| 1 | |
| 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 six 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 | * weighted interleave
|
| 23 | * Allocate memory interleaved over a set of nodes based on
|
| 24 | * a set of weights (per-node), with normal fallback if it
|
| 25 | * fails. Otherwise operates the same as interleave.
|
| 26 | * Example: nodeset(0,1) & weights (2,1) - 2 pages allocated
|
| 27 | * on node 0 for every 1 page allocated on node 1.
|
| 28 | *
|
| 29 | * bind Only allocate memory on a specific set of nodes,
|
| 30 | * no fallback.
|
| 31 | * FIXME: memory is allocated starting with the first node
|
| 32 | * to the last. It would be better if bind would truly restrict
|
| 33 | * the allocation to memory nodes instead
|
| 34 | *
|
| 35 | * preferred Try a specific node first before normal fallback.
|
| 36 | * As a special case NUMA_NO_NODE here means do the allocation
|
| 37 | * on the local CPU. This is normally identical to default,
|
| 38 | * but useful to set in a VMA when you have a non default
|
| 39 | * process policy.
|
| 40 | *
|
| 41 | * preferred many Try a set of nodes first before normal fallback. This is
|
| 42 | * similar to preferred without the special case.
|
| 43 | *
|
| 44 | * default Allocate on the local node first, or when on a VMA
|
| 45 | * use the process policy. This is what Linux always did
|
| 46 | * in a NUMA aware kernel and still does by, ahem, default.
|
| 47 | *
|
| 48 | * The process policy is applied for most non interrupt memory allocations
|
| 49 | * in that process' context. Interrupts ignore the policies and always
|
| 50 | * try to allocate on the local CPU. The VMA policy is only applied for memory
|
| 51 | * allocations for a VMA in the VM.
|
| 52 | *
|
| 53 | * Currently there are a few corner cases in swapping where the policy
|
| 54 | * is not applied, but the majority should be handled. When process policy
|
| 55 | * is used it is not remembered over swap outs/swap ins.
|
| 56 | *
|
| 57 | * Only the highest zone in the zone hierarchy gets policied. Allocations
|
| 58 | * requesting a lower zone just use default policy. This implies that
|
| 59 | * on systems with highmem kernel lowmem allocation don't get policied.
|
| 60 | * Same with GFP_DMA allocations.
|
| 61 | *
|
| 62 | * For shmem/tmpfs shared memory the policy is shared between
|
| 63 | * all users and remembered even when nobody has memory mapped.
|
| 64 |
|
| 65 |
|
| 66 | |
| 67 | fix mmap readahead to honour policy and enable policy for any page cache
|
| 68 | object
|
| 69 | statistics for bigpages
|
| 70 | global policy for page cache? currently it uses process policy. Requires
|
| 71 | first item above.
|
| 72 | handle mremap for shared memory (currently ignored for the policy)
|
| 73 | grows down?
|
| 74 | make bind policy root only? It can trigger oom much faster and the
|
| 75 | kernel is not always grateful with that.
|
| 76 |
|
| 77 |
|
| 78 | #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
|
| 79 |
|
| 80 | #include <linux/mempolicy.h>
|
| 81 | #include <linux/pagewalk.h>
|
| 82 | #include <linux/highmem.h>
|
| 83 | #include <linux/hugetlb.h>
|
| 84 | #include <linux/kernel.h>
|
| 85 | #include <linux/sched.h>
|
| 86 | #include <linux/sched/mm.h>
|
| 87 | #include <linux/sched/numa_balancing.h>
|
| 88 | #include <linux/sched/sysctl.h>
|
| 89 | #include <linux/sched/task.h>
|
| 90 | #include <linux/nodemask.h>
|
| 91 | #include <linux/cpuset.h>
|
| 92 | #include <linux/slab.h>
|
| 93 | #include <linux/string.h>
|
| 94 | #include <linux/export.h>
|
| 95 | #include <linux/nsproxy.h>
|
| 96 | #include <linux/interrupt.h>
|
| 97 | #include <linux/init.h>
|
| 98 | #include <linux/compat.h>
|
| 99 | #include <linux/ptrace.h>
|
| 100 | #include <linux/swap.h>
|
| 101 | #include <linux/seq_file.h>
|
| 102 | #include <linux/proc_fs.h>
|
| 103 | #include <linux/memory-tiers.h>
|
| 104 | #include <linux/migrate.h>
|
| 105 | #include <linux/ksm.h>
|
| 106 | #include <linux/rmap.h>
|
| 107 | #include <linux/security.h>
|
| 108 | #include <linux/syscalls.h>
|
| 109 | #include <linux/ctype.h>
|
| 110 | #include <linux/mm_inline.h>
|
| 111 | #include <linux/mmu_notifier.h>
|
| 112 | #include <linux/printk.h>
|
| 113 | #include <linux/leafops.h>
|
| 114 | #include <linux/gcd.h>
|
| 115 |
|
| 116 | #include <asm/tlbflush.h>
|
| 117 | #include <asm/tlb.h>
|
| 118 | #include <linux/uaccess.h>
|
| 119 | #include <linux/memory.h>
|
| 120 |
|
| 121 | #include "internal.h"
|
| 122 | #include "page_alloc.h"
|
| 123 |
|
| 124 |
|
| 125 | #define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0)
|
| 126 | #define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1)
|
| 127 | #define MPOL_MF_WRLOCK (MPOL_MF_INTERNAL << 2)
|
| 128 |
|
| 129 | static struct kmem_cache *policy_cache;
|
| 130 | static struct kmem_cache *sn_cache;
|
| 131 |
|
| 132 | |
| 133 | policied.
|
| 134 | enum zone_type policy_zone = 0;
|
| 135 |
|
| 136 | |
| 137 | * run-time system-wide default policy => local allocation
|
| 138 |
|
| 139 | static struct mempolicy default_policy = {
|
| 140 | .refcnt = ATOMIC_INIT(1),
|
| 141 | .mode = MPOL_LOCAL,
|
| 142 | };
|
| 143 |
|
| 144 | static struct mempolicy preferred_node_policy[MAX_NUMNODES];
|
| 145 |
|
| 146 | |
| 147 | * weightiness balances the tradeoff between small weights (cycles through nodes
|
| 148 | * faster, more fair/even distribution) and large weights (smaller errors
|
| 149 | * between actual bandwidth ratios and weight ratios). 32 is a number that has
|
| 150 | * been found to perform at a reasonable compromise between the two goals.
|
| 151 |
|
| 152 | static const int weightiness = 32;
|
| 153 |
|
| 154 | |
| 155 | * A null weighted_interleave_state is interpreted as having .mode="auto",
|
| 156 | * and .iw_table is interpreted as an array of 1s with length nr_node_ids.
|
| 157 |
|
| 158 | struct weighted_interleave_state {
|
| 159 | bool mode_auto;
|
| 160 | u8 iw_table[];
|
| 161 | };
|
| 162 | static struct weighted_interleave_state __rcu *wi_state;
|
| 163 | static unsigned int *node_bw_table;
|
| 164 |
|
| 165 | |
| 166 | * wi_state_lock protects both wi_state and node_bw_table.
|
| 167 | * node_bw_table is only used by writers to update wi_state.
|
| 168 |
|
| 169 | static DEFINE_MUTEX(wi_state_lock);
|
| 170 |
|
| 171 | static u8 get_il_weight(int node)
|
| 172 | {
|
| 173 | struct weighted_interleave_state *state;
|
| 174 | u8 weight = 1;
|
| 175 |
|
| 176 | rcu_read_lock();
|
| 177 | state = rcu_dereference(wi_state);
|
| 178 | if (state)
|
| 179 | weight = state->iw_table[node];
|
| 180 | rcu_read_unlock();
|
| 181 | return weight;
|
| 182 | }
|
| 183 |
|
| 184 | |
| 185 | * Convert bandwidth values into weighted interleave weights.
|
| 186 | * Call with wi_state_lock.
|
| 187 |
|
| 188 | static void reduce_interleave_weights(unsigned int *bw, u8 *new_iw)
|
| 189 | {
|
| 190 | u64 sum_bw = 0;
|
| 191 | unsigned int cast_sum_bw, scaling_factor = 1, iw_gcd = 0;
|
| 192 | int nid;
|
| 193 |
|
| 194 | for_each_node_state(nid, N_MEMORY)
|
| 195 | sum_bw += bw[nid];
|
| 196 |
|
| 197 |
|
| 198 | for_each_node_state(nid, N_MEMORY) {
|
| 199 | |
| 200 | * Try not to perform 64-bit division.
|
| 201 | * If sum_bw < scaling_factor, then sum_bw < U32_MAX.
|
| 202 | * If sum_bw > scaling_factor, then round the weight up to 1.
|
| 203 |
|
| 204 | scaling_factor = weightiness * bw[nid];
|
| 205 | if (bw[nid] && sum_bw < scaling_factor) {
|
| 206 | cast_sum_bw = (unsigned int)sum_bw;
|
| 207 | new_iw[nid] = scaling_factor / cast_sum_bw;
|
| 208 | } else {
|
| 209 | new_iw[nid] = 1;
|
| 210 | }
|
| 211 | if (!iw_gcd)
|
| 212 | iw_gcd = new_iw[nid];
|
| 213 | iw_gcd = gcd(iw_gcd, new_iw[nid]);
|
| 214 | }
|
| 215 |
|
| 216 |
|
| 217 | for_each_node_state(nid, N_MEMORY)
|
| 218 | new_iw[nid] /= iw_gcd;
|
| 219 | }
|
| 220 |
|
| 221 | int mempolicy_set_node_perf(unsigned int node, struct access_coordinate *coords)
|
| 222 | {
|
| 223 | struct weighted_interleave_state *new_wi_state, *old_wi_state = NULL;
|
| 224 | unsigned int *old_bw, *new_bw;
|
| 225 | unsigned int bw_val;
|
| 226 | int i;
|
| 227 |
|
| 228 | bw_val = min(coords->read_bandwidth, coords->write_bandwidth);
|
| 229 | new_bw = kcalloc(nr_node_ids, sizeof(unsigned int), GFP_KERNEL);
|
| 230 | if (!new_bw)
|
| 231 | return -ENOMEM;
|
| 232 |
|
| 233 | new_wi_state = kmalloc_flex(*new_wi_state, iw_table, nr_node_ids);
|
| 234 | if (!new_wi_state) {
|
| 235 | kfree(new_bw);
|
| 236 | return -ENOMEM;
|
| 237 | }
|
| 238 | new_wi_state->mode_auto = true;
|
| 239 | for (i = 0; i < nr_node_ids; i++)
|
| 240 | new_wi_state->iw_table[i] = 1;
|
| 241 |
|
| 242 | |
| 243 | * Update bandwidth info, even in manual mode. That way, when switching
|
| 244 | * to auto mode in the future, iw_table can be overwritten using
|
| 245 | * accurate bw data.
|
| 246 |
|
| 247 | mutex_lock(&wi_state_lock);
|
| 248 |
|
| 249 | old_bw = node_bw_table;
|
| 250 | if (old_bw)
|
| 251 | memcpy(new_bw, old_bw, nr_node_ids * sizeof(*old_bw));
|
| 252 | new_bw[node] = bw_val;
|
| 253 | node_bw_table = new_bw;
|
| 254 |
|
| 255 | old_wi_state = rcu_dereference_protected(wi_state,
|
| 256 | lockdep_is_held(&wi_state_lock));
|
| 257 | if (old_wi_state && !old_wi_state->mode_auto) {
|
| 258 |
|
| 259 | mutex_unlock(&wi_state_lock);
|
| 260 | kfree(new_wi_state);
|
| 261 | goto out;
|
| 262 | }
|
| 263 |
|
| 264 |
|
| 265 | reduce_interleave_weights(new_bw, new_wi_state->iw_table);
|
| 266 | rcu_assign_pointer(wi_state, new_wi_state);
|
| 267 |
|
| 268 | mutex_unlock(&wi_state_lock);
|
| 269 | if (old_wi_state) {
|
| 270 | synchronize_rcu();
|
| 271 | kfree(old_wi_state);
|
| 272 | }
|
| 273 | out:
|
| 274 | kfree(old_bw);
|
| 275 | return 0;
|
| 276 | }
|
| 277 |
|
| 278 | |
| 279 | * numa_nearest_node - Find nearest node by state
|
| 280 | * @node: Node id to start the search
|
| 281 | * @state: State to filter the search
|
| 282 | *
|
| 283 | * Lookup the closest node by distance if @nid is not in state.
|
| 284 | *
|
| 285 | * Return: this @node if it is in state, otherwise the closest node by distance
|
| 286 |
|
| 287 | int numa_nearest_node(int node, unsigned int state)
|
| 288 | {
|
| 289 | int min_dist = INT_MAX, dist, n, min_node;
|
| 290 |
|
| 291 | if (state >= NR_NODE_STATES)
|
| 292 | return -EINVAL;
|
| 293 |
|
| 294 | if (node == NUMA_NO_NODE || node_state(node, state))
|
| 295 | return node;
|
| 296 |
|
| 297 | min_node = node;
|
| 298 | for_each_node_state(n, state) {
|
| 299 | dist = node_distance(node, n);
|
| 300 | if (dist < min_dist) {
|
| 301 | min_dist = dist;
|
| 302 | min_node = n;
|
| 303 | }
|
| 304 | }
|
| 305 |
|
| 306 | return min_node;
|
| 307 | }
|
| 308 | EXPORT_SYMBOL_GPL(numa_nearest_node);
|
| 309 |
|
| 310 | |
| 311 | * nearest_node_nodemask - Find the node in @mask at the nearest distance
|
| 312 | * from @node.
|
| 313 | *
|
| 314 | * @node: a valid node ID to start the search from.
|
| 315 | * @mask: a pointer to a nodemask representing the allowed nodes.
|
| 316 | *
|
| 317 | * This function iterates over all nodes in @mask and calculates the
|
| 318 | * distance from the starting @node, then it returns the node ID that is
|
| 319 | * the closest to @node, or MAX_NUMNODES if no node is found.
|
| 320 | *
|
| 321 | * Note that @node must be a valid node ID usable with node_distance(),
|
| 322 | * providing an invalid node ID (e.g., NUMA_NO_NODE) may result in crashes
|
| 323 | * or unexpected behavior.
|
| 324 |
|
| 325 | int nearest_node_nodemask(int node, nodemask_t *mask)
|
| 326 | {
|
| 327 | int dist, n, min_dist = INT_MAX, min_node = MAX_NUMNODES;
|
| 328 |
|
| 329 | for_each_node_mask(n, *mask) {
|
| 330 | dist = node_distance(node, n);
|
| 331 | if (dist < min_dist) {
|
| 332 | min_dist = dist;
|
| 333 | min_node = n;
|
| 334 | }
|
| 335 | }
|
| 336 |
|
| 337 | return min_node;
|
| 338 | }
|
| 339 | EXPORT_SYMBOL_GPL(nearest_node_nodemask);
|
| 340 |
|
| 341 | struct mempolicy *get_task_policy(struct task_struct *p)
|
| 342 | {
|
| 343 | struct mempolicy *pol = p->mempolicy;
|
| 344 | int node;
|
| 345 |
|
| 346 | if (pol)
|
| 347 | return pol;
|
| 348 |
|
| 349 | node = numa_node_id();
|
| 350 | if (node != NUMA_NO_NODE) {
|
| 351 | pol = &preferred_node_policy[node];
|
| 352 |
|
| 353 | if (pol->mode)
|
| 354 | return pol;
|
| 355 | }
|
| 356 |
|
| 357 | return &default_policy;
|
| 358 | }
|
| 359 | EXPORT_SYMBOL_FOR_MODULES(get_task_policy, "kvm");
|
| 360 |
|
| 361 | static const struct mempolicy_operations {
|
| 362 | int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
|
| 363 | void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes);
|
| 364 | } mpol_ops[MPOL_MAX];
|
| 365 |
|
| 366 | static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
|
| 367 | {
|
| 368 | return pol->flags & MPOL_USER_NODEMASK_FLAGS;
|
| 369 | }
|
| 370 |
|
| 371 | static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
|
| 372 | const nodemask_t *rel)
|
| 373 | {
|
| 374 | nodemask_t tmp;
|
| 375 | nodes_fold(tmp, *orig, nodes_weight(*rel));
|
| 376 | nodes_onto(*ret, tmp, *rel);
|
| 377 | }
|
| 378 |
|
| 379 | static int mpol_new_nodemask(struct mempolicy *pol, const nodemask_t *nodes)
|
| 380 | {
|
| 381 | if (nodes_empty(*nodes))
|
| 382 | return -EINVAL;
|
| 383 | pol->nodes = *nodes;
|
| 384 | return 0;
|
| 385 | }
|
| 386 |
|
| 387 | static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
|
| 388 | {
|
| 389 | if (nodes_empty(*nodes))
|
| 390 | return -EINVAL;
|
| 391 |
|
| 392 | nodes_clear(pol->nodes);
|
| 393 | node_set(first_node(*nodes), pol->nodes);
|
| 394 | return 0;
|
| 395 | }
|
| 396 |
|
| 397 | |
| 398 | * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
|
| 399 | * any, for the new policy. mpol_new() has already validated the nodes
|
| 400 | * parameter with respect to the policy mode and flags.
|
| 401 | *
|
| 402 | * Must be called holding task's alloc_lock to protect task's mems_allowed
|
| 403 | * and mempolicy. May also be called holding the mmap_lock for write.
|
| 404 |
|
| 405 | static int mpol_set_nodemask(struct mempolicy *pol,
|
| 406 | const nodemask_t *nodes, struct nodemask_scratch *nsc)
|
| 407 | {
|
| 408 | int ret;
|
| 409 |
|
| 410 | |
| 411 | * Default (pol==NULL) resp. local memory policies are not a
|
| 412 | * subject of any remapping. They also do not need any special
|
| 413 | * constructor.
|
| 414 |
|
| 415 | if (!pol || pol->mode == MPOL_LOCAL)
|
| 416 | return 0;
|
| 417 |
|
| 418 |
|
| 419 | nodes_and(nsc->mask1,
|
| 420 | cpuset_current_mems_allowed, node_states[N_MEMORY]);
|
| 421 |
|
| 422 | VM_BUG_ON(!nodes);
|
| 423 |
|
| 424 | if (pol->flags & MPOL_F_RELATIVE_NODES)
|
| 425 | mpol_relative_nodemask(&nsc->mask2, nodes, &nsc->mask1);
|
| 426 | else
|
| 427 | nodes_and(nsc->mask2, *nodes, nsc->mask1);
|
| 428 |
|
| 429 | if (mpol_store_user_nodemask(pol))
|
| 430 | pol->w.user_nodemask = *nodes;
|
| 431 | else
|
| 432 | pol->w.cpuset_mems_allowed = cpuset_current_mems_allowed;
|
| 433 |
|
| 434 | ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
|
| 435 | return ret;
|
| 436 | }
|
| 437 |
|
| 438 | |
| 439 | * This function just creates a new policy, does some check and simple
|
| 440 | * initialization. You must invoke mpol_set_nodemask() to set nodes.
|
| 441 |
|
| 442 | static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
|
| 443 | nodemask_t *nodes)
|
| 444 | {
|
| 445 | struct mempolicy *policy;
|
| 446 |
|
| 447 | if (mode == MPOL_DEFAULT) {
|
| 448 | if (nodes && !nodes_empty(*nodes))
|
| 449 | return ERR_PTR(-EINVAL);
|
| 450 | return NULL;
|
| 451 | }
|
| 452 | VM_BUG_ON(!nodes);
|
| 453 |
|
| 454 | |
| 455 | * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
|
| 456 | * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
|
| 457 | * All other modes require a valid pointer to a non-empty nodemask.
|
| 458 |
|
| 459 | if (mode == MPOL_PREFERRED) {
|
| 460 | if (nodes_empty(*nodes)) {
|
| 461 | if (((flags & MPOL_F_STATIC_NODES) ||
|
| 462 | (flags & MPOL_F_RELATIVE_NODES)))
|
| 463 | return ERR_PTR(-EINVAL);
|
| 464 |
|
| 465 | mode = MPOL_LOCAL;
|
| 466 | }
|
| 467 | } else if (mode == MPOL_LOCAL) {
|
| 468 | if (!nodes_empty(*nodes) ||
|
| 469 | (flags & MPOL_F_STATIC_NODES) ||
|
| 470 | (flags & MPOL_F_RELATIVE_NODES))
|
| 471 | return ERR_PTR(-EINVAL);
|
| 472 | } else if (nodes_empty(*nodes))
|
| 473 | return ERR_PTR(-EINVAL);
|
| 474 |
|
| 475 | policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
|
| 476 | if (!policy)
|
| 477 | return ERR_PTR(-ENOMEM);
|
| 478 | atomic_set(&policy->refcnt, 1);
|
| 479 | policy->mode = mode;
|
| 480 | policy->flags = flags;
|
| 481 | policy->home_node = NUMA_NO_NODE;
|
| 482 |
|
| 483 | return policy;
|
| 484 | }
|
| 485 |
|
| 486 |
|
| 487 | void __mpol_put(struct mempolicy *pol)
|
| 488 | {
|
| 489 | if (!atomic_dec_and_test(&pol->refcnt))
|
| 490 | return;
|
| 491 | |
| 492 | * Required to allow mmap_lock_speculative*() access, see for example
|
| 493 | * futex_key_to_node_opt(). All accesses are serialized by mmap_lock,
|
| 494 | * however the speculative lock section unbound by the normal lock
|
| 495 | * boundaries, requiring RCU freeing.
|
| 496 |
|
| 497 | kfree_rcu(pol, rcu);
|
| 498 | }
|
| 499 | EXPORT_SYMBOL_FOR_MODULES(__mpol_put, "kvm");
|
| 500 |
|
| 501 | static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes)
|
| 502 | {
|
| 503 | }
|
| 504 |
|
| 505 | static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes)
|
| 506 | {
|
| 507 | nodemask_t tmp;
|
| 508 |
|
| 509 | if (pol->flags & MPOL_F_STATIC_NODES)
|
| 510 | nodes_and(tmp, pol->w.user_nodemask, *nodes);
|
| 511 | else if (pol->flags & MPOL_F_RELATIVE_NODES)
|
| 512 | mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
|
| 513 | else {
|
| 514 | nodes_remap(tmp, pol->nodes, pol->w.cpuset_mems_allowed,
|
| 515 | *nodes);
|
| 516 | pol->w.cpuset_mems_allowed = *nodes;
|
| 517 | }
|
| 518 |
|
| 519 | if (nodes_empty(tmp))
|
| 520 | tmp = *nodes;
|
| 521 |
|
| 522 | pol->nodes = tmp;
|
| 523 | }
|
| 524 |
|
| 525 | static void mpol_rebind_preferred(struct mempolicy *pol,
|
| 526 | const nodemask_t *nodes)
|
| 527 | {
|
| 528 | pol->w.cpuset_mems_allowed = *nodes;
|
| 529 | }
|
| 530 |
|
| 531 | |
| 532 | * mpol_rebind_policy - Migrate a policy to a different set of nodes
|
| 533 | *
|
| 534 | * Per-vma policies are protected by mmap_lock. Allocations using per-task
|
| 535 | * policies are protected by task->mems_allowed_seq to prevent a premature
|
| 536 | * OOM/allocation failure due to parallel nodemask modification.
