1 /* 2 * fs/fs-writeback.c 3 * 4 * Copyright (C) 2002, Linus Torvalds. 5 * 6 * Contains all the functions related to writing back and waiting 7 * upon dirty inodes against superblocks, and writing back dirty 8 * pages against inodes. ie: data writeback. Writeout of the 9 * inode itself is not handled here. 10 * 11 * 10Apr2002 Andrew Morton 12 * Split out of fs/inode.c 13 * Additions for address_space-based writeback 14 */ 15 16 #include <linux/kernel.h> 17 #include <linux/export.h> 18 #include <linux/spinlock.h> 19 #include <linux/slab.h> 20 #include <linux/sched.h> 21 #include <linux/fs.h> 22 #include <linux/mm.h> 23 #include <linux/pagemap.h> 24 #include <linux/kthread.h> 25 #include <linux/writeback.h> 26 #include <linux/blkdev.h> 27 #include <linux/backing-dev.h> 28 #include <linux/tracepoint.h> 29 #include <linux/device.h> 30 #include <linux/memcontrol.h> 31 #include "internal.h" 32 33 /* 34 * 4MB minimal write chunk size 35 */ 36 #define MIN_WRITEBACK_PAGES (4096UL >> (PAGE_CACHE_SHIFT - 10)) 37 38 struct wb_completion { 39 atomic_t cnt; 40 }; 41 42 /* 43 * Passed into wb_writeback(), essentially a subset of writeback_control 44 */ 45 struct wb_writeback_work { 46 long nr_pages; 47 struct super_block *sb; 48 unsigned long *older_than_this; 49 enum writeback_sync_modes sync_mode; 50 unsigned int tagged_writepages:1; 51 unsigned int for_kupdate:1; 52 unsigned int range_cyclic:1; 53 unsigned int for_background:1; 54 unsigned int for_sync:1; /* sync(2) WB_SYNC_ALL writeback */ 55 unsigned int auto_free:1; /* free on completion */ 56 unsigned int single_wait:1; 57 unsigned int single_done:1; 58 enum wb_reason reason; /* why was writeback initiated? */ 59 60 struct list_head list; /* pending work list */ 61 struct wb_completion *done; /* set if the caller waits */ 62 }; 63 64 /* 65 * If one wants to wait for one or more wb_writeback_works, each work's 66 * ->done should be set to a wb_completion defined using the following 67 * macro. Once all work items are issued with wb_queue_work(), the caller 68 * can wait for the completion of all using wb_wait_for_completion(). Work 69 * items which are waited upon aren't freed automatically on completion. 70 */ 71 #define DEFINE_WB_COMPLETION_ONSTACK(cmpl) \ 72 struct wb_completion cmpl = { \ 73 .cnt = ATOMIC_INIT(1), \ 74 } 75 76 77 /* 78 * If an inode is constantly having its pages dirtied, but then the 79 * updates stop dirtytime_expire_interval seconds in the past, it's 80 * possible for the worst case time between when an inode has its 81 * timestamps updated and when they finally get written out to be two 82 * dirtytime_expire_intervals. We set the default to 12 hours (in 83 * seconds), which means most of the time inodes will have their 84 * timestamps written to disk after 12 hours, but in the worst case a 85 * few inodes might not their timestamps updated for 24 hours. 86 */ 87 unsigned int dirtytime_expire_interval = 12 * 60 * 60; 88 89 static inline struct inode *wb_inode(struct list_head *head) 90 { 91 return list_entry(head, struct inode, i_wb_list); 92 } 93 94 /* 95 * Include the creation of the trace points after defining the 96 * wb_writeback_work structure and inline functions so that the definition 97 * remains local to this file. 98 */ 99 #define CREATE_TRACE_POINTS 100 #include <trace/events/writeback.h> 101 102 EXPORT_TRACEPOINT_SYMBOL_GPL(wbc_writepage); 103 104 static bool wb_io_lists_populated(struct bdi_writeback *wb) 105 { 106 if (wb_has_dirty_io(wb)) { 107 return false; 108 } else { 109 set_bit(WB_has_dirty_io, &wb->state); 110 WARN_ON_ONCE(!wb->avg_write_bandwidth); 111 atomic_long_add(wb->avg_write_bandwidth, 112 &wb->bdi->tot_write_bandwidth); 113 return true; 114 } 115 } 116 117 static void wb_io_lists_depopulated(struct bdi_writeback *wb) 118 { 119 if (wb_has_dirty_io(wb) && list_empty(&wb->b_dirty) && 120 list_empty(&wb->b_io) && list_empty(&wb->b_more_io)) { 121 clear_bit(WB_has_dirty_io, &wb->state); 122 WARN_ON_ONCE(atomic_long_sub_return(wb->avg_write_bandwidth, 123 &wb->bdi->tot_write_bandwidth) < 0); 124 } 125 } 126 127 /** 128 * inode_wb_list_move_locked - move an inode onto a bdi_writeback IO list 129 * @inode: inode to be moved 130 * @wb: target bdi_writeback 131 * @head: one of @wb->b_{dirty|io|more_io} 132 * 133 * Move @inode->i_wb_list to @list of @wb and set %WB_has_dirty_io. 134 * Returns %true if @inode is the first occupant of the !dirty_time IO 135 * lists; otherwise, %false. 136 */ 137 static bool inode_wb_list_move_locked(struct inode *inode, 138 struct bdi_writeback *wb, 139 struct list_head *head) 140 { 141 assert_spin_locked(&wb->list_lock); 142 143 list_move(&inode->i_wb_list, head); 144 145 /* dirty_time doesn't count as dirty_io until expiration */ 146 if (head != &wb->b_dirty_time) 147 return wb_io_lists_populated(wb); 148 149 wb_io_lists_depopulated(wb); 150 return false; 151 } 152 153 /** 154 * inode_wb_list_del_locked - remove an inode from its bdi_writeback IO list 155 * @inode: inode to be removed 156 * @wb: bdi_writeback @inode is being removed from 157 * 158 * Remove @inode which may be on one of @wb->b_{dirty|io|more_io} lists and 159 * clear %WB_has_dirty_io if all are empty afterwards. 160 */ 161 static void inode_wb_list_del_locked(struct inode *inode, 162 struct bdi_writeback *wb) 163 { 164 assert_spin_locked(&wb->list_lock); 165 166 list_del_init(&inode->i_wb_list); 167 wb_io_lists_depopulated(wb); 168 } 169 170 static void wb_wakeup(struct bdi_writeback *wb) 171 { 172 spin_lock_bh(&wb->work_lock); 173 if (test_bit(WB_registered, &wb->state)) 174 mod_delayed_work(bdi_wq, &wb->dwork, 0); 175 spin_unlock_bh(&wb->work_lock); 176 } 177 178 static void wb_queue_work(struct bdi_writeback *wb, 179 struct wb_writeback_work *work) 180 { 181 trace_writeback_queue(wb->bdi, work); 182 183 spin_lock_bh(&wb->work_lock); 184 if (!test_bit(WB_registered, &wb->state)) { 185 if (work->single_wait) 186 work->single_done = 1; 187 goto out_unlock; 188 } 189 if (work->done) 190 atomic_inc(&work->done->cnt); 191 list_add_tail(&work->list, &wb->work_list); 192 mod_delayed_work(bdi_wq, &wb->dwork, 0); 193 out_unlock: 194 spin_unlock_bh(&wb->work_lock); 195 } 196 197 /** 198 * wb_wait_for_completion - wait for completion of bdi_writeback_works 199 * @bdi: bdi work items were issued to 200 * @done: target wb_completion 201 * 202 * Wait for one or more work items issued to @bdi with their ->done field 203 * set to @done, which should have been defined with 204 * DEFINE_WB_COMPLETION_ONSTACK(). This function returns after all such 205 * work items are completed. Work items which are waited upon aren't freed 206 * automatically on completion. 207 */ 208 static void wb_wait_for_completion(struct backing_dev_info *bdi, 209 struct wb_completion *done) 210 { 211 atomic_dec(&done->cnt); /* put down the initial count */ 212 wait_event(bdi->wb_waitq, !atomic_read(&done->cnt)); 213 } 214 215 #ifdef CONFIG_CGROUP_WRITEBACK 216 217 /* parameters for foreign inode detection, see wb_detach_inode() */ 218 #define WB_FRN_TIME_SHIFT 13 /* 1s = 2^13, upto 8 secs w/ 16bit */ 219 #define WB_FRN_TIME_AVG_SHIFT 3 /* avg = avg * 7/8 + new * 1/8 */ 220 #define WB_FRN_TIME_CUT_DIV 2 /* ignore rounds < avg / 2 */ 221 #define WB_FRN_TIME_PERIOD (2 * (1 << WB_FRN_TIME_SHIFT)) /* 2s */ 222 223 #define WB_FRN_HIST_SLOTS 16 /* inode->i_wb_frn_history is 16bit */ 224 #define WB_FRN_HIST_UNIT (WB_FRN_TIME_PERIOD / WB_FRN_HIST_SLOTS) 225 /* each slot's duration is 2s / 16 */ 226 #define WB_FRN_HIST_THR_SLOTS (WB_FRN_HIST_SLOTS / 2) 227 /* if foreign slots >= 8, switch */ 228 #define WB_FRN_HIST_MAX_SLOTS (WB_FRN_HIST_THR_SLOTS / 2 + 1) 229 /* one round can affect upto 5 slots */ 230 231 void __inode_attach_wb(struct inode *inode, struct page *page) 232 { 233 struct backing_dev_info *bdi = inode_to_bdi(inode); 234 struct bdi_writeback *wb = NULL; 235 236 if (inode_cgwb_enabled(inode)) { 237 struct cgroup_subsys_state *memcg_css; 238 239 if (page) { 240 memcg_css = mem_cgroup_css_from_page(page); 241 wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC); 242 } else { 243 /* must pin memcg_css, see wb_get_create() */ 244 memcg_css = task_get_css(current, memory_cgrp_id); 245 wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC); 246 css_put(memcg_css); 247 } 248 } 249 250 if (!wb) 251 wb = &bdi->wb; 252 253 /* 254 * There may be multiple instances of this function racing to 255 * update the same inode. Use cmpxchg() to tell the winner. 256 */ 257 if (unlikely(cmpxchg(&inode->i_wb, NULL, wb))) 258 wb_put(wb); 259 } 260 261 /** 262 * locked_inode_to_wb_and_lock_list - determine a locked inode's wb and lock it 263 * @inode: inode of interest with i_lock held 264 * 265 * Returns @inode's wb with its list_lock held. @inode->i_lock must be 266 * held on entry and is released on return. The returned wb is guaranteed 267 * to stay @inode's associated wb until its list_lock is released. 