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 #include "blk-ioprio.h" 35 36 /* 37 * blkcg_pol_mutex protects blkcg_policy[] and policy [de]activation. 38 * blkcg_pol_register_mutex nests outside of it and synchronizes entire 39 * policy [un]register operations including cgroup file additions / 40 * removals. Putting cgroup file registration outside blkcg_pol_mutex 41 * allows grabbing it from cgroup callbacks. 42 */ 43 static DEFINE_MUTEX(blkcg_pol_register_mutex); 44 static DEFINE_MUTEX(blkcg_pol_mutex); 45 46 struct blkcg blkcg_root; 47 EXPORT_SYMBOL_GPL(blkcg_root); 48 49 struct cgroup_subsys_state * const blkcg_root_css = &blkcg_root.css; 50 EXPORT_SYMBOL_GPL(blkcg_root_css); 51 52 static struct blkcg_policy *blkcg_policy[BLKCG_MAX_POLS]; 53 54 static LIST_HEAD(all_blkcgs); /* protected by blkcg_pol_mutex */ 55 56 bool blkcg_debug_stats = false; 57 static struct workqueue_struct *blkcg_punt_bio_wq; 58 59 #define BLKG_DESTROY_BATCH_SIZE 64 60 61 static bool blkcg_policy_enabled(struct request_queue *q, 62 const struct blkcg_policy *pol) 63 { 64 return pol && test_bit(pol->plid, q->blkcg_pols); 65 } 66 67 /** 68 * blkg_free - free a blkg 69 * @blkg: blkg to free 70 * 71 * Free @blkg which may be partially allocated. 72 */ 73 static void blkg_free(struct blkcg_gq *blkg) 74 { 75 int i; 76 77 if (!blkg) 78 return; 79 80 for (i = 0; i < BLKCG_MAX_POLS; i++) 81 if (blkg->pd[i]) 82 blkcg_policy[i]->pd_free_fn(blkg->pd[i]); 83 84 free_percpu(blkg->iostat_cpu); 85 percpu_ref_exit(&blkg->refcnt); 86 kfree(blkg); 87 } 88 89 static void __blkg_release(struct rcu_head *rcu) 90 { 91 struct blkcg_gq *blkg = container_of(rcu, struct blkcg_gq, rcu_head); 92 93 WARN_ON(!bio_list_empty(&blkg->async_bios)); 94 95 /* release the blkcg and parent blkg refs this blkg has been holding */ 96 css_put(&blkg->blkcg->css); 97 if (blkg->parent) 98 blkg_put(blkg->parent); 99 blkg_free(blkg); 100 } 101 102 /* 103 * A group is RCU protected, but having an rcu lock does not mean that one 104 * can access all the fields of blkg and assume these are valid. For 105 * example, don't try to follow throtl_data and request queue links. 106 * 107 * Having a reference to blkg under an rcu allows accesses to only values 108 * local to groups like group stats and group rate limits. 109 */ 110 static void blkg_release(struct percpu_ref *ref) 111 { 112 struct blkcg_gq *blkg = container_of(ref, struct blkcg_gq, refcnt); 113 114 call_rcu(&blkg->rcu_head, __blkg_release); 115 } 116 117 static void blkg_async_bio_workfn(struct work_struct *work) 118 { 119 struct blkcg_gq *blkg = container_of(work, struct blkcg_gq, 120 async_bio_work); 121 struct bio_list bios = BIO_EMPTY_LIST; 122 struct bio *bio; 123 struct blk_plug plug; 124 bool need_plug = false; 125 126 /* as long as there are pending bios, @blkg can't go away */ 127 spin_lock_bh(&blkg->async_bio_lock); 128 bio_list_merge(&bios, &blkg->async_bios); 129 bio_list_init(&blkg->async_bios); 130 spin_unlock_bh(&blkg->async_bio_lock); 131 132 /* start plug only when bio_list contains at least 2 bios */ 133 if (bios.head && bios.head->bi_next) { 134 need_plug = true; 135 blk_start_plug(&plug); 136 } 137 while ((bio = bio_list_pop(&bios))) 138 submit_bio(bio); 139 if (need_plug) 140 blk_finish_plug(&plug); 141 } 142 143 /** 144 * blkg_alloc - allocate a blkg 145 * @blkcg: block cgroup the new blkg is associated with 146 * @q: request_queue the new blkg is associated with 147 * @gfp_mask: allocation mask to use 148 * 149 * Allocate a new blkg assocating @blkcg and @q. 150 */ 151 static struct blkcg_gq *blkg_alloc(struct blkcg *blkcg, struct request_queue *q, 152 gfp_t gfp_mask) 153 { 154 struct blkcg_gq *blkg; 155 int i, cpu; 156 157 /* alloc and init base part */ 158 blkg = kzalloc_node(sizeof(*blkg), gfp_mask, q->node); 159 if (!blkg) 160 return NULL; 161 162 if (percpu_ref_init(&blkg->refcnt, blkg_release, 0, gfp_mask)) 163 goto err_free; 164 165 blkg->iostat_cpu = alloc_percpu_gfp(struct blkg_iostat_set, gfp_mask); 166 if (!blkg->iostat_cpu) 167 goto err_free; 168 169 blkg->q = q; 170 INIT_LIST_HEAD(&blkg->q_node); 171 spin_lock_init(&blkg->async_bio_lock); 172 bio_list_init(&blkg->async_bios); 173 INIT_WORK(&blkg->async_bio_work, blkg_async_bio_workfn); 174 blkg->blkcg = blkcg; 175 176 u64_stats_init(&blkg->iostat.sync); 177 for_each_possible_cpu(cpu) 178 u64_stats_init(&per_cpu_ptr(blkg->iostat_cpu, cpu)->sync); 179 180 for (i = 0; i < BLKCG_MAX_POLS; i++) { 181 struct blkcg_policy *pol = blkcg_policy[i]; 182 struct blkg_policy_data *pd; 183 184 if (!blkcg_policy_enabled(q, pol)) 185 continue; 186 187 /* alloc per-policy data and attach it to blkg */ 188 pd = pol->pd_alloc_fn(gfp_mask, q, blkcg); 189 if (!pd) 190 goto err_free; 191 192 blkg->pd[i] = pd; 193 pd->blkg = blkg; 194 pd->plid = i; 195 } 196 197 return blkg; 198 199 err_free: 200 blkg_free(blkg); 201 return NULL; 202 } 203 204 struct blkcg_gq *blkg_lookup_slowpath(struct blkcg *blkcg, 205 struct request_queue *q, bool update_hint) 206 { 207 struct blkcg_gq *blkg; 208 209 /* 210 * Hint didn't match. Look up from the radix tree. Note that the 211 * hint can only be updated under queue_lock as otherwise @blkg 212 * could have already been removed from blkg_tree. The caller is 213 * responsible for grabbing queue_lock if @update_hint. 214 */ 215 blkg = radix_tree_lookup(&blkcg->blkg_tree, q->id); 216 if (blkg && blkg->q == q) { 217 if (update_hint) { 218 lockdep_assert_held(&q->queue_lock); 219 rcu_assign_pointer(blkcg->blkg_hint, blkg); 220 } 221 return blkg; 222 } 223 224 return NULL; 225 } 226 EXPORT_SYMBOL_GPL(blkg_lookup_slowpath); 227 228 /* 229 * If @new_blkg is %NULL, this function tries to allocate a new one as 230 * necessary using %GFP_NOWAIT. @new_blkg is always consumed on return. 231 */ 232 static struct blkcg_gq *blkg_create(struct blkcg *blkcg, 233 struct request_queue *q, 234 struct blkcg_gq *new_blkg) 235 { 236 struct blkcg_gq *blkg; 237 int i, ret; 238 239 WARN_ON_ONCE(!rcu_read_lock_held()); 240 lockdep_assert_held(&q->queue_lock); 241 242 /* request_queue is dying, do not create/recreate a blkg */ 243 if (blk_queue_dying(q)) { 244 ret = -ENODEV; 245 goto err_free_blkg; 246 } 247 248 /* blkg holds a reference to blkcg */ 249 if (!css_tryget_online(&blkcg->css)) { 250 ret = -ENODEV; 251 goto err_free_blkg; 252 } 253 254 /* allocate */ 255 if (!new_blkg) { 256 new_blkg = blkg_alloc(blkcg, q, GFP_NOWAIT | __GFP_NOWARN); 257 if (unlikely(!new_blkg)) { 258 ret = -ENOMEM; 259 goto err_put_css; 260 } 261 } 262 blkg = new_blkg; 263 264 /* link parent */ 265 if (blkcg_parent(blkcg)) { 266 blkg->parent = __blkg_lookup(blkcg_parent(blkcg), q, false); 267 if (WARN_ON_ONCE(!blkg->parent)) { 268 ret = -ENODEV; 269 goto err_put_css; 270 } 271 blkg_get(blkg->parent); 272 } 273 274 /* invoke per-policy init */ 275 for (i = 0; i < BLKCG_MAX_POLS; i++) { 276 struct blkcg_policy *pol = blkcg_policy[i]; 277 278 if (blkg->pd[i] && pol->pd_init_fn) 279 pol->pd_init_fn(blkg->pd[i]); 280 } 281 282 /* insert */ 283 