Martin KaFai Lau | 3a08c2f | 2016-11-11 10:55:06 -0800 | [diff] [blame] | 1 | /* Copyright (c) 2016 Facebook |
| 2 | * |
| 3 | * This program is free software; you can redistribute it and/or |
| 4 | * modify it under the terms of version 2 of the GNU General Public |
| 5 | * License as published by the Free Software Foundation. |
| 6 | */ |
| 7 | #include <linux/cpumask.h> |
| 8 | #include <linux/spinlock.h> |
| 9 | #include <linux/percpu.h> |
| 10 | |
| 11 | #include "bpf_lru_list.h" |
| 12 | |
| 13 | #define LOCAL_FREE_TARGET (128) |
| 14 | #define LOCAL_NR_SCANS LOCAL_FREE_TARGET |
| 15 | |
Martin KaFai Lau | 695ba26 | 2017-04-14 10:30:29 -0700 | [diff] [blame] | 16 | #define PERCPU_FREE_TARGET (4) |
Martin KaFai Lau | 961578b | 2016-11-11 10:55:07 -0800 | [diff] [blame] | 17 | #define PERCPU_NR_SCANS PERCPU_FREE_TARGET |
| 18 | |
Martin KaFai Lau | 3a08c2f | 2016-11-11 10:55:06 -0800 | [diff] [blame] | 19 | /* Helpers to get the local list index */ |
| 20 | #define LOCAL_LIST_IDX(t) ((t) - BPF_LOCAL_LIST_T_OFFSET) |
| 21 | #define LOCAL_FREE_LIST_IDX LOCAL_LIST_IDX(BPF_LRU_LOCAL_LIST_T_FREE) |
| 22 | #define LOCAL_PENDING_LIST_IDX LOCAL_LIST_IDX(BPF_LRU_LOCAL_LIST_T_PENDING) |
| 23 | #define IS_LOCAL_LIST_TYPE(t) ((t) >= BPF_LOCAL_LIST_T_OFFSET) |
| 24 | |
| 25 | static int get_next_cpu(int cpu) |
| 26 | { |
| 27 | cpu = cpumask_next(cpu, cpu_possible_mask); |
| 28 | if (cpu >= nr_cpu_ids) |
| 29 | cpu = cpumask_first(cpu_possible_mask); |
| 30 | return cpu; |
| 31 | } |
| 32 | |
| 33 | /* Local list helpers */ |
| 34 | static struct list_head *local_free_list(struct bpf_lru_locallist *loc_l) |
| 35 | { |
| 36 | return &loc_l->lists[LOCAL_FREE_LIST_IDX]; |
| 37 | } |
| 38 | |
| 39 | static struct list_head *local_pending_list(struct bpf_lru_locallist *loc_l) |
| 40 | { |
| 41 | return &loc_l->lists[LOCAL_PENDING_LIST_IDX]; |
| 42 | } |
| 43 | |
| 44 | /* bpf_lru_node helpers */ |
| 45 | static bool bpf_lru_node_is_ref(const struct bpf_lru_node *node) |
| 46 | { |
| 47 | return node->ref; |
| 48 | } |
| 49 | |
| 50 | static void bpf_lru_list_count_inc(struct bpf_lru_list *l, |
| 51 | enum bpf_lru_list_type type) |
| 52 | { |
| 53 | if (type < NR_BPF_LRU_LIST_COUNT) |
| 54 | l->counts[type]++; |
| 55 | } |
| 56 | |
| 57 | static void bpf_lru_list_count_dec(struct bpf_lru_list *l, |
| 58 | enum bpf_lru_list_type type) |
| 59 | { |
| 60 | if (type < NR_BPF_LRU_LIST_COUNT) |
| 61 | l->counts[type]--; |
| 62 | } |
| 63 | |
| 64 | static void __bpf_lru_node_move_to_free(struct bpf_lru_list *l, |
| 65 | struct bpf_lru_node *node, |
| 66 | struct list_head *free_list, |
| 67 | enum bpf_lru_list_type tgt_free_type) |
| 68 | { |
| 69 | if (WARN_ON_ONCE(IS_LOCAL_LIST_TYPE(node->type))) |
| 70 | return; |
| 71 | |
| 72 | /* If the removing node is the next_inactive_rotation candidate, |
| 73 | * move the next_inactive_rotation pointer also. |
| 74 | */ |
| 75 | if (&node->list == l->next_inactive_rotation) |
| 76 | l->next_inactive_rotation = l->next_inactive_rotation->prev; |
| 77 | |
| 78 | bpf_lru_list_count_dec(l, node->type); |
| 79 | |
| 80 | node->type = tgt_free_type; |
| 81 | list_move(&node->list, free_list); |
| 82 | } |
| 83 | |
| 84 | /* Move nodes from local list to the LRU list */ |
