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