spin_lock(&blkcg->lock); 284 ret = radix_tree_insert(&blkcg->blkg_tree, q->id, blkg); 285 if (likely(!ret)) { 286 hlist_add_head_rcu(&blkg->blkcg_node, &blkcg->blkg_list); 287 list_add(&blkg->q_node, &q->blkg_list); 288 289 for (i = 0; i < BLKCG_MAX_POLS; i++) { 290 struct blkcg_policy *pol = blkcg_policy[i]; 291 292 if (blkg->pd[i] && pol->pd_online_fn) 293 pol->pd_online_fn(blkg->pd[i]); 294 } 295 } 296 blkg->online = true; 297 spin_unlock(&blkcg->lock); 298 299 if (!ret) 300 return blkg; 301 302 /* @blkg failed fully initialized, use the usual release path */ 303 blkg_put(blkg); 304 return ERR_PTR(ret); 305 306 err_put_css: 307 css_put(&blkcg->css); 308 err_free_blkg: 309 blkg_free(new_blkg); 310 return ERR_PTR(ret); 311 } 312 313 /** 314 * blkg_lookup_create - lookup blkg, try to create one if not there 315 * @blkcg: blkcg of interest 316 * @q: request_queue of interest 317 * 318 * Lookup blkg for the @blkcg - @q pair. If it doesn't exist, try to 319 * create one. blkg creation is performed recursively from blkcg_root such 320 * that all non-root blkg's have access to the parent blkg. This function 321 * should be called under RCU read lock and takes @q->queue_lock. 322 * 323 * Returns the blkg or the closest blkg if blkg_create() fails as it walks 324 * down from root. 325 */ 326 static struct blkcg_gq *blkg_lookup_create(struct blkcg *blkcg, 327 struct request_queue *q) 328 { 329 struct blkcg_gq *blkg; 330 unsigned long flags; 331 332 WARN_ON_ONCE(!rcu_read_lock_held()); 333 334 blkg = blkg_lookup(blkcg, q); 335 if (blkg) 336 return blkg; 337 338 spin_lock_irqsave(&q->queue_lock, flags); 339 blkg = __blkg_lookup(blkcg, q, true); 340 if (blkg) 341 goto found; 342 343 /* 344 * Create blkgs walking down from blkcg_root to @blkcg, so that all 345 * non-root blkgs have access to their parents. Returns the closest 346 * blkg to the intended blkg should blkg_create() fail. 347 */ 348 while (true) { 349 struct blkcg *pos = blkcg; 350 struct blkcg *parent = blkcg_parent(blkcg); 351 struct blkcg_gq *ret_blkg = q->root_blkg; 352 353 while (parent) { 354 blkg = __blkg_lookup(parent, q, false); 355 if (blkg) { 356 /* remember closest blkg */ 357 ret_blkg = blkg; 358 break; 359 } 360 pos = parent; 361 parent = blkcg_parent(parent); 362 } 363 364 blkg = blkg_create(pos, q, NULL); 365 if (IS_ERR(blkg)) { 366 blkg = ret_blkg; 367 break; 368 } 369 if (pos == blkcg) 370 break; 371 } 372 373 found: 374 spin_unlock_irqrestore(&q->queue_lock, flags); 375 return blkg; 376 } 377 378 static void blkg_destroy(struct blkcg_gq *blkg) 379 { 380 struct blkcg *blkcg = blkg->blkcg; 381 int i; 382 383 lockdep_assert_held(&blkg->q->queue_lock); 384 lockdep_assert_held(&blkcg->lock); 385 386 /* Something wrong if we are trying to remove same group twice */ 387 WARN_ON_ONCE(list_empty(&blkg->q_node)); 388 WARN_ON_ONCE(hlist_unhashed(&blkg->blkcg_node)); 389 390 for (i = 0; i < BLKCG_MAX_POLS; i++) { 391 struct blkcg_policy *pol = blkcg_policy[i]; 392 393 if (blkg->pd[i] && pol->pd_offline_fn) 394 pol->pd_offline_fn(blkg->pd[i]); 395 } 396 397 blkg->online = false; 398 399 radix_tree_delete(&blkcg->blkg_tree, blkg->q->id); 400 list_del_init(&blkg->q_node); 401 hlist_del_init_rcu(&blkg->blkcg_node); 402 403 /* 404 * Both setting lookup hint to and clearing it from @blkg are done 405 * under queue_lock. If it's not pointing to @blkg now, it never 406 * will. Hint assignment itself can race safely. 407 */ 408 if (rcu_access_pointer(blkcg->blkg_hint) == blkg) 409 rcu_assign_pointer(blkcg->blkg_hint, NULL); 410 411 /* 412 * Put the reference taken at the time of creation so that when all 413 * queues are gone, group can be destroyed. 414 */ 415 percpu_ref_kill(&blkg->refcnt); 416 } 417 418 /** 419 * blkg_destroy_all - destroy all blkgs associated with a request_queue 420 * @q: request_queue of interest 421 * 422 * Destroy all blkgs associated with @q. 423 */ 424 static void blkg_destroy_all(struct request_queue *q) 425 { 426 struct blkcg_gq *blkg, *n; 427 int count = BLKG_DESTROY_BATCH_SIZE; 428 429 restart: 430 spin_lock_irq(&q->queue_lock); 431 list_for_each_entry_safe(blkg, n, &q->blkg_list, q_node) { 432 struct blkcg *blkcg = blkg->blkcg; 433 434 spin_lock(&blkcg->lock); 435 blkg_destroy(blkg); 436 spin_unlock(&blkcg->lock); 437 438 /* 439 * in order to avoid holding the spin lock for too long, release 440 * it when a batch of blkgs are destroyed. 441 */ 442 if (!(--count)) { 443 count = BLKG_DESTROY_BATCH_SIZE; 444 spin_unlock_irq(&q->queue_lock); 445 cond_resched(); 446 goto restart; 447 } 448 } 449 450 q->root_blkg = NULL; 451 spin_unlock_irq(&q->queue_lock); 452 } 453 454 static int blkcg_reset_stats(struct cgroup_subsys_state *css, 455 struct cftype *cftype, u64 val) 456 { 457 struct blkcg *blkcg = css_to_blkcg(css); 458 struct blkcg_gq *blkg; 459 int i, cpu; 460 461 mutex_lock(&blkcg_pol_mutex); 462 spin_lock_irq(&blkcg->lock); 463 464 /* 465 * Note that stat reset is racy - it doesn't synchronize against 466 * stat updates. This is a debug feature which shouldn't exist 467 * anyway. If you get hit by a race, retry. 468 */ 469 hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) { 470 for_each_possible_cpu(cpu) { 471 struct blkg_iostat_set *bis = 472 per_cpu_ptr(blkg->iostat_cpu, cpu); 473 memset(bis, 0, sizeof(*bis)); 474 } 475 memset(&blkg->iostat, 0, sizeof(blkg->iostat)); 476 477 for (i = 0; i < BLKCG_MAX_POLS; i++) { 478 struct blkcg_policy *pol = blkcg_policy[i]; 479 480 if (blkg->pd[i] && pol->pd_reset_stats_fn) 481 pol->pd_reset_stats_fn(blkg->pd[i]); 482 } 483 } 484 485 spin_unlock_irq(&blkcg->lock); 486 mutex_unlock(&blkcg_pol_mutex); 487 return 0; 488 } 489 490 const char *blkg_dev_name(struct blkcg_gq *blkg) 491 { 492 /* some drivers (floppy) instantiate a queue w/o disk registered */ 493 if (blkg->q->backing_dev_info->dev) 494 return bdi_dev_name(blkg->q->backing_dev_info); 495 return NULL; 496 } 497 498 /** 499 * blkcg_print_blkgs - helper for printing per-blkg data 500 * @sf: seq_file to print to 501 * @blkcg: blkcg of interest 502 * @prfill: fill function to print out a blkg 503 * @pol: policy in question 504 * @data: data to be passed to @prfill 505 * @show_total: to print out sum of prfill return values or not 506 * 507 * This function invokes @prfill on each blkg of @blkcg if pd for the 508 * policy specified by @pol exists. @prfill is invoked with @sf, the 509 * policy data and @data and the matching queue lock held. If @show_total 510 * is %true, the sum of the return values from @prfill is printed with 511 * "Total" label at the end. 512 * 513 * This is to be used to construct print functions for 514 * cftype->read_seq_string method. 