| 1 | // SPDX-License-Identifier: GPL-2.0-only |
| 2 | /* |
| 3 | * Simple NUMA memory policy for the Linux kernel. |
| 4 | * |
| 5 | * Copyright 2003,2004 Andi Kleen, SuSE Labs. |
| 6 | * (C) Copyright 2005 Christoph Lameter, Silicon Graphics, Inc. |
| 7 | * |
| 8 | * NUMA policy allows the user to give hints in which node(s) memory should |
| 9 | * be allocated. |
| 10 | * |
| 11 | * Support four policies per VMA and per process: |
| 12 | * |
| 13 | * The VMA policy has priority over the process policy for a page fault. |
| 14 | * |
| 15 | * interleave Allocate memory interleaved over a set of nodes, |
| 16 | * with normal fallback if it fails. |
| 17 | * For VMA based allocations this interleaves based on the |
| 18 | * offset into the backing object or offset into the mapping |
| 19 | * for anonymous memory. For process policy an process counter |
| 20 | * is used. |
| 21 | * |
| 22 | * bind Only allocate memory on a specific set of nodes, |
| 23 | * no fallback. |
| 24 | * FIXME: memory is allocated starting with the first node |
| 25 | * to the last. It would be better if bind would truly restrict |
| 26 | * the allocation to memory nodes instead |
| 27 | * |
| 28 | * preferred Try a specific node first before normal fallback. |
| 29 | * As a special case NUMA_NO_NODE here means do the allocation |
| 30 | * on the local CPU. This is normally identical to default, |
| 31 | * but useful to set in a VMA when you have a non default |
| 32 | * process policy. |
| 33 | * |
| 34 | * default Allocate on the local node first, or when on a VMA |
| 35 | * use the process policy. This is what Linux always did |
| 36 | * in a NUMA aware kernel and still does by, ahem, default. |
| 37 | * |
| 38 | * The process policy is applied for most non interrupt memory allocations |
| 39 | * in that process' context. Interrupts ignore the policies and always |
| 40 | * try to allocate on the local CPU. The VMA policy is only applied for memory |
| 41 | * allocations for a VMA in the VM. |
| 42 | * |
| 43 | * Currently there are a few corner cases in swapping where the policy |
| 44 | * is not applied, but the majority should be handled. When process policy |
| 45 | * is used it is not remembered over swap outs/swap ins. |
| 46 | * |
| 47 | * Only the highest zone in the zone hierarchy gets policied. Allocations |
| 48 | * requesting a lower zone just use default policy. This implies that |
| 49 | * on systems with highmem kernel lowmem allocation don't get policied. |
| 50 | * Same with GFP_DMA allocations. |
| 51 | * |
| 52 | * For shmfs/tmpfs/hugetlbfs shared memory the policy is shared between |
| 53 | * all users and remembered even when nobody has memory mapped. |
| 54 | */ |
| 55 | |
| 56 | /* Notebook: |
| 57 | fix mmap readahead to honour policy and enable policy for any page cache |
| 58 | object |
| 59 | statistics for bigpages |
| 60 | global policy for page cache? currently it uses process policy. Requires |
| 61 | first item above. |
| 62 | handle mremap for shared memory (currently ignored for the policy) |
| 63 | grows down? |
| 64 | make bind policy root only? It can trigger oom much faster and the |
| 65 | kernel is not always grateful with that. |
| 66 | */ |
| 67 | |
| 68 | #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt |
| 69 | |
| 70 | #include <linux/mempolicy.h> |
| 71 | #include <linux/pagewalk.h> |
| 72 | #include <linux/highmem.h> |
| 73 | #include <linux/hugetlb.h> |
| 74 | #include <linux/kernel.h> |
| 75 | #include <linux/sched.h> |
| 76 | #include <linux/sched/mm.h> |
| 77 | #include <linux/sched/numa_balancing.h> |
| 78 | #include <linux/sched/task.h> |
| 79 | #include <linux/nodemask.h> |
| 80 | #include <linux/cpuset.h> |
| 81 | #include <linux/slab.h> |
| 82 | #include <linux/string.h> |
| 83 | #include <linux/export.h> |
| 84 | #include <linux/nsproxy.h> |
| 85 | #include <linux/interrupt.h> |
| 86 | #include <linux/init.h> |
| 87 | #include <linux/compat.h> |
| 88 | #include <linux/ptrace.h> |
| 89 | #include <linux/swap.h> |
| 90 | #include <linux/seq_file.h> |
| 91 | #include <linux/proc_fs.h> |
| 92 | #include <linux/migrate.h> |
| 93 | #include <linux/ksm.h> |
| 94 | #include <linux/rmap.h> |
| 95 | #include <linux/security.h> |
| 96 | #include <linux/syscalls.h> |
| 97 | #include <linux/ctype.h> |
| 98 | #include <linux/mm_inline.h> |
| 99 | #include <linux/mmu_notifier.h> |
| 100 | #include <linux/printk.h> |
| 101 | #include <linux/swapops.h> |
| 102 | |
| 103 | #include <asm/tlbflush.h> |
| 104 | #include <linux/uaccess.h> |
| 105 | |
| 106 | #include "internal.h" |
| 107 | |
| 108 | /* Internal flags */ |
| 109 | #define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0) /* Skip checks for continuous vmas */ |
| 110 | #define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1) /* Invert check for nodemask */ |
| 111 | |
| 112 | static struct kmem_cache *policy_cache; |
| 113 | static struct kmem_cache *sn_cache; |
| 114 | |
| 115 | /* Highest zone. An specific allocation for a zone below that is not |
| 116 | policied. */ |
| 117 | enum zone_type policy_zone = 0; |
| 118 | |
| 119 | /* |
| 120 | * run-time system-wide default policy => local allocation |
| 121 | */ |
| 122 | static struct mempolicy default_policy = { |
| 123 | .refcnt = ATOMIC_INIT(1), /* never free it */ |
| 124 | .mode = MPOL_PREFERRED, |
| 125 | .flags = MPOL_F_LOCAL, |
| 126 | }; |
| 127 | |
| 128 | static struct mempolicy preferred_node_policy[MAX_NUMNODES]; |
| 129 | |
| 130 | /** |
| 131 | * numa_map_to_online_node - Find closest online node |
| 132 | * @nid: Node id to start the search |
| 133 | * |
| 134 | * Lookup the next closest node by distance if @nid is not online. |
| 135 | */ |
| 136 | int numa_map_to_online_node(int node) |
| 137 | { |
| 138 | int min_dist = INT_MAX, dist, n, min_node; |
| 139 | |
| 140 | if (node == NUMA_NO_NODE || node_online(node)) |
| 141 | return node; |
| 142 | |
| 143 | min_node = node; |
| 144 | for_each_online_node(n) { |
| 145 | dist = node_distance(node, n); |
| 146 | if (dist < min_dist) { |
| 147 | min_dist = dist; |
| 148 | min_node = n; |
| 149 | } |
| 150 | } |
| 151 | |
| 152 | return min_node; |
| 153 | } |
| 154 | EXPORT_SYMBOL_GPL(numa_map_to_online_node); |
| 155 | |
| 156 | struct mempolicy *get_task_policy(struct task_struct *p) |
| 157 | { |
| 158 | struct mempolicy *pol = p->mempolicy; |
| 159 | int node; |
| 160 | |
| 161 | if (pol) |
| 162 | return pol; |
| 163 | |
| 164 | node = numa_node_id(); |
| 165 | if (node != NUMA_NO_NODE) { |
| 166 | pol = &preferred_node_policy[node]; |
| 167 | /* preferred_node_policy is not initialised early in boot */ |
| 168 | if (pol->mode) |
| 169 | return pol; |
| 170 | } |
| 171 | |
| 172 | return &default_policy; |
| 173 | } |
| 174 | |
| 175 | static const struct mempolicy_operations { |
| 176 | int (*create)(struct mempolicy *pol, const nodemask_t *nodes); |
| 177 | void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes); |
| 178 | } mpol_ops[MPOL_MAX]; |
| 179 | |
| 180 | static inline int mpol_store_user_nodemask(const struct mempolicy *pol) |
| 181 | { |
| 182 | return pol->flags & MPOL_MODE_FLAGS; |
| 183 | } |
| 184 | |
| 185 | static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig, |
| 186 | const nodemask_t *rel) |
| 187 | { |
| 188 | nodemask_t tmp; |
| 189 | nodes_fold(tmp, *orig, nodes_weight(*rel)); |
| 190 | nodes_onto(*ret, tmp, *rel); |
| 191 | } |
| 192 | |
| 193 | static int mpol_new_interleave(struct mempolicy *pol, const nodemask_t *nodes) |
| 194 | { |
| 195 | if (nodes_empty(*nodes)) |
| 196 | return -EINVAL; |
| 197 | pol->v.nodes = *nodes; |
| 198 | return 0; |
| 199 | } |
| 200 | |
| 201 | static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes) |
| 202 | { |
| 203 | if (!nodes) |
| 204 | pol->flags |= MPOL_F_LOCAL; /* local allocation */ |
| 205 | else if (nodes_empty(*nodes)) |
| 206 | return -EINVAL; /* no allowed nodes */ |
| 207 | else |
| 208 | pol->v.preferred_node = first_node(*nodes); |
| 209 | return 0; |
| 210 | } |
| 211 | |
| 212 | static int mpol_new_bind(struct mempolicy *pol, const nodemask_t *nodes) |
| 213 | { |
| 214 | if (nodes_empty(*nodes)) |
| 215 | return -EINVAL; |
| 216 | pol->v.nodes = *nodes; |
| 217 | return 0; |
| 218 | } |
| 219 | |
| 220 | /* |
| 221 | * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if |
| 222 | * any, for the new policy. mpol_new() has already validated the nodes |
| 223 | * parameter with respect to the policy mode and flags. But, we need to |
| 224 | * handle an empty nodemask with MPOL_PREFERRED here. |
| 225 | * |
| 226 | * Must be called holding task's alloc_lock to protect task's mems_allowed |
| 227 | * and mempolicy. May also be called holding the mmap_semaphore for write. |
| 228 | */ |
| 229 | static int mpol_set_nodemask(struct mempolicy *pol, |
| 230 | const nodemask_t *nodes, struct nodemask_scratch *nsc) |
| 231 | { |
| 232 | int ret; |
| 233 | |
| 234 | /* if mode is MPOL_DEFAULT, pol is NULL. This is right. */ |
| 235 | if (pol == NULL) |
| 236 | return 0; |
| 237 | /* Check N_MEMORY */ |
| 238 | nodes_and(nsc->mask1, |
| 239 | cpuset_current_mems_allowed, node_states[N_MEMORY]); |
| 240 | |
| 241 | VM_BUG_ON(!nodes); |
| 242 | if (pol->mode == MPOL_PREFERRED && nodes_empty(*nodes)) |
| 243 | nodes = NULL; /* explicit local allocation */ |
| 244 | else { |
| 245 | if (pol->flags & MPOL_F_RELATIVE_NODES) |
| 246 | mpol_relative_nodemask(&nsc->mask2, nodes, &nsc->mask1); |
| 247 | else |
| 248 | nodes_and(nsc->mask2, *nodes, nsc->mask1); |
| 249 | |
| 250 | if (mpol_store_user_nodemask(pol)) |
| 251 | pol->w.user_nodemask = *nodes; |
| 252 | else |
| 253 | pol->w.cpuset_mems_allowed = |
| 254 | cpuset_current_mems_allowed; |
| 255 | } |
| 256 | |
| 257 | if (nodes) |
| 258 | ret = mpol_ops[pol->mode].create(pol, &nsc->mask2); |
| 259 | else |
| 260 | ret = mpol_ops[pol->mode].create(pol, NULL); |
| 261 | return ret; |
| 262 | } |
| 263 | |
| 264 | /* |
| 265 | * This function just creates a new policy, does some check and simple |
| 266 | * initialization. You must invoke mpol_set_nodemask() to set nodes. |
| 267 | */ |
| 268 | static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags, |
| 269 | nodemask_t *nodes) |
| 270 | { |
| 271 | struct mempolicy *policy; |
| 272 | |
| 273 | pr_debug("setting mode %d flags %d nodes[0] %lx\n", |
| 274 | mode, flags, nodes ? nodes_addr(*nodes)[0] : NUMA_NO_NODE); |
| 275 | |
| 276 | if (mode == MPOL_DEFAULT) { |
| 277 | if (nodes && !nodes_empty(*nodes)) |
| 278 | return ERR_PTR(-EINVAL); |
| 279 | return NULL; |
| 280 | } |
| 281 | VM_BUG_ON(!nodes); |
| 282 | |
| 283 | /* |
| 284 | * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or |
| 285 | * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation). |
| 286 | * All other modes require a valid pointer to a non-empty nodemask. |
| 287 | */ |
| 288 | if (mode == MPOL_PREFERRED) { |
| 289 | if (nodes_empty(*nodes)) { |
| 290 | if (((flags & MPOL_F_STATIC_NODES) || |
| 291 | (flags & MPOL_F_RELATIVE_NODES))) |
| 292 | return ERR_PTR(-EINVAL); |
| 293 | } |
| 294 | } else if (mode == MPOL_LOCAL) { |
| 295 | if (!nodes_empty(*nodes) || |
| 296 | (flags & MPOL_F_STATIC_NODES) || |
| 297 | (flags & MPOL_F_RELATIVE_NODES)) |
| 298 | return ERR_PTR(-EINVAL); |
| 299 | mode = MPOL_PREFERRED; |
| 300 | } else if (nodes_empty(*nodes)) |
| 301 | return ERR_PTR(-EINVAL); |
| 302 | policy = kmem_cache_alloc(policy_cache, GFP_KERNEL); |
| 303 | if (!policy) |
| 304 | return ERR_PTR(-ENOMEM); |
| 305 | atomic_set(&policy->refcnt, 1); |
| 306 | policy->mode = mode; |
| 307 | policy->flags = flags; |
| 308 | |
| 309 | return policy; |
| 310 | } |
| 311 | |
| 312 | /* Slow path of a mpol destructor. */ |
| 313 | void __mpol_put(struct mempolicy *p) |
| 314 | { |
| 315 | if (!atomic_dec_and_test(&p->refcnt)) |
| 316 | return; |
| 317 | kmem_cache_free(policy_cache, p); |
| 318 | } |
| 319 | |
| 320 | static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes) |
| 321 | { |
| 322 | } |
| 323 | |
| 324 | static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes) |
| 325 | { |
| 326 | nodemask_t tmp; |
| 327 | |
| 328 | if (pol->flags & MPOL_F_STATIC_NODES) |
| 329 | nodes_and(tmp, pol->w.user_nodemask, *nodes); |
| 330 | else if (pol->flags & MPOL_F_RELATIVE_NODES) |
| 331 | mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes); |
| 332 | else { |
| 333 | nodes_remap(tmp, pol->v.nodes,pol->w.cpuset_mems_allowed, |
| 334 | *nodes); |
| 335 | pol->w.cpuset_mems_allowed = *nodes; |
| 336 | } |
| 337 | |
| 338 | if (nodes_empty(tmp)) |
| 339 | tmp = *nodes; |
| 340 | |
| 341 | pol->v.nodes = tmp; |
| 342 | } |
| 343 | |
| 344 | static void mpol_rebind_preferred(struct mempolicy *pol, |
| 345 | const nodemask_t *nodes) |
| 346 | { |
| 347 | nodemask_t tmp; |
| 348 | |
| 349 | if (pol->flags & MPOL_F_STATIC_NODES) { |
| 350 | int node = first_node(pol->w.user_nodemask); |
| 351 | |
| 352 | if (node_isset(node, *nodes)) { |
| 353 | pol->v.preferred_node = node; |
| 354 | pol->flags &= ~MPOL_F_LOCAL; |
| 355 | } else |
| 356 | pol->flags |= MPOL_F_LOCAL; |
| 357 | } else if (pol->flags & MPOL_F_RELATIVE_NODES) { |
| 358 | mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes); |