268 */ 269 static struct bdi_writeback * 270 locked_inode_to_wb_and_lock_list(struct inode *inode) 271 __releases(&inode->i_lock) 272 __acquires(&wb->list_lock) 273 { 274 while (true) { 275 struct bdi_writeback *wb = inode_to_wb(inode); 276 277 /* 278 * inode_to_wb() association is protected by both 279 * @inode->i_lock and @wb->list_lock but list_lock nests 280 * outside i_lock. Drop i_lock and verify that the 281 * association hasn't changed after acquiring list_lock. 282 */ 283 wb_get(wb); 284 spin_unlock(&inode->i_lock); 285 spin_lock(&wb->list_lock); 286 wb_put(wb); /* not gonna deref it anymore */ 287 288 /* i_wb may have changed inbetween, can't use inode_to_wb() */ 289 if (likely(wb == inode->i_wb)) 290 return wb; /* @inode already has ref */ 291 292 spin_unlock(&wb->list_lock); 293 cpu_relax(); 294 spin_lock(&inode->i_lock); 295 } 296 } 297 298 /** 299 * inode_to_wb_and_lock_list - determine an inode's wb and lock it 300 * @inode: inode of interest 301 * 302 * Same as locked_inode_to_wb_and_lock_list() but @inode->i_lock isn't held 303 * on entry. 304 */ 305 static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode) 306 __acquires(&wb->list_lock) 307 { 308 spin_lock(&inode->i_lock); 309 return locked_inode_to_wb_and_lock_list(inode); 310 } 311 312 struct inode_switch_wbs_context { 313 struct inode *inode; 314 struct bdi_writeback *new_wb; 315 316 struct rcu_head rcu_head; 317 struct work_struct work; 318 }; 319 320 static void inode_switch_wbs_work_fn(struct work_struct *work) 321 { 322 struct inode_switch_wbs_context *isw = 323 container_of(work, struct inode_switch_wbs_context, work); 324 struct inode *inode = isw->inode; 325 struct address_space *mapping = inode->i_mapping; 326 struct bdi_writeback *old_wb = inode->i_wb; 327 struct bdi_writeback *new_wb = isw->new_wb; 328 struct radix_tree_iter iter; 329 bool switched = false; 330 void **slot; 331 332 /* 333 * By the time control reaches here, RCU grace period has passed 334 * since I_WB_SWITCH assertion and all wb stat update transactions 335 * between unlocked_inode_to_wb_begin/end() are guaranteed to be 336 * synchronizing against mapping->tree_lock. 337 * 338 * Grabbing old_wb->list_lock, inode->i_lock and mapping->tree_lock 339 * gives us exclusion against all wb related operations on @inode 340 * including IO list manipulations and stat updates. 341 */ 342 if (old_wb < new_wb) { 343 spin_lock(&old_wb->list_lock); 344 spin_lock_nested(&new_wb->list_lock, SINGLE_DEPTH_NESTING); 345 } else { 346 spin_lock(&new_wb->list_lock); 347 spin_lock_nested(&old_wb->list_lock, SINGLE_DEPTH_NESTING); 348 } 349 spin_lock(&inode->i_lock); 350 spin_lock_irq(&mapping->tree_lock); 351 352 /* 353 * Once I_FREEING is visible under i_lock, the eviction path owns 354 * the inode and we shouldn't modify ->i_wb_list. 355 */ 356 if (unlikely(inode->i_state & I_FREEING)) 357 goto skip_switch; 358 359 /* 360 * Count and transfer stats. Note that PAGECACHE_TAG_DIRTY points 361 * to possibly dirty pages while PAGECACHE_TAG_WRITEBACK points to 362 * pages actually under underwriteback. 363 */ 364 radix_tree_for_each_tagged(slot, &mapping->page_tree, &iter, 0, 365 PAGECACHE_TAG_DIRTY) { 366 struct page *page = radix_tree_deref_slot_protected(slot, 367 &mapping->tree_lock); 368 if (likely(page) && PageDirty(page)) { 369 __dec_wb_stat(old_wb, WB_RECLAIMABLE); 370 __inc_wb_stat(new_wb, WB_RECLAIMABLE); 371 } 372 } 373 374 radix_tree_for_each_tagged(slot, &mapping->page_tree, &iter, 0, 375 PAGECACHE_TAG_WRITEBACK) { 376 struct page *page = radix_tree_deref_slot_protected(slot, 377 &mapping->tree_lock); 378 if (likely(page)) { 379 WARN_ON_ONCE(!PageWriteback(page)); 380 __dec_wb_stat(old_wb, WB_WRITEBACK); 381 __inc_wb_stat(new_wb, WB_WRITEBACK); 382 } 383 } 384 385 wb_get(new_wb); 386 387 /* 388 * Transfer to @new_wb's IO list if necessary. The specific list 389 * @inode was on is ignored and the inode is put on ->b_dirty which 390 * is always correct including from ->b_dirty_time. The transfer 391 * preserves @inode->dirtied_when ordering. 392 */ 393 if (!list_empty(&inode->i_wb_list)) { 394 struct inode *pos; 395 396 inode_wb_list_del_locked(inode, old_wb); 397 inode->i_wb = new_wb; 398 list_for_each_entry(pos, &new_wb->b_dirty, i_wb_list) 399 if (time_after_eq(inode->dirtied_when, 400 pos->dirtied_when)) 401 break; 402 inode_wb_list_move_locked(inode, new_wb, pos->i_wb_list.prev); 403 } else { 404 inode->i_wb = new_wb; 405 } 406 407 /* ->i_wb_frn updates may race wbc_detach_inode() but doesn't matter */ 408 inode->i_wb_frn_winner = 0; 409 inode->i_wb_frn_avg_time = 0; 410 inode->i_wb_frn_history = 0; 411 switched = true; 412 skip_switch: 413 /* 414 * Paired with load_acquire in unlocked_inode_to_wb_begin() and 415 * ensures that the new wb is visible if they see !I_WB_SWITCH. 416 */ 417 smp_store_release(&inode->i_state, inode->i_state & ~I_WB_SWITCH); 418 419 spin_unlock_irq(&mapping->tree_lock); 420 spin_unlock(&inode->i_lock); 421 spin_unlock(&new_wb->list_lock); 422 spin_unlock(&old_wb->list_lock); 423 424 if (switched) { 425 wb_wakeup(new_wb); 426 wb_put(old_wb); 427 } 428 wb_put(new_wb); 429 430 iput(inode); 431 kfree(isw); 432 } 433 434 static void inode_switch_wbs_rcu_fn(struct rcu_head *rcu_head) 435 { 436 struct inode_switch_wbs_context *isw = container_of(rcu_head, 437 struct inode_switch_wbs_context, rcu_head); 438 439 /* needs to grab bh-unsafe locks, bounce to work item */ 440 INIT_WORK(&isw->work, inode_switch_wbs_work_fn); 441 schedule_work(&isw->work); 442 } 443 444 /** 445 * inode_switch_wbs - change the wb association of an inode 446 * @inode: target inode 447 * @new_wb_id: ID of the new wb 448 * 449 * Switch @inode's wb association to the wb identified by @new_wb_id. The 450 * switching is performed asynchronously and may fail silently. 451 */ 452 static void inode_switch_wbs(struct inode *inode, int new_wb_id) 453 { 454 struct backing_dev_info *bdi = inode_to_bdi(inode); 455 struct cgroup_subsys_state *memcg_css; 456 struct inode_switch_wbs_context *isw; 457 458 /* noop if seems to be already in progress */ 459 if (inode->i_state & I_WB_SWITCH) 460 return; 461 462 isw = kzalloc(sizeof(*isw), GFP_ATOMIC); 463 if (!isw) 464 return; 465 466 /* find and pin the new wb */ 467 rcu_read_lock(); 468 memcg_css = css_from_id(new_wb_id, &memory_cgrp_subsys); 469 if (memcg_css) 470 isw->new_wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC); 471 rcu_read_unlock(); 472 if (!isw->new_wb) 473 goto out_free; 474 475 /* while holding I_WB_SWITCH, no one else can update the association */ 476 spin_lock(&inode->i_lock); 477 if (inode->i_state & (I_WB_SWITCH | I_FREEING) || 478 inode_to_wb(inode) == isw->new_wb) { 479 spin_unlock(&inode->i_lock); 480 goto out_free; 481 } 482 inode->i_state |= I_WB_SWITCH; 483 spin_unlock(&inode->i_lock); 484 485 ihold(inode); 486 isw->inode = inode; 487 488 /* 489 * In addition to synchronizing among switchers, I_WB_SWITCH tells 490 * the RCU protected stat update paths to grab the mapping's 491 * tree_lock so that stat transfer can synchronize against them. 492 * Let's continue after I_WB_SWITCH is guaranteed to be visible. 493 */ 494 call_rcu(&isw->rcu_head, inode_switch_wbs_rcu_fn); 495 return; 496 497 out_free: 498 if (isw->new_wb) 499 wb_put(isw->new_wb); 500 kfree(isw); 501 } 502 503 /** 504 * wbc_attach_and_unlock_inode - associate wbc with target inode and unlock it 505 * @wbc: writeback_control of interest 506 * @inode: target inode 507 * 508 * @inode is locked and about to be written back under the control of @wbc. 509 * Record @inode's writeback context into @wbc and unlock the i_lock. On 510 * writeback completion, wbc_detach_inode() should be called. This is used 511 * to track the cgroup writeback context. 512 */ 513 void wbc_attach_and_unlock_inode(struct writeback_control *wbc, 514 struct inode *inode) 515 { 516 if (!inode_cgwb_enabled(inode)) { 517 spin_unlock(&inode->i_lock); 518 return; 519 } 520 521 wbc->wb = inode_to_wb(inode); 522 wbc->inode = inode; 523 524 wbc->wb_id = wbc->wb->memcg_css->id; 525 wbc->wb_lcand_id = inode->i_wb_frn_winner; 526 wbc->wb_tcand_id = 0; 527 wbc->wb_bytes = 0; 528 wbc->wb_lcand_bytes = 0; 529 wbc->wb_tcand_bytes = 0; 530 531 wb_get(wbc->wb); 532 spin_unlock(&inode->i_lock); 533 534 /* 535 * A dying wb indicates that the memcg-blkcg mapping has changed 536 * and a new wb is already serving the memcg. Switch immediately. 537 */ 538 if (unlikely(wb_dying(wbc->wb))) 539 inode_switch_wbs(inode, wbc->wb_id); 540 } 541 542 /** 543 * wbc_detach_inode - disassociate wbc from inode and perform foreign detection 544 * @wbc: writeback_control of the just finished writeback 545 * 546 * To be called after a writeback attempt of an inode finishes and undoes 547 * wbc_attach_and_unlock_inode(). Can be called under any context. 548 * 549 * As concurrent write sharing of an inode is expected to be very rare and 550 * memcg only tracks page ownership on first-use basis severely confining 551 * the usefulness of such sharing, cgroup writeback tracks ownership 552 * per-inode. While the support for concurrent write sharing of an inode 553 * is deemed unnecessary, an inode being written to by different cgroups at 554 * different points in time is a lot more common, and, more importantly, 555 * charging only by first-use can too readily lead to grossly incorrect 556 * behaviors (single foreign page can lead to gigabytes of writeback to be 557 * incorrectly attributed). 558 * 559 * To resolve this issue, cgroup writeback detects the majority dirtier of 560 * an inode and transfers the ownership to it. To avoid unnnecessary 561 * oscillation, the detection mechanism keeps track of history and gives 562 * out the switch verdict only if the foreign usage pattern is stable over 563 * a certain amount of time and/or writeback attempts. 