1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Common Block IO controller cgroup interface 4 * 5 * Based on ideas and code from CFQ, CFS and BFQ: 6 * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk> 7 * 8 * Copyright (C) 2008 Fabio Checconi <fabio@gandalf.sssup.it> 9 * Paolo Valente <paolo.valente@unimore.it> 10 * 11 * Copyright (C) 2009 Vivek Goyal <vgoyal@redhat.com> 12 * Nauman Rafique <nauman@google.com> 13 * 14 * For policy-specific per-blkcg data: 15 * Copyright (C) 2015 Paolo Valente <paolo.valente@unimore.it> 16 * Arianna Avanzini <avanzini.arianna@gmail.com> 17 */ 18 #include <linux/ioprio.h> 19 #include <linux/kdev_t.h> 20 #include <linux/module.h> 21 #include <linux/sched/signal.h> 22 #include <linux/err.h> 23 #include <linux/blkdev.h> 24 #include <linux/backing-dev.h> 25 #include <linux/slab.h> 26 #include <linux/genhd.h> 27 #include <linux/delay.h> 28 #include <linux/atomic.h> 29 #include <linux/ctype.h> 30 #include <linux/blk-cgroup.h> 31 #include <linux/tracehook.h> 32 #include <linux/psi.h> 33 #include "blk.h" 34 35 #define MAX_KEY_LEN 100 36 37 /* 38 * blkcg_pol_mutex protects blkcg_policy[] and policy [de]activation. 39 * blkcg_pol_register_mutex nests outside of it and synchronizes entire 40 * policy [un]register operations including cgroup file additions / 41 * removals. Putting cgroup file registration outside blkcg_pol_mutex 42 * allows grabbing it from cgroup callbacks. 43 */ 44 static DEFINE_MUTEX(blkcg_pol_register_mutex); 45 static DEFINE_MUTEX(blkcg_pol_mutex); 46 47 struct blkcg blkcg_root; 48 EXPORT_SYMBOL_GPL(blkcg_root); 49 50 struct cgroup_subsys_state * const blkcg_root_css = &blkcg_root.css; 51 EXPORT_SYMBOL_GPL(blkcg_root_css); 52 53 static struct blkcg_policy *blkcg_policy[BLKCG_MAX_POLS]; 54 55 static LIST_HEAD(all_blkcgs); /* protected by blkcg_pol_mutex */ 56 57 bool blkcg_debug_stats = false; 58 static struct workqueue_struct *blkcg_punt_bio_wq; 59 60 static bool blkcg_policy_enabled(struct request_queue *q, 61 const struct blkcg_policy *pol) 62 { 63 return pol && test_bit(pol->plid, q->blkcg_pols); 64 } 65 66 /** 67 * blkg_free - free a blkg 68 * @blkg: blkg to free 69 * 70 * Free @blkg which may be partially allocated. 71 */ 72 static void blkg_free(struct blkcg_gq *blkg) 73 { 74 int i; 75 76 if (!blkg) 77 return; 78 79 for (i = 0; i < BLKCG_MAX_POLS; i++) 80 if (blkg->pd[i]) 81 blkcg_policy[i]->pd_free_fn(blkg->pd[i]); 82 83 free_percpu(blkg->iostat_cpu); 84 percpu_ref_exit(&blkg->refcnt); 85 kfree(blkg); 86 } 87 88 static void __blkg_release(struct rcu_head *rcu) 89 { 90 struct blkcg_gq *blkg = container_of(rcu, struct blkcg_gq, rcu_head); 91 92 WARN_ON(!bio_list_empty(&blkg->async_bios)); 93 94 /* release the blkcg and parent blkg refs this blkg has been holding */ 95 css_put(&blkg->blkcg->css); 96 if (blkg->parent) 97 blkg_put(blkg->parent); 98 blkg_free(blkg); 99 } 100 101 /* 102 * A group is RCU protected, but having an rcu lock does not mean that one 103 * can access all the fields of blkg and assume these are valid. For 104 * example, don't try to follow throtl_data and request queue links. 105 * 106 * Having a reference to blkg under an rcu allows accesses to only values 107 * local to groups like group stats and group rate limits. 108 */ 109 static void blkg_release(struct percpu_ref *ref) 110 { 111 struct blkcg_gq *blkg = container_of(ref, struct blkcg_gq, refcnt); 112 113 call_rcu(&blkg->rcu_head, __blkg_release); 114 } 115 116 static void blkg_async_bio_workfn(struct work_struct *work) 117 { 118 struct blkcg_gq *blkg = container_of(work, struct blkcg_gq, 119 async_bio_work); 120 struct bio_list bios = BIO_EMPTY_LIST; 121 struct bio *bio; 122 struct blk_plug plug; 123 bool need_plug = false; 124 125 /* as long as there are pending bios, @blkg can't go away */ 126 spin_lock_bh(&blkg->async_bio_lock); 127 bio_list_merge(&bios, &blkg->async_bios); 128 bio_list_init(&blkg->async_bios); 129 spin_unlock_bh(&blkg->async_bio_lock); 130 131 /* start plug only when bio_list contains at least 2 bios */ 132 if (bios.head && bios.head->bi_next) { 133 need_plug = true; 134 blk_start_plug(&plug); 135 } 136 while ((bio = bio_list_pop(&bios))) 137 submit_bio(bio); 138 if (need_plug) 139 blk_finish_plug(&plug); 140 } 141 142 /** 143 * blkg_alloc - allocate a blkg 144 * @blkcg: block cgroup the new blkg is associated with 145 * @q: request_queue the new blkg is associated with 146 * @gfp_mask: allocation mask to use 147 * 148 * Allocate a new blkg assocating @blkcg and @q. 149 */ 150 static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct request_queue *q, 151 gfp_t gfp_mask) 152 { 153 struct blkcg_gq *blkg; 154 int i, cpu; 155 156 /* alloc and init base part */ 157 blkg = kzalloc_node(sizeof(*blkg), gfp_mask, q->node); 158 if (!blkg) 159 return NULL; 160 161 if (percpu_ref_init(&blkg->refcnt, blkg_release, 0, gfp_mask)) 162 goto err_free; 163 164 blkg->iostat_cpu = alloc_percpu_gfp(struct blkg_iostat_set, gfp_mask); 165 if (!blkg->iostat_cpu) 166 goto err_free; 167 168 blkg->q = q; 169 INIT_LIST_HEAD(&blkg->q_node); 170 spin_lock_init(&blkg->async_bio_lock); 171 bio_list_init(&blkg->async_bios); 172 INIT_WORK(&blkg->async_bio_work, blkg_async_bio_workfn); 173 blkg->blkcg = blkcg; 174 175 u64_stats_init(&blkg->iostat.sync); 176 for_each_possible_cpu(cpu) 177 u64_stats_init(&per_cpu_ptr(blkg->iostat_cpu, cpu)->sync); 178 179 for (i = 0; i < BLKCG_MAX_POLS; i++) { 180 struct blkcg_policy *pol = blkcg_policy[i]; 181 struct blkg_policy_data *pd; 182 183 if (!blkcg_policy_enabled(q, pol)) 184 continue; 185 186 /* alloc per-policy data and attach it to blkg */ 187 pd = pol->pd_alloc_fn(gfp_mask, q, blkcg); 188 if (!pd) 189 goto err_free; 190 191 blkg->pd[i] = pd; 192 pd->blkg = blkg; 193 pd->plid = i; 194 } 195 196 return blkg; 197 198 err_free: 199 blkg_free(blkg); 200 return NULL; 201 } 202 203 struct blkcg_gq *blkg_lookup_slowpath(struct blkcg *blkcg, 204 struct request_queue *q, bool update_hint) 205 { 206 struct blkcg_gq *blkg; 207 208 /* 209 * Hint didn't match. Look up from the radix tree. Note that the 210 * hint can only be updated under queue_lock as otherwise @blkg 211 * could have already been removed from blkg_tree. The caller is 212 * responsible for grabbing queue_lock if @update_hint. 213 */ 214 blkg = radix_tree_lookup(&blkcg->blkg_tree, q->id); 215 if (blkg && blkg->q == q) { 216 if (update_hint) { 217 lockdep_assert_held(&q->queue_lock); 218 rcu_assign_pointer(blkcg->blkg_hint, blkg); 219 } 220 return blkg; 221 } 222 223 return NULL; 224 } 225 EXPORT_SYMBOL_GPL(blkg_lookup_slowpath); 226 227 /* 228 * If @new_blkg is %NULL, this function tries to allocate a new one as 229 * necessary using %GFP_NOWAIT. @new_blkg is always consumed on return. 