| 359 | pol->v.preferred_node = first_node(tmp); |
| 360 | } else if (!(pol->flags & MPOL_F_LOCAL)) { |
| 361 | pol->v.preferred_node = node_remap(pol->v.preferred_node, |
| 362 | pol->w.cpuset_mems_allowed, |
| 363 | *nodes); |
| 364 | pol->w.cpuset_mems_allowed = *nodes; |
| 365 | } |
| 366 | } |
| 367 | |
| 368 | /* |
| 369 | * mpol_rebind_policy - Migrate a policy to a different set of nodes |
| 370 | * |
| 371 | * Per-vma policies are protected by mmap_sem. Allocations using per-task |
| 372 | * policies are protected by task->mems_allowed_seq to prevent a premature |
| 373 | * OOM/allocation failure due to parallel nodemask modification. |
| 374 | */ |
| 375 | static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask) |
| 376 | { |
| 377 | if (!pol) |
| 378 | return; |
| 379 | if (!mpol_store_user_nodemask(pol) && !(pol->flags & MPOL_F_LOCAL) && |
| 380 | nodes_equal(pol->w.cpuset_mems_allowed, *newmask)) |
| 381 | return; |
| 382 | |
| 383 | mpol_ops[pol->mode].rebind(pol, newmask); |
| 384 | } |
| 385 | |
| 386 | /* |
| 387 | * Wrapper for mpol_rebind_policy() that just requires task |
| 388 | * pointer, and updates task mempolicy. |
| 389 | * |
| 390 | * Called with task's alloc_lock held. |
| 391 | */ |
| 392 | |
| 393 | void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new) |
| 394 | { |
| 395 | mpol_rebind_policy(tsk->mempolicy, new); |
| 396 | } |
| 397 | |
| 398 | /* |
| 399 | * Rebind each vma in mm to new nodemask. |
| 400 | * |
| 401 | * Call holding a reference to mm. Takes mm->mmap_sem during call. |
| 402 | */ |
| 403 | |
| 404 | void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new) |
| 405 | { |
| 406 | struct vm_area_struct *vma; |
| 407 | |
| 408 | down_write(&mm->mmap_sem); |
| 409 | for (vma = mm->mmap; vma; vma = vma->vm_next) |
| 410 | mpol_rebind_policy(vma->vm_policy, new); |
| 411 | up_write(&mm->mmap_sem); |
| 412 | } |
| 413 | |
| 414 | static const struct mempolicy_operations mpol_ops[MPOL_MAX] = { |
| 415 | [MPOL_DEFAULT] = { |
| 416 | .rebind = mpol_rebind_default, |
| 417 | }, |
| 418 | [MPOL_INTERLEAVE] = { |
| 419 | .create = mpol_new_interleave, |
| 420 | .rebind = mpol_rebind_nodemask, |
| 421 | }, |
| 422 | [MPOL_PREFERRED] = { |
| 423 | .create = mpol_new_preferred, |
| 424 | .rebind = mpol_rebind_preferred, |
| 425 | }, |
| 426 | [MPOL_BIND] = { |
| 427 | .create = mpol_new_bind, |
| 428 | .rebind = mpol_rebind_nodemask, |
| 429 | }, |
| 430 | }; |
| 431 | |
| 432 | static int migrate_page_add(struct page *page, struct list_head *pagelist, |
| 433 | unsigned long flags); |
| 434 | |
| 435 | struct queue_pages { |
| 436 | struct list_head *pagelist; |
| 437 | unsigned long flags; |
| 438 | nodemask_t *nmask; |
| 439 | unsigned long start; |
| 440 | unsigned long end; |
| 441 | struct vm_area_struct *first; |
| 442 | }; |
| 443 | |
| 444 | /* |
| 445 | * Check if the page's nid is in qp->nmask. |
| 446 | * |
| 447 | * If MPOL_MF_INVERT is set in qp->flags, check if the nid is |
| 448 | * in the invert of qp->nmask. |
| 449 | */ |
| 450 | static inline bool queue_pages_required(struct page *page, |
| 451 | struct queue_pages *qp) |
| 452 | { |
| 453 | int nid = page_to_nid(page); |
| 454 | unsigned long flags = qp->flags; |
| 455 | |
| 456 | return node_isset(nid, *qp->nmask) == !(flags & MPOL_MF_INVERT); |
| 457 | } |
| 458 | |
| 459 | /* |
| 460 | * queue_pages_pmd() has four possible return values: |
| 461 | * 0 - pages are placed on the right node or queued successfully. |
| 462 | * 1 - there is unmovable page, and MPOL_MF_MOVE* & MPOL_MF_STRICT were |
| 463 | * specified. |
| 464 | * 2 - THP was split. |
| 465 | * -EIO - is migration entry or only MPOL_MF_STRICT was specified and an |
| 466 | * existing page was already on a node that does not follow the |
| 467 | * policy. |
| 468 | */ |
| 469 | static int queue_pages_pmd(pmd_t *pmd, spinlock_t *ptl, unsigned long addr, |
| 470 | unsigned long end, struct mm_walk *walk) |
| 471 | __releases(ptl) |
| 472 | { |
| 473 | int ret = 0; |
| 474 | struct page *page; |
| 475 | struct queue_pages *qp = walk->private; |
| 476 | unsigned long flags; |
| 477 | |
| 478 | if (unlikely(is_pmd_migration_entry(*pmd))) { |
| 479 | ret = -EIO; |
| 480 | goto unlock; |
| 481 | } |
| 482 | page = pmd_page(*pmd); |
| 483 | if (is_huge_zero_page(page)) { |
| 484 | spin_unlock(ptl); |
| 485 | __split_huge_pmd(walk->vma, pmd, addr, false, NULL); |
| 486 | ret = 2; |
| 487 | goto out; |
| 488 | } |
| 489 | if (!queue_pages_required(page, qp)) |
| 490 | goto unlock; |
| 491 | |
| 492 | flags = qp->flags; |
| 493 | /* go to thp migration */ |
| 494 | if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) { |
| 495 | if (!vma_migratable(walk->vma) || |
| 496 | migrate_page_add(page, qp->pagelist, flags)) { |
| 497 | ret = 1; |
| 498 | goto unlock; |
| 499 | } |
| 500 | } else |
| 501 | ret = -EIO; |
| 502 | unlock: |
| 503 | spin_unlock(ptl); |
| 504 | out: |
| 505 | return ret; |
| 506 | } |
| 507 | |
| 508 | /* |
| 509 | * Scan through pages checking if pages follow certain conditions, |
| 510 | * and move them to the pagelist if they do. |
| 511 | * |
| 512 | * queue_pages_pte_range() has three possible return values: |
| 513 | * 0 - pages are placed on the right node or queued successfully. |
| 514 | * 1 - there is unmovable page, and MPOL_MF_MOVE* & MPOL_MF_STRICT were |
| 515 | * specified. |
| 516 | * -EIO - only MPOL_MF_STRICT was specified and an existing page was already |
| 517 | * on a node that does not follow the policy. |
| 518 | */ |
| 519 | static int queue_pages_pte_range(pmd_t *pmd, unsigned long addr, |
| 520 | unsigned long end, struct mm_walk *walk) |
| 521 | { |
| 522 | struct vm_area_struct *vma = walk->vma; |
| 523 | struct page *page; |
| 524 | struct queue_pages *qp = walk->private; |
| 525 | unsigned long flags = qp->flags; |
| 526 | int ret; |
| 527 | bool has_unmovable = false; |
| 528 | pte_t *pte; |
| 529 | spinlock_t *ptl; |
| 530 | |
| 531 | ptl = pmd_trans_huge_lock(pmd, vma); |
| 532 | if (ptl) { |
| 533 | ret = queue_pages_pmd(pmd, ptl, addr, end, walk); |
| 534 | if (ret != 2) |
| 535 | return ret; |
| 536 | } |
| 537 | /* THP was split, fall through to pte walk */ |
| 538 | |
| 539 | if (pmd_trans_unstable(pmd)) |
| 540 | return 0; |
| 541 | |
| 542 | pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl); |
| 543 | for (; addr != end; pte++, addr += PAGE_SIZE) { |
| 544 | if (!pte_present(*pte)) |
| 545 | continue; |
| 546 | page = vm_normal_page(vma, addr, *pte); |
| 547 | if (!page) |
| 548 | continue; |
| 549 | /* |
| 550 | * vm_normal_page() filters out zero pages, but there might |
| 551 | * still be PageReserved pages to skip, perhaps in a VDSO. |
| 552 | */ |
| 553 | if (PageReserved(page)) |
| 554 | continue; |
| 555 | if (!queue_pages_required(page, qp)) |
| 556 | continue; |
| 557 | if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) { |
| 558 | /* MPOL_MF_STRICT must be specified if we get here */ |
| 559 | if (!vma_migratable(vma)) { |
| 560 | has_unmovable = true; |
| 561 | break; |
| 562 | } |
| 563 | |
| 564 | /* |
| 565 | * Do not abort immediately since there may be |
| 566 | * temporary off LRU pages in the range. Still |
| 567 | * need migrate other LRU pages. |
| 568 | */ |
| 569 | if (migrate_page_add(page, qp->pagelist, flags)) |
| 570 | has_unmovable = true; |
| 571 | } else |
| 572 | break; |
| 573 | } |
| 574 | pte_unmap_unlock(pte - 1, ptl); |
| 575 | cond_resched(); |
| 576 | |
| 577 | if (has_unmovable) |
| 578 | return 1; |
| 579 | |
| 580 | return addr != end ? -EIO : 0; |
| 581 | } |
| 582 | |
| 583 | static int queue_pages_hugetlb(pte_t *pte, unsigned long hmask, |
| 584 | unsigned long addr, unsigned long end, |
| 585 | struct mm_walk *walk) |
| 586 | { |
| 587 | int ret = 0; |
| 588 | #ifdef CONFIG_HUGETLB_PAGE |
| 589 | struct queue_pages *qp = walk->private; |
| 590 | unsigned long flags = (qp->flags & MPOL_MF_VALID); |
| 591 | struct page *page; |
| 592 | spinlock_t *ptl; |
| 593 | pte_t entry; |
| 594 | |
| 595 | ptl = huge_pte_lock(hstate_vma(walk->vma), walk->mm, pte); |
| 596 | entry = huge_ptep_get(pte); |
| 597 | if (!pte_present(entry)) |
| 598 | goto unlock; |
| 599 | page = pte_page(entry); |
| 600 | if (!queue_pages_required(page, qp)) |
| 601 | goto unlock; |
| 602 | |
| 603 | if (flags == MPOL_MF_STRICT) { |
| 604 | /* |
| 605 | * STRICT alone means only detecting misplaced page and no |
| 606 | * need to further check other vma. |
| 607 | */ |
| 608 | ret = -EIO; |
| 609 | goto unlock; |
| 610 | } |
| 611 | |
| 612 | if (!vma_migratable(walk->vma)) { |
| 613 | /* |
| 614 | * Must be STRICT with MOVE*, otherwise .test_walk() have |
| 615 | * stopped walking current vma. |
| 616 | * Detecting misplaced page but allow migrating pages which |
| 617 | * have been queued. |
| 618 | */ |
| 619 | ret = 1; |
| 620 | goto unlock; |
| 621 | } |
| 622 | |
| 623 | /* With MPOL_MF_MOVE, we migrate only unshared hugepage. */ |
| 624 | if (flags & (MPOL_MF_MOVE_ALL) || |
| 625 | (flags & MPOL_MF_MOVE && page_mapcount(page) == 1)) { |
| 626 | if (!isolate_huge_page(page, qp->pagelist) && |
| 627 | (flags & MPOL_MF_STRICT)) |
| 628 | /* |
| 629 | * Failed to isolate page but allow migrating pages |
| 630 | * which have been queued. |
| 631 | */ |
| 632 | ret = 1; |
| 633 | } |
| 634 | unlock: |
| 635 | spin_unlock(ptl); |
| 636 | #else |
| 637 | BUG(); |
| 638 | #endif |
| 639 | return ret; |
| 640 | } |
| 641 | |
| 642 | #ifdef CONFIG_NUMA_BALANCING |
| 643 | /* |
| 644 | * This is used to mark a range of virtual addresses to be inaccessible. |
| 645 | * These are later cleared by a NUMA hinting fault. Depending on these |
| 646 | * faults, pages may be migrated for better NUMA placement. |
| 647 | * |
| 648 | * This is assuming that NUMA faults are handled using PROT_NONE. If |
| 649 | * an architecture makes a different choice, it will need further |
| 650 | * changes to the core. |
| 651 | */ |
| 652 | unsigned long change_prot_numa(struct vm_area_struct *vma, |
| 653 | unsigned long addr, unsigned long end) |
| 654 | { |
| 655 | int nr_updated; |
| 656 | |
| 657 | nr_updated = change_protection(vma, addr, end, PAGE_NONE, MM_CP_PROT_NUMA); |
| 658 | if (nr_updated) |
| 659 | count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated); |
| 660 | |
| 661 | return nr_updated; |
| 662 | } |
| 663 | #else |
| 664 | static unsigned long change_prot_numa(struct vm_area_struct *vma, |
| 665 | unsigned long addr, unsigned long end) |
| 666 | { |
| 667 | return 0; |
| 668 | } |
| 669 | #endif /* CONFIG_NUMA_BALANCING */ |
| 670 | |
| 671 | static int queue_pages_test_walk(unsigned long start, unsigned long end, |
| 672 | struct mm_walk *walk) |
| 673 | { |
| 674 | struct vm_area_struct *vma = walk->vma; |
| 675 | struct queue_pages *qp = walk->private; |
| 676 | unsigned long endvma = vma->vm_end; |
| 677 | unsigned long flags = qp->flags; |
| 678 | |
| 679 | /* range check first */ |
| 680 | VM_BUG_ON_VMA((vma->vm_start > start) || (vma->vm_end < end), vma); |
| 681 | |
| 682 | if (!qp->first) { |
| 683 | qp->first = vma; |
| 684 | if (!(flags & MPOL_MF_DISCONTIG_OK) && |
| 685 | (qp->start < vma->vm_start)) |
| 686 | /* hole at head side of range */ |
| 687 | return -EFAULT; |
| 688 | } |
| 689 | if (!(flags & MPOL_MF_DISCONTIG_OK) && |
| 690 | ((vma->vm_end < qp->end) && |
| 691 | (!vma->vm_next || vma->vm_end < vma->vm_next->vm_start))) |
| 692 | /* hole at middle or tail of range */ |
| 693 | return -EFAULT; |
| 694 | |
| 695 | /* |
| 696 | * Need check MPOL_MF_STRICT to return -EIO if possible |
| 697 | * regardless of vma_migratable |
| 698 | */ |
| 699 | if (!vma_migratable(vma) && |
| 700 | !(flags & MPOL_MF_STRICT)) |
| 701 | return 1; |
| 702 | |
| 703 | if (endvma > end) |
| 704 | endvma = end; |
| 705 | |
| 706 | if (flags & MPOL_MF_LAZY) { |
| 707 | /* Similar to task_numa_work, skip inaccessible VMAs */ |
| 708 | if (!is_vm_hugetlb_page(vma) && vma_is_accessible(vma) && |
| 709 | !(vma->vm_flags & VM_MIXEDMAP)) |
| 710 | change_prot_numa(vma, start, endvma); |
| 711 | return 1; |
| 712 | } |
| 713 | |
| 714 | /* queue pages from current vma */ |
| 715 | if (flags & MPOL_MF_VALID) |
| 716 | return 0; |
| 717 | return 1; |
| 718 | } |
| 719 | |
| 720 | static const struct mm_walk_ops queue_pages_walk_ops = { |
| 721 | .hugetlb_entry = queue_pages_hugetlb, |
| 722 | .pmd_entry = queue_pages_pte_range, |
| 723 | .test_walk = queue_pages_test_walk, |
| 724 | }; |
| 725 | |
| 726 | /* |
| 727 | * Walk through page tables and collect pages to be migrated. |
| 728 | * |
| 729 | * If pages found in a given range are on a set of nodes (determined by |
| 730 | * @nodes and @flags,) it's isolated and queued to the pagelist which is |
| 731 | * passed via @private. |
| 732 | * |
| 733 | * queue_pages_range() has three possible return values: |
| 734 | * 1 - there is unmovable page, but MPOL_MF_MOVE* & MPOL_MF_STRICT were |
| 735 | * specified. |
| 736 | * 0 - queue pages successfully or no misplaced page. |
| 737 | * errno - i.e. misplaced pages with MPOL_MF_STRICT specified (-EIO) or |
| 738 | * memory range specified by nodemask and maxnode points outside |
| 739 | * your accessible address space (-EFAULT) |
| 740 | */ |
| 741 | static int |