564 * 565 * On each writeback attempt, @wbc tries to detect the majority writer 566 * using Boyer-Moore majority vote algorithm. In addition to the byte 567 * count from the majority voting, it also counts the bytes written for the 568 * current wb and the last round's winner wb (max of last round's current 569 * wb, the winner from two rounds ago, and the last round's majority 570 * candidate). Keeping track of the historical winner helps the algorithm 571 * to semi-reliably detect the most active writer even when it's not the 572 * absolute majority. 573 * 574 * Once the winner of the round is determined, whether the winner is 575 * foreign or not and how much IO time the round consumed is recorded in 576 * inode->i_wb_frn_history. If the amount of recorded foreign IO time is 577 * over a certain threshold, the switch verdict is given. 578 */ 579 void wbc_detach_inode(struct writeback_control *wbc) 580 { 581 struct bdi_writeback *wb = wbc->wb; 582 struct inode *inode = wbc->inode; 583 unsigned long avg_time, max_bytes, max_time; 584 u16 history; 585 int max_id; 586 587 if (!wb) 588 return; 589 590 history = inode->i_wb_frn_history; 591 avg_time = inode->i_wb_frn_avg_time; 592 593 /* pick the winner of this round */ 594 if (wbc->wb_bytes >= wbc->wb_lcand_bytes && 595 wbc->wb_bytes >= wbc->wb_tcand_bytes) { 596 max_id = wbc->wb_id; 597 max_bytes = wbc->wb_bytes; 598 } else if (wbc->wb_lcand_bytes >= wbc->wb_tcand_bytes) { 599 max_id = wbc->wb_lcand_id; 600 max_bytes = wbc->wb_lcand_bytes; 601 } else { 602 max_id = wbc->wb_tcand_id; 603 max_bytes = wbc->wb_tcand_bytes; 604 } 605 606 /* 607 * Calculate the amount of IO time the winner consumed and fold it 608 * into the running average kept per inode. If the consumed IO 609 * time is lower than avag / WB_FRN_TIME_CUT_DIV, ignore it for 610 * deciding whether to switch or not. This is to prevent one-off 611 * small dirtiers from skewing the verdict. 612 */ 613 max_time = DIV_ROUND_UP((max_bytes >> PAGE_SHIFT) << WB_FRN_TIME_SHIFT, 614 wb->avg_write_bandwidth); 615 if (avg_time) 616 avg_time += (max_time >> WB_FRN_TIME_AVG_SHIFT) - 617 (avg_time >> WB_FRN_TIME_AVG_SHIFT); 618 else 619 avg_time = max_time; /* immediate catch up on first run */ 620 621 if (max_time >= avg_time / WB_FRN_TIME_CUT_DIV) { 622 int slots; 623 624 /* 625 * The switch verdict is reached if foreign wb's consume 626 * more than a certain proportion of IO time in a 627 * WB_FRN_TIME_PERIOD. This is loosely tracked by 16 slot 628 * history mask where each bit represents one sixteenth of 629 * the period. Determine the number of slots to shift into 630 * history from @max_time. 631 */ 632 slots = min(DIV_ROUND_UP(max_time, WB_FRN_HIST_UNIT), 633 (unsigned long)WB_FRN_HIST_MAX_SLOTS); 634 history <<= slots; 635 if (wbc->wb_id != max_id) 636 history |= (1U << slots) - 1; 637 638 /* 639 * Switch if the current wb isn't the consistent winner. 640 * If there are multiple closely competing dirtiers, the 641 * inode may switch across them repeatedly over time, which 642 * is okay. The main goal is avoiding keeping an inode on 643 * the wrong wb for an extended period of time. 644 */ 645 if (hweight32(history) > WB_FRN_HIST_THR_SLOTS) 646 inode_switch_wbs(inode, max_id); 647 } 648 649 /* 650 * Multiple instances of this function may race to update the 651 * following fields but we don't mind occassional inaccuracies. 652 */ 653 inode->i_wb_frn_winner = max_id; 654 inode->i_wb_frn_avg_time = min(avg_time, (unsigned long)U16_MAX); 655 inode->i_wb_frn_history = history; 656 657 wb_put(wbc->wb); 658 wbc->wb = NULL; 659 } 660 661 /** 662 * wbc_account_io - account IO issued during writeback 663 * @wbc: writeback_control of the writeback in progress 664 * @page: page being written out 665 * @bytes: number of bytes being written out 666 * 667 * @bytes from @page are about to written out during the writeback 668 * controlled by @wbc. Keep the book for foreign inode detection. See 669 * wbc_detach_inode(). 670 */ 671 void wbc_account_io(struct writeback_control *wbc, struct page *page, 672 size_t bytes) 673 { 674 int id; 675 676 /* 677 * pageout() path doesn't attach @wbc to the inode being written 678 * out. This is intentional as we don't want the function to block 679 * behind a slow cgroup. Ultimately, we want pageout() to kick off 680 * regular writeback instead of writing things out itself. 681 */ 682 if (!wbc->wb) 683 return; 684 685 rcu_read_lock(); 686 id = mem_cgroup_css_from_page(page)->id; 687 rcu_read_unlock(); 688 689 if (id == wbc->wb_id) { 690 wbc->wb_bytes += bytes; 691 return; 692 } 693 694 if (id == wbc->wb_lcand_id) 695 wbc->wb_lcand_bytes += bytes; 696 697 /* Boyer-Moore majority vote algorithm */ 698 if (!wbc->wb_tcand_bytes) 699 wbc->wb_tcand_id = id; 700 if (id == wbc->wb_tcand_id) 701 wbc->wb_tcand_bytes += bytes; 702 else 703 wbc->wb_tcand_bytes -= min(bytes, wbc->wb_tcand_bytes); 704 } 705 706 /** 707 * inode_congested - test whether an inode is congested 708 * @inode: inode to test for congestion 709 * @cong_bits: mask of WB_[a]sync_congested bits to test 710 * 711 * Tests whether @inode is congested. @cong_bits is the mask of congestion 712 * bits to test and the return value is the mask of set bits. 713 * 714 * If cgroup writeback is enabled for @inode, the congestion state is 715 * determined by whether the cgwb (cgroup bdi_writeback) for the blkcg 716 * associated with @inode is congested; otherwise, the root wb's congestion 717 * state is used. 718 */ 719 int inode_congested(struct inode *inode, int cong_bits) 720 { 721 /* 722 * Once set, ->i_wb never becomes NULL while the inode is alive. 723 * Start transaction iff ->i_wb is visible. 724 */ 725 if (inode && inode_to_wb_is_valid(inode)) { 726 struct bdi_writeback *wb; 727 bool locked, congested; 728 729 wb = unlocked_inode_to_wb_begin(inode, &locked); 730 congested = wb_congested(wb, cong_bits); 731 unlocked_inode_to_wb_end(inode, locked); 732 return congested; 733 } 734 735 return wb_congested(&inode_to_bdi(inode)->wb, cong_bits); 736 } 737 EXPORT_SYMBOL_GPL(inode_congested); 738 739 /** 740 * wb_wait_for_single_work - wait for completion of a single bdi_writeback_work 741 * @bdi: bdi the work item was issued to 742 * @work: work item to wait for 743 * 744 * Wait for the completion of @work which was issued to one of @bdi's 745 * bdi_writeback's. The caller must have set @work->single_wait before 746 * issuing it. This wait operates independently fo 747 * wb_wait_for_completion() and also disables automatic freeing of @work. 748 */ 749 static void wb_wait_for_single_work(struct backing_dev_info *bdi, 750 struct wb_writeback_work *work) 751 { 752 if (WARN_ON_ONCE(!work->single_wait)) 753 return; 754 755 wait_event(bdi->wb_waitq, work->single_done); 756 757 /* 758 * Paired with smp_wmb() in wb_do_writeback() and ensures that all 759 * modifications to @work prior to assertion of ->single_done is 760 * visible to the caller once this function returns. 761 */ 762 smp_rmb(); 763 } 764 765 /** 766 * wb_split_bdi_pages - split nr_pages to write according to bandwidth 767 * @wb: target bdi_writeback to split @nr_pages to 768 * @nr_pages: number of pages to write for the whole bdi 769 * 770 * Split @wb's portion of @nr_pages according to @wb's write bandwidth in 771 * relation to the total write bandwidth of all wb's w/ dirty inodes on 772 * @wb->bdi. 773 */ 774 static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages) 775 { 776 unsigned long this_bw = wb->avg_write_bandwidth; 777 unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth); 778 779 if (nr_pages == LONG_MAX) 780 return LONG_MAX; 781 782 /* 783 * This may be called on clean wb's and proportional distribution 784 * may not make sense, just use the original @nr_pages in those 785 * cases. In general, we wanna err on the side of writing more. 786 */ 787 if (!tot_bw || this_bw >= tot_bw) 788 return nr_pages; 789 else 790 return DIV_ROUND_UP_ULL((u64)nr_pages * this_bw, tot_bw); 791 } 792 793 /** 794 * wb_clone_and_queue_work - clone a wb_writeback_work and issue it to a wb 795 * @wb: target bdi_writeback 796 * @base_work: source wb_writeback_work 797 * 798 * Try to make a clone of @base_work and issue it to @wb. If cloning 799 * succeeds, %true is returned; otherwise, @base_work is issued directly 800 * and %false is returned. In the latter case, the caller is required to 801 * wait for @base_work's completion using wb_wait_for_single_work(). 