515 */ 516 void blkcg_print_blkgs(struct seq_file *sf, struct blkcg *blkcg, 517 u64 (*prfill)(struct seq_file *, 518 struct blkg_policy_data *, int), 519 const struct blkcg_policy *pol, int data, 520 bool show_total) 521 { 522 struct blkcg_gq *blkg; 523 u64 total = 0; 524 525 rcu_read_lock(); 526 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) { 527 spin_lock_irq(&blkg->q->queue_lock); 528 if (blkcg_policy_enabled(blkg->q, pol)) 529 total += prfill(sf, blkg->pd[pol->plid], data); 530 spin_unlock_irq(&blkg->q->queue_lock); 531 } 532 rcu_read_unlock(); 533 534 if (show_total) 535 seq_printf(sf, "Total %llu\n", (unsigned long long)total); 536 } 537 EXPORT_SYMBOL_GPL(blkcg_print_blkgs); 538 539 /** 540 * __blkg_prfill_u64 - prfill helper for a single u64 value 541 * @sf: seq_file to print to 542 * @pd: policy private data of interest 543 * @v: value to print 544 * 545 * Print @v to @sf for the device assocaited with @pd. 546 */ 547 u64 __blkg_prfill_u64(struct seq_file *sf, struct blkg_policy_data *pd, u64 v) 548 { 549 const char *dname = blkg_dev_name(pd->blkg); 550 551 if (!dname) 552 return 0; 553 554 seq_printf(sf, "%s %llu\n", dname, (unsigned long long)v); 555 return v; 556 } 557 EXPORT_SYMBOL_GPL(__blkg_prfill_u64); 558 559 /* Performs queue bypass and policy enabled checks then looks up blkg. */ 560 static struct blkcg_gq *blkg_lookup_check(struct blkcg *blkcg, 561 const struct blkcg_policy *pol, 562 struct request_queue *q) 563 { 564 WARN_ON_ONCE(!rcu_read_lock_held()); 565 lockdep_assert_held(&q->queue_lock); 566 567 if (!blkcg_policy_enabled(q, pol)) 568 return ERR_PTR(-EOPNOTSUPP); 569 return __blkg_lookup(blkcg, q, true /* update_hint */); 570 } 571 572 /** 573 * blkcg_conf_open_bdev - parse and open bdev for per-blkg config update 574 * @inputp: input string pointer 575 * 576 * Parse the device node prefix part, MAJ:MIN, of per-blkg config update 577 * from @input and get and return the matching bdev. *@inputp is 578 * updated to point past the device node prefix. Returns an ERR_PTR() 579 * value on error. 580 * 581 * Use this function iff blkg_conf_prep() can't be used for some reason. 582 */ 583 struct block_device *blkcg_conf_open_bdev(char **inputp) 584 { 585 char *input = *inputp; 586 unsigned int major, minor; 587 struct block_device *bdev; 588 int key_len; 589 590 if (sscanf(input, "%u:%u%n", &major, &minor, &key_len) != 2) 591 return ERR_PTR(-EINVAL); 592 593 input += key_len; 594 if (!isspace(*input)) 595 return ERR_PTR(-EINVAL); 596 input = skip_spaces(input); 597 598 bdev = blkdev_get_no_open(MKDEV(major, minor)); 599 if (!bdev) 600 return ERR_PTR(-ENODEV); 601 if (bdev_is_partition(bdev)) { 602 blkdev_put_no_open(bdev); 603 return ERR_PTR(-ENODEV); 604 } 605 606 *inputp = input; 607 return bdev; 608 } 609 610 /** 611 * blkg_conf_prep - parse and prepare for per-blkg config update 612 * @blkcg: target block cgroup 613 * @pol: target policy 614 * @input: input string 615 * @ctx: blkg_conf_ctx to be filled 616 * 617 * Parse per-blkg config update from @input and initialize @ctx with the 618 * result. @ctx->blkg points to the blkg to be updated and @ctx->body the 619 * part of @input following MAJ:MIN. This function returns with RCU read 620 * lock and queue lock held and must be paired with blkg_conf_finish(). 621 */ 622 int blkg_conf_prep(struct blkcg *blkcg, const struct blkcg_policy *pol, 623 char *input, struct blkg_conf_ctx *ctx) 624 __acquires(rcu) __acquires(&bdev->bd_disk->queue->queue_lock) 625 { 626 struct block_device *bdev; 627 struct request_queue *q; 628 struct blkcg_gq *blkg; 629 int ret; 630 631 bdev = blkcg_conf_open_bdev(&input); 632 if (IS_ERR(bdev)) 633 return PTR_ERR(bdev); 634 635 q = bdev->bd_disk->queue; 636 637 rcu_read_lock(); 638 spin_lock_irq(&q->queue_lock); 639 640 blkg = blkg_lookup_check(blkcg, pol, q); 641 if (IS_ERR(blkg)) { 642 ret = PTR_ERR(blkg); 643 goto fail_unlock; 644 } 645 646 if (blkg) 647 goto success; 648 649 /* 650 * Create blkgs walking down from blkcg_root to @blkcg, so that all 651 * non-root blkgs have access to their parents. 652 */ 653 while (true) { 654 struct blkcg *pos = blkcg; 655 struct blkcg *parent; 656 struct blkcg_gq *new_blkg; 657 658 parent = blkcg_parent(blkcg); 659 while (parent && !__blkg_lookup(parent, q, false)) { 660 pos = parent; 661 parent = blkcg_parent(parent); 662 } 663 664 /* Drop locks to do new blkg allocation with GFP_KERNEL. */ 665 spin_unlock_irq(&q->queue_lock); 666 rcu_read_unlock(); 667 668 new_blkg = blkg_alloc(pos, q, GFP_KERNEL); 669 if (unlikely(!new_blkg)) { 670 ret = -ENOMEM; 671 goto fail; 672 } 673 674 if (radix_tree_preload(GFP_KERNEL)) { 675 blkg_free(new_blkg); 676 ret = -ENOMEM; 677 goto fail; 678 } 679 680 rcu_read_lock(); 681 spin_lock_irq(&q->queue_lock); 682 683 blkg = blkg_lookup_check(pos, pol, q); 684 if (IS_ERR(blkg)) { 685 ret = PTR_ERR(blkg); 686 blkg_free(new_blkg); 687 goto fail_preloaded; 688 } 689 690 if (blkg) { 691 blkg_free(new_blkg); 692 } else { 693 blkg = blkg_create(pos, q, new_blkg); 694 if (IS_ERR(blkg)) { 695 ret = PTR_ERR(blkg); 696 goto fail_preloaded; 697 } 698 } 699 700 radix_tree_preload_end(); 701 702 if (pos == blkcg) 703 goto success; 704 } 705 success: 706 ctx->bdev = bdev; 707 ctx->blkg = blkg; 708 ctx->body = input; 709 return 0; 710 711 fail_preloaded: 712 radix_tree_preload_end(); 713 fail_unlock: 714 spin_unlock_irq(&q->queue_lock); 715 rcu_read_unlock(); 716 fail: 717 blkdev_put_no_open(bdev); 718 /* 719 * If queue was bypassing, we should retry. Do so after a 720 * short msleep(). It isn't strictly necessary but queue 721 * can be bypassing for some time and it's always nice to 722 * avoid busy looping. 723 */ 724 if (ret == -EBUSY) { 725 msleep(10); 726 ret = restart_syscall(); 727 } 728 return ret; 729 } 730 EXPORT_SYMBOL_GPL(blkg_conf_prep); 731 732 /** 733 * blkg_conf_finish - finish up per-blkg config update 734 * @ctx: blkg_conf_ctx intiailized by blkg_conf_prep() 735 * 736 * Finish up after per-blkg config update. This function must be paired 737 * with blkg_conf_prep(). 738 */ 739 void blkg_conf_finish(struct blkg_conf_ctx *ctx) 740 __releases(&ctx->bdev->bd_disk->queue->queue_lock) __releases(rcu) 741 { 742 spin_unlock_irq(&ctx->bdev->bd_disk->queue->queue_lock); 743 rcu_read_unlock(); 744 blkdev_put_no_open(ctx->bdev); 745 } 746 EXPORT_SYMBOL_GPL(blkg_conf_finish); 747 748 static void blkg_iostat_set(struct blkg_iostat *dst, struct blkg_iostat *src) 749 { 750 int i; 751 752 for (i = 0; i < BLKG_IOSTAT_NR; i++) { 753 dst->bytes[i] = src->bytes[i]; 754 dst->ios[i] = src->ios[i]; 755 } 756 } 757 758 static void blkg_iostat_add(struct blkg_iostat *dst, struct blkg_iostat *src) 759 { 760 int i; 761 762 for (i = 0; i < BLKG_IOSTAT_NR; i++) { 763 dst->bytes[i] += src->bytes[i]; 764 dst->ios[i] += src->ios[i]; 765 } 766 } 767 768 static void blkg_iostat_sub(struct blkg_iostat *dst, struct blkg_iostat *src) 769 { 770 int i; 771 772 for (i = 0; i < BLKG_IOSTAT_NR; i++) { 773 dst->bytes[i] -= src->bytes[i]; 774 dst->ios[i] -= src->ios[i]; 775 } 776 } 777 778 static void blkcg_rstat_flush(struct cgroup_subsys_state *css, int cpu) 779 { 780 struct blkcg *blkcg = css_to_blkcg(css); 781 struct blkcg_gq *blkg; 782 783 /* Root-level stats are sourced from system-wide IO stats */ 784 if (!cgroup_parent(css->cgroup)) 785 return; 786 787 rcu_read_lock(); 788 789 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) { 