8d3 LS |
1665 | /* |
1666 | * handled MPOL_F_LOCAL above | |
1667 | */ | |
1668 | return policy->v.preferred_node; | |
765c4507 | 1669 | |
dc85da15 CL |
1670 | case MPOL_INTERLEAVE: |
1671 | return interleave_nodes(policy); | |
1672 | ||
dd1a239f | 1673 | case MPOL_BIND: { |
c33d6c06 MG |
1674 | struct zoneref *z; |
1675 | ||
dc85da15 CL |
1676 | /* |
1677 | * Follow bind policy behavior and start allocation at the | |
1678 | * first node. | |
1679 | */ | |
19770b32 | 1680 | struct zonelist *zonelist; |
19770b32 | 1681 | enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL); |
c9634cf0 | 1682 | zonelist = &NODE_DATA(node)->node_zonelists[ZONELIST_FALLBACK]; |
c33d6c06 MG |
1683 | z = first_zones_zonelist(zonelist, highest_zoneidx, |
1684 | &policy->v.nodes); | |
1685 | return z->zone ? z->zone->node : node; | |
dd1a239f | 1686 | } |
dc85da15 | 1687 | |
dc85da15 | 1688 | default: |
bea904d5 | 1689 | BUG(); |
dc85da15 CL |
1690 | } |
1691 | } | |
1692 | ||
fee83b3a AM |
1693 | /* |
1694 | * Do static interleaving for a VMA with known offset @n. Returns the n'th | |
1695 | * node in pol->v.nodes (starting from n=0), wrapping around if n exceeds the | |
1696 | * number of present nodes. | |
1697 | */ | |
1da177e4 | 1698 | static unsigned offset_il_node(struct mempolicy *pol, |
fee83b3a | 1699 | struct vm_area_struct *vma, unsigned long n) |
1da177e4 | 1700 | { |
dfcd3c0d | 1701 | unsigned nnodes = nodes_weight(pol->v.nodes); |
f5b087b5 | 1702 | unsigned target; |
fee83b3a AM |
1703 | int i; |
1704 | int nid; | |
1da177e4 | 1705 | |
f5b087b5 DR |
1706 | if (!nnodes) |
1707 | return numa_node_id(); | |
fee83b3a AM |
1708 | target = (unsigned int)n % nnodes; |
1709 | nid = first_node(pol->v.nodes); | |
1710 | for (i = 0; i < target; i++) | |
dfcd3c0d | 1711 | nid = next_node(nid, pol->v.nodes); |
1da177e4 LT |
1712 | return nid; |
1713 | } | |
1714 | ||
5da7ca86 CL |
1715 | /* Determine a node number for interleave */ |
1716 | static inline unsigned interleave_nid(struct mempolicy *pol, | |
1717 | struct vm_area_struct *vma, unsigned long addr, int shift) | |
1718 | { | |
1719 | if (vma) { | |
1720 | unsigned long off; | |
1721 | ||
3b98b087 NA |
1722 | /* |
1723 | * for small pages, there is no difference between | |
1724 | * shift and PAGE_SHIFT, so the bit-shift is safe. | |
1725 | * for huge pages, since vm_pgoff is in units of small | |
1726 | * pages, we need to shift off the always 0 bits to get | |
1727 | * a useful offset. | |
1728 | */ | |
1729 | BUG_ON(shift < PAGE_SHIFT); | |
1730 | off = vma->vm_pgoff >> (shift - PAGE_SHIFT); | |
5da7ca86 CL |
1731 | off += (addr - vma->vm_start) >> shift; |
1732 | return offset_il_node(pol, vma, off); | |
1733 | } else | |
1734 | return interleave_nodes(pol); | |
1735 | } | |
1736 | ||
00ac59ad | 1737 | #ifdef CONFIG_HUGETLBFS |
480eccf9 | 1738 | /* |
04ec6264 | 1739 | * huge_node(@vma, @addr, @gfp_flags, @mpol) |
b46e14ac FF |
1740 | * @vma: virtual memory area whose policy is sought |
1741 | * @addr: address in @vma for shared policy lookup and interleave policy | |
1742 | * @gfp_flags: for requested zone | |
1743 | * @mpol: pointer to mempolicy pointer for reference counted mempolicy | |
1744 | * @nodemask: pointer to nodemask pointer for MPOL_BIND nodemask | |
480eccf9 | 1745 | * |
04ec6264 | 1746 | * Returns a nid suitable for a huge page allocation and a pointer |
52cd3b07 LS |
1747 | * to the struct mempolicy for conditional unref after allocation. |
1748 | * If the effective policy is 'BIND, returns a pointer to the mempolicy's | |
1749 | * @nodemask for filtering the zonelist. | |
c0ff7453 | 1750 | * |
d26914d1 | 1751 | * Must be protected by read_mems_allowed_begin() |
480eccf9 | 1752 | */ |
04ec6264 VB |
1753 | int huge_node(struct vm_area_struct *vma, unsigned long addr, gfp_t gfp_flags, |
1754 | struct mempolicy **mpol, nodemask_t **nodemask) | |
5da7ca86 | 1755 | { |
04ec6264 | 1756 | int nid; |
5da7ca86 | 1757 | |
dd6eecb9 | 1758 | *mpol = get_vma_policy(vma, addr); |
19770b32 | 1759 | *nodemask = NULL; /* assume !MPOL_BIND */ |
5da7ca86 | 1760 | |
52cd3b07 | 1761 | if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) { |
04ec6264 VB |
1762 | nid = interleave_nid(*mpol, vma, addr, |
1763 | huge_page_shift(hstate_vma(vma))); | |
52cd3b07 | 1764 | } else { |
04ec6264 | 1765 | nid = policy_node(gfp_flags, *mpol, numa_node_id()); |
52cd3b07 LS |
1766 | if ((*mpol)->mode == MPOL_BIND) |
1767 | *nodemask = &(*mpol)->v.nodes; | |
480eccf9 | 1768 | } |
04ec6264 | 1769 | return nid; |
5da7ca86 | 1770 | } |
06808b08 LS |
1771 | |
1772 | /* | |
1773 | * init_nodemask_of_mempolicy | |
1774 | * | |
1775 | * If the current task's mempolicy is "default" [NULL], return 'false' | |
1776 | * to indicate default policy. Otherwise, extract the policy nodemask | |
1777 | * for 'bind' or 'interleave' policy into the argument nodemask, or | |
1778 | * initialize the argument nodemask to contain the single node for | |
1779 | * 'preferred' or 'local' policy and return 'true' to indicate presence | |
1780 | * of non-default mempolicy. | |
1781 | * | |
1782 | * We don't bother with reference counting the mempolicy [mpol_get/put] | |
1783 | * because the current task is examining it's own mempolicy and a task's | |