230 */ 231 static struct blkcg_gq *blkg_create(struct blkcg *blkcg, 232 struct request_queue *q, 233 struct blkcg_gq *new_blkg) 234 { 235 struct blkcg_gq *blkg; 236 int i, ret; 237 238 WARN_ON_ONCE(!rcu_read_lock_held()); 239 lockdep_assert_held(&q->queue_lock); 240 241 /* request_queue is dying, do not create/recreate a blkg */ 242 if (blk_queue_dying(q)) { 243 ret = -ENODEV; 244 goto err_free_blkg; 245 } 246 247 /* blkg holds a reference to blkcg */ 248 if (!css_tryget_online(&blkcg->css)) { 249 ret = -ENODEV; 250 goto err_free_blkg; 251 } 252 253 /* allocate */ 254 if (!new_blkg) { 255 new_blkg = blkg_alloc(blkcg, q, GFP_NOWAIT | __GFP_NOWARN); 256 if (unlikely(!new_blkg)) { 257 ret = -ENOMEM; 258 goto err_put_css; 259 } 260 } 261 blkg = new_blkg; 262 263 /* link parent */ 264 if (blkcg_parent(blkcg)) { 265 blkg->parent = __blkg_lookup(blkcg_parent(blkcg), q, false); 266 if (WARN_ON_ONCE(!blkg->parent)) { 267 ret = -ENODEV; 268 goto err_put_css; 269 } 270 blkg_get(blkg->parent); 271 } 272 273 /* invoke per-policy init */ 274 for (i = 0; i < BLKCG_MAX_POLS; i++) { 275 struct blkcg_policy *pol = blkcg_policy[i]; 276 277 if (blkg->pd[i] && pol->pd_init_fn) 278 pol->pd_init_fn(blkg->pd[i]); 279 } 280 281 /* insert */ 282 spin_lock(&blkcg->lock); 283 ret = radix_tree_insert(&blkcg->blkg_tree, q->id, blkg); 284 if (likely(!ret)) { 285 hlist_add_head_rcu(&blkg->blkcg_node, &blkcg->blkg_list); 286 list_add(&blkg->q_node, &q->blkg_list); 287 288 for (i = 0; i < BLKCG_MAX_POLS; i++) { 289 struct blkcg_policy *pol = blkcg_policy[i]; 290 291 if (blkg->pd[i] && pol->pd_online_fn) 292 pol->pd_online_fn(blkg->pd[i]); 293 } 294 } 295 blkg->online = true; 296 spin_unlock(&blkcg->lock); 297 298 if (!ret) 299 return blkg; 300 301 /* @blkg failed fully initialized, use the usual release path */ 302 blkg_put(blkg); 303 return ERR_PTR(ret); 304 305 err_put_css: 306 css_put(&blkcg->css); 307 err_free_blkg: 308 blkg_free(new_blkg); 309 return ERR_PTR(ret); 310 } 311 312 /** 313 * blkg_lookup_create - lookup blkg, try to create one if not there 314 * @blkcg: blkcg of interest 315 * @q: request_queue of interest 316 * 317 * Lookup blkg for the @blkcg - @q pair. If it doesn't exist, try to 318 * create one. blkg creation is performed recursively from blkcg_root such 319 * that all non-root blkg's have access to the parent blkg. This function 320 * should be called under RCU read lock and takes @q->queue_lock. 321 * 322 * Returns the blkg or the closest blkg if blkg_create() fails as it walks 323 * down from root. 324 */ 325 static struct blkcg_gq *blkg_lookup_create(struct blkcg *blkcg, 326 struct request_queue *q) 327 { 328 struct blkcg_gq *blkg; 329 unsigned long flags; 330 331 WARN_ON_ONCE(!rcu_read_lock_held()); 332 333 blkg = blkg_lookup(blkcg, q); 334 if (blkg) 335 return blkg; 336 337 spin_lock_irqsave(&q->queue_lock, flags); 338 blkg = __blkg_lookup(blkcg, q, true); 339 if (blkg) 340 goto found; 341 342 /* 343 * Create blkgs walking down from blkcg_root to @blkcg, so that all 344 * non-root blkgs have access to their parents. Returns the closest 345 * blkg to the intended blkg should blkg_create() fail. 346 */ 347 while (true) { 348 struct blkcg *pos = blkcg; 349 struct blkcg *parent = blkcg_parent(blkcg); 350 struct blkcg_gq *ret_blkg = q->root_blkg; 351 352 while (parent) { 353 blkg = __blkg_lookup(parent, q, false); 354 if (blkg) { 355 /* remember closest blkg */ 356 ret_blkg = blkg; 357 break; 358 } 359 pos = parent; 360 parent = blkcg_parent(parent); 361 } 362 363 blkg = blkg_create(pos, q, NULL); 364 if (IS_ERR(blkg)) { 365 blkg = ret_blkg; 366 break; 367 } 368 if (pos == blkcg) 369 break; 370 } 371 372 found: 373 spin_unlock_irqrestore(&q->queue_lock, flags); 374 return blkg; 375 } 376 377 static void blkg_destroy(struct blkcg_gq *blkg) 378 { 379 struct blkcg *blkcg = blkg->blkcg; 380 int i; 381 382 lockdep_assert_held(&blkg->q->queue_lock); 383 lockdep_assert_held(&blkcg->lock); 384 385 /* Something wrong if we are trying to remove same group twice */ 386 WARN_ON_ONCE(list_empty(&blkg->q_node)); 387 WARN_ON_ONCE(hlist_unhashed(&blkg->blkcg_node)); 388 389 for (i = 0; i < BLKCG_MAX_POLS; i++) { 390 struct blkcg_policy *pol = blkcg_policy[i]; 391 392 if (blkg->pd[i] && pol->pd_offline_fn) 393 pol->pd_offline_fn(blkg->pd[i]); 394 } 395 396 blkg->online = false; 397 398 radix_tree_delete(&blkcg->blkg_tree, blkg->q->id); 399 list_del_init(&blkg->q_node); 400 hlist_del_init_rcu(&blkg->blkcg_node); 401 402 /* 403 * Both setting lookup hint to and clearing it from @blkg are done 404 * under queue_lock. If it's not pointing to @blkg now, it never 405 * will. Hint assignment itself can race safely. 406 */ 407 if (rcu_access_pointer(blkcg->blkg_hint) == blkg) 408 rcu_assign_pointer(blkcg->blkg_hint, NULL); 409 410 /* 411 * Put the reference taken at the time of creation so that when all 412 * queues are gone, group can be destroyed. 413 */ 414 percpu_ref_kill(&blkg->refcnt); 415 } 416 417 /** 418 * blkg_destroy_all - destroy all blkgs associated with a request_queue 419 * @q: request_queue of interest 420 * 421 * Destroy all blkgs associated with @q. 422 */ 423 static void blkg_destroy_all(struct request_queue *q) 424 { 425 struct blkcg_gq *blkg, *n; 426 427 spin_lock_irq(&q->queue_lock); 428 list_for_each_entry_safe(blkg, n, &q->blkg_list, q_node) { 429 struct blkcg *blkcg = blkg->blkcg; 430 431 spin_lock(&blkcg->lock); 432 blkg_destroy(blkg); 433 spin_unlock(&blkcg->lock); 434 } 435 436 q->root_blkg = NULL; 437 spin_unlock_irq(&q->queue_lock); 438 } 439 440 static int blkcg_reset_stats(struct cgroup_subsys_state *css, 441 struct cftype *cftype, u64 val) 442 { 443 struct blkcg *blkcg = css_to_blkcg(css); 444 struct blkcg_gq *blkg; 445 int i, cpu; 446 447 mutex_lock(&blkcg_pol_mutex); 448 spin_lock_irq(&blkcg->lock); 449 450 /* 451 * Note that stat reset is racy - it doesn't synchronize against 452 * stat updates. This is a debug feature which shouldn't exist 453 * anyway. If you get hit by a race, retry. 454 */ 455 hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) { 456 for_each_possible_cpu(cpu) { 457 struct blkg_iostat_set *bis = 458 per_cpu_ptr(blkg->iostat_cpu, cpu); 459 memset(bis, 0, sizeof(*bis)); 460 } 461 memset(&blkg->iostat, 0, sizeof(blkg->iostat)); 462 463 for (i = 0; i < BLKCG_MAX_POLS; i++) { 464 struct blkcg_policy *pol = blkcg_policy[i]; 465 466 if (blkg->pd[i] && pol->pd_reset_stats_fn) 467 pol->pd_reset_stats_fn(blkg->pd[i]); 468 } 469 } 470 471 spin_unlock_irq(&blkcg->lock); 472 mutex_unlock(&blkcg_pol_mutex); 473 return 0; 474 } 475 476 const char *blkg_dev_name(struct blkcg_gq *blkg) 477 { 478 /* some drivers (floppy) instantiate a queue w/o disk registered */ 479 if (blkg->q->backing_dev_info->dev) 480 return bdi_dev_name(blkg->q->backing_dev_info); 481 return NULL; 482 } 483 484 /** 485 * blkcg_print_blkgs - helper for printing per-blkg data 486 * @sf: seq_file to print to 487 * @blkcg: blkcg of interest 488 * @prfill: fill function to print out a blkg 489 * @pol: policy in question 490 * @data: data to be passed to @prfill 491 * @show_total: to print out sum of prfill return values or not 492 * 493 * This function invokes @prfill on each blkg of @blkcg if pd for the 494 * policy specified by @pol exists. @prfill is invoked with @sf, the 495 * policy data and @data and the matching queue lock held. If @show_total 496 * is %true, the sum of the return values from @prfill is printed with 497 * "Total" label at the end. 498 * 499 * This is to be used to construct print functions for 500 * cftype->read_seq_string method. 