790 struct blkcg_gq *parent = blkg->parent; 791 struct blkg_iostat_set *bisc = per_cpu_ptr(blkg->iostat_cpu, cpu); 792 struct blkg_iostat cur, delta; 793 unsigned long flags; 794 unsigned int seq; 795 796 /* fetch the current per-cpu values */ 797 do { 798 seq = u64_stats_fetch_begin(&bisc->sync); 799 blkg_iostat_set(&cur, &bisc->cur); 800 } while (u64_stats_fetch_retry(&bisc->sync, seq)); 801 802 /* propagate percpu delta to global */ 803 flags = u64_stats_update_begin_irqsave(&blkg->iostat.sync); 804 blkg_iostat_set(&delta, &cur); 805 blkg_iostat_sub(&delta, &bisc->last); 806 blkg_iostat_add(&blkg->iostat.cur, &delta); 807 blkg_iostat_add(&bisc->last, &delta); 808 u64_stats_update_end_irqrestore(&blkg->iostat.sync, flags); 809 810 /* propagate global delta to parent (unless that's root) */ 811 if (parent && parent->parent) { 812 flags = u64_stats_update_begin_irqsave(&parent->iostat.sync); 813 blkg_iostat_set(&delta, &blkg->iostat.cur); 814 blkg_iostat_sub(&delta, &blkg->iostat.last); 815 blkg_iostat_add(&parent->iostat.cur, &delta); 816 blkg_iostat_add(&blkg->iostat.last, &delta); 817 u64_stats_update_end_irqrestore(&parent->iostat.sync, flags); 818 } 819 } 820 821 rcu_read_unlock(); 822 } 823 824 /* 825 * We source root cgroup stats from the system-wide stats to avoid 826 * tracking the same information twice and incurring overhead when no 827 * cgroups are defined. For that reason, cgroup_rstat_flush in 828 * blkcg_print_stat does not actually fill out the iostat in the root 829 * cgroup's blkcg_gq. 830 * 831 * However, we would like to re-use the printing code between the root and 832 * non-root cgroups to the extent possible. For that reason, we simulate 833 * flushing the root cgroup's stats by explicitly filling in the iostat 834 * with disk level statistics. 835 */ 836 static void blkcg_fill_root_iostats(void) 837 { 838 struct class_dev_iter iter; 839 struct device *dev; 840 841 class_dev_iter_init(&iter, &block_class, NULL, &disk_type); 842 while ((dev = class_dev_iter_next(&iter))) { 843 struct block_device *bdev = dev_to_bdev(dev); 844 struct blkcg_gq *blkg = 845 blk_queue_root_blkg(bdev->bd_disk->queue); 846 struct blkg_iostat tmp; 847 int cpu; 848 849 memset(&tmp, 0, sizeof(tmp)); 850 for_each_possible_cpu(cpu) { 851 struct disk_stats *cpu_dkstats; 852 unsigned long flags; 853 854 cpu_dkstats = per_cpu_ptr(bdev->bd_stats, cpu); 855 tmp.ios[BLKG_IOSTAT_READ] += 856 cpu_dkstats->ios[STAT_READ]; 857 tmp.ios[BLKG_IOSTAT_WRITE] += 858 cpu_dkstats->ios[STAT_WRITE]; 859 tmp.ios[BLKG_IOSTAT_DISCARD] += 860 cpu_dkstats->ios[STAT_DISCARD]; 861 // convert sectors to bytes 862 tmp.bytes[BLKG_IOSTAT_READ] += 863 cpu_dkstats->sectors[STAT_READ] << 9; 864 tmp.bytes[BLKG_IOSTAT_WRITE] += 865 cpu_dkstats->sectors[STAT_WRITE] << 9; 866 tmp.bytes[BLKG_IOSTAT_DISCARD] += 867 cpu_dkstats->sectors[STAT_DISCARD] << 9; 868 869 flags = u64_stats_update_begin_irqsave(&blkg->iostat.sync); 870 blkg_iostat_set(&blkg->iostat.cur, &tmp); 871 u64_stats_update_end_irqrestore(&blkg->iostat.sync, flags); 872 } 873 } 874 } 875 876 static int blkcg_print_stat(struct seq_file *sf, void *v) 877 { 878 struct blkcg *blkcg = css_to_blkcg(seq_css(sf)); 879 struct blkcg_gq *blkg; 880 881 if (!seq_css(sf)->parent) 882 blkcg_fill_root_iostats(); 883 else 884 cgroup_rstat_flush(blkcg->css.cgroup); 885 886 rcu_read_lock(); 887 888 hlist_for_each_entry_rcu(blkg, &blkcg->blkg_list, blkcg_node) { 889 struct blkg_iostat_set *bis = &blkg->iostat; 890 const char *dname; 891 char *buf; 892 u64 rbytes, wbytes, rios, wios, dbytes, dios; 893 size_t size = seq_get_buf(sf, &buf), off = 0; 894 int i; 895 bool has_stats = false; 896 unsigned seq; 897 898 spin_lock_irq(&blkg->q->queue_lock); 899 900 if (!blkg->online) 901 goto skip; 902 903 dname = blkg_dev_name(blkg); 904 if (!dname) 905 goto skip; 906 907 /* 908 * Hooray string manipulation, count is the size written NOT 909 * INCLUDING THE \0, so size is now count+1 less than what we 910 * had before, but we want to start writing the next bit from 911 * the \0 so we only add count to buf. 912 */ 913 off += scnprintf(buf+off, size-off, "%s ", dname); 914 915 do { 916 seq = u64_stats_fetch_begin(&bis->sync); 917 918 rbytes = bis->cur.bytes[BLKG_IOSTAT_READ]; 919 wbytes = bis->cur.bytes[BLKG_IOSTAT_WRITE]; 920 dbytes = bis->cur.bytes[BLKG_IOSTAT_DISCARD]; 921 rios = bis->cur.ios[BLKG_IOSTAT_READ]; 922 wios = bis->cur.ios[BLKG_IOSTAT_WRITE]; 923 dios = bis->cur.ios[BLKG_IOSTAT_DISCARD]; 924 } while (u64_stats_fetch_retry(&bis->sync, seq)); 925 926 if (rbytes || wbytes || rios || wios) { 927 has_stats = true; 928 off += scnprintf(buf+off, size-off, 929 "rbytes=%llu wbytes=%llu rios=%llu wios=%llu dbytes=%llu dios=%llu", 930 rbytes, wbytes, rios, wios, 931 dbytes, dios); 932 } 933 934 if (blkcg_debug_stats && atomic_read(&blkg->use_delay)) { 935 has_stats = true; 936 off += scnprintf(buf+off, size-off, 937 " use_delay=%d delay_nsec=%llu", 938 atomic_read(&blkg->use_delay), 939 (unsigned long long)atomic64_read(&blkg->delay_nsec)); 940 } 941 942 for (i = 0; i < BLKCG_MAX_POLS; i++) { 943 struct blkcg_policy *pol = blkcg_policy[i]; 944 size_t written; 945 946 if (!blkg->pd[i] || !pol->pd_stat_fn) 947 continue; 948 949 written = pol->pd_stat_fn(blkg->pd[i], buf+off, size-off); 950 if (written) 951 has_stats = true; 952 off += written; 953 } 954 955 if (has_stats) { 956 if (off < size - 1) { 957 off += scnprintf(buf+off, size-off, "\n"); 958 seq_commit(sf, off); 959 } else { 960 seq_commit(sf, -1); 961 } 962 } 963 skip: 964 spin_unlock_irq(&blkg->q->queue_lock); 965 } 966 967 rcu_read_unlock(); 968 return 0; 969 } 970 971 static struct cftype blkcg_files[] = { 972 { 973 .name = "stat", 974 .seq_show = blkcg_print_stat, 975 }, 976 { } /* terminate */ 977 }; 978 979 static struct cftype blkcg_legacy_files[] = { 980 { 981 .name = "reset_stats", 982 .write_u64 = blkcg_reset_stats, 983 }, 984 { } /* terminate */ 985 }; 986 987 /* 988 * blkcg destruction is a three-stage process. 989 * 990 * 1. Destruction starts. The blkcg_css_offline() callback is invoked 991 * which offlines writeback. Here we tie the next stage of blkg destruction 992 * to the completion of writeback associated with the blkcg. This lets us 993 * avoid punting potentially large amounts of outstanding writeback to root 994 * while maintaining any ongoing policies. The next stage is triggered when 995 * the nr_cgwbs count goes to zero. 996 * 997 * 2. When the nr_cgwbs count goes to zero, blkcg_destroy_blkgs() is called 998 * and handles the destruction of blkgs. Here the css reference held by 999 * the blkg is put back eventually allowing blkcg_css_free() to be called. 1000 * This work may occur in cgwb_release_workfn() on the cgwb_release 1001 * workqueue. Any submitted ios that fail to get the blkg ref will be 1002 * punted to the root_blkg. 1003 * 1004 * 3. Once the blkcg ref count goes to zero, blkcg_css_free() is called. 1005 * This finally frees the blkcg. 