1784 | * mempolicy is only ever changed by the task itself. | |
1785 | * | |
1786 | * N.B., it is the caller's responsibility to free a returned nodemask. | |
1787 | */ | |
1788 | bool init_nodemask_of_mempolicy(nodemask_t *mask) | |
1789 | { | |
1790 | struct mempolicy *mempolicy; | |
1791 | int nid; | |
1792 | ||
1793 | if (!(mask && current->mempolicy)) | |
1794 | return false; | |
1795 | ||
c0ff7453 | 1796 | task_lock(current); |
06808b08 LS |
1797 | mempolicy = current->mempolicy; |
1798 | switch (mempolicy->mode) { | |
1799 | case MPOL_PREFERRED: | |
1800 | if (mempolicy->flags & MPOL_F_LOCAL) | |
1801 | nid = numa_node_id(); | |
1802 | else | |
1803 | nid = mempolicy->v.preferred_node; | |
1804 | init_nodemask_of_node(mask, nid); | |
1805 | break; | |
1806 | ||
1807 | case MPOL_BIND: | |
1808 | /* Fall through */ | |
1809 | case MPOL_INTERLEAVE: | |
1810 | *mask = mempolicy->v.nodes; | |
1811 | break; | |
1812 | ||
1813 | default: | |
1814 | BUG(); | |
1815 | } | |
c0ff7453 | 1816 | task_unlock(current); |
06808b08 LS |
1817 | |
1818 | return true; | |
1819 | } | |
00ac59ad | 1820 | #endif |
5da7ca86 | 1821 | |
6f48d0eb DR |
1822 | /* |
1823 | * mempolicy_nodemask_intersects | |
1824 | * | |
1825 | * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default | |
1826 | * policy. Otherwise, check for intersection between mask and the policy | |
1827 | * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local' | |
1828 | * policy, always return true since it may allocate elsewhere on fallback. | |
1829 | * | |
1830 | * Takes task_lock(tsk) to prevent freeing of its mempolicy. | |
1831 | */ | |
1832 | bool mempolicy_nodemask_intersects(struct task_struct *tsk, | |
1833 | const nodemask_t *mask) | |
1834 | { | |
1835 | struct mempolicy *mempolicy; | |
1836 | bool ret = true; | |
1837 | ||
1838 | if (!mask) | |
1839 | return ret; | |
1840 | task_lock(tsk); | |
1841 | mempolicy = tsk->mempolicy; | |
1842 | if (!mempolicy) | |
1843 | goto out; | |
1844 | ||
1845 | switch (mempolicy->mode) { | |
1846 | case MPOL_PREFERRED: | |
1847 | /* | |
1848 | * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to | |
1849 | * allocate from, they may fallback to other nodes when oom. | |
1850 | * Thus, it's possible for tsk to have allocated memory from | |
1851 | * nodes in mask. | |
1852 | */ | |
1853 | break; | |
1854 | case MPOL_BIND: | |
1855 | case MPOL_INTERLEAVE: | |
1856 | ret = nodes_intersects(mempolicy->v.nodes, *mask); | |
1857 | break; | |
1858 | default: | |
1859 | BUG(); | |
1860 | } | |
1861 | out: | |
1862 | task_unlock(tsk); | |
1863 | return ret; | |
1864 | } | |
1865 | ||
1da177e4 LT |
1866 | /* Allocate a page in interleaved policy. |
1867 | Own path because it needs to do special accounting. */ | |
662f3a0b AK |
1868 | static struct page *alloc_page_interleave(gfp_t gfp, unsigned order, |
1869 | unsigned nid) | |
1da177e4 | 1870 | { |
1da177e4 LT |
1871 | struct page *page; |
1872 | ||
04ec6264 VB |
1873 | page = __alloc_pages(gfp, order, nid); |
1874 | if (page && page_to_nid(page) == nid) | |
ca889e6c | 1875 | inc_zone_page_state(page, NUMA_INTERLEAVE_HIT); |
1da177e4 LT |
1876 | return page; |
1877 | } | |
1878 | ||
1879 | /** | |
0bbbc0b3 | 1880 | * alloc_pages_vma - Allocate a page for a VMA. |
1da177e4 LT |
1881 | * |
1882 | * @gfp: | |
1883 | * %GFP_USER user allocation. | |
1884 | * %GFP_KERNEL kernel allocations, | |
1885 | * %GFP_HIGHMEM highmem/user allocations, | |
1886 | * %GFP_FS allocation should not call back into a file system. | |
1887 | * %GFP_ATOMIC don't sleep. | |
1888 | * | |
0bbbc0b3 | 1889 | * @order:Order of the GFP allocation. |
1da177e4 LT |
1890 | * @vma: Pointer to VMA or NULL if not available. |
1891 | * @addr: Virtual Address of the allocation. Must be inside the VMA. | |
be97a41b VB |
1892 | * @node: Which node to prefer for allocation (modulo policy). |
1893 | * @hugepage: for hugepages try only the preferred node if possible | |
1da177e4 LT |
1894 | * |
1895 | * This function allocates a page from the kernel page pool and applies | |
1896 | * a NUMA policy associated with the VMA or the current process. | |
1897 | * When VMA is not NULL caller must hold down_read on the mmap_sem of the | |
1898 | * mm_struct of the VMA to prevent it from going away. Should be used for | |
be97a41b VB |
1899 | * all allocations for pages that will be mapped into user space. Returns |
1900 | * NULL when no page can be allocated. | |
1da177e4 LT |
1901 | */ |
1902 | struct page * | |
0bbbc0b3 | 1903 | alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma, |
be97a41b | 1904 | unsigned long addr, int node, bool hugepage) |
1da177e4 | 1905 | { |
cc9a6c87 | 1906 | struct mempolicy *pol; |
c0ff7453 | 1907 | struct page *page; |
04ec6264 | 1908 | int preferred_nid; |
be97a41b | 1909 | nodemask_t *nmask; |
cc9a6c87 | 1910 | |
dd6eecb9 | 1911 | pol = get_vma_policy(vma, addr); |
1da177e4 | 1912 | |
0867a57c VB |
1913 | if (pol->mode == MPOL_INTERLEAVE) { |
1914 | unsigned nid; | |
1915 | ||
1916 | nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order); | |
1917 | mpol_cond_put(pol); | |
1918 | page = alloc_page_interleave(gfp, order, nid); | |
1919 | goto out; | |
1920 | } | |
1921 | ||