|
| 537 |
|
| 538 | static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask)
|
| 539 | {
|
| 540 | if (!pol || pol->mode == MPOL_LOCAL)
|
| 541 | return;
|
| 542 | if (!mpol_store_user_nodemask(pol) &&
|
| 543 | nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
|
| 544 | return;
|
| 545 |
|
| 546 | mpol_ops[pol->mode].rebind(pol, newmask);
|
| 547 | }
|
| 548 |
|
| 549 | |
| 550 | * Wrapper for mpol_rebind_policy() that just requires task
|
| 551 | * pointer, and updates task mempolicy.
|
| 552 | *
|
| 553 | * Called with task's alloc_lock held.
|
| 554 |
|
| 555 | void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new)
|
| 556 | {
|
| 557 | mpol_rebind_policy(tsk->mempolicy, new);
|
| 558 | }
|
| 559 |
|
| 560 | |
| 561 | * Rebind each vma in mm to new nodemask.
|
| 562 | *
|
| 563 | * Call holding a reference to mm. Takes mm->mmap_lock during call.
|
| 564 |
|
| 565 | void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
|
| 566 | {
|
| 567 | struct vm_area_struct *vma;
|
| 568 | VMA_ITERATOR(vmi, mm, 0);
|
| 569 |
|
| 570 | mmap_write_lock(mm);
|
| 571 | for_each_vma(vmi, vma) {
|
| 572 | vma_start_write(vma);
|
| 573 | mpol_rebind_policy(vma->vm_policy, new);
|
| 574 | }
|
| 575 | mmap_write_unlock(mm);
|
| 576 | }
|
| 577 |
|
| 578 | static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
|
| 579 | [MPOL_DEFAULT] = {
|
| 580 | .rebind = mpol_rebind_default,
|
| 581 | },
|
| 582 | [MPOL_INTERLEAVE] = {
|
| 583 | .create = mpol_new_nodemask,
|
| 584 | .rebind = mpol_rebind_nodemask,
|
| 585 | },
|
| 586 | [MPOL_PREFERRED] = {
|
| 587 | .create = mpol_new_preferred,
|
| 588 | .rebind = mpol_rebind_preferred,
|
| 589 | },
|
| 590 | [MPOL_BIND] = {
|
| 591 | .create = mpol_new_nodemask,
|
| 592 | .rebind = mpol_rebind_nodemask,
|
| 593 | },
|
| 594 | [MPOL_LOCAL] = {
|
| 595 | .rebind = mpol_rebind_default,
|
| 596 | },
|
| 597 | [MPOL_PREFERRED_MANY] = {
|
| 598 | .create = mpol_new_nodemask,
|
| 599 | .rebind = mpol_rebind_preferred,
|
| 600 | },
|
| 601 | [MPOL_WEIGHTED_INTERLEAVE] = {
|
| 602 | .create = mpol_new_nodemask,
|
| 603 | .rebind = mpol_rebind_nodemask,
|
| 604 | },
|
| 605 | };
|
| 606 |
|
| 607 | static bool migrate_folio_add(struct folio *folio, struct list_head *foliolist,
|
| 608 | unsigned long flags);
|
| 609 | static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *pol,
|
| 610 | pgoff_t ilx, int *nid);
|
| 611 |
|
| 612 | static bool strictly_unmovable(unsigned long flags)
|
| 613 | {
|
| 614 | |
| 615 | * STRICT without MOVE flags lets do_mbind() fail immediately with -EIO
|
| 616 | * if any misplaced page is found.
|
| 617 |
|
| 618 | return (flags & (MPOL_MF_STRICT | MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) ==
|
| 619 | MPOL_MF_STRICT;
|
| 620 | }
|
| 621 |
|
| 622 | struct migration_mpol {
|
| 623 | struct mempolicy *pol;
|
| 624 | pgoff_t ilx;
|
| 625 | };
|
| 626 |
|
| 627 | struct queue_pages {
|
| 628 | struct list_head *pagelist;
|
| 629 | unsigned long flags;
|
| 630 | nodemask_t *nmask;
|
| 631 | unsigned long start;
|
| 632 | unsigned long end;
|
| 633 | struct vm_area_struct *first;
|
| 634 | struct folio *large;
|
| 635 | long nr_failed;
|
| 636 | };
|
| 637 |
|
| 638 | |
| 639 | * Check if the folio's nid is in qp->nmask.
|
| 640 | *
|
| 641 | * If MPOL_MF_INVERT is set in qp->flags, check if the nid is
|
| 642 | * in the invert of qp->nmask.
|
| 643 |
|
| 644 | static inline bool queue_folio_required(struct folio *folio,
|
| 645 | struct queue_pages *qp)
|
| 646 | {
|
| 647 | int nid = folio_nid(folio);
|
| 648 | unsigned long flags = qp->flags;
|
| 649 |
|
| 650 | return node_isset(nid, *qp->nmask) == !(flags & MPOL_MF_INVERT);
|
| 651 | }
|
| 652 |
|
| 653 | static void queue_folios_pmd(pmd_t *pmd, struct mm_walk *walk)
|
| 654 | {
|
| 655 | struct folio *folio;
|
| 656 | struct queue_pages *qp = walk->private;
|
| 657 | pmd_t pmdval = pmdp_get(pmd);
|
| 658 |
|
| 659 | if (unlikely(!pmd_present(pmdval))) {
|
| 660 | if (pmd_is_migration_entry(pmdval))
|
| 661 | qp->nr_failed++;
|
| 662 | return;
|
| 663 | }
|
| 664 | folio = pmd_folio(pmdval);
|
| 665 | if (is_huge_zero_folio(folio)) {
|
| 666 | walk->action = ACTION_CONTINUE;
|
| 667 | return;
|
| 668 | }
|
| 669 | if (!queue_folio_required(folio, qp))
|
| 670 | return;
|
| 671 | if (!(qp->flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) ||
|
| 672 | !vma_migratable(walk->vma) ||
|
| 673 | !migrate_folio_add(folio, qp->pagelist, qp->flags))
|
| 674 | qp->nr_failed++;
|
| 675 | }
|
| 676 |
|
| 677 | |
| 678 | * Scan through folios, checking if they satisfy the required conditions,
|
| 679 | * moving them from LRU to local pagelist for migration if they do (or not).
|
| 680 | *
|
| 681 | * queue_folios_pte_range() has two possible return values:
|
| 682 | * 0 - continue walking to scan for more, even if an existing folio on the
|
| 683 | * wrong node could not be isolated and queued for migration.
|
| 684 | * -EIO - only MPOL_MF_STRICT was specified, without MPOL_MF_MOVE or ..._ALL,
|
| 685 | * and an existing folio was on a node that does not follow the policy.
|
| 686 |
|
| 687 | static int queue_folios_pte_range(pmd_t *pmd, unsigned long addr,
|
| 688 | unsigned long end, struct mm_walk *walk)
|
| 689 | {
|
| 690 | struct vm_area_struct *vma = walk->vma;
|
| 691 | struct folio *folio;
|
| 692 | struct queue_pages *qp = walk->private;
|
| 693 | unsigned long flags = qp->flags;
|
| 694 | pte_t *pte, *mapped_pte;
|
| 695 | pte_t ptent;
|
| 696 | spinlock_t *ptl;
|
| 697 | int max_nr, nr;
|
| 698 |
|
| 699 | ptl = pmd_trans_huge_lock(pmd, vma);
|
| 700 | if (ptl) {
|
| 701 | queue_folios_pmd(pmd, walk);
|
| 702 | spin_unlock(ptl);
|
| 703 | goto out;
|
| 704 | }
|
| 705 |
|
| 706 | mapped_pte = pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
|
| 707 | if (!pte) {
|
| 708 | walk->action = ACTION_AGAIN;
|
| 709 | return 0;
|
| 710 | }
|
| 711 | for (; addr != end; pte += nr, addr += nr * PAGE_SIZE) {
|
| 712 | max_nr = (end - addr) >> PAGE_SHIFT;
|
| 713 | nr = 1;
|
| 714 | ptent = ptep_get(pte);
|
| 715 | if (pte_none(ptent))
|
| 716 | continue;
|
| 717 | if (!pte_present(ptent)) {
|
| 718 | const softleaf_t entry = softleaf_from_pte(ptent);
|
| 719 |
|
| 720 | if (softleaf_is_migration(entry))
|
| 721 | qp->nr_failed++;
|
| 722 | continue;
|
| 723 | }
|
| 724 | folio = vm_normal_folio(vma, addr, ptent);
|
| 725 | if (!folio || folio_is_zone_device(folio))
|
| 726 | continue;
|
| 727 | if (folio_test_large(folio) && max_nr != 1)
|
| 728 | nr = folio_pte_batch(folio, pte, ptent, max_nr);
|
| 729 | |
| 730 | * vm_normal_folio() filters out zero pages, but there might
|
| 731 | * still be reserved folios to skip, perhaps in a VDSO.
|
| 732 |
|
| 733 | if (folio_test_reserved(folio))
|
| 734 | continue;
|
| 735 | if (!queue_folio_required(folio, qp))
|
| 736 | continue;
|
| 737 | if (folio_test_large(folio)) {
|
| 738 | |
| 739 | * A large folio can only be isolated from LRU once,
|
| 740 | * but may be mapped by many PTEs (and Copy-On-Write may
|
| 741 | * intersperse PTEs of other, order 0, folios). This is
|
| 742 | * a common case, so don't mistake it for failure (but
|
| 743 | * there can be other cases of multi-mapped pages which
|
| 744 | * this quick check does not help to filter out - and a
|
| 745 | * search of the pagelist might grow to be prohibitive).
|
| 746 | *
|
| 747 | * migrate_pages(&pagelist) returns nr_failed folios, so
|
| 748 | * check "large" now so that queue_pages_range() returns
|
| 749 | * a comparable nr_failed folios. This does imply that
|
| 750 | * if folio could not be isolated for some racy reason
|
| 751 | * at its first PTE, later PTEs will not give it another
|
| 752 | * chance of isolation; but keeps the accounting simple.
|
| 753 |
|
| 754 | if (folio == qp->large)
|
| 755 | continue;
|
| 756 | qp->large = folio;
|
| 757 | }
|
| 758 | if (!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) ||
|
| 759 | !vma_migratable(vma) ||
|
| 760 | !migrate_folio_add(folio, qp->pagelist, flags)) {
|
| 761 | qp->nr_failed += nr;
|
| 762 | if (strictly_unmovable(flags))
|
| 763 | break;
|
| 764 | }
|
| 765 | }
|
| 766 | pte_unmap_unlock(mapped_pte, ptl);
|
| 767 | cond_resched();
|
| 768 | out:
|
| 769 | if (qp->nr_failed && strictly_unmovable(flags))
|
| 770 | return -EIO;
|
| 771 | return 0;
|
| 772 | }
|
| 773 |
|
| 774 | static int queue_folios_hugetlb(pte_t *pte, unsigned long hmask,
|
| 775 | unsigned long addr, unsigned long end,
|
| 776 | struct mm_walk *walk)
|
| 777 | {
|
| 778 | #ifdef CONFIG_HUGETLB_PAGE
|
| 779 | struct queue_pages *qp = walk->private;
|
| 780 | unsigned long flags = qp->flags;
|
| 781 | struct folio *folio;
|
| 782 | spinlock_t *ptl;
|
| 783 | pte_t ptep;
|
| 784 |
|
| 785 | ptl = huge_pte_lock(hstate_vma(walk->vma), walk->mm, pte);
|
| 786 | ptep = huge_ptep_get(walk->mm, addr, pte);
|
| 787 | if (!pte_present(ptep)) {
|
| 788 | if (!huge_pte_none(ptep)) {
|
| 789 | const softleaf_t entry = softleaf_from_pte(ptep);
|
| 790 |
|
| 791 | if (unlikely(softleaf_is_migration(entry)))
|
| 792 | qp->nr_failed++;
|
| 793 | }
|
| 794 |
|
| 795 | goto unlock;
|
| 796 | }
|
| 797 | folio = pfn_folio(pte_pfn(ptep));
|
| 798 | if (!queue_folio_required(folio, qp))
|
| 799 | goto unlock;
|
| 800 | if (!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) ||
|
| 801 | !vma_migratable(walk->vma)) {
|
| 802 | qp->nr_failed++;
|
| 803 | goto unlock;
|
| 804 | }
|
| 805 | |
| 806 | * Unless MPOL_MF_MOVE_ALL, we try to avoid migrating a shared folio.
|
| 807 | * Choosing not to migrate a shared folio is not counted as a failure.
|
| 808 | *
|
| 809 | * See folio_maybe_mapped_shared() on possible imprecision when we
|
| 810 | * cannot easily detect if a folio is shared.
|
| 811 |
|
| 812 | if ((flags & MPOL_MF_MOVE_ALL) ||
|
| 813 | (!folio_maybe_mapped_shared(folio) && !hugetlb_pmd_shared(pte)))
|
| 814 | if (!folio_isolate_hugetlb(folio, qp->pagelist))
|
| 815 | qp->nr_failed++;
|
| 816 | unlock:
|
| 817 | spin_unlock(ptl);
|
| 818 | if (qp->nr_failed && strictly_unmovable(flags))
|
| 819 | return -EIO;
|
| 820 | #endif
|
| 821 | return 0;
|
| 822 | }
|
| 823 |
|
| 824 | #ifdef CONFIG_NUMA_BALANCING
|
| 825 | |
| 826 | * folio_can_map_prot_numa() - check whether the folio can map prot numa
|
| 827 | * @folio: The folio whose mapping considered for being made NUMA hintable
|
| 828 | * @vma: The VMA that the folio belongs to.
|
| 829 | * @is_private_single_threaded: Is this a single-threaded private VMA or not
|
| 830 | *
|
| 831 | * This function checks to see if the folio actually indicates that
|
| 832 | * we need to make the mapping one which causes a NUMA hinting fault,
|
| 833 | * as there are cases where it's simply unnecessary, and the folio's
|
| 834 | * access time is adjusted for memory tiering if prot numa needed.
|
| 835 | *
|
| 836 | * Return: True if the mapping of the folio needs to be changed, false otherwise.
|
| 837 |
|
| 838 | bool folio_can_map_prot_numa(struct folio *folio, struct vm_area_struct *vma,
|
| 839 | bool is_private_single_threaded)
|
| 840 | {
|
| 841 | int nid;
|
| 842 |
|
| 843 | if (!folio || folio_is_zone_device(folio) || folio_test_ksm(folio))
|
| 844 | return false;
|
| 845 |
|
| 846 |
|
| 847 | if (vma_is_cow_mapping(vma) && folio_maybe_mapped_shared(folio))
|
| 848 | return false;
|
| 849 |
|
| 850 |
|
| 851 | if (folio_maybe_dma_pinned(folio))
|
| 852 | return false;
|
| 853 |
|
| 854 | |
| 855 | * While migration can move some dirty folios,
|
| 856 | * it cannot move them all from MIGRATE_ASYNC
|
| 857 | * context.
|
| 858 |
|
| 859 | if (folio_is_file_lru(folio) && folio_test_dirty(folio))
|
| 860 | return false;
|
| 861 |
|
| 862 | |
| 863 | * Don't mess with PTEs if folio is already on the node
|
| 864 | * a single-threaded process is running on.
|
| 865 |
|
| 866 | nid = folio_nid(folio);
|
| 867 | if (is_private_single_threaded && (nid == numa_node_id()))
|
| 868 | return false;
|
| 869 |
|
| 870 | |
| 871 | * Skip scanning top tier node if normal numa
|
| 872 | * balancing is disabled
|
| 873 |
|
| 874 | if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_NORMAL) &&
|
| 875 | node_is_toptier(nid))
|
| 876 | return false;
|
| 877 |
|
| 878 | if (folio_use_access_time(folio))
|
| 879 | folio_xchg_access_time(folio, jiffies_to_msecs(jiffies));
|
| 880 |
|
| 881 | return true;
|
| 882 | }
|
| 883 |
|
| 884 | |
| 885 | * This is used to mark a range of virtual addresses to be inaccessible.
|
| 886 | * These are later cleared by a NUMA hinting fault. Depending on these
|
| 887 | * faults, pages may be migrated for better NUMA placement.
|
| 888 | *
|
| 889 | * This is assuming that NUMA faults are handled using PROT_NONE. If
|
| 890 | * an architecture makes a different choice, it will need further
|
| 891 | * changes to the core.
|
| 892 |
|
| 893 | unsigned long change_prot_numa(struct vm_area_struct *vma,
|
| 894 | unsigned long addr, unsigned long end)
|
| 895 | {
|
| 896 | struct mmu_gather tlb;
|
| 897 | long nr_updated;
|
| 898 |
|
| 899 | tlb_gather_mmu(&tlb, vma->vm_mm);
|
| 900 |
|
| 901 | nr_updated = change_protection(&tlb, vma, addr, end, MM_CP_PROT_NUMA);
|
| 902 | if (nr_updated > 0) {
|
| 903 | count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
|
| 904 | count_memcg_events_mm(vma->vm_mm, NUMA_PTE_UPDATES, nr_updated);
|
| 905 | }
|
| 906 |
|
| 907 | tlb_finish_mmu(&tlb);
|
| 908 |
|
| 909 | return nr_updated;
|
| 910 | }
|
| 911 | #endif
|
| 912 |
|
| 913 | static int queue_pages_test_walk(unsigned long start, unsigned long end,
|
| 914 | struct mm_walk *walk)
|
| 915 | {
|
| 916 | struct vm_area_struct *next, *vma = walk->vma;
|
| 917 | struct queue_pages *qp = walk->private;
|
| 918 | unsigned long flags = qp->flags;
|
| 919 |
|
| 920 |
|
| 921 | VM_BUG_ON_VMA(!range_in_vma(vma, start, end), vma);
|
| 922 |
|
| 923 | if (!qp->first) {
|
| 924 | qp->first = vma;
|
| 925 | if (!(flags & MPOL_MF_DISCONTIG_OK) &&
|
| 926 | (qp->start < vma->vm_start))
|
| 927 |
|
| 928 | return -EFAULT;
|
| 929 | }
|
| 930 | next = find_vma(vma->vm_mm, vma->vm_end);
|
| 931 | if (!(flags & MPOL_MF_DISCONTIG_OK) &&
|
| 932 | ((vma->vm_end < qp->end) &&
|
| 933 | (!next || vma->vm_end < next->vm_start)))
|
| 934 |
|
| 935 | return -EFAULT;
|
| 936 |
|
| 937 | |
| 938 | * Need check MPOL_MF_STRICT to return -EIO if possible
|
| 939 | * regardless of vma_migratable
|
| 940 |
|
| 941 | if (!vma_migratable(vma) &&
|
| 942 | !(flags & MPOL_MF_STRICT))
|
| 943 | return 1;
|
| 944 |
|
| 945 | |
| 946 | * Check page nodes, and queue pages to move, in the current vma.
|
| 947 | * But if no moving, and no strict checking, the scan can be skipped.
|
| 948 |
|
| 949 | if (flags & (MPOL_MF_STRICT | MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
|
| 950 | return 0;
|
| 951 | return 1;
|
| 952 | }
|
| 953 |
|
| 954 | static const struct mm_walk_ops queue_pages_walk_ops = {
|
| 955 | .hugetlb_entry = queue_folios_hugetlb,
|
| 956 | .pmd_entry = queue_folios_pte_range,
|
| 957 | .test_walk = queue_pages_test_walk,
|
| 958 | .walk_lock = PGWALK_RDLOCK,
|
| 959 | };
|
| 960 |
|
| 961 | static const struct mm_walk_ops queue_pages_lock_vma_walk_ops = {
|
| 962 | .hugetlb_entry = queue_folios_hugetlb,
|
| 963 | .pmd_entry = queue_folios_pte_range,
|
| 964 | .test_walk = queue_pages_test_walk,
|
| 965 | .walk_lock = PGWALK_WRLOCK,
|
| 966 | };
|
| 967 |
|
| 968 | |
| 969 | * Walk through page tables and collect pages to be migrated.
|
| 970 | *
|
| 971 | * If pages found in a given range are not on the required set of @nodes,
|
| 972 | * and migration is allowed, they are isolated and queued to @pagelist.
|
| 973 | *
|
| 974 | * queue_pages_range() may return:
|
| 975 | * 0 - all pages already on the right node, or successfully queued for moving
|
| 976 | * (or neither strict checking nor moving requested: only range checking).
|
| 977 | * >0 - this number of misplaced folios could not be queued for moving
|
| 978 | * (a hugetlbfs page or a transparent huge page being counted as 1).
|
| 979 | * -EIO - a misplaced page found, when MPOL_MF_STRICT specified without MOVEs.
|
| 980 | * -EFAULT - a hole in the memory range, when MPOL_MF_DISCONTIG_OK unspecified.
|
| 981 |
|
| 982 | static long
|
| 983 | queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
|
| 984 | nodemask_t *nodes, unsigned long flags,
|
| 985 | struct list_head *pagelist)
|
| 986 | {
|
| 987 | int err;
|
| 988 | struct queue_pages qp = {
|
| 989 | .pagelist = pagelist,
|
| 990 | .flags = flags,
|
| 991 | .nmask = nodes,
|
| 992 | .start = start,
|
| 993 | .end = end,
|
| 994 | .first = NULL,
|
| 995 | };
|
| 996 | const struct mm_walk_ops *ops = (flags & MPOL_MF_WRLOCK) ?
|
| 997 | &queue_pages_lock_vma_walk_ops : &queue_pages_walk_ops;
|
| 998 |
|
| 999 | err = walk_page_range(mm, start, end, ops, &qp);
|
| 1000 |
|
| 1001 | if (!qp.first)
|
| 1002 |
|
| 1003 | err = -EFAULT;
|
| 1004 |
|
| 1005 | return err ? : qp.nr_failed;
|
| 1006 | }
|
| 1007 |
|
| 1008 | |
| 1009 | * Apply policy to a single VMA
|
| 1010 | * This must be called with the mmap_lock held for writing.