1006 */ 1007 1008 /** 1009 * blkcg_css_offline - cgroup css_offline callback 1010 * @css: css of interest 1011 * 1012 * This function is called when @css is about to go away. Here the cgwbs are 1013 * offlined first and only once writeback associated with the blkcg has 1014 * finished do we start step 2 (see above). 1015 */ 1016 static void blkcg_css_offline(struct cgroup_subsys_state *css) 1017 { 1018 struct blkcg *blkcg = css_to_blkcg(css); 1019 1020 /* this prevents anyone from attaching or migrating to this blkcg */ 1021 wb_blkcg_offline(blkcg); 1022 1023 /* put the base online pin allowing step 2 to be triggered */ 1024 blkcg_unpin_online(blkcg); 1025 } 1026 1027 /** 1028 * blkcg_destroy_blkgs - responsible for shooting down blkgs 1029 * @blkcg: blkcg of interest 1030 * 1031 * blkgs should be removed while holding both q and blkcg locks. As blkcg lock 1032 * is nested inside q lock, this function performs reverse double lock dancing. 1033 * Destroying the blkgs releases the reference held on the blkcg's css allowing 1034 * blkcg_css_free to eventually be called. 1035 * 1036 * This is the blkcg counterpart of ioc_release_fn(). 1037 */ 1038 void blkcg_destroy_blkgs(struct blkcg *blkcg) 1039 { 1040 might_sleep(); 1041 1042 spin_lock_irq(&blkcg->lock); 1043 1044 while (!hlist_empty(&blkcg->blkg_list)) { 1045 struct blkcg_gq *blkg = hlist_entry(blkcg->blkg_list.first, 1046 struct blkcg_gq, blkcg_node); 1047 struct request_queue *q = blkg->q; 1048 1049 if (need_resched() || !spin_trylock(&q->queue_lock)) { 1050 /* 1051 * Given that the system can accumulate a huge number 1052 * of blkgs in pathological cases, check to see if we 1053 * need to rescheduling to avoid softlockup. 1054 */ 1055 spin_unlock_irq(&blkcg->lock); 1056 cond_resched(); 1057 spin_lock_irq(&blkcg->lock); 1058 continue; 1059 } 1060 1061 blkg_destroy(blkg); 1062 spin_unlock(&q->queue_lock); 1063 } 1064 1065 spin_unlock_irq(&blkcg->lock); 1066 } 1067 1068 static void blkcg_css_free(struct cgroup_subsys_state *css) 1069 { 1070 struct blkcg *blkcg = css_to_blkcg(css); 1071 int i; 1072 1073 mutex_lock(&blkcg_pol_mutex); 1074 1075 list_del(&blkcg->all_blkcgs_node); 1076 1077 for (i = 0; i < BLKCG_MAX_POLS; i++) 1078 if (blkcg->cpd[i]) 1079 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]); 1080 1081 mutex_unlock(&blkcg_pol_mutex); 1082 1083 kfree(blkcg); 1084 } 1085 1086 static struct cgroup_subsys_state * 1087 blkcg_css_alloc(struct cgroup_subsys_state *parent_css) 1088 { 1089 struct blkcg *blkcg; 1090 struct cgroup_subsys_state *ret; 1091 int i; 1092 1093 mutex_lock(&blkcg_pol_mutex); 1094 1095 if (!parent_css) { 1096 blkcg = &blkcg_root; 1097 } else { 1098 blkcg = kzalloc(sizeof(*blkcg), GFP_KERNEL); 1099 if (!blkcg) { 1100 ret = ERR_PTR(-ENOMEM); 1101 goto unlock; 1102 } 1103 } 1104 1105 for (i = 0; i < BLKCG_MAX_POLS ; i++) { 1106 struct blkcg_policy *pol = blkcg_policy[i]; 1107 struct blkcg_policy_data *cpd; 1108 1109 /* 1110 * If the policy hasn't been attached yet, wait for it 1111 * to be attached before doing anything else. Otherwise, 1112 * check if the policy requires any specific per-cgroup 1113 * data: if it does, allocate and initialize it. 1114 */ 1115 if (!pol || !pol->cpd_alloc_fn) 1116 continue; 1117 1118 cpd = pol->cpd_alloc_fn(GFP_KERNEL); 1119 if (!cpd) { 1120 ret = ERR_PTR(-ENOMEM); 1121 goto free_pd_blkcg; 1122 } 1123 blkcg->cpd[i] = cpd; 1124 cpd->blkcg = blkcg; 1125 cpd->plid = i; 1126 if (pol->cpd_init_fn) 1127 pol->cpd_init_fn(cpd); 1128 } 1129 1130 spin_lock_init(&blkcg->lock); 1131 refcount_set(&blkcg->online_pin, 1); 1132 INIT_RADIX_TREE(&blkcg->blkg_tree, GFP_NOWAIT | __GFP_NOWARN); 1133 INIT_HLIST_HEAD(&blkcg->blkg_list); 1134 #ifdef CONFIG_CGROUP_WRITEBACK 1135 INIT_LIST_HEAD(&blkcg->cgwb_list); 1136 #endif 1137 list_add_tail(&blkcg->all_blkcgs_node, &all_blkcgs); 1138 1139 mutex_unlock(&blkcg_pol_mutex); 1140 return &blkcg->css; 1141 1142 free_pd_blkcg: 1143 for (i--; i >= 0; i--) 1144 if (blkcg->cpd[i]) 1145 blkcg_policy[i]->cpd_free_fn(blkcg->cpd[i]); 1146 1147 if (blkcg != &blkcg_root) 1148 kfree(blkcg); 1149 unlock: 1150 mutex_unlock(&blkcg_pol_mutex); 1151 return ret; 1152 } 1153 1154 static int blkcg_css_online(struct cgroup_subsys_state *css) 1155 { 1156 struct blkcg *blkcg = css_to_blkcg(css); 1157 struct blkcg *parent = blkcg_parent(blkcg); 1158 1159 /* 1160 * blkcg_pin_online() is used to delay blkcg offline so that blkgs 1161 * don't go offline while cgwbs are still active on them. Pin the 1162 * parent so that offline always happens towards the root. 1163 */ 1164 if (parent) 1165 blkcg_pin_online(parent); 1166 return 0; 1167 } 1168 1169 /** 1170 * blkcg_init_queue - initialize blkcg part of request queue 1171 * @q: request_queue to initialize 1172 * 1173 * Called from blk_alloc_queue(). Responsible for initializing blkcg 1174 * part of new request_queue @q. 1175 * 1176 * RETURNS: 1177 * 0 on success, -errno on failure. 1178 */ 1179 int blkcg_init_queue(struct request_queue *q) 1180 { 1181 struct blkcg_gq *new_blkg, *blkg; 1182 bool preloaded; 1183 int ret; 1184 1185 new_blkg = blkg_alloc(&blkcg_root, q, GFP_KERNEL); 1186 if (!new_blkg) 1187 return -ENOMEM; 1188 1189 preloaded = !radix_tree_preload(GFP_KERNEL); 1190 1191 /* Make sure the root blkg exists. */ 1192 rcu_read_lock(); 1193 spin_lock_irq(&q->queue_lock); 1194 blkg = blkg_create(&blkcg_root, q, new_blkg); 1195 if (IS_ERR(blkg)) 1196 goto err_unlock; 1197 q->root_blkg = blkg; 1198 spin_unlock_irq(&q->queue_lock); 1199 rcu_read_unlock(); 1200 1201 if (preloaded) 1202 radix_tree_preload_end(); 1203 1204 ret = blk_iolatency_init(q); 1205 if (ret) 1206 goto err_destroy_all; 1207 1208 ret = blk_ioprio_init(q); 1209 if (ret) 1210 goto err_destroy_all; 1211 1212 ret = blk_throtl_init(q); 1213 if (ret) 1214 goto err_destroy_all; 1215 1216 return 0; 1217 1218 err_destroy_all: 1219 blkg_destroy_all(q); 1220 return ret; 1221 err_unlock: 1222 spin_unlock_irq(&q->queue_lock); 1223 rcu_read_unlock(); 1224 if (preloaded) 1225 radix_tree_preload_end(); 1226 return PTR_ERR(blkg); 1227 } 1228 1229 /** 1230 * blkcg_exit_queue - exit and release blkcg part of request_queue 1231 * @q: request_queue being released 1232 * 1233 * Called from blk_exit_queue(). Responsible for exiting blkcg part. 1234 */ 1235 void blkcg_exit_queue(struct request_queue *q) 1236 { 1237 blkg_destroy_all(q); 1238 blk_throtl_exit(q); 1239 } 1240 1241 static void blkcg_bind(struct cgroup_subsys_state *root_css) 1242 { 1243 int i; 1244 1245 mutex_lock(&blkcg_pol_mutex); 1246 1247 for (i = 0; i < BLKCG_MAX_POLS; i++) { 1248 struct blkcg_policy *pol = blkcg_policy[i]; 1249 struct blkcg *blkcg; 1250 1251 if (!pol || !pol->cpd_bind_fn) 1252 continue; 1253 1254 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) 1255 if (blkcg->cpd[pol->plid]) 1256 pol->cpd_bind_fn(blkcg->cpd[pol->plid]); 1257 } 1258 mutex_unlock(&blkcg_pol_mutex); 1259 } 1260 1261 static void blkcg_exit(struct task_struct *tsk) 1262 { 1263 if (tsk->throttle_queue) 1264 blk_put_queue(tsk->throttle_queue); 1265 tsk->throttle_queue = NULL; 1266 } 1267 1268 struct cgroup_subsys io_cgrp_subsys = { 1269 .css_alloc = blkcg_css_alloc, 1270 .css_online = blkcg_css_online, 1271 .css_offline = blkcg_css_offline, 1272 .css_free = blkcg_css_free, 1273 .css_rstat_flush = blkcg_rstat_flush, 1274 .bind = blkcg_bind, 1275 .dfl_cftypes = blkcg_files, 1276 .legacy_cftypes = blkcg_legacy_files, 1277 .legacy_name = "blkio", 1278 .exit = blkcg_exit, 1279 #ifdef CONFIG_MEMCG 1280 /* 1281 * This ensures that, if available, memcg is automatically enabled 1282 * together on the default hierarchy so that the owner cgroup can 1283 * be retrieved from writeback pages. 