1922 | if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && hugepage)) { | |
1923 | int hpage_node = node; | |
1924 | ||
be97a41b VB |
1925 | /* |
1926 | * For hugepage allocation and non-interleave policy which | |
0867a57c VB |
1927 | * allows the current node (or other explicitly preferred |
1928 | * node) we only try to allocate from the current/preferred | |
1929 | * node and don't fall back to other nodes, as the cost of | |
1930 | * remote accesses would likely offset THP benefits. | |
be97a41b VB |
1931 | * |
1932 | * If the policy is interleave, or does not allow the current | |
1933 | * node in its nodemask, we allocate the standard way. | |
1934 | */ | |
0867a57c VB |
1935 | if (pol->mode == MPOL_PREFERRED && |
1936 | !(pol->flags & MPOL_F_LOCAL)) | |
1937 | hpage_node = pol->v.preferred_node; | |
1938 | ||
be97a41b | 1939 | nmask = policy_nodemask(gfp, pol); |
0867a57c | 1940 | if (!nmask || node_isset(hpage_node, *nmask)) { |
be97a41b | 1941 | mpol_cond_put(pol); |
96db800f | 1942 | page = __alloc_pages_node(hpage_node, |
5265047a | 1943 | gfp | __GFP_THISNODE, order); |
be97a41b VB |
1944 | goto out; |
1945 | } | |
1946 | } | |
1947 | ||
be97a41b | 1948 | nmask = policy_nodemask(gfp, pol); |
04ec6264 VB |
1949 | preferred_nid = policy_node(gfp, pol, node); |
1950 | page = __alloc_pages_nodemask(gfp, order, preferred_nid, nmask); | |
d51e9894 | 1951 | mpol_cond_put(pol); |
be97a41b | 1952 | out: |
c0ff7453 | 1953 | return page; |
1da177e4 LT |
1954 | } |
1955 | ||
1956 | /** | |
1957 | * alloc_pages_current - Allocate pages. | |
1958 | * | |
1959 | * @gfp: | |
1960 | * %GFP_USER user allocation, | |
1961 | * %GFP_KERNEL kernel allocation, | |
1962 | * %GFP_HIGHMEM highmem allocation, | |
1963 | * %GFP_FS don't call back into a file system. | |
1964 | * %GFP_ATOMIC don't sleep. | |
1965 | * @order: Power of two of allocation size in pages. 0 is a single page. | |
1966 | * | |
1967 | * Allocate a page from the kernel page pool. When not in | |
1968 | * interrupt context and apply the current process NUMA policy. | |
1969 | * Returns NULL when no page can be allocated. | |
1da177e4 | 1970 | */ |
dd0fc66f | 1971 | struct page *alloc_pages_current(gfp_t gfp, unsigned order) |
1da177e4 | 1972 | { |
8d90274b | 1973 | struct mempolicy *pol = &default_policy; |
c0ff7453 | 1974 | struct page *page; |
1da177e4 | 1975 | |
8d90274b ON |
1976 | if (!in_interrupt() && !(gfp & __GFP_THISNODE)) |
1977 | pol = get_task_policy(current); | |
52cd3b07 LS |
1978 | |
1979 | /* | |
1980 | * No reference counting needed for current->mempolicy | |
1981 | * nor system default_policy | |
1982 | */ | |
45c4745a | 1983 | if (pol->mode == MPOL_INTERLEAVE) |
c0ff7453 MX |
1984 | page = alloc_page_interleave(gfp, order, interleave_nodes(pol)); |
1985 | else | |
1986 | page = __alloc_pages_nodemask(gfp, order, | |
04ec6264 | 1987 | policy_node(gfp, pol, numa_node_id()), |
5c4b4be3 | 1988 | policy_nodemask(gfp, pol)); |
cc9a6c87 | 1989 | |
c0ff7453 | 1990 | return page; |
1da177e4 LT |
1991 | } |
1992 | EXPORT_SYMBOL(alloc_pages_current); | |
1993 | ||
ef0855d3 ON |
1994 | int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst) |
1995 | { | |
1996 | struct mempolicy *pol = mpol_dup(vma_policy(src)); | |
1997 | ||
1998 | if (IS_ERR(pol)) | |
1999 | return PTR_ERR(pol); | |
2000 | dst->vm_policy = pol; | |
2001 | return 0; | |
2002 | } | |
2003 | ||
4225399a | 2004 | /* |
846a16bf | 2005 | * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it |
4225399a PJ |
2006 | * rebinds the mempolicy its copying by calling mpol_rebind_policy() |
2007 | * with the mems_allowed returned by cpuset_mems_allowed(). This | |
2008 | * keeps mempolicies cpuset relative after its cpuset moves. See | |
2009 | * further kernel/cpuset.c update_nodemask(). | |
708c1bbc MX |
2010 | * |
2011 | * current's mempolicy may be rebinded by the other task(the task that changes | |
2012 | * cpuset's mems), so we needn't do rebind work for current task. | |
4225399a | 2013 | */ |
4225399a | 2014 | |
846a16bf LS |
2015 | /* Slow path of a mempolicy duplicate */ |
2016 | struct mempolicy *__mpol_dup(struct mempolicy *old) | |
1da177e4 LT |
2017 | { |
2018 | struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL); | |
2019 | ||
2020 | if (!new) | |
2021 | return ERR_PTR(-ENOMEM); | |
708c1bbc MX |
2022 | |
2023 | /* task's mempolicy is protected by alloc_lock */ | |
2024 | if (old == current->mempolicy) { | |
2025 | task_lock(current); | |
2026 | *new = *old; | |
2027 | task_unlock(current); | |
2028 | } else | |
2029 | *new = *old; | |
2030 | ||
4225399a PJ |
2031 | if (current_cpuset_is_being_rebound()) { |
2032 | nodemask_t mems = cpuset_mems_allowed(current); | |
213980c0 | 2033 | mpol_rebind_policy(new, &mems); |
4225399a | 2034 | } |
1da177e4 | 2035 | atomic_set(&new->refcnt, 1); |
1da177e4 LT |
2036 | return new; |
2037 | } | |
2038 | ||
2039 | /* Slow path of a mempolicy comparison */ | |
fcfb4dcc | 2040 | bool __mpol_equal(struct mempolicy *a, struct mempolicy *b) |
1da177e4 LT |
2041 | { |
2042 | if (!a || !b) | |
fcfb4dcc | 2043 | return false; |
45c4745a | 2044 | if (a->mode != b->mode) |
fcfb4dcc | 2045 | return false; |
19800502 | 2046 | if (a->flags != b->flags) |
fcfb4dcc | 2047 | return false; |