501 */ 502 void blkcg_print_blkgs(struct seq_file *sf, struct blkcg *blkcg, 503 u64 (*prfill)(struct seq_file *, 504 struct blkg_policy_data *, int), 505 const struct blkcg_policy *pol, int data, 506 bool show_total) 507 { 508 struct blkcg_gq *blkg; 509 u64 total = 0; 510 511 rcu_read_lock(); 512 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) { 513 spin_lock_irq(&blkg->q->queue_lock); 514 if (blkcg_policy_enabled(blkg->q, pol)) 515 total += prfill(sf, blkg->pd[pol->plid], data); 516 spin_unlock_irq(&blkg->q->queue_lock); 517 } 518 rcu_read_unlock(); 519 520 if (show_total) 521 seq_printf(sf, "Total %llu\n", (unsigned long long)total); 522 } 523 EXPORT_SYMBOL_GPL(blkcg_print_blkgs); 524 525 /** 526 * __blkg_prfill_u64 - prfill helper for a single u64 value 527 * @sf: seq_file to print to 528 * @pd: policy private data of interest 529 * @v: value to print 530 * 531 * Print @v to @sf for the device assocaited with @pd. 532 */ 533 u64 __blkg_prfill_u64(struct seq_file *sf, struct blkg_policy_data *pd, u64 v) 534 { 535 const char *dname = blkg_dev_name(pd->blkg); 536 537 if (!dname) 538 return 0; 539 540 seq_printf(sf, "%s %llu\n", dname, (unsigned long long)v); 541 return v; 542 } 543 EXPORT_SYMBOL_GPL(__blkg_prfill_u64); 544 545 /* Performs queue bypass and policy enabled checks then looks up blkg. */ 546 static struct blkcg_gq *blkg_lookup_check(struct blkcg *blkcg, 547 const struct blkcg_policy *pol, 548 struct request_queue *q) 549 { 550 WARN_ON_ONCE(!rcu_read_lock_held()); 551 lockdep_assert_held(&q->queue_lock); 552 553 if (!blkcg_policy_enabled(q, pol)) 554 return ERR_PTR(-EOPNOTSUPP); 555 return __blkg_lookup(blkcg, q, true /* update_hint */); 556 } 557 558 /** 559 * blkg_conf_prep - parse and prepare for per-blkg config update 560 * @inputp: input string pointer 561 * 562 * Parse the device node prefix part, MAJ:MIN, of per-blkg config update 563 * from @input and get and return the matching gendisk. *@inputp is 564 * updated to point past the device node prefix. Returns an ERR_PTR() 565 * value on error. 566 * 567 * Use this function iff blkg_conf_prep() can't be used for some reason. 568 */ 569 struct gendisk *blkcg_conf_get_disk(char **inputp) 570 { 571 char *input = *inputp; 572 unsigned int major, minor; 573 struct gendisk *disk; 574 int key_len, part; 575 576 if (sscanf(input, "%u:%u%n", &major, &minor, &key_len) != 2) 577 return ERR_PTR(-EINVAL); 578 579 input += key_len; 580 if (!isspace(*input)) 581 return ERR_PTR(-EINVAL); 582 input = skip_spaces(input); 583 584 disk = get_gendisk(MKDEV(major, minor), &part); 585 if (!disk) 586 return ERR_PTR(-ENODEV); 587 if (part) { 588 put_disk_and_module(disk); 589 return ERR_PTR(-ENODEV); 590 } 591 592 *inputp = input; 593 return disk; 594 } 595 596 /** 597 * blkg_conf_prep - parse and prepare for per-blkg config update 598 * @blkcg: target block cgroup 599 * @pol: target policy 600 * @input: input string 601 * @ctx: blkg_conf_ctx to be filled 602 * 603 * Parse per-blkg config update from @input and initialize @ctx with the 604 * result. @ctx->blkg points to the blkg to be updated and @ctx->body the 605 * part of @input following MAJ:MIN. This function returns with RCU read 606 * lock and queue lock held and must be paired with blkg_conf_finish(). 607 */ 608 int blkg_conf_prep(struct blkcg *blkcg, const struct blkcg_policy *pol, 609 char *input, struct blkg_conf_ctx *ctx) 610 __acquires(rcu) __acquires(&disk->queue->queue_lock) 611 { 612 struct gendisk *disk; 613 struct request_queue *q; 614 struct blkcg_gq *blkg; 615 int ret; 616 617 disk = blkcg_conf_get_disk(&input); 618 if (IS_ERR(disk)) 619 return PTR_ERR(disk); 620 621 q = disk->queue; 622 623 rcu_read_lock(); 624 spin_lock_irq(&q->queue_lock); 625 626 blkg = blkg_lookup_check(blkcg, pol, q); 627 if (IS_ERR(blkg)) { 628 ret = PTR_ERR(blkg); 629 goto fail_unlock; 630 } 631 632 if (blkg) 633 goto success; 634 635 /* 636 * Create blkgs walking down from blkcg_root to @blkcg, so that all 637 * non-root blkgs have access to their parents. 638 */ 639 while (true) { 640 struct blkcg *pos = blkcg; 641 struct blkcg *parent; 642 struct blkcg_gq *new_blkg; 643 644 parent = blkcg_parent(blkcg); 645 while (parent && !__blkg_lookup(parent, q, false)) { 646 pos = parent; 647 parent = blkcg_parent(parent); 648 } 649 650 /* Drop locks to do new blkg allocation with GFP_KERNEL. */ 651 spin_unlock_irq(&q->queue_lock); 652 rcu_read_unlock(); 653 654 new_blkg = blkg_alloc(pos, q, GFP_KERNEL); 655 if (unlikely(!new_blkg)) { 656 ret = -ENOMEM; 657 goto fail; 658 } 659 660 rcu_read_lock(); 661 spin_lock_irq(&q->queue_lock); 662 663 blkg = blkg_lookup_check(pos, pol, q); 664 if (IS_ERR(blkg)) { 665 ret = PTR_ERR(blkg); 666 goto fail_unlock; 667 } 668 669 if (blkg) { 670 blkg_free(new_blkg); 671 } else { 672 blkg = blkg_create(pos, q, new_blkg); 673 if (IS_ERR(blkg)) { 674 ret = PTR_ERR(blkg); 675 goto fail_unlock; 676 } 677 } 678 679 if (pos == blkcg) 680 goto success; 681 } 682 success: 683 ctx->disk = disk; 684 ctx->blkg = blkg; 685 ctx->body = input; 686 return 0; 687 688 fail_unlock: 689 spin_unlock_irq(&q->queue_lock); 690 rcu_read_unlock(); 691 fail: 692 put_disk_and_module(disk); 693 /* 694 * If queue was bypassing, we should retry. Do so after a 695 * short msleep(). It isn't strictly necessary but queue 696 * can be bypassing for some time and it's always nice to 697 * avoid busy looping. 698 */ 699 if (ret == -EBUSY) { 700 msleep(10); 701 ret = restart_syscall(); 702 } 703 return ret; 704 } 705 EXPORT_SYMBOL_GPL(blkg_conf_prep); 706 707 /** 708 * blkg_conf_finish - finish up per-blkg config update 709 * @ctx: blkg_conf_ctx intiailized by blkg_conf_prep() 710 * 711 * Finish up after per-blkg config update. This function must be paired 712 * with blkg_conf_prep(). 713 */ 714 void blkg_conf_finish(struct blkg_conf_ctx *ctx) 715 __releases(&ctx->disk->queue->queue_lock) __releases(rcu) 716 { 717 spin_unlock_irq(&ctx->disk->queue->queue_lock); 718 rcu_read_unlock(); 719 put_disk_and_module(ctx->disk); 720 } 721 EXPORT_SYMBOL_GPL(blkg_conf_finish); 722 723 static void blkg_iostat_set(struct blkg_iostat *dst, struct blkg_iostat *src) 724 { 725 int i; 726 727 for (i = 0; i < BLKG_IOSTAT_NR; i++) { 728 dst->bytes[i] = src->bytes[i]; 729 dst->ios[i] = src->ios[i]; 730 } 731 } 732 733 static void blkg_iostat_add(struct blkg_iostat *dst, struct blkg_iostat *src) 734 { 735 int i; 736 737 for (i = 0; i < BLKG_IOSTAT_NR; i++) { 738 dst->bytes[i] += src->bytes[i]; 739 dst->ios[i] += src->ios[i]; 740 } 741 } 742 743 static void blkg_iostat_sub(struct blkg_iostat *dst, struct blkg_iostat *src) 744 { 745 int i; 746 747 for (i = 0; i < BLKG_IOSTAT_NR; i++) { 748 dst->bytes[i] -= src->bytes[i]; 749 dst->ios[i] -= src->ios[i]; 750 } 751 } 752 753 static void blkcg_rstat_flush(struct cgroup_subsys_state *css, int cpu) 754 { 755 struct blkcg *blkcg = css_to_blkcg(css); 756 struct blkcg_gq *blkg; 757 758 rcu_read_lock(); 759 760 