|
| 1011 |
|
| 1012 | static int vma_replace_policy(struct vm_area_struct *vma,
|
| 1013 | struct mempolicy *pol)
|
| 1014 | {
|
| 1015 | int err;
|
| 1016 | struct mempolicy *old;
|
| 1017 | struct mempolicy *new;
|
| 1018 |
|
| 1019 | vma_assert_write_locked(vma);
|
| 1020 |
|
| 1021 | new = mpol_dup(pol);
|
| 1022 | if (IS_ERR(new))
|
| 1023 | return PTR_ERR(new);
|
| 1024 |
|
| 1025 | if (vma->vm_ops && vma->vm_ops->set_policy) {
|
| 1026 | err = vma->vm_ops->set_policy(vma, new);
|
| 1027 | if (err)
|
| 1028 | goto err_out;
|
| 1029 | }
|
| 1030 |
|
| 1031 | old = vma->vm_policy;
|
| 1032 | WRITE_ONCE(vma->vm_policy, new);
|
| 1033 | mpol_put(old);
|
| 1034 |
|
| 1035 | return 0;
|
| 1036 | err_out:
|
| 1037 | mpol_put(new);
|
| 1038 | return err;
|
| 1039 | }
|
| 1040 |
|
| 1041 |
|
| 1042 | static int mbind_range(struct vma_iterator *vmi, struct vm_area_struct *vma,
|
| 1043 | struct vm_area_struct **prev, unsigned long start,
|
| 1044 | unsigned long end, struct mempolicy *new_pol)
|
| 1045 | {
|
| 1046 | unsigned long vmstart, vmend;
|
| 1047 |
|
| 1048 | vmend = min(end, vma->vm_end);
|
| 1049 | if (start > vma->vm_start) {
|
| 1050 | *prev = vma;
|
| 1051 | vmstart = start;
|
| 1052 | } else {
|
| 1053 | vmstart = vma->vm_start;
|
| 1054 | }
|
| 1055 |
|
| 1056 | if (mpol_equal(vma->vm_policy, new_pol)) {
|
| 1057 | *prev = vma;
|
| 1058 | return 0;
|
| 1059 | }
|
| 1060 |
|
| 1061 | vma = vma_modify_policy(vmi, *prev, vma, vmstart, vmend, new_pol);
|
| 1062 | if (IS_ERR(vma))
|
| 1063 | return PTR_ERR(vma);
|
| 1064 |
|
| 1065 | *prev = vma;
|
| 1066 | return vma_replace_policy(vma, new_pol);
|
| 1067 | }
|
| 1068 |
|
| 1069 |
|
| 1070 | static long do_set_mempolicy(unsigned short mode, unsigned short flags,
|
| 1071 | nodemask_t *nodes)
|
| 1072 | {
|
| 1073 | struct mempolicy *new, *old;
|
| 1074 | NODEMASK_SCRATCH(scratch);
|
| 1075 | int ret;
|
| 1076 |
|
| 1077 | if (!scratch)
|
| 1078 | return -ENOMEM;
|
| 1079 |
|
| 1080 | new = mpol_new(mode, flags, nodes);
|
| 1081 | if (IS_ERR(new)) {
|
| 1082 | ret = PTR_ERR(new);
|
| 1083 | goto out;
|
| 1084 | }
|
| 1085 |
|
| 1086 | task_lock(current);
|
| 1087 | ret = mpol_set_nodemask(new, nodes, scratch);
|
| 1088 | if (ret) {
|
| 1089 | task_unlock(current);
|
| 1090 | mpol_put(new);
|
| 1091 | goto out;
|
| 1092 | }
|
| 1093 |
|
| 1094 | old = current->mempolicy;
|
| 1095 | current->mempolicy = new;
|
| 1096 | if (new && (new->mode == MPOL_INTERLEAVE ||
|
| 1097 | new->mode == MPOL_WEIGHTED_INTERLEAVE)) {
|
| 1098 | current->il_prev = MAX_NUMNODES-1;
|
| 1099 | current->il_weight = 0;
|
| 1100 | }
|
| 1101 | task_unlock(current);
|
| 1102 | mpol_put(old);
|
| 1103 | ret = 0;
|
| 1104 | out:
|
| 1105 | NODEMASK_SCRATCH_FREE(scratch);
|
| 1106 | return ret;
|
| 1107 | }
|
| 1108 |
|
| 1109 | |
| 1110 | * Return nodemask for policy for get_mempolicy() query
|
| 1111 | *
|
| 1112 | * Called with task's alloc_lock held
|
| 1113 |
|
| 1114 | static void get_policy_nodemask(struct mempolicy *pol, nodemask_t *nodes)
|
| 1115 | {
|
| 1116 | nodes_clear(*nodes);
|
| 1117 | if (pol == &default_policy)
|
| 1118 | return;
|
| 1119 |
|
| 1120 | switch (pol->mode) {
|
| 1121 | case MPOL_BIND:
|
| 1122 | case MPOL_INTERLEAVE:
|
| 1123 | case MPOL_PREFERRED:
|
| 1124 | case MPOL_PREFERRED_MANY:
|
| 1125 | case MPOL_WEIGHTED_INTERLEAVE:
|
| 1126 | *nodes = pol->nodes;
|
| 1127 | break;
|
| 1128 | case MPOL_LOCAL:
|
| 1129 |
|
| 1130 | break;
|
| 1131 | default:
|
| 1132 | BUG();
|
| 1133 | }
|
| 1134 | }
|
| 1135 |
|
| 1136 | static int lookup_node(struct mm_struct *mm, unsigned long addr)
|
| 1137 | {
|
| 1138 | struct page *p = NULL;
|
| 1139 | int ret;
|
| 1140 |
|
| 1141 | ret = get_user_pages_fast(addr & PAGE_MASK, 1, 0, &p);
|
| 1142 | if (ret > 0) {
|
| 1143 | ret = page_to_nid(p);
|
| 1144 | put_page(p);
|
| 1145 | }
|
| 1146 | return ret;
|
| 1147 | }
|
| 1148 |
|
| 1149 |
|
| 1150 | static long do_get_mempolicy(int *policy, nodemask_t *nmask,
|
| 1151 | unsigned long addr, unsigned long flags)
|
| 1152 | {
|
| 1153 | int err;
|
| 1154 | struct mm_struct *mm = current->mm;
|
| 1155 | struct vm_area_struct *vma = NULL;
|
| 1156 | struct mempolicy *pol = current->mempolicy, *pol_refcount = NULL;
|
| 1157 |
|
| 1158 | if (flags &
|
| 1159 | ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
|
| 1160 | return -EINVAL;
|
| 1161 |
|
| 1162 | if (flags & MPOL_F_MEMS_ALLOWED) {
|
| 1163 | if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
|
| 1164 | return -EINVAL;
|
| 1165 | *policy = 0;
|
| 1166 | task_lock(current);
|
| 1167 | *nmask = cpuset_current_mems_allowed;
|
| 1168 | task_unlock(current);
|
| 1169 | return 0;
|
| 1170 | }
|
| 1171 |
|
| 1172 | if (flags & MPOL_F_ADDR) {
|
| 1173 | pgoff_t ilx;
|
| 1174 | |
| 1175 | * Do NOT fall back to task policy if the
|
| 1176 | * vma/shared policy at addr is NULL. We
|
| 1177 | * want to return MPOL_DEFAULT in this case.
|
| 1178 |
|
| 1179 | mmap_read_lock(mm);
|
| 1180 | vma = vma_lookup(mm, addr);
|
| 1181 | if (!vma) {
|
| 1182 | mmap_read_unlock(mm);
|
| 1183 | return -EFAULT;
|
| 1184 | }
|
| 1185 | pol = __get_vma_policy(vma, addr, &ilx);
|
| 1186 | } else if (addr)
|
| 1187 | return -EINVAL;
|
| 1188 |
|
| 1189 | if (!pol)
|
| 1190 | pol = &default_policy;
|
| 1191 |
|
| 1192 | if (flags & MPOL_F_NODE) {
|
| 1193 | if (flags & MPOL_F_ADDR) {
|
| 1194 | |
| 1195 | * Take a refcount on the mpol, because we are about to
|
| 1196 | * drop the mmap_lock, after which only "pol" remains
|
| 1197 | * valid, "vma" is stale.
|
| 1198 |
|
| 1199 | pol_refcount = pol;
|
| 1200 | vma = NULL;
|
| 1201 | mpol_get(pol);
|
| 1202 | mmap_read_unlock(mm);
|
| 1203 | err = lookup_node(mm, addr);
|
| 1204 | if (err < 0)
|
| 1205 | goto out;
|
| 1206 | *policy = err;
|
| 1207 | } else if (pol == current->mempolicy &&
|
| 1208 | pol->mode == MPOL_INTERLEAVE) {
|
| 1209 | *policy = next_node_in(current->il_prev, pol->nodes);
|
| 1210 | } else if (pol == current->mempolicy &&
|
| 1211 | pol->mode == MPOL_WEIGHTED_INTERLEAVE) {
|
| 1212 | if (current->il_weight)
|
| 1213 | *policy = current->il_prev;
|
| 1214 | else
|
| 1215 | *policy = next_node_in(current->il_prev,
|
| 1216 | pol->nodes);
|
| 1217 | } else {
|
| 1218 | err = -EINVAL;
|
| 1219 | goto out;
|
| 1220 | }
|
| 1221 | } else {
|
| 1222 | *policy = pol == &default_policy ? MPOL_DEFAULT :
|
| 1223 | pol->mode;
|
| 1224 | |
| 1225 | * Internal mempolicy flags must be masked off before exposing
|
| 1226 | * the policy to userspace.
|
| 1227 |
|
| 1228 | *policy |= (pol->flags & MPOL_MODE_FLAGS);
|
| 1229 | }
|
| 1230 |
|
| 1231 | err = 0;
|
| 1232 | if (nmask) {
|
| 1233 | if (mpol_store_user_nodemask(pol)) {
|
| 1234 | *nmask = pol->w.user_nodemask;
|
| 1235 | } else {
|
| 1236 | task_lock(current);
|
| 1237 | get_policy_nodemask(pol, nmask);
|
| 1238 | task_unlock(current);
|
| 1239 | }
|
| 1240 | }
|
| 1241 |
|
| 1242 | out:
|
| 1243 | mpol_cond_put(pol);
|
| 1244 | if (vma)
|
| 1245 | mmap_read_unlock(mm);
|
| 1246 | if (pol_refcount)
|
| 1247 | mpol_put(pol_refcount);
|
| 1248 | return err;
|
| 1249 | }
|
| 1250 |
|
| 1251 | #ifdef CONFIG_NUMA_MIGRATION
|
| 1252 | static bool migrate_folio_add(struct folio *folio, struct list_head *foliolist,
|
| 1253 | unsigned long flags)
|
| 1254 | {
|
| 1255 | |
| 1256 | * Unless MPOL_MF_MOVE_ALL, we try to avoid migrating a shared folio.
|
| 1257 | * Choosing not to migrate a shared folio is not counted as a failure.
|
| 1258 | *
|
| 1259 | * See folio_maybe_mapped_shared() on possible imprecision when we
|
| 1260 | * cannot easily detect if a folio is shared.
|
| 1261 |
|
| 1262 | if ((flags & MPOL_MF_MOVE_ALL) || !folio_maybe_mapped_shared(folio)) {
|
| 1263 | if (folio_isolate_lru(folio)) {
|
| 1264 | list_add_tail(&folio->lru, foliolist);
|
| 1265 | node_stat_mod_folio(folio,
|
| 1266 | NR_ISOLATED_ANON + folio_is_file_lru(folio),
|
| 1267 | folio_nr_pages(folio));
|
| 1268 | } else {
|
| 1269 | |
| 1270 | * Non-movable folio may reach here. And, there may be
|
| 1271 | * temporary off LRU folios or non-LRU movable folios.
|
| 1272 | * Treat them as unmovable folios since they can't be
|
| 1273 | * isolated, so they can't be moved at the moment.
|
| 1274 |
|
| 1275 | return false;
|
| 1276 | }
|
| 1277 | }
|
| 1278 | return true;
|
| 1279 | }
|
| 1280 |
|
| 1281 | |
| 1282 | * Migrate pages from one node to a target node.
|
| 1283 | * Returns error or the number of pages not migrated.
|
| 1284 |
|
| 1285 | static long migrate_to_node(struct mm_struct *mm, int source, int dest,
|
| 1286 | int flags)
|
| 1287 | {
|
| 1288 | nodemask_t nmask;
|
| 1289 | struct vm_area_struct *vma;
|
| 1290 | LIST_HEAD(pagelist);
|
| 1291 | long nr_failed;
|
| 1292 | long err = 0;
|
| 1293 | struct migration_target_control mtc = {
|
| 1294 | .nid = dest,
|
| 1295 | .gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE,
|
| 1296 | .reason = MR_SYSCALL,
|
| 1297 | };
|
| 1298 |
|
| 1299 | nodes_clear(nmask);
|
| 1300 | node_set(source, nmask);
|
| 1301 |
|
| 1302 | VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
|
| 1303 |
|
| 1304 | mmap_read_lock(mm);
|
| 1305 | vma = find_vma(mm, 0);
|
| 1306 | if (unlikely(!vma)) {
|
| 1307 | mmap_read_unlock(mm);
|
| 1308 | return 0;
|
| 1309 | }
|
| 1310 |
|
| 1311 | |
| 1312 | * This does not migrate the range, but isolates all pages that
|
| 1313 | * need migration. Between passing in the full user address
|
| 1314 | * space range and MPOL_MF_DISCONTIG_OK, this call cannot fail,
|
| 1315 | * but passes back the count of pages which could not be isolated.
|
| 1316 |
|
| 1317 | nr_failed = queue_pages_range(mm, vma->vm_start, mm->task_size, &nmask,
|
| 1318 | flags | MPOL_MF_DISCONTIG_OK, &pagelist);
|
| 1319 | mmap_read_unlock(mm);
|
| 1320 |
|
| 1321 | if (!list_empty(&pagelist)) {
|
| 1322 | err = migrate_pages(&pagelist, alloc_migration_target, NULL,
|
| 1323 | (unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL, NULL);
|
| 1324 | if (err)
|
| 1325 | putback_movable_pages(&pagelist);
|
| 1326 | }
|
| 1327 |
|
| 1328 | if (err >= 0)
|
| 1329 | err += nr_failed;
|
| 1330 | return err;
|
| 1331 | }
|
| 1332 |
|
| 1333 | |
| 1334 | * Move pages between the two nodesets so as to preserve the physical
|
| 1335 | * layout as much as possible.
|
| 1336 | *
|
| 1337 | * Returns the number of page that could not be moved.
|
| 1338 |
|
| 1339 | int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
|
| 1340 | const nodemask_t *to, int flags)
|
| 1341 | {
|
| 1342 | long nr_failed = 0;
|
| 1343 | long err = 0;
|
| 1344 | nodemask_t tmp;
|
| 1345 |
|
| 1346 | lru_cache_disable();
|
| 1347 |
|
| 1348 | |
| 1349 | * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
|
| 1350 | * bit in 'to' is not also set in 'tmp'. Clear the found 'source'
|
| 1351 | * bit in 'tmp', and return that <source, dest> pair for migration.
|
| 1352 | * The pair of nodemasks 'to' and 'from' define the map.
|
| 1353 | *
|
| 1354 | * If no pair of bits is found that way, fallback to picking some
|
| 1355 | * pair of 'source' and 'dest' bits that are not the same. If the
|
| 1356 | * 'source' and 'dest' bits are the same, this represents a node
|
| 1357 | * that will be migrating to itself, so no pages need move.
|
| 1358 | *
|
| 1359 | * If no bits are left in 'tmp', or if all remaining bits left
|
| 1360 | * in 'tmp' correspond to the same bit in 'to', return false
|
| 1361 | * (nothing left to migrate).
|
| 1362 | *
|
| 1363 | * This lets us pick a pair of nodes to migrate between, such that
|
| 1364 | * if possible the dest node is not already occupied by some other
|
| 1365 | * source node, minimizing the risk of overloading the memory on a
|
| 1366 | * node that would happen if we migrated incoming memory to a node
|
| 1367 | * before migrating outgoing memory source that same node.
|
| 1368 | *
|
| 1369 | * A single scan of tmp is sufficient. As we go, we remember the
|
| 1370 | * most recent <s, d> pair that moved (s != d). If we find a pair
|
| 1371 | * that not only moved, but what's better, moved to an empty slot
|
| 1372 | * (d is not set in tmp), then we break out then, with that pair.
|
| 1373 | * Otherwise when we finish scanning from_tmp, we at least have the
|
| 1374 | * most recent <s, d> pair that moved. If we get all the way through
|
| 1375 | * the scan of tmp without finding any node that moved, much less
|
| 1376 | * moved to an empty node, then there is nothing left worth migrating.
|
| 1377 |
|
| 1378 |
|
| 1379 | tmp = *from;
|
| 1380 | while (!nodes_empty(tmp)) {
|
| 1381 | int s, d;
|
| 1382 | int source = NUMA_NO_NODE;
|
| 1383 | int dest = 0;
|
| 1384 |
|
| 1385 | for_each_node_mask(s, tmp) {
|
| 1386 |
|
| 1387 | |
| 1388 | * do_migrate_pages() tries to maintain the relative
|
| 1389 | * node relationship of the pages established between
|
| 1390 | * threads and memory areas.
|
| 1391 | *
|
| 1392 | * However if the number of source nodes is not equal to
|
| 1393 | * the number of destination nodes we can not preserve
|
| 1394 | * this node relative relationship. In that case, skip
|
| 1395 | * copying memory from a node that is in the destination
|
| 1396 | * mask.
|
| 1397 | *
|
| 1398 | * Example: [2,3,4] -> [3,4,5] moves everything.
|
| 1399 | * [0-7] - > [3,4,5] moves only 0,1,2,6,7.
|
| 1400 |
|
| 1401 |
|
| 1402 | if ((nodes_weight(*from) != nodes_weight(*to)) &&
|
| 1403 | (node_isset(s, *to)))
|
| 1404 | continue;
|
| 1405 |
|
| 1406 | d = node_remap(s, *from, *to);
|
| 1407 | if (s == d)
|
| 1408 | continue;
|
| 1409 |
|
| 1410 | source = s;
|
| 1411 | dest = d;
|
| 1412 |
|
| 1413 |
|
| 1414 | if (!node_isset(dest, tmp))
|
| 1415 | break;
|
| 1416 | }
|
| 1417 | if (source == NUMA_NO_NODE)
|
| 1418 | break;
|
| 1419 |
|
| 1420 | node_clear(source, tmp);
|
| 1421 | err = migrate_to_node(mm, source, dest, flags);
|
| 1422 | if (err > 0)
|
| 1423 | nr_failed += err;
|
| 1424 | if (err < 0)
|
| 1425 | break;
|
| 1426 | }
|
| 1427 |
|
| 1428 | lru_cache_enable();
|
| 1429 | if (err < 0)
|
| 1430 | return err;
|
| 1431 | return (nr_failed < INT_MAX) ? nr_failed : INT_MAX;
|
| 1432 | }
|
| 1433 |
|
| 1434 | |
| 1435 | * Allocate a new folio for page migration, according to NUMA mempolicy.
|
| 1436 |
|
| 1437 | static struct folio *alloc_migration_target_by_mpol(struct folio *src,
|
| 1438 | unsigned long private)
|
| 1439 | {
|
| 1440 | struct migration_mpol *mmpol = (struct migration_mpol *)private;
|
| 1441 | struct mempolicy *pol = mmpol->pol;
|
| 1442 | pgoff_t ilx = mmpol->ilx;
|
| 1443 | unsigned int order;
|
| 1444 | int nid = numa_node_id();
|
| 1445 | gfp_t gfp;
|
| 1446 |
|
| 1447 | order = folio_order(src);
|
| 1448 | ilx += src->index >> order;
|
| 1449 |
|
| 1450 | if (folio_test_hugetlb(src)) {
|
| 1451 | nodemask_t *nodemask;
|
| 1452 | struct hstate *h;
|
| 1453 |
|
| 1454 | h = folio_hstate(src);
|
| 1455 | gfp = htlb_alloc_mask(h);
|
| 1456 | nodemask = policy_nodemask(gfp, pol, ilx, &nid);
|
| 1457 | return alloc_hugetlb_folio_nodemask(h, nid, nodemask, gfp,
|
| 1458 | htlb_allow_alloc_fallback(MR_MEMPOLICY_MBIND));
|
| 1459 | }
|
| 1460 |
|
| 1461 | if (folio_test_large(src))
|
| 1462 | gfp = GFP_TRANSHUGE;
|
| 1463 | else
|
| 1464 | gfp = GFP_HIGHUSER_MOVABLE | __GFP_RETRY_MAYFAIL | __GFP_COMP;
|
| 1465 |
|
| 1466 | return folio_alloc_mpol(gfp, order, pol, ilx, nid);
|
| 1467 | }
|
| 1468 | #else
|
| 1469 |
|
| 1470 | static bool migrate_folio_add(struct folio *folio, struct list_head *foliolist,
|
| 1471 | unsigned long flags)
|
| 1472 | {
|
| 1473 | return false;
|
| 1474 | }
|
| 1475 |
|
| 1476 | int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
|
| 1477 | const nodemask_t *to, int flags)
|
| 1478 | {
|
| 1479 | return -ENOSYS;
|
| 1480 | }
|
| 1481 |
|
| 1482 | static struct folio *alloc_migration_target_by_mpol(struct folio *src,
|
| 1483 | unsigned long private)
|
| 1484 | {
|
| 1485 | return NULL;
|
| 1486 | }
|
| 1487 | #endif
|
| 1488 |
|
| 1489 | static long do_mbind(unsigned long start, unsigned long len,
|
| 1490 | unsigned short mode, unsigned short mode_flags,
|
| 1491 | nodemask_t *nmask, unsigned long flags)
|
| 1492 | {
|
| 1493 | struct mm_struct *mm = current->mm;
|
| 1494 | struct vm_area_struct *vma, *prev;
|
| 1495 | struct vma_iterator vmi;
|
| 1496 | struct migration_mpol mmpol;
|
| 1497 | struct mempolicy *new;
|
| 1498 | unsigned long end;
|
| 1499 | long err;
|
| 1500 | long nr_failed;
|
| 1501 | LIST_HEAD(pagelist);
|
| 1502 |
|
| 1503 | if (flags & ~(unsigned long)MPOL_MF_VALID)
|
| 1504 | return -EINVAL;
|
| 1505 | if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
|
| 1506 | return -EPERM;
|
| 1507 |
|
| 1508 | if (start & ~PAGE_MASK)
|
| 1509 | return -EINVAL;
|
| 1510 |
|
| 1511 | if (mode == MPOL_DEFAULT)
|
| 1512 | flags &= ~MPOL_MF_STRICT;
|
| 1513 |
|
| 1514 | len = PAGE_ALIGN(len);
|
| 1515 | end = start + len;
|
| 1516 |
|
| 1517 | if (end < start)
|
| 1518 | return -EINVAL;
|
| 1519 | if (end == start)
|
| 1520 | return 0;
|
| 1521 |
|
| 1522 | new = mpol_new(mode, mode_flags, nmask);
|
| 1523 | if (IS_ERR(new))
|
| 1524 | return PTR_ERR(new);
|
| 1525 |
|
| 1526 | |
| 1527 | * If we are using the default policy then operation
|
| 1528 | * on discontinuous address spaces is okay after all
|
| 1529 |
|
| 1530 | if (!new)
|
| 1531 | flags |= MPOL_MF_DISCONTIG_OK;
|
| 1532 |
|
| 1533 | if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
|
| 1534 | lru_cache_disable();
|
| 1535 | {
|
| 1536 | NODEMASK_SCRATCH(scratch);
|
| 1537 | if (scratch) {
|
| 1538 | mmap_write_lock(mm);
|
| 1539 | err = mpol_set_nodemask(new, nmask, scratch);
|
| 1540 | if (err)
|
| 1541 | mmap_write_unlock(mm);
|
| 1542 | } else
|
| 1543 | err = -ENOMEM;
|
| 1544 | NODEMASK_SCRATCH_FREE(scratch);
|
| 1545 | }
|
| 1546 | if (err)
|
| 1547 | goto mpol_out;
|
| 1548 |
|
| 1549 | |
| 1550 | * Lock the VMAs before scanning for pages to migrate,
|
| 1551 | * to ensure we don't miss a concurrently inserted page.