19800502 BL |
2048 | if (mpol_store_user_nodemask(a)) |
2049 | if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask)) | |
fcfb4dcc | 2050 | return false; |
19800502 | 2051 | |
45c4745a | 2052 | switch (a->mode) { |
19770b32 MG |
2053 | case MPOL_BIND: |
2054 | /* Fall through */ | |
1da177e4 | 2055 | case MPOL_INTERLEAVE: |
fcfb4dcc | 2056 | return !!nodes_equal(a->v.nodes, b->v.nodes); |
1da177e4 | 2057 | case MPOL_PREFERRED: |
75719661 | 2058 | return a->v.preferred_node == b->v.preferred_node; |
1da177e4 LT |
2059 | default: |
2060 | BUG(); | |
fcfb4dcc | 2061 | return false; |
1da177e4 LT |
2062 | } |
2063 | } | |
2064 | ||
1da177e4 LT |
2065 | /* |
2066 | * Shared memory backing store policy support. | |
2067 | * | |
2068 | * Remember policies even when nobody has shared memory mapped. | |
2069 | * The policies are kept in Red-Black tree linked from the inode. | |
4a8c7bb5 | 2070 | * They are protected by the sp->lock rwlock, which should be held |
1da177e4 LT |
2071 | * for any accesses to the tree. |
2072 | */ | |
2073 | ||
4a8c7bb5 NZ |
2074 | /* |
2075 | * lookup first element intersecting start-end. Caller holds sp->lock for | |
2076 | * reading or for writing | |
2077 | */ | |
1da177e4 LT |
2078 | static struct sp_node * |
2079 | sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end) | |
2080 | { | |
2081 | struct rb_node *n = sp->root.rb_node; | |
2082 | ||
2083 | while (n) { | |
2084 | struct sp_node *p = rb_entry(n, struct sp_node, nd); | |
2085 | ||
2086 | if (start >= p->end) | |
2087 | n = n->rb_right; | |
2088 | else if (end <= p->start) | |
2089 | n = n->rb_left; | |
2090 | else | |
2091 | break; | |
2092 | } | |
2093 | if (!n) | |
2094 | return NULL; | |
2095 | for (;;) { | |
2096 | struct sp_node *w = NULL; | |
2097 | struct rb_node *prev = rb_prev(n); | |
2098 | if (!prev) | |
2099 | break; | |
2100 | w = rb_entry(prev, struct sp_node, nd); | |
2101 | if (w->end <= start) | |
2102 | break; | |
2103 | n = prev; | |
2104 | } | |
2105 | return rb_entry(n, struct sp_node, nd); | |
2106 | } | |
2107 | ||
4a8c7bb5 NZ |
2108 | /* |
2109 | * Insert a new shared policy into the list. Caller holds sp->lock for | |
2110 | * writing. | |
2111 | */ | |
1da177e4 LT |
2112 | static void sp_insert(struct shared_policy *sp, struct sp_node *new) |
2113 | { | |
2114 | struct rb_node **p = &sp->root.rb_node; | |
2115 | struct rb_node *parent = NULL; | |
2116 | struct sp_node *nd; | |
2117 | ||
2118 | while (*p) { | |
2119 | parent = *p; | |
2120 | nd = rb_entry(parent, struct sp_node, nd); | |
2121 | if (new->start < nd->start) | |
2122 | p = &(*p)->rb_left; | |
2123 | else if (new->end > nd->end) | |
2124 | p = &(*p)->rb_right; | |
2125 | else | |
2126 | BUG(); | |
2127 | } | |
2128 | rb_link_node(&new->nd, parent, p); | |
2129 | rb_insert_color(&new->nd, &sp->root); | |
140d5a49 | 2130 | pr_debug("inserting %lx-%lx: %d\n", new->start, new->end, |
45c4745a | 2131 | new->policy ? new->policy->mode : 0); |
1da177e4 LT |
2132 | } |
2133 | ||
2134 | /* Find shared policy intersecting idx */ | |
2135 | struct mempolicy * | |
2136 | mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx) | |
2137 | { | |
2138 | struct mempolicy *pol = NULL; | |
2139 | struct sp_node *sn; | |
2140 | ||
2141 | if (!sp->root.rb_node) | |
2142 | return NULL; | |
4a8c7bb5 | 2143 | read_lock(&sp->lock); |
1da177e4 LT |
2144 | sn = sp_lookup(sp, idx, idx+1); |
2145 | if (sn) { | |
2146 | mpol_get(sn->policy); | |
2147 | pol = sn->policy; | |
2148 | } | |
4a8c7bb5 | 2149 | read_unlock(&sp->lock); |
1da177e4 LT |
2150 | return pol; |
2151 | } | |
2152 | ||
63f74ca2 KM |
2153 | static void sp_free(struct sp_node *n) |
2154 | { | |
2155 | mpol_put(n->policy); | |
2156 | kmem_cache_free(sn_cache, n); | |
2157 | } | |
2158 | ||
771fb4d8 LS |
2159 | /** |
2160 | * mpol_misplaced - check whether current page node is valid in policy | |
2161 | * | |
b46e14ac FF |
2162 | * @page: page to be checked |
2163 | * @vma: vm area where page mapped | |
2164 | * @addr: virtual address where page mapped | |
771fb4d8 LS |
2165 | * |
2166 | * Lookup current policy node id for vma,addr and "compare to" page's | |
2167 | * node id. | |
2168 | * | |
2169 | * Returns: | |
2170 | * -1 - not misplaced, page is in the right node | |
2171 | * node - node id where the page should be | |
2172 | * | |
2173 | * Policy determination "mimics" alloc_page_vma(). | |
2174 | * Called from fault path where we know the vma and faulting address. | |
2175 | */ | |
2176 | int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr) | |
2177 | { | |
2178 | struct mempolicy *pol; | |
c33d6c06 | 2179 | struct zoneref *z; |
771fb4d8 LS |
2180 | int curnid = page_to_nid(page); |
2181 | unsigned long pgoff; | |
90572890 PZ |
2182 | int thiscpu = raw_smp_processor_id(); |
2183 | int thisnid = cpu_to_node(thiscpu); | |
771fb4d8 LS |
2184 | int polnid = -1; |
2185 | int ret = -1; | |
2186 | ||
2187 | BUG_ON(!vma); | |
2188 | ||
dd6eecb9 | 2189 | pol = get_vma_policy(vma, addr); |
771fb4d8 LS |
2190 | if (!(pol->flags & MPOL_F_MOF)) |
2191 | goto out; | |
2192 | ||
2193 | switch (pol->mode) { | |
2194 | case MPOL_INTERLEAVE: | |
2195 | BUG_ON(addr >= vma->vm_end); | |
2196 | BUG_ON(addr < vma->vm_start); | |
2197 | ||