| 742 | queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end, |
| 743 | nodemask_t *nodes, unsigned long flags, |
| 744 | struct list_head *pagelist) |
| 745 | { |
| 746 | int err; |
| 747 | struct queue_pages qp = { |
| 748 | .pagelist = pagelist, |
| 749 | .flags = flags, |
| 750 | .nmask = nodes, |
| 751 | .start = start, |
| 752 | .end = end, |
| 753 | .first = NULL, |
| 754 | }; |
| 755 | |
| 756 | err = walk_page_range(mm, start, end, &queue_pages_walk_ops, &qp); |
| 757 | |
| 758 | if (!qp.first) |
| 759 | /* whole range in hole */ |
| 760 | err = -EFAULT; |
| 761 | |
| 762 | return err; |
| 763 | } |
| 764 | |
| 765 | /* |
| 766 | * Apply policy to a single VMA |
| 767 | * This must be called with the mmap_sem held for writing. |
| 768 | */ |
| 769 | static int vma_replace_policy(struct vm_area_struct *vma, |
| 770 | struct mempolicy *pol) |
| 771 | { |
| 772 | int err; |
| 773 | struct mempolicy *old; |
| 774 | struct mempolicy *new; |
| 775 | |
| 776 | pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n", |
| 777 | vma->vm_start, vma->vm_end, vma->vm_pgoff, |
| 778 | vma->vm_ops, vma->vm_file, |
| 779 | vma->vm_ops ? vma->vm_ops->set_policy : NULL); |
| 780 | |
| 781 | new = mpol_dup(pol); |
| 782 | if (IS_ERR(new)) |
| 783 | return PTR_ERR(new); |
| 784 | |
| 785 | if (vma->vm_ops && vma->vm_ops->set_policy) { |
| 786 | err = vma->vm_ops->set_policy(vma, new); |
| 787 | if (err) |
| 788 | goto err_out; |
| 789 | } |
| 790 | |
| 791 | old = vma->vm_policy; |
| 792 | vma->vm_policy = new; /* protected by mmap_sem */ |
| 793 | mpol_put(old); |
| 794 | |
| 795 | return 0; |
| 796 | err_out: |
| 797 | mpol_put(new); |
| 798 | return err; |
| 799 | } |
| 800 | |
| 801 | /* Step 2: apply policy to a range and do splits. */ |
| 802 | static int mbind_range(struct mm_struct *mm, unsigned long start, |
| 803 | unsigned long end, struct mempolicy *new_pol) |
| 804 | { |
| 805 | struct vm_area_struct *next; |
| 806 | struct vm_area_struct *prev; |
| 807 | struct vm_area_struct *vma; |
| 808 | int err = 0; |
| 809 | pgoff_t pgoff; |
| 810 | unsigned long vmstart; |
| 811 | unsigned long vmend; |
| 812 | |
| 813 | vma = find_vma(mm, start); |
| 814 | VM_BUG_ON(!vma); |
| 815 | |
| 816 | prev = vma->vm_prev; |
| 817 | if (start > vma->vm_start) |
| 818 | prev = vma; |
| 819 | |
| 820 | for (; vma && vma->vm_start < end; prev = vma, vma = next) { |
| 821 | next = vma->vm_next; |
| 822 | vmstart = max(start, vma->vm_start); |
| 823 | vmend = min(end, vma->vm_end); |
| 824 | |
| 825 | if (mpol_equal(vma_policy(vma), new_pol)) |
| 826 | continue; |
| 827 | |
| 828 | pgoff = vma->vm_pgoff + |
| 829 | ((vmstart - vma->vm_start) >> PAGE_SHIFT); |
| 830 | prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags, |
| 831 | vma->anon_vma, vma->vm_file, pgoff, |
| 832 | new_pol, vma->vm_userfaultfd_ctx); |
| 833 | if (prev) { |
| 834 | vma = prev; |
| 835 | next = vma->vm_next; |
| 836 | if (mpol_equal(vma_policy(vma), new_pol)) |
| 837 | continue; |
| 838 | /* vma_merge() joined vma && vma->next, case 8 */ |
| 839 | goto replace; |
| 840 | } |
| 841 | if (vma->vm_start != vmstart) { |
| 842 | err = split_vma(vma->vm_mm, vma, vmstart, 1); |
| 843 | if (err) |
| 844 | goto out; |
| 845 | } |
| 846 | if (vma->vm_end != vmend) { |
| 847 | err = split_vma(vma->vm_mm, vma, vmend, 0); |
| 848 | if (err) |
| 849 | goto out; |
| 850 | } |
| 851 | replace: |
| 852 | err = vma_replace_policy(vma, new_pol); |
| 853 | if (err) |
| 854 | goto out; |
| 855 | } |
| 856 | |
| 857 | out: |
| 858 | return err; |
| 859 | } |
| 860 | |
| 861 | /* Set the process memory policy */ |
| 862 | static long do_set_mempolicy(unsigned short mode, unsigned short flags, |
| 863 | nodemask_t *nodes) |
| 864 | { |
| 865 | struct mempolicy *new, *old; |
| 866 | NODEMASK_SCRATCH(scratch); |
| 867 | int ret; |
| 868 | |
| 869 | if (!scratch) |
| 870 | return -ENOMEM; |
| 871 | |
| 872 | new = mpol_new(mode, flags, nodes); |
| 873 | if (IS_ERR(new)) { |
| 874 | ret = PTR_ERR(new); |
| 875 | goto out; |
| 876 | } |
| 877 | |
| 878 | task_lock(current); |
| 879 | ret = mpol_set_nodemask(new, nodes, scratch); |
| 880 | if (ret) { |
| 881 | task_unlock(current); |
| 882 | mpol_put(new); |
| 883 | goto out; |
| 884 | } |
| 885 | old = current->mempolicy; |
| 886 | current->mempolicy = new; |
| 887 | if (new && new->mode == MPOL_INTERLEAVE) |
| 888 | current->il_prev = MAX_NUMNODES-1; |
| 889 | task_unlock(current); |
| 890 | mpol_put(old); |
| 891 | ret = 0; |
| 892 | out: |
| 893 | NODEMASK_SCRATCH_FREE(scratch); |
| 894 | return ret; |
| 895 | } |
| 896 | |
| 897 | /* |
| 898 | * Return nodemask for policy for get_mempolicy() query |
| 899 | * |
| 900 | * Called with task's alloc_lock held |
| 901 | */ |
| 902 | static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes) |
| 903 | { |
| 904 | nodes_clear(*nodes); |
| 905 | if (p == &default_policy) |
| 906 | return; |
| 907 | |
| 908 | switch (p->mode) { |
| 909 | case MPOL_BIND: |
| 910 | case MPOL_INTERLEAVE: |
| 911 | *nodes = p->v.nodes; |
| 912 | break; |
| 913 | case MPOL_PREFERRED: |
| 914 | if (!(p->flags & MPOL_F_LOCAL)) |
| 915 | node_set(p->v.preferred_node, *nodes); |
| 916 | /* else return empty node mask for local allocation */ |
| 917 | break; |
| 918 | default: |
| 919 | BUG(); |
| 920 | } |
| 921 | } |
| 922 | |
| 923 | static int lookup_node(struct mm_struct *mm, unsigned long addr) |
| 924 | { |
| 925 | struct page *p = NULL; |
| 926 | int err; |
| 927 | |
| 928 | int locked = 1; |
| 929 | err = get_user_pages_locked(addr & PAGE_MASK, 1, 0, &p, &locked); |
| 930 | if (err == 0) { |
| 931 | /* E.g. GUP interrupted by fatal signal */ |
| 932 | err = -EFAULT; |
| 933 | } else if (err > 0) { |
| 934 | err = page_to_nid(p); |
| 935 | put_page(p); |
| 936 | } |
| 937 | if (locked) |
| 938 | up_read(&mm->mmap_sem); |
| 939 | return err; |
| 940 | } |
| 941 | |
| 942 | /* Retrieve NUMA policy */ |
| 943 | static long do_get_mempolicy(int *policy, nodemask_t *nmask, |
| 944 | unsigned long addr, unsigned long flags) |
| 945 | { |
| 946 | int err; |
| 947 | struct mm_struct *mm = current->mm; |
| 948 | struct vm_area_struct *vma = NULL; |
| 949 | struct mempolicy *pol = current->mempolicy, *pol_refcount = NULL; |
| 950 | |
| 951 | if (flags & |
| 952 | ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED)) |
| 953 | return -EINVAL; |
| 954 | |
| 955 | if (flags & MPOL_F_MEMS_ALLOWED) { |
| 956 | if (flags & (MPOL_F_NODE|MPOL_F_ADDR)) |
| 957 | return -EINVAL; |
| 958 | *policy = 0; /* just so it's initialized */ |
| 959 | task_lock(current); |
| 960 | *nmask = cpuset_current_mems_allowed; |
| 961 | task_unlock(current); |
| 962 | return 0; |
| 963 | } |
| 964 | |
| 965 | if (flags & MPOL_F_ADDR) { |
| 966 | /* |
| 967 | * Do NOT fall back to task policy if the |
| 968 | * vma/shared policy at addr is NULL. We |
| 969 | * want to return MPOL_DEFAULT in this case. |
| 970 | */ |
| 971 | down_read(&mm->mmap_sem); |
| 972 | vma = find_vma_intersection(mm, addr, addr+1); |
| 973 | if (!vma) { |
| 974 | up_read(&mm->mmap_sem); |
| 975 | return -EFAULT; |
| 976 | } |
| 977 | if (vma->vm_ops && vma->vm_ops->get_policy) |
| 978 | pol = vma->vm_ops->get_policy(vma, addr); |
| 979 | else |
| 980 | pol = vma->vm_policy; |
| 981 | } else if (addr) |
| 982 | return -EINVAL; |
| 983 | |
| 984 | if (!pol) |
| 985 | pol = &default_policy; /* indicates default behavior */ |
| 986 | |
| 987 | if (flags & MPOL_F_NODE) { |
| 988 | if (flags & MPOL_F_ADDR) { |
| 989 | /* |
| 990 | * Take a refcount on the mpol, lookup_node() |
| 991 | * wil drop the mmap_sem, so after calling |
| 992 | * lookup_node() only "pol" remains valid, "vma" |
| 993 | * is stale. |
| 994 | */ |
| 995 | pol_refcount = pol; |
| 996 | vma = NULL; |
| 997 | mpol_get(pol); |
| 998 | err = lookup_node(mm, addr); |
| 999 | if (err < 0) |
| 1000 | goto out; |
| 1001 | *policy = err; |
| 1002 | } else if (pol == current->mempolicy && |
| 1003 | pol->mode == MPOL_INTERLEAVE) { |
| 1004 | *policy = next_node_in(current->il_prev, pol->v.nodes); |
| 1005 | } else { |
| 1006 | err = -EINVAL; |
| 1007 | goto out; |
| 1008 | } |
| 1009 | } else { |
| 1010 | *policy = pol == &default_policy ? MPOL_DEFAULT : |
| 1011 | pol->mode; |
| 1012 | /* |
| 1013 | * Internal mempolicy flags must be masked off before exposing |
| 1014 | * the policy to userspace. |
| 1015 | */ |
| 1016 | *policy |= (pol->flags & MPOL_MODE_FLAGS); |
| 1017 | } |
| 1018 | |
| 1019 | err = 0; |
| 1020 | if (nmask) { |
| 1021 | if (mpol_store_user_nodemask(pol)) { |
| 1022 | *nmask = pol->w.user_nodemask; |
| 1023 | } else { |
| 1024 | task_lock(current); |
| 1025 | get_policy_nodemask(pol, nmask); |
| 1026 | task_unlock(current); |
| 1027 | } |
| 1028 | } |
| 1029 | |
| 1030 | out: |
| 1031 | mpol_cond_put(pol); |
| 1032 | if (vma) |
| 1033 | up_read(&mm->mmap_sem); |
| 1034 | if (pol_refcount) |
| 1035 | mpol_put(pol_refcount); |
| 1036 | return err; |
| 1037 | } |
| 1038 | |
| 1039 | #ifdef CONFIG_MIGRATION |
| 1040 | /* |
| 1041 | * page migration, thp tail pages can be passed. |
| 1042 | */ |
| 1043 | static int migrate_page_add(struct page *page, struct list_head *pagelist, |
| 1044 | unsigned long flags) |
| 1045 | { |
| 1046 | struct page *head = compound_head(page); |
| 1047 | /* |
| 1048 | * Avoid migrating a page that is shared with others. |
| 1049 | */ |
| 1050 | if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(head) == 1) { |
| 1051 | if (!isolate_lru_page(head)) { |
| 1052 | list_add_tail(&head->lru, pagelist); |
| 1053 | mod_node_page_state(page_pgdat(head), |
| 1054 | NR_ISOLATED_ANON + page_is_file_lru(head), |
| 1055 | hpage_nr_pages(head)); |
| 1056 | } else if (flags & MPOL_MF_STRICT) { |
| 1057 | /* |
| 1058 | * Non-movable page may reach here. And, there may be |
| 1059 | * temporary off LRU pages or non-LRU movable pages. |
| 1060 | * Treat them as unmovable pages since they can't be |
| 1061 | * isolated, so they can't be moved at the moment. It |
| 1062 | * should return -EIO for this case too. |
| 1063 | */ |
| 1064 | return -EIO; |
| 1065 | } |
| 1066 | } |
| 1067 | |
| 1068 | return 0; |
| 1069 | } |
| 1070 | |
| 1071 | /* page allocation callback for NUMA node migration */ |
| 1072 | struct page *alloc_new_node_page(struct page *page, unsigned long node) |
| 1073 | { |
| 1074 | if (PageHuge(page)) |
| 1075 | return alloc_huge_page_node(page_hstate(compound_head(page)), |
| 1076 | node); |
| 1077 | else if (PageTransHuge(page)) { |
| 1078 | struct page *thp; |
| 1079 | |
| 1080 | thp = alloc_pages_node(node, |
| 1081 | (GFP_TRANSHUGE | __GFP_THISNODE), |
| 1082 | HPAGE_PMD_ORDER); |
| 1083 | if (!thp) |
| 1084 | return NULL; |
| 1085 | prep_transhuge_page(thp); |
| 1086 | return thp; |
| 1087 | } else |
| 1088 | return __alloc_pages_node(node, GFP_HIGHUSER_MOVABLE | |
| 1089 | __GFP_THISNODE, 0); |
| 1090 | } |
| 1091 | |
| 1092 | /* |
| 1093 | * Migrate pages from one node to a target node. |
| 1094 | * Returns error or the number of pages not migrated. |
| 1095 | */ |
| 1096 | static int migrate_to_node(struct mm_struct *mm, int source, int dest, |
| 1097 | int flags) |
| 1098 | { |
| 1099 | nodemask_t nmask; |
| 1100 | LIST_HEAD(pagelist); |
| 1101 | int err = 0; |
| 1102 | |
| 1103 | nodes_clear(nmask); |
| 1104 | node_set(source, nmask); |
| 1105 | |
| 1106 | /* |
| 1107 | * This does not "check" the range but isolates all pages that |
| 1108 | * need migration. Between passing in the full user address |
| 1109 | * space range and MPOL_MF_DISCONTIG_OK, this call can not fail. |
| 1110 | */ |
| 1111 | VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))); |
| 1112 | queue_pages_range(mm, mm->mmap->vm_start, mm->task_size, &nmask, |
| 1113 | flags | MPOL_MF_DISCONTIG_OK, &pagelist); |
| 1114 | |
| 1115 | if (!list_empty(&pagelist)) { |
| 1116 | err = migrate_pages(&pagelist, alloc_new_node_page, NULL, dest, |
| 1117 | MIGRATE_SYNC, MR_SYSCALL); |
| 1118 | if (err) |
| 1119 | putback_movable_pages(&pagelist); |
| 1120 | } |
| 1121 | |
| 1122 | return err; |
| 1123 | } |
| 1124 | |
| 1125 | /* |
| 1126 | * Move pages between the two nodesets so as to preserve the physical |
| 1127 | * layout as much as possible. |
| 1128 | * |
| 1129 | * Returns the number of page that could not be moved. |
| 1130 | */ |
| 1131 | int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from, |
| 1132 | const nodemask_t *to, int flags) |
| 1133 | { |
| 1134 | int busy = 0; |
| 1135 | int err; |
| 1136 | nodemask_t tmp; |
| 1137 | |
| 1138 | err = migrate_prep(); |
| 1139 | if (err) |
| 1140 | return err; |
| 1141 | |
| 1142 | down_read(&mm->mmap_sem); |
| 1143 | |
| 1144 | /* |
| 1145 | * Find a 'source' bit set in 'tmp' whose corresponding 'dest' |
| 1146 | * bit in 'to' is not also set in 'tmp'. Clear the found 'source' |
| 1147 | * bit in 'tmp', and return that <source, dest> pair for migration. |
| 1148 | * The pair of nodemasks 'to' and 'from' define the map. |
| 1149 | * |
| 1150 | * If no pair of bits is found that way, fallback to picking some |
| 1151 | * pair of 'source' and 'dest' bits that are not the same. If the |
| 1152 | * 'source' and 'dest' bits are the same, this represents a node |
| 1153 | * that will be migrating to itself, so no pages need move. |