|
| 1552 |
|
| 1553 | nr_failed = queue_pages_range(mm, start, end, nmask,
|
| 1554 | flags | MPOL_MF_INVERT | MPOL_MF_WRLOCK, &pagelist);
|
| 1555 |
|
| 1556 | if (nr_failed < 0) {
|
| 1557 | err = nr_failed;
|
| 1558 | nr_failed = 0;
|
| 1559 | } else {
|
| 1560 | vma_iter_init(&vmi, mm, start);
|
| 1561 | prev = vma_prev(&vmi);
|
| 1562 | for_each_vma_range(vmi, vma, end) {
|
| 1563 | err = mbind_range(&vmi, vma, &prev, start, end, new);
|
| 1564 | if (err)
|
| 1565 | break;
|
| 1566 | }
|
| 1567 | }
|
| 1568 |
|
| 1569 | if (!err && !list_empty(&pagelist)) {
|
| 1570 |
|
| 1571 | if (!new) {
|
| 1572 | new = get_task_policy(current);
|
| 1573 | mpol_get(new);
|
| 1574 | }
|
| 1575 | mmpol.pol = new;
|
| 1576 | mmpol.ilx = 0;
|
| 1577 |
|
| 1578 | |
| 1579 | * In the interleaved case, attempt to allocate on exactly the
|
| 1580 | * targeted nodes, for the first VMA to be migrated; for later
|
| 1581 | * VMAs, the nodes will still be interleaved from the targeted
|
| 1582 | * nodemask, but one by one may be selected differently.
|
| 1583 |
|
| 1584 | if (new->mode == MPOL_INTERLEAVE ||
|
| 1585 | new->mode == MPOL_WEIGHTED_INTERLEAVE) {
|
| 1586 | struct folio *folio;
|
| 1587 | unsigned int order;
|
| 1588 | unsigned long addr = -EFAULT;
|
| 1589 |
|
| 1590 | list_for_each_entry(folio, &pagelist, lru) {
|
| 1591 | if (!folio_test_ksm(folio))
|
| 1592 | break;
|
| 1593 | }
|
| 1594 | if (!list_entry_is_head(folio, &pagelist, lru)) {
|
| 1595 | vma_iter_init(&vmi, mm, start);
|
| 1596 | for_each_vma_range(vmi, vma, end) {
|
| 1597 | addr = page_address_in_vma(folio,
|
| 1598 | folio_page(folio, 0), vma);
|
| 1599 | if (addr != -EFAULT)
|
| 1600 | break;
|
| 1601 | }
|
| 1602 | }
|
| 1603 | if (addr != -EFAULT) {
|
| 1604 | order = folio_order(folio);
|
| 1605 |
|
| 1606 | mpol_cond_put(get_vma_policy(vma, addr, order,
|
| 1607 | &mmpol.ilx));
|
| 1608 |
|
| 1609 | mmpol.ilx -= folio->index >> order;
|
| 1610 | }
|
| 1611 | }
|
| 1612 | }
|
| 1613 |
|
| 1614 | mmap_write_unlock(mm);
|
| 1615 |
|
| 1616 | if (!err && !list_empty(&pagelist)) {
|
| 1617 | nr_failed |= migrate_pages(&pagelist,
|
| 1618 | alloc_migration_target_by_mpol, NULL,
|
| 1619 | (unsigned long)&mmpol, MIGRATE_SYNC,
|
| 1620 | MR_MEMPOLICY_MBIND, NULL);
|
| 1621 | }
|
| 1622 |
|
| 1623 | if (nr_failed && (flags & MPOL_MF_STRICT))
|
| 1624 | err = -EIO;
|
| 1625 | if (!list_empty(&pagelist))
|
| 1626 | putback_movable_pages(&pagelist);
|
| 1627 | mpol_out:
|
| 1628 | mpol_put(new);
|
| 1629 | if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
|
| 1630 | lru_cache_enable();
|
| 1631 | return err;
|
| 1632 | }
|
| 1633 |
|
| 1634 | |
| 1635 | * User space interface with variable sized bitmaps for nodelists.
|
| 1636 |
|
| 1637 | static int get_bitmap(unsigned long *mask, const unsigned long __user *nmask,
|
| 1638 | unsigned long maxnode)
|
| 1639 | {
|
| 1640 | unsigned long nlongs = BITS_TO_LONGS(maxnode);
|
| 1641 | int ret;
|
| 1642 |
|
| 1643 | if (in_compat_syscall())
|
| 1644 | ret = compat_get_bitmap(mask,
|
| 1645 | (const compat_ulong_t __user *)nmask,
|
| 1646 | maxnode);
|
| 1647 | else
|
| 1648 | ret = copy_from_user(mask, nmask,
|
| 1649 | nlongs * sizeof(unsigned long));
|
| 1650 |
|
| 1651 | if (ret)
|
| 1652 | return -EFAULT;
|
| 1653 |
|
| 1654 | if (maxnode % BITS_PER_LONG)
|
| 1655 | mask[nlongs - 1] &= (1UL << (maxnode % BITS_PER_LONG)) - 1;
|
| 1656 |
|
| 1657 | return 0;
|
| 1658 | }
|
| 1659 |
|
| 1660 |
|
| 1661 | static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
|
| 1662 | unsigned long maxnode)
|
| 1663 | {
|
| 1664 | --maxnode;
|
| 1665 | nodes_clear(*nodes);
|
| 1666 | if (maxnode == 0 || !nmask)
|
| 1667 | return 0;
|
| 1668 | if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
|
| 1669 | return -EINVAL;
|
| 1670 |
|
| 1671 | |
| 1672 | * When the user specified more nodes than supported just check
|
| 1673 | * if the non supported part is all zero, one word at a time,
|
| 1674 | * starting at the end.
|
| 1675 |
|
| 1676 | while (maxnode > MAX_NUMNODES) {
|
| 1677 | unsigned long bits = min_t(unsigned long, maxnode, BITS_PER_LONG);
|
| 1678 | unsigned long t;
|
| 1679 |
|
| 1680 | if (get_bitmap(&t, &nmask[(maxnode - 1) / BITS_PER_LONG], bits))
|
| 1681 | return -EFAULT;
|
| 1682 |
|
| 1683 | if (maxnode - bits >= MAX_NUMNODES) {
|
| 1684 | maxnode -= bits;
|
| 1685 | } else {
|
| 1686 | maxnode = MAX_NUMNODES;
|
| 1687 | t &= ~((1UL << (MAX_NUMNODES % BITS_PER_LONG)) - 1);
|
| 1688 | }
|
| 1689 | if (t)
|
| 1690 | return -EINVAL;
|
| 1691 | }
|
| 1692 |
|
| 1693 | return get_bitmap(nodes_addr(*nodes), nmask, maxnode);
|
| 1694 | }
|
| 1695 |
|
| 1696 |
|
| 1697 | static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
|
| 1698 | nodemask_t *nodes)
|
| 1699 | {
|
| 1700 | unsigned long copy = ALIGN(maxnode-1, 64) / 8;
|
| 1701 | unsigned int nbytes = BITS_TO_LONGS(nr_node_ids) * sizeof(long);
|
| 1702 | bool compat = in_compat_syscall();
|
| 1703 |
|
| 1704 | if (compat)
|
| 1705 | nbytes = BITS_TO_COMPAT_LONGS(nr_node_ids) * sizeof(compat_long_t);
|
| 1706 |
|
| 1707 | if (copy > nbytes) {
|
| 1708 | if (copy > PAGE_SIZE)
|
| 1709 | return -EINVAL;
|
| 1710 | if (clear_user((char __user *)mask + nbytes, copy - nbytes))
|
| 1711 | return -EFAULT;
|
| 1712 | copy = nbytes;
|
| 1713 | maxnode = nr_node_ids;
|
| 1714 | }
|
| 1715 |
|
| 1716 | if (compat)
|
| 1717 | return compat_put_bitmap((compat_ulong_t __user *)mask,
|
| 1718 | nodes_addr(*nodes), maxnode);
|
| 1719 |
|
| 1720 | return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
|
| 1721 | }
|
| 1722 |
|
| 1723 |
|
| 1724 | static inline int sanitize_mpol_flags(int *mode, unsigned short *flags)
|
| 1725 | {
|
| 1726 | *flags = *mode & MPOL_MODE_FLAGS;
|
| 1727 | *mode &= ~MPOL_MODE_FLAGS;
|
| 1728 |
|
| 1729 | if ((unsigned int)(*mode) >= MPOL_MAX)
|
| 1730 | return -EINVAL;
|
| 1731 | if ((*flags & MPOL_F_STATIC_NODES) && (*flags & MPOL_F_RELATIVE_NODES))
|
| 1732 | return -EINVAL;
|
| 1733 | if (*flags & MPOL_F_NUMA_BALANCING) {
|
| 1734 | if (*mode == MPOL_BIND || *mode == MPOL_PREFERRED_MANY)
|
| 1735 | *flags |= (MPOL_F_MOF | MPOL_F_MORON);
|
| 1736 | else
|
| 1737 | return -EINVAL;
|
| 1738 | }
|
| 1739 | return 0;
|
| 1740 | }
|
| 1741 |
|
| 1742 | static long kernel_mbind(unsigned long start, unsigned long len,
|
| 1743 | unsigned long mode, const unsigned long __user *nmask,
|
| 1744 | unsigned long maxnode, unsigned int flags)
|
| 1745 | {
|
| 1746 | unsigned short mode_flags;
|
| 1747 | nodemask_t nodes;
|
| 1748 | int lmode = mode;
|
| 1749 | int err;
|
| 1750 |
|
| 1751 | start = untagged_addr(start);
|
| 1752 | err = sanitize_mpol_flags(&lmode, &mode_flags);
|
| 1753 | if (err)
|
| 1754 | return err;
|
| 1755 |
|
| 1756 | err = get_nodes(&nodes, nmask, maxnode);
|
| 1757 | if (err)
|
| 1758 | return err;
|
| 1759 |
|
| 1760 | return do_mbind(start, len, lmode, mode_flags, &nodes, flags);
|
| 1761 | }
|
| 1762 |
|
| 1763 | SYSCALL_DEFINE4(set_mempolicy_home_node, unsigned long, start, unsigned long, len,
|
| 1764 | unsigned long, home_node, unsigned long, flags)
|
| 1765 | {
|
| 1766 | struct mm_struct *mm = current->mm;
|
| 1767 | struct vm_area_struct *vma, *prev;
|
| 1768 | struct mempolicy *new, *old;
|
| 1769 | unsigned long end;
|
| 1770 | int err = -ENOENT;
|
| 1771 | VMA_ITERATOR(vmi, mm, start);
|
| 1772 |
|
| 1773 | start = untagged_addr(start);
|
| 1774 | if (start & ~PAGE_MASK)
|
| 1775 | return -EINVAL;
|
| 1776 | |
| 1777 | * flags is used for future extension if any.
|
| 1778 |
|
| 1779 | if (flags != 0)
|
| 1780 | return -EINVAL;
|
| 1781 |
|
| 1782 | |
| 1783 | * Check home_node is online to avoid accessing uninitialized
|
| 1784 | * NODE_DATA.
|
| 1785 |
|
| 1786 | if (home_node >= MAX_NUMNODES || !node_online(home_node))
|
| 1787 | return -EINVAL;
|
| 1788 |
|
| 1789 | len = PAGE_ALIGN(len);
|
| 1790 | end = start + len;
|
| 1791 |
|
| 1792 | if (end < start)
|
| 1793 | return -EINVAL;
|
| 1794 | if (end == start)
|
| 1795 | return 0;
|
| 1796 | mmap_write_lock(mm);
|
| 1797 | prev = vma_prev(&vmi);
|
| 1798 | for_each_vma_range(vmi, vma, end) {
|
| 1799 | |
| 1800 | * If any vma in the range got policy other than MPOL_BIND
|
| 1801 | * or MPOL_PREFERRED_MANY we return error. We don't reset
|
| 1802 | * the home node for vmas we already updated before.
|
| 1803 |
|
| 1804 | old = vma_policy(vma);
|
| 1805 | if (!old) {
|
| 1806 | prev = vma;
|
| 1807 | continue;
|
| 1808 | }
|
| 1809 | if (old->mode != MPOL_BIND && old->mode != MPOL_PREFERRED_MANY) {
|
| 1810 | err = -EOPNOTSUPP;
|
| 1811 | break;
|
| 1812 | }
|
| 1813 | new = mpol_dup(old);
|
| 1814 | if (IS_ERR(new)) {
|
| 1815 | err = PTR_ERR(new);
|
| 1816 | break;
|
| 1817 | }
|
| 1818 |
|
| 1819 | vma_start_write(vma);
|
| 1820 | new->home_node = home_node;
|
| 1821 | err = mbind_range(&vmi, vma, &prev, start, end, new);
|
| 1822 | mpol_put(new);
|
| 1823 | if (err)
|
| 1824 | break;
|
| 1825 | }
|
| 1826 | mmap_write_unlock(mm);
|
| 1827 | return err;
|
| 1828 | }
|
| 1829 |
|
| 1830 | SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
|
| 1831 | unsigned long, mode, const unsigned long __user *, nmask,
|
| 1832 | unsigned long, maxnode, unsigned int, flags)
|
| 1833 | {
|
| 1834 | return kernel_mbind(start, len, mode, nmask, maxnode, flags);
|
| 1835 | }
|
| 1836 |
|
| 1837 |
|
| 1838 | static long kernel_set_mempolicy(int mode, const unsigned long __user *nmask,
|
| 1839 | unsigned long maxnode)
|
| 1840 | {
|
| 1841 | unsigned short mode_flags;
|
| 1842 | nodemask_t nodes;
|
| 1843 | int lmode = mode;
|
| 1844 | int err;
|
| 1845 |
|
| 1846 | err = sanitize_mpol_flags(&lmode, &mode_flags);
|
| 1847 | if (err)
|
| 1848 | return err;
|
| 1849 |
|
| 1850 | err = get_nodes(&nodes, nmask, maxnode);
|
| 1851 | if (err)
|
| 1852 | return err;
|
| 1853 |
|
| 1854 | return do_set_mempolicy(lmode, mode_flags, &nodes);
|
| 1855 | }
|
| 1856 |
|
| 1857 | SYSCALL_DEFINE3(set_mempolicy, int, mode, const unsigned long __user *, nmask,
|
| 1858 | unsigned long, maxnode)
|
| 1859 | {
|
| 1860 | return kernel_set_mempolicy(mode, nmask, maxnode);
|
| 1861 | }
|
| 1862 |
|
| 1863 | static int kernel_migrate_pages(pid_t pid, unsigned long maxnode,
|
| 1864 | const unsigned long __user *old_nodes,
|
| 1865 | const unsigned long __user *new_nodes)
|
| 1866 | {
|
| 1867 | struct mm_struct *mm = NULL;
|
| 1868 | struct task_struct *task;
|
| 1869 | nodemask_t task_nodes;
|
| 1870 | int err;
|
| 1871 | nodemask_t *old;
|
| 1872 | nodemask_t *new;
|
| 1873 | NODEMASK_SCRATCH(scratch);
|
| 1874 |
|
| 1875 | if (!scratch)
|
| 1876 | return -ENOMEM;
|
| 1877 |
|
| 1878 | old = &scratch->mask1;
|
| 1879 | new = &scratch->mask2;
|
| 1880 |
|
| 1881 | err = get_nodes(old, old_nodes, maxnode);
|
| 1882 | if (err)
|
| 1883 | goto out;
|
| 1884 |
|
| 1885 | err = get_nodes(new, new_nodes, maxnode);
|
| 1886 | if (err)
|
| 1887 | goto out;
|
| 1888 |
|
| 1889 |
|
| 1890 | rcu_read_lock();
|
| 1891 | task = pid ? find_task_by_vpid(pid) : current;
|
| 1892 | if (!task) {
|
| 1893 | rcu_read_unlock();
|
| 1894 | err = -ESRCH;
|
| 1895 | goto out;
|
| 1896 | }
|
| 1897 | get_task_struct(task);
|
| 1898 |
|
| 1899 | err = -EINVAL;
|
| 1900 |
|
| 1901 | |
| 1902 | * Check if this process has the right to modify the specified process.
|
| 1903 | * Use the regular "ptrace_may_access()" checks.
|
| 1904 |
|
| 1905 | if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
|
| 1906 | rcu_read_unlock();
|
| 1907 | err = -EPERM;
|
| 1908 | goto out_put;
|
| 1909 | }
|
| 1910 | rcu_read_unlock();
|
| 1911 |
|
| 1912 | task_nodes = cpuset_mems_allowed(task);
|
| 1913 |
|
| 1914 | if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
|
| 1915 | err = -EPERM;
|
| 1916 | goto out_put;
|
| 1917 | }
|
| 1918 |
|
| 1919 | task_nodes = cpuset_mems_allowed(current);
|
| 1920 | if (!nodes_and(*new, *new, task_nodes))
|
| 1921 | goto out_put;
|
| 1922 |
|
| 1923 | err = security_task_movememory(task);
|
| 1924 | if (err)
|
| 1925 | goto out_put;
|
| 1926 |
|
| 1927 | mm = get_task_mm(task);
|
| 1928 | put_task_struct(task);
|
| 1929 |
|
| 1930 | if (!mm) {
|
| 1931 | err = -EINVAL;
|
| 1932 | goto out;
|
| 1933 | }
|
| 1934 |
|
| 1935 | err = do_migrate_pages(mm, old, new,
|
| 1936 | capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
|
| 1937 |
|
| 1938 | mmput(mm);
|
| 1939 | out:
|
| 1940 | NODEMASK_SCRATCH_FREE(scratch);
|
| 1941 |
|
| 1942 | return err;
|
| 1943 |
|
| 1944 | out_put:
|
| 1945 | put_task_struct(task);
|
| 1946 | goto out;
|
| 1947 | }
|
| 1948 |
|
| 1949 | SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
|
| 1950 | const unsigned long __user *, old_nodes,
|
| 1951 | const unsigned long __user *, new_nodes)
|
| 1952 | {
|
| 1953 | return kernel_migrate_pages(pid, maxnode, old_nodes, new_nodes);
|
| 1954 | }
|
| 1955 |
|
| 1956 |
|
| 1957 | static int kernel_get_mempolicy(int __user *policy,
|
| 1958 | unsigned long __user *nmask,
|
| 1959 | unsigned long maxnode,
|
| 1960 | unsigned long addr,
|
| 1961 | unsigned long flags)
|
| 1962 | {
|
| 1963 | int err;
|
| 1964 | int pval;
|
| 1965 | nodemask_t nodes;
|
| 1966 |
|
| 1967 | if (nmask != NULL && maxnode < nr_node_ids)
|
| 1968 | return -EINVAL;
|
| 1969 |
|
| 1970 | addr = untagged_addr(addr);
|
| 1971 |
|
| 1972 | err = do_get_mempolicy(&pval, &nodes, addr, flags);
|
| 1973 |
|
| 1974 | if (err)
|
| 1975 | return err;
|
| 1976 |
|
| 1977 | if (policy && put_user(pval, policy))
|
| 1978 | return -EFAULT;
|
| 1979 |
|
| 1980 | if (nmask)
|
| 1981 | err = copy_nodes_to_user(nmask, maxnode, &nodes);
|
| 1982 |
|
| 1983 | return err;
|
| 1984 | }
|
| 1985 |
|
| 1986 | SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
|
| 1987 | unsigned long __user *, nmask, unsigned long, maxnode,
|
| 1988 | unsigned long, addr, unsigned long, flags)
|
| 1989 | {
|
| 1990 | return kernel_get_mempolicy(policy, nmask, maxnode, addr, flags);
|
| 1991 | }
|
| 1992 |
|
| 1993 | bool vma_migratable(struct vm_area_struct *vma)
|
| 1994 | {
|
| 1995 | if (vma->vm_flags & (VM_IO | VM_PFNMAP))
|
| 1996 | return false;
|
| 1997 |
|
| 1998 | |
| 1999 | * DAX device mappings require predictable access latency, so avoid
|
| 2000 | * incurring periodic faults.
|
| 2001 |
|
| 2002 | if (vma_is_dax(vma))
|
| 2003 | return false;
|
| 2004 |
|
| 2005 | if (is_vm_hugetlb_page(vma) &&
|
| 2006 | !hugepage_migration_supported(hstate_vma(vma)))
|
| 2007 | return false;
|
| 2008 |
|
| 2009 | |
| 2010 | * Migration allocates pages in the highest zone. If we cannot
|
| 2011 | * do so then migration (at least from node to node) is not
|
| 2012 | * possible.