1284 */ 1285 .depends_on = 1 << memory_cgrp_id, 1286 #endif 1287 }; 1288 EXPORT_SYMBOL_GPL(io_cgrp_subsys); 1289 1290 /** 1291 * blkcg_activate_policy - activate a blkcg policy on a request_queue 1292 * @q: request_queue of interest 1293 * @pol: blkcg policy to activate 1294 * 1295 * Activate @pol on @q. Requires %GFP_KERNEL context. @q goes through 1296 * bypass mode to populate its blkgs with policy_data for @pol. 1297 * 1298 * Activation happens with @q bypassed, so nobody would be accessing blkgs 1299 * from IO path. Update of each blkg is protected by both queue and blkcg 1300 * locks so that holding either lock and testing blkcg_policy_enabled() is 1301 * always enough for dereferencing policy data. 1302 * 1303 * The caller is responsible for synchronizing [de]activations and policy 1304 * [un]registerations. Returns 0 on success, -errno on failure. 1305 */ 1306 int blkcg_activate_policy(struct request_queue *q, 1307 const struct blkcg_policy *pol) 1308 { 1309 struct blkg_policy_data *pd_prealloc = NULL; 1310 struct blkcg_gq *blkg, *pinned_blkg = NULL; 1311 int ret; 1312 1313 if (blkcg_policy_enabled(q, pol)) 1314 return 0; 1315 1316 if (queue_is_mq(q)) 1317 blk_mq_freeze_queue(q); 1318 retry: 1319 spin_lock_irq(&q->queue_lock); 1320 1321 /* blkg_list is pushed at the head, reverse walk to allocate parents first */ 1322 list_for_each_entry_reverse(blkg, &q->blkg_list, q_node) { 1323 struct blkg_policy_data *pd; 1324 1325 if (blkg->pd[pol->plid]) 1326 continue; 1327 1328 /* If prealloc matches, use it; otherwise try GFP_NOWAIT */ 1329 if (blkg == pinned_blkg) { 1330 pd = pd_prealloc; 1331 pd_prealloc = NULL; 1332 } else { 1333 pd = pol->pd_alloc_fn(GFP_NOWAIT | __GFP_NOWARN, q, 1334 blkg->blkcg); 1335 } 1336 1337 if (!pd) { 1338 /* 1339 * GFP_NOWAIT failed. Free the existing one and 1340 * prealloc for @blkg w/ GFP_KERNEL. 1341 */ 1342 if (pinned_blkg) 1343 blkg_put(pinned_blkg); 1344 blkg_get(blkg); 1345 pinned_blkg = blkg; 1346 1347 spin_unlock_irq(&q->queue_lock); 1348 1349 if (pd_prealloc) 1350 pol->pd_free_fn(pd_prealloc); 1351 pd_prealloc = pol->pd_alloc_fn(GFP_KERNEL, q, 1352 blkg->blkcg); 1353 if (pd_prealloc) 1354 goto retry; 1355 else 1356 goto enomem; 1357 } 1358 1359 blkg->pd[pol->plid] = pd; 1360 pd->blkg = blkg; 1361 pd->plid = pol->plid; 1362 } 1363 1364 /* all allocated, init in the same order */ 1365 if (pol->pd_init_fn) 1366 list_for_each_entry_reverse(blkg, &q->blkg_list, q_node) 1367 pol->pd_init_fn(blkg->pd[pol->plid]); 1368 1369 __set_bit(pol->plid, q->blkcg_pols); 1370 ret = 0; 1371 1372 spin_unlock_irq(&q->queue_lock); 1373 out: 1374 if (queue_is_mq(q)) 1375 blk_mq_unfreeze_queue(q); 1376 if (pinned_blkg) 1377 blkg_put(pinned_blkg); 1378 if (pd_prealloc) 1379 pol->pd_free_fn(pd_prealloc); 1380 return ret; 1381 1382 enomem: 1383 /* alloc failed, nothing's initialized yet, free everything */ 1384 spin_lock_irq(&q->queue_lock); 1385 list_for_each_entry(blkg, &q->blkg_list, q_node) { 1386 if (blkg->pd[pol->plid]) { 1387 pol->pd_free_fn(blkg->pd[pol->plid]); 1388 blkg->pd[pol->plid] = NULL; 1389 } 1390 } 1391 spin_unlock_irq(&q->queue_lock); 1392 ret = -ENOMEM; 1393 goto out; 1394 } 1395 EXPORT_SYMBOL_GPL(blkcg_activate_policy); 1396 1397 /** 1398 * blkcg_deactivate_policy - deactivate a blkcg policy on a request_queue 1399 * @q: request_queue of interest 1400 * @pol: blkcg policy to deactivate 1401 * 1402 * Deactivate @pol on @q. Follows the same synchronization rules as 1403 * blkcg_activate_policy(). 1404 */ 1405 void blkcg_deactivate_policy(struct request_queue *q, 1406 const struct blkcg_policy *pol) 1407 { 1408 struct blkcg_gq *blkg; 1409 1410 if (!blkcg_policy_enabled(q, pol)) 1411 return; 1412 1413 if (queue_is_mq(q)) 1414 blk_mq_freeze_queue(q); 1415 1416 spin_lock_irq(&q->queue_lock); 1417 1418 __clear_bit(pol->plid, q->blkcg_pols); 1419 1420 list_for_each_entry(blkg, &q->blkg_list, q_node) { 1421 if (blkg->pd[pol->plid]) { 1422 if (pol->pd_offline_fn) 1423 pol->pd_offline_fn(blkg->pd[pol->plid]); 1424 pol->pd_free_fn(blkg->pd[pol->plid]); 1425 blkg->pd[pol->plid] = NULL; 1426 } 1427 } 1428 1429 spin_unlock_irq(&q->queue_lock); 1430 1431 if (queue_is_mq(q)) 1432 blk_mq_unfreeze_queue(q); 1433 } 1434 EXPORT_SYMBOL_GPL(blkcg_deactivate_policy); 1435 1436 /** 1437 * blkcg_policy_register - register a blkcg policy 1438 * @pol: blkcg policy to register 1439 * 1440 * Register @pol with blkcg core. Might sleep and @pol may be modified on 1441 * successful registration. Returns 0 on success and -errno on failure. 1442 */ 1443 int blkcg_policy_register(struct blkcg_policy *pol) 1444 { 1445 struct blkcg *blkcg; 1446 int i, ret; 1447 1448 mutex_lock(&blkcg_pol_register_mutex); 1449 mutex_lock(&blkcg_pol_mutex); 1450 1451 /* find an empty slot */ 1452 ret = -ENOSPC; 1453 for (i = 0; i < BLKCG_MAX_POLS; i++) 1454 if (!blkcg_policy[i]) 1455 break; 1456 if (i >= BLKCG_MAX_POLS) { 1457 pr_warn("blkcg_policy_register: BLKCG_MAX_POLS too small\n"); 1458 goto err_unlock; 1459 } 1460 1461 /* Make sure cpd/pd_alloc_fn and cpd/pd_free_fn in pairs */ 1462 if ((!pol->cpd_alloc_fn ^ !pol->cpd_free_fn) || 1463 (!pol->pd_alloc_fn ^ !pol->pd_free_fn)) 1464 goto err_unlock; 1465 1466 /* register @pol */ 1467 pol->plid = i; 1468 blkcg_policy[pol->plid] = pol; 1469 1470 /* allocate and install cpd's */ 1471 if (pol->cpd_alloc_fn) { 1472 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) { 1473 struct blkcg_policy_data *cpd; 1474 1475 cpd = pol->cpd_alloc_fn(GFP_KERNEL); 1476 if (!cpd) 1477 goto err_free_cpds; 1478 1479 blkcg->cpd[pol->plid] = cpd; 1480 cpd->blkcg = blkcg; 1481 cpd->plid = pol->plid; 1482 if (pol->cpd_init_fn) 1483 pol->cpd_init_fn(cpd); 1484 } 1485 } 1486 1487 mutex_unlock(&blkcg_pol_mutex); 1488 1489 /* everything is in place, add intf files for the new policy */ 1490 if (pol->dfl_cftypes) 1491 WARN_ON(cgroup_add_dfl_cftypes(&io_cgrp_subsys, 1492 pol->dfl_cftypes)); 1493 if (pol->legacy_cftypes) 1494 WARN_ON(cgroup_add_legacy_cftypes(&io_cgrp_subsys, 1495 pol->legacy_cftypes)); 1496 mutex_unlock(&blkcg_pol_register_mutex); 1497 return 0; 1498 1499 err_free_cpds: 1500 if (pol->cpd_free_fn) { 1501 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) { 1502 if (blkcg->cpd[pol->plid]) { 1503 pol->cpd_free_fn(blkcg->cpd[pol->plid]); 1504 blkcg->cpd[pol->plid] = NULL; 1505 } 1506 } 1507 } 1508 blkcg_policy[pol->plid] = NULL; 1509 err_unlock: 1510 mutex_unlock(&blkcg_pol_mutex); 1511 mutex_unlock(&blkcg_pol_register_mutex); 1512 return ret; 1513 } 1514 EXPORT_SYMBOL_GPL(blkcg_policy_register); 1515 1516 /** 1517 * blkcg_policy_unregister - unregister a blkcg policy 1518 * @pol: blkcg policy to unregister 1519 * 1520 * Undo blkcg_policy_register(@pol). Might sleep. 