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) { 761 struct blkcg_gq *parent = blkg->parent; 762 struct blkg_iostat_set *bisc = per_cpu_ptr(blkg->iostat_cpu, cpu); 763 struct blkg_iostat cur, delta; 764 unsigned int seq; 765 766 /* fetch the current per-cpu values */ 767 do { 768 seq = u64_stats_fetch_begin(&bisc->sync); 769 blkg_iostat_set(&cur, &bisc->cur); 770 } while (u64_stats_fetch_retry(&bisc->sync, seq)); 771 772 /* propagate percpu delta to global */ 773 u64_stats_update_begin(&blkg->iostat.sync); 774 blkg_iostat_set(&delta, &cur); 775 blkg_iostat_sub(&delta, &bisc->last); 776 blkg_iostat_add(&blkg->iostat.cur, &delta); 777 blkg_iostat_add(&bisc->last, &delta); 778 u64_stats_update_end(&blkg->iostat.sync); 779 780 /* propagate global delta to parent */ 781 if (parent) { 782 u64_stats_update_begin(&parent->iostat.sync); 783 blkg_iostat_set(&delta, &blkg->iostat.cur); 784 blkg_iostat_sub(&delta, &blkg->iostat.last); 785 blkg_iostat_add(&parent->iostat.cur, &delta); 786 blkg_iostat_add(&blkg->iostat.last, &delta); 787 u64_stats_update_end(&parent->iostat.sync); 788 } 789 } 790 791 rcu_read_unlock(); 792 } 793 794 /* 795 * The rstat algorithms intentionally don't handle the root cgroup to avoid 796 * incurring overhead when no cgroups are defined. For that reason, 797 * cgroup_rstat_flush in blkcg_print_stat does not actually fill out the 798 * iostat in the root cgroup's blkcg_gq. 799 * 800 * However, we would like to re-use the printing code between the root and 801 * non-root cgroups to the extent possible. For that reason, we simulate 802 * flushing the root cgroup's stats by explicitly filling in the iostat 803 * with disk level statistics. 804 */ 805 static void blkcg_fill_root_iostats(void) 806 { 807 struct class_dev_iter iter; 808 struct device *dev; 809 810 class_dev_iter_init(&iter, &block_class, NULL, &disk_type); 811 while ((dev = class_dev_iter_next(&iter))) { 812 struct gendisk *disk = dev_to_disk(dev); 813 struct hd_struct *part = disk_get_part(disk, 0); 814 struct blkcg_gq *blkg = blk_queue_root_blkg(disk->queue); 815 struct blkg_iostat tmp; 816 int cpu; 817 818 memset(&tmp, 0, sizeof(tmp)); 819 for_each_possible_cpu(cpu) { 820 struct disk_stats *cpu_dkstats; 821 822 cpu_dkstats = per_cpu_ptr(part->dkstats, cpu); 823 tmp.ios[BLKG_IOSTAT_READ] += 824 cpu_dkstats->ios[STAT_READ]; 825 tmp.ios[BLKG_IOSTAT_WRITE] += 826 cpu_dkstats->ios[STAT_WRITE]; 827 tmp.ios[BLKG_IOSTAT_DISCARD] += 828 cpu_dkstats->ios[STAT_DISCARD]; 829 // convert sectors to bytes 830 tmp.bytes[BLKG_IOSTAT_READ] += 831 cpu_dkstats->sectors[STAT_READ] << 9; 832 tmp.bytes[BLKG_IOSTAT_WRITE] += 833 cpu_dkstats->sectors[STAT_WRITE] << 9; 834 tmp.bytes[BLKG_IOSTAT_DISCARD] += 835 cpu_dkstats->sectors[STAT_DISCARD] << 9; 836 837 u64_stats_update_begin(&blkg->iostat.sync); 838 blkg_iostat_set(&blkg->iostat.cur, &tmp); 839 u64_stats_update_end(&blkg->iostat.sync); 840 } 841 } 842 } 843 844 static int blkcg_print_stat(struct seq_file *sf, void *v) 845 { 846 struct blkcg *blkcg = css_to_blkcg(seq_css(sf)); 847 struct blkcg_gq *blkg; 848 849 if (!seq_css(sf)->parent) 850 blkcg_fill_root_iostats(); 851 else 852 cgroup_rstat_flush(blkcg->css.cgroup); 853 854 rcu_read_lock(); 855 856 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) { 857 struct blkg_iostat_set *bis = &blkg->iostat; 858 const char *dname; 859 char *buf; 860 u64 rbytes, wbytes, rios, wios, dbytes, dios; 861 size_t size = seq_get_buf(sf, &buf), off = 0; 862 int i; 863 bool has_stats = false; 864 unsigned seq; 865 866 spin_lock_irq(&blkg->q->queue_lock); 867 868 if (!blkg->online) 869 goto skip; 870 871 dname = blkg_dev_name(blkg); 872 if (!dname) 873 goto skip; 874 875 /* 876 * Hooray string manipulation, count is the size written NOT 877 * INCLUDING THE \0, so size is now count+1 less than what we 878 * had before, but we want to start writing the next bit from 879 * the \0 so we only add count to buf. 880 */ 881 off += scnprintf(buf+off, size-off, "%s ", dname); 882 883 do { 884 seq = u64_stats_fetch_begin(&bis->sync); 885 886 rbytes = bis->cur.bytes[BLKG_IOSTAT_READ]; 887 wbytes = bis->cur.bytes[BLKG_IOSTAT_WRITE]; 888 dbytes = bis->cur.bytes[BLKG_IOSTAT_DISCARD]; 889 rios = bis->cur.ios[BLKG_IOSTAT_READ]; 890 wios = bis->cur.ios[BLKG_IOSTAT_WRITE]; 891 dios = bis->cur.ios[BLKG_IOSTAT_DISCARD]; 892 } while (u64_stats_fetch_retry(&bis->sync, seq)); 893 894 if (rbytes || wbytes || rios || wios) { 895 has_stats = true; 896 off += scnprintf(buf+off, size-off, 897 "rbytes=%llu wbytes=%llu rios=%llu wios=%llu dbytes=%llu dios=%llu", 898 rbytes, wbytes, rios, wios, 899 dbytes, dios); 900 } 901 902 if (blkcg_debug_stats && atomic_read(&blkg->use_delay)) { 903 has_stats = true; 904 off += scnprintf(buf+off, size-off, 905 " use_delay=%d delay_nsec=%llu", 906 atomic_read(&blkg->use_delay), 907 (unsigned long long)atomic64_read(&blkg->delay_nsec)); 908 } 909 910 for (i = 0; i < BLKCG_MAX_POLS; i++) { 911 struct blkcg_policy *pol = blkcg_policy[i]; 912 size_t written; 913 914 if (!blkg->pd[i] || !pol->pd_stat_fn) 915 continue; 916 917 written = pol->pd_stat_fn(blkg->pd[i], buf+off, size-off); 918 if (written) 919 has_stats = true; 920 off += written; 921 } 922 923 if (has_stats) { 924 if (off < size - 1) { 925 off += scnprintf(buf+off, size-off, "\n"); 926 seq_commit(sf, off); 927 } else { 928 seq_commit(sf, -1); 929 } 930 } 931 skip: 932 spin_unlock_irq(&blkg->q->queue_lock); 933 } 934 935 rcu_read_unlock(); 936 return 0; 937 } 938 939 static struct cftype blkcg_files[] = { 940 { 941 .name = "stat", 942 .seq_show = blkcg_print_stat, 943 }, 944 { } /* terminate */ 945 }; 946 947 static struct cftype blkcg_legacy_files[] = { 948 { 949 .name = "reset_stats", 950 .write_u64 = blkcg_reset_stats, 951 }, 952 { } /* terminate */ 953 }; 954 955 /* 956 * blkcg destruction is a three-stage process. 957 * 958 * 1. Destruction starts. The blkcg_css_offline() callback is invoked 959 * which offlines writeback. Here we tie the next stage of blkg destruction 960 * to the completion of writeback associated with the blkcg. This lets us 961 * avoid punting potentially large amounts of outstanding writeback to root 962 * while maintaining any ongoing policies. The next stage is triggered when 963 * the nr_cgwbs count goes to zero. 964 * 965 * 2. When the nr_cgwbs count goes to zero, blkcg_destroy_blkgs() is called 966 * and handles the destruction of blkgs. Here the css reference held by 967 * the blkg is put back eventually allowing blkcg_css_free() to be called. 968 * This work may occur in cgwb_release_workfn() on the cgwb_release 969 * workqueue. Any submitted ios that fail to get the blkg ref will be 970 * punted to the root_blkg. 971 * 972 * 3. Once the blkcg ref count goes to zero, blkcg_css_free() is called. 973 * This finally frees the blkcg. 974 */ 975 976 /** 977 * blkcg_css_offline - cgroup css_offline callback 978 * @css: css of interest 979 * 980 * This function is called when @css is about to go away. Here the cgwbs are 981 * offlined first and only once writeback associated with the blkcg has 982 * finished do we start step 2 (see above). 