| 1154 | * |
| 1155 | * If no bits are left in 'tmp', or if all remaining bits left |
| 1156 | * in 'tmp' correspond to the same bit in 'to', return false |
| 1157 | * (nothing left to migrate). |
| 1158 | * |
| 1159 | * This lets us pick a pair of nodes to migrate between, such that |
| 1160 | * if possible the dest node is not already occupied by some other |
| 1161 | * source node, minimizing the risk of overloading the memory on a |
| 1162 | * node that would happen if we migrated incoming memory to a node |
| 1163 | * before migrating outgoing memory source that same node. |
| 1164 | * |
| 1165 | * A single scan of tmp is sufficient. As we go, we remember the |
| 1166 | * most recent <s, d> pair that moved (s != d). If we find a pair |
| 1167 | * that not only moved, but what's better, moved to an empty slot |
| 1168 | * (d is not set in tmp), then we break out then, with that pair. |
| 1169 | * Otherwise when we finish scanning from_tmp, we at least have the |
| 1170 | * most recent <s, d> pair that moved. If we get all the way through |
| 1171 | * the scan of tmp without finding any node that moved, much less |
| 1172 | * moved to an empty node, then there is nothing left worth migrating. |
| 1173 | */ |
| 1174 | |
| 1175 | tmp = *from; |
| 1176 | while (!nodes_empty(tmp)) { |
| 1177 | int s,d; |
| 1178 | int source = NUMA_NO_NODE; |
| 1179 | int dest = 0; |
| 1180 | |
| 1181 | for_each_node_mask(s, tmp) { |
| 1182 | |
| 1183 | /* |
| 1184 | * do_migrate_pages() tries to maintain the relative |
| 1185 | * node relationship of the pages established between |
| 1186 | * threads and memory areas. |
| 1187 | * |
| 1188 | * However if the number of source nodes is not equal to |
| 1189 | * the number of destination nodes we can not preserve |
| 1190 | * this node relative relationship. In that case, skip |
| 1191 | * copying memory from a node that is in the destination |
| 1192 | * mask. |
| 1193 | * |
| 1194 | * Example: [2,3,4] -> [3,4,5] moves everything. |
| 1195 | * [0-7] - > [3,4,5] moves only 0,1,2,6,7. |
| 1196 | */ |
| 1197 | |
| 1198 | if ((nodes_weight(*from) != nodes_weight(*to)) && |
| 1199 | (node_isset(s, *to))) |
| 1200 | continue; |
| 1201 | |
| 1202 | d = node_remap(s, *from, *to); |
| 1203 | if (s == d) |
| 1204 | continue; |
| 1205 | |
| 1206 | source = s; /* Node moved. Memorize */ |
| 1207 | dest = d; |
| 1208 | |
| 1209 | /* dest not in remaining from nodes? */ |
| 1210 | if (!node_isset(dest, tmp)) |
| 1211 | break; |
| 1212 | } |
| 1213 | if (source == NUMA_NO_NODE) |
| 1214 | break; |
| 1215 | |
| 1216 | node_clear(source, tmp); |
| 1217 | err = migrate_to_node(mm, source, dest, flags); |
| 1218 | if (err > 0) |
| 1219 | busy += err; |
| 1220 | if (err < 0) |
| 1221 | break; |
| 1222 | } |
| 1223 | up_read(&mm->mmap_sem); |
| 1224 | if (err < 0) |
| 1225 | return err; |
| 1226 | return busy; |
| 1227 | |
| 1228 | } |
| 1229 | |
| 1230 | /* |
| 1231 | * Allocate a new page for page migration based on vma policy. |
| 1232 | * Start by assuming the page is mapped by the same vma as contains @start. |
| 1233 | * Search forward from there, if not. N.B., this assumes that the |
| 1234 | * list of pages handed to migrate_pages()--which is how we get here-- |
| 1235 | * is in virtual address order. |
| 1236 | */ |
| 1237 | static struct page *new_page(struct page *page, unsigned long start) |
| 1238 | { |
| 1239 | struct vm_area_struct *vma; |
| 1240 | unsigned long uninitialized_var(address); |
| 1241 | |
| 1242 | vma = find_vma(current->mm, start); |
| 1243 | while (vma) { |
| 1244 | address = page_address_in_vma(page, vma); |
| 1245 | if (address != -EFAULT) |
| 1246 | break; |
| 1247 | vma = vma->vm_next; |
| 1248 | } |
| 1249 | |
| 1250 | if (PageHuge(page)) { |
| 1251 | return alloc_huge_page_vma(page_hstate(compound_head(page)), |
| 1252 | vma, address); |
| 1253 | } else if (PageTransHuge(page)) { |
| 1254 | struct page *thp; |
| 1255 | |
| 1256 | thp = alloc_hugepage_vma(GFP_TRANSHUGE, vma, address, |
| 1257 | HPAGE_PMD_ORDER); |
| 1258 | if (!thp) |
| 1259 | return NULL; |
| 1260 | prep_transhuge_page(thp); |
| 1261 | return thp; |
| 1262 | } |
| 1263 | /* |
| 1264 | * if !vma, alloc_page_vma() will use task or system default policy |
| 1265 | */ |
| 1266 | return alloc_page_vma(GFP_HIGHUSER_MOVABLE | __GFP_RETRY_MAYFAIL, |
| 1267 | vma, address); |
| 1268 | } |
| 1269 | #else |
| 1270 | |
| 1271 | static int migrate_page_add(struct page *page, struct list_head *pagelist, |
| 1272 | unsigned long flags) |
| 1273 | { |
| 1274 | return -EIO; |
| 1275 | } |
| 1276 | |
| 1277 | int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from, |
| 1278 | const nodemask_t *to, int flags) |
| 1279 | { |
| 1280 | return -ENOSYS; |
| 1281 | } |
| 1282 | |
| 1283 | static struct page *new_page(struct page *page, unsigned long start) |
| 1284 | { |
| 1285 | return NULL; |
| 1286 | } |
| 1287 | #endif |
| 1288 | |
| 1289 | static long do_mbind(unsigned long start, unsigned long len, |
| 1290 | unsigned short mode, unsigned short mode_flags, |
| 1291 | nodemask_t *nmask, unsigned long flags) |
| 1292 | { |
| 1293 | struct mm_struct *mm = current->mm; |
| 1294 | struct mempolicy *new; |
| 1295 | unsigned long end; |
| 1296 | int err; |
| 1297 | int ret; |
| 1298 | LIST_HEAD(pagelist); |
| 1299 | |
| 1300 | if (flags & ~(unsigned long)MPOL_MF_VALID) |
| 1301 | return -EINVAL; |
| 1302 | if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE)) |
| 1303 | return -EPERM; |
| 1304 | |
| 1305 | if (start & ~PAGE_MASK) |
| 1306 | return -EINVAL; |
| 1307 | |
| 1308 | if (mode == MPOL_DEFAULT) |
| 1309 | flags &= ~MPOL_MF_STRICT; |
| 1310 | |
| 1311 | len = (len + PAGE_SIZE - 1) & PAGE_MASK; |
| 1312 | end = start + len; |
| 1313 | |
| 1314 | if (end < start) |
| 1315 | return -EINVAL; |
| 1316 | if (end == start) |
| 1317 | return 0; |
| 1318 | |
| 1319 | new = mpol_new(mode, mode_flags, nmask); |
| 1320 | if (IS_ERR(new)) |
| 1321 | return PTR_ERR(new); |
| 1322 | |
| 1323 | if (flags & MPOL_MF_LAZY) |
| 1324 | new->flags |= MPOL_F_MOF; |
| 1325 | |
| 1326 | /* |
| 1327 | * If we are using the default policy then operation |
| 1328 | * on discontinuous address spaces is okay after all |
| 1329 | */ |
| 1330 | if (!new) |
| 1331 | flags |= MPOL_MF_DISCONTIG_OK; |
| 1332 | |
| 1333 | pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n", |
| 1334 | start, start + len, mode, mode_flags, |
| 1335 | nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE); |
| 1336 | |
| 1337 | if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) { |
| 1338 | |
| 1339 | err = migrate_prep(); |
| 1340 | if (err) |
| 1341 | goto mpol_out; |
| 1342 | } |
| 1343 | { |
| 1344 | NODEMASK_SCRATCH(scratch); |
| 1345 | if (scratch) { |
| 1346 | down_write(&mm->mmap_sem); |
| 1347 | task_lock(current); |
| 1348 | err = mpol_set_nodemask(new, nmask, scratch); |
| 1349 | task_unlock(current); |
| 1350 | if (err) |
| 1351 | up_write(&mm->mmap_sem); |
| 1352 | } else |
| 1353 | err = -ENOMEM; |
| 1354 | NODEMASK_SCRATCH_FREE(scratch); |
| 1355 | } |
| 1356 | if (err) |
| 1357 | goto mpol_out; |
| 1358 | |
| 1359 | ret = queue_pages_range(mm, start, end, nmask, |
| 1360 | flags | MPOL_MF_INVERT, &pagelist); |
| 1361 | |
| 1362 | if (ret < 0) { |
| 1363 | err = ret; |
| 1364 | goto up_out; |
| 1365 | } |
| 1366 | |
| 1367 | err = mbind_range(mm, start, end, new); |
| 1368 | |
| 1369 | if (!err) { |
| 1370 | int nr_failed = 0; |
| 1371 | |
| 1372 | if (!list_empty(&pagelist)) { |
| 1373 | WARN_ON_ONCE(flags & MPOL_MF_LAZY); |
| 1374 | nr_failed = migrate_pages(&pagelist, new_page, NULL, |
| 1375 | start, MIGRATE_SYNC, MR_MEMPOLICY_MBIND); |
| 1376 | if (nr_failed) |
| 1377 | putback_movable_pages(&pagelist); |
| 1378 | } |
| 1379 | |
| 1380 | if ((ret > 0) || (nr_failed && (flags & MPOL_MF_STRICT))) |
| 1381 | err = -EIO; |
| 1382 | } else { |
| 1383 | up_out: |
| 1384 | if (!list_empty(&pagelist)) |
| 1385 | putback_movable_pages(&pagelist); |
| 1386 | } |
| 1387 | |
| 1388 | up_write(&mm->mmap_sem); |
| 1389 | mpol_out: |
| 1390 | mpol_put(new); |
| 1391 | return err; |
| 1392 | } |
| 1393 | |
| 1394 | /* |
| 1395 | * User space interface with variable sized bitmaps for nodelists. |
| 1396 | */ |
| 1397 | |
| 1398 | /* Copy a node mask from user space. */ |
| 1399 | static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask, |
| 1400 | unsigned long maxnode) |
| 1401 | { |
| 1402 | unsigned long k; |
| 1403 | unsigned long t; |
| 1404 | unsigned long nlongs; |
| 1405 | unsigned long endmask; |
| 1406 | |
| 1407 | --maxnode; |
| 1408 | nodes_clear(*nodes); |
| 1409 | if (maxnode == 0 || !nmask) |
| 1410 | return 0; |
| 1411 | if (maxnode > PAGE_SIZE*BITS_PER_BYTE) |
| 1412 | return -EINVAL; |
| 1413 | |
| 1414 | nlongs = BITS_TO_LONGS(maxnode); |
| 1415 | if ((maxnode % BITS_PER_LONG) == 0) |
| 1416 | endmask = ~0UL; |
| 1417 | else |
| 1418 | endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1; |
| 1419 | |
| 1420 | /* |
| 1421 | * When the user specified more nodes than supported just check |
| 1422 | * if the non supported part is all zero. |
| 1423 | * |
| 1424 | * If maxnode have more longs than MAX_NUMNODES, check |
| 1425 | * the bits in that area first. And then go through to |
| 1426 | * check the rest bits which equal or bigger than MAX_NUMNODES. |
| 1427 | * Otherwise, just check bits [MAX_NUMNODES, maxnode). |
| 1428 | */ |
| 1429 | if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) { |
| 1430 | for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) { |
| 1431 | if (get_user(t, nmask + k)) |
| 1432 | return -EFAULT; |
| 1433 | if (k == nlongs - 1) { |
| 1434 | if (t & endmask) |
| 1435 | return -EINVAL; |
| 1436 | } else if (t) |
| 1437 | return -EINVAL; |
| 1438 | } |
| 1439 | nlongs = BITS_TO_LONGS(MAX_NUMNODES); |
| 1440 | endmask = ~0UL; |
| 1441 | } |
| 1442 | |
| 1443 | if (maxnode > MAX_NUMNODES && MAX_NUMNODES % BITS_PER_LONG != 0) { |
| 1444 | unsigned long valid_mask = endmask; |
| 1445 | |
| 1446 | valid_mask &= ~((1UL << (MAX_NUMNODES % BITS_PER_LONG)) - 1); |
| 1447 | if (get_user(t, nmask + nlongs - 1)) |
| 1448 | return -EFAULT; |
| 1449 | if (t & valid_mask) |
| 1450 | return -EINVAL; |
| 1451 | } |
| 1452 | |
| 1453 | if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long))) |
| 1454 | return -EFAULT; |
| 1455 | nodes_addr(*nodes)[nlongs-1] &= endmask; |
| 1456 | return 0; |
| 1457 | } |
| 1458 | |
| 1459 | /* Copy a kernel node mask to user space */ |
| 1460 | static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode, |
| 1461 | nodemask_t *nodes) |
| 1462 | { |
| 1463 | unsigned long copy = ALIGN(maxnode-1, 64) / 8; |
| 1464 | unsigned int nbytes = BITS_TO_LONGS(nr_node_ids) * sizeof(long); |
| 1465 | |
| 1466 | if (copy > nbytes) { |
| 1467 | if (copy > PAGE_SIZE) |
| 1468 | return -EINVAL; |
| 1469 | if (clear_user((char __user *)mask + nbytes, copy - nbytes)) |
| 1470 | return -EFAULT; |
| 1471 | copy = nbytes; |
| 1472 | } |
| 1473 | return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0; |
| 1474 | } |
| 1475 | |
| 1476 | static long kernel_mbind(unsigned long start, unsigned long len, |
| 1477 | unsigned long mode, const unsigned long __user *nmask, |
| 1478 | unsigned long maxnode, unsigned int flags) |
| 1479 | { |
| 1480 | nodemask_t nodes; |
| 1481 | int err; |
| 1482 | unsigned short mode_flags; |
| 1483 | |
| 1484 | start = untagged_addr(start); |
| 1485 | mode_flags = mode & MPOL_MODE_FLAGS; |
| 1486 | mode &= ~MPOL_MODE_FLAGS; |
| 1487 | if (mode >= MPOL_MAX) |
| 1488 | return -EINVAL; |
| 1489 | if ((mode_flags & MPOL_F_STATIC_NODES) && |
| 1490 | (mode_flags & MPOL_F_RELATIVE_NODES)) |
| 1491 | return -EINVAL; |
| 1492 | err = get_nodes(&nodes, nmask, maxnode); |
| 1493 | if (err) |
| 1494 | return err; |
| 1495 | return do_mbind(start, len, mode, mode_flags, &nodes, flags); |
| 1496 | } |
| 1497 | |
| 1498 | SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len, |
| 1499 | unsigned long, mode, const unsigned long __user *, nmask, |
| 1500 | unsigned long, maxnode, unsigned int, flags) |
| 1501 | { |
| 1502 | return kernel_mbind(start, len, mode, nmask, maxnode, flags); |
| 1503 | } |
| 1504 | |
| 1505 | /* Set the process memory policy */ |
| 1506 | static long kernel_set_mempolicy(int mode, const unsigned long __user *nmask, |
| 1507 | unsigned long maxnode) |
| 1508 | { |
| 1509 | int err; |
| 1510 | nodemask_t nodes; |
| 1511 | unsigned short flags; |
| 1512 | |
| 1513 | flags = mode & MPOL_MODE_FLAGS; |
| 1514 | mode &= ~MPOL_MODE_FLAGS; |
| 1515 | if ((unsigned int)mode >= MPOL_MAX) |
| 1516 | return -EINVAL; |
| 1517 | if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES)) |
| 1518 | return -EINVAL; |
| 1519 | err = get_nodes(&nodes, nmask, maxnode); |
| 1520 | if (err) |
| 1521 | return err; |
| 1522 | return do_set_mempolicy(mode, flags, &nodes); |
| 1523 | } |
| 1524 | |
| 1525 | SYSCALL_DEFINE3(set_mempolicy, int, mode, const unsigned long __user *, nmask, |
| 1526 | unsigned long, maxnode) |
| 1527 | { |
| 1528 | return kernel_set_mempolicy(mode, nmask, maxnode); |
| 1529 | } |
| 1530 | |
| 1531 | static int kernel_migrate_pages(pid_t pid, unsigned long maxnode, |
| 1532 | const unsigned long __user *old_nodes, |