| 85 | static void __bpf_lru_node_move_in(struct bpf_lru_list *l, |
| 86 | struct bpf_lru_node *node, |
| 87 | enum bpf_lru_list_type tgt_type) |
| 88 | { |
| 89 | if (WARN_ON_ONCE(!IS_LOCAL_LIST_TYPE(node->type)) || |
| 90 | WARN_ON_ONCE(IS_LOCAL_LIST_TYPE(tgt_type))) |
| 91 | return; |
| 92 | |
| 93 | bpf_lru_list_count_inc(l, tgt_type); |
| 94 | node->type = tgt_type; |
| 95 | node->ref = 0; |
| 96 | list_move(&node->list, &l->lists[tgt_type]); |
| 97 | } |
| 98 | |
| 99 | /* Move nodes between or within active and inactive list (like |
| 100 | * active to inactive, inactive to active or tail of active back to |
| 101 | * the head of active). |
| 102 | */ |
| 103 | static void __bpf_lru_node_move(struct bpf_lru_list *l, |
| 104 | struct bpf_lru_node *node, |
| 105 | enum bpf_lru_list_type tgt_type) |
| 106 | { |
| 107 | if (WARN_ON_ONCE(IS_LOCAL_LIST_TYPE(node->type)) || |
| 108 | WARN_ON_ONCE(IS_LOCAL_LIST_TYPE(tgt_type))) |
| 109 | return; |
| 110 | |
| 111 | if (node->type != tgt_type) { |
| 112 | bpf_lru_list_count_dec(l, node->type); |
| 113 | bpf_lru_list_count_inc(l, tgt_type); |
| 114 | node->type = tgt_type; |
| 115 | } |
| 116 | node->ref = 0; |
| 117 | |
| 118 | /* If the moving node is the next_inactive_rotation candidate, |
| 119 | * move the next_inactive_rotation pointer also. |
| 120 | */ |
| 121 | if (&node->list == l->next_inactive_rotation) |
| 122 | l->next_inactive_rotation = l->next_inactive_rotation->prev; |
| 123 | |
| 124 | list_move(&node->list, &l->lists[tgt_type]); |
| 125 | } |
| 126 | |
| 127 | static bool bpf_lru_list_inactive_low(const struct bpf_lru_list *l) |
| 128 | { |
| 129 | return l->counts[BPF_LRU_LIST_T_INACTIVE] < |
| 130 | l->counts[BPF_LRU_LIST_T_ACTIVE]; |
| 131 | } |
| 132 | |
| 133 | /* Rotate the active list: |
| 134 | * 1. Start from tail |
| 135 | * 2. If the node has the ref bit set, it will be rotated |
| 136 | * back to the head of active list with the ref bit cleared. |
| 137 | * Give this node one more chance to survive in the active list. |
| 138 | * 3. If the ref bit is not set, move it to the head of the |
| 139 | * inactive list. |
| 140 | * 4. It will at most scan nr_scans nodes |
| 141 | */ |
| 142 | static void __bpf_lru_list_rotate_active(struct bpf_lru *lru, |
| 143 | struct bpf_lru_list *l) |
| 144 | { |
| 145 | struct list_head *active = &l->lists[BPF_LRU_LIST_T_ACTIVE]; |
| 146 | struct bpf_lru_node *node, *tmp_node, *first_node; |
| 147 | unsigned int i = 0; |
| 148 | |
| 149 | first_node = list_first_entry(active, struct bpf_lru_node, list); |
| 150 | list_for_each_entry_safe_reverse(node, tmp_node, active, list) { |
| 151 | if (bpf_lru_node_is_ref(node)) |
| 152 | __bpf_lru_node_move(l, node, BPF_LRU_LIST_T_ACTIVE); |
| 153 | else |
| 154 | __bpf_lru_node_move(l, node, BPF_LRU_LIST_T_INACTIVE); |
| 155 | |
| 156 | if (++i == lru->nr_scans || node == first_node) |
| 157 | break; |
| 158 | } |
| 159 | } |
| 160 | |
| 161 | /* Rotate the inactive list. It starts from the next_inactive_rotation |
| 162 | * 1. If the node has ref bit set, it will be moved to the head |
| 163 | * of active list with the ref bit cleared. |
| 164 | * 2. If the node does not have ref bit set, it will leave it |
| 165 | * at its current location (i.e. do nothing) so that it can |