983 */ 984 static void blkcg_css_offline(struct cgroup_subsys_state *css) 985 { 986 struct blkcg *blkcg = css_to_blkcg(css); 987 988 /* this prevents anyone from attaching or migrating to this blkcg */ 989 wb_blkcg_offline(blkcg); 990 991 /* put the base online pin allowing step 2 to be triggered */ 992 blkcg_unpin_online(blkcg); 993 } 994 995 /** 996 * blkcg_destroy_blkgs - responsible for shooting down blkgs 997 * @blkcg: blkcg of interest 998 * 999 * blkgs should be removed while holding both q and blkcg locks. As blkcg lock 1000 * is nested inside q lock, this function performs reverse double lock dancing. 1001 * Destroying the blkgs releases the reference held on the blkcg's css allowing 1002 * blkcg_css_free to eventually be called. 1003 * 1004 * This is the blkcg counterpart of ioc_release_fn(). 1005 */ 1006 void blkcg_destroy_blkgs(struct blkcg *blkcg) 1007 { 1008 spin_lock_irq(&blkcg->lock); 1009 1010 while (!hlist_empty(&blkcg->blkg_list)) { 1011 struct blkcg_gq *blkg = hlist_entry(blkcg->blkg_list.first, 1012 struct blkcg_gq, blkcg_node); 1013 struct request_queue *q = blkg->q; 1014 1015 if (spin_trylock(&q->queue_lock)) { 1016 blkg_destroy(blkg); 1017 spin_unlock(&q->queue_lock); 1018 } else { 1019 spin_unlock_irq(&blkcg->lock); 1020 cpu_relax(); 1021 spin_lock_irq(&blkcg->lock); 1022 } 1023 } 1024 1025 spin_unlock_irq(&blkcg->lock); 1026 } 1027 1028 static void blkcg_css_free(struct cgroup_subsys_state *css) 1029 { 1030 struct blkcg *blkcg = css_to_blkcg(css); 1031 int i; 1032 1033 mutex_lock(&blkcg_pol_mutex); 1034 1035 list_del(&blkcg->all_blkcgs_node); 1036 1037 for (i = 0; i < BLKCG_MAX_POLS; i++) 1038 if (blkcg->cpd[i]) 1039 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]); 1040 1041 mutex_unlock(&blkcg_pol_mutex); 1042 1043 kfree(blkcg); 1044 } 1045 1046 static struct cgroup_subsys_state * 1047 blkcg_css_alloc(struct cgroup_subsys_state *parent_css) 1048 { 1049 struct blkcg *blkcg; 1050 struct cgroup_subsys_state *ret; 1051 int i; 1052 1053 mutex_lock(&blkcg_pol_mutex); 1054 1055 if (!parent_css) { 1056 blkcg = &blkcg_root; 1057 } else { 1058 blkcg = kzalloc(sizeof(*blkcg), GFP_KERNEL); 1059 if (!blkcg) { 1060 ret = ERR_PTR(-ENOMEM); 1061 goto unlock; 1062 } 1063 } 1064 1065 for (i = 0; i < BLKCG_MAX_POLS ; i++) { 1066 struct blkcg_policy *pol = blkcg_policy[i]; 1067 struct blkcg_policy_data *cpd; 1068 1069 /* 1070 * If the policy hasn't been attached yet, wait for it 1071 * to be attached before doing anything else. Otherwise, 1072 * check if the policy requires any specific per-cgroup 1073 * data: if it does, allocate and initialize it. 1074 */ 1075 if (!pol || !pol->cpd_alloc_fn) 1076 continue; 1077 1078 cpd = pol->cpd_alloc_fn(GFP_KERNEL); 1079 if (!cpd) { 1080 ret = ERR_PTR(-ENOMEM); 1081 goto free_pd_blkcg; 1082 } 1083 blkcg->cpd[i] = cpd; 1084 cpd->blkcg = blkcg; 1085 cpd->plid = i; 1086 if (pol->cpd_init_fn) 1087 pol->cpd_init_fn(cpd); 1088 } 1089 1090 spin_lock_init(&blkcg->lock); 1091 refcount_set(&blkcg->online_pin, 1); 1092 INIT_RADIX_TREE(&blkcg->blkg_tree, GFP_NOWAIT | __GFP_NOWARN); 1093 INIT_HLIST_HEAD(&blkcg->blkg_list); 1094 #ifdef CONFIG_CGROUP_WRITEBACK 1095 INIT_LIST_HEAD(&blkcg->cgwb_list); 1096 #endif 1097 list_add_tail(&blkcg->all_blkcgs_node, &all_blkcgs); 1098 1099 mutex_unlock(&blkcg_pol_mutex); 1100 return &blkcg->css; 1101 1102 free_pd_blkcg: 1103 for (i--; i >= 0; i--) 1104 if (blkcg->cpd[i]) 1105 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]); 1106 1107 if (blkcg != &blkcg_root) 1108 kfree(blkcg); 1109 unlock: 1110 mutex_unlock(&blkcg_pol_mutex); 1111 return ret; 1112 } 1113 1114 static int blkcg_css_online(struct cgroup_subsys_state *css) 1115 { 1116 struct blkcg *blkcg = css_to_blkcg(css); 1117 struct blkcg *parent = blkcg_parent(blkcg); 1118 1119 /* 1120 * blkcg_pin_online() is used to delay blkcg offline so that blkgs 1121 * don't go offline while cgwbs are still active on them. Pin the 1122 * parent so that offline always happens towards the root. 1123 */ 1124 if (parent) 1125 blkcg_pin_online(parent); 1126 return 0; 1127 } 1128 1129 /** 1130 * blkcg_init_queue - initialize blkcg part of request queue 1131 * @q: request_queue to initialize 1132 * 1133 * Called from blk_alloc_queue(). Responsible for initializing blkcg 1134 * part of new request_queue @q. 1135 * 1136 * RETURNS: 1137 * 0 on success, -errno on failure. 1138 */ 1139 int blkcg_init_queue(struct request_queue *q) 1140 { 1141 struct blkcg_gq *new_blkg, *blkg; 1142 bool preloaded; 1143 int ret; 1144 1145 new_blkg = blkg_alloc(&blkcg_root, q, GFP_KERNEL); 1146 if (!new_blkg) 1147 return -ENOMEM; 1148 1149 preloaded = !radix_tree_preload(GFP_KERNEL); 1150 1151 /* Make sure the root blkg exists. */ 1152 rcu_read_lock(); 1153 spin_lock_irq(&q->queue_lock); 1154 blkg = blkg_create(&blkcg_root, q, new_blkg); 1155 if (IS_ERR(blkg)) 1156 goto err_unlock; 1157 q->root_blkg = blkg; 1158 spin_unlock_irq(&q->queue_lock); 1159 rcu_read_unlock(); 1160 1161 if (preloaded) 1162 radix_tree_preload_end(); 1163 1164 ret = blk_throtl_init(q); 1165 if (ret) 1166 goto err_destroy_all; 1167 1168 ret = blk_iolatency_init(q); 1169 if (ret) { 1170 blk_throtl_exit(q); 1171 goto err_destroy_all; 1172 } 1173 return 0; 1174 1175 err_destroy_all: 1176 blkg_destroy_all(q); 1177 return ret; 1178 err_unlock: 1179 spin_unlock_irq(&q->queue_lock); 1180 rcu_read_unlock(); 1181 if (preloaded) 1182 radix_tree_preload_end(); 1183 return PTR_ERR(blkg); 1184 } 1185 1186 /** 1187 * blkcg_exit_queue - exit and release blkcg part of request_queue 1188 * @q: request_queue being released 1189 * 1190 * Called from blk_exit_queue(). Responsible for exiting blkcg part. 1191 */ 1192 void blkcg_exit_queue(struct request_queue *q) 1193 { 1194 blkg_destroy_all(q); 1195 blk_throtl_exit(q); 1196 } 1197 1198 /* 1199 * We cannot support shared io contexts, as we have no mean to support 1200 * two tasks with the same ioc in two different groups without major rework 1201 * of the main cic data structures. For now we allow a task to change 1202 * its cgroup only if it's the only owner of its ioc. 1203 */ 1204 static int blkcg_can_attach(struct cgroup_taskset *tset) 1205 { 1206 struct task_struct *task; 1207 struct cgroup_subsys_state *dst_css; 1208 struct io_context *ioc; 1209 int ret = 0; 1210 1211 /* task_lock() is needed to avoid races with exit_io_context() */ 1212 cgroup_taskset_for_each(task, dst_css, tset) { 1213 task_lock(task); 1214 ioc = task->io_context; 1215 if (ioc && atomic_read(&ioc->nr_tasks) > 1) 1216 ret = -EINVAL; 1217 task_unlock(task); 1218 if (ret) 1219 break; 1220 } 1221 return ret; 1222 } 1223 1224 static void blkcg_bind(struct cgroup_subsys_state *root_css) 1225 { 1226 int i; 1227 1228 mutex_lock(&blkcg_pol_mutex); 1229 1230 for (i = 0; i < BLKCG_MAX_POLS; i++) { 1231 struct blkcg_policy *pol = blkcg_policy[i]; 1232 struct blkcg *blkcg; 1233 1234 if (!pol || !pol->cpd_bind_fn) 1235 continue; 1236 1237 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) 1238 if (blkcg->cpd[pol->plid]) 1239 pol->cpd_bind_fn(blkcg->cpd[pol->plid]); 1240 } 1241 mutex_unlock(&blkcg_pol_mutex); 1242 } 1243 1244 static void blkcg_exit(struct task_struct *tsk) 1245 { 1246 if (tsk->throttle_queue) 1247 blk_put_queue(tsk->throttle_queue); 1248 tsk->throttle_queue = NULL; 1249 } 1250 1251 struct cgroup_subsys io_cgrp_subsys = { 1252 .css_alloc = blkcg_css_alloc, 1253 .css_online = blkcg_css_online, 1254 .css_offline = blkcg_css_offline, 1255 .css_free = blkcg_css_free, 1256 .can_attach = blkcg_can_attach, 1257 .css_rstat_flush = blkcg_rstat_flush, 1258 .bind = blkcg_bind, 1259 .dfl_cftypes = blkcg_files, 1260 .legacy_cftypes = blkcg_legacy_files, 1261 .legacy_name = "blkio", 1262 .exit = blkcg_exit, 1263 #ifdef CONFIG_MEMCG 1264 /* 1265 * This ensures that, if available, memcg is automatically enabled 1266 * together on the default hierarchy so that the owner cgroup can 1267 * be retrieved from writeback pages. 