|
| 2013 |
|
| 2014 | if (vma->vm_file &&
|
| 2015 | gfp_zone(mapping_gfp_mask(vma->vm_file->f_mapping))
|
| 2016 | < policy_zone)
|
| 2017 | return false;
|
| 2018 | return true;
|
| 2019 | }
|
| 2020 |
|
| 2021 | struct mempolicy *__get_vma_policy(struct vm_area_struct *vma,
|
| 2022 | unsigned long addr, pgoff_t *ilx)
|
| 2023 | {
|
| 2024 | *ilx = 0;
|
| 2025 | return (vma->vm_ops && vma->vm_ops->get_policy) ?
|
| 2026 | vma->vm_ops->get_policy(vma, addr, ilx) : vma->vm_policy;
|
| 2027 | }
|
| 2028 |
|
| 2029 | |
| 2030 | * get_vma_policy(@vma, @addr, @order, @ilx)
|
| 2031 | * @vma: virtual memory area whose policy is sought
|
| 2032 | * @addr: address in @vma for shared policy lookup
|
| 2033 | * @order: 0, or appropriate huge_page_order for interleaving
|
| 2034 | * @ilx: interleave index (output), for use only when MPOL_INTERLEAVE or
|
| 2035 | * MPOL_WEIGHTED_INTERLEAVE
|
| 2036 | *
|
| 2037 | * Returns effective policy for a VMA at specified address.
|
| 2038 | * Falls back to current->mempolicy or system default policy, as necessary.
|
| 2039 | * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
|
| 2040 | * count--added by the get_policy() vm_op, as appropriate--to protect against
|
| 2041 | * freeing by another task. It is the caller's responsibility to free the
|
| 2042 | * extra reference for shared policies.
|
| 2043 |
|
| 2044 | struct mempolicy *get_vma_policy(struct vm_area_struct *vma,
|
| 2045 | unsigned long addr, int order, pgoff_t *ilx)
|
| 2046 | {
|
| 2047 | struct mempolicy *pol;
|
| 2048 |
|
| 2049 | pol = __get_vma_policy(vma, addr, ilx);
|
| 2050 | if (!pol)
|
| 2051 | pol = get_task_policy(current);
|
| 2052 | if (pol->mode == MPOL_INTERLEAVE ||
|
| 2053 | pol->mode == MPOL_WEIGHTED_INTERLEAVE) {
|
| 2054 | *ilx += vma_start_pgoff(vma) >> order;
|
| 2055 | *ilx += linear_page_delta(vma, addr) >> order;
|
| 2056 | }
|
| 2057 | return pol;
|
| 2058 | }
|
| 2059 |
|
| 2060 | bool vma_policy_mof(struct vm_area_struct *vma)
|
| 2061 | {
|
| 2062 | struct mempolicy *pol;
|
| 2063 | pgoff_t ilx;
|
| 2064 | bool mof;
|
| 2065 |
|
| 2066 | pol = __get_vma_policy(vma, vma->vm_start, &ilx);
|
| 2067 | if (!pol)
|
| 2068 | pol = get_task_policy(current);
|
| 2069 | mof = pol->flags & MPOL_F_MOF;
|
| 2070 | mpol_cond_put(pol);
|
| 2071 | return mof;
|
| 2072 | }
|
| 2073 |
|
| 2074 | bool apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
|
| 2075 | {
|
| 2076 | enum zone_type dynamic_policy_zone = policy_zone;
|
| 2077 |
|
| 2078 | BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
|
| 2079 |
|
| 2080 | |
| 2081 | * if policy->nodes has movable memory only,
|
| 2082 | * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
|
| 2083 | *
|
| 2084 | * policy->nodes is intersect with node_states[N_MEMORY].
|
| 2085 | * so if the following test fails, it implies
|
| 2086 | * policy->nodes has movable memory only.
|
| 2087 |
|
| 2088 | if (!nodes_intersects(policy->nodes, node_states[N_HIGH_MEMORY]))
|
| 2089 | dynamic_policy_zone = ZONE_MOVABLE;
|
| 2090 |
|
| 2091 | return zone >= dynamic_policy_zone;
|
| 2092 | }
|
| 2093 |
|
| 2094 | static unsigned int weighted_interleave_nodes(struct mempolicy *policy)
|
| 2095 | {
|
| 2096 | unsigned int node;
|
| 2097 | unsigned int cpuset_mems_cookie;
|
| 2098 |
|
| 2099 | retry:
|
| 2100 |
|
| 2101 | cpuset_mems_cookie = read_mems_allowed_begin();
|
| 2102 | node = current->il_prev;
|
| 2103 | if (!current->il_weight || !node_isset(node, policy->nodes)) {
|
| 2104 | node = next_node_in(node, policy->nodes);
|
| 2105 | if (read_mems_allowed_retry(cpuset_mems_cookie))
|
| 2106 | goto retry;
|
| 2107 | if (node == MAX_NUMNODES)
|
| 2108 | return node;
|
| 2109 | current->il_prev = node;
|
| 2110 | current->il_weight = get_il_weight(node);
|
| 2111 | }
|
| 2112 | current->il_weight--;
|
| 2113 | return node;
|
| 2114 | }
|
| 2115 |
|
| 2116 |
|
| 2117 | static unsigned int interleave_nodes(struct mempolicy *policy)
|
| 2118 | {
|
| 2119 | unsigned int nid;
|
| 2120 | unsigned int cpuset_mems_cookie;
|
| 2121 |
|
| 2122 |
|
| 2123 | do {
|
| 2124 | cpuset_mems_cookie = read_mems_allowed_begin();
|
| 2125 | nid = next_node_in(current->il_prev, policy->nodes);
|
| 2126 | } while (read_mems_allowed_retry(cpuset_mems_cookie));
|
| 2127 |
|
| 2128 | if (nid < MAX_NUMNODES)
|
| 2129 | current->il_prev = nid;
|
| 2130 | return nid;
|
| 2131 | }
|
| 2132 |
|
| 2133 | |
| 2134 | * Depending on the memory policy provide a node from which to allocate the
|
| 2135 | * next slab entry.
|
| 2136 |
|
| 2137 | unsigned int mempolicy_slab_node(void)
|
| 2138 | {
|
| 2139 | struct mempolicy *policy;
|
| 2140 | int node = numa_mem_id();
|
| 2141 |
|
| 2142 | if (!in_task())
|
| 2143 | return node;
|
| 2144 |
|
| 2145 | policy = current->mempolicy;
|
| 2146 | if (!policy)
|
| 2147 | return node;
|
| 2148 |
|
| 2149 | switch (policy->mode) {
|
| 2150 | case MPOL_PREFERRED:
|
| 2151 | return first_node(policy->nodes);
|
| 2152 |
|
| 2153 | case MPOL_INTERLEAVE:
|
| 2154 | return interleave_nodes(policy);
|
| 2155 |
|
| 2156 | case MPOL_WEIGHTED_INTERLEAVE:
|
| 2157 | return weighted_interleave_nodes(policy);
|
| 2158 |
|
| 2159 | case MPOL_BIND:
|
| 2160 | case MPOL_PREFERRED_MANY:
|
| 2161 | {
|
| 2162 | struct zoneref *z;
|
| 2163 |
|
| 2164 | |
| 2165 | * Follow bind policy behavior and start allocation at the
|
| 2166 | * first node.
|
| 2167 |
|
| 2168 | struct zonelist *zonelist;
|
| 2169 | enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
|
| 2170 | zonelist = &NODE_DATA(node)->node_zonelists[ZONELIST_FALLBACK];
|
| 2171 | z = first_zones_zonelist(zonelist, highest_zoneidx,
|
| 2172 | &policy->nodes);
|
| 2173 | return zonelist_zone(z) ? zonelist_node_idx(z) : node;
|
| 2174 | }
|
| 2175 | case MPOL_LOCAL:
|
| 2176 | return node;
|
| 2177 |
|
| 2178 | default:
|
| 2179 | BUG();
|
| 2180 | }
|
| 2181 | }
|
| 2182 |
|
| 2183 | static unsigned int read_once_policy_nodemask(struct mempolicy *pol,
|
| 2184 | nodemask_t *mask)
|
| 2185 | {
|
| 2186 | |
| 2187 | * barrier stabilizes the nodemask locally so that it can be iterated
|
| 2188 | * over safely without concern for changes. Allocators validate node
|
| 2189 | * selection does not violate mems_allowed, so this is safe.
|
| 2190 |
|
| 2191 | barrier();
|
| 2192 | memcpy(mask, &pol->nodes, sizeof(nodemask_t));
|
| 2193 | barrier();
|
| 2194 | return nodes_weight(*mask);
|
| 2195 | }
|
| 2196 |
|
| 2197 | static unsigned int weighted_interleave_nid(struct mempolicy *pol, pgoff_t ilx)
|
| 2198 | {
|
| 2199 | struct weighted_interleave_state *state;
|
| 2200 | nodemask_t nodemask;
|
| 2201 | unsigned int target, nr_nodes;
|
| 2202 | u8 *table = NULL;
|
| 2203 | unsigned int weight_total = 0;
|
| 2204 | u8 weight;
|
| 2205 | int nid = 0;
|
| 2206 |
|
| 2207 | nr_nodes = read_once_policy_nodemask(pol, &nodemask);
|
| 2208 | if (!nr_nodes)
|
| 2209 | return numa_node_id();
|
| 2210 |
|
| 2211 | rcu_read_lock();
|
| 2212 |
|
| 2213 | state = rcu_dereference(wi_state);
|
| 2214 |
|
| 2215 | if (state)
|
| 2216 | table = state->iw_table;
|
| 2217 |
|
| 2218 |
|
| 2219 | for_each_node_mask(nid, nodemask)
|
| 2220 | weight_total += table ? table[nid] : 1;
|
| 2221 |
|
| 2222 |
|
| 2223 | target = ilx % weight_total;
|
| 2224 | nid = first_node(nodemask);
|
| 2225 | while (target) {
|
| 2226 |
|
| 2227 | weight = table ? table[nid] : 1;
|
| 2228 | if (target < weight)
|
| 2229 | break;
|
| 2230 | target -= weight;
|
| 2231 | nid = next_node_in(nid, nodemask);
|
| 2232 | }
|
| 2233 | rcu_read_unlock();
|
| 2234 | return nid;
|
| 2235 | }
|
| 2236 |
|
| 2237 | |
| 2238 | * Do static interleaving for interleave index @ilx. Returns the ilx'th
|
| 2239 | * node in pol->nodes (starting from ilx=0), wrapping around if ilx
|
| 2240 | * exceeds the number of present nodes.
|
| 2241 |
|
| 2242 | static unsigned int interleave_nid(struct mempolicy *pol, pgoff_t ilx)
|
| 2243 | {
|
| 2244 | nodemask_t nodemask;
|
| 2245 | unsigned int target, nnodes;
|
| 2246 | int i;
|
| 2247 | int nid;
|
| 2248 |
|
| 2249 | nnodes = read_once_policy_nodemask(pol, &nodemask);
|
| 2250 | if (!nnodes)
|
| 2251 | return numa_node_id();
|
| 2252 | target = ilx % nnodes;
|
| 2253 | nid = first_node(nodemask);
|
| 2254 | for (i = 0; i < target; i++)
|
| 2255 | nid = next_node(nid, nodemask);
|
| 2256 | return nid;
|
| 2257 | }
|
| 2258 |
|
| 2259 | |
| 2260 | * Return a nodemask representing a mempolicy for filtering nodes for
|
| 2261 | * page allocation, together with preferred node id (or the input node id).
|
| 2262 |
|
| 2263 | static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *pol,
|
| 2264 | pgoff_t ilx, int *nid)
|
| 2265 | {
|
| 2266 | nodemask_t *nodemask = NULL;
|
| 2267 |
|
| 2268 | switch (pol->mode) {
|
| 2269 | case MPOL_PREFERRED:
|
| 2270 |
|
| 2271 | *nid = first_node(pol->nodes);
|
| 2272 | break;
|
| 2273 | case MPOL_PREFERRED_MANY:
|
| 2274 | nodemask = &pol->nodes;
|
| 2275 | if (pol->home_node != NUMA_NO_NODE)
|
| 2276 | *nid = pol->home_node;
|
| 2277 | break;
|
| 2278 | case MPOL_BIND:
|
| 2279 |
|
| 2280 | if (apply_policy_zone(pol, gfp_zone(gfp)) &&
|
| 2281 | cpuset_nodemask_valid_mems_allowed(&pol->nodes))
|
| 2282 | nodemask = &pol->nodes;
|
| 2283 | if (pol->home_node != NUMA_NO_NODE)
|
| 2284 | *nid = pol->home_node;
|
| 2285 | |
| 2286 | * __GFP_THISNODE shouldn't even be used with the bind policy
|
| 2287 | * because we might easily break the expectation to stay on the
|
| 2288 | * requested node and not break the policy.
|
| 2289 |
|
| 2290 | WARN_ON_ONCE(gfp & __GFP_THISNODE);
|
| 2291 | break;
|
| 2292 | case MPOL_INTERLEAVE:
|
| 2293 |
|
| 2294 | *nid = (ilx == NO_INTERLEAVE_INDEX) ?
|
| 2295 | interleave_nodes(pol) : interleave_nid(pol, ilx);
|
| 2296 | break;
|
| 2297 | case MPOL_WEIGHTED_INTERLEAVE:
|
| 2298 | *nid = (ilx == NO_INTERLEAVE_INDEX) ?
|
| 2299 | weighted_interleave_nodes(pol) :
|
| 2300 | weighted_interleave_nid(pol, ilx);
|
| 2301 | break;
|
| 2302 | }
|
| 2303 |
|
| 2304 | return nodemask;
|
| 2305 | }
|
| 2306 |
|
| 2307 | #ifdef CONFIG_HUGETLBFS
|
| 2308 | |
| 2309 | * huge_node(@vma, @addr, @gfp_flags, @mpol)
|
| 2310 | * @vma: virtual memory area whose policy is sought
|
| 2311 | * @addr: address in @vma for shared policy lookup and interleave policy
|
| 2312 | * @gfp_flags: for requested zone
|
| 2313 | * @mpol: pointer to mempolicy pointer for reference counted mempolicy
|
| 2314 | * @nodemask: pointer to nodemask pointer for 'bind' and 'prefer-many' policy
|
| 2315 | *
|
| 2316 | * Returns a nid suitable for a huge page allocation and a pointer
|
| 2317 | * to the struct mempolicy for conditional unref after allocation.
|
| 2318 | * If the effective policy is 'bind' or 'prefer-many', returns a pointer
|
| 2319 | * to the mempolicy's @nodemask for filtering the zonelist.
|
| 2320 |
|
| 2321 | int huge_node(struct vm_area_struct *vma, unsigned long addr, gfp_t gfp_flags,
|
| 2322 | struct mempolicy **mpol, nodemask_t **nodemask)
|
| 2323 | {
|
| 2324 | pgoff_t ilx;
|
| 2325 | int nid;
|
| 2326 |
|
| 2327 | nid = numa_node_id();
|
| 2328 | *mpol = get_vma_policy(vma, addr, hstate_vma(vma)->order, &ilx);
|
| 2329 | *nodemask = policy_nodemask(gfp_flags, *mpol, ilx, &nid);
|
| 2330 | return nid;
|
| 2331 | }
|
| 2332 |
|
| 2333 | |
| 2334 | * init_nodemask_of_mempolicy
|
| 2335 | *
|
| 2336 | * If the current task's mempolicy is "default" [NULL], return 'false'
|
| 2337 | * to indicate default policy. Otherwise, extract the policy nodemask
|
| 2338 | * for 'bind' or 'interleave' policy into the argument nodemask, or
|
| 2339 | * initialize the argument nodemask to contain the single node for
|
| 2340 | * 'preferred' or 'local' policy and return 'true' to indicate presence
|
| 2341 | * of non-default mempolicy.
|
| 2342 | *
|
| 2343 | * We don't bother with reference counting the mempolicy [mpol_get/put]
|
| 2344 | * because the current task is examining it's own mempolicy and a task's
|
| 2345 | * mempolicy is only ever changed by the task itself.
|
| 2346 | *
|
| 2347 | * N.B., it is the caller's responsibility to free a returned nodemask.
|
| 2348 |
|
| 2349 | bool init_nodemask_of_mempolicy(nodemask_t *mask)
|
| 2350 | {
|
| 2351 | struct mempolicy *mempolicy;
|
| 2352 |
|
| 2353 | if (!(mask && current->mempolicy))
|
| 2354 | return false;
|
| 2355 |
|
| 2356 | task_lock(current);
|
| 2357 | mempolicy = current->mempolicy;
|
| 2358 | switch (mempolicy->mode) {
|
| 2359 | case MPOL_PREFERRED:
|
| 2360 | case MPOL_PREFERRED_MANY:
|
| 2361 | case MPOL_BIND:
|
| 2362 | case MPOL_INTERLEAVE:
|
| 2363 | case MPOL_WEIGHTED_INTERLEAVE:
|
| 2364 | *mask = mempolicy->nodes;
|
| 2365 | break;
|
| 2366 |
|
| 2367 | case MPOL_LOCAL:
|
| 2368 | init_nodemask_of_node(mask, numa_node_id());
|
| 2369 | break;
|
| 2370 |
|
| 2371 | default:
|
| 2372 | BUG();
|
| 2373 | }
|
| 2374 | task_unlock(current);
|
| 2375 |
|
| 2376 | return true;
|
| 2377 | }
|
| 2378 | #endif
|
| 2379 |
|
| 2380 | |
| 2381 | * mempolicy_in_oom_domain
|
| 2382 | *
|
| 2383 | * If tsk's mempolicy is "bind", check for intersection between mask and
|
| 2384 | * the policy nodemask. Otherwise, return true for all other policies
|
| 2385 | * including "interleave", as a tsk with "interleave" policy may have
|
| 2386 | * memory allocated from all nodes in system.
|
| 2387 | *
|
| 2388 | * Takes task_lock(tsk) to prevent freeing of its mempolicy.
|
| 2389 |
|
| 2390 | bool mempolicy_in_oom_domain(struct task_struct *tsk,
|
| 2391 | const nodemask_t *mask)
|
| 2392 | {
|
| 2393 | struct mempolicy *mempolicy;
|
| 2394 | bool ret = true;
|
| 2395 |
|
| 2396 | if (!mask)
|
| 2397 | return ret;
|
| 2398 |
|
| 2399 | task_lock(tsk);
|
| 2400 | mempolicy = tsk->mempolicy;
|
| 2401 | if (mempolicy && mempolicy->mode == MPOL_BIND)
|
| 2402 | ret = nodes_intersects(mempolicy->nodes, *mask);
|
| 2403 | task_unlock(tsk);
|
| 2404 |
|
| 2405 | return ret;
|
| 2406 | }
|
| 2407 |
|
| 2408 | static struct page *alloc_pages_preferred_many(gfp_t gfp, unsigned int order,
|
| 2409 | int nid, nodemask_t *nodemask)
|
| 2410 | {
|
| 2411 | struct page *page;
|
| 2412 | gfp_t preferred_gfp;
|
| 2413 |
|
| 2414 | |
| 2415 | * This is a two pass approach. The first pass will only try the
|
| 2416 | * preferred nodes but skip the direct reclaim and allow the
|
| 2417 | * allocation to fail, while the second pass will try all the
|
| 2418 | * nodes in system.
|
| 2419 |
|
| 2420 | preferred_gfp = gfp | __GFP_NOWARN;
|
| 2421 | preferred_gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
|
| 2422 | page = __alloc_frozen_pages_noprof(preferred_gfp, order, nid, nodemask,
|
| 2423 | ALLOC_DEFAULT);
|
| 2424 | if (!page)
|
| 2425 | page = __alloc_frozen_pages_noprof(gfp, order, nid, NULL,
|
| 2426 | ALLOC_DEFAULT);
|
| 2427 |
|
| 2428 | return page;
|
| 2429 | }
|
| 2430 |
|
| 2431 | |
| 2432 | * alloc_pages_mpol - Allocate pages according to NUMA mempolicy.
|
| 2433 | * @gfp: GFP flags.
|
| 2434 | * @order: Order of the page allocation.
|
| 2435 | * @pol: Pointer to the NUMA mempolicy.
|
| 2436 | * @ilx: Index for interleave mempolicy (also distinguishes alloc_pages()).
|
| 2437 | * @nid: Preferred node (usually numa_node_id() but @mpol may override it).
|
| 2438 | *
|
| 2439 | * Return: The page on success or NULL if allocation fails.
|
| 2440 |
|
| 2441 | static struct page *alloc_pages_mpol(gfp_t gfp, unsigned int order,
|
| 2442 | struct mempolicy *pol, pgoff_t ilx, int nid)
|
| 2443 | {
|
| 2444 | nodemask_t *nodemask;
|
| 2445 | struct page *page;
|
| 2446 |
|
| 2447 | nodemask = policy_nodemask(gfp, pol, ilx, &nid);
|
| 2448 |
|
| 2449 | if (pol->mode == MPOL_PREFERRED_MANY)
|
| 2450 | return alloc_pages_preferred_many(gfp, order, nid, nodemask);
|
| 2451 |
|
| 2452 | if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) &&
|
| 2453 |
|
| 2454 | is_pmd_order(order) && ilx != NO_INTERLEAVE_INDEX) {
|
| 2455 | |
| 2456 | * For hugepage allocation and non-interleave policy which
|
| 2457 | * allows the current node (or other explicitly preferred
|
| 2458 | * node) we only try to allocate from the current/preferred
|
| 2459 | * node and don't fall back to other nodes, as the cost of
|
| 2460 | * remote accesses would likely offset THP benefits.
|
| 2461 | *
|
| 2462 | * If the policy is interleave or does not allow the current
|
| 2463 | * node in its nodemask, we allocate the standard way.
|
| 2464 |
|
| 2465 | if (pol->mode != MPOL_INTERLEAVE &&
|
| 2466 | pol->mode != MPOL_WEIGHTED_INTERLEAVE &&
|
| 2467 | (!nodemask || node_isset(nid, *nodemask))) {
|
| 2468 | |
| 2469 | * First, try to allocate THP only on local node, but
|
| 2470 | * don't reclaim unnecessarily, just compact.
|
| 2471 |
|
| 2472 | page = __alloc_frozen_pages_noprof(
|
| 2473 | gfp | __GFP_THISNODE | __GFP_NORETRY, order,
|
| 2474 | nid, NULL, ALLOC_DEFAULT);
|
| 2475 | if (page || !(gfp & __GFP_DIRECT_RECLAIM))
|
| 2476 | return page;
|
| 2477 | |
| 2478 | * If hugepage allocations are configured to always
|
| 2479 | * synchronous compact or the vma has been madvised
|
| 2480 | * to prefer hugepage backing, retry allowing remote
|
| 2481 | * memory with both reclaim and compact as well.