| 166 | * be considered during the next inactive_shrink. |
| 167 | * 3. It will at most scan nr_scans nodes |
| 168 | */ |
| 169 | static void __bpf_lru_list_rotate_inactive(struct bpf_lru *lru, |
| 170 | struct bpf_lru_list *l) |
| 171 | { |
| 172 | struct list_head *inactive = &l->lists[BPF_LRU_LIST_T_INACTIVE]; |
Martin KaFai Lau | 2874aa2 | 2016-11-15 11:00:04 -0800 | [diff] [blame] | 173 | struct list_head *cur, *last, *next = inactive; |
Martin KaFai Lau | 3a08c2f | 2016-11-11 10:55:06 -0800 | [diff] [blame] | 174 | struct bpf_lru_node *node; |
| 175 | unsigned int i = 0; |
| 176 | |
| 177 | if (list_empty(inactive)) |
| 178 | return; |
| 179 | |
| 180 | last = l->next_inactive_rotation->next; |
| 181 | if (last == inactive) |
| 182 | last = last->next; |
| 183 | |
| 184 | cur = l->next_inactive_rotation; |
| 185 | while (i < lru->nr_scans) { |
| 186 | if (cur == inactive) { |
| 187 | cur = cur->prev; |
| 188 | continue; |
| 189 | } |
| 190 | |
| 191 | node = list_entry(cur, struct bpf_lru_node, list); |
| 192 | next = cur->prev; |
| 193 | if (bpf_lru_node_is_ref(node)) |
| 194 | __bpf_lru_node_move(l, node, BPF_LRU_LIST_T_ACTIVE); |
| 195 | if (cur == last) |
| 196 | break; |
| 197 | cur = next; |
| 198 | i++; |
| 199 | } |
| 200 | |
| 201 | l->next_inactive_rotation = next; |
| 202 | } |
| 203 | |
| 204 | /* Shrink the inactive list. It starts from the tail of the |
| 205 | * inactive list and only move the nodes without the ref bit |
| 206 | * set to the designated free list. |
| 207 | */ |
| 208 | static unsigned int |
| 209 | __bpf_lru_list_shrink_inactive(struct bpf_lru *lru, |
| 210 | struct bpf_lru_list *l, |
| 211 | unsigned int tgt_nshrink, |
| 212 | struct list_head *free_list, |
| 213 | enum bpf_lru_list_type tgt_free_type) |
| 214 | { |
| 215 | struct list_head *inactive = &l->lists[BPF_LRU_LIST_T_INACTIVE]; |
Tobias Klauser | a5ef01a | 2017-01-10 15:02:07 +0100 | [diff] [blame] | 216 | struct bpf_lru_node *node, *tmp_node; |
Martin KaFai Lau | 3a08c2f | 2016-11-11 10:55:06 -0800 | [diff] [blame] | 217 | unsigned int nshrinked = 0; |
| 218 | unsigned int i = 0; |
| 219 | |
Martin KaFai Lau | 3a08c2f | 2016-11-11 10:55:06 -0800 | [diff] [blame] | 220 | list_for_each_entry_safe_reverse(node, tmp_node, inactive, list) { |
| 221 | if (bpf_lru_node_is_ref(node)) { |
| 222 | __bpf_lru_node_move(l, node, BPF_LRU_LIST_T_ACTIVE); |
| 223 | } else if (lru->del_from_htab(lru->del_arg, node)) { |
| 224 | __bpf_lru_node_move_to_free(l, node, free_list, |
| 225 | tgt_free_type); |
| 226 | if (++nshrinked == tgt_nshrink) |
| 227 | break; |
| 228 | } |
| 229 | |
| 230 | if (++i == lru->nr_scans) |
| 231 | break; |
| 232 | } |
| 233 | |
| 234 | return nshrinked; |
| 235 | } |
| 236 | |
| 237 | /* 1. Rotate the active list (if needed) |
| 238 | * 2. Always rotate the inactive list |
| 239 | */ |
| 240 | static void __bpf_lru_list_rotate(struct bpf_lru *lru, struct bpf_lru_list *l) |
| 241 | { |
| 242 | if (bpf_lru_list_inactive_low(l)) |
| 243 | __bpf_lru_list_rotate_active(lru, l); |
| 244 | |
| 245 | __bpf_lru_list_rotate_inactive(lru, l); |
| 246 | } |
| 247 | |
| 248 | /* Calls __bpf_lru_list_shrink_inactive() to shrink some |
| 249 | * ref-bit-cleared nodes and move them to the designated |
| 250 | * free list. |
| 251 | * |