802 * 803 * A clone is auto-freed on completion. @base_work never is. 804 */ 805 static bool wb_clone_and_queue_work(struct bdi_writeback *wb, 806 struct wb_writeback_work *base_work) 807 { 808 struct wb_writeback_work *work; 809 810 work = kmalloc(sizeof(*work), GFP_ATOMIC); 811 if (work) { 812 *work = *base_work; 813 work->auto_free = 1; 814 work->single_wait = 0; 815 } else { 816 work = base_work; 817 work->auto_free = 0; 818 work->single_wait = 1; 819 } 820 work->single_done = 0; 821 wb_queue_work(wb, work); 822 return work != base_work; 823 } 824 825 /** 826 * bdi_split_work_to_wbs - split a wb_writeback_work to all wb's of a bdi 827 * @bdi: target backing_dev_info 828 * @base_work: wb_writeback_work to issue 829 * @skip_if_busy: skip wb's which already have writeback in progress 830 * 831 * Split and issue @base_work to all wb's (bdi_writeback's) of @bdi which 832 * have dirty inodes. If @base_work->nr_page isn't %LONG_MAX, it's 833 * distributed to the busy wbs according to each wb's proportion in the 834 * total active write bandwidth of @bdi. 835 */ 836 static void bdi_split_work_to_wbs(struct backing_dev_info *bdi, 837 struct wb_writeback_work *base_work, 838 bool skip_if_busy) 839 { 840 long nr_pages = base_work->nr_pages; 841 int next_blkcg_id = 0; 842 struct bdi_writeback *wb; 843 struct wb_iter iter; 844 845 might_sleep(); 846 847 if (!bdi_has_dirty_io(bdi)) 848 return; 849 restart: 850 rcu_read_lock(); 851 bdi_for_each_wb(wb, bdi, &iter, next_blkcg_id) { 852 if (!wb_has_dirty_io(wb) || 853 (skip_if_busy && writeback_in_progress(wb))) 854 continue; 855 856 base_work->nr_pages = wb_split_bdi_pages(wb, nr_pages); 857 if (!wb_clone_and_queue_work(wb, base_work)) { 858 next_blkcg_id = wb->blkcg_css->id + 1; 859 rcu_read_unlock(); 860 wb_wait_for_single_work(bdi, base_work); 861 goto restart; 862 } 863 } 864 rcu_read_unlock(); 865 } 866 867 #else /* CONFIG_CGROUP_WRITEBACK */ 868 869 static struct bdi_writeback * 870 locked_inode_to_wb_and_lock_list(struct inode *inode) 871 __releases(&inode->i_lock) 872 __acquires(&wb->list_lock) 873 { 874 struct bdi_writeback *wb = inode_to_wb(inode); 875 876 spin_unlock(&inode->i_lock); 877 spin_lock(&wb->list_lock); 878 return wb; 879 } 880 881 static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode) 882 __acquires(&wb->list_lock) 883 { 884 struct bdi_writeback *wb = inode_to_wb(inode); 885 886 spin_lock(&wb->list_lock); 887 return wb; 888 } 889 890 static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages) 891 { 892 return nr_pages; 893 } 894 895 static void bdi_split_work_to_wbs(struct backing_dev_info *bdi, 896 struct wb_writeback_work *base_work, 897 bool skip_if_busy) 898 { 899 might_sleep(); 900 901 if (bdi_has_dirty_io(bdi) && 902 (!skip_if_busy || !writeback_in_progress(&bdi->wb))) { 903 base_work->auto_free = 0; 904 base_work->single_wait = 0; 905 base_work->single_done = 0; 906 wb_queue_work(&bdi->wb, base_work); 907 } 908 } 909 910 #endif /* CONFIG_CGROUP_WRITEBACK */ 911 912 void wb_start_writeback(struct bdi_writeback *wb, long nr_pages, 913 bool range_cyclic, enum wb_reason reason) 914 { 915 struct wb_writeback_work *work; 916 917 if (!wb_has_dirty_io(wb)) 918 return; 919 920 /* 921 * This is WB_SYNC_NONE writeback, so if allocation fails just 922 * wakeup the thread for old dirty data writeback 923 */ 924 work = kzalloc(sizeof(*work), GFP_ATOMIC); 925 if (!work) { 926 trace_writeback_nowork(wb->bdi); 927 wb_wakeup(wb); 928 return; 929 } 930 931 work->sync_mode = WB_SYNC_NONE; 932 work->nr_pages = nr_pages; 933 work->range_cyclic = range_cyclic; 934 work->reason = reason; 935 work->auto_free = 1; 936 937 wb_queue_work(wb, work); 938 } 939 940 /** 941 * wb_start_background_writeback - start background writeback 942 * @wb: bdi_writback to write from 943 * 944 * Description: 945 * This makes sure WB_SYNC_NONE background writeback happens. When 946 * this function returns, it is only guaranteed that for given wb 947 * some IO is happening if we are over background dirty threshold. 948 * Caller need not hold sb s_umount semaphore. 949 */ 950 void wb_start_background_writeback(struct bdi_writeback *wb) 951 { 952 /* 953 * We just wake up the flusher thread. It will perform background 954 * writeback as soon as there is no other work to do. 955 */ 956 trace_writeback_wake_background(wb->bdi); 957 wb_wakeup(wb); 958 } 959 960 /* 961 * Remove the inode from the writeback list it is on. 962 */ 963 void inode_wb_list_del(struct inode *inode) 964 { 965 struct bdi_writeback *wb; 966 967 wb = inode_to_wb_and_lock_list(inode); 968 inode_wb_list_del_locked(inode, wb); 969 spin_unlock(&wb->list_lock); 970 } 971 972 /* 973 * Redirty an inode: set its when-it-was dirtied timestamp and move it to the 974 * furthest end of its superblock's dirty-inode list. 975 * 976 * Before stamping the inode's ->dirtied_when, we check to see whether it is 977 * already the most-recently-dirtied inode on the b_dirty list. If that is 978 * the case then the inode must have been redirtied while it was being written 979 * out and we don't reset its dirtied_when. 980 */ 981 static void redirty_tail(struct inode *inode, struct bdi_writeback *wb) 982 { 983 if (!list_empty(&wb->b_dirty)) { 984 struct inode *tail; 985 986 tail = wb_inode(wb->b_dirty.next); 987 if (time_before(inode->dirtied_when, tail->dirtied_when)) 988 inode->dirtied_when = jiffies; 989 } 990 inode_wb_list_move_locked(inode, wb, &wb->b_dirty); 991 } 992 993 /* 994 * requeue inode for re-scanning after bdi->b_io list is exhausted. 995 */ 996 static void requeue_io(struct inode *inode, struct bdi_writeback *wb) 997 { 998 inode_wb_list_move_locked(inode, wb, &wb->b_more_io); 999 } 1000 1001 static void inode_sync_complete(struct inode *inode) 1002 { 1003 inode->i_state &= ~I_SYNC; 1004 /* If inode is clean an unused, put it into LRU now... */ 1005 inode_add_lru(inode); 1006 /* Waiters must see I_SYNC cleared before being woken up */ 1007 smp_mb(); 1008 wake_up_bit(&inode->i_state, __I_SYNC); 1009 } 1010 1011 static bool inode_dirtied_after(struct inode *inode, unsigned long t) 1012 { 1013 bool ret = time_after(inode->dirtied_when, t); 1014 #ifndef CONFIG_64BIT 1015 /* 1016 * For inodes being constantly redirtied, dirtied_when can get stuck. 1017 * It _appears_ to be in the future, but is actually in distant past. 1018 * This test is necessary to prevent such wrapped-around relative times 1019 * from permanently stopping the whole bdi writeback. 1020 */ 1021 ret = ret && time_before_eq(inode->dirtied_when, jiffies); 1022 #endif 1023 return ret; 1024 } 1025 1026 #define EXPIRE_DIRTY_ATIME 0x0001 1027 1028 /* 1029 * Move expired (dirtied before work->older_than_this) dirty inodes from 1030 * @delaying_queue to @dispatch_queue. 1031 */ 1032 static int move_expired_inodes(struct list_head *delaying_queue, 1033 struct list_head *dispatch_queue, 1034 int flags, 1035 struct wb_writeback_work *work) 1036 { 1037 unsigned long *older_than_this = NULL; 1038 unsigned long expire_time; 1039 LIST_HEAD(tmp); 1040 struct list_head *pos, *node; 1041 struct super_block *sb = NULL; 1042 struct inode *inode; 1043 int do_sb_sort = 0; 1044 int moved = 0; 1045 1046 if ((flags & EXPIRE_DIRTY_ATIME) == 0) 1047 older_than_this = work->older_than_this; 1048 else if (!work->for_sync) { 1049 expire_time = jiffies - (dirtytime_expire_interval * HZ); 1050 older_than_this = &expire_time; 1051 } 1052 while (!list_empty(delaying_queue)) { 1053 inode = wb_inode(delaying_queue->prev); 1054 if (older_than_this && 1055 inode_dirtied_after(inode, *older_than_this)) 1056 break; 1057 list_move(&inode->i_wb_list, &tmp); 1058 moved++; 1059 if (flags & EXPIRE_DIRTY_ATIME) 1060 set_bit(__I_DIRTY_TIME_EXPIRED, &inode->i_state); 1061 if (sb_is_blkdev_sb(inode->i_sb)) 1062 continue; 1063 if (sb && sb != inode->i_sb) 1064 do_sb_sort = 1; 1065 sb = inode->i_sb; 1066 } 1067 1068 /* just one sb in list, splice to dispatch_queue and we're done */ 1069 if (!do_sb_sort) { 1070 list_splice(&tmp, dispatch_queue); 1071 goto out; 1072 } 1073 1074 /* Move inodes from one superblock together */ 1075 while (!list_empty(&tmp)) { 1076 sb = wb_inode(tmp.prev)->i_sb; 1077 list_for_each_prev_safe(pos, node, &tmp) { 1078 inode = wb_inode(pos); 1079 if (inode->i_sb == sb) 1080 list_move(&inode->i_wb_list, dispatch_queue); 1081 } 1082 } 1083 out: 1084 return moved; 1085 } 1086 1087 /* 1088 * Queue all expired dirty inodes for io, eldest first. 1089 * Before 1090 * newly dirtied b_dirty b_io b_more_io 1091 * =============> gf edc BA 1092 * After 1093 * newly dirtied b_dirty b_io b_more_io 1094 * =============> g fBAedc 1095 * | 1096 * +--> dequeue for IO 1097 */ 1098 static void queue_io(struct bdi_writeback *wb, struct wb_writeback_work *work) 1099 { 1100 int moved; 1101 1102 assert_spin_locked(&wb->list_lock); 1103 list_splice_init(&wb->b_more_io, &wb->b_io); 1104 moved = move_expired_inodes(&wb->b_dirty, &wb->b_io, 0, work); 1105 moved += move_expired_inodes(&wb->b_dirty_time, &wb->b_io, 1106 EXPIRE_DIRTY_ATIME, work); 1107 if (moved) 1108 wb_io_lists_populated(wb); 1109 trace_writeback_queue_io(wb, work, moved); 1110 } 1111 1112 static int write_inode(struct inode *inode, struct writeback_control *wbc) 1113 { 1114 int ret; 1115 1116 if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode)) { 1117 trace_writeback_write_inode_start(inode, wbc); 1118 ret = inode->i_sb->s_op->write_inode(inode, wbc); 1119 trace_writeback_write_inode(inode, wbc); 1120 return ret; 1121 } 1122 return 0; 1123 } 1124 1125 /* 1126 * Wait for writeback on an inode to complete. Called with i_lock held. 1127 * Caller must make sure inode cannot go away when we drop i_lock. 1128 */ 1129 static void __inode_wait_for_writeback(struct inode *inode) 1130 __releases(inode->i_lock) 1131 __acquires(inode->i_lock) 1132 { 1133 DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC); 1134 wait_queue_head_t *wqh; 1135 1136 wqh = bit_waitqueue(&inode->i_state, __I_SYNC); 1137 while (inode->i_state & I_SYNC) { 1138 spin_unlock(&inode->i_lock); 1139 __wait_on_bit(wqh, &wq, bit_wait, 1140 TASK_UNINTERRUPTIBLE); 1141 spin_lock(&inode->i_lock); 1142 } 1143 } 1144 1145 /* 1146 * Wait for writeback on an inode to complete. Caller must have inode pinned. 