2198 | pgoff = vma->vm_pgoff; | |
2199 | pgoff += (addr - vma->vm_start) >> PAGE_SHIFT; | |
2200 | polnid = offset_il_node(pol, vma, pgoff); | |
2201 | break; | |
2202 | ||
2203 | case MPOL_PREFERRED: | |
2204 | if (pol->flags & MPOL_F_LOCAL) | |
2205 | polnid = numa_node_id(); | |
2206 | else | |
2207 | polnid = pol->v.preferred_node; | |
2208 | break; | |
2209 | ||
2210 | case MPOL_BIND: | |
c33d6c06 | 2211 | |
771fb4d8 LS |
2212 | /* |
2213 | * allows binding to multiple nodes. | |
2214 | * use current page if in policy nodemask, | |
2215 | * else select nearest allowed node, if any. | |
2216 | * If no allowed nodes, use current [!misplaced]. | |
2217 | */ | |
2218 | if (node_isset(curnid, pol->v.nodes)) | |
2219 | goto out; | |
c33d6c06 | 2220 | z = first_zones_zonelist( |
771fb4d8 LS |
2221 | node_zonelist(numa_node_id(), GFP_HIGHUSER), |
2222 | gfp_zone(GFP_HIGHUSER), | |
c33d6c06 MG |
2223 | &pol->v.nodes); |
2224 | polnid = z->zone->node; | |
771fb4d8 LS |
2225 | break; |
2226 | ||
2227 | default: | |
2228 | BUG(); | |
2229 | } | |
5606e387 MG |
2230 | |
2231 | /* Migrate the page towards the node whose CPU is referencing it */ | |
e42c8ff2 | 2232 | if (pol->flags & MPOL_F_MORON) { |
90572890 | 2233 | polnid = thisnid; |
5606e387 | 2234 | |
10f39042 | 2235 | if (!should_numa_migrate_memory(current, page, curnid, thiscpu)) |
de1c9ce6 | 2236 | goto out; |
e42c8ff2 MG |
2237 | } |
2238 | ||
771fb4d8 LS |
2239 | if (curnid != polnid) |
2240 | ret = polnid; | |
2241 | out: | |
2242 | mpol_cond_put(pol); | |
2243 | ||
2244 | return ret; | |
2245 | } | |
2246 | ||
c11600e4 DR |
2247 | /* |
2248 | * Drop the (possibly final) reference to task->mempolicy. It needs to be | |
2249 | * dropped after task->mempolicy is set to NULL so that any allocation done as | |
2250 | * part of its kmem_cache_free(), such as by KASAN, doesn't reference a freed | |
2251 | * policy. | |
2252 | */ | |
2253 | void mpol_put_task_policy(struct task_struct *task) | |
2254 | { | |
2255 | struct mempolicy *pol; | |
2256 | ||
2257 | task_lock(task); | |
2258 | pol = task->mempolicy; | |
2259 | task->mempolicy = NULL; | |
2260 | task_unlock(task); | |
2261 | mpol_put(pol); | |
2262 | } | |
2263 | ||
1da177e4 LT |
2264 | static void sp_delete(struct shared_policy *sp, struct sp_node *n) |
2265 | { | |
140d5a49 | 2266 | pr_debug("deleting %lx-l%lx\n", n->start, n->end); |
1da177e4 | 2267 | rb_erase(&n->nd, &sp->root); |
63f74ca2 | 2268 | sp_free(n); |
1da177e4 LT |
2269 | } |
2270 | ||
42288fe3 MG |
2271 | static void sp_node_init(struct sp_node *node, unsigned long start, |
2272 | unsigned long end, struct mempolicy *pol) | |
2273 | { | |
2274 | node->start = start; | |
2275 | node->end = end; | |
2276 | node->policy = pol; | |
2277 | } | |
2278 | ||
dbcb0f19 AB |
2279 | static struct sp_node *sp_alloc(unsigned long start, unsigned long end, |
2280 | struct mempolicy *pol) | |
1da177e4 | 2281 | { |
869833f2 KM |
2282 | struct sp_node *n; |
2283 | struct mempolicy *newpol; | |
1da177e4 | 2284 | |
869833f2 | 2285 | n = kmem_cache_alloc(sn_cache, GFP_KERNEL); |
1da177e4 LT |
2286 | if (!n) |
2287 | return NULL; | |
869833f2 KM |
2288 | |
2289 | newpol = mpol_dup(pol); | |
2290 | if (IS_ERR(newpol)) { | |
2291 | kmem_cache_free(sn_cache, n); | |
2292 | return NULL; | |
2293 | } | |
2294 | newpol->flags |= MPOL_F_SHARED; | |
42288fe3 | 2295 | sp_node_init(n, start, end, newpol); |
869833f2 | 2296 | |
1da177e4 LT |
2297 | return n; |
2298 | } | |
2299 | ||
2300 | /* Replace a policy range. */ | |
2301 | static int shared_policy_replace(struct shared_policy *sp, unsigned long start, | |
2302 | unsigned long end, struct sp_node *new) | |
2303 | { | |
b22d127a | 2304 | struct sp_node *n; |
42288fe3 MG |
2305 | struct sp_node *n_new = NULL; |
2306 | struct mempolicy *mpol_new = NULL; | |
b22d127a | 2307 | int ret = 0; |
1da177e4 | 2308 | |
42288fe3 | 2309 | restart: |
4a8c7bb5 | 2310 | write_lock(&sp->lock); |
1da177e4 LT |
2311 | n = sp_lookup(sp, start, end); |
2312 | /* Take care of old policies in the same range. */ | |
2313 | while (n && n->start < end) { | |
2314 | struct rb_node *next = rb_next(&n->nd); | |
2315 | if (n->start >= start) { | |
2316 | if (n->end <= end) | |
2317 | sp_delete(sp, n); | |
2318 | else | |
2319 | n->start = end; | |
2320 | } else { | |
2321 | /* Old policy spanning whole new range. */ | |
2322 | if (n->end > end) { | |
42288fe3 MG |
2323 | if (!n_new) |
2324 | goto alloc_new; | |
2325 | ||
2326 | *mpol_new = *n->policy; | |
2327 | atomic_set(&mpol_new->refcnt, 1); | |
7880639c | 2328 | sp_node_init(n_new, end, n->end, mpol_new); |
1da177e4 | 2329 | n->end = start; |
5ca39575 | 2330 | sp_insert(sp, n_new); |
42288fe3 MG |
2331 | n_new = NULL; |
2332 | mpol_new = NULL; | |
1da177e4 LT |
2333 | break; |
2334 | } else | |
2335 | n->end = start; | |
2336 | } | |
2337 | if (!next) | |
2338 | break; | |
2339 | n = rb_entry(next, struct sp_node, nd); | |
2340 | } | |
2341 | if (new) | |
2342 | sp_insert(sp, new); | |
4a8c7bb5 | 2343 | write_unlock(&sp->lock); |
42288fe3 MG |
2344 | ret = 0; |
2345 | ||
2346 | err_out: | |
2347 | if (mpol_new) | |
2348 | mpol_put(mpol_new); | |
2349 | if (n_new) | |
2350 | kmem_cache_free(sn_cache, n_new); | |
2351 | ||