| 252 | * If it cannot get a free node after calling |
| 253 | * __bpf_lru_list_shrink_inactive(). It will just remove |
| 254 | * one node from either inactive or active list without |
| 255 | * honoring the ref-bit. It prefers inactive list to active |
| 256 | * list in this situation. |
| 257 | */ |
| 258 | static unsigned int __bpf_lru_list_shrink(struct bpf_lru *lru, |
| 259 | struct bpf_lru_list *l, |
| 260 | unsigned int tgt_nshrink, |
| 261 | struct list_head *free_list, |
| 262 | enum bpf_lru_list_type tgt_free_type) |
| 263 | |
| 264 | { |
| 265 | struct bpf_lru_node *node, *tmp_node; |
| 266 | struct list_head *force_shrink_list; |
| 267 | unsigned int nshrinked; |
| 268 | |
| 269 | nshrinked = __bpf_lru_list_shrink_inactive(lru, l, tgt_nshrink, |
| 270 | free_list, tgt_free_type); |
| 271 | if (nshrinked) |
| 272 | return nshrinked; |
| 273 | |
| 274 | /* Do a force shrink by ignoring the reference bit */ |
| 275 | if (!list_empty(&l->lists[BPF_LRU_LIST_T_INACTIVE])) |
| 276 | force_shrink_list = &l->lists[BPF_LRU_LIST_T_INACTIVE]; |
| 277 | else |
| 278 | force_shrink_list = &l->lists[BPF_LRU_LIST_T_ACTIVE]; |
| 279 | |
| 280 | list_for_each_entry_safe_reverse(node, tmp_node, force_shrink_list, |
| 281 | list) { |
| 282 | if (lru->del_from_htab(lru->del_arg, node)) { |
| 283 | __bpf_lru_node_move_to_free(l, node, free_list, |
| 284 | tgt_free_type); |
| 285 | return 1; |
| 286 | } |
| 287 | } |
| 288 | |
| 289 | return 0; |
| 290 | } |
| 291 | |
| 292 | /* Flush the nodes from the local pending list to the LRU list */ |
| 293 | static void __local_list_flush(struct bpf_lru_list *l, |
| 294 | struct bpf_lru_locallist *loc_l) |
| 295 | { |
| 296 | struct bpf_lru_node *node, *tmp_node; |
| 297 | |
| 298 | list_for_each_entry_safe_reverse(node, tmp_node, |
| 299 | local_pending_list(loc_l), list) { |
| 300 | if (bpf_lru_node_is_ref(node)) |
| 301 | __bpf_lru_node_move_in(l, node, BPF_LRU_LIST_T_ACTIVE); |
| 302 | else |
| 303 | __bpf_lru_node_move_in(l, node, |
| 304 | BPF_LRU_LIST_T_INACTIVE); |
| 305 | } |
| 306 | } |
| 307 | |
| 308 | static void bpf_lru_list_push_free(struct bpf_lru_list *l, |
| 309 | struct bpf_lru_node *node) |
| 310 | { |
| 311 | unsigned long flags; |
| 312 | |
| 313 | if (WARN_ON_ONCE(IS_LOCAL_LIST_TYPE(node->type))) |
| 314 | return; |
| 315 | |
| 316 | raw_spin_lock_irqsave(&l->lock, flags); |
| 317 | __bpf_lru_node_move(l, node, BPF_LRU_LIST_T_FREE); |
| 318 | raw_spin_unlock_irqrestore(&l->lock, flags); |
| 319 | } |
| 320 | |
| 321 | static void bpf_lru_list_pop_free_to_local(struct bpf_lru *lru, |
| 322 | struct bpf_lru_locallist *loc_l) |
| 323 | { |
| 324 | struct bpf_lru_list *l = &lru->common_lru.lru_list; |
| 325 | struct bpf_lru_node *node, *tmp_node; |
| 326 | unsigned int nfree = 0; |
| 327 | |
| 328 | raw_spin_lock(&l->lock); |
| 329 | |
| 330 | __local_list_flush(l, loc_l); |
| 331 | |
| 332 | __bpf_lru_list_rotate(lru, l); |
| 333 | |
| 334 | list_for_each_entry_safe(node, tmp_node, &l->lists[BPF_LRU_LIST_T_FREE], |
| 335 | list) { |
| 336 | __bpf_lru_node_move_to_free(l, node, local_free_list(loc_l), |
| 337 | BPF_LRU_LOCAL_LIST_T_FREE); |
| 338 | if (++nfree == LOCAL_FREE_TARGET) |
| 339 | break; |
| 340 | } |
| 341 | |
| 342 | if (nfree < LOCAL_FREE_TARGET) |