|
| 2482 |
|
| 2483 | }
|
| 2484 | }
|
| 2485 |
|
| 2486 | page = __alloc_frozen_pages_noprof(gfp, order, nid, nodemask, ALLOC_DEFAULT);
|
| 2487 |
|
| 2488 | if (unlikely(pol->mode == MPOL_INTERLEAVE ||
|
| 2489 | pol->mode == MPOL_WEIGHTED_INTERLEAVE) && page) {
|
| 2490 |
|
| 2491 | if (static_branch_likely(&vm_numa_stat_key) &&
|
| 2492 | page_to_nid(page) == nid) {
|
| 2493 | preempt_disable();
|
| 2494 | __count_numa_event(page_zone(page), NUMA_INTERLEAVE_HIT);
|
| 2495 | preempt_enable();
|
| 2496 | }
|
| 2497 | }
|
| 2498 |
|
| 2499 | return page;
|
| 2500 | }
|
| 2501 |
|
| 2502 | struct folio *folio_alloc_mpol_noprof(gfp_t gfp, unsigned int order,
|
| 2503 | struct mempolicy *pol, pgoff_t ilx, int nid)
|
| 2504 | {
|
| 2505 | struct page *page = alloc_pages_mpol(gfp | __GFP_COMP, order, pol,
|
| 2506 | ilx, nid);
|
| 2507 | if (!page)
|
| 2508 | return NULL;
|
| 2509 |
|
| 2510 | set_page_refcounted(page);
|
| 2511 | return page_rmappable_folio(page);
|
| 2512 | }
|
| 2513 |
|
| 2514 | |
| 2515 | * vma_alloc_folio - Allocate a folio for a VMA.
|
| 2516 | * @gfp: GFP flags.
|
| 2517 | * @order: Order of the folio.
|
| 2518 | * @vma: Pointer to VMA.
|
| 2519 | * @addr: Virtual address of the allocation. Must be inside @vma.
|
| 2520 | *
|
| 2521 | * Allocate a folio for a specific address in @vma, using the appropriate
|
| 2522 | * NUMA policy. The caller must hold the mmap_lock of the mm_struct of the
|
| 2523 | * VMA to prevent it from going away. Should be used for all allocations
|
| 2524 | * for folios that will be mapped into user space, excepting hugetlbfs, and
|
| 2525 | * excepting where direct use of folio_alloc_mpol() is more appropriate.
|
| 2526 | *
|
| 2527 | * Return: The folio on success or NULL if allocation fails.
|
| 2528 |
|
| 2529 | struct folio *vma_alloc_folio_noprof(gfp_t gfp, int order, struct vm_area_struct *vma,
|
| 2530 | unsigned long addr)
|
| 2531 | {
|
| 2532 | struct mempolicy *pol;
|
| 2533 | pgoff_t ilx;
|
| 2534 | struct folio *folio;
|
| 2535 |
|
| 2536 | if (vma->vm_flags & VM_DROPPABLE)
|
| 2537 | gfp |= __GFP_NOWARN;
|
| 2538 |
|
| 2539 | pol = get_vma_policy(vma, addr, order, &ilx);
|
| 2540 | folio = folio_alloc_mpol_noprof(gfp, order, pol, ilx, numa_node_id());
|
| 2541 | mpol_cond_put(pol);
|
| 2542 | return folio;
|
| 2543 | }
|
| 2544 | EXPORT_SYMBOL(vma_alloc_folio_noprof);
|
| 2545 |
|
| 2546 | struct page *alloc_frozen_pages_noprof(gfp_t gfp, unsigned order)
|
| 2547 | {
|
| 2548 | struct mempolicy *pol = &default_policy;
|
| 2549 |
|
| 2550 | |
| 2551 | * No reference counting needed for current->mempolicy
|
| 2552 | * nor system default_policy
|
| 2553 |
|
| 2554 | if (!in_interrupt() && !(gfp & __GFP_THISNODE))
|
| 2555 | pol = get_task_policy(current);
|
| 2556 |
|
| 2557 | return alloc_pages_mpol(gfp, order, pol, NO_INTERLEAVE_INDEX,
|
| 2558 | numa_node_id());
|
| 2559 | }
|
| 2560 |
|
| 2561 | |
| 2562 | * alloc_pages - Allocate pages.
|
| 2563 | * @gfp: GFP flags.
|
| 2564 | * @order: Power of two of number of pages to allocate.
|
| 2565 | *
|
| 2566 | * Allocate 1 << @order contiguous pages. The physical address of the
|
| 2567 | * first page is naturally aligned (eg an order-3 allocation will be aligned
|
| 2568 | * to a multiple of 8 * PAGE_SIZE bytes). The NUMA policy of the current
|
| 2569 | * process is honoured when in process context.
|
| 2570 | *
|
| 2571 | * Context: Can be called from any context, providing the appropriate GFP
|
| 2572 | * flags are used.
|
| 2573 | * Return: The page on success or NULL if allocation fails.
|
| 2574 |
|
| 2575 | struct page *alloc_pages_noprof(gfp_t gfp, unsigned int order)
|
| 2576 | {
|
| 2577 | struct page *page = alloc_frozen_pages_noprof(gfp, order);
|
| 2578 |
|
| 2579 | if (page)
|
| 2580 | set_page_refcounted(page);
|
| 2581 | return page;
|
| 2582 | }
|
| 2583 | EXPORT_SYMBOL(alloc_pages_noprof);
|
| 2584 |
|
| 2585 | struct folio *folio_alloc_noprof(gfp_t gfp, unsigned int order)
|
| 2586 | {
|
| 2587 | return page_rmappable_folio(alloc_pages_noprof(gfp | __GFP_COMP, order));
|
| 2588 | }
|
| 2589 | EXPORT_SYMBOL(folio_alloc_noprof);
|
| 2590 |
|
| 2591 | static unsigned long alloc_pages_bulk_interleave(gfp_t gfp,
|
| 2592 | struct mempolicy *pol, unsigned long nr_pages,
|
| 2593 | struct page **page_array)
|
| 2594 | {
|
| 2595 | int nodes;
|
| 2596 | unsigned long nr_pages_per_node;
|
| 2597 | int delta;
|
| 2598 | int i;
|
| 2599 | unsigned long nr_allocated;
|
| 2600 | unsigned long total_allocated = 0;
|
| 2601 |
|
| 2602 | nodes = nodes_weight(pol->nodes);
|
| 2603 | nr_pages_per_node = nr_pages / nodes;
|
| 2604 | delta = nr_pages - nodes * nr_pages_per_node;
|
| 2605 |
|
| 2606 | for (i = 0; i < nodes; i++) {
|
| 2607 | if (delta) {
|
| 2608 | nr_allocated = alloc_pages_bulk_noprof(gfp,
|
| 2609 | interleave_nodes(pol), NULL,
|
| 2610 | nr_pages_per_node + 1,
|
| 2611 | page_array);
|
| 2612 | delta--;
|
| 2613 | } else {
|
| 2614 | nr_allocated = alloc_pages_bulk_noprof(gfp,
|
| 2615 | interleave_nodes(pol), NULL,
|
| 2616 | nr_pages_per_node, page_array);
|
| 2617 | }
|
| 2618 |
|
| 2619 | page_array += nr_allocated;
|
| 2620 | total_allocated += nr_allocated;
|
| 2621 | }
|
| 2622 |
|
| 2623 | return total_allocated;
|
| 2624 | }
|
| 2625 |
|
| 2626 | static unsigned long alloc_pages_bulk_weighted_interleave(gfp_t gfp,
|
| 2627 | struct mempolicy *pol, unsigned long nr_pages,
|
| 2628 | struct page **page_array)
|
| 2629 | {
|
| 2630 | struct weighted_interleave_state *state;
|
| 2631 | struct task_struct *me = current;
|
| 2632 | unsigned int cpuset_mems_cookie;
|
| 2633 | unsigned long total_allocated = 0;
|
| 2634 | unsigned long nr_allocated = 0;
|
| 2635 | unsigned long rounds;
|
| 2636 | unsigned long node_pages, delta;
|
| 2637 | u8 *weights, weight;
|
| 2638 | unsigned int weight_total = 0;
|
| 2639 | unsigned long rem_pages = nr_pages;
|
| 2640 | nodemask_t nodes;
|
| 2641 | int nnodes, node;
|
| 2642 | int resume_node = MAX_NUMNODES - 1;
|
| 2643 | u8 resume_weight = 0;
|
| 2644 | int prev_node;
|
| 2645 | int i;
|
| 2646 |
|
| 2647 | if (!nr_pages)
|
| 2648 | return 0;
|
| 2649 |
|
| 2650 |
|
| 2651 | do {
|
| 2652 | cpuset_mems_cookie = read_mems_allowed_begin();
|
| 2653 | nnodes = read_once_policy_nodemask(pol, &nodes);
|
| 2654 | } while (read_mems_allowed_retry(cpuset_mems_cookie));
|
| 2655 |
|
| 2656 |
|
| 2657 | if (!nnodes)
|
| 2658 | return 0;
|
| 2659 |
|
| 2660 |
|
| 2661 | node = me->il_prev;
|
| 2662 | weight = me->il_weight;
|
| 2663 | if (weight && node_isset(node, nodes)) {
|
| 2664 | node_pages = min(rem_pages, weight);
|
| 2665 | nr_allocated = __alloc_pages_bulk(gfp, node, NULL, node_pages,
|
| 2666 | page_array);
|
| 2667 | page_array += nr_allocated;
|
| 2668 | total_allocated += nr_allocated;
|
| 2669 |
|
| 2670 | if (rem_pages <= weight) {
|
| 2671 | me->il_weight -= rem_pages;
|
| 2672 | return total_allocated;
|
| 2673 | }
|
| 2674 |
|
| 2675 | rem_pages -= weight;
|
| 2676 | }
|
| 2677 |
|
| 2678 | me->il_weight = 0;
|
| 2679 | prev_node = node;
|
| 2680 |
|
| 2681 |
|
| 2682 | weights = kmalloc(nr_node_ids, gfp & GFP_RECLAIM_MASK);
|
| 2683 | if (!weights)
|
| 2684 | return total_allocated;
|
| 2685 |
|
| 2686 | rcu_read_lock();
|
| 2687 | state = rcu_dereference(wi_state);
|
| 2688 | if (state) {
|
| 2689 | memcpy(weights, state->iw_table, nr_node_ids * sizeof(u8));
|
| 2690 | rcu_read_unlock();
|
| 2691 | } else {
|
| 2692 | rcu_read_unlock();
|
| 2693 | for (i = 0; i < nr_node_ids; i++)
|
| 2694 | weights[i] = 1;
|
| 2695 | }
|
| 2696 |
|
| 2697 |
|
| 2698 | for_each_node_mask(node, nodes)
|
| 2699 | weight_total += weights[node];
|
| 2700 |
|
| 2701 | |
| 2702 | * Calculate rounds/partial rounds to minimize __alloc_pages_bulk calls.
|
| 2703 | * Track which node weighted interleave should resume from.
|
| 2704 | *
|
| 2705 | * if (rounds > 0) and (delta == 0), resume_node will always be
|
| 2706 | * the node following prev_node and its weight.
|
| 2707 |
|
| 2708 | rounds = rem_pages / weight_total;
|
| 2709 | delta = rem_pages % weight_total;
|
| 2710 | resume_node = next_node_in(prev_node, nodes);
|
| 2711 | resume_weight = weights[resume_node];
|
| 2712 | for (i = 0; i < nnodes; i++) {
|
| 2713 | node = next_node_in(prev_node, nodes);
|
| 2714 | weight = weights[node];
|
| 2715 | node_pages = weight * rounds;
|
| 2716 |
|
| 2717 | if (delta > weight) {
|
| 2718 | node_pages += weight;
|
| 2719 | delta -= weight;
|
| 2720 | } else if (delta) {
|
| 2721 |
|
| 2722 | node_pages += delta;
|
| 2723 | resume_node = node;
|
| 2724 | resume_weight = weight - delta;
|
| 2725 | delta = 0;
|
| 2726 | }
|
| 2727 |
|
| 2728 | if (!node_pages)
|
| 2729 | break;
|
| 2730 | nr_allocated = __alloc_pages_bulk(gfp, node, NULL, node_pages,
|
| 2731 | page_array);
|
| 2732 | page_array += nr_allocated;
|
| 2733 | total_allocated += nr_allocated;
|
| 2734 | if (total_allocated == nr_pages)
|
| 2735 | break;
|
| 2736 | prev_node = node;
|
| 2737 | }
|
| 2738 | me->il_prev = resume_node;
|
| 2739 | me->il_weight = resume_weight;
|
| 2740 | kfree(weights);
|
| 2741 | return total_allocated;
|
| 2742 | }
|
| 2743 |
|
| 2744 | static unsigned long alloc_pages_bulk_preferred_many(gfp_t gfp, int nid,
|
| 2745 | struct mempolicy *pol, unsigned long nr_pages,
|
| 2746 | struct page **page_array)
|
| 2747 | {
|
| 2748 | gfp_t preferred_gfp;
|
| 2749 | unsigned long nr_allocated = 0;
|
| 2750 |
|
| 2751 | preferred_gfp = gfp | __GFP_NOWARN;
|
| 2752 | preferred_gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
|
| 2753 |
|
| 2754 | nr_allocated = alloc_pages_bulk_noprof(preferred_gfp, nid, &pol->nodes,
|
| 2755 | nr_pages, page_array);
|
| 2756 |
|
| 2757 | if (nr_allocated < nr_pages)
|
| 2758 | nr_allocated += alloc_pages_bulk_noprof(gfp, numa_node_id(), NULL,
|
| 2759 | nr_pages - nr_allocated,
|
| 2760 | page_array + nr_allocated);
|
| 2761 | return nr_allocated;
|
| 2762 | }
|
| 2763 |
|
| 2764 | |
| 2765 | * same time in some situation such as vmalloc.
|
| 2766 | *
|
| 2767 | * It can accelerate memory allocation especially interleaving
|
| 2768 | * allocate memory.
|
| 2769 |
|
| 2770 | unsigned long alloc_pages_bulk_mempolicy_noprof(gfp_t gfp,
|
| 2771 | unsigned long nr_pages, struct page **page_array)
|
| 2772 | {
|
| 2773 | struct mempolicy *pol = &default_policy;
|
| 2774 | nodemask_t *nodemask;
|
| 2775 | int nid;
|
| 2776 |
|
| 2777 | if (!in_interrupt() && !(gfp & __GFP_THISNODE))
|
| 2778 | pol = get_task_policy(current);
|
| 2779 |
|
| 2780 | if (pol->mode == MPOL_INTERLEAVE)
|
| 2781 | return alloc_pages_bulk_interleave(gfp, pol,
|
| 2782 | nr_pages, page_array);
|
| 2783 |
|
| 2784 | if (pol->mode == MPOL_WEIGHTED_INTERLEAVE)
|
| 2785 | return alloc_pages_bulk_weighted_interleave(
|
| 2786 | gfp, pol, nr_pages, page_array);
|
| 2787 |
|
| 2788 | if (pol->mode == MPOL_PREFERRED_MANY)
|
| 2789 | return alloc_pages_bulk_preferred_many(gfp,
|
| 2790 | numa_node_id(), pol, nr_pages, page_array);
|
| 2791 |
|
| 2792 | nid = numa_node_id();
|
| 2793 | nodemask = policy_nodemask(gfp, pol, NO_INTERLEAVE_INDEX, &nid);
|
| 2794 | return alloc_pages_bulk_noprof(gfp, nid, nodemask,
|
| 2795 | nr_pages, page_array);
|
| 2796 | }
|
| 2797 |
|
| 2798 | int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
|
| 2799 | {
|
| 2800 | struct mempolicy *pol = mpol_dup(src->vm_policy);
|
| 2801 |
|
| 2802 | if (IS_ERR(pol))
|
| 2803 | return PTR_ERR(pol);
|
| 2804 | dst->vm_policy = pol;
|
| 2805 | return 0;
|
| 2806 | }
|
| 2807 |
|
| 2808 | |
| 2809 | * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
|
| 2810 | * rebinds the mempolicy its copying by calling mpol_rebind_policy()
|
| 2811 | * with the mems_allowed returned by cpuset_mems_allowed(). This
|
| 2812 | * keeps mempolicies cpuset relative after its cpuset moves. See
|
| 2813 | * further kernel/cpuset.c update_nodemask().
|
| 2814 | *
|
| 2815 | * current's mempolicy may be rebinded by the other task(the task that changes
|
| 2816 | * cpuset's mems), so we needn't do rebind work for current task.
|
| 2817 |
|
| 2818 |
|
| 2819 |
|
| 2820 | struct mempolicy *__mpol_dup(struct mempolicy *old)
|
| 2821 | {
|
| 2822 | struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
|
| 2823 |
|
| 2824 | if (!new)
|
| 2825 | return ERR_PTR(-ENOMEM);
|
| 2826 |
|
| 2827 |
|
| 2828 | if (old == current->mempolicy) {
|
| 2829 | task_lock(current);
|
| 2830 | *new = *old;
|
| 2831 | task_unlock(current);
|
| 2832 | } else
|
| 2833 | *new = *old;
|
| 2834 |
|
| 2835 | if (current_cpuset_is_being_rebound()) {
|
| 2836 | nodemask_t mems = cpuset_mems_allowed(current);
|
| 2837 | mpol_rebind_policy(new, &mems);
|
| 2838 | }
|
| 2839 | atomic_set(&new->refcnt, 1);
|
| 2840 | return new;
|
| 2841 | }
|
| 2842 |
|
| 2843 |
|
| 2844 | bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
|
| 2845 | {
|
| 2846 | if (!a || !b)
|
| 2847 | return false;
|
| 2848 | if (a->mode != b->mode)
|
| 2849 | return false;
|
| 2850 | if (a->flags != b->flags)
|
| 2851 | return false;
|
| 2852 | if (a->home_node != b->home_node)
|
| 2853 | return false;
|
| 2854 | if (mpol_store_user_nodemask(a))
|
| 2855 | if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
|
| 2856 | return false;
|
| 2857 |
|
| 2858 | switch (a->mode) {
|
| 2859 | case MPOL_BIND:
|
| 2860 | case MPOL_INTERLEAVE:
|
| 2861 | case MPOL_PREFERRED:
|
| 2862 | case MPOL_PREFERRED_MANY:
|
| 2863 | case MPOL_WEIGHTED_INTERLEAVE:
|
| 2864 | return nodes_equal(a->nodes, b->nodes);
|
| 2865 | case MPOL_LOCAL:
|
| 2866 | return true;
|
| 2867 | default:
|
| 2868 | BUG();
|
| 2869 | return false;
|
| 2870 | }
|
| 2871 | }
|
| 2872 |
|
| 2873 | |
| 2874 | * Shared memory backing store policy support.
|
| 2875 | *
|
| 2876 | * Remember policies even when nobody has shared memory mapped.
|
| 2877 | * The policies are kept in Red-Black tree linked from the inode.
|
| 2878 | * They are protected by the sp->lock rwlock, which should be held
|
| 2879 | * for any accesses to the tree.
|
| 2880 |
|
| 2881 |
|
| 2882 | |
| 2883 | * lookup first element intersecting start-end. Caller holds sp->lock for
|
| 2884 | * reading or for writing
|
| 2885 |
|
| 2886 | static struct sp_node *sp_lookup(struct shared_policy *sp,
|
| 2887 | pgoff_t start, pgoff_t end)
|
| 2888 | {
|
| 2889 | struct rb_node *n = sp->root.rb_node;
|
| 2890 |
|
| 2891 | while (n) {
|
| 2892 | struct sp_node *p = rb_entry(n, struct sp_node, nd);
|
| 2893 |
|
| 2894 | if (start >= p->end)
|
| 2895 | n = n->rb_right;
|
| 2896 | else if (end <= p->start)
|
| 2897 | n = n->rb_left;
|
| 2898 | else
|
| 2899 | break;
|
| 2900 | }
|
| 2901 | if (!n)
|
| 2902 | return NULL;
|
| 2903 | for (;;) {
|
| 2904 | struct sp_node *w = NULL;
|
| 2905 | struct rb_node *prev = rb_prev(n);
|
| 2906 | if (!prev)
|
| 2907 | break;
|
| 2908 | w = rb_entry(prev, struct sp_node, nd);
|
| 2909 | if (w->end <= start)
|
| 2910 | break;
|
| 2911 | n = prev;
|
| 2912 | }
|
| 2913 | return rb_entry(n, struct sp_node, nd);
|
| 2914 | }
|
| 2915 |
|
| 2916 | |
| 2917 | * Insert a new shared policy into the list. Caller holds sp->lock for
|
| 2918 | * writing.
|
| 2919 |
|
| 2920 | static void sp_insert(struct shared_policy *sp, struct sp_node *new)
|
| 2921 | {
|
| 2922 | struct rb_node **p = &sp->root.rb_node;
|
| 2923 | struct rb_node *parent = NULL;
|
| 2924 | struct sp_node *nd;
|
| 2925 |
|
| 2926 | while (*p) {
|
| 2927 | parent = *p;
|
| 2928 | nd = rb_entry(parent, struct sp_node, nd);
|
| 2929 | if (new->start < nd->start)
|
| 2930 | p = &(*p)->rb_left;
|
| 2931 | else if (new->end > nd->end)
|
| 2932 | p = &(*p)->rb_right;
|
| 2933 | else
|
| 2934 | BUG();
|
| 2935 | }
|
| 2936 | rb_link_node(&new->nd, parent, p);
|
| 2937 | rb_insert_color(&new->nd, &sp->root);
|
| 2938 | }
|
| 2939 |
|
| 2940 |
|
| 2941 | struct mempolicy *mpol_shared_policy_lookup(struct shared_policy *sp,
|
| 2942 | pgoff_t idx)
|
| 2943 | {
|
| 2944 | struct mempolicy *pol = NULL;
|
| 2945 | struct sp_node *sn;
|
| 2946 |
|
| 2947 | if (!sp->root.rb_node)
|
| 2948 | return NULL;
|
| 2949 | read_lock(&sp->lock);
|
| 2950 | sn = sp_lookup(sp, idx, idx+1);
|
| 2951 | if (sn) {
|
| 2952 | mpol_get(sn->policy);
|
| 2953 | pol = sn->policy;
|
| 2954 | }
|
| 2955 | read_unlock(&sp->lock);
|
| 2956 | return pol;
|
| 2957 | }
|
| 2958 | EXPORT_SYMBOL_FOR_MODULES(mpol_shared_policy_lookup, "kvm");
|
| 2959 |
|
| 2960 | static void sp_free(struct sp_node *n)
|
| 2961 | {
|
| 2962 | mpol_put(n->policy);
|
| 2963 | kmem_cache_free(sn_cache, n);
|
| 2964 | }
|
| 2965 |
|
| 2966 | |
| 2967 | * mpol_misplaced - check whether current folio node is valid in policy
|
| 2968 | *
|
| 2969 | * @folio: folio to be checked
|
| 2970 | * @vmf: structure describing the fault
|
| 2971 | * @addr: virtual address in @vma for shared policy lookup and interleave policy
|
| 2972 | *
|
| 2973 | * Lookup current policy node id for vma,addr and "compare to" folio's
|
| 2974 | * node id. Policy determination "mimics" alloc_page_vma().