| 1533 | const unsigned long __user *new_nodes) |
| 1534 | { |
| 1535 | struct mm_struct *mm = NULL; |
| 1536 | struct task_struct *task; |
| 1537 | nodemask_t task_nodes; |
| 1538 | int err; |
| 1539 | nodemask_t *old; |
| 1540 | nodemask_t *new; |
| 1541 | NODEMASK_SCRATCH(scratch); |
| 1542 | |
| 1543 | if (!scratch) |
| 1544 | return -ENOMEM; |
| 1545 | |
| 1546 | old = &scratch->mask1; |
| 1547 | new = &scratch->mask2; |
| 1548 | |
| 1549 | err = get_nodes(old, old_nodes, maxnode); |
| 1550 | if (err) |
| 1551 | goto out; |
| 1552 | |
| 1553 | err = get_nodes(new, new_nodes, maxnode); |
| 1554 | if (err) |
| 1555 | goto out; |
| 1556 | |
| 1557 | /* Find the mm_struct */ |
| 1558 | rcu_read_lock(); |
| 1559 | task = pid ? find_task_by_vpid(pid) : current; |
| 1560 | if (!task) { |
| 1561 | rcu_read_unlock(); |
| 1562 | err = -ESRCH; |
| 1563 | goto out; |
| 1564 | } |
| 1565 | get_task_struct(task); |
| 1566 | |
| 1567 | err = -EINVAL; |
| 1568 | |
| 1569 | /* |
| 1570 | * Check if this process has the right to modify the specified process. |
| 1571 | * Use the regular "ptrace_may_access()" checks. |
| 1572 | */ |
| 1573 | if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) { |
| 1574 | rcu_read_unlock(); |
| 1575 | err = -EPERM; |
| 1576 | goto out_put; |
| 1577 | } |
| 1578 | rcu_read_unlock(); |
| 1579 | |
| 1580 | task_nodes = cpuset_mems_allowed(task); |
| 1581 | /* Is the user allowed to access the target nodes? */ |
| 1582 | if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) { |
| 1583 | err = -EPERM; |
| 1584 | goto out_put; |
| 1585 | } |
| 1586 | |
| 1587 | task_nodes = cpuset_mems_allowed(current); |
| 1588 | nodes_and(*new, *new, task_nodes); |
| 1589 | if (nodes_empty(*new)) |
| 1590 | goto out_put; |
| 1591 | |
| 1592 | err = security_task_movememory(task); |
| 1593 | if (err) |
| 1594 | goto out_put; |
| 1595 | |
| 1596 | mm = get_task_mm(task); |
| 1597 | put_task_struct(task); |
| 1598 | |
| 1599 | if (!mm) { |
| 1600 | err = -EINVAL; |
| 1601 | goto out; |
| 1602 | } |
| 1603 | |
| 1604 | err = do_migrate_pages(mm, old, new, |
| 1605 | capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE); |
| 1606 | |
| 1607 | mmput(mm); |
| 1608 | out: |
| 1609 | NODEMASK_SCRATCH_FREE(scratch); |
| 1610 | |
| 1611 | return err; |
| 1612 | |
| 1613 | out_put: |
| 1614 | put_task_struct(task); |
| 1615 | goto out; |
| 1616 | |
| 1617 | } |
| 1618 | |
| 1619 | SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode, |
| 1620 | const unsigned long __user *, old_nodes, |
| 1621 | const unsigned long __user *, new_nodes) |
| 1622 | { |
| 1623 | return kernel_migrate_pages(pid, maxnode, old_nodes, new_nodes); |
| 1624 | } |
| 1625 | |
| 1626 | |
| 1627 | /* Retrieve NUMA policy */ |
| 1628 | static int kernel_get_mempolicy(int __user *policy, |
| 1629 | unsigned long __user *nmask, |
| 1630 | unsigned long maxnode, |
| 1631 | unsigned long addr, |
| 1632 | unsigned long flags) |
| 1633 | { |
| 1634 | int err; |
| 1635 | int uninitialized_var(pval); |
| 1636 | nodemask_t nodes; |
| 1637 | |
| 1638 | addr = untagged_addr(addr); |
| 1639 | |
| 1640 | if (nmask != NULL && maxnode < nr_node_ids) |
| 1641 | return -EINVAL; |
| 1642 | |
| 1643 | err = do_get_mempolicy(&pval, &nodes, addr, flags); |
| 1644 | |
| 1645 | if (err) |
| 1646 | return err; |
| 1647 | |
| 1648 | if (policy && put_user(pval, policy)) |
| 1649 | return -EFAULT; |
| 1650 | |
| 1651 | if (nmask) |
| 1652 | err = copy_nodes_to_user(nmask, maxnode, &nodes); |
| 1653 | |
| 1654 | return err; |
| 1655 | } |
| 1656 | |
| 1657 | SYSCALL_DEFINE5(get_mempolicy, int __user *, policy, |
| 1658 | unsigned long __user *, nmask, unsigned long, maxnode, |
| 1659 | unsigned long, addr, unsigned long, flags) |
| 1660 | { |
| 1661 | return kernel_get_mempolicy(policy, nmask, maxnode, addr, flags); |
| 1662 | } |
| 1663 | |
| 1664 | #ifdef CONFIG_COMPAT |
| 1665 | |
| 1666 | COMPAT_SYSCALL_DEFINE5(get_mempolicy, int __user *, policy, |
| 1667 | compat_ulong_t __user *, nmask, |
| 1668 | compat_ulong_t, maxnode, |
| 1669 | compat_ulong_t, addr, compat_ulong_t, flags) |
| 1670 | { |
| 1671 | long err; |
| 1672 | unsigned long __user *nm = NULL; |
| 1673 | unsigned long nr_bits, alloc_size; |
| 1674 | DECLARE_BITMAP(bm, MAX_NUMNODES); |
| 1675 | |
| 1676 | nr_bits = min_t(unsigned long, maxnode-1, nr_node_ids); |
| 1677 | alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8; |
| 1678 | |
| 1679 | if (nmask) |
| 1680 | nm = compat_alloc_user_space(alloc_size); |
| 1681 | |
| 1682 | err = kernel_get_mempolicy(policy, nm, nr_bits+1, addr, flags); |
| 1683 | |
| 1684 | if (!err && nmask) { |
| 1685 | unsigned long copy_size; |
| 1686 | copy_size = min_t(unsigned long, sizeof(bm), alloc_size); |
| 1687 | err = copy_from_user(bm, nm, copy_size); |
| 1688 | /* ensure entire bitmap is zeroed */ |
| 1689 | err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8); |
| 1690 | err |= compat_put_bitmap(nmask, bm, nr_bits); |
| 1691 | } |
| 1692 | |
| 1693 | return err; |
| 1694 | } |
| 1695 | |
| 1696 | COMPAT_SYSCALL_DEFINE3(set_mempolicy, int, mode, compat_ulong_t __user *, nmask, |
| 1697 | compat_ulong_t, maxnode) |
| 1698 | { |
| 1699 | unsigned long __user *nm = NULL; |
| 1700 | unsigned long nr_bits, alloc_size; |
| 1701 | DECLARE_BITMAP(bm, MAX_NUMNODES); |
| 1702 | |
| 1703 | nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES); |
| 1704 | alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8; |
| 1705 | |
| 1706 | if (nmask) { |
| 1707 | if (compat_get_bitmap(bm, nmask, nr_bits)) |
| 1708 | return -EFAULT; |
| 1709 | nm = compat_alloc_user_space(alloc_size); |
| 1710 | if (copy_to_user(nm, bm, alloc_size)) |
| 1711 | return -EFAULT; |
| 1712 | } |
| 1713 | |
| 1714 | return kernel_set_mempolicy(mode, nm, nr_bits+1); |
| 1715 | } |
| 1716 | |
| 1717 | COMPAT_SYSCALL_DEFINE6(mbind, compat_ulong_t, start, compat_ulong_t, len, |
| 1718 | compat_ulong_t, mode, compat_ulong_t __user *, nmask, |
| 1719 | compat_ulong_t, maxnode, compat_ulong_t, flags) |
| 1720 | { |
| 1721 | unsigned long __user *nm = NULL; |
| 1722 | unsigned long nr_bits, alloc_size; |
| 1723 | nodemask_t bm; |
| 1724 | |
| 1725 | nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES); |
| 1726 | alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8; |
| 1727 | |
| 1728 | if (nmask) { |
| 1729 | if (compat_get_bitmap(nodes_addr(bm), nmask, nr_bits)) |
| 1730 | return -EFAULT; |
| 1731 | nm = compat_alloc_user_space(alloc_size); |
| 1732 | if (copy_to_user(nm, nodes_addr(bm), alloc_size)) |
| 1733 | return -EFAULT; |
| 1734 | } |
| 1735 | |
| 1736 | return kernel_mbind(start, len, mode, nm, nr_bits+1, flags); |
| 1737 | } |
| 1738 | |
| 1739 | COMPAT_SYSCALL_DEFINE4(migrate_pages, compat_pid_t, pid, |
| 1740 | compat_ulong_t, maxnode, |
| 1741 | const compat_ulong_t __user *, old_nodes, |
| 1742 | const compat_ulong_t __user *, new_nodes) |
| 1743 | { |
| 1744 | unsigned long __user *old = NULL; |
| 1745 | unsigned long __user *new = NULL; |
| 1746 | nodemask_t tmp_mask; |
| 1747 | unsigned long nr_bits; |
| 1748 | unsigned long size; |
| 1749 | |
| 1750 | nr_bits = min_t(unsigned long, maxnode - 1, MAX_NUMNODES); |
| 1751 | size = ALIGN(nr_bits, BITS_PER_LONG) / 8; |
| 1752 | if (old_nodes) { |
| 1753 | if (compat_get_bitmap(nodes_addr(tmp_mask), old_nodes, nr_bits)) |
| 1754 | return -EFAULT; |
| 1755 | old = compat_alloc_user_space(new_nodes ? size * 2 : size); |
| 1756 | if (new_nodes) |
| 1757 | new = old + size / sizeof(unsigned long); |
| 1758 | if (copy_to_user(old, nodes_addr(tmp_mask), size)) |
| 1759 | return -EFAULT; |
| 1760 | } |
| 1761 | if (new_nodes) { |
| 1762 | if (compat_get_bitmap(nodes_addr(tmp_mask), new_nodes, nr_bits)) |
| 1763 | return -EFAULT; |
| 1764 | if (new == NULL) |
| 1765 | new = compat_alloc_user_space(size); |
| 1766 | if (copy_to_user(new, nodes_addr(tmp_mask), size)) |
| 1767 | return -EFAULT; |
| 1768 | } |
| 1769 | return kernel_migrate_pages(pid, nr_bits + 1, old, new); |
| 1770 | } |
| 1771 | |
| 1772 | #endif /* CONFIG_COMPAT */ |
| 1773 | |
| 1774 | bool vma_migratable(struct vm_area_struct *vma) |
| 1775 | { |
| 1776 | if (vma->vm_flags & (VM_IO | VM_PFNMAP)) |
| 1777 | return false; |
| 1778 | |
| 1779 | /* |
| 1780 | * DAX device mappings require predictable access latency, so avoid |
| 1781 | * incurring periodic faults. |
| 1782 | */ |
| 1783 | if (vma_is_dax(vma)) |
| 1784 | return false; |
| 1785 | |
| 1786 | if (is_vm_hugetlb_page(vma) && |
| 1787 | !hugepage_migration_supported(hstate_vma(vma))) |
| 1788 | return false; |
| 1789 | |
| 1790 | /* |
| 1791 | * Migration allocates pages in the highest zone. If we cannot |
| 1792 | * do so then migration (at least from node to node) is not |
| 1793 | * possible. |
| 1794 | */ |
| 1795 | if (vma->vm_file && |
| 1796 | gfp_zone(mapping_gfp_mask(vma->vm_file->f_mapping)) |
| 1797 | < policy_zone) |
| 1798 | return false; |
| 1799 | return true; |
| 1800 | } |
| 1801 | |
| 1802 | struct mempolicy *__get_vma_policy(struct vm_area_struct *vma, |
| 1803 | unsigned long addr) |
| 1804 | { |
| 1805 | struct mempolicy *pol = NULL; |
| 1806 | |
| 1807 | if (vma) { |
| 1808 | if (vma->vm_ops && vma->vm_ops->get_policy) { |
| 1809 | pol = vma->vm_ops->get_policy(vma, addr); |
| 1810 | } else if (vma->vm_policy) { |
| 1811 | pol = vma->vm_policy; |
| 1812 | |
| 1813 | /* |
| 1814 | * shmem_alloc_page() passes MPOL_F_SHARED policy with |
| 1815 | * a pseudo vma whose vma->vm_ops=NULL. Take a reference |
| 1816 | * count on these policies which will be dropped by |
| 1817 | * mpol_cond_put() later |
| 1818 | */ |
| 1819 | if (mpol_needs_cond_ref(pol)) |
| 1820 | mpol_get(pol); |
| 1821 | } |
| 1822 | } |
| 1823 | |
| 1824 | return pol; |
| 1825 | } |
| 1826 | |
| 1827 | /* |
| 1828 | * get_vma_policy(@vma, @addr) |
| 1829 | * @vma: virtual memory area whose policy is sought |
| 1830 | * @addr: address in @vma for shared policy lookup |
| 1831 | * |
| 1832 | * Returns effective policy for a VMA at specified address. |
| 1833 | * Falls back to current->mempolicy or system default policy, as necessary. |
| 1834 | * Shared policies [those marked as MPOL_F_SHARED] require an extra reference |
| 1835 | * count--added by the get_policy() vm_op, as appropriate--to protect against |
| 1836 | * freeing by another task. It is the caller's responsibility to free the |
| 1837 | * extra reference for shared policies. |
| 1838 | */ |
| 1839 | static struct mempolicy *get_vma_policy(struct vm_area_struct *vma, |
| 1840 | unsigned long addr) |
| 1841 | { |
| 1842 | struct mempolicy *pol = __get_vma_policy(vma, addr); |
| 1843 | |
| 1844 | if (!pol) |
| 1845 | pol = get_task_policy(current); |
| 1846 | |
| 1847 | return pol; |
| 1848 | } |
| 1849 | |
| 1850 | bool vma_policy_mof(struct vm_area_struct *vma) |
| 1851 | { |
| 1852 | struct mempolicy *pol; |
| 1853 | |
| 1854 | if (vma->vm_ops && vma->vm_ops->get_policy) { |
| 1855 | bool ret = false; |
| 1856 | |
| 1857 | pol = vma->vm_ops->get_policy(vma, vma->vm_start); |
| 1858 | if (pol && (pol->flags & MPOL_F_MOF)) |
| 1859 | ret = true; |
| 1860 | mpol_cond_put(pol); |
| 1861 | |
| 1862 | return ret; |
| 1863 | } |
| 1864 | |
| 1865 | pol = vma->vm_policy; |
| 1866 | if (!pol) |
| 1867 | pol = get_task_policy(current); |
| 1868 | |
| 1869 | return pol->flags & MPOL_F_MOF; |
| 1870 | } |
| 1871 | |
| 1872 | static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone) |
| 1873 | { |
| 1874 | enum zone_type dynamic_policy_zone = policy_zone; |
| 1875 | |
| 1876 | BUG_ON(dynamic_policy_zone == ZONE_MOVABLE); |
| 1877 | |
| 1878 | /* |
| 1879 | * if policy->v.nodes has movable memory only, |
| 1880 | * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only. |
| 1881 | * |
| 1882 | * policy->v.nodes is intersect with node_states[N_MEMORY]. |
| 1883 | * so if the following test faile, it implies |
| 1884 | * policy->v.nodes has movable memory only. |
| 1885 | */ |
| 1886 | if (!nodes_intersects(policy->v.nodes, node_states[N_HIGH_MEMORY])) |
| 1887 | dynamic_policy_zone = ZONE_MOVABLE; |
| 1888 | |
| 1889 | return zone >= dynamic_policy_zone; |
| 1890 | } |
| 1891 | |
| 1892 | /* |
| 1893 | * Return a nodemask representing a mempolicy for filtering nodes for |
| 1894 | * page allocation |
| 1895 | */ |
| 1896 | static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy) |
| 1897 | { |
| 1898 | /* Lower zones don't get a nodemask applied for MPOL_BIND */ |
| 1899 | if (unlikely(policy->mode == MPOL_BIND) && |
| 1900 | apply_policy_zone(policy, gfp_zone(gfp)) && |
| 1901 | cpuset_nodemask_valid_mems_allowed(&policy->v.nodes)) |
| 1902 | return &policy->v.nodes; |
| 1903 | |
| 1904 | return NULL; |
| 1905 | } |
| 1906 | |
| 1907 | /* Return the node id preferred by the given mempolicy, or the given id */ |
| 1908 | static int policy_node(gfp_t gfp, struct mempolicy *policy, |
| 1909 | int nd) |
| 1910 | { |
| 1911 | if (policy->mode == MPOL_PREFERRED && !(policy->flags & MPOL_F_LOCAL)) |
| 1912 | nd = policy->v.preferred_node; |
| 1913 | else { |
| 1914 | /* |
| 1915 | * __GFP_THISNODE shouldn't even be used with the bind policy |
| 1916 | * because we might easily break the expectation to stay on the |
| 1917 | * requested node and not break the policy. |