| 343 | __bpf_lru_list_shrink(lru, l, LOCAL_FREE_TARGET - nfree, |
| 344 | local_free_list(loc_l), |
| 345 | BPF_LRU_LOCAL_LIST_T_FREE); |
| 346 | |
| 347 | raw_spin_unlock(&l->lock); |
| 348 | } |
| 349 | |
| 350 | static void __local_list_add_pending(struct bpf_lru *lru, |
| 351 | struct bpf_lru_locallist *loc_l, |
| 352 | int cpu, |
| 353 | struct bpf_lru_node *node, |
| 354 | u32 hash) |
| 355 | { |
| 356 | *(u32 *)((void *)node + lru->hash_offset) = hash; |
| 357 | node->cpu = cpu; |
| 358 | node->type = BPF_LRU_LOCAL_LIST_T_PENDING; |
| 359 | node->ref = 0; |
| 360 | list_add(&node->list, local_pending_list(loc_l)); |
| 361 | } |
| 362 | |
Tobias Klauser | 3bf0033 | 2017-01-10 15:02:16 +0100 | [diff] [blame] | 363 | static struct bpf_lru_node * |
| 364 | __local_list_pop_free(struct bpf_lru_locallist *loc_l) |
Martin KaFai Lau | 3a08c2f | 2016-11-11 10:55:06 -0800 | [diff] [blame] | 365 | { |
| 366 | struct bpf_lru_node *node; |
| 367 | |
| 368 | node = list_first_entry_or_null(local_free_list(loc_l), |
| 369 | struct bpf_lru_node, |
| 370 | list); |
| 371 | if (node) |
| 372 | list_del(&node->list); |
| 373 | |
| 374 | return node; |
| 375 | } |
| 376 | |
Tobias Klauser | 3bf0033 | 2017-01-10 15:02:16 +0100 | [diff] [blame] | 377 | static struct bpf_lru_node * |
| 378 | __local_list_pop_pending(struct bpf_lru *lru, struct bpf_lru_locallist *loc_l) |
Martin KaFai Lau | 3a08c2f | 2016-11-11 10:55:06 -0800 | [diff] [blame] | 379 | { |
| 380 | struct bpf_lru_node *node; |
| 381 | bool force = false; |
| 382 | |
| 383 | ignore_ref: |
| 384 | /* Get from the tail (i.e. older element) of the pending list. */ |
| 385 | list_for_each_entry_reverse(node, local_pending_list(loc_l), |
| 386 | list) { |
| 387 | if ((!bpf_lru_node_is_ref(node) || force) && |
| 388 | lru->del_from_htab(lru->del_arg, node)) { |
| 389 | list_del(&node->list); |
| 390 | return node; |
| 391 | } |
| 392 | } |
| 393 | |
| 394 | if (!force) { |
| 395 | force = true; |
| 396 | goto ignore_ref; |
| 397 | } |
| 398 | |
| 399 | return NULL; |
| 400 | } |
| 401 | |
Martin KaFai Lau | 961578b | 2016-11-11 10:55:07 -0800 | [diff] [blame] | 402 | static struct bpf_lru_node *bpf_percpu_lru_pop_free(struct bpf_lru *lru, |
| 403 | u32 hash) |
| 404 | { |
| 405 | struct list_head *free_list; |
| 406 | struct bpf_lru_node *node = NULL; |
| 407 | struct bpf_lru_list *l; |
| 408 | unsigned long flags; |
| 409 | int cpu = raw_smp_processor_id(); |
| 410 | |
| 411 | l = per_cpu_ptr(lru->percpu_lru, cpu); |
| 412 | |
| 413 | raw_spin_lock_irqsave(&l->lock, flags); |
| 414 | |
| 415 | __bpf_lru_list_rotate(lru, l); |
| 416 | |
| 417 | free_list = &l->lists[BPF_LRU_LIST_T_FREE]; |
| 418 | if (list_empty(free_list)) |
| 419 | __bpf_lru_list_shrink(lru, l, PERCPU_FREE_TARGET, free_list, |
| 420 | BPF_LRU_LIST_T_FREE); |
| 421 | |
| 422 | if (!list_empty(free_list)) { |
| 423 | node = list_first_entry(free_list, struct bpf_lru_node, list); |
| 424 | *(u32 *)((void *)node + lru->hash_offset) = hash; |
| 425 | node->ref = 0; |
| 426 | __bpf_lru_node_move(l, node, BPF_LRU_LIST_T_INACTIVE); |
| 427 | } |
| 428 | |
| 429 | raw_spin_unlock_irqrestore(&l->lock, flags); |
| 430 | |
| 431 | return node; |
| 432 | } |
| 433 | |
| 434 | static struct bpf_lru_node *bpf_common_lru_pop_free(struct bpf_lru *lru, |