1521 */ 1522 void blkcg_policy_unregister(struct blkcg_policy *pol) 1523 { 1524 struct blkcg *blkcg; 1525 1526 mutex_lock(&blkcg_pol_register_mutex); 1527 1528 if (WARN_ON(blkcg_policy[pol->plid] != pol)) 1529 goto out_unlock; 1530 1531 /* kill the intf files first */ 1532 if (pol->dfl_cftypes) 1533 cgroup_rm_cftypes(pol->dfl_cftypes); 1534 if (pol->legacy_cftypes) 1535 cgroup_rm_cftypes(pol->legacy_cftypes); 1536 1537 /* remove cpds and unregister */ 1538 mutex_lock(&blkcg_pol_mutex); 1539 1540 if (pol->cpd_free_fn) { 1541 list_for_each_entry(blkcg, &all_blkcgs, all_blkcgs_node) { 1542 if (blkcg->cpd[pol->plid]) { 1543 pol->cpd_free_fn(blkcg->cpd[pol->plid]); 1544 blkcg->cpd[pol->plid] = NULL; 1545 } 1546 } 1547 } 1548 blkcg_policy[pol->plid] = NULL; 1549 1550 mutex_unlock(&blkcg_pol_mutex); 1551 out_unlock: 1552 mutex_unlock(&blkcg_pol_register_mutex); 1553 } 1554 EXPORT_SYMBOL_GPL(blkcg_policy_unregister); 1555 1556 bool __blkcg_punt_bio_submit(struct bio *bio) 1557 { 1558 struct blkcg_gq *blkg = bio->bi_blkg; 1559 1560 /* consume the flag first */ 1561 bio->bi_opf &= ~REQ_CGROUP_PUNT; 1562 1563 /* never bounce for the root cgroup */ 1564 if (!blkg->parent) 1565 return false; 1566 1567 spin_lock_bh(&blkg->async_bio_lock); 1568 bio_list_add(&blkg->async_bios, bio); 1569 spin_unlock_bh(&blkg->async_bio_lock); 1570 1571 queue_work(blkcg_punt_bio_wq, &blkg->async_bio_work); 1572 return true; 1573 } 1574 1575 /* 1576 * Scale the accumulated delay based on how long it has been since we updated 1577 * the delay. We only call this when we are adding delay, in case it's been a 1578 * while since we added delay, and when we are checking to see if we need to 1579 * delay a task, to account for any delays that may have occurred. 1580 */ 1581 static void blkcg_scale_delay(struct blkcg_gq *blkg, u64 now) 1582 { 1583 u64 old = atomic64_read(&blkg->delay_start); 1584 1585 /* negative use_delay means no scaling, see blkcg_set_delay() */ 1586 if (atomic_read(&blkg->use_delay) < 0) 1587 return; 1588 1589 /* 1590 * We only want to scale down every second. The idea here is that we 1591 * want to delay people for min(delay_nsec, NSEC_PER_SEC) in a certain 1592 * time window. We only want to throttle tasks for recent delay that 1593 * has occurred, in 1 second time windows since that's the maximum 1594 * things can be throttled. We save the current delay window in 1595 * blkg->last_delay so we know what amount is still left to be charged 1596 * to the blkg from this point onward. blkg->last_use keeps track of 1597 * the use_delay counter. The idea is if we're unthrottling the blkg we 1598 * are ok with whatever is happening now, and we can take away more of 1599 * the accumulated delay as we've already throttled enough that 1600 * everybody is happy with their IO latencies. 1601 */ 1602 if (time_before64(old + NSEC_PER_SEC, now) && 1603 atomic64_cmpxchg(&blkg->delay_start, old, now) == old) { 1604 u64 cur = atomic64_read(&blkg->delay_nsec); 1605 u64 sub = min_t(u64, blkg->last_delay, now - old); 1606 int cur_use = atomic_read(&blkg->use_delay); 1607 1608 /* 1609 * We've been unthrottled, subtract a larger chunk of our 1610 * accumulated delay. 1611 */ 1612 if (cur_use < blkg->last_use) 1613 sub = max_t(u64, sub, blkg->last_delay >> 1); 1614 1615 /* 1616 * This shouldn't happen, but handle it anyway. Our delay_nsec 1617 * should only ever be growing except here where we subtract out 1618 * min(last_delay, 1 second), but lord knows bugs happen and I'd 1619 * rather not end up with negative numbers. 1620 */ 1621 if (unlikely(cur < sub)) { 1622 atomic64_set(&blkg->delay_nsec, 0); 1623 blkg->last_delay = 0; 1624 } else { 1625 atomic64_sub(sub, &blkg->delay_nsec); 1626 blkg->last_delay = cur - sub; 1627 } 1628 blkg->last_use = cur_use; 1629 } 1630 } 1631 1632 /* 1633 * This is called when we want to actually walk up the hierarchy and check to 1634 * see if we need to throttle, and then actually throttle if there is some 1635 * accumulated delay. This should only be called upon return to user space so 1636 * we're not holding some lock that would induce a priority inversion. 1637 */ 1638 static void blkcg_maybe_throttle_blkg(struct blkcg_gq *blkg, bool use_memdelay) 1639 { 1640 unsigned long pflags; 1641 bool clamp; 1642 u64 now = ktime_to_ns(ktime_get()); 1643 u64 exp; 1644 u64 delay_nsec = 0; 1645 int tok; 1646 1647 while (blkg->parent) { 1648 int use_delay = atomic_read(&blkg->use_delay); 1649 1650 if (use_delay) { 1651 u64 this_delay; 1652 1653 blkcg_scale_delay(blkg, now); 1654 this_delay = atomic64_read(&blkg->delay_nsec); 1655 if (this_delay > delay_nsec) { 1656 delay_nsec = this_delay; 1657 clamp = use_delay > 0; 1658 } 1659 } 1660 blkg = blkg->parent; 1661 } 1662 1663 if (!delay_nsec) 1664 return; 1665 1666 /* 1667 * Let's not sleep for all eternity if we've amassed a huge delay. 1668 * Swapping or metadata IO can accumulate 10's of seconds worth of 1669 * delay, and we want userspace to be able to do _something_ so cap the 1670 * delays at 0.25s. If there's 10's of seconds worth of delay then the 1671 * tasks will be delayed for 0.25 second for every syscall. If 1672 * blkcg_set_delay() was used as indicated by negative use_delay, the 1673 * caller is responsible for regulating the range. 1674 */ 1675 if (clamp) 1676 delay_nsec = min_t(u64, delay_nsec, 250 * NSEC_PER_MSEC); 1677 1678 if (use_memdelay) 1679 psi_memstall_enter(&pflags); 1680 1681 exp = ktime_add_ns(now, delay_nsec); 1682 tok = io_schedule_prepare(); 1683 do { 1684 __set_current_state(TASK_KILLABLE); 1685 if (!schedule_hrtimeout(&exp, HRTIMER_MODE_ABS)) 1686 break; 1687 } while (!fatal_signal_pending(current)); 1688 io_schedule_finish(tok); 1689 1690 if (use_memdelay) 1691 psi_memstall_leave(&pflags); 1692 } 1693 1694 /** 1695 * blkcg_maybe_throttle_current - throttle the current task if it has been marked 1696 * 1697 * This is only called if we've been marked with set_notify_resume(). Obviously 1698 * we can be set_notify_resume() for reasons other than blkcg throttling, so we 1699 * check to see if current->throttle_queue is set and if not this doesn't do 1700 * anything. This should only ever be called by the resume code, it's not meant 1701 * to be called by people willy-nilly as it will actually do the work to 1702 * throttle the task if it is setup for throttling. 