1268 */ 1269 .depends_on = 1 << memory_cgrp_id, 1270 #endif 1271 }; 1272 EXPORT_SYMBOL_GPL(io_cgrp_subsys); 1273 1274 /** 1275 * blkcg_activate_policy - activate a blkcg policy on a request_queue 1276 * @q: request_queue of interest 1277 * @pol: blkcg policy to activate 1278 * 1279 * Activate @pol on @q. Requires %GFP_KERNEL context. @q goes through 1280 * bypass mode to populate its blkgs with policy_data for @pol. 1281 * 1282 * Activation happens with @q bypassed, so nobody would be accessing blkgs 1283 * from IO path. Update of each blkg is protected by both queue and blkcg 1284 * locks so that holding either lock and testing blkcg_policy_enabled() is 1285 * always enough for dereferencing policy data. 1286 * 1287 * The caller is responsible for synchronizing [de]activations and policy 1288 * [un]registerations. Returns 0 on success, -errno on failure. 1289 */ 1290 int blkcg_activate_policy(struct request_queue *q, 1291 const struct blkcg_policy *pol) 1292 { 1293 struct blkg_policy_data *pd_prealloc = NULL; 1294 struct blkcg_gq *blkg, *pinned_blkg = NULL; 1295 int ret; 1296 1297 if (blkcg_policy_enabled(q, pol)) 1298 return 0; 1299 1300 if (queue_is_mq(q)) 1301 blk_mq_freeze_queue(q); 1302 retry: 1303 spin_lock_irq(&q->queue_lock); 1304 1305 /* blkg_list is pushed at the head, reverse walk to allocate parents first */ 1306 list_for_each_entry_reverse(blkg, &q->blkg_list, q_node) { 1307 struct blkg_policy_data *pd; 1308 1309 if (blkg->pd[pol->plid]) 1310 continue; 1311 1312 /* If prealloc matches, use it; otherwise try GFP_NOWAIT */ 1313 if (blkg == pinned_blkg) { 1314 pd = pd_prealloc; 1315 pd_prealloc = NULL; 1316 } else { 1317 pd = pol->pd_alloc_fn(GFP_NOWAIT | __GFP_NOWARN, q, 1318 blkg->blkcg); 1319 } 1320 1321 if (!pd) { 1322 /* 1323 * GFP_NOWAIT failed. Free the existing one and 1324 * prealloc for @blkg w/ GFP_KERNEL. 1325 */ 1326 if (pinned_blkg) 1327 blkg_put(pinned_blkg); 1328 blkg_get(blkg); 1329 pinned_blkg = blkg; 1330 1331 spin_unlock_irq(&q->queue_lock); 1332 1333 if (pd_prealloc) 1334 pol->pd_free_fn(pd_prealloc); 1335 pd_prealloc = pol->pd_alloc_fn(GFP_KERNEL, q, 1336 blkg->blkcg); 1337 if (pd_prealloc) 1338 goto retry; 1339 else 1340 goto enomem; 1341 } 1342 1343 blkg->pd[pol->plid] = pd; 1344 pd->blkg = blkg; 1345 pd->plid = pol->plid; 1346 } 1347 1348 /* all allocated, init in the same order */ 1349 if (pol->pd_init_fn) 1350 list_for_each_entry_reverse(blkg, &q->blkg_list, q_node) 1351 pol->pd_init_fn(blkg->pd[pol->plid]); 1352 1353 __set_bit(pol->plid, q->blkcg_pols); 1354 ret = 0; 1355 1356 spin_unlock_irq(&q->queue_lock); 1357 out: 1358 if (queue_is_mq(q)) 1359 blk_mq_unfreeze_queue(q); 1360 if (pinned_blkg) 1361 blkg_put(pinned_blkg); 1362 if (pd_prealloc) 1363 pol->pd_free_fn(pd_prealloc); 1364 return ret; 1365 1366 enomem: 1367 /* alloc failed, nothing's initialized yet, free everything */ 1368 spin_lock_irq(&q->queue_lock); 1369 list_for_each_entry(blkg, &q->blkg_list, q_node) { 1370 if (blkg->pd[pol->plid]) { 1371 pol->pd_free_fn(blkg->pd[pol->plid]); 1372 blkg->pd[pol->plid] = NULL; 1373 } 1374 } 1375 spin_unlock_irq(&q->queue_lock); 1376 ret = -ENOMEM; 1377 goto out; 1378 } 1379 EXPORT_SYMBOL_GPL(blkcg_activate_policy); 1380 1381 /** 1382 * blkcg_deactivate_policy - deactivate a blkcg policy on a request_queue 1383 * @q: request_queue of interest 1384 * @pol: blkcg policy to deactivate 1385 * 1386 * Deactivate @pol on @q. Follows the same synchronization rules as 1387 * blkcg_activate_policy(). 1388 */ 1389 void blkcg_deactivate_policy(struct request_queue *q, 1390 const struct blkcg_policy *pol) 1391 { 1392 struct blkcg_gq *blkg; 1393 1394 if (!blkcg_policy_enabled(q, pol)) 1395 return; 1396 1397 if (queue_is_mq(q)) 1398 blk_mq_freeze_queue(q); 1399 1400 spin_lock_irq(&q->queue_lock); 1401 1402 __clear_bit(pol->plid, q->blkcg_pols); 1403 1404 list_for_each_entry(blkg, &q->blkg_list, q_node) { 1405 if (blkg->pd[pol->plid]) { 1406 if (pol->pd_offline_fn) 1407 pol->pd_offline_fn(blkg->pd[pol->plid]); 1408 pol->pd_free_fn(blkg->pd[pol->plid]); 1409 blkg->pd[pol->plid] = NULL; 1410 } 1411 } 1412 1413 spin_unlock_irq(&q->queue_lock); 1414 1415 if (queue_is_mq(q)) 1416 blk_mq_unfreeze_queue(q); 1417 } 1418 EXPORT_SYMBOL_GPL(blkcg_deactivate_policy); 1419 1420 /** 1421 * blkcg_policy_register - register a blkcg policy 1422 * @pol: blkcg policy to register 1423 * 1424 * Register @pol with blkcg core. Might sleep and @pol may be modified on 1425 * successful registration. Returns 0 on success and -errno on failure. 1426 */ 1427 int blkcg_policy_register(struct blkcg_policy *pol) 1428 { 1429 struct blkcg *blkcg; 1430 int i, ret; 1431 1432 mutex_lock(&blkcg_pol_register_mutex); 1433 mutex_lock(&blkcg_pol_mutex); 1434 1435 /* find an empty slot */ 1436 ret = -ENOSPC; 1437 for (i = 0; i < BLKCG_MAX_POLS; i++) 1438 if (!blkcg_policy[i]) 1439 break; 1440 if (i >= BLKCG_MAX_POLS) { 1441 pr_warn("blkcg_policy_register: BLKCG_MAX_POLS too small\n"); 1442 goto err_unlock; 1443 } 1444 1445 /* Make sure cpd/pd_alloc_fn and cpd/pd_free_fn in pairs */ 1446 if ((!pol->cpd_alloc_fn ^ !pol->cpd_free_fn) || 1447 (!pol->pd_alloc_fn ^ !pol->pd_free_fn)) 1448 goto err_unlock; 1449 1450 /* register @pol */ 1451 pol->plid = i; 1452 blkcg_policy[pol->plid] = pol; 1453 1454 /* allocate and install cpd's */ 1455 if (pol->cpd_alloc_fn) { 1456 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) { 1457 struct blkcg_policy_data *cpd; 1458 1459 cpd = pol->cpd_alloc_fn(GFP_KERNEL); 1460 if (!cpd) 1461 goto err_free_cpds; 1462 1463 blkcg->cpd[pol->plid] = cpd; 1464 cpd->blkcg = blkcg; 1465 cpd->plid = pol->plid; 1466 if (pol->cpd_init_fn) 1467 pol->cpd_init_fn(cpd); 1468 } 1469 } 1470 1471 mutex_unlock(&blkcg_pol_mutex); 1472 1473 /* everything is in place, add intf files for the new policy */ 1474 if (pol->dfl_cftypes) 1475 WARN_ON(cgroup_add_dfl_cftypes(&io_cgrp_subsys, 1476 pol->dfl_cftypes)); 1477 if (pol->legacy_cftypes) 1478 WARN_ON(cgroup_add_legacy_cftypes(&io_cgrp_subsys, 1479 pol->legacy_cftypes)); 1480 mutex_unlock(&blkcg_pol_register_mutex); 1481 return 0; 1482 1483 err_free_cpds: 1484 if (pol->cpd_free_fn) { 1485 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) { 1486 if (blkcg->cpd[pol->plid]) { 1487 pol->cpd_free_fn(blkcg->cpd[pol->plid]); 1488 blkcg->cpd[pol->plid] = NULL; 1489 } 1490 } 1491 } 1492 blkcg_policy[pol->plid] = NULL; 1493 err_unlock: 1494 mutex_unlock(&blkcg_pol_mutex); 1495 mutex_unlock(&blkcg_pol_register_mutex); 1496 return ret; 1497 } 1498 EXPORT_SYMBOL_GPL(blkcg_policy_register); 1499 1500 /** 1501 * blkcg_policy_unregister - unregister a blkcg policy 1502 * @pol: blkcg policy to unregister 1503 * 1504 * Undo blkcg_policy_register(@pol). Might sleep. 