1147 */ 1148 void inode_wait_for_writeback(struct inode *inode) 1149 { 1150 spin_lock(&inode->i_lock); 1151 __inode_wait_for_writeback(inode); 1152 spin_unlock(&inode->i_lock); 1153 } 1154 1155 /* 1156 * Sleep until I_SYNC is cleared. This function must be called with i_lock 1157 * held and drops it. It is aimed for callers not holding any inode reference 1158 * so once i_lock is dropped, inode can go away. 1159 */ 1160 static void inode_sleep_on_writeback(struct inode *inode) 1161 __releases(inode->i_lock) 1162 { 1163 DEFINE_WAIT(wait); 1164 wait_queue_head_t *wqh = bit_waitqueue(&inode->i_state, __I_SYNC); 1165 int sleep; 1166 1167 prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE); 1168 sleep = inode->i_state & I_SYNC; 1169 spin_unlock(&inode->i_lock); 1170 if (sleep) 1171 schedule(); 1172 finish_wait(wqh, &wait); 1173 } 1174 1175 /* 1176 * Find proper writeback list for the inode depending on its current state and 1177 * possibly also change of its state while we were doing writeback. Here we 1178 * handle things such as livelock prevention or fairness of writeback among 1179 * inodes. This function can be called only by flusher thread - noone else 1180 * processes all inodes in writeback lists and requeueing inodes behind flusher 1181 * thread's back can have unexpected consequences. 1182 */ 1183 static void requeue_inode(struct inode *inode, struct bdi_writeback *wb, 1184 struct writeback_control *wbc) 1185 { 1186 if (inode->i_state & I_FREEING) 1187 return; 1188 1189 /* 1190 * Sync livelock prevention. Each inode is tagged and synced in one 1191 * shot. If still dirty, it will be redirty_tail()'ed below. Update 1192 * the dirty time to prevent enqueue and sync it again. 1193 */ 1194 if ((inode->i_state & I_DIRTY) && 1195 (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)) 1196 inode->dirtied_when = jiffies; 1197 1198 if (wbc->pages_skipped) { 1199 /* 1200 * writeback is not making progress due to locked 1201 * buffers. Skip this inode for now. 1202 */ 1203 redirty_tail(inode, wb); 1204 return; 1205 } 1206 1207 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_DIRTY)) { 1208 /* 1209 * We didn't write back all the pages. nfs_writepages() 1210 * sometimes bales out without doing anything. 1211 */ 1212 if (wbc->nr_to_write <= 0) { 1213 /* Slice used up. Queue for next turn. */ 1214 requeue_io(inode, wb); 1215 } else { 1216 /* 1217 * Writeback blocked by something other than 1218 * congestion. Delay the inode for some time to 1219 * avoid spinning on the CPU (100% iowait) 1220 * retrying writeback of the dirty page/inode 1221 * that cannot be performed immediately. 1222 */ 1223 redirty_tail(inode, wb); 1224 } 1225 } else if (inode->i_state & I_DIRTY) { 1226 /* 1227 * Filesystems can dirty the inode during writeback operations, 1228 * such as delayed allocation during submission or metadata 1229 * updates after data IO completion. 1230 */ 1231 redirty_tail(inode, wb); 1232 } else if (inode->i_state & I_DIRTY_TIME) { 1233 inode->dirtied_when = jiffies; 1234 inode_wb_list_move_locked(inode, wb, &wb->b_dirty_time); 1235 } else { 1236 /* The inode is clean. Remove from writeback lists. */ 1237 inode_wb_list_del_locked(inode, wb); 1238 } 1239 } 1240 1241 /* 1242 * Write out an inode and its dirty pages. Do not update the writeback list 1243 * linkage. That is left to the caller. The caller is also responsible for 1244 * setting I_SYNC flag and calling inode_sync_complete() to clear it. 1245 */ 1246 static int 1247 __writeback_single_inode(struct inode *inode, struct writeback_control *wbc) 1248 { 1249 struct address_space *mapping = inode->i_mapping; 1250 long nr_to_write = wbc->nr_to_write; 1251 unsigned dirty; 1252 int ret; 1253 1254 WARN_ON(!(inode->i_state & I_SYNC)); 1255 1256 trace_writeback_single_inode_start(inode, wbc, nr_to_write); 1257 1258 ret = do_writepages(mapping, wbc); 1259 1260 /* 1261 * Make sure to wait on the data before writing out the metadata. 1262 * This is important for filesystems that modify metadata on data 1263 * I/O completion. We don't do it for sync(2) writeback because it has a 1264 * separate, external IO completion path and ->sync_fs for guaranteeing 1265 * inode metadata is written back correctly. 1266 */ 1267 if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync) { 1268 int err = filemap_fdatawait(mapping); 1269 if (ret == 0) 1270 ret = err; 1271 } 1272 1273 /* 1274 * Some filesystems may redirty the inode during the writeback 1275 * due to delalloc, clear dirty metadata flags right before 1276 * write_inode() 1277 */ 1278 spin_lock(&inode->i_lock); 1279 1280 dirty = inode->i_state & I_DIRTY; 1281 if (inode->i_state & I_DIRTY_TIME) { 1282 if ((dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) || 1283 unlikely(inode->i_state & I_DIRTY_TIME_EXPIRED) || 1284 unlikely(time_after(jiffies, 1285 (inode->dirtied_time_when + 1286 dirtytime_expire_interval * HZ)))) { 1287 dirty |= I_DIRTY_TIME | I_DIRTY_TIME_EXPIRED; 1288 trace_writeback_lazytime(inode); 1289 } 1290 } else 1291 inode->i_state &= ~I_DIRTY_TIME_EXPIRED; 1292 inode->i_state &= ~dirty; 1293 1294 /* 1295 * Paired with smp_mb() in __mark_inode_dirty(). This allows 1296 * __mark_inode_dirty() to test i_state without grabbing i_lock - 1297 * either they see the I_DIRTY bits cleared or we see the dirtied 1298 * inode. 1299 * 1300 * I_DIRTY_PAGES is always cleared together above even if @mapping 1301 * still has dirty pages. The flag is reinstated after smp_mb() if 1302 * necessary. This guarantees that either __mark_inode_dirty() 1303 * sees clear I_DIRTY_PAGES or we see PAGECACHE_TAG_DIRTY. 1304 */ 1305 smp_mb(); 1306 1307 if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) 1308 inode->i_state |= I_DIRTY_PAGES; 1309 1310 spin_unlock(&inode->i_lock); 1311 1312 if (dirty & I_DIRTY_TIME) 1313 mark_inode_dirty_sync(inode); 1314 /* Don't write the inode if only I_DIRTY_PAGES was set */ 1315 if (dirty & ~I_DIRTY_PAGES) { 1316 int err = write_inode(inode, wbc); 1317 if (ret == 0) 1318 ret = err; 1319 } 1320 trace_writeback_single_inode(inode, wbc, nr_to_write); 1321 return ret; 1322 } 1323 1324 /* 1325 * Write out an inode's dirty pages. Either the caller has an active reference 1326 * on the inode or the inode has I_WILL_FREE set. 1327 * 1328 * This function is designed to be called for writing back one inode which 1329 * we go e.g. from filesystem. Flusher thread uses __writeback_single_inode() 1330 * and does more profound writeback list handling in writeback_sb_inodes(). 1331 */ 1332 static int 1333 writeback_single_inode(struct inode *inode, struct bdi_writeback *wb, 1334 struct writeback_control *wbc) 1335 { 1336 int ret = 0; 1337 1338 spin_lock(&inode->i_lock); 1339 if (!atomic_read(&inode->i_count)) 1340 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING))); 1341 else 1342 WARN_ON(inode->i_state & I_WILL_FREE); 1343 1344 if (inode->i_state & I_SYNC) { 1345 if (wbc->sync_mode != WB_SYNC_ALL) 1346 goto out; 1347 /* 1348 * It's a data-integrity sync. We must wait. Since callers hold 1349 * inode reference or inode has I_WILL_FREE set, it cannot go 1350 * away under us. 1351 */ 1352 __inode_wait_for_writeback(inode); 1353 } 1354 WARN_ON(inode->i_state & I_SYNC); 1355 /* 1356 * Skip inode if it is clean and we have no outstanding writeback in 1357 * WB_SYNC_ALL mode. We don't want to mess with writeback lists in this 1358 * function since flusher thread may be doing for example sync in 1359 * parallel and if we move the inode, it could get skipped. So here we 1360 * make sure inode is on some writeback list and leave it there unless 1361 * we have completely cleaned the inode. 1362 */ 1363 if (!(inode->i_state & I_DIRTY_ALL) && 1364 (wbc->sync_mode != WB_SYNC_ALL || 1365 !mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK))) 1366 goto out; 1367 inode->i_state |= I_SYNC; 1368 wbc_attach_and_unlock_inode(wbc, inode); 1369 1370 ret = __writeback_single_inode(inode, wbc); 1371 1372 wbc_detach_inode(wbc); 1373 spin_lock(&wb->list_lock); 1374 spin_lock(&inode->i_lock); 1375 /* 1376 * If inode is clean, remove it from writeback lists. Otherwise don't 1377 * touch it. See comment above for explanation. 1378 */ 1379 if (!(inode->i_state & I_DIRTY_ALL)) 1380 inode_wb_list_del_locked(inode, wb); 1381 spin_unlock(&wb->list_lock); 1382 inode_sync_complete(inode); 1383 out: 1384 spin_unlock(&inode->i_lock); 1385 return ret; 1386 } 1387 1388 static long writeback_chunk_size(struct bdi_writeback *wb, 1389 struct wb_writeback_work *work) 1390 { 1391 long pages; 1392 1393 /* 1394 * WB_SYNC_ALL mode does livelock avoidance by syncing dirty 1395 * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX 1396 * here avoids calling into writeback_inodes_wb() more than once. 1397 * 1398 * The intended call sequence for WB_SYNC_ALL writeback is: 1399 * 1400 * wb_writeback() 1401 * writeback_sb_inodes() <== called only once 1402 * write_cache_pages() <== called once for each inode 1403 * (quickly) tag currently dirty pages 1404 * (maybe slowly) sync all tagged pages 1405 */ 1406 if (work->sync_mode == WB_SYNC_ALL || work->tagged_writepages) 1407 pages = LONG_MAX; 1408 else { 1409 pages = min(wb->avg_write_bandwidth / 2, 1410 global_wb_domain.dirty_limit / DIRTY_SCOPE); 1411 pages = min(pages, work->nr_pages); 1412 pages = round_down(pages + MIN_WRITEBACK_PAGES, 1413 MIN_WRITEBACK_PAGES); 1414 } 1415 1416 return pages; 1417 } 1418 1419 /* 1420 * Write a portion of b_io inodes which belong to @sb. 1421 * 1422 * Return the number of pages and/or inodes written. 