b22d127a | 2352 | return ret; |
42288fe3 MG |
2353 | |
2354 | alloc_new: | |
4a8c7bb5 | 2355 | write_unlock(&sp->lock); |
42288fe3 MG |
2356 | ret = -ENOMEM; |
2357 | n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL); | |
2358 | if (!n_new) | |
2359 | goto err_out; | |
2360 | mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL); | |
2361 | if (!mpol_new) | |
2362 | goto err_out; | |
2363 | goto restart; | |
1da177e4 LT |
2364 | } |
2365 | ||
71fe804b LS |
2366 | /** |
2367 | * mpol_shared_policy_init - initialize shared policy for inode | |
2368 | * @sp: pointer to inode shared policy | |
2369 | * @mpol: struct mempolicy to install | |
2370 | * | |
2371 | * Install non-NULL @mpol in inode's shared policy rb-tree. | |
2372 | * On entry, the current task has a reference on a non-NULL @mpol. | |
2373 | * This must be released on exit. | |
4bfc4495 | 2374 | * This is called at get_inode() calls and we can use GFP_KERNEL. |
71fe804b LS |
2375 | */ |
2376 | void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol) | |
2377 | { | |
58568d2a MX |
2378 | int ret; |
2379 | ||
71fe804b | 2380 | sp->root = RB_ROOT; /* empty tree == default mempolicy */ |
4a8c7bb5 | 2381 | rwlock_init(&sp->lock); |
71fe804b LS |
2382 | |
2383 | if (mpol) { | |
2384 | struct vm_area_struct pvma; | |
2385 | struct mempolicy *new; | |
4bfc4495 | 2386 | NODEMASK_SCRATCH(scratch); |
71fe804b | 2387 | |
4bfc4495 | 2388 | if (!scratch) |
5c0c1654 | 2389 | goto put_mpol; |
71fe804b LS |
2390 | /* contextualize the tmpfs mount point mempolicy */ |
2391 | new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask); | |
15d77835 | 2392 | if (IS_ERR(new)) |
0cae3457 | 2393 | goto free_scratch; /* no valid nodemask intersection */ |
58568d2a MX |
2394 | |
2395 | task_lock(current); | |
4bfc4495 | 2396 | ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch); |
58568d2a | 2397 | task_unlock(current); |
15d77835 | 2398 | if (ret) |
5c0c1654 | 2399 | goto put_new; |
71fe804b LS |
2400 | |
2401 | /* Create pseudo-vma that contains just the policy */ | |
2402 | memset(&pvma, 0, sizeof(struct vm_area_struct)); | |
2403 | pvma.vm_end = TASK_SIZE; /* policy covers entire file */ | |
2404 | mpol_set_shared_policy(sp, &pvma, new); /* adds ref */ | |
15d77835 | 2405 | |
5c0c1654 | 2406 | put_new: |
71fe804b | 2407 | mpol_put(new); /* drop initial ref */ |
0cae3457 | 2408 | free_scratch: |
4bfc4495 | 2409 | NODEMASK_SCRATCH_FREE(scratch); |
5c0c1654 LS |
2410 | put_mpol: |
2411 | mpol_put(mpol); /* drop our incoming ref on sb mpol */ | |
7339ff83 RH |
2412 | } |
2413 | } | |
2414 | ||
1da177e4 LT |
2415 | int mpol_set_shared_policy(struct shared_policy *info, |
2416 | struct vm_area_struct *vma, struct mempolicy *npol) | |
2417 | { | |
2418 | int err; | |
2419 | struct sp_node *new = NULL; | |
2420 | unsigned long sz = vma_pages(vma); | |
2421 | ||
028fec41 | 2422 | pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n", |
1da177e4 | 2423 | vma->vm_pgoff, |
45c4745a | 2424 | sz, npol ? npol->mode : -1, |
028fec41 | 2425 | npol ? npol->flags : -1, |
00ef2d2f | 2426 | npol ? nodes_addr(npol->v.nodes)[0] : NUMA_NO_NODE); |
1da177e4 LT |
2427 | |
2428 | if (npol) { | |
2429 | new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol); | |
2430 | if (!new) | |
2431 | return -ENOMEM; | |
2432 | } | |
2433 | err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new); | |
2434 | if (err && new) | |
63f74ca2 | 2435 | sp_free(new); |
1da177e4 LT |
2436 | return err; |
2437 | } | |
2438 | ||
2439 | /* Free a backing policy store on inode delete. */ | |
2440 | void mpol_free_shared_policy(struct shared_policy *p) | |
2441 | { | |
2442 | struct sp_node *n; | |
2443 | struct rb_node *next; | |
2444 | ||
2445 | if (!p->root.rb_node) | |
2446 | return; | |
4a8c7bb5 | 2447 | write_lock(&p->lock); |
1da177e4 LT |
2448 | next = rb_first(&p->root); |
2449 | while (next) { | |
2450 | n = rb_entry(next, struct sp_node, nd); | |
2451 | next = rb_next(&n->nd); | |
63f74ca2 | 2452 | sp_delete(p, n); |
1da177e4 | 2453 | } |
4a8c7bb5 | 2454 | write_unlock(&p->lock); |
1da177e4 LT |
2455 | } |
2456 | ||
1a687c2e | 2457 | #ifdef CONFIG_NUMA_BALANCING |
c297663c | 2458 | static int __initdata numabalancing_override; |
1a687c2e MG |
2459 | |
2460 | static void __init check_numabalancing_enable(void) | |
2461 | { | |
2462 | bool numabalancing_default = false; | |
2463 | ||
2464 | if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED)) | |
2465 | numabalancing_default = true; | |
2466 | ||
c297663c MG |
2467 | /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */ |
2468 | if (numabalancing_override) | |
2469 | set_numabalancing_state(numabalancing_override == 1); | |
2470 | ||
b0dc2b9b | 2471 | if (num_online_nodes() > 1 && !numabalancing_override) { |
756a025f | 2472 | pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n", |
c297663c | 2473 | numabalancing_default ? "Enabling" : "Disabling"); |
1a687c2e MG |
2474 | set_numabalancing_state(numabalancing_default); |
2475 | } | |
2476 | } | |
2477 | ||
2478 | static int __init setup_numabalancing(char *str) | |
2479 | { | |
2480 | int ret = 0; | |
2481 | if (!str) | |
2482 | goto out; | |
1a687c2e MG |
2483 | |
2484 | if (!strcmp(str, "enable")) { | |