1505 */ 1506 void blkcg_policy_unregister(struct blkcg_policy *pol) 1507 { 1508 struct blkcg *blkcg; 1509 1510 mutex_lock(&blkcg_pol_register_mutex); 1511 1512 if (WARN_ON(blkcg_policy[pol->plid] != pol)) 1513 goto out_unlock; 1514 1515 /* kill the intf files first */ 1516 if (pol->dfl_cftypes) 1517 cgroup_rm_cftypes(pol->dfl_cftypes); 1518 if (pol->legacy_cftypes) 1519 cgroup_rm_cftypes(pol->legacy_cftypes); 1520 1521 /* remove cpds and unregister */ 1522 mutex_lock(&blkcg_pol_mutex); 1523 1524 if (pol->cpd_free_fn) { 1525 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) { 1526 if (blkcg->cpd[pol->plid]) { 1527 pol->cpd_free_fn(blkcg->cpd[pol->plid]); 1528 blkcg->cpd[pol->plid] = NULL; 1529 } 1530 } 1531 } 1532 blkcg_policy[pol->plid] = NULL; 1533 1534 mutex_unlock(&blkcg_pol_mutex); 1535 out_unlock: 1536 mutex_unlock(&blkcg_pol_register_mutex); 1537 } 1538 EXPORT_SYMBOL_GPL(blkcg_policy_unregister); 1539 1540 bool __blkcg_punt_bio_submit(struct bio *bio) 1541 { 1542 struct blkcg_gq *blkg = bio->bi_blkg; 1543 1544 /* consume the flag first */ 1545 bio->bi_opf &= ~REQ_CGROUP_PUNT; 1546 1547 /* never bounce for the root cgroup */ 1548 if (!blkg->parent) 1549 return false; 1550 1551 spin_lock_bh(&blkg->async_bio_lock); 1552 bio_list_add(&blkg->async_bios, bio); 1553 spin_unlock_bh(&blkg->async_bio_lock); 1554 1555 queue_work(blkcg_punt_bio_wq, &blkg->async_bio_work); 1556 return true; 1557 } 1558 1559 /* 1560 * Scale the accumulated delay based on how long it has been since we updated 1561 * the delay. We only call this when we are adding delay, in case it's been a 1562 * while since we added delay, and when we are checking to see if we need to 1563 * delay a task, to account for any delays that may have occurred. 1564 */ 1565 static void blkcg_scale_delay(struct blkcg_gq *blkg, u64 now) 1566 { 1567 u64 old = atomic64_read(&blkg->delay_start); 1568 1569 /* negative use_delay means no scaling, see blkcg_set_delay() */ 1570 if (atomic_read(&blkg->use_delay) < 0) 1571 return; 1572 1573 /* 1574 * We only want to scale down every second. The idea here is that we 1575 * want to delay people for min(delay_nsec, NSEC_PER_SEC) in a certain 1576 * time window. We only want to throttle tasks for recent delay that 1577 * has occurred, in 1 second time windows since that's the maximum 1578 * things can be throttled. We save the current delay window in 1579 * blkg->last_delay so we know what amount is still left to be charged 1580 * to the blkg from this point onward. blkg->last_use keeps track of 1581 * the use_delay counter. The idea is if we're unthrottling the blkg we 1582 * are ok with whatever is happening now, and we can take away more of 1583 * the accumulated delay as we've already throttled enough that 1584 * everybody is happy with their IO latencies. 1585 */ 1586 if (time_before64(old + NSEC_PER_SEC, now) && 1587 atomic64_cmpxchg(&blkg->delay_start, old, now) == old) { 1588 u64 cur = atomic64_read(&blkg->delay_nsec); 1589 u64 sub = min_t(u64, blkg->last_delay, now - old); 1590 int cur_use = atomic_read(&blkg->use_delay); 1591 1592 /* 1593 * We've been unthrottled, subtract a larger chunk of our 1594 * accumulated delay. 1595 */ 1596 if (cur_use < blkg->last_use) 1597 sub = max_t(u64, sub, blkg->last_delay >> 1); 1598 1599 /* 1600 * This shouldn't happen, but handle it anyway. Our delay_nsec 1601 * should only ever be growing except here where we subtract out 1602 * min(last_delay, 1 second), but lord knows bugs happen and I'd 1603 * rather not end up with negative numbers. 1604 */ 1605 if (unlikely(cur < sub)) { 1606 atomic64_set(&blkg->delay_nsec, 0); 1607 blkg->last_delay = 0; 1608 } else { 1609 atomic64_sub(sub, &blkg->delay_nsec); 1610 blkg->last_delay = cur - sub; 1611 } 1612 blkg->last_use = cur_use; 1613 } 1614 } 1615 1616 /* 1617 * This is called when we want to actually walk up the hierarchy and check to 1618 * see if we need to throttle, and then actually throttle if there is some 1619 * accumulated delay. This should only be called upon return to user space so 1620 * we're not holding some lock that would induce a priority inversion. 1621 */ 1622 static void blkcg_maybe_throttle_blkg(struct blkcg_gq *blkg, bool use_memdelay) 1623 { 1624 unsigned long pflags; 1625 bool clamp; 1626 u64 now = ktime_to_ns(ktime_get()); 1627 u64 exp; 1628 u64 delay_nsec = 0; 1629 int tok; 1630 1631 while (blkg->parent) { 1632 int use_delay = atomic_read(&blkg->use_delay); 1633 1634 if (use_delay) { 1635 u64 this_delay; 1636 1637 blkcg_scale_delay(blkg, now); 1638 this_delay = atomic64_read(&blkg->delay_nsec); 1639 if (this_delay > delay_nsec) { 1640 delay_nsec = this_delay; 1641 clamp = use_delay > 0; 1642 } 1643 } 1644 blkg = blkg->parent; 1645 } 1646 1647 if (!delay_nsec) 1648 return; 1649 1650 /* 1651 * Let's not sleep for all eternity if we've amassed a huge delay. 1652 * Swapping or metadata IO can accumulate 10's of seconds worth of 1653 * delay, and we want userspace to be able to do _something_ so cap the 1654 * delays at 0.25s. If there's 10's of seconds worth of delay then the 1655 * tasks will be delayed for 0.25 second for every syscall. If 1656 * blkcg_set_delay() was used as indicated by negative use_delay, the 1657 * caller is responsible for regulating the range. 1658 */ 1659 if (clamp) 1660 delay_nsec = min_t(u64, delay_nsec, 250 * NSEC_PER_MSEC); 1661 1662 if (use_memdelay) 1663 psi_memstall_enter(&pflags); 1664 1665 exp = ktime_add_ns(now, delay_nsec); 1666 tok = io_schedule_prepare(); 1667 do { 1668 __set_current_state(TASK_KILLABLE); 1669 if (!schedule_hrtimeout(&exp, HRTIMER_MODE_ABS)) 1670 break; 1671 } while (!fatal_signal_pending(current)); 1672 io_schedule_finish(tok); 1673 1674 if (use_memdelay) 1675 psi_memstall_leave(&pflags); 1676 } 1677 1678 /** 1679 * blkcg_maybe_throttle_current - throttle the current task if it has been marked 1680 * 1681 * This is only called if we've been marked with set_notify_resume(). Obviously 1682 * we can be set_notify_resume() for reasons other than blkcg throttling, so we 1683 * check to see if current->throttle_queue is set and if not this doesn't do 1684 * anything. This should only ever be called by the resume code, it's not meant 1685 * to be called by people willy-nilly as it will actually do the work to 1686 * throttle the task if it is setup for throttling. 