|
| 2975 | * Called from fault path where we know the vma and faulting address.
|
| 2976 | *
|
| 2977 | * Return: NUMA_NO_NODE if the page is in a node that is valid for this
|
| 2978 | * policy, or a suitable node ID to allocate a replacement folio from.
|
| 2979 |
|
| 2980 | int mpol_misplaced(struct folio *folio, struct vm_fault *vmf,
|
| 2981 | unsigned long addr)
|
| 2982 | {
|
| 2983 | struct mempolicy *pol;
|
| 2984 | pgoff_t ilx;
|
| 2985 | struct zoneref *z;
|
| 2986 | int curnid = folio_nid(folio);
|
| 2987 | struct vm_area_struct *vma = vmf->vma;
|
| 2988 | int thiscpu = raw_smp_processor_id();
|
| 2989 | int thisnid = numa_node_id();
|
| 2990 | int polnid = NUMA_NO_NODE;
|
| 2991 | int ret = NUMA_NO_NODE;
|
| 2992 |
|
| 2993 | |
| 2994 | * Make sure ptl is held so that we don't preempt and we
|
| 2995 | * have a stable smp processor id
|
| 2996 |
|
| 2997 | lockdep_assert_held(vmf->ptl);
|
| 2998 | pol = get_vma_policy(vma, addr, folio_order(folio), &ilx);
|
| 2999 | if (!(pol->flags & MPOL_F_MOF))
|
| 3000 | goto out;
|
| 3001 |
|
| 3002 | switch (pol->mode) {
|
| 3003 | case MPOL_INTERLEAVE:
|
| 3004 | polnid = interleave_nid(pol, ilx);
|
| 3005 | break;
|
| 3006 |
|
| 3007 | case MPOL_WEIGHTED_INTERLEAVE:
|
| 3008 | polnid = weighted_interleave_nid(pol, ilx);
|
| 3009 | break;
|
| 3010 |
|
| 3011 | case MPOL_PREFERRED:
|
| 3012 | if (node_isset(curnid, pol->nodes))
|
| 3013 | goto out;
|
| 3014 | polnid = first_node(pol->nodes);
|
| 3015 | break;
|
| 3016 |
|
| 3017 | case MPOL_LOCAL:
|
| 3018 | polnid = numa_node_id();
|
| 3019 | break;
|
| 3020 |
|
| 3021 | case MPOL_BIND:
|
| 3022 | case MPOL_PREFERRED_MANY:
|
| 3023 | |
| 3024 | * Even though MPOL_PREFERRED_MANY can allocate pages outside
|
| 3025 | * policy nodemask we don't allow numa migration to nodes
|
| 3026 | * outside policy nodemask for now. This is done so that if we
|
| 3027 | * want demotion to slow memory to happen, before allocating
|
| 3028 | * from some DRAM node say 'x', we will end up using a
|
| 3029 | * MPOL_PREFERRED_MANY mask excluding node 'x'. In such scenario
|
| 3030 | * we should not promote to node 'x' from slow memory node.
|
| 3031 |
|
| 3032 | if (pol->flags & MPOL_F_MORON) {
|
| 3033 | |
| 3034 | * Optimize placement among multiple nodes
|
| 3035 | * via NUMA balancing
|
| 3036 |
|
| 3037 | if (node_isset(thisnid, pol->nodes))
|
| 3038 | break;
|
| 3039 | goto out;
|
| 3040 | }
|
| 3041 |
|
| 3042 | |
| 3043 | * use current page if in policy nodemask,
|
| 3044 | * else select nearest allowed node, if any.
|
| 3045 | * If no allowed nodes, use current [!misplaced].
|
| 3046 |
|
| 3047 | if (node_isset(curnid, pol->nodes))
|
| 3048 | goto out;
|
| 3049 | z = first_zones_zonelist(
|
| 3050 | node_zonelist(thisnid, GFP_HIGHUSER),
|
| 3051 | gfp_zone(GFP_HIGHUSER),
|
| 3052 | &pol->nodes);
|
| 3053 | polnid = zonelist_node_idx(z);
|
| 3054 | break;
|
| 3055 |
|
| 3056 | default:
|
| 3057 | BUG();
|
| 3058 | }
|
| 3059 |
|
| 3060 |
|
| 3061 | if (pol->flags & MPOL_F_MORON) {
|
| 3062 | polnid = thisnid;
|
| 3063 |
|
| 3064 | if (!should_numa_migrate_memory(current, folio, curnid,
|
| 3065 | thiscpu))
|
| 3066 | goto out;
|
| 3067 | }
|
| 3068 |
|
| 3069 | if (curnid != polnid)
|
| 3070 | ret = polnid;
|
| 3071 | out:
|
| 3072 | mpol_cond_put(pol);
|
| 3073 |
|
| 3074 | return ret;
|
| 3075 | }
|
| 3076 |
|
| 3077 | |
| 3078 | * Drop the (possibly final) reference to task->mempolicy. It needs to be
|
| 3079 | * dropped after task->mempolicy is set to NULL so that any allocation done as
|
| 3080 | * part of its kmem_cache_free(), such as by KASAN, doesn't reference a freed
|
| 3081 | * policy.
|
| 3082 |
|
| 3083 | void mpol_put_task_policy(struct task_struct *task)
|
| 3084 | {
|
| 3085 | struct mempolicy *pol;
|
| 3086 |
|
| 3087 | task_lock(task);
|
| 3088 | pol = task->mempolicy;
|
| 3089 | task->mempolicy = NULL;
|
| 3090 | task_unlock(task);
|
| 3091 | mpol_put(pol);
|
| 3092 | }
|
| 3093 |
|
| 3094 | static void sp_delete(struct shared_policy *sp, struct sp_node *n)
|
| 3095 | {
|
| 3096 | rb_erase(&n->nd, &sp->root);
|
| 3097 | sp_free(n);
|
| 3098 | }
|
| 3099 |
|
| 3100 | static void sp_node_init(struct sp_node *node, unsigned long start,
|
| 3101 | unsigned long end, struct mempolicy *pol)
|
| 3102 | {
|
| 3103 | node->start = start;
|
| 3104 | node->end = end;
|
| 3105 | node->policy = pol;
|
| 3106 | }
|
| 3107 |
|
| 3108 | static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
|
| 3109 | struct mempolicy *pol)
|
| 3110 | {
|
| 3111 | struct sp_node *n;
|
| 3112 | struct mempolicy *newpol;
|
| 3113 |
|
| 3114 | n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
|
| 3115 | if (!n)
|
| 3116 | return NULL;
|
| 3117 |
|
| 3118 | newpol = mpol_dup(pol);
|
| 3119 | if (IS_ERR(newpol)) {
|
| 3120 | kmem_cache_free(sn_cache, n);
|
| 3121 | return NULL;
|
| 3122 | }
|
| 3123 | newpol->flags |= MPOL_F_SHARED;
|
| 3124 | sp_node_init(n, start, end, newpol);
|
| 3125 |
|
| 3126 | return n;
|
| 3127 | }
|
| 3128 |
|
| 3129 |
|
| 3130 | static int shared_policy_replace(struct shared_policy *sp, pgoff_t start,
|
| 3131 | pgoff_t end, struct sp_node *new)
|
| 3132 | {
|
| 3133 | struct sp_node *n;
|
| 3134 | struct sp_node *n_new = NULL;
|
| 3135 | struct mempolicy *mpol_new = NULL;
|
| 3136 | int ret = 0;
|
| 3137 |
|
| 3138 | restart:
|
| 3139 | write_lock(&sp->lock);
|
| 3140 | n = sp_lookup(sp, start, end);
|
| 3141 |
|
| 3142 | while (n && n->start < end) {
|
| 3143 | struct rb_node *next = rb_next(&n->nd);
|
| 3144 | if (n->start >= start) {
|
| 3145 | if (n->end <= end)
|
| 3146 | sp_delete(sp, n);
|
| 3147 | else
|
| 3148 | n->start = end;
|
| 3149 | } else {
|
| 3150 |
|
| 3151 | if (n->end > end) {
|
| 3152 | if (!n_new)
|
| 3153 | goto alloc_new;
|
| 3154 |
|
| 3155 | *mpol_new = *n->policy;
|
| 3156 | atomic_set(&mpol_new->refcnt, 1);
|
| 3157 | sp_node_init(n_new, end, n->end, mpol_new);
|
| 3158 | n->end = start;
|
| 3159 | sp_insert(sp, n_new);
|
| 3160 | n_new = NULL;
|
| 3161 | mpol_new = NULL;
|
| 3162 | break;
|
| 3163 | } else
|
| 3164 | n->end = start;
|
| 3165 | }
|
| 3166 | if (!next)
|
| 3167 | break;
|
| 3168 | n = rb_entry(next, struct sp_node, nd);
|
| 3169 | }
|
| 3170 | if (new)
|
| 3171 | sp_insert(sp, new);
|
| 3172 | write_unlock(&sp->lock);
|
| 3173 | ret = 0;
|
| 3174 |
|
| 3175 | err_out:
|
| 3176 | if (mpol_new)
|
| 3177 | mpol_put(mpol_new);
|
| 3178 | if (n_new)
|
| 3179 | kmem_cache_free(sn_cache, n_new);
|
| 3180 |
|
| 3181 | return ret;
|
| 3182 |
|
| 3183 | alloc_new:
|
| 3184 | write_unlock(&sp->lock);
|
| 3185 | ret = -ENOMEM;
|
| 3186 | n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
|
| 3187 | if (!n_new)
|
| 3188 | goto err_out;
|
| 3189 | mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
|
| 3190 | if (!mpol_new)
|
| 3191 | goto err_out;
|
| 3192 | atomic_set(&mpol_new->refcnt, 1);
|
| 3193 | goto restart;
|
| 3194 | }
|
| 3195 |
|
| 3196 | |
| 3197 | * mpol_shared_policy_init - initialize shared policy for inode
|
| 3198 | * @sp: pointer to inode shared policy
|
| 3199 | * @mpol: struct mempolicy to install
|
| 3200 | *
|
| 3201 | * Install non-NULL @mpol in inode's shared policy rb-tree.
|
| 3202 | * On entry, the current task has a reference on a non-NULL @mpol.
|
| 3203 | * This must be released on exit.
|
| 3204 | * This is called at get_inode() calls and we can use GFP_KERNEL.
|
| 3205 |
|
| 3206 | void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
|
| 3207 | {
|
| 3208 | int ret;
|
| 3209 |
|
| 3210 | sp->root = RB_ROOT;
|
| 3211 | rwlock_init(&sp->lock);
|
| 3212 |
|
| 3213 | if (mpol) {
|
| 3214 | struct sp_node *sn;
|
| 3215 | struct mempolicy *npol;
|
| 3216 | NODEMASK_SCRATCH(scratch);
|
| 3217 |
|
| 3218 | if (!scratch)
|
| 3219 | goto put_mpol;
|
| 3220 |
|
| 3221 |
|
| 3222 | npol = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
|
| 3223 | if (IS_ERR(npol))
|
| 3224 | goto free_scratch;
|
| 3225 |
|
| 3226 | task_lock(current);
|
| 3227 | ret = mpol_set_nodemask(npol, &mpol->w.user_nodemask, scratch);
|
| 3228 | task_unlock(current);
|
| 3229 | if (ret)
|
| 3230 | goto put_npol;
|
| 3231 |
|
| 3232 |
|
| 3233 | sn = sp_alloc(0, MAX_LFS_FILESIZE >> PAGE_SHIFT, npol);
|
| 3234 | if (sn)
|
| 3235 | sp_insert(sp, sn);
|
| 3236 | put_npol:
|
| 3237 | mpol_put(npol);
|
| 3238 | free_scratch:
|
| 3239 | NODEMASK_SCRATCH_FREE(scratch);
|
| 3240 | put_mpol:
|
| 3241 | mpol_put(mpol);
|
| 3242 | }
|
| 3243 | }
|
| 3244 | EXPORT_SYMBOL_FOR_MODULES(mpol_shared_policy_init, "kvm");
|
| 3245 |
|
| 3246 | int mpol_set_shared_policy(struct shared_policy *sp,
|
| 3247 | struct vm_area_struct *vma, struct mempolicy *pol)
|
| 3248 | {
|
| 3249 | const pgoff_t pgoff = vma_start_pgoff(vma);
|
| 3250 | const pgoff_t pgoff_end = vma_end_pgoff(vma);
|
| 3251 | struct sp_node *new = NULL;
|
| 3252 | int err;
|
| 3253 |
|
| 3254 | if (pol) {
|
| 3255 | new = sp_alloc(pgoff, pgoff_end, pol);
|
| 3256 | if (!new)
|
| 3257 | return -ENOMEM;
|
| 3258 | }
|
| 3259 | err = shared_policy_replace(sp, pgoff, pgoff_end, new);
|
| 3260 | if (err && new)
|
| 3261 | sp_free(new);
|
| 3262 | return err;
|
| 3263 | }
|
| 3264 | EXPORT_SYMBOL_FOR_MODULES(mpol_set_shared_policy, "kvm");
|
| 3265 |
|
| 3266 |
|
| 3267 | void mpol_free_shared_policy(struct shared_policy *sp)
|
| 3268 | {
|
| 3269 | struct sp_node *n;
|
| 3270 | struct rb_node *next;
|
| 3271 |
|
| 3272 | if (!sp->root.rb_node)
|
| 3273 | return;
|
| 3274 | write_lock(&sp->lock);
|
| 3275 | next = rb_first(&sp->root);
|
| 3276 | while (next) {
|
| 3277 | n = rb_entry(next, struct sp_node, nd);
|
| 3278 | next = rb_next(&n->nd);
|
| 3279 | sp_delete(sp, n);
|
| 3280 | }
|
| 3281 | write_unlock(&sp->lock);
|
| 3282 | }
|
| 3283 | EXPORT_SYMBOL_FOR_MODULES(mpol_free_shared_policy, "kvm");
|
| 3284 |
|
| 3285 | #ifdef CONFIG_NUMA_BALANCING
|
| 3286 | static int __initdata numabalancing_override;
|
| 3287 |
|
| 3288 | static void __init check_numabalancing_enable(void)
|
| 3289 | {
|
| 3290 | bool numabalancing_default = false;
|
| 3291 |
|
| 3292 | if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
|
| 3293 | numabalancing_default = true;
|
| 3294 |
|
| 3295 |
|
| 3296 | if (numabalancing_override)
|
| 3297 | set_numabalancing_state(numabalancing_override == 1);
|
| 3298 |
|
| 3299 | if (num_online_nodes() > 1 && !numabalancing_override) {
|
| 3300 | pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
|
| 3301 | numabalancing_default ? "Enabling" : "Disabling");
|
| 3302 | set_numabalancing_state(numabalancing_default);
|
| 3303 | }
|
| 3304 | }
|
| 3305 |
|
| 3306 | static int __init setup_numabalancing(char *str)
|
| 3307 | {
|
| 3308 | int ret = 0;
|
| 3309 | if (!str)
|
| 3310 | goto out;
|
| 3311 |
|
| 3312 | if (!strcmp(str, "enable")) {
|
| 3313 | numabalancing_override = 1;
|
| 3314 | ret = 1;
|
| 3315 | } else if (!strcmp(str, "disable")) {
|
| 3316 | numabalancing_override = -1;
|
| 3317 | ret = 1;
|
| 3318 | }
|
| 3319 | out:
|
| 3320 | if (!ret)
|
| 3321 | pr_warn("Unable to parse numa_balancing=\n");
|
| 3322 |
|
| 3323 | return ret;
|
| 3324 | }
|
| 3325 | __setup("numa_balancing=", setup_numabalancing);
|
| 3326 | #else
|
| 3327 | static inline void __init check_numabalancing_enable(void)
|
| 3328 | {
|
| 3329 | }
|
| 3330 | #endif
|
| 3331 |
|
| 3332 | void __init numa_policy_init(void)
|
| 3333 | {
|
| 3334 | nodemask_t interleave_nodes;
|
| 3335 | unsigned long largest = 0;
|
| 3336 | int nid, prefer = 0;
|
| 3337 |
|
| 3338 | policy_cache = kmem_cache_create("numa_policy",
|
| 3339 | sizeof(struct mempolicy),
|
| 3340 | 0, SLAB_PANIC, NULL);
|
| 3341 |
|
| 3342 | sn_cache = kmem_cache_create("shared_policy_node",
|
| 3343 | sizeof(struct sp_node),
|
| 3344 | 0, SLAB_PANIC, NULL);
|
| 3345 |
|
| 3346 | for_each_node(nid) {
|
| 3347 | preferred_node_policy[nid] = (struct mempolicy) {
|
| 3348 | .refcnt = ATOMIC_INIT(1),
|
| 3349 | .mode = MPOL_PREFERRED,
|
| 3350 | .flags = MPOL_F_MOF | MPOL_F_MORON,
|
| 3351 | .nodes = nodemask_of_node(nid),
|
| 3352 | };
|
| 3353 | }
|
| 3354 |
|
| 3355 | |
| 3356 | * Set interleaving policy for system init. Interleaving is only
|
| 3357 | * enabled across suitably sized nodes (default is >= 16MB), or
|
| 3358 | * fall back to the largest node if they're all smaller.
|
| 3359 |
|
| 3360 | nodes_clear(interleave_nodes);
|
| 3361 | for_each_node_state(nid, N_MEMORY) {
|
| 3362 | unsigned long total_pages = node_present_pages(nid);
|
| 3363 |
|
| 3364 |
|
| 3365 | if (largest < total_pages) {
|
| 3366 | largest = total_pages;
|
| 3367 | prefer = nid;
|
| 3368 | }
|
| 3369 |
|
| 3370 |
|
| 3371 | if ((total_pages << PAGE_SHIFT) >= (16 << 20))
|
| 3372 | node_set(nid, interleave_nodes);
|
| 3373 | }
|
| 3374 |
|
| 3375 |
|
| 3376 | if (unlikely(nodes_empty(interleave_nodes)))
|
| 3377 | node_set(prefer, interleave_nodes);
|
| 3378 |
|
| 3379 | if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
|
| 3380 | pr_err("%s: interleaving failed\n", __func__);
|
| 3381 |
|
| 3382 | check_numabalancing_enable();
|
| 3383 | }
|
| 3384 |
|
| 3385 |
|
| 3386 | void numa_default_policy(void)
|
| 3387 | {
|
| 3388 | do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
|
| 3389 | }
|
| 3390 |
|
| 3391 | |
| 3392 | * Parse and format mempolicy from/to strings
|
| 3393 |
|
| 3394 | static const char * const policy_modes[] =
|
| 3395 | {
|
| 3396 | [MPOL_DEFAULT] = "default",
|
| 3397 | [MPOL_PREFERRED] = "prefer",
|
| 3398 | [MPOL_BIND] = "bind",
|
| 3399 | [MPOL_INTERLEAVE] = "interleave",
|
| 3400 | [MPOL_WEIGHTED_INTERLEAVE] = "weighted interleave",
|
| 3401 | [MPOL_LOCAL] = "local",
|
| 3402 | [MPOL_PREFERRED_MANY] = "prefer (many)",
|
| 3403 | };
|
| 3404 |
|
| 3405 | #ifdef CONFIG_TMPFS
|
| 3406 | |
| 3407 | * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
|
| 3408 | * @str: string containing mempolicy to parse
|
| 3409 | * @mpol: pointer to struct mempolicy pointer, returned on success.
|
| 3410 | *
|
| 3411 | * Format of input:
|
| 3412 | * <mode>[=<flags>][:<nodelist>]
|
| 3413 | *
|
| 3414 | * Return: %0 on success, else %1
|
| 3415 |
|
| 3416 | int mpol_parse_str(char *str, struct mempolicy **mpol)
|
| 3417 | {
|
| 3418 | struct mempolicy *new = NULL;
|
| 3419 | unsigned short mode_flags;
|
| 3420 | nodemask_t nodes;
|
| 3421 | char *nodelist = strchr(str, ':');
|
| 3422 | char *flags = strchr(str, '=');
|
| 3423 | int err = 1, mode;
|
| 3424 |
|
| 3425 | if (flags)
|
| 3426 | *flags++ = '\0';
|
| 3427 |
|
| 3428 | if (nodelist) {
|
| 3429 |
|
| 3430 | *nodelist++ = '\0';
|
| 3431 | if (nodelist_parse(nodelist, nodes))
|
| 3432 | goto out;
|
| 3433 | if (!nodes_subset(nodes, node_states[N_MEMORY]))
|
| 3434 | goto out;
|
| 3435 | } else
|
| 3436 | nodes_clear(nodes);
|
| 3437 |
|
| 3438 | mode = match_string(policy_modes, MPOL_MAX, str);
|
| 3439 | if (mode < 0)
|
| 3440 | goto out;
|
| 3441 |
|
| 3442 | switch (mode) {
|
| 3443 | case MPOL_PREFERRED:
|
| 3444 | |
| 3445 | * Insist on a nodelist of one node only, although later
|
| 3446 | * we use first_node(nodes) to grab a single node, so here
|
| 3447 | * nodelist (or nodes) cannot be empty.