| 435 | u32 hash) |
Martin KaFai Lau | 3a08c2f | 2016-11-11 10:55:06 -0800 | [diff] [blame] | 436 | { |
| 437 | struct bpf_lru_locallist *loc_l, *steal_loc_l; |
| 438 | struct bpf_common_lru *clru = &lru->common_lru; |
| 439 | struct bpf_lru_node *node; |
| 440 | int steal, first_steal; |
| 441 | unsigned long flags; |
| 442 | int cpu = raw_smp_processor_id(); |
| 443 | |
| 444 | loc_l = per_cpu_ptr(clru->local_list, cpu); |
| 445 | |
| 446 | raw_spin_lock_irqsave(&loc_l->lock, flags); |
| 447 | |
| 448 | node = __local_list_pop_free(loc_l); |
| 449 | if (!node) { |
| 450 | bpf_lru_list_pop_free_to_local(lru, loc_l); |
| 451 | node = __local_list_pop_free(loc_l); |
| 452 | } |
| 453 | |
| 454 | if (node) |
| 455 | __local_list_add_pending(lru, loc_l, cpu, node, hash); |
| 456 | |
| 457 | raw_spin_unlock_irqrestore(&loc_l->lock, flags); |
| 458 | |
| 459 | if (node) |
| 460 | return node; |
| 461 | |
| 462 | /* No free nodes found from the local free list and |
| 463 | * the global LRU list. |
| 464 | * |
| 465 | * Steal from the local free/pending list of the |
| 466 | * current CPU and remote CPU in RR. It starts |
| 467 | * with the loc_l->next_steal CPU. |
| 468 | */ |
| 469 | |
| 470 | first_steal = loc_l->next_steal; |
| 471 | steal = first_steal; |
| 472 | do { |
| 473 | steal_loc_l = per_cpu_ptr(clru->local_list, steal); |
| 474 | |
| 475 | raw_spin_lock_irqsave(&steal_loc_l->lock, flags); |
| 476 | |
| 477 | node = __local_list_pop_free(steal_loc_l); |
| 478 | if (!node) |
| 479 | node = __local_list_pop_pending(lru, steal_loc_l); |
| 480 | |
| 481 | raw_spin_unlock_irqrestore(&steal_loc_l->lock, flags); |
| 482 | |
| 483 | steal = get_next_cpu(steal); |
| 484 | } while (!node && steal != first_steal); |
| 485 | |
| 486 | loc_l->next_steal = steal; |
| 487 | |
| 488 | if (node) { |
| 489 | raw_spin_lock_irqsave(&loc_l->lock, flags); |
| 490 | __local_list_add_pending(lru, loc_l, cpu, node, hash); |
| 491 | raw_spin_unlock_irqrestore(&loc_l->lock, flags); |
| 492 | } |
| 493 | |
| 494 | return node; |
| 495 | } |
| 496 | |
Martin KaFai Lau | 961578b | 2016-11-11 10:55:07 -0800 | [diff] [blame] | 497 | struct bpf_lru_node *bpf_lru_pop_free(struct bpf_lru *lru, u32 hash) |
| 498 | { |
| 499 | if (lru->percpu) |
| 500 | return bpf_percpu_lru_pop_free(lru, hash); |
| 501 | else |
| 502 | return bpf_common_lru_pop_free(lru, hash); |
| 503 | } |
| 504 | |
| 505 | static void bpf_common_lru_push_free(struct bpf_lru *lru, |
| 506 | struct bpf_lru_node *node) |
Martin KaFai Lau | 3a08c2f | 2016-11-11 10:55:06 -0800 | [diff] [blame] | 507 | { |
| 508 | unsigned long flags; |
| 509 | |
| 510 | if (WARN_ON_ONCE(node->type == BPF_LRU_LIST_T_FREE) || |
| 511 | WARN_ON_ONCE(node->type == BPF_LRU_LOCAL_LIST_T_FREE)) |
| 512 | return; |
| 513 | |
| 514 | if (node->type == BPF_LRU_LOCAL_LIST_T_PENDING) { |
| 515 | struct bpf_lru_locallist *loc_l; |
| 516 | |
| 517 | loc_l = per_cpu_ptr(lru->common_lru.local_list, node->cpu); |
| 518 | |
| 519 | raw_spin_lock_irqsave(&loc_l->lock, flags); |
| 520 | |
| 521 | if (unlikely(node->type != BPF_LRU_LOCAL_LIST_T_PENDING)) { |
| 522 | raw_spin_unlock_irqrestore(&loc_l->lock, flags); |
| 523 | goto check_lru_list; |
| 524 | } |
| 525 | |
| 526 | node->type = BPF_LRU_LOCAL_LIST_T_FREE; |