| 1918 | */ |
| 1919 | WARN_ON_ONCE(policy->mode == MPOL_BIND && (gfp & __GFP_THISNODE)); |
| 1920 | } |
| 1921 | |
| 1922 | return nd; |
| 1923 | } |
| 1924 | |
| 1925 | /* Do dynamic interleaving for a process */ |
| 1926 | static unsigned interleave_nodes(struct mempolicy *policy) |
| 1927 | { |
| 1928 | unsigned next; |
| 1929 | struct task_struct *me = current; |
| 1930 | |
| 1931 | next = next_node_in(me->il_prev, policy->v.nodes); |
| 1932 | if (next < MAX_NUMNODES) |
| 1933 | me->il_prev = next; |
| 1934 | return next; |
| 1935 | } |
| 1936 | |
| 1937 | /* |
| 1938 | * Depending on the memory policy provide a node from which to allocate the |
| 1939 | * next slab entry. |
| 1940 | */ |
| 1941 | unsigned int mempolicy_slab_node(void) |
| 1942 | { |
| 1943 | struct mempolicy *policy; |
| 1944 | int node = numa_mem_id(); |
| 1945 | |
| 1946 | if (in_interrupt()) |
| 1947 | return node; |
| 1948 | |
| 1949 | policy = current->mempolicy; |
| 1950 | if (!policy || policy->flags & MPOL_F_LOCAL) |
| 1951 | return node; |
| 1952 | |
| 1953 | switch (policy->mode) { |
| 1954 | case MPOL_PREFERRED: |
| 1955 | /* |
| 1956 | * handled MPOL_F_LOCAL above |
| 1957 | */ |
| 1958 | return policy->v.preferred_node; |
| 1959 | |
| 1960 | case MPOL_INTERLEAVE: |
| 1961 | return interleave_nodes(policy); |
| 1962 | |
| 1963 | case MPOL_BIND: { |
| 1964 | struct zoneref *z; |
| 1965 | |
| 1966 | /* |
| 1967 | * Follow bind policy behavior and start allocation at the |
| 1968 | * first node. |
| 1969 | */ |
| 1970 | struct zonelist *zonelist; |
| 1971 | enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL); |
| 1972 | zonelist = &NODE_DATA(node)->node_zonelists[ZONELIST_FALLBACK]; |
| 1973 | z = first_zones_zonelist(zonelist, highest_zoneidx, |
| 1974 | &policy->v.nodes); |
| 1975 | return z->zone ? zone_to_nid(z->zone) : node; |
| 1976 | } |
| 1977 | |
| 1978 | default: |
| 1979 | BUG(); |
| 1980 | } |
| 1981 | } |
| 1982 | |
| 1983 | /* |
| 1984 | * Do static interleaving for a VMA with known offset @n. Returns the n'th |
| 1985 | * node in pol->v.nodes (starting from n=0), wrapping around if n exceeds the |
| 1986 | * number of present nodes. |
| 1987 | */ |
| 1988 | static unsigned offset_il_node(struct mempolicy *pol, unsigned long n) |
| 1989 | { |
| 1990 | unsigned nnodes = nodes_weight(pol->v.nodes); |
| 1991 | unsigned target; |
| 1992 | int i; |
| 1993 | int nid; |
| 1994 | |
| 1995 | if (!nnodes) |
| 1996 | return numa_node_id(); |
| 1997 | target = (unsigned int)n % nnodes; |
| 1998 | nid = first_node(pol->v.nodes); |
| 1999 | for (i = 0; i < target; i++) |
| 2000 | nid = next_node(nid, pol->v.nodes); |
| 2001 | return nid; |
| 2002 | } |
| 2003 | |
| 2004 | /* Determine a node number for interleave */ |
| 2005 | static inline unsigned interleave_nid(struct mempolicy *pol, |
| 2006 | struct vm_area_struct *vma, unsigned long addr, int shift) |
| 2007 | { |
| 2008 | if (vma) { |
| 2009 | unsigned long off; |
| 2010 | |
| 2011 | /* |
| 2012 | * for small pages, there is no difference between |
| 2013 | * shift and PAGE_SHIFT, so the bit-shift is safe. |
| 2014 | * for huge pages, since vm_pgoff is in units of small |
| 2015 | * pages, we need to shift off the always 0 bits to get |
| 2016 | * a useful offset. |
| 2017 | */ |
| 2018 | BUG_ON(shift < PAGE_SHIFT); |
| 2019 | off = vma->vm_pgoff >> (shift - PAGE_SHIFT); |
| 2020 | off += (addr - vma->vm_start) >> shift; |
| 2021 | return offset_il_node(pol, off); |
| 2022 | } else |
| 2023 | return interleave_nodes(pol); |
| 2024 | } |
| 2025 | |
| 2026 | #ifdef CONFIG_HUGETLBFS |
| 2027 | /* |
| 2028 | * huge_node(@vma, @addr, @gfp_flags, @mpol) |
| 2029 | * @vma: virtual memory area whose policy is sought |
| 2030 | * @addr: address in @vma for shared policy lookup and interleave policy |
| 2031 | * @gfp_flags: for requested zone |
| 2032 | * @mpol: pointer to mempolicy pointer for reference counted mempolicy |
| 2033 | * @nodemask: pointer to nodemask pointer for MPOL_BIND nodemask |
| 2034 | * |
| 2035 | * Returns a nid suitable for a huge page allocation and a pointer |
| 2036 | * to the struct mempolicy for conditional unref after allocation. |
| 2037 | * If the effective policy is 'BIND, returns a pointer to the mempolicy's |
| 2038 | * @nodemask for filtering the zonelist. |
| 2039 | * |
| 2040 | * Must be protected by read_mems_allowed_begin() |
| 2041 | */ |
| 2042 | int huge_node(struct vm_area_struct *vma, unsigned long addr, gfp_t gfp_flags, |
| 2043 | struct mempolicy **mpol, nodemask_t **nodemask) |
| 2044 | { |
| 2045 | int nid; |
| 2046 | |
| 2047 | *mpol = get_vma_policy(vma, addr); |
| 2048 | *nodemask = NULL; /* assume !MPOL_BIND */ |
| 2049 | |
| 2050 | if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) { |
| 2051 | nid = interleave_nid(*mpol, vma, addr, |
| 2052 | huge_page_shift(hstate_vma(vma))); |
| 2053 | } else { |
| 2054 | nid = policy_node(gfp_flags, *mpol, numa_node_id()); |
| 2055 | if ((*mpol)->mode == MPOL_BIND) |
| 2056 | *nodemask = &(*mpol)->v.nodes; |
| 2057 | } |
| 2058 | return nid; |
| 2059 | } |
| 2060 | |
| 2061 | /* |
| 2062 | * init_nodemask_of_mempolicy |
| 2063 | * |
| 2064 | * If the current task's mempolicy is "default" [NULL], return 'false' |
| 2065 | * to indicate default policy. Otherwise, extract the policy nodemask |
| 2066 | * for 'bind' or 'interleave' policy into the argument nodemask, or |
| 2067 | * initialize the argument nodemask to contain the single node for |
| 2068 | * 'preferred' or 'local' policy and return 'true' to indicate presence |
| 2069 | * of non-default mempolicy. |
| 2070 | * |
| 2071 | * We don't bother with reference counting the mempolicy [mpol_get/put] |
| 2072 | * because the current task is examining it's own mempolicy and a task's |
| 2073 | * mempolicy is only ever changed by the task itself. |
| 2074 | * |
| 2075 | * N.B., it is the caller's responsibility to free a returned nodemask. |
| 2076 | */ |
| 2077 | bool init_nodemask_of_mempolicy(nodemask_t *mask) |
| 2078 | { |
| 2079 | struct mempolicy *mempolicy; |
| 2080 | int nid; |
| 2081 | |
| 2082 | if (!(mask && current->mempolicy)) |
| 2083 | return false; |
| 2084 | |
| 2085 | task_lock(current); |
| 2086 | mempolicy = current->mempolicy; |
| 2087 | switch (mempolicy->mode) { |
| 2088 | case MPOL_PREFERRED: |
| 2089 | if (mempolicy->flags & MPOL_F_LOCAL) |
| 2090 | nid = numa_node_id(); |
| 2091 | else |
| 2092 | nid = mempolicy->v.preferred_node; |
| 2093 | init_nodemask_of_node(mask, nid); |
| 2094 | break; |
| 2095 | |
| 2096 | case MPOL_BIND: |
| 2097 | case MPOL_INTERLEAVE: |
| 2098 | *mask = mempolicy->v.nodes; |
| 2099 | break; |
| 2100 | |
| 2101 | default: |
| 2102 | BUG(); |
| 2103 | } |
| 2104 | task_unlock(current); |
| 2105 | |
| 2106 | return true; |
| 2107 | } |
| 2108 | #endif |
| 2109 | |
| 2110 | /* |
| 2111 | * mempolicy_nodemask_intersects |
| 2112 | * |
| 2113 | * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default |
| 2114 | * policy. Otherwise, check for intersection between mask and the policy |
| 2115 | * nodemask for 'bind' or 'interleave' policy. For 'perferred' or 'local' |
| 2116 | * policy, always return true since it may allocate elsewhere on fallback. |
| 2117 | * |
| 2118 | * Takes task_lock(tsk) to prevent freeing of its mempolicy. |
| 2119 | */ |
| 2120 | bool mempolicy_nodemask_intersects(struct task_struct *tsk, |
| 2121 | const nodemask_t *mask) |
| 2122 | { |
| 2123 | struct mempolicy *mempolicy; |
| 2124 | bool ret = true; |
| 2125 | |
| 2126 | if (!mask) |
| 2127 | return ret; |
| 2128 | task_lock(tsk); |
| 2129 | mempolicy = tsk->mempolicy; |
| 2130 | if (!mempolicy) |
| 2131 | goto out; |
| 2132 | |
| 2133 | switch (mempolicy->mode) { |
| 2134 | case MPOL_PREFERRED: |
| 2135 | /* |
| 2136 | * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to |
| 2137 | * allocate from, they may fallback to other nodes when oom. |
| 2138 | * Thus, it's possible for tsk to have allocated memory from |
| 2139 | * nodes in mask. |
| 2140 | */ |
| 2141 | break; |
| 2142 | case MPOL_BIND: |
| 2143 | case MPOL_INTERLEAVE: |
| 2144 | ret = nodes_intersects(mempolicy->v.nodes, *mask); |
| 2145 | break; |
| 2146 | default: |
| 2147 | BUG(); |
| 2148 | } |
| 2149 | out: |
| 2150 | task_unlock(tsk); |
| 2151 | return ret; |
| 2152 | } |
| 2153 | |
| 2154 | /* Allocate a page in interleaved policy. |
| 2155 | Own path because it needs to do special accounting. */ |
| 2156 | static struct page *alloc_page_interleave(gfp_t gfp, unsigned order, |
| 2157 | unsigned nid) |
| 2158 | { |
| 2159 | struct page *page; |
| 2160 | |
| 2161 | page = __alloc_pages(gfp, order, nid); |
| 2162 | /* skip NUMA_INTERLEAVE_HIT counter update if numa stats is disabled */ |
| 2163 | if (!static_branch_likely(&vm_numa_stat_key)) |
| 2164 | return page; |
| 2165 | if (page && page_to_nid(page) == nid) { |
| 2166 | preempt_disable(); |
| 2167 | __inc_numa_state(page_zone(page), NUMA_INTERLEAVE_HIT); |
| 2168 | preempt_enable(); |
| 2169 | } |
| 2170 | return page; |
| 2171 | } |
| 2172 | |
| 2173 | /** |
| 2174 | * alloc_pages_vma - Allocate a page for a VMA. |
| 2175 | * |
| 2176 | * @gfp: |
| 2177 | * %GFP_USER user allocation. |
| 2178 | * %GFP_KERNEL kernel allocations, |
| 2179 | * %GFP_HIGHMEM highmem/user allocations, |
| 2180 | * %GFP_FS allocation should not call back into a file system. |
| 2181 | * %GFP_ATOMIC don't sleep. |
| 2182 | * |
| 2183 | * @order:Order of the GFP allocation. |
| 2184 | * @vma: Pointer to VMA or NULL if not available. |
| 2185 | * @addr: Virtual Address of the allocation. Must be inside the VMA. |
| 2186 | * @node: Which node to prefer for allocation (modulo policy). |
| 2187 | * @hugepage: for hugepages try only the preferred node if possible |
| 2188 | * |
| 2189 | * This function allocates a page from the kernel page pool and applies |
| 2190 | * a NUMA policy associated with the VMA or the current process. |
| 2191 | * When VMA is not NULL caller must hold down_read on the mmap_sem of the |
| 2192 | * mm_struct of the VMA to prevent it from going away. Should be used for |
| 2193 | * all allocations for pages that will be mapped into user space. Returns |
| 2194 | * NULL when no page can be allocated. |
| 2195 | */ |
| 2196 | struct page * |
| 2197 | alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma, |
| 2198 | unsigned long addr, int node, bool hugepage) |
| 2199 | { |
| 2200 | struct mempolicy *pol; |
| 2201 | struct page *page; |
| 2202 | int preferred_nid; |
| 2203 | nodemask_t *nmask; |
| 2204 | |
| 2205 | pol = get_vma_policy(vma, addr); |
| 2206 | |
| 2207 | if (pol->mode == MPOL_INTERLEAVE) { |
| 2208 | unsigned nid; |
| 2209 | |
| 2210 | nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order); |
| 2211 | mpol_cond_put(pol); |
| 2212 | page = alloc_page_interleave(gfp, order, nid); |
| 2213 | goto out; |
| 2214 | } |
| 2215 | |
| 2216 | if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && hugepage)) { |
| 2217 | int hpage_node = node; |
| 2218 | |
| 2219 | /* |
| 2220 | * For hugepage allocation and non-interleave policy which |
| 2221 | * allows the current node (or other explicitly preferred |
| 2222 | * node) we only try to allocate from the current/preferred |
| 2223 | * node and don't fall back to other nodes, as the cost of |
| 2224 | * remote accesses would likely offset THP benefits. |
| 2225 | * |
| 2226 | * If the policy is interleave, or does not allow the current |
| 2227 | * node in its nodemask, we allocate the standard way. |
| 2228 | */ |
| 2229 | if (pol->mode == MPOL_PREFERRED && !(pol->flags & MPOL_F_LOCAL)) |
| 2230 | hpage_node = pol->v.preferred_node; |
| 2231 | |
| 2232 | nmask = policy_nodemask(gfp, pol); |
| 2233 | if (!nmask || node_isset(hpage_node, *nmask)) { |
| 2234 | mpol_cond_put(pol); |
| 2235 | /* |
| 2236 | * First, try to allocate THP only on local node, but |
| 2237 | * don't reclaim unnecessarily, just compact. |
| 2238 | */ |
| 2239 | page = __alloc_pages_node(hpage_node, |
| 2240 | gfp | __GFP_THISNODE | __GFP_NORETRY, order); |
| 2241 | |
| 2242 | /* |
| 2243 | * If hugepage allocations are configured to always |
| 2244 | * synchronous compact or the vma has been madvised |
| 2245 | * to prefer hugepage backing, retry allowing remote |
| 2246 | * memory with both reclaim and compact as well. |
| 2247 | */ |
| 2248 | if (!page && (gfp & __GFP_DIRECT_RECLAIM)) |
| 2249 | page = __alloc_pages_node(hpage_node, |
| 2250 | gfp, order); |
| 2251 | |
| 2252 | goto out; |
| 2253 | } |
| 2254 | } |
| 2255 | |
| 2256 | nmask = policy_nodemask(gfp, pol); |
| 2257 | preferred_nid = policy_node(gfp, pol, node); |
| 2258 | page = __alloc_pages_nodemask(gfp, order, preferred_nid, nmask); |
| 2259 | mpol_cond_put(pol); |
| 2260 | out: |
| 2261 | return page; |
| 2262 | } |
| 2263 | EXPORT_SYMBOL(alloc_pages_vma); |
| 2264 | |
| 2265 | /** |
| 2266 | * alloc_pages_current - Allocate pages. |
| 2267 | * |
| 2268 | * @gfp: |
| 2269 | * %GFP_USER user allocation, |
| 2270 | * %GFP_KERNEL kernel allocation, |
| 2271 | * %GFP_HIGHMEM highmem allocation, |
| 2272 | * %GFP_FS don't call back into a file system. |
| 2273 | * %GFP_ATOMIC don't sleep. |
| 2274 | * @order: Power of two of allocation size in pages. 0 is a single page. |