1687 */ 1688 void blkcg_maybe_throttle_current(void) 1689 { 1690 struct request_queue *q = current->throttle_queue; 1691 struct cgroup_subsys_state *css; 1692 struct blkcg *blkcg; 1693 struct blkcg_gq *blkg; 1694 bool use_memdelay = current->use_memdelay; 1695 1696 if (!q) 1697 return; 1698 1699 current->throttle_queue = NULL; 1700 current->use_memdelay = false; 1701 1702 rcu_read_lock(); 1703 css = kthread_blkcg(); 1704 if (css) 1705 blkcg = css_to_blkcg(css); 1706 else 1707 blkcg = css_to_blkcg(task_css(current, io_cgrp_id)); 1708 1709 if (!blkcg) 1710 goto out; 1711 blkg = blkg_lookup(blkcg, q); 1712 if (!blkg) 1713 goto out; 1714 if (!blkg_tryget(blkg)) 1715 goto out; 1716 rcu_read_unlock(); 1717 1718 blkcg_maybe_throttle_blkg(blkg, use_memdelay); 1719 blkg_put(blkg); 1720 blk_put_queue(q); 1721 return; 1722 out: 1723 rcu_read_unlock(); 1724 blk_put_queue(q); 1725 } 1726 1727 /** 1728 * blkcg_schedule_throttle - this task needs to check for throttling 1729 * @q: the request queue IO was submitted on 1730 * @use_memdelay: do we charge this to memory delay for PSI 1731 * 1732 * This is called by the IO controller when we know there's delay accumulated 1733 * for the blkg for this task. We do not pass the blkg because there are places 1734 * we call this that may not have that information, the swapping code for 1735 * instance will only have a request_queue at that point. This set's the 1736 * notify_resume for the task to check and see if it requires throttling before 1737 * returning to user space. 1738 * 1739 * We will only schedule once per syscall. You can call this over and over 1740 * again and it will only do the check once upon return to user space, and only 1741 * throttle once. If the task needs to be throttled again it'll need to be 1742 * re-set at the next time we see the task. 1743 */ 1744 void blkcg_schedule_throttle(struct request_queue *q, bool use_memdelay) 1745 { 1746 if (unlikely(current->flags & PF_KTHREAD)) 1747 return; 1748 1749 if (!blk_get_queue(q)) 1750 return; 1751 1752 if (current->throttle_queue) 1753 blk_put_queue(current->throttle_queue); 1754 current->throttle_queue = q; 1755 if (use_memdelay) 1756 current->use_memdelay = use_memdelay; 1757 set_notify_resume(current); 1758 } 1759 1760 /** 1761 * blkcg_add_delay - add delay to this blkg 1762 * @blkg: blkg of interest 1763 * @now: the current time in nanoseconds 1764 * @delta: how many nanoseconds of delay to add 1765 * 1766 * Charge @delta to the blkg's current delay accumulation. This is used to 1767 * throttle tasks if an IO controller thinks we need more throttling. 1768 */ 1769 void blkcg_add_delay(struct blkcg_gq *blkg, u64 now, u64 delta) 1770 { 1771 if (WARN_ON_ONCE(atomic_read(&blkg->use_delay) < 0)) 1772 return; 1773 blkcg_scale_delay(blkg, now); 1774 atomic64_add(delta, &blkg->delay_nsec); 1775 } 1776 1777 /** 1778 * blkg_tryget_closest - try and get a blkg ref on the closet blkg 1779 * @bio: target bio 1780 * @css: target css 1781 * 1782 * As the failure mode here is to walk up the blkg tree, this ensure that the 1783 * blkg->parent pointers are always valid. This returns the blkg that it ended 1784 * up taking a reference on or %NULL if no reference was taken. 1785 */ 1786 static inline struct blkcg_gq *blkg_tryget_closest(struct bio *bio, 1787 struct cgroup_subsys_state *css) 1788 { 1789 struct blkcg_gq *blkg, *ret_blkg = NULL; 1790 1791 rcu_read_lock(); 1792 blkg = blkg_lookup_create(css_to_blkcg(css), bio->bi_disk->queue); 1793 while (blkg) { 1794 if (blkg_tryget(blkg)) { 1795 ret_blkg = blkg; 1796 break; 1797 } 1798 blkg = blkg->parent; 1799 } 1800 rcu_read_unlock(); 1801 1802 return ret_blkg; 1803 } 1804 1805 /** 1806 * bio_associate_blkg_from_css - associate a bio with a specified css 1807 * @bio: target bio 1808 * @css: target css 1809 * 1810 * Associate @bio with the blkg found by combining the css's blkg and the 1811 * request_queue of the @bio. An association failure is handled by walking up 1812 * the blkg tree. Therefore, the blkg associated can be anything between @blkg 1813 * and q->root_blkg. This situation only happens when a cgroup is dying and 1814 * then the remaining bios will spill to the closest alive blkg. 1815 * 1816 * A reference will be taken on the blkg and will be released when @bio is 1817 * freed. 1818 */ 1819 void bio_associate_blkg_from_css(struct bio *bio, 1820 struct cgroup_subsys_state *css) 1821 { 1822 if (bio->bi_blkg) 1823 blkg_put(bio->bi_blkg); 1824 1825 if (css && css->parent) { 1826 bio->bi_blkg = blkg_tryget_closest(bio, css); 1827 } else { 1828 blkg_get(bio->bi_disk->queue->root_blkg); 1829 bio->bi_blkg = bio->bi_disk->queue->root_blkg; 1830 } 1831 } 1832 EXPORT_SYMBOL_GPL(bio_associate_blkg_from_css); 1833 1834 /** 1835 * bio_associate_blkg - associate a bio with a blkg 1836 * @bio: target bio 1837 * 1838 * Associate @bio with the blkg found from the bio's css and request_queue. 1839 * If one is not found, bio_lookup_blkg() creates the blkg. If a blkg is 1840 * already associated, the css is reused and association redone as the 1841 * request_queue may have changed. 1842 */ 1843 void bio_associate_blkg(struct bio *bio) 1844 { 1845 struct cgroup_subsys_state *css; 1846 1847 rcu_read_lock(); 1848 1849 if (bio->bi_blkg) 1850 css = &bio_blkcg(bio)->css; 1851 else 1852 css = blkcg_css(); 1853 1854 bio_associate_blkg_from_css(bio, css); 1855 1856 rcu_read_unlock(); 1857 } 1858 EXPORT_SYMBOL_GPL(bio_associate_blkg); 1859 1860 /** 1861 * bio_clone_blkg_association - clone blkg association from src to dst bio 1862 * @dst: destination bio 1863 * @src: source bio 1864 */ 1865 void bio_clone_blkg_association(struct bio *dst, struct bio *src) 1866 { 1867 if (src->bi_blkg) { 1868 if (dst->bi_blkg) 1869 blkg_put(dst->bi_blkg); 1870 blkg_get(src->bi_blkg); 1871 dst->bi_blkg = src->bi_blkg; 1872 } 1873 } 1874 EXPORT_SYMBOL_GPL(bio_clone_blkg_association); 1875 1876 static int blk_cgroup_io_type(struct bio *bio) 1877 { 1878 if (op_is_discard(bio->bi_opf)) 1879 return BLKG_IOSTAT_DISCARD; 1880 if (op_is_write(bio->bi_opf)) 1881 return BLKG_IOSTAT_WRITE; 1882 return BLKG_IOSTAT_READ; 1883 } 1884 1885 void blk_cgroup_bio_start(struct bio *bio) 1886 { 1887 int rwd = blk_cgroup_io_type(bio), cpu; 1888 struct blkg_iostat_set *bis; 1889 1890 cpu = get_cpu(); 1891 bis = per_cpu_ptr(bio->bi_blkg->iostat_cpu, cpu); 1892 u64_stats_update_begin(&bis->sync); 1893 1894 /* 1895 * If the bio is flagged with BIO_CGROUP_ACCT it means this is a split 1896 * bio and we would have already accounted for the size of the bio. 1897 */ 1898 if (!bio_flagged(bio, BIO_CGROUP_ACCT)) { 1899 bio_set_flag(bio, BIO_CGROUP_ACCT); 1900 bis->cur.bytes[rwd] += bio->bi_iter.bi_size; 1901 } 1902 bis->cur.ios[rwd]++; 1903 1904 u64_stats_update_end(&bis->sync); 1905 if (cgroup_subsys_on_dfl(io_cgrp_subsys)) 1906 cgroup_rstat_updated(bio->bi_blkg->blkcg->css.cgroup, cpu); 1907 put_cpu(); 1908 } 1909 1910 static int __init blkcg_init(void) 1911 { 1912 blkcg_punt_bio_wq = alloc_workqueue("blkcg_punt_bio", 1913 WQ_MEM_RECLAIM | WQ_FREEZABLE | 1914 WQ_UNBOUND | WQ_SYSFS, 0); 1915 if (!blkcg_punt_bio_wq) 1916 return -ENOMEM; 1917 return 0; 1918 } 1919 subsys_initcall(blkcg_init); 1920 1921 module_param(blkcg_debug_stats, bool, 0644); 1922 MODULE_PARM_DESC(blkcg_debug_stats, "True if you want debug stats, false if not"); 1923