c297663c | 2485 | numabalancing_override = 1; |
1a687c2e MG |
2486 | ret = 1; |
2487 | } else if (!strcmp(str, "disable")) { | |
c297663c | 2488 | numabalancing_override = -1; |
1a687c2e MG |
2489 | ret = 1; |
2490 | } | |
2491 | out: | |
2492 | if (!ret) | |
4a404bea | 2493 | pr_warn("Unable to parse numa_balancing=\n"); |
1a687c2e MG |
2494 | |
2495 | return ret; | |
2496 | } | |
2497 | __setup("numa_balancing=", setup_numabalancing); | |
2498 | #else | |
2499 | static inline void __init check_numabalancing_enable(void) | |
2500 | { | |
2501 | } | |
2502 | #endif /* CONFIG_NUMA_BALANCING */ | |
2503 | ||
1da177e4 LT |
2504 | /* assumes fs == KERNEL_DS */ |
2505 | void __init numa_policy_init(void) | |
2506 | { | |
b71636e2 PM |
2507 | nodemask_t interleave_nodes; |
2508 | unsigned long largest = 0; | |
2509 | int nid, prefer = 0; | |
2510 | ||
1da177e4 LT |
2511 | policy_cache = kmem_cache_create("numa_policy", |
2512 | sizeof(struct mempolicy), | |
20c2df83 | 2513 | 0, SLAB_PANIC, NULL); |
1da177e4 LT |
2514 | |
2515 | sn_cache = kmem_cache_create("shared_policy_node", | |
2516 | sizeof(struct sp_node), | |
20c2df83 | 2517 | 0, SLAB_PANIC, NULL); |
1da177e4 | 2518 | |
5606e387 MG |
2519 | for_each_node(nid) { |
2520 | preferred_node_policy[nid] = (struct mempolicy) { | |
2521 | .refcnt = ATOMIC_INIT(1), | |
2522 | .mode = MPOL_PREFERRED, | |
2523 | .flags = MPOL_F_MOF | MPOL_F_MORON, | |
2524 | .v = { .preferred_node = nid, }, | |
2525 | }; | |
2526 | } | |
2527 | ||
b71636e2 PM |
2528 | /* |
2529 | * Set interleaving policy for system init. Interleaving is only | |
2530 | * enabled across suitably sized nodes (default is >= 16MB), or | |
2531 | * fall back to the largest node if they're all smaller. | |
2532 | */ | |
2533 | nodes_clear(interleave_nodes); | |
01f13bd6 | 2534 | for_each_node_state(nid, N_MEMORY) { |
b71636e2 PM |
2535 | unsigned long total_pages = node_present_pages(nid); |
2536 | ||
2537 | /* Preserve the largest node */ | |
2538 | if (largest < total_pages) { | |
2539 | largest = total_pages; | |
2540 | prefer = nid; | |
2541 | } | |
2542 | ||
2543 | /* Interleave this node? */ | |
2544 | if ((total_pages << PAGE_SHIFT) >= (16 << 20)) | |
2545 | node_set(nid, interleave_nodes); | |
2546 | } | |
2547 | ||
2548 | /* All too small, use the largest */ | |
2549 | if (unlikely(nodes_empty(interleave_nodes))) | |
2550 | node_set(prefer, interleave_nodes); | |
1da177e4 | 2551 | |
028fec41 | 2552 | if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes)) |
b1de0d13 | 2553 | pr_err("%s: interleaving failed\n", __func__); |
1a687c2e MG |
2554 | |
2555 | check_numabalancing_enable(); | |
1da177e4 LT |
2556 | } |
2557 | ||
8bccd85f | 2558 | /* Reset policy of current process to default */ |
1da177e4 LT |
2559 | void numa_default_policy(void) |
2560 | { | |
028fec41 | 2561 | do_set_mempolicy(MPOL_DEFAULT, 0, NULL); |
1da177e4 | 2562 | } |
68860ec1 | 2563 | |
095f1fc4 LS |
2564 | /* |
2565 | * Parse and format mempolicy from/to strings | |
2566 | */ | |
2567 | ||
1a75a6c8 | 2568 | /* |
f2a07f40 | 2569 | * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag. |
1a75a6c8 | 2570 | */ |
345ace9c LS |
2571 | static const char * const policy_modes[] = |
2572 | { | |
2573 | [MPOL_DEFAULT] = "default", | |
2574 | [MPOL_PREFERRED] = "prefer", | |
2575 | [MPOL_BIND] = "bind", | |
2576 | [MPOL_INTERLEAVE] = "interleave", | |
d3a71033 | 2577 | [MPOL_LOCAL] = "local", |
345ace9c | 2578 | }; |
1a75a6c8 | 2579 | |
095f1fc4 LS |
2580 | |
2581 | #ifdef CONFIG_TMPFS | |
2582 | /** | |
f2a07f40 | 2583 | * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option. |
095f1fc4 | 2584 | * @str: string containing mempolicy to parse |
71fe804b | 2585 | * @mpol: pointer to struct mempolicy pointer, returned on success. |
095f1fc4 LS |
2586 | * |
2587 | * Format of input: | |
2588 | * <mode>[=<flags>][:<nodelist>] | |
2589 | * | |
71fe804b | 2590 | * On success, returns 0, else 1 |
095f1fc4 | 2591 | */ |
a7a88b23 | 2592 | int mpol_parse_str(char *str, struct mempolicy **mpol) |
095f1fc4 | 2593 | { |
71fe804b | 2594 | struct mempolicy *new = NULL; |
b4652e84 | 2595 | unsigned short mode; |
f2a07f40 | 2596 | unsigned short mode_flags; |
71fe804b | 2597 | nodemask_t nodes; |
095f1fc4 LS |
2598 | char *nodelist = strchr(str, ':'); |
2599 | char *flags = strchr(str, '='); | |
095f1fc4 LS |
2600 | int err = 1; |
2601 | ||
2602 | if (nodelist) { | |
2603 | /* NUL-terminate mode or flags string */ | |
2604 | *nodelist++ = '\0'; | |
71fe804b | 2605 | if (nodelist_parse(nodelist, nodes)) |
095f1fc4 | 2606 | goto out; |
01f13bd6 | 2607 | if (!nodes_subset(nodes, node_states[N_MEMORY])) |
095f1fc4 | 2608 | goto out; |
71fe804b LS |
2609 | } else |
2610 | nodes_clear(nodes); | |
2611 | ||
095f1fc4 LS |
2612 | if (flags) |
2613 | *flags++ = '\0'; /* terminate mode string */ | |
2614 | ||
479e2802 | 2615 | for (mode = 0; mode < MPOL_MAX; mode++) { |
345ace9c | 2616 | if (!strcmp(str, policy_modes[mode])) { |
095f1fc4 LS |
2617 | break; |
2618 | } | |
2619 | } | |
a720094d | 2620 | if (mode >= MPOL_MAX) |
095f1fc4 LS |
2621 | goto out; |
2622 | ||
71fe804b | 2623 | switch (mode) { |
095f1fc4 | 2624 | case MPOL_PREFERRED: |
71fe804b LS |
2625 | /* |
2626 | * Insist on a nodelist of one node only | |