| 2275 | * |
| 2276 | * Allocate a page from the kernel page pool. When not in |
| 2277 | * interrupt context and apply the current process NUMA policy. |
| 2278 | * Returns NULL when no page can be allocated. |
| 2279 | */ |
| 2280 | struct page *alloc_pages_current(gfp_t gfp, unsigned order) |
| 2281 | { |
| 2282 | struct mempolicy *pol = &default_policy; |
| 2283 | struct page *page; |
| 2284 | |
| 2285 | if (!in_interrupt() && !(gfp & __GFP_THISNODE)) |
| 2286 | pol = get_task_policy(current); |
| 2287 | |
| 2288 | /* |
| 2289 | * No reference counting needed for current->mempolicy |
| 2290 | * nor system default_policy |
| 2291 | */ |
| 2292 | if (pol->mode == MPOL_INTERLEAVE) |
| 2293 | page = alloc_page_interleave(gfp, order, interleave_nodes(pol)); |
| 2294 | else |
| 2295 | page = __alloc_pages_nodemask(gfp, order, |
| 2296 | policy_node(gfp, pol, numa_node_id()), |
| 2297 | policy_nodemask(gfp, pol)); |
| 2298 | |
| 2299 | return page; |
| 2300 | } |
| 2301 | EXPORT_SYMBOL(alloc_pages_current); |
| 2302 | |
| 2303 | int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst) |
| 2304 | { |
| 2305 | struct mempolicy *pol = mpol_dup(vma_policy(src)); |
| 2306 | |
| 2307 | if (IS_ERR(pol)) |
| 2308 | return PTR_ERR(pol); |
| 2309 | dst->vm_policy = pol; |
| 2310 | return 0; |
| 2311 | } |
| 2312 | |
| 2313 | /* |
| 2314 | * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it |
| 2315 | * rebinds the mempolicy its copying by calling mpol_rebind_policy() |
| 2316 | * with the mems_allowed returned by cpuset_mems_allowed(). This |
| 2317 | * keeps mempolicies cpuset relative after its cpuset moves. See |
| 2318 | * further kernel/cpuset.c update_nodemask(). |
| 2319 | * |
| 2320 | * current's mempolicy may be rebinded by the other task(the task that changes |
| 2321 | * cpuset's mems), so we needn't do rebind work for current task. |
| 2322 | */ |
| 2323 | |
| 2324 | /* Slow path of a mempolicy duplicate */ |
| 2325 | struct mempolicy *__mpol_dup(struct mempolicy *old) |
| 2326 | { |
| 2327 | struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL); |
| 2328 | |
| 2329 | if (!new) |
| 2330 | return ERR_PTR(-ENOMEM); |
| 2331 | |
| 2332 | /* task's mempolicy is protected by alloc_lock */ |
| 2333 | if (old == current->mempolicy) { |
| 2334 | task_lock(current); |
| 2335 | *new = *old; |
| 2336 | task_unlock(current); |
| 2337 | } else |
| 2338 | *new = *old; |
| 2339 | |
| 2340 | if (current_cpuset_is_being_rebound()) { |
| 2341 | nodemask_t mems = cpuset_mems_allowed(current); |
| 2342 | mpol_rebind_policy(new, &mems); |
| 2343 | } |
| 2344 | atomic_set(&new->refcnt, 1); |
| 2345 | return new; |
| 2346 | } |
| 2347 | |
| 2348 | /* Slow path of a mempolicy comparison */ |
| 2349 | bool __mpol_equal(struct mempolicy *a, struct mempolicy *b) |
| 2350 | { |
| 2351 | if (!a || !b) |
| 2352 | return false; |
| 2353 | if (a->mode != b->mode) |
| 2354 | return false; |
| 2355 | if (a->flags != b->flags) |
| 2356 | return false; |
| 2357 | if (mpol_store_user_nodemask(a)) |
| 2358 | if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask)) |
| 2359 | return false; |
| 2360 | |
| 2361 | switch (a->mode) { |
| 2362 | case MPOL_BIND: |
| 2363 | case MPOL_INTERLEAVE: |
| 2364 | return !!nodes_equal(a->v.nodes, b->v.nodes); |
| 2365 | case MPOL_PREFERRED: |
| 2366 | /* a's ->flags is the same as b's */ |
| 2367 | if (a->flags & MPOL_F_LOCAL) |
| 2368 | return true; |
| 2369 | return a->v.preferred_node == b->v.preferred_node; |
| 2370 | default: |
| 2371 | BUG(); |
| 2372 | return false; |
| 2373 | } |
| 2374 | } |
| 2375 | |
| 2376 | /* |
| 2377 | * Shared memory backing store policy support. |
| 2378 | * |
| 2379 | * Remember policies even when nobody has shared memory mapped. |
| 2380 | * The policies are kept in Red-Black tree linked from the inode. |
| 2381 | * They are protected by the sp->lock rwlock, which should be held |
| 2382 | * for any accesses to the tree. |
| 2383 | */ |
| 2384 | |
| 2385 | /* |
| 2386 | * lookup first element intersecting start-end. Caller holds sp->lock for |
| 2387 | * reading or for writing |
| 2388 | */ |
| 2389 | static struct sp_node * |
| 2390 | sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end) |
| 2391 | { |
| 2392 | struct rb_node *n = sp->root.rb_node; |
| 2393 | |
| 2394 | while (n) { |
| 2395 | struct sp_node *p = rb_entry(n, struct sp_node, nd); |
| 2396 | |
| 2397 | if (start >= p->end) |
| 2398 | n = n->rb_right; |
| 2399 | else if (end <= p->start) |
| 2400 | n = n->rb_left; |
| 2401 | else |
| 2402 | break; |
| 2403 | } |
| 2404 | if (!n) |
| 2405 | return NULL; |
| 2406 | for (;;) { |
| 2407 | struct sp_node *w = NULL; |
| 2408 | struct rb_node *prev = rb_prev(n); |
| 2409 | if (!prev) |
| 2410 | break; |
| 2411 | w = rb_entry(prev, struct sp_node, nd); |
| 2412 | if (w->end <= start) |
| 2413 | break; |
| 2414 | n = prev; |
| 2415 | } |
| 2416 | return rb_entry(n, struct sp_node, nd); |
| 2417 | } |
| 2418 | |
| 2419 | /* |
| 2420 | * Insert a new shared policy into the list. Caller holds sp->lock for |
| 2421 | * writing. |
| 2422 | */ |
| 2423 | static void sp_insert(struct shared_policy *sp, struct sp_node *new) |
| 2424 | { |
| 2425 | struct rb_node **p = &sp->root.rb_node; |
| 2426 | struct rb_node *parent = NULL; |
| 2427 | struct sp_node *nd; |
| 2428 | |
| 2429 | while (*p) { |
| 2430 | parent = *p; |
| 2431 | nd = rb_entry(parent, struct sp_node, nd); |
| 2432 | if (new->start < nd->start) |
| 2433 | p = &(*p)->rb_left; |
| 2434 | else if (new->end > nd->end) |
| 2435 | p = &(*p)->rb_right; |
| 2436 | else |
| 2437 | BUG(); |
| 2438 | } |
| 2439 | rb_link_node(&new->nd, parent, p); |
| 2440 | rb_insert_color(&new->nd, &sp->root); |
| 2441 | pr_debug("inserting %lx-%lx: %d\n", new->start, new->end, |
| 2442 | new->policy ? new->policy->mode : 0); |
| 2443 | } |
| 2444 | |
| 2445 | /* Find shared policy intersecting idx */ |
| 2446 | struct mempolicy * |
| 2447 | mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx) |
| 2448 | { |
| 2449 | struct mempolicy *pol = NULL; |
| 2450 | struct sp_node *sn; |
| 2451 | |
| 2452 | if (!sp->root.rb_node) |
| 2453 | return NULL; |
| 2454 | read_lock(&sp->lock); |
| 2455 | sn = sp_lookup(sp, idx, idx+1); |
| 2456 | if (sn) { |
| 2457 | mpol_get(sn->policy); |
| 2458 | pol = sn->policy; |
| 2459 | } |
| 2460 | read_unlock(&sp->lock); |
| 2461 | return pol; |
| 2462 | } |
| 2463 | |
| 2464 | static void sp_free(struct sp_node *n) |
| 2465 | { |
| 2466 | mpol_put(n->policy); |
| 2467 | kmem_cache_free(sn_cache, n); |
| 2468 | } |
| 2469 | |
| 2470 | /** |
| 2471 | * mpol_misplaced - check whether current page node is valid in policy |
| 2472 | * |
| 2473 | * @page: page to be checked |
| 2474 | * @vma: vm area where page mapped |
| 2475 | * @addr: virtual address where page mapped |
| 2476 | * |
| 2477 | * Lookup current policy node id for vma,addr and "compare to" page's |
| 2478 | * node id. |
| 2479 | * |
| 2480 | * Returns: |
| 2481 | * -1 - not misplaced, page is in the right node |
| 2482 | * node - node id where the page should be |
| 2483 | * |
| 2484 | * Policy determination "mimics" alloc_page_vma(). |
| 2485 | * Called from fault path where we know the vma and faulting address. |
| 2486 | */ |
| 2487 | int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr) |
| 2488 | { |
| 2489 | struct mempolicy *pol; |
| 2490 | struct zoneref *z; |
| 2491 | int curnid = page_to_nid(page); |
| 2492 | unsigned long pgoff; |
| 2493 | int thiscpu = raw_smp_processor_id(); |
| 2494 | int thisnid = cpu_to_node(thiscpu); |
| 2495 | int polnid = NUMA_NO_NODE; |
| 2496 | int ret = -1; |
| 2497 | |
| 2498 | pol = get_vma_policy(vma, addr); |
| 2499 | if (!(pol->flags & MPOL_F_MOF)) |
| 2500 | goto out; |
| 2501 | |
| 2502 | switch (pol->mode) { |
| 2503 | case MPOL_INTERLEAVE: |
| 2504 | pgoff = vma->vm_pgoff; |
| 2505 | pgoff += (addr - vma->vm_start) >> PAGE_SHIFT; |
| 2506 | polnid = offset_il_node(pol, pgoff); |
| 2507 | break; |
| 2508 | |
| 2509 | case MPOL_PREFERRED: |
| 2510 | if (pol->flags & MPOL_F_LOCAL) |
| 2511 | polnid = numa_node_id(); |
| 2512 | else |
| 2513 | polnid = pol->v.preferred_node; |
| 2514 | break; |
| 2515 | |
| 2516 | case MPOL_BIND: |
| 2517 | |
| 2518 | /* |
| 2519 | * allows binding to multiple nodes. |
| 2520 | * use current page if in policy nodemask, |
| 2521 | * else select nearest allowed node, if any. |
| 2522 | * If no allowed nodes, use current [!misplaced]. |
| 2523 | */ |
| 2524 | if (node_isset(curnid, pol->v.nodes)) |
| 2525 | goto out; |
| 2526 | z = first_zones_zonelist( |
| 2527 | node_zonelist(numa_node_id(), GFP_HIGHUSER), |
| 2528 | gfp_zone(GFP_HIGHUSER), |
| 2529 | &pol->v.nodes); |
| 2530 | polnid = zone_to_nid(z->zone); |
| 2531 | break; |
| 2532 | |
| 2533 | default: |
| 2534 | BUG(); |
| 2535 | } |
| 2536 | |
| 2537 | /* Migrate the page towards the node whose CPU is referencing it */ |
| 2538 | if (pol->flags & MPOL_F_MORON) { |
| 2539 | polnid = thisnid; |
| 2540 | |
| 2541 | if (!should_numa_migrate_memory(current, page, curnid, thiscpu)) |
| 2542 | goto out; |
| 2543 | } |
| 2544 | |
| 2545 | if (curnid != polnid) |
| 2546 | ret = polnid; |
| 2547 | out: |
| 2548 | mpol_cond_put(pol); |
| 2549 | |
| 2550 | return ret; |
| 2551 | } |
| 2552 | |
| 2553 | /* |
| 2554 | * Drop the (possibly final) reference to task->mempolicy. It needs to be |
| 2555 | * dropped after task->mempolicy is set to NULL so that any allocation done as |
| 2556 | * part of its kmem_cache_free(), such as by KASAN, doesn't reference a freed |
| 2557 | * policy. |
| 2558 | */ |
| 2559 | void mpol_put_task_policy(struct task_struct *task) |
| 2560 | { |
| 2561 | struct mempolicy *pol; |
| 2562 | |
| 2563 | task_lock(task); |
| 2564 | pol = task->mempolicy; |
| 2565 | task->mempolicy = NULL; |
| 2566 | task_unlock(task); |
| 2567 | mpol_put(pol); |
| 2568 | } |
| 2569 | |
| 2570 | static void sp_delete(struct shared_policy *sp, struct sp_node *n) |
| 2571 | { |
| 2572 | pr_debug("deleting %lx-l%lx\n", n->start, n->end); |
| 2573 | rb_erase(&n->nd, &sp->root); |
| 2574 | sp_free(n); |
| 2575 | } |
| 2576 | |
| 2577 | static void sp_node_init(struct sp_node *node, unsigned long start, |
| 2578 | unsigned long end, struct mempolicy *pol) |
| 2579 | { |
| 2580 | node->start = start; |
| 2581 | node->end = end; |
| 2582 | node->policy = pol; |
| 2583 | } |
| 2584 | |
| 2585 | static struct sp_node *sp_alloc(unsigned long start, unsigned long end, |
| 2586 | struct mempolicy *pol) |
| 2587 | { |
| 2588 | struct sp_node *n; |
| 2589 | struct mempolicy *newpol; |
| 2590 | |
| 2591 | n = kmem_cache_alloc(sn_cache, GFP_KERNEL); |
| 2592 | if (!n) |
| 2593 | return NULL; |
| 2594 | |
| 2595 | newpol = mpol_dup(pol); |
| 2596 | if (IS_ERR(newpol)) { |
| 2597 | kmem_cache_free(sn_cache, n); |
| 2598 | return NULL; |
| 2599 | } |
| 2600 | newpol->flags |= MPOL_F_SHARED; |
| 2601 | sp_node_init(n, start, end, newpol); |
| 2602 | |
| 2603 | return n; |
| 2604 | } |
| 2605 | |
| 2606 | /* Replace a policy range. */ |
| 2607 | static int shared_policy_replace(struct shared_policy *sp, unsigned long start, |
| 2608 | unsigned long end, struct sp_node *new) |
| 2609 | { |
| 2610 | struct sp_node *n; |
| 2611 | struct sp_node *n_new = NULL; |
| 2612 | struct mempolicy *mpol_new = NULL; |
| 2613 | int ret = 0; |
| 2614 | |
| 2615 | restart: |
| 2616 | write_lock(&sp->lock); |
| 2617 | n = sp_lookup(sp, start, end); |
| 2618 | /* Take care of old policies in the same range. */ |
| 2619 | while (n && n->start < end) { |
| 2620 | struct rb_node *next = rb_next(&n->nd); |
| 2621 | if (n->start >= start) { |
| 2622 | if (n->end <= end) |
| 2623 | sp_delete(sp, n); |
| 2624 | else |
| 2625 | n->start = end; |
| 2626 | } else { |
| 2627 | /* Old policy spanning whole new range. */ |
| 2628 | if (n->end > end) { |
| 2629 | if (!n_new) |
| 2630 | goto alloc_new; |
| 2631 | |
| 2632 | *mpol_new = *n->policy; |
| 2633 | atomic_set(&mpol_new->refcnt, 1); |
| 2634 | sp_node_init(n_new, end, n->end, mpol_new); |
| 2635 | n->end = start; |
| 2636 | sp_insert(sp, n_new); |
| 2637 | n_new = NULL; |
| 2638 | mpol_new = NULL; |
| 2639 | break; |
| 2640 | } else |
| 2641 | n->end = start; |
| 2642 | } |
| 2643 | if (!next) |
| 2644 | break; |
| 2645 | n = rb_entry(next, struct sp_node, nd); |
| 2646 | } |
| 2647 | if (new) |
| 2648 | sp_insert(sp, new); |
| 2649 | write_unlock(&sp->lock); |
| 2650 | ret = 0; |
| 2651 | |
| 2652 | err_out: |
| 2653 | if (mpol_new) |
| 2654 | mpol_put(mpol_new); |
| 2655 | if (n_new) |
| 2656 | kmem_cache_free(sn_cache, n_new); |
| 2657 | |
| 2658 | return ret; |
| 2659 | |
| 2660 | alloc_new: |
| 2661 | write_unlock(&sp->lock); |
| 2662 | ret = -ENOMEM; |
| 2663 | n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL); |
| 2664 | if (!n_new) |
| 2665 | goto err_out; |
| 2666 | mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL); |
| 2667 | if (!mpol_new) |
| 2668 | goto err_out; |
| 2669 | goto restart; |
| 2670 | } |
| 2671 | |
| 2672 | /** |
| 2673 | * mpol_shared_policy_init - initialize shared policy for inode |
| 2674 | * @sp: pointer to inode shared policy |
| 2675 | * @mpol: struct mempolicy to install |
| 2676 | * |