| 527 | node->ref = 0; |
| 528 | list_move(&node->list, local_free_list(loc_l)); |
| 529 | |
| 530 | raw_spin_unlock_irqrestore(&loc_l->lock, flags); |
| 531 | return; |
| 532 | } |
| 533 | |
| 534 | check_lru_list: |
| 535 | bpf_lru_list_push_free(&lru->common_lru.lru_list, node); |
| 536 | } |
| 537 | |
Martin KaFai Lau | 961578b | 2016-11-11 10:55:07 -0800 | [diff] [blame] | 538 | static void bpf_percpu_lru_push_free(struct bpf_lru *lru, |
| 539 | struct bpf_lru_node *node) |
| 540 | { |
| 541 | struct bpf_lru_list *l; |
| 542 | unsigned long flags; |
| 543 | |
| 544 | l = per_cpu_ptr(lru->percpu_lru, node->cpu); |
| 545 | |
| 546 | raw_spin_lock_irqsave(&l->lock, flags); |
| 547 | |
| 548 | __bpf_lru_node_move(l, node, BPF_LRU_LIST_T_FREE); |
| 549 | |
| 550 | raw_spin_unlock_irqrestore(&l->lock, flags); |
| 551 | } |
| 552 | |
| 553 | void bpf_lru_push_free(struct bpf_lru *lru, struct bpf_lru_node *node) |
| 554 | { |
| 555 | if (lru->percpu) |
| 556 | bpf_percpu_lru_push_free(lru, node); |
| 557 | else |
| 558 | bpf_common_lru_push_free(lru, node); |
| 559 | } |
| 560 | |
Tobias Klauser | 3bf0033 | 2017-01-10 15:02:16 +0100 | [diff] [blame] | 561 | static void bpf_common_lru_populate(struct bpf_lru *lru, void *buf, |
| 562 | u32 node_offset, u32 elem_size, |
| 563 | u32 nr_elems) |
Martin KaFai Lau | 3a08c2f | 2016-11-11 10:55:06 -0800 | [diff] [blame] | 564 | { |
| 565 | struct bpf_lru_list *l = &lru->common_lru.lru_list; |
| 566 | u32 i; |
| 567 | |
| 568 | for (i = 0; i < nr_elems; i++) { |
| 569 | struct bpf_lru_node *node; |
| 570 | |
| 571 | node = (struct bpf_lru_node *)(buf + node_offset); |
| 572 | node->type = BPF_LRU_LIST_T_FREE; |
| 573 | node->ref = 0; |
| 574 | list_add(&node->list, &l->lists[BPF_LRU_LIST_T_FREE]); |
| 575 | buf += elem_size; |
| 576 | } |
| 577 | } |
| 578 | |
Tobias Klauser | 3bf0033 | 2017-01-10 15:02:16 +0100 | [diff] [blame] | 579 | static void bpf_percpu_lru_populate(struct bpf_lru *lru, void *buf, |
| 580 | u32 node_offset, u32 elem_size, |
| 581 | u32 nr_elems) |
Martin KaFai Lau | 961578b | 2016-11-11 10:55:07 -0800 | [diff] [blame] | 582 | { |
| 583 | u32 i, pcpu_entries; |
| 584 | int cpu; |
| 585 | struct bpf_lru_list *l; |
| 586 | |
| 587 | pcpu_entries = nr_elems / num_possible_cpus(); |
| 588 | |
| 589 | i = 0; |
| 590 | |
| 591 | for_each_possible_cpu(cpu) { |
| 592 | struct bpf_lru_node *node; |
| 593 | |
| 594 | l = per_cpu_ptr(lru->percpu_lru, cpu); |
| 595 | again: |
| 596 | node = (struct bpf_lru_node *)(buf + node_offset); |
| 597 | node->cpu = cpu; |
| 598 | node->type = BPF_LRU_LIST_T_FREE; |
| 599 | node->ref = 0; |
| 600 | list_add(&node->list, &l->lists[BPF_LRU_LIST_T_FREE]); |
| 601 | i++; |
| 602 | buf += elem_size; |
| 603 | if (i == nr_elems) |
| 604 | break; |
| 605 | if (i % pcpu_entries) |
| 606 | goto again; |
| 607 | } |
| 608 | } |
| 609 | |
| 610 | void bpf_lru_populate(struct bpf_lru *lru, void *buf, u32 node_offset, |
| 611 | u32 elem_size, u32 nr_elems) |
| 612 | { |
| 613 | if (lru->percpu) |
| 614 | bpf_percpu_lru_populate(lru, buf, node_offset, elem_size, |
| 615 | nr_elems); |
| 616 | else |
| 617 | bpf_common_lru_populate(lru, buf, node_offset, elem_size, |
| 618 | nr_elems); |
| 619 | } |
| 620 | |