1423 */ 1424 static long writeback_sb_inodes(struct super_block *sb, 1425 struct bdi_writeback *wb, 1426 struct wb_writeback_work *work) 1427 { 1428 struct writeback_control wbc = { 1429 .sync_mode = work->sync_mode, 1430 .tagged_writepages = work->tagged_writepages, 1431 .for_kupdate = work->for_kupdate, 1432 .for_background = work->for_background, 1433 .for_sync = work->for_sync, 1434 .range_cyclic = work->range_cyclic, 1435 .range_start = 0, 1436 .range_end = LLONG_MAX, 1437 }; 1438 unsigned long start_time = jiffies; 1439 long write_chunk; 1440 long wrote = 0; /* count both pages and inodes */ 1441 1442 while (!list_empty(&wb->b_io)) { 1443 struct inode *inode = wb_inode(wb->b_io.prev); 1444 1445 if (inode->i_sb != sb) { 1446 if (work->sb) { 1447 /* 1448 * We only want to write back data for this 1449 * superblock, move all inodes not belonging 1450 * to it back onto the dirty list. 1451 */ 1452 redirty_tail(inode, wb); 1453 continue; 1454 } 1455 1456 /* 1457 * The inode belongs to a different superblock. 1458 * Bounce back to the caller to unpin this and 1459 * pin the next superblock. 1460 */ 1461 break; 1462 } 1463 1464 /* 1465 * Don't bother with new inodes or inodes being freed, first 1466 * kind does not need periodic writeout yet, and for the latter 1467 * kind writeout is handled by the freer. 1468 */ 1469 spin_lock(&inode->i_lock); 1470 if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) { 1471 spin_unlock(&inode->i_lock); 1472 redirty_tail(inode, wb); 1473 continue; 1474 } 1475 if ((inode->i_state & I_SYNC) && wbc.sync_mode != WB_SYNC_ALL) { 1476 /* 1477 * If this inode is locked for writeback and we are not 1478 * doing writeback-for-data-integrity, move it to 1479 * b_more_io so that writeback can proceed with the 1480 * other inodes on s_io. 1481 * 1482 * We'll have another go at writing back this inode 1483 * when we completed a full scan of b_io. 1484 */ 1485 spin_unlock(&inode->i_lock); 1486 requeue_io(inode, wb); 1487 trace_writeback_sb_inodes_requeue(inode); 1488 continue; 1489 } 1490 spin_unlock(&wb->list_lock); 1491 1492 /* 1493 * We already requeued the inode if it had I_SYNC set and we 1494 * are doing WB_SYNC_NONE writeback. So this catches only the 1495 * WB_SYNC_ALL case. 1496 */ 1497 if (inode->i_state & I_SYNC) { 1498 /* Wait for I_SYNC. This function drops i_lock... */ 1499 inode_sleep_on_writeback(inode); 1500 /* Inode may be gone, start again */ 1501 spin_lock(&wb->list_lock); 1502 continue; 1503 } 1504 inode->i_state |= I_SYNC; 1505 wbc_attach_and_unlock_inode(&wbc, inode); 1506 1507 write_chunk = writeback_chunk_size(wb, work); 1508 wbc.nr_to_write = write_chunk; 1509 wbc.pages_skipped = 0; 1510 1511 /* 1512 * We use I_SYNC to pin the inode in memory. While it is set 1513 * evict_inode() will wait so the inode cannot be freed. 1514 */ 1515 __writeback_single_inode(inode, &wbc); 1516 1517 wbc_detach_inode(&wbc); 1518 work->nr_pages -= write_chunk - wbc.nr_to_write; 1519 wrote += write_chunk - wbc.nr_to_write; 1520 spin_lock(&wb->list_lock); 1521 spin_lock(&inode->i_lock); 1522 if (!(inode->i_state & I_DIRTY_ALL)) 1523 wrote++; 1524 requeue_inode(inode, wb, &wbc); 1525 inode_sync_complete(inode); 1526 spin_unlock(&inode->i_lock); 1527 cond_resched_lock(&wb->list_lock); 1528 /* 1529 * bail out to wb_writeback() often enough to check 1530 * background threshold and other termination conditions. 1531 */ 1532 if (wrote) { 1533 if (time_is_before_jiffies(start_time + HZ / 10UL)) 1534 break; 1535 if (work->nr_pages <= 0) 1536 break; 1537 } 1538 } 1539 return wrote; 1540 } 1541 1542 static long __writeback_inodes_wb(struct bdi_writeback *wb, 1543 struct wb_writeback_work *work) 1544 { 1545 unsigned long start_time = jiffies; 1546 long wrote = 0; 1547 1548 while (!list_empty(&wb->b_io)) { 1549 struct inode *inode = wb_inode(wb->b_io.prev); 1550 struct super_block *sb = inode->i_sb; 1551 1552 if (!trylock_super(sb)) { 1553 /* 1554 * trylock_super() may fail consistently due to 1555 * s_umount being grabbed by someone else. Don't use 1556 * requeue_io() to avoid busy retrying the inode/sb. 1557 */ 1558 redirty_tail(inode, wb); 1559 continue; 1560 } 1561 wrote += writeback_sb_inodes(sb, wb, work); 1562 up_read(&sb->s_umount); 1563 1564 /* refer to the same tests at the end of writeback_sb_inodes */ 1565 if (wrote) { 1566 if (time_is_before_jiffies(start_time + HZ / 10UL)) 1567 break; 1568 if (work->nr_pages <= 0) 1569 break; 1570 } 1571 } 1572 /* Leave any unwritten inodes on b_io */ 1573 return wrote; 1574 } 1575 1576 static long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages, 1577 enum wb_reason reason) 1578 { 1579 struct wb_writeback_work work = { 1580 .nr_pages = nr_pages, 1581 .sync_mode = WB_SYNC_NONE, 1582 .range_cyclic = 1, 1583 .reason = reason, 1584 }; 1585 1586 spin_lock(&wb->list_lock); 1587 if (list_empty(&wb->b_io)) 1588 queue_io(wb, &work); 1589 __writeback_inodes_wb(wb, &work); 1590 spin_unlock(&wb->list_lock); 1591 1592 return nr_pages - work.nr_pages; 1593 } 1594 1595 /* 1596 * Explicit flushing or periodic writeback of "old" data. 1597 * 1598 * Define "old": the first time one of an inode's pages is dirtied, we mark the 1599 * dirtying-time in the inode's address_space. So this periodic writeback code 1600 * just walks the superblock inode list, writing back any inodes which are 1601 * older than a specific point in time. 1602 * 1603 * Try to run once per dirty_writeback_interval. But if a writeback event 1604 * takes longer than a dirty_writeback_interval interval, then leave a 1605 * one-second gap. 1606 * 1607 * older_than_this takes precedence over nr_to_write. So we'll only write back 1608 * all dirty pages if they are all attached to "old" mappings. 1609 */ 1610 static long wb_writeback(struct bdi_writeback *wb, 1611 struct wb_writeback_work *work) 1612 { 1613 unsigned long wb_start = jiffies; 1614 long nr_pages = work->nr_pages; 1615 unsigned long oldest_jif; 1616 struct inode *inode; 1617 long progress; 1618 1619 oldest_jif = jiffies; 1620 work->older_than_this = &oldest_jif; 1621 1622 spin_lock(&wb->list_lock); 1623 for (;;) { 1624 /* 1625 * Stop writeback when nr_pages has been consumed 1626 */ 1627 if (work->nr_pages <= 0) 1628 break; 1629 1630 /* 1631 * Background writeout and kupdate-style writeback may 1632 * run forever. Stop them if there is other work to do 1633 * so that e.g. sync can proceed. They'll be restarted 1634 * after the other works are all done. 1635 */ 1636 if ((work->for_background || work->for_kupdate) && 1637 !list_empty(&wb->work_list)) 1638 break; 1639 1640 /* 1641 * For background writeout, stop when we are below the 1642 * background dirty threshold 1643 */ 1644 if (work->for_background && !wb_over_bg_thresh(wb)) 1645 break; 1646 1647 /* 1648 * Kupdate and background works are special and we want to 1649 * include all inodes that need writing. Livelock avoidance is 1650 * handled by these works yielding to any other work so we are 1651 * safe. 1652 */ 1653 if (work->for_kupdate) { 1654 oldest_jif = jiffies - 1655 msecs_to_jiffies(dirty_expire_interval * 10); 1656 } else if (work->for_background) 1657 oldest_jif = jiffies; 1658 1659 trace_writeback_start(wb->bdi, work); 1660 if (list_empty(&wb->b_io)) 1661 queue_io(wb, work); 1662 if (work->sb) 1663 progress = writeback_sb_inodes(work->sb, wb, work); 1664 else 1665 progress = __writeback_inodes_wb(wb, work); 1666 trace_writeback_written(wb->bdi, work); 1667 1668 wb_update_bandwidth(wb, wb_start); 1669 1670 /* 1671 * Did we write something? Try for more 1672 * 1673 * Dirty inodes are moved to b_io for writeback in batches. 1674 * The completion of the current batch does not necessarily 1675 * mean the overall work is done. So we keep looping as long 1676 * as made some progress on cleaning pages or inodes. 1677 */ 1678 if (progress) 1679 continue; 1680 /* 1681 * No more inodes for IO, bail 1682 */ 1683 if (list_empty(&wb->b_more_io)) 1684 break; 1685 /* 1686 * Nothing written. Wait for some inode to 1687 * become available for writeback. Otherwise 1688 * we'll just busyloop. 1689 */ 1690 if (!list_empty(&wb->b_more_io)) { 1691 trace_writeback_wait(wb->bdi, work); 1692 inode = wb_inode(wb->b_more_io.prev); 1693 spin_lock(&inode->i_lock); 1694 spin_unlock(&wb->list_lock); 1695 /* This function drops i_lock... */ 1696 inode_sleep_on_writeback(inode); 1697 spin_lock(&wb->list_lock); 1698 } 1699 } 1700 spin_unlock(&wb->list_lock); 1701 1702 return nr_pages - work->nr_pages; 1703 } 1704 1705 /* 1706 * Return the next wb_writeback_work struct that hasn't been processed yet. 1707 */ 1708 static struct wb_writeback_work *get_next_work_item(struct bdi_writeback *wb) 1709 { 1710 struct wb_writeback_work *work = NULL; 1711 1712 spin_lock_bh(&wb->work_lock); 1713 if (!list_empty(&wb->work_list)) { 1714 work = list_entry(wb->work_list.next, 1715 struct wb_writeback_work, list); 1716 list_del_init(&work->list); 1717 } 1718 spin_unlock_bh(&wb->work_lock); 1719 return work; 1720 } 1721 1722 /* 1723 * Add in the number of potentially dirty inodes, because each inode 1724 * write can dirty pagecache in the underlying blockdev. 