| 2677 | * Install non-NULL @mpol in inode's shared policy rb-tree. |
| 2678 | * On entry, the current task has a reference on a non-NULL @mpol. |
| 2679 | * This must be released on exit. |
| 2680 | * This is called at get_inode() calls and we can use GFP_KERNEL. |
| 2681 | */ |
| 2682 | void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol) |
| 2683 | { |
| 2684 | int ret; |
| 2685 | |
| 2686 | sp->root = RB_ROOT; /* empty tree == default mempolicy */ |
| 2687 | rwlock_init(&sp->lock); |
| 2688 | |
| 2689 | if (mpol) { |
| 2690 | struct vm_area_struct pvma; |
| 2691 | struct mempolicy *new; |
| 2692 | NODEMASK_SCRATCH(scratch); |
| 2693 | |
| 2694 | if (!scratch) |
| 2695 | goto put_mpol; |
| 2696 | /* contextualize the tmpfs mount point mempolicy */ |
| 2697 | new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask); |
| 2698 | if (IS_ERR(new)) |
| 2699 | goto free_scratch; /* no valid nodemask intersection */ |
| 2700 | |
| 2701 | task_lock(current); |
| 2702 | ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch); |
| 2703 | task_unlock(current); |
| 2704 | if (ret) |
| 2705 | goto put_new; |
| 2706 | |
| 2707 | /* Create pseudo-vma that contains just the policy */ |
| 2708 | vma_init(&pvma, NULL); |
| 2709 | pvma.vm_end = TASK_SIZE; /* policy covers entire file */ |
| 2710 | mpol_set_shared_policy(sp, &pvma, new); /* adds ref */ |
| 2711 | |
| 2712 | put_new: |
| 2713 | mpol_put(new); /* drop initial ref */ |
| 2714 | free_scratch: |
| 2715 | NODEMASK_SCRATCH_FREE(scratch); |
| 2716 | put_mpol: |
| 2717 | mpol_put(mpol); /* drop our incoming ref on sb mpol */ |
| 2718 | } |
| 2719 | } |
| 2720 | |
| 2721 | int mpol_set_shared_policy(struct shared_policy *info, |
| 2722 | struct vm_area_struct *vma, struct mempolicy *npol) |
| 2723 | { |
| 2724 | int err; |
| 2725 | struct sp_node *new = NULL; |
| 2726 | unsigned long sz = vma_pages(vma); |
| 2727 | |
| 2728 | pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n", |
| 2729 | vma->vm_pgoff, |
| 2730 | sz, npol ? npol->mode : -1, |
| 2731 | npol ? npol->flags : -1, |
| 2732 | npol ? nodes_addr(npol->v.nodes)[0] : NUMA_NO_NODE); |
| 2733 | |
| 2734 | if (npol) { |
| 2735 | new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol); |
| 2736 | if (!new) |
| 2737 | return -ENOMEM; |
| 2738 | } |
| 2739 | err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new); |
| 2740 | if (err && new) |
| 2741 | sp_free(new); |
| 2742 | return err; |
| 2743 | } |
| 2744 | |
| 2745 | /* Free a backing policy store on inode delete. */ |
| 2746 | void mpol_free_shared_policy(struct shared_policy *p) |
| 2747 | { |
| 2748 | struct sp_node *n; |
| 2749 | struct rb_node *next; |
| 2750 | |
| 2751 | if (!p->root.rb_node) |
| 2752 | return; |
| 2753 | write_lock(&p->lock); |
| 2754 | next = rb_first(&p->root); |
| 2755 | while (next) { |
| 2756 | n = rb_entry(next, struct sp_node, nd); |
| 2757 | next = rb_next(&n->nd); |
| 2758 | sp_delete(p, n); |
| 2759 | } |
| 2760 | write_unlock(&p->lock); |
| 2761 | } |
| 2762 | |
| 2763 | #ifdef CONFIG_NUMA_BALANCING |
| 2764 | static int __initdata numabalancing_override; |
| 2765 | |
| 2766 | static void __init check_numabalancing_enable(void) |
| 2767 | { |
| 2768 | bool numabalancing_default = false; |
| 2769 | |
| 2770 | if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED)) |
| 2771 | numabalancing_default = true; |
| 2772 | |
| 2773 | /* Parsed by setup_numabalancing. override == 1 enables, -1 disables */ |
| 2774 | if (numabalancing_override) |
| 2775 | set_numabalancing_state(numabalancing_override == 1); |
| 2776 | |
| 2777 | if (num_online_nodes() > 1 && !numabalancing_override) { |
| 2778 | pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n", |
| 2779 | numabalancing_default ? "Enabling" : "Disabling"); |
| 2780 | set_numabalancing_state(numabalancing_default); |
| 2781 | } |
| 2782 | } |
| 2783 | |
| 2784 | static int __init setup_numabalancing(char *str) |
| 2785 | { |
| 2786 | int ret = 0; |
| 2787 | if (!str) |
| 2788 | goto out; |
| 2789 | |
| 2790 | if (!strcmp(str, "enable")) { |
| 2791 | numabalancing_override = 1; |
| 2792 | ret = 1; |
| 2793 | } else if (!strcmp(str, "disable")) { |
| 2794 | numabalancing_override = -1; |
| 2795 | ret = 1; |
| 2796 | } |
| 2797 | out: |
| 2798 | if (!ret) |
| 2799 | pr_warn("Unable to parse numa_balancing=\n"); |
| 2800 | |
| 2801 | return ret; |
| 2802 | } |
| 2803 | __setup("numa_balancing=", setup_numabalancing); |
| 2804 | #else |
| 2805 | static inline void __init check_numabalancing_enable(void) |
| 2806 | { |
| 2807 | } |
| 2808 | #endif /* CONFIG_NUMA_BALANCING */ |
| 2809 | |
| 2810 | /* assumes fs == KERNEL_DS */ |
| 2811 | void __init numa_policy_init(void) |
| 2812 | { |
| 2813 | nodemask_t interleave_nodes; |
| 2814 | unsigned long largest = 0; |
| 2815 | int nid, prefer = 0; |
| 2816 | |
| 2817 | policy_cache = kmem_cache_create("numa_policy", |
| 2818 | sizeof(struct mempolicy), |
| 2819 | 0, SLAB_PANIC, NULL); |
| 2820 | |
| 2821 | sn_cache = kmem_cache_create("shared_policy_node", |
| 2822 | sizeof(struct sp_node), |
| 2823 | 0, SLAB_PANIC, NULL); |
| 2824 | |
| 2825 | for_each_node(nid) { |
| 2826 | preferred_node_policy[nid] = (struct mempolicy) { |
| 2827 | .refcnt = ATOMIC_INIT(1), |
| 2828 | .mode = MPOL_PREFERRED, |
| 2829 | .flags = MPOL_F_MOF | MPOL_F_MORON, |
| 2830 | .v = { .preferred_node = nid, }, |
| 2831 | }; |
| 2832 | } |
| 2833 | |
| 2834 | /* |
| 2835 | * Set interleaving policy for system init. Interleaving is only |
| 2836 | * enabled across suitably sized nodes (default is >= 16MB), or |
| 2837 | * fall back to the largest node if they're all smaller. |
| 2838 | */ |
| 2839 | nodes_clear(interleave_nodes); |
| 2840 | for_each_node_state(nid, N_MEMORY) { |
| 2841 | unsigned long total_pages = node_present_pages(nid); |
| 2842 | |
| 2843 | /* Preserve the largest node */ |
| 2844 | if (largest < total_pages) { |
| 2845 | largest = total_pages; |
| 2846 | prefer = nid; |
| 2847 | } |
| 2848 | |
| 2849 | /* Interleave this node? */ |
| 2850 | if ((total_pages << PAGE_SHIFT) >= (16 << 20)) |
| 2851 | node_set(nid, interleave_nodes); |
| 2852 | } |
| 2853 | |
| 2854 | /* All too small, use the largest */ |
| 2855 | if (unlikely(nodes_empty(interleave_nodes))) |
| 2856 | node_set(prefer, interleave_nodes); |
| 2857 | |
| 2858 | if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes)) |
| 2859 | pr_err("%s: interleaving failed\n", __func__); |
| 2860 | |
| 2861 | check_numabalancing_enable(); |
| 2862 | } |
| 2863 | |
| 2864 | /* Reset policy of current process to default */ |
| 2865 | void numa_default_policy(void) |
| 2866 | { |
| 2867 | do_set_mempolicy(MPOL_DEFAULT, 0, NULL); |
| 2868 | } |
| 2869 | |
| 2870 | /* |
| 2871 | * Parse and format mempolicy from/to strings |
| 2872 | */ |
| 2873 | |
| 2874 | /* |
| 2875 | * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag. |
| 2876 | */ |
| 2877 | static const char * const policy_modes[] = |
| 2878 | { |
| 2879 | [MPOL_DEFAULT] = "default", |
| 2880 | [MPOL_PREFERRED] = "prefer", |
| 2881 | [MPOL_BIND] = "bind", |
| 2882 | [MPOL_INTERLEAVE] = "interleave", |
| 2883 | [MPOL_LOCAL] = "local", |
| 2884 | }; |
| 2885 | |
| 2886 | |
| 2887 | #ifdef CONFIG_TMPFS |
| 2888 | /** |
| 2889 | * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option. |
| 2890 | * @str: string containing mempolicy to parse |
| 2891 | * @mpol: pointer to struct mempolicy pointer, returned on success. |
| 2892 | * |
| 2893 | * Format of input: |
| 2894 | * <mode>[=<flags>][:<nodelist>] |
| 2895 | * |
| 2896 | * On success, returns 0, else 1 |
| 2897 | */ |
| 2898 | int mpol_parse_str(char *str, struct mempolicy **mpol) |
| 2899 | { |
| 2900 | struct mempolicy *new = NULL; |
| 2901 | unsigned short mode_flags; |
| 2902 | nodemask_t nodes; |
| 2903 | char *nodelist = strchr(str, ':'); |
| 2904 | char *flags = strchr(str, '='); |
| 2905 | int err = 1, mode; |
| 2906 | |
| 2907 | if (flags) |
| 2908 | *flags++ = '\0'; /* terminate mode string */ |
| 2909 | |
| 2910 | if (nodelist) { |
| 2911 | /* NUL-terminate mode or flags string */ |
| 2912 | *nodelist++ = '\0'; |
| 2913 | if (nodelist_parse(nodelist, nodes)) |
| 2914 | goto out; |
| 2915 | if (!nodes_subset(nodes, node_states[N_MEMORY])) |
| 2916 | goto out; |
| 2917 | } else |
| 2918 | nodes_clear(nodes); |
| 2919 | |
| 2920 | mode = match_string(policy_modes, MPOL_MAX, str); |
| 2921 | if (mode < 0) |
| 2922 | goto out; |
| 2923 | |
| 2924 | switch (mode) { |
| 2925 | case MPOL_PREFERRED: |
| 2926 | /* |
| 2927 | * Insist on a nodelist of one node only, although later |
| 2928 | * we use first_node(nodes) to grab a single node, so here |
| 2929 | * nodelist (or nodes) cannot be empty. |
| 2930 | */ |
| 2931 | if (nodelist) { |
| 2932 | char *rest = nodelist; |
| 2933 | while (isdigit(*rest)) |
| 2934 | rest++; |
| 2935 | if (*rest) |
| 2936 | goto out; |
| 2937 | if (nodes_empty(nodes)) |
| 2938 | goto out; |
| 2939 | } |
| 2940 | break; |
| 2941 | case MPOL_INTERLEAVE: |
| 2942 | /* |
| 2943 | * Default to online nodes with memory if no nodelist |
| 2944 | */ |
| 2945 | if (!nodelist) |
| 2946 | nodes = node_states[N_MEMORY]; |
| 2947 | break; |
| 2948 | case MPOL_LOCAL: |
| 2949 | /* |
| 2950 | * Don't allow a nodelist; mpol_new() checks flags |
| 2951 | */ |
| 2952 | if (nodelist) |
| 2953 | goto out; |
| 2954 | mode = MPOL_PREFERRED; |
| 2955 | break; |
| 2956 | case MPOL_DEFAULT: |
| 2957 | /* |
| 2958 | * Insist on a empty nodelist |
| 2959 | */ |
| 2960 | if (!nodelist) |
| 2961 | err = 0; |
| 2962 | goto out; |
| 2963 | case MPOL_BIND: |
| 2964 | /* |
| 2965 | * Insist on a nodelist |
| 2966 | */ |
| 2967 | if (!nodelist) |
| 2968 | goto out; |
| 2969 | } |
| 2970 | |
| 2971 | mode_flags = 0; |
| 2972 | if (flags) { |
| 2973 | /* |
| 2974 | * Currently, we only support two mutually exclusive |
| 2975 | * mode flags. |
| 2976 | */ |
| 2977 | if (!strcmp(flags, "static")) |
| 2978 | mode_flags |= MPOL_F_STATIC_NODES; |
| 2979 | else if (!strcmp(flags, "relative")) |
| 2980 | mode_flags |= MPOL_F_RELATIVE_NODES; |
| 2981 | else |
| 2982 | goto out; |
| 2983 | } |
| 2984 | |
| 2985 | new = mpol_new(mode, mode_flags, &nodes); |
| 2986 | if (IS_ERR(new)) |
| 2987 | goto out; |
| 2988 | |
| 2989 | /* |
| 2990 | * Save nodes for mpol_to_str() to show the tmpfs mount options |
| 2991 | * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo. |
| 2992 | */ |
| 2993 | if (mode != MPOL_PREFERRED) |
| 2994 | new->v.nodes = nodes; |
| 2995 | else if (nodelist) |
| 2996 | new->v.preferred_node = first_node(nodes); |
| 2997 | else |
| 2998 | new->flags |= MPOL_F_LOCAL; |
| 2999 | |
| 3000 | /* |
| 3001 | * Save nodes for contextualization: this will be used to "clone" |
| 3002 | * the mempolicy in a specific context [cpuset] at a later time. |
| 3003 | */ |
| 3004 | new->w.user_nodemask = nodes; |
| 3005 | |
| 3006 | err = 0; |
| 3007 | |
| 3008 | out: |
| 3009 | /* Restore string for error message */ |
| 3010 | if (nodelist) |
| 3011 | *--nodelist = ':'; |
| 3012 | if (flags) |
| 3013 | *--flags = '='; |
| 3014 | if (!err) |
| 3015 | *mpol = new; |
| 3016 | return err; |
| 3017 | } |
| 3018 | #endif /* CONFIG_TMPFS */ |
| 3019 | |
| 3020 | /** |
| 3021 | * mpol_to_str - format a mempolicy structure for printing |
| 3022 | * @buffer: to contain formatted mempolicy string |
| 3023 | * @maxlen: length of @buffer |
| 3024 | * @pol: pointer to mempolicy to be formatted |
| 3025 | * |
| 3026 | * Convert @pol into a string. If @buffer is too short, truncate the string. |
| 3027 | * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the |
| 3028 | * longest flag, "relative", and to display at least a few node ids. |
| 3029 | */ |
| 3030 | void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol) |
| 3031 | { |
| 3032 | char *p = buffer; |
| 3033 | nodemask_t nodes = NODE_MASK_NONE; |
| 3034 | unsigned short mode = MPOL_DEFAULT; |
| 3035 | unsigned short flags = 0; |
| 3036 | |
| 3037 | if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) { |
| 3038 | mode = pol->mode; |
| 3039 | flags = pol->flags; |
| 3040 | } |
| 3041 | |
| 3042 | switch (mode) { |
| 3043 | case MPOL_DEFAULT: |
| 3044 | break; |
| 3045 | case MPOL_PREFERRED: |
| 3046 | if (flags & MPOL_F_LOCAL) |
| 3047 | mode = MPOL_LOCAL; |
| 3048 | else |
| 3049 | node_set(pol->v.preferred_node, nodes); |
| 3050 | break; |
| 3051 | case MPOL_BIND: |
| 3052 | case MPOL_INTERLEAVE: |
| 3053 | nodes = pol->v.nodes; |
| 3054 | break; |
| 3055 | default: |
| 3056 | WARN_ON_ONCE(1); |
| 3057 | snprintf(p, maxlen, "unknown"); |
| 3058 | return; |
| 3059 | } |
| 3060 | |
| 3061 | p += snprintf(p, maxlen, "%s", policy_modes[mode]); |
| 3062 | |
| 3063 | if (flags & MPOL_MODE_FLAGS) { |
| 3064 | p += snprintf(p, buffer + maxlen - p, "="); |
| 3065 | |
| 3066 | /* |
| 3067 | * Currently, the only defined flags are mutually exclusive |
| 3068 | */ |
| 3069 | if (flags & MPOL_F_STATIC_NODES) |
| 3070 | p += snprintf(p, buffer + maxlen - p, "static"); |
| 3071 | else if (flags & MPOL_F_RELATIVE_NODES) |
| 3072 | p += snprintf(p, buffer + maxlen - p, "relative"); |
| 3073 | } |
| 3074 | |
| 3075 | if (!nodes_empty(nodes)) |
| 3076 | p += scnprintf(p, buffer + maxlen - p, ":%*pbl", |
| 3077 | nodemask_pr_args(&nodes)); |
| 3078 | } |