1703 */ 1704 void blkcg_maybe_throttle_current(void) 1705 { 1706 struct request_queue *q = current->throttle_queue; 1707 struct cgroup_subsys_state *css; 1708 struct blkcg *blkcg; 1709 struct blkcg_gq *blkg; 1710 bool use_memdelay = current->use_memdelay; 1711 1712 if (!q) 1713 return; 1714 1715 current->throttle_queue = NULL; 1716 current->use_memdelay = false; 1717 1718 rcu_read_lock(); 1719 css = kthread_blkcg(); 1720 if (css) 1721 blkcg = css_to_blkcg(css); 1722 else 1723 blkcg = css_to_blkcg(task_css(current, io_cgrp_id)); 1724 1725 if (!blkcg) 1726 goto out; 1727 blkg = blkg_lookup(blkcg, q); 1728 if (!blkg) 1729 goto out; 1730 if (!blkg_tryget(blkg)) 1731 goto out; 1732 rcu_read_unlock(); 1733 1734 blkcg_maybe_throttle_blkg(blkg, use_memdelay); 1735 blkg_put(blkg); 1736 blk_put_queue(q); 1737 return; 1738 out: 1739 rcu_read_unlock(); 1740 blk_put_queue(q); 1741 } 1742 1743 /** 1744 * blkcg_schedule_throttle - this task needs to check for throttling 1745 * @q: the request queue IO was submitted on 1746 * @use_memdelay: do we charge this to memory delay for PSI 1747 * 1748 * This is called by the IO controller when we know there's delay accumulated 1749 * for the blkg for this task. We do not pass the blkg because there are places 1750 * we call this that may not have that information, the swapping code for 1751 * instance will only have a request_queue at that point. This set's the 1752 * notify_resume for the task to check and see if it requires throttling before 1753 * returning to user space. 1754 * 1755 * We will only schedule once per syscall. You can call this over and over 1756 * again and it will only do the check once upon return to user space, and only 1757 * throttle once. If the task needs to be throttled again it'll need to be 1758 * re-set at the next time we see the task. 1759 */ 1760 void blkcg_schedule_throttle(struct request_queue *q, bool use_memdelay) 1761 { 1762 if (unlikely(current->flags & PF_KTHREAD)) 1763 return; 1764 1765 if (current->throttle_queue != q) { 1766 if (!blk_get_queue(q)) 1767 return; 1768 1769 if (current->throttle_queue) 1770 blk_put_queue(current->throttle_queue); 1771 current->throttle_queue = q; 1772 } 1773 1774 if (use_memdelay) 1775 current->use_memdelay = use_memdelay; 1776 set_notify_resume(current); 1777 } 1778 1779 /** 1780 * blkcg_add_delay - add delay to this blkg 1781 * @blkg: blkg of interest 1782 * @now: the current time in nanoseconds 1783 * @delta: how many nanoseconds of delay to add 1784 * 1785 * Charge @delta to the blkg's current delay accumulation. This is used to 1786 * throttle tasks if an IO controller thinks we need more throttling. 1787 */ 1788 void blkcg_add_delay(struct blkcg_gq *blkg, u64 now, u64 delta) 1789 { 1790 if (WARN_ON_ONCE(atomic_read(&blkg->use_delay) < 0)) 1791 return; 1792 blkcg_scale_delay(blkg, now); 1793 atomic64_add(delta, &blkg->delay_nsec); 1794 } 1795 1796 /** 1797 * blkg_tryget_closest - try and get a blkg ref on the closet blkg 1798 * @bio: target bio 1799 * @css: target css 1800 * 1801 * As the failure mode here is to walk up the blkg tree, this ensure that the 1802 * blkg->parent pointers are always valid. This returns the blkg that it ended 1803 * up taking a reference on or %NULL if no reference was taken. 1804 */ 1805 static inline struct blkcg_gq *blkg_tryget_closest(struct bio *bio, 1806 struct cgroup_subsys_state *css) 1807 { 1808 struct blkcg_gq *blkg, *ret_blkg = NULL; 1809 1810 rcu_read_lock(); 1811 blkg = blkg_lookup_create(css_to_blkcg(css), 1812 bio->bi_bdev->bd_disk->queue); 1813 while (blkg) { 1814 if (blkg_tryget(blkg)) { 1815 ret_blkg = blkg; 1816 break; 1817 } 1818 blkg = blkg->parent; 1819 } 1820 rcu_read_unlock(); 1821 1822 return ret_blkg; 1823 } 1824 1825 /** 1826 * bio_associate_blkg_from_css - associate a bio with a specified css 1827 * @bio: target bio 1828 * @css: target css 1829 * 1830 * Associate @bio with the blkg found by combining the css's blkg and the 1831 * request_queue of the @bio. An association failure is handled by walking up 1832 * the blkg tree. Therefore, the blkg associated can be anything between @blkg 1833 * and q->root_blkg. This situation only happens when a cgroup is dying and 1834 * then the remaining bios will spill to the closest alive blkg. 1835 * 1836 * A reference will be taken on the blkg and will be released when @bio is 1837 * freed. 1838 */ 1839 void bio_associate_blkg_from_css(struct bio *bio, 1840 struct cgroup_subsys_state *css) 1841 { 1842 if (bio->bi_blkg) 1843 blkg_put(bio->bi_blkg); 1844 1845 if (css && css->parent) { 1846 bio->bi_blkg = blkg_tryget_closest(bio, css); 1847 } else { 1848 blkg_get(bio->bi_bdev->bd_disk->queue->root_blkg); 1849 bio->bi_blkg = bio->bi_bdev->bd_disk->queue->root_blkg; 1850 } 1851 } 1852 EXPORT_SYMBOL_GPL(bio_associate_blkg_from_css); 1853 1854 /** 1855 * bio_associate_blkg - associate a bio with a blkg 1856 * @bio: target bio 1857 * 1858 * Associate @bio with the blkg found from the bio's css and request_queue. 1859 * If one is not found, bio_lookup_blkg() creates the blkg. If a blkg is 1860 * already associated, the css is reused and association redone as the 1861 * request_queue may have changed. 1862 */ 1863 void bio_associate_blkg(struct bio *bio) 1864 { 1865 struct cgroup_subsys_state *css; 1866 1867 rcu_read_lock(); 1868 1869 if (bio->bi_blkg) 1870 css = &bio_blkcg(bio)->css; 1871 else 1872 css = blkcg_css(); 1873 1874 bio_associate_blkg_from_css(bio, css); 1875 1876 rcu_read_unlock(); 1877 } 1878 EXPORT_SYMBOL_GPL(bio_associate_blkg); 1879 1880 /** 1881 * bio_clone_blkg_association - clone blkg association from src to dst bio 1882 * @dst: destination bio 1883 * @src: source bio 1884 */ 1885 void bio_clone_blkg_association(struct bio *dst, struct bio *src) 1886 { 1887 if (src->bi_blkg) { 1888 if (dst->bi_blkg) 1889 blkg_put(dst->bi_blkg); 1890 blkg_get(src->bi_blkg); 1891 dst->bi_blkg = src->bi_blkg; 1892 } 1893 } 1894 EXPORT_SYMBOL_GPL(bio_clone_blkg_association); 1895 1896 static int blk_cgroup_io_type(struct bio *bio) 1897 { 1898 if (op_is_discard(bio->bi_opf)) 1899 return BLKG_IOSTAT_DISCARD; 1900 if (op_is_write(bio->bi_opf)) 1901 return BLKG_IOSTAT_WRITE; 1902 return BLKG_IOSTAT_READ; 1903 } 1904 1905 void blk_cgroup_bio_start(struct bio *bio) 1906 { 1907 int rwd = blk_cgroup_io_type(bio), cpu; 1908 struct blkg_iostat_set *bis; 1909 1910 cpu = get_cpu(); 1911 bis = per_cpu_ptr(bio->bi_blkg->iostat_cpu, cpu); 1912 u64_stats_update_begin(&bis->sync); 1913 1914 /* 1915 * If the bio is flagged with BIO_CGROUP_ACCT it means this is a split 1916 * bio and we would have already accounted for the size of the bio. 1917 */ 1918 if (!bio_flagged(bio, BIO_CGROUP_ACCT)) { 1919 bio_set_flag(bio, BIO_CGROUP_ACCT); 1920 bis->cur.bytes[rwd] += bio->bi_iter.bi_size; 1921 } 1922 bis->cur.ios[rwd]++; 1923 1924 u64_stats_update_end(&bis->sync); 1925 if (cgroup_subsys_on_dfl(io_cgrp_subsys)) 1926 cgroup_rstat_updated(bio->bi_blkg->blkcg->css.cgroup, cpu); 1927 put_cpu(); 1928 } 1929 1930 static int __init blkcg_init(void) 1931 { 1932 blkcg_punt_bio_wq = alloc_workqueue("blkcg_punt_bio", 1933 WQ_MEM_RECLAIM | WQ_FREEZABLE | 1934 WQ_UNBOUND | WQ_SYSFS, 0); 1935 if (!blkcg_punt_bio_wq) 1936 return -ENOMEM; 1937 return 0; 1938 } 1939 subsys_initcall(blkcg_init); 1940 1941 module_param(blkcg_debug_stats, bool, 0644); 1942 MODULE_PARM_DESC(blkcg_debug_stats, "True if you want debug stats, false if not"); 1943