1725 */ 1726 static unsigned long get_nr_dirty_pages(void) 1727 { 1728 return global_page_state(NR_FILE_DIRTY) + 1729 global_page_state(NR_UNSTABLE_NFS) + 1730 get_nr_dirty_inodes(); 1731 } 1732 1733 static long wb_check_background_flush(struct bdi_writeback *wb) 1734 { 1735 if (wb_over_bg_thresh(wb)) { 1736 1737 struct wb_writeback_work work = { 1738 .nr_pages = LONG_MAX, 1739 .sync_mode = WB_SYNC_NONE, 1740 .for_background = 1, 1741 .range_cyclic = 1, 1742 .reason = WB_REASON_BACKGROUND, 1743 }; 1744 1745 return wb_writeback(wb, &work); 1746 } 1747 1748 return 0; 1749 } 1750 1751 static long wb_check_old_data_flush(struct bdi_writeback *wb) 1752 { 1753 unsigned long expired; 1754 long nr_pages; 1755 1756 /* 1757 * When set to zero, disable periodic writeback 1758 */ 1759 if (!dirty_writeback_interval) 1760 return 0; 1761 1762 expired = wb->last_old_flush + 1763 msecs_to_jiffies(dirty_writeback_interval * 10); 1764 if (time_before(jiffies, expired)) 1765 return 0; 1766 1767 wb->last_old_flush = jiffies; 1768 nr_pages = get_nr_dirty_pages(); 1769 1770 if (nr_pages) { 1771 struct wb_writeback_work work = { 1772 .nr_pages = nr_pages, 1773 .sync_mode = WB_SYNC_NONE, 1774 .for_kupdate = 1, 1775 .range_cyclic = 1, 1776 .reason = WB_REASON_PERIODIC, 1777 }; 1778 1779 return wb_writeback(wb, &work); 1780 } 1781 1782 return 0; 1783 } 1784 1785 /* 1786 * Retrieve work items and do the writeback they describe 1787 */ 1788 static long wb_do_writeback(struct bdi_writeback *wb) 1789 { 1790 struct wb_writeback_work *work; 1791 long wrote = 0; 1792 1793 set_bit(WB_writeback_running, &wb->state); 1794 while ((work = get_next_work_item(wb)) != NULL) { 1795 struct wb_completion *done = work->done; 1796 bool need_wake_up = false; 1797 1798 trace_writeback_exec(wb->bdi, work); 1799 1800 wrote += wb_writeback(wb, work); 1801 1802 if (work->single_wait) { 1803 WARN_ON_ONCE(work->auto_free); 1804 /* paired w/ rmb in wb_wait_for_single_work() */ 1805 smp_wmb(); 1806 work->single_done = 1; 1807 need_wake_up = true; 1808 } else if (work->auto_free) { 1809 kfree(work); 1810 } 1811 1812 if (done && atomic_dec_and_test(&done->cnt)) 1813 need_wake_up = true; 1814 1815 if (need_wake_up) 1816 wake_up_all(&wb->bdi->wb_waitq); 1817 } 1818 1819 /* 1820 * Check for periodic writeback, kupdated() style 1821 */ 1822 wrote += wb_check_old_data_flush(wb); 1823 wrote += wb_check_background_flush(wb); 1824 clear_bit(WB_writeback_running, &wb->state); 1825 1826 return wrote; 1827 } 1828 1829 /* 1830 * Handle writeback of dirty data for the device backed by this bdi. Also 1831 * reschedules periodically and does kupdated style flushing. 1832 */ 1833 void wb_workfn(struct work_struct *work) 1834 { 1835 struct bdi_writeback *wb = container_of(to_delayed_work(work), 1836 struct bdi_writeback, dwork); 1837 long pages_written; 1838 1839 set_worker_desc("flush-%s", dev_name(wb->bdi->dev)); 1840 current->flags |= PF_SWAPWRITE; 1841 1842 if (likely(!current_is_workqueue_rescuer() || 1843 !test_bit(WB_registered, &wb->state))) { 1844 /* 1845 * The normal path. Keep writing back @wb until its 1846 * work_list is empty. Note that this path is also taken 1847 * if @wb is shutting down even when we're running off the 1848 * rescuer as work_list needs to be drained. 1849 */ 1850 do { 1851 pages_written = wb_do_writeback(wb); 1852 trace_writeback_pages_written(pages_written); 1853 } while (!list_empty(&wb->work_list)); 1854 } else { 1855 /* 1856 * bdi_wq can't get enough workers and we're running off 1857 * the emergency worker. Don't hog it. Hopefully, 1024 is 1858 * enough for efficient IO. 1859 */ 1860 pages_written = writeback_inodes_wb(wb, 1024, 1861 WB_REASON_FORKER_THREAD); 1862 trace_writeback_pages_written(pages_written); 1863 } 1864 1865 if (!list_empty(&wb->work_list)) 1866 mod_delayed_work(bdi_wq, &wb->dwork, 0); 1867 else if (wb_has_dirty_io(wb) && dirty_writeback_interval) 1868 wb_wakeup_delayed(wb); 1869 1870 current->flags &= ~PF_SWAPWRITE; 1871 } 1872 1873 /* 1874 * Start writeback of `nr_pages' pages. If `nr_pages' is zero, write back 1875 * the whole world. 1876 */ 1877 void wakeup_flusher_threads(long nr_pages, enum wb_reason reason) 1878 { 1879 struct backing_dev_info *bdi; 1880 1881 if (!nr_pages) 1882 nr_pages = get_nr_dirty_pages(); 1883 1884 rcu_read_lock(); 1885 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) { 1886 struct bdi_writeback *wb; 1887 struct wb_iter iter; 1888 1889 if (!bdi_has_dirty_io(bdi)) 1890 continue; 1891 1892 bdi_for_each_wb(wb, bdi, &iter, 0) 1893 wb_start_writeback(wb, wb_split_bdi_pages(wb, nr_pages), 1894 false, reason); 1895 } 1896 rcu_read_unlock(); 1897 } 1898 1899 /* 1900 * Wake up bdi's periodically to make sure dirtytime inodes gets 1901 * written back periodically. We deliberately do *not* check the 1902 * b_dirtytime list in wb_has_dirty_io(), since this would cause the 1903 * kernel to be constantly waking up once there are any dirtytime 1904 * inodes on the system. So instead we define a separate delayed work 1905 * function which gets called much more rarely. (By default, only 1906 * once every 12 hours.) 1907 * 1908 * If there is any other write activity going on in the file system, 1909 * this function won't be necessary. But if the only thing that has 1910 * happened on the file system is a dirtytime inode caused by an atime 1911 * update, we need this infrastructure below to make sure that inode 1912 * eventually gets pushed out to disk. 1913 */ 1914 static void wakeup_dirtytime_writeback(struct work_struct *w); 1915 static DECLARE_DELAYED_WORK(dirtytime_work, wakeup_dirtytime_writeback); 1916 1917 static void wakeup_dirtytime_writeback(struct work_struct *w) 1918 { 1919 struct backing_dev_info *bdi; 1920 1921 rcu_read_lock(); 1922 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) { 1923 struct bdi_writeback *wb; 1924 struct wb_iter iter; 1925 1926 bdi_for_each_wb(wb, bdi, &iter, 0) 1927 if (!list_empty(&bdi->wb.b_dirty_time)) 1928 wb_wakeup(&bdi->wb); 1929 } 1930 rcu_read_unlock(); 1931 schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ); 1932 } 1933 1934 static int __init start_dirtytime_writeback(void) 1935 { 1936 schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ); 1937 return 0; 1938 } 1939 __initcall(start_dirtytime_writeback); 1940 1941 int dirtytime_interval_handler(struct ctl_table *table, int write, 1942 void __user *buffer, size_t *lenp, loff_t *ppos) 1943 { 1944 int ret; 1945 1946 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); 1947 if (ret == 0 && write) 1948 mod_delayed_work(system_wq, &dirtytime_work, 0); 1949 return ret; 1950 } 1951 1952 static noinline void block_dump___mark_inode_dirty(struct inode *inode) 1953 { 1954 if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) { 1955 struct dentry *dentry; 1956 const char *name = "?"; 1957 1958 dentry = d_find_alias(inode); 1959 if (dentry) { 1960 spin_lock(&dentry->d_lock); 1961 name = (const char *) dentry->d_name.name; 1962 } 1963 printk(KERN_DEBUG 1964 "%s(%d): dirtied inode %lu (%s) on %s\n", 1965 current->comm, task_pid_nr(current), inode->i_ino, 1966 name, inode->i_sb->s_id); 1967 if (dentry) { 1968 spin_unlock(&dentry->d_lock); 1969 dput(dentry); 1970 } 1971 } 1972 } 1973 1974 /** 1975 * __mark_inode_dirty - internal function 1976 * @inode: inode to mark 1977 * @flags: what kind of dirty (i.e. I_DIRTY_SYNC) 1978 * Mark an inode as dirty. Callers should use mark_inode_dirty or 1979 * mark_inode_dirty_sync. 1980 * 1981 * Put the inode on the super block's dirty list. 1982 * 1983 * CAREFUL! We mark it dirty unconditionally, but move it onto the 1984 * dirty list only if it is hashed or if it refers to a blockdev. 1985 * If it was not hashed, it will never be added to the dirty list 1986 * even if it is later hashed, as it will have been marked dirty already. 1987 * 1988 * In short, make sure you hash any inodes _before_ you start marking 1989 * them dirty. 1990 * 1991 * Note that for blockdevs, inode->dirtied_when represents the dirtying time of 1992 * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of 1993 * the kernel-internal blockdev inode represents the dirtying time of the 1994 * blockdev's pages. This is why for I_DIRTY_PAGES we always use 1995 * page->mapping->host, so the page-dirtying time is recorded in the internal 1996 * blockdev inode. 1997 */ 1998 #define I_DIRTY_INODE (I_DIRTY_SYNC | I_DIRTY_DATASYNC) 1999 void __mark_inode_dirty(struct inode *inode, int flags) 2000 { 2001 struct super_block *sb = inode->i_sb; 2002 int dirtytime; 2003 2004 trace_writeback_mark_inode_dirty(inode, flags); 2005 2006 /* 2007 * Don't do this for I_DIRTY_PAGES - that doesn't actually 2008 * dirty the inode itself 2009 */ 2010 if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC | I_DIRTY_TIME)) { 2011 trace_writeback_dirty_inode_start(inode, flags); 2012 2013 if (sb->s_op->dirty_inode) 2014 sb->s_op->dirty_inode(inode, flags); 2015 2016 trace_writeback_dirty_inode(inode, flags); 2017 } 2018 if (flags & I_DIRTY_INODE) 2019 flags &= ~I_DIRTY_TIME; 2020 dirtytime = flags & I_DIRTY_TIME; 2021 2022 /* 2023 * Paired with smp_mb() in __writeback_single_inode() for the 2024 * following lockless i_state test. See there for details. 2025 */ 2026 smp_mb(); 2027 2028 if (((inode->i_state & flags) == flags) || 2029 (dirtytime && (inode->i_state & I_DIRTY_INODE))) 2030 return; 2031 2032 if (unlikely(block_dump)) 2033 block_dump___mark_inode_dirty(inode); 2034 2035 spin_lock(&inode->i_lock); 2036 if (dirtytime && (inode->i_state & I_DIRTY_INODE)) 2037 goto out_unlock_inode; 2038 if ((inode->i_state & flags) != flags) { 2039 const int was_dirty = inode->i_state & I_DIRTY; 2040 2041 inode_attach_wb(inode, NULL); 2042 2043 if (flags & I_DIRTY_INODE) 2044 inode->i_state &= ~I_DIRTY_TIME; 2045 inode->i_state |= flags; 2046 2047 /* 2048 * If the inode is being synced, just update its dirty state. 2049 * The unlocker will place the inode on the appropriate 2050 * superblock list, based upon its state. 