2627 | */ | |
095f1fc4 LS |
2628 | if (nodelist) { |
2629 | char *rest = nodelist; | |
2630 | while (isdigit(*rest)) | |
2631 | rest++; | |
926f2ae0 KM |
2632 | if (*rest) |
2633 | goto out; | |
095f1fc4 LS |
2634 | } |
2635 | break; | |
095f1fc4 LS |
2636 | case MPOL_INTERLEAVE: |
2637 | /* | |
2638 | * Default to online nodes with memory if no nodelist | |
2639 | */ | |
2640 | if (!nodelist) | |
01f13bd6 | 2641 | nodes = node_states[N_MEMORY]; |
3f226aa1 | 2642 | break; |
71fe804b | 2643 | case MPOL_LOCAL: |
3f226aa1 | 2644 | /* |
71fe804b | 2645 | * Don't allow a nodelist; mpol_new() checks flags |
3f226aa1 | 2646 | */ |
71fe804b | 2647 | if (nodelist) |
3f226aa1 | 2648 | goto out; |
71fe804b | 2649 | mode = MPOL_PREFERRED; |
3f226aa1 | 2650 | break; |
413b43de RT |
2651 | case MPOL_DEFAULT: |
2652 | /* | |
2653 | * Insist on a empty nodelist | |
2654 | */ | |
2655 | if (!nodelist) | |
2656 | err = 0; | |
2657 | goto out; | |
d69b2e63 KM |
2658 | case MPOL_BIND: |
2659 | /* | |
2660 | * Insist on a nodelist | |
2661 | */ | |
2662 | if (!nodelist) | |
2663 | goto out; | |
095f1fc4 LS |
2664 | } |
2665 | ||
71fe804b | 2666 | mode_flags = 0; |
095f1fc4 LS |
2667 | if (flags) { |
2668 | /* | |
2669 | * Currently, we only support two mutually exclusive | |
2670 | * mode flags. | |
2671 | */ | |
2672 | if (!strcmp(flags, "static")) | |
71fe804b | 2673 | mode_flags |= MPOL_F_STATIC_NODES; |
095f1fc4 | 2674 | else if (!strcmp(flags, "relative")) |
71fe804b | 2675 | mode_flags |= MPOL_F_RELATIVE_NODES; |
095f1fc4 | 2676 | else |
926f2ae0 | 2677 | goto out; |
095f1fc4 | 2678 | } |
71fe804b LS |
2679 | |
2680 | new = mpol_new(mode, mode_flags, &nodes); | |
2681 | if (IS_ERR(new)) | |
926f2ae0 KM |
2682 | goto out; |
2683 | ||
f2a07f40 HD |
2684 | /* |
2685 | * Save nodes for mpol_to_str() to show the tmpfs mount options | |
2686 | * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo. | |
2687 | */ | |
2688 | if (mode != MPOL_PREFERRED) | |
2689 | new->v.nodes = nodes; | |
2690 | else if (nodelist) | |
2691 | new->v.preferred_node = first_node(nodes); | |
2692 | else | |
2693 | new->flags |= MPOL_F_LOCAL; | |
2694 | ||
2695 | /* | |
2696 | * Save nodes for contextualization: this will be used to "clone" | |
2697 | * the mempolicy in a specific context [cpuset] at a later time. | |
2698 | */ | |
2699 | new->w.user_nodemask = nodes; | |
2700 | ||
926f2ae0 | 2701 | err = 0; |
71fe804b | 2702 | |
095f1fc4 LS |
2703 | out: |
2704 | /* Restore string for error message */ | |
2705 | if (nodelist) | |
2706 | *--nodelist = ':'; | |
2707 | if (flags) | |
2708 | *--flags = '='; | |
71fe804b LS |
2709 | if (!err) |
2710 | *mpol = new; | |
095f1fc4 LS |
2711 | return err; |
2712 | } | |
2713 | #endif /* CONFIG_TMPFS */ | |
2714 | ||
71fe804b LS |
2715 | /** |
2716 | * mpol_to_str - format a mempolicy structure for printing | |
2717 | * @buffer: to contain formatted mempolicy string | |
2718 | * @maxlen: length of @buffer | |
2719 | * @pol: pointer to mempolicy to be formatted | |
71fe804b | 2720 | * |
948927ee DR |
2721 | * Convert @pol into a string. If @buffer is too short, truncate the string. |
2722 | * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the | |
2723 | * longest flag, "relative", and to display at least a few node ids. | |
1a75a6c8 | 2724 | */ |
948927ee | 2725 | void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol) |
1a75a6c8 CL |
2726 | { |
2727 | char *p = buffer; | |
948927ee DR |
2728 | nodemask_t nodes = NODE_MASK_NONE; |
2729 | unsigned short mode = MPOL_DEFAULT; | |
2730 | unsigned short flags = 0; | |
2291990a | 2731 | |
8790c71a | 2732 | if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) { |
bea904d5 | 2733 | mode = pol->mode; |
948927ee DR |
2734 | flags = pol->flags; |
2735 | } | |
bea904d5 | 2736 | |
1a75a6c8 CL |
2737 | switch (mode) { |
2738 | case MPOL_DEFAULT: | |
1a75a6c8 | 2739 | break; |
1a75a6c8 | 2740 | case MPOL_PREFERRED: |
fc36b8d3 | 2741 | if (flags & MPOL_F_LOCAL) |
f2a07f40 | 2742 | mode = MPOL_LOCAL; |
53f2556b | 2743 | else |
fc36b8d3 | 2744 | node_set(pol->v.preferred_node, nodes); |
1a75a6c8 | 2745 | break; |
1a75a6c8 | 2746 | case MPOL_BIND: |
1a75a6c8 | 2747 | case MPOL_INTERLEAVE: |
f2a07f40 | 2748 | nodes = pol->v.nodes; |
1a75a6c8 | 2749 | break; |
1a75a6c8 | 2750 | default: |
948927ee DR |
2751 | WARN_ON_ONCE(1); |
2752 | snprintf(p, maxlen, "unknown"); | |
2753 | return; | |
1a75a6c8 CL |
2754 | } |
2755 | ||
b7a9f420 | 2756 | p += snprintf(p, maxlen, "%s", policy_modes[mode]); |
1a75a6c8 | 2757 | |
fc36b8d3 | 2758 | if (flags & MPOL_MODE_FLAGS) { |
948927ee | 2759 | p += snprintf(p, buffer + maxlen - p, "="); |
f5b087b5 | 2760 | |
2291990a LS |
2761 | /* |
2762 | * Currently, the only defined flags are mutually exclusive | |
2763 | */ | |
f5b087b5 | 2764 | if (flags & MPOL_F_STATIC_NODES) |
2291990a LS |
2765 | p += snprintf(p, buffer + maxlen - p, "static"); |
2766 | else if (flags & MPOL_F_RELATIVE_NODES) | |
2767 | p += snprintf(p, buffer + maxlen - p, "relative"); | |
f5b087b5 DR |
2768 | } |
2769 | ||
9e763e0f TH |
2770 | if (!nodes_empty(nodes)) |
2771 | p += scnprintf(p, buffer + maxlen - p, ":%*pbl", | |
2772 | nodemask_pr_args(&nodes)); | |
1a75a6c8 | 2773 | } |