|
| 3448 |
|
| 3449 | if (nodelist) {
|
| 3450 | char *rest = nodelist;
|
| 3451 | while (isdigit(*rest))
|
| 3452 | rest++;
|
| 3453 | if (*rest)
|
| 3454 | goto out;
|
| 3455 | if (nodes_empty(nodes))
|
| 3456 | goto out;
|
| 3457 | }
|
| 3458 | break;
|
| 3459 | case MPOL_INTERLEAVE:
|
| 3460 | case MPOL_WEIGHTED_INTERLEAVE:
|
| 3461 | |
| 3462 | * Default to online nodes with memory if no nodelist
|
| 3463 |
|
| 3464 | if (!nodelist)
|
| 3465 | nodes = node_states[N_MEMORY];
|
| 3466 | break;
|
| 3467 | case MPOL_LOCAL:
|
| 3468 | |
| 3469 | * Don't allow a nodelist; mpol_new() checks flags
|
| 3470 |
|
| 3471 | if (nodelist)
|
| 3472 | goto out;
|
| 3473 | break;
|
| 3474 | case MPOL_DEFAULT:
|
| 3475 | |
| 3476 | * Insist on a empty nodelist
|
| 3477 |
|
| 3478 | if (!nodelist)
|
| 3479 | err = 0;
|
| 3480 | goto out;
|
| 3481 | case MPOL_PREFERRED_MANY:
|
| 3482 | case MPOL_BIND:
|
| 3483 | |
| 3484 | * Insist on a nodelist
|
| 3485 |
|
| 3486 | if (!nodelist)
|
| 3487 | goto out;
|
| 3488 | }
|
| 3489 |
|
| 3490 | mode_flags = 0;
|
| 3491 | if (flags) {
|
| 3492 | |
| 3493 | * Currently, we only support two mutually exclusive
|
| 3494 | * mode flags.
|
| 3495 |
|
| 3496 | if (!strcmp(flags, "static"))
|
| 3497 | mode_flags |= MPOL_F_STATIC_NODES;
|
| 3498 | else if (!strcmp(flags, "relative"))
|
| 3499 | mode_flags |= MPOL_F_RELATIVE_NODES;
|
| 3500 | else
|
| 3501 | goto out;
|
| 3502 | }
|
| 3503 |
|
| 3504 | new = mpol_new(mode, mode_flags, &nodes);
|
| 3505 | if (IS_ERR(new))
|
| 3506 | goto out;
|
| 3507 |
|
| 3508 | |
| 3509 | * Save nodes for mpol_to_str() to show the tmpfs mount options
|
| 3510 | * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
|
| 3511 |
|
| 3512 | if (mode != MPOL_PREFERRED) {
|
| 3513 | new->nodes = nodes;
|
| 3514 | } else if (nodelist) {
|
| 3515 | nodes_clear(new->nodes);
|
| 3516 | node_set(first_node(nodes), new->nodes);
|
| 3517 | } else {
|
| 3518 | new->mode = MPOL_LOCAL;
|
| 3519 | }
|
| 3520 |
|
| 3521 | |
| 3522 | * Save nodes for contextualization: this will be used to "clone"
|
| 3523 | * the mempolicy in a specific context [cpuset] at a later time.
|
| 3524 |
|
| 3525 | new->w.user_nodemask = nodes;
|
| 3526 |
|
| 3527 | err = 0;
|
| 3528 |
|
| 3529 | out:
|
| 3530 |
|
| 3531 | if (nodelist)
|
| 3532 | *--nodelist = ':';
|
| 3533 | if (flags)
|
| 3534 | *--flags = '=';
|
| 3535 | if (!err)
|
| 3536 | *mpol = new;
|
| 3537 | return err;
|
| 3538 | }
|
| 3539 | #endif
|
| 3540 |
|
| 3541 | |
| 3542 | * mpol_to_str - format a mempolicy structure for printing
|
| 3543 | * @buffer: to contain formatted mempolicy string
|
| 3544 | * @maxlen: length of @buffer
|
| 3545 | * @pol: pointer to mempolicy to be formatted
|
| 3546 | *
|
| 3547 | * Convert @pol into a string. If @buffer is too short, truncate the string.
|
| 3548 | * Recommend a @maxlen of at least 51 for the longest mode, "weighted
|
| 3549 | * interleave", plus the longest flag flags, "relative|balancing", and to
|
| 3550 | * display at least a few node ids.
|
| 3551 |
|
| 3552 | void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
|
| 3553 | {
|
| 3554 | char *p = buffer;
|
| 3555 | nodemask_t nodes = NODE_MASK_NONE;
|
| 3556 | unsigned short mode = MPOL_DEFAULT;
|
| 3557 | unsigned short flags = 0;
|
| 3558 |
|
| 3559 | if (pol &&
|
| 3560 | pol != &default_policy &&
|
| 3561 | !(pol >= &preferred_node_policy[0] &&
|
| 3562 | pol <= &preferred_node_policy[ARRAY_SIZE(preferred_node_policy) - 1])) {
|
| 3563 | mode = pol->mode;
|
| 3564 | flags = pol->flags;
|
| 3565 | }
|
| 3566 |
|
| 3567 | switch (mode) {
|
| 3568 | case MPOL_DEFAULT:
|
| 3569 | case MPOL_LOCAL:
|
| 3570 | break;
|
| 3571 | case MPOL_PREFERRED:
|
| 3572 | case MPOL_PREFERRED_MANY:
|
| 3573 | case MPOL_BIND:
|
| 3574 | case MPOL_INTERLEAVE:
|
| 3575 | case MPOL_WEIGHTED_INTERLEAVE:
|
| 3576 | nodes = pol->nodes;
|
| 3577 | break;
|
| 3578 | default:
|
| 3579 | WARN_ON_ONCE(1);
|
| 3580 | snprintf(p, maxlen, "unknown");
|
| 3581 | return;
|
| 3582 | }
|
| 3583 |
|
| 3584 | p += snprintf(p, maxlen, "%s", policy_modes[mode]);
|
| 3585 |
|
| 3586 | if (flags & MPOL_MODE_FLAGS) {
|
| 3587 | p += snprintf(p, buffer + maxlen - p, "=");
|
| 3588 |
|
| 3589 | |
| 3590 | * Static and relative are mutually exclusive.
|
| 3591 |
|
| 3592 | if (flags & MPOL_F_STATIC_NODES)
|
| 3593 | p += snprintf(p, buffer + maxlen - p, "static");
|
| 3594 | else if (flags & MPOL_F_RELATIVE_NODES)
|
| 3595 | p += snprintf(p, buffer + maxlen - p, "relative");
|
| 3596 |
|
| 3597 | if (flags & MPOL_F_NUMA_BALANCING) {
|
| 3598 | if (!is_power_of_2(flags & MPOL_MODE_FLAGS))
|
| 3599 | p += snprintf(p, buffer + maxlen - p, "|");
|
| 3600 | p += snprintf(p, buffer + maxlen - p, "balancing");
|
| 3601 | }
|
| 3602 | }
|
| 3603 |
|
| 3604 | if (!nodes_empty(nodes))
|
| 3605 | p += scnprintf(p, buffer + maxlen - p, ":%*pbl",
|
| 3606 | nodemask_pr_args(&nodes));
|
| 3607 | }
|
| 3608 |
|
| 3609 | #ifdef CONFIG_SYSFS
|
| 3610 | struct iw_node_attr {
|
| 3611 | struct kobj_attribute kobj_attr;
|
| 3612 | int nid;
|
| 3613 | };
|
| 3614 |
|
| 3615 | struct sysfs_wi_group {
|
| 3616 | struct kobject wi_kobj;
|
| 3617 | struct mutex kobj_lock;
|
| 3618 | struct iw_node_attr *nattrs[];
|
| 3619 | };
|
| 3620 |
|
| 3621 | static struct sysfs_wi_group *wi_group;
|
| 3622 |
|
| 3623 | static ssize_t node_show(struct kobject *kobj, struct kobj_attribute *attr,
|
| 3624 | char *buf)
|
| 3625 | {
|
| 3626 | struct iw_node_attr *node_attr;
|
| 3627 | u8 weight;
|
| 3628 |
|
| 3629 | node_attr = container_of(attr, struct iw_node_attr, kobj_attr);
|
| 3630 | weight = get_il_weight(node_attr->nid);
|
| 3631 | return sysfs_emit(buf, "%d\n", weight);
|
| 3632 | }
|
| 3633 |
|
| 3634 | static ssize_t node_store(struct kobject *kobj, struct kobj_attribute *attr,
|
| 3635 | const char *buf, size_t count)
|
| 3636 | {
|
| 3637 | struct weighted_interleave_state *new_wi_state, *old_wi_state = NULL;
|
| 3638 | struct iw_node_attr *node_attr;
|
| 3639 | u8 weight = 0;
|
| 3640 | int i;
|
| 3641 |
|
| 3642 | node_attr = container_of(attr, struct iw_node_attr, kobj_attr);
|
| 3643 | if (count == 0 || sysfs_streq(buf, "") ||
|
| 3644 | kstrtou8(buf, 0, &weight) || weight == 0)
|
| 3645 | return -EINVAL;
|
| 3646 |
|
| 3647 | new_wi_state = kzalloc_flex(*new_wi_state, iw_table, nr_node_ids);
|
| 3648 | if (!new_wi_state)
|
| 3649 | return -ENOMEM;
|
| 3650 |
|
| 3651 | mutex_lock(&wi_state_lock);
|
| 3652 | old_wi_state = rcu_dereference_protected(wi_state,
|
| 3653 | lockdep_is_held(&wi_state_lock));
|
| 3654 | if (old_wi_state) {
|
| 3655 | memcpy(new_wi_state->iw_table, old_wi_state->iw_table,
|
| 3656 | nr_node_ids * sizeof(u8));
|
| 3657 | } else {
|
| 3658 | for (i = 0; i < nr_node_ids; i++)
|
| 3659 | new_wi_state->iw_table[i] = 1;
|
| 3660 | }
|
| 3661 | new_wi_state->iw_table[node_attr->nid] = weight;
|
| 3662 | new_wi_state->mode_auto = false;
|
| 3663 |
|
| 3664 | rcu_assign_pointer(wi_state, new_wi_state);
|
| 3665 | mutex_unlock(&wi_state_lock);
|
| 3666 | if (old_wi_state) {
|
| 3667 | synchronize_rcu();
|
| 3668 | kfree(old_wi_state);
|
| 3669 | }
|
| 3670 | return count;
|
| 3671 | }
|
| 3672 |
|
| 3673 | static ssize_t weighted_interleave_auto_show(struct kobject *kobj,
|
| 3674 | struct kobj_attribute *attr, char *buf)
|
| 3675 | {
|
| 3676 | struct weighted_interleave_state *state;
|
| 3677 | bool wi_auto = true;
|
| 3678 |
|
| 3679 | rcu_read_lock();
|
| 3680 | state = rcu_dereference(wi_state);
|
| 3681 | if (state)
|
| 3682 | wi_auto = state->mode_auto;
|
| 3683 | rcu_read_unlock();
|
| 3684 |
|
| 3685 | return sysfs_emit(buf, "%s\n", str_true_false(wi_auto));
|
| 3686 | }
|
| 3687 |
|
| 3688 | static ssize_t weighted_interleave_auto_store(struct kobject *kobj,
|
| 3689 | struct kobj_attribute *attr, const char *buf, size_t count)
|
| 3690 | {
|
| 3691 | struct weighted_interleave_state *new_wi_state, *old_wi_state = NULL;
|
| 3692 | unsigned int *bw;
|
| 3693 | bool input;
|
| 3694 | int i;
|
| 3695 |
|
| 3696 | if (kstrtobool(buf, &input))
|
| 3697 | return -EINVAL;
|
| 3698 |
|
| 3699 | new_wi_state = kzalloc_flex(*new_wi_state, iw_table, nr_node_ids);
|
| 3700 | if (!new_wi_state)
|
| 3701 | return -ENOMEM;
|
| 3702 | for (i = 0; i < nr_node_ids; i++)
|
| 3703 | new_wi_state->iw_table[i] = 1;
|
| 3704 |
|
| 3705 | mutex_lock(&wi_state_lock);
|
| 3706 | old_wi_state = rcu_dereference_protected(wi_state,
|
| 3707 | lockdep_is_held(&wi_state_lock));
|
| 3708 |
|
| 3709 | if (old_wi_state && input == old_wi_state->mode_auto) {
|
| 3710 | mutex_unlock(&wi_state_lock);
|
| 3711 | kfree(new_wi_state);
|
| 3712 | return count;
|
| 3713 | }
|
| 3714 |
|
| 3715 | if (!input) {
|
| 3716 | if (old_wi_state)
|
| 3717 | memcpy(new_wi_state->iw_table, old_wi_state->iw_table,
|
| 3718 | nr_node_ids * sizeof(u8));
|
| 3719 | goto update_wi_state;
|
| 3720 | }
|
| 3721 |
|
| 3722 | bw = node_bw_table;
|
| 3723 | if (!bw) {
|
| 3724 | mutex_unlock(&wi_state_lock);
|
| 3725 | kfree(new_wi_state);
|
| 3726 | return -ENODEV;
|
| 3727 | }
|
| 3728 |
|
| 3729 | new_wi_state->mode_auto = true;
|
| 3730 | reduce_interleave_weights(bw, new_wi_state->iw_table);
|
| 3731 |
|
| 3732 | update_wi_state:
|
| 3733 | rcu_assign_pointer(wi_state, new_wi_state);
|
| 3734 | mutex_unlock(&wi_state_lock);
|
| 3735 | if (old_wi_state) {
|
| 3736 | synchronize_rcu();
|
| 3737 | kfree(old_wi_state);
|
| 3738 | }
|
| 3739 | return count;
|
| 3740 | }
|
| 3741 |
|
| 3742 | static void sysfs_wi_node_delete(int nid)
|
| 3743 | {
|
| 3744 | struct iw_node_attr *attr;
|
| 3745 |
|
| 3746 | if (nid < 0 || nid >= nr_node_ids)
|
| 3747 | return;
|
| 3748 |
|
| 3749 | mutex_lock(&wi_group->kobj_lock);
|
| 3750 | attr = wi_group->nattrs[nid];
|
| 3751 | if (!attr) {
|
| 3752 | mutex_unlock(&wi_group->kobj_lock);
|
| 3753 | return;
|
| 3754 | }
|
| 3755 |
|
| 3756 | wi_group->nattrs[nid] = NULL;
|
| 3757 | mutex_unlock(&wi_group->kobj_lock);
|
| 3758 |
|
| 3759 | sysfs_remove_file(&wi_group->wi_kobj, &attr->kobj_attr.attr);
|
| 3760 | kfree(attr->kobj_attr.attr.name);
|
| 3761 | kfree(attr);
|
| 3762 | }
|
| 3763 |
|
| 3764 | static void sysfs_wi_node_delete_all(void)
|
| 3765 | {
|
| 3766 | int nid;
|
| 3767 |
|
| 3768 | for (nid = 0; nid < nr_node_ids; nid++)
|
| 3769 | sysfs_wi_node_delete(nid);
|
| 3770 | }
|
| 3771 |
|
| 3772 | static void wi_state_free(void)
|
| 3773 | {
|
| 3774 | struct weighted_interleave_state *old_wi_state;
|
| 3775 |
|
| 3776 | mutex_lock(&wi_state_lock);
|
| 3777 | old_wi_state = rcu_dereference_protected(wi_state,
|
| 3778 | lockdep_is_held(&wi_state_lock));
|
| 3779 | rcu_assign_pointer(wi_state, NULL);
|
| 3780 | mutex_unlock(&wi_state_lock);
|
| 3781 |
|
| 3782 | if (old_wi_state) {
|
| 3783 | synchronize_rcu();
|
| 3784 | kfree(old_wi_state);
|
| 3785 | }
|
| 3786 | }
|
| 3787 |
|
| 3788 | static struct kobj_attribute wi_auto_attr = {
|
| 3789 | .attr = { .name = "auto", .mode = 0664 },
|
| 3790 | .show = weighted_interleave_auto_show,
|
| 3791 | .store = weighted_interleave_auto_store,
|
| 3792 | };
|
| 3793 |
|
| 3794 | static void wi_cleanup(void) {
|
| 3795 | sysfs_remove_file(&wi_group->wi_kobj, &wi_auto_attr.attr);
|
| 3796 | sysfs_wi_node_delete_all();
|
| 3797 | wi_state_free();
|
| 3798 | }
|
| 3799 |
|
| 3800 | static void wi_kobj_release(struct kobject *wi_kobj)
|
| 3801 | {
|
| 3802 | kfree(wi_group);
|
| 3803 | }
|
| 3804 |
|
| 3805 | static const struct kobj_type wi_ktype = {
|
| 3806 | .sysfs_ops = &kobj_sysfs_ops,
|
| 3807 | .release = wi_kobj_release,
|
| 3808 | };
|
| 3809 |
|
| 3810 | static int sysfs_wi_node_add(int nid)
|
| 3811 | {
|
| 3812 | int ret;
|
| 3813 | char *name;
|
| 3814 | struct iw_node_attr *new_attr;
|
| 3815 |
|
| 3816 | if (nid < 0 || nid >= nr_node_ids) {
|
| 3817 | pr_err("invalid node id: %d\n", nid);
|
| 3818 | return -EINVAL;
|
| 3819 | }
|
| 3820 |
|
| 3821 | new_attr = kzalloc_obj(*new_attr);
|
| 3822 | if (!new_attr)
|
| 3823 | return -ENOMEM;
|
| 3824 |
|
| 3825 | name = kasprintf(GFP_KERNEL, "node%d", nid);
|
| 3826 | if (!name) {
|
| 3827 | kfree(new_attr);
|
| 3828 | return -ENOMEM;
|
| 3829 | }
|
| 3830 |
|
| 3831 | sysfs_attr_init(&new_attr->kobj_attr.attr);
|
| 3832 | new_attr->kobj_attr.attr.name = name;
|
| 3833 | new_attr->kobj_attr.attr.mode = 0644;
|
| 3834 | new_attr->kobj_attr.show = node_show;
|
| 3835 | new_attr->kobj_attr.store = node_store;
|
| 3836 | new_attr->nid = nid;
|
| 3837 |
|
| 3838 | mutex_lock(&wi_group->kobj_lock);
|
| 3839 | if (wi_group->nattrs[nid]) {
|
| 3840 | mutex_unlock(&wi_group->kobj_lock);
|
| 3841 | ret = -EEXIST;
|
| 3842 | goto out;
|
| 3843 | }
|
| 3844 |
|
| 3845 | ret = sysfs_create_file(&wi_group->wi_kobj, &new_attr->kobj_attr.attr);
|
| 3846 | if (ret) {
|
| 3847 | mutex_unlock(&wi_group->kobj_lock);
|
| 3848 | goto out;
|
| 3849 | }
|
| 3850 | wi_group->nattrs[nid] = new_attr;
|
| 3851 | mutex_unlock(&wi_group->kobj_lock);
|
| 3852 | return 0;
|
| 3853 |
|
| 3854 | out:
|
| 3855 | kfree(new_attr->kobj_attr.attr.name);
|
| 3856 | kfree(new_attr);
|
| 3857 | return ret;
|
| 3858 | }
|
| 3859 |
|
| 3860 | static int wi_node_notifier(struct notifier_block *nb,
|
| 3861 | unsigned long action, void *data)
|
| 3862 | {
|
| 3863 | int err;
|
| 3864 | struct node_notify *nn = data;
|
| 3865 | int nid = nn->nid;
|
| 3866 |
|
| 3867 | switch (action) {
|
| 3868 | case NODE_ADDED_FIRST_MEMORY:
|
| 3869 | err = sysfs_wi_node_add(nid);
|
| 3870 | if (err)
|
| 3871 | pr_err("failed to add sysfs for node%d during hotplug: %d\n",
|
| 3872 | nid, err);
|
| 3873 | break;
|
| 3874 | case NODE_REMOVED_LAST_MEMORY:
|
| 3875 | sysfs_wi_node_delete(nid);
|
| 3876 | break;
|
| 3877 | }
|
| 3878 |
|
| 3879 | return NOTIFY_OK;
|
| 3880 | }
|
| 3881 |
|
| 3882 | static int __init add_weighted_interleave_group(struct kobject *mempolicy_kobj)
|
| 3883 | {
|
| 3884 | int nid, err;
|
| 3885 |
|
| 3886 | wi_group = kzalloc_flex(*wi_group, nattrs, nr_node_ids);
|
| 3887 | if (!wi_group)
|
| 3888 | return -ENOMEM;
|
| 3889 | mutex_init(&wi_group->kobj_lock);
|
| 3890 |
|
| 3891 | err = kobject_init_and_add(&wi_group->wi_kobj, &wi_ktype, mempolicy_kobj,
|
| 3892 | "weighted_interleave");
|
| 3893 | if (err)
|
| 3894 | goto err_put_kobj;
|
| 3895 |
|
| 3896 | err = sysfs_create_file(&wi_group->wi_kobj, &wi_auto_attr.attr);
|
| 3897 | if (err)
|
| 3898 | goto err_put_kobj;
|
| 3899 |
|
| 3900 | for_each_online_node(nid) {
|
| 3901 | if (!node_state(nid, N_MEMORY))
|
| 3902 | continue;
|
| 3903 |
|
| 3904 | err = sysfs_wi_node_add(nid);
|
| 3905 | if (err) {
|
| 3906 | pr_err("failed to add sysfs for node%d during init: %d\n",
|
| 3907 | nid, err);
|
| 3908 | goto err_cleanup_kobj;
|
| 3909 | }
|
| 3910 | }
|
| 3911 |
|
| 3912 | hotplug_node_notifier(wi_node_notifier, DEFAULT_CALLBACK_PRI);
|
| 3913 | return 0;
|
| 3914 |
|
| 3915 | err_cleanup_kobj:
|
| 3916 | wi_cleanup();
|
| 3917 | kobject_del(&wi_group->wi_kobj);
|
| 3918 | err_put_kobj:
|
| 3919 | kobject_put(&wi_group->wi_kobj);
|
| 3920 | return err;
|
| 3921 | }
|
| 3922 |
|
| 3923 | static int __init mempolicy_sysfs_init(void)
|
| 3924 | {
|
| 3925 | int err;
|
| 3926 | static struct kobject *mempolicy_kobj;
|
| 3927 |
|
| 3928 | mempolicy_kobj = kobject_create_and_add("mempolicy", mm_kobj);
|
| 3929 | if (!mempolicy_kobj)
|
| 3930 | return -ENOMEM;
|
| 3931 |
|
| 3932 | err = add_weighted_interleave_group(mempolicy_kobj);
|
| 3933 | if (err)
|
| 3934 | goto err_kobj;
|
| 3935 |
|
| 3936 | return 0;
|
| 3937 |
|
| 3938 | err_kobj:
|
| 3939 | kobject_del(mempolicy_kobj);
|
| 3940 | kobject_put(mempolicy_kobj);
|
| 3941 | return err;
|
| 3942 | }
|
| 3943 |
|
| 3944 | late_initcall(mempolicy_sysfs_init);
|
| 3945 | #endif
|