2051 */ 2052 if (inode->i_state & I_SYNC) 2053 goto out_unlock_inode; 2054 2055 /* 2056 * Only add valid (hashed) inodes to the superblock's 2057 * dirty list. Add blockdev inodes as well. 2058 */ 2059 if (!S_ISBLK(inode->i_mode)) { 2060 if (inode_unhashed(inode)) 2061 goto out_unlock_inode; 2062 } 2063 if (inode->i_state & I_FREEING) 2064 goto out_unlock_inode; 2065 2066 /* 2067 * If the inode was already on b_dirty/b_io/b_more_io, don't 2068 * reposition it (that would break b_dirty time-ordering). 2069 */ 2070 if (!was_dirty) { 2071 struct bdi_writeback *wb; 2072 struct list_head *dirty_list; 2073 bool wakeup_bdi = false; 2074 2075 wb = locked_inode_to_wb_and_lock_list(inode); 2076 2077 WARN(bdi_cap_writeback_dirty(wb->bdi) && 2078 !test_bit(WB_registered, &wb->state), 2079 "bdi-%s not registered\n", wb->bdi->name); 2080 2081 inode->dirtied_when = jiffies; 2082 if (dirtytime) 2083 inode->dirtied_time_when = jiffies; 2084 2085 if (inode->i_state & (I_DIRTY_INODE | I_DIRTY_PAGES)) 2086 dirty_list = &wb->b_dirty; 2087 else 2088 dirty_list = &wb->b_dirty_time; 2089 2090 wakeup_bdi = inode_wb_list_move_locked(inode, wb, 2091 dirty_list); 2092 2093 spin_unlock(&wb->list_lock); 2094 trace_writeback_dirty_inode_enqueue(inode); 2095 2096 /* 2097 * If this is the first dirty inode for this bdi, 2098 * we have to wake-up the corresponding bdi thread 2099 * to make sure background write-back happens 2100 * later. 2101 */ 2102 if (bdi_cap_writeback_dirty(wb->bdi) && wakeup_bdi) 2103 wb_wakeup_delayed(wb); 2104 return; 2105 } 2106 } 2107 out_unlock_inode: 2108 spin_unlock(&inode->i_lock); 2109 2110 } 2111 EXPORT_SYMBOL(__mark_inode_dirty); 2112 2113 static void wait_sb_inodes(struct super_block *sb) 2114 { 2115 struct inode *inode, *old_inode = NULL; 2116 2117 /* 2118 * We need to be protected against the filesystem going from 2119 * r/o to r/w or vice versa. 2120 */ 2121 WARN_ON(!rwsem_is_locked(&sb->s_umount)); 2122 2123 spin_lock(&inode_sb_list_lock); 2124 2125 /* 2126 * Data integrity sync. Must wait for all pages under writeback, 2127 * because there may have been pages dirtied before our sync 2128 * call, but which had writeout started before we write it out. 2129 * In which case, the inode may not be on the dirty list, but 2130 * we still have to wait for that writeout. 2131 */ 2132 list_for_each_entry(inode, &sb->s_inodes, i_sb_list) { 2133 struct address_space *mapping = inode->i_mapping; 2134 2135 spin_lock(&inode->i_lock); 2136 if ((inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) || 2137 (mapping->nrpages == 0)) { 2138 spin_unlock(&inode->i_lock); 2139 continue; 2140 } 2141 __iget(inode); 2142 spin_unlock(&inode->i_lock); 2143 spin_unlock(&inode_sb_list_lock); 2144 2145 /* 2146 * We hold a reference to 'inode' so it couldn't have been 2147 * removed from s_inodes list while we dropped the 2148 * inode_sb_list_lock. We cannot iput the inode now as we can 2149 * be holding the last reference and we cannot iput it under 2150 * inode_sb_list_lock. So we keep the reference and iput it 2151 * later. 2152 */ 2153 iput(old_inode); 2154 old_inode = inode; 2155 2156 filemap_fdatawait(mapping); 2157 2158 cond_resched(); 2159 2160 spin_lock(&inode_sb_list_lock); 2161 } 2162 spin_unlock(&inode_sb_list_lock); 2163 iput(old_inode); 2164 } 2165 2166 static void __writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr, 2167 enum wb_reason reason, bool skip_if_busy) 2168 { 2169 DEFINE_WB_COMPLETION_ONSTACK(done); 2170 struct wb_writeback_work work = { 2171 .sb = sb, 2172 .sync_mode = WB_SYNC_NONE, 2173 .tagged_writepages = 1, 2174 .done = &done, 2175 .nr_pages = nr, 2176 .reason = reason, 2177 }; 2178 struct backing_dev_info *bdi = sb->s_bdi; 2179 2180 if (!bdi_has_dirty_io(bdi) || bdi == &noop_backing_dev_info) 2181 return; 2182 WARN_ON(!rwsem_is_locked(&sb->s_umount)); 2183 2184 bdi_split_work_to_wbs(sb->s_bdi, &work, skip_if_busy); 2185 wb_wait_for_completion(bdi, &done); 2186 } 2187 2188 /** 2189 * writeback_inodes_sb_nr - writeback dirty inodes from given super_block 2190 * @sb: the superblock 2191 * @nr: the number of pages to write 2192 * @reason: reason why some writeback work initiated 2193 * 2194 * Start writeback on some inodes on this super_block. No guarantees are made 2195 * on how many (if any) will be written, and this function does not wait 2196 * for IO completion of submitted IO. 2197 */ 2198 void writeback_inodes_sb_nr(struct super_block *sb, 2199 unsigned long nr, 2200 enum wb_reason reason) 2201 { 2202 __writeback_inodes_sb_nr(sb, nr, reason, false); 2203 } 2204 EXPORT_SYMBOL(writeback_inodes_sb_nr); 2205 2206 /** 2207 * writeback_inodes_sb - writeback dirty inodes from given super_block 2208 * @sb: the superblock 2209 * @reason: reason why some writeback work was initiated 2210 * 2211 * Start writeback on some inodes on this super_block. No guarantees are made 2212 * on how many (if any) will be written, and this function does not wait 2213 * for IO completion of submitted IO. 2214 */ 2215 void writeback_inodes_sb(struct super_block *sb, enum wb_reason reason) 2216 { 2217 return writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason); 2218 } 2219 EXPORT_SYMBOL(writeback_inodes_sb); 2220 2221 /** 2222 * try_to_writeback_inodes_sb_nr - try to start writeback if none underway 2223 * @sb: the superblock 2224 * @nr: the number of pages to write 2225 * @reason: the reason of writeback 2226 * 2227 * Invoke writeback_inodes_sb_nr if no writeback is currently underway. 2228 * Returns 1 if writeback was started, 0 if not. 2229 */ 2230 bool try_to_writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr, 2231 enum wb_reason reason) 2232 { 2233 if (!down_read_trylock(&sb->s_umount)) 2234 return false; 2235 2236 __writeback_inodes_sb_nr(sb, nr, reason, true); 2237 up_read(&sb->s_umount); 2238 return true; 2239 } 2240 EXPORT_SYMBOL(try_to_writeback_inodes_sb_nr); 2241 2242 /** 2243 * try_to_writeback_inodes_sb - try to start writeback if none underway 2244 * @sb: the superblock 2245 * @reason: reason why some writeback work was initiated 2246 * 2247 * Implement by try_to_writeback_inodes_sb_nr() 2248 * Returns 1 if writeback was started, 0 if not. 2249 */ 2250 bool try_to_writeback_inodes_sb(struct super_block *sb, enum wb_reason reason) 2251 { 2252 return try_to_writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason); 2253 } 2254 EXPORT_SYMBOL(try_to_writeback_inodes_sb); 2255 2256 /** 2257 * sync_inodes_sb - sync sb inode pages 2258 * @sb: the superblock 2259 * 2260 * This function writes and waits on any dirty inode belonging to this 2261 * super_block. 2262 */ 2263 void sync_inodes_sb(struct super_block *sb) 2264 { 2265 DEFINE_WB_COMPLETION_ONSTACK(done); 2266 struct wb_writeback_work work = { 2267 .sb = sb, 2268 .sync_mode = WB_SYNC_ALL, 2269 .nr_pages = LONG_MAX, 2270 .range_cyclic = 0, 2271 .done = &done, 2272 .reason = WB_REASON_SYNC, 2273 .for_sync = 1, 2274 }; 2275 struct backing_dev_info *bdi = sb->s_bdi; 2276 2277 /* Nothing to do? */ 2278 if (!bdi_has_dirty_io(bdi) || bdi == &noop_backing_dev_info) 2279 return; 2280 WARN_ON(!rwsem_is_locked(&sb->s_umount)); 2281 2282 bdi_split_work_to_wbs(bdi, &work, false); 2283 wb_wait_for_completion(bdi, &done); 2284 2285 wait_sb_inodes(sb); 2286 } 2287 EXPORT_SYMBOL(sync_inodes_sb); 2288 2289 /** 2290 * write_inode_now - write an inode to disk 2291 * @inode: inode to write to disk 2292 * @sync: whether the write should be synchronous or not 2293 * 2294 * This function commits an inode to disk immediately if it is dirty. This is 2295 * primarily needed by knfsd. 2296 * 2297 * The caller must either have a ref on the inode or must have set I_WILL_FREE. 2298 */ 2299 int write_inode_now(struct inode *inode, int sync) 2300 { 2301 struct bdi_writeback *wb = &inode_to_bdi(inode)->wb; 2302 struct writeback_control wbc = { 2303 .nr_to_write = LONG_MAX, 2304 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE, 2305 .range_start = 0, 2306 .range_end = LLONG_MAX, 2307 }; 2308 2309 if (!mapping_cap_writeback_dirty(inode->i_mapping)) 2310 wbc.nr_to_write = 0; 2311 2312 might_sleep(); 2313 return writeback_single_inode(inode, wb, &wbc); 2314 } 2315 EXPORT_SYMBOL(write_inode_now); 2316 2317 /** 2318 * sync_inode - write an inode and its pages to disk. 2319 * @inode: the inode to sync 2320 * @wbc: controls the writeback mode 2321 * 2322 * sync_inode() will write an inode and its pages to disk. It will also 2323 * correctly update the inode on its superblock's dirty inode lists and will 2324 * update inode->i_state. 2325 * 2326 * The caller must have a ref on the inode. 2327 */ 2328 int sync_inode(struct inode *inode, struct writeback_control *wbc) 2329 { 2330 return writeback_single_inode(inode, &inode_to_bdi(inode)->wb, wbc); 2331 } 2332 EXPORT_SYMBOL(sync_inode); 2333 2334 /** 2335 * sync_inode_metadata - write an inode to disk 2336 * @inode: the inode to sync 2337 * @wait: wait for I/O to complete. 2338 * 2339 * Write an inode to disk and adjust its dirty state after completion. 2340 * 2341 * Note: only writes the actual inode, no associated data or other metadata. 2342 */ 2343 int sync_inode_metadata(struct inode *inode, int wait) 2344 { 2345 struct writeback_control wbc = { 2346 .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE, 2347 .nr_to_write = 0, /* metadata-only */ 2348 }; 2349 2350 return sync_inode(inode, &wbc); 2351 } 2352 EXPORT_SYMBOL(sync_inode_metadata); 2353