Martin KaFai Lau | 3a08c2f | 2016-11-11 10:55:06 -0800 | [diff] [blame] | 621 | static void bpf_lru_locallist_init(struct bpf_lru_locallist *loc_l, int cpu) |
| 622 | { |
| 623 | int i; |
| 624 | |
| 625 | for (i = 0; i < NR_BPF_LRU_LOCAL_LIST_T; i++) |
| 626 | INIT_LIST_HEAD(&loc_l->lists[i]); |
| 627 | |
| 628 | loc_l->next_steal = cpu; |
| 629 | |
| 630 | raw_spin_lock_init(&loc_l->lock); |
| 631 | } |
| 632 | |
| 633 | static void bpf_lru_list_init(struct bpf_lru_list *l) |
| 634 | { |
| 635 | int i; |
| 636 | |
| 637 | for (i = 0; i < NR_BPF_LRU_LIST_T; i++) |
| 638 | INIT_LIST_HEAD(&l->lists[i]); |
| 639 | |
| 640 | for (i = 0; i < NR_BPF_LRU_LIST_COUNT; i++) |
| 641 | l->counts[i] = 0; |
| 642 | |
| 643 | l->next_inactive_rotation = &l->lists[BPF_LRU_LIST_T_INACTIVE]; |
| 644 | |
| 645 | raw_spin_lock_init(&l->lock); |
| 646 | } |
| 647 | |
Martin KaFai Lau | 961578b | 2016-11-11 10:55:07 -0800 | [diff] [blame] | 648 | int bpf_lru_init(struct bpf_lru *lru, bool percpu, u32 hash_offset, |
Martin KaFai Lau | 3a08c2f | 2016-11-11 10:55:06 -0800 | [diff] [blame] | 649 | del_from_htab_func del_from_htab, void *del_arg) |
| 650 | { |
| 651 | int cpu; |
Martin KaFai Lau | 3a08c2f | 2016-11-11 10:55:06 -0800 | [diff] [blame] | 652 | |
Martin KaFai Lau | 961578b | 2016-11-11 10:55:07 -0800 | [diff] [blame] | 653 | if (percpu) { |
| 654 | lru->percpu_lru = alloc_percpu(struct bpf_lru_list); |
| 655 | if (!lru->percpu_lru) |
| 656 | return -ENOMEM; |
Martin KaFai Lau | 3a08c2f | 2016-11-11 10:55:06 -0800 | [diff] [blame] | 657 | |
Martin KaFai Lau | 961578b | 2016-11-11 10:55:07 -0800 | [diff] [blame] | 658 | for_each_possible_cpu(cpu) { |
| 659 | struct bpf_lru_list *l; |
Martin KaFai Lau | 3a08c2f | 2016-11-11 10:55:06 -0800 | [diff] [blame] | 660 | |
Martin KaFai Lau | 961578b | 2016-11-11 10:55:07 -0800 | [diff] [blame] | 661 | l = per_cpu_ptr(lru->percpu_lru, cpu); |
| 662 | bpf_lru_list_init(l); |
| 663 | } |
| 664 | lru->nr_scans = PERCPU_NR_SCANS; |
| 665 | } else { |
| 666 | struct bpf_common_lru *clru = &lru->common_lru; |
| 667 | |
| 668 | clru->local_list = alloc_percpu(struct bpf_lru_locallist); |
| 669 | if (!clru->local_list) |
| 670 | return -ENOMEM; |
| 671 | |
| 672 | for_each_possible_cpu(cpu) { |
| 673 | struct bpf_lru_locallist *loc_l; |
| 674 | |
| 675 | loc_l = per_cpu_ptr(clru->local_list, cpu); |
| 676 | bpf_lru_locallist_init(loc_l, cpu); |
| 677 | } |
| 678 | |
| 679 | bpf_lru_list_init(&clru->lru_list); |
| 680 | lru->nr_scans = LOCAL_NR_SCANS; |
Martin KaFai Lau | 3a08c2f | 2016-11-11 10:55:06 -0800 | [diff] [blame] | 681 | } |
| 682 | |
Martin KaFai Lau | 961578b | 2016-11-11 10:55:07 -0800 | [diff] [blame] | 683 | lru->percpu = percpu; |
Martin KaFai Lau | 3a08c2f | 2016-11-11 10:55:06 -0800 | [diff] [blame] | 684 | lru->del_from_htab = del_from_htab; |
| 685 | lru->del_arg = del_arg; |
| 686 | lru->hash_offset = hash_offset; |
| 687 | |
| 688 | return 0; |
| 689 | } |
| 690 | |
| 691 | void bpf_lru_destroy(struct bpf_lru *lru) |
| 692 | { |
Martin KaFai Lau | 961578b | 2016-11-11 10:55:07 -0800 | [diff] [blame] | 693 | if (lru->percpu) |
| 694 | free_percpu(lru->percpu_lru); |
| 695 | else |
| 696 | free_percpu(lru->common_lru.local_list); |
Martin KaFai Lau | 3a08c2f | 2016-11-11 10:55:06 -0800 | [diff] [blame] | 697 | } |