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_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 enum wb_reason reason; /* why was writeback initiated? */ 57 58 struct list_head list; /* pending work list */ 59 struct wb_completion *done; /* set if the caller waits */ 60 }; 61 62 /* 63 * If one wants to wait for one or more wb_writeback_works, each work's 64 * ->done should be set to a wb_completion defined using the following 65 * macro. Once all work items are issued with wb_queue_work(), the caller 66 * can wait for the completion of all using wb_wait_for_completion(). Work 67 * items which are waited upon aren't freed automatically on completion. 68 */ 69 #define DEFINE_WB_COMPLETION_ONSTACK(cmpl) \ 70 struct wb_completion cmpl = { \ 71 .cnt = ATOMIC_INIT(1), \ 72 } 73 74 75 /* 76 * If an inode is constantly having its pages dirtied, but then the 77 * updates stop dirtytime_expire_interval seconds in the past, it's 78 * possible for the worst case time between when an inode has its 79 * timestamps updated and when they finally get written out to be two 80 * dirtytime_expire_intervals. We set the default to 12 hours (in 81 * seconds), which means most of the time inodes will have their 82 * timestamps written to disk after 12 hours, but in the worst case a 83 * few inodes might not their timestamps updated for 24 hours. 84 */ 85 unsigned int dirtytime_expire_interval = 12 * 60 * 60; 86 87 static inline struct inode *wb_inode(struct list_head *head) 88 { 89 return list_entry(head, struct inode, i_io_list); 90 } 91 92 /* 93 * Include the creation of the trace points after defining the 94 * wb_writeback_work structure and inline functions so that the definition 95 * remains local to this file. 96 */ 97 #define CREATE_TRACE_POINTS 98 #include <trace/events/writeback.h> 99 100 EXPORT_TRACEPOINT_SYMBOL_GPL(wbc_writepage); 101 102 static bool wb_io_lists_populated(struct bdi_writeback *wb) 103 { 104 if (wb_has_dirty_io(wb)) { 105 return false; 106 } else { 107 set_bit(WB_has_dirty_io, &wb->state); 108 WARN_ON_ONCE(!wb->avg_write_bandwidth); 109 atomic_long_add(wb->avg_write_bandwidth, 110 &wb->bdi->tot_write_bandwidth); 111 return true; 112 } 113 } 114 115 static void wb_io_lists_depopulated(struct bdi_writeback *wb) 116 { 117 if (wb_has_dirty_io(wb) && list_empty(&wb->b_dirty) && 118 list_empty(&wb->b_io) && list_empty(&wb->b_more_io)) { 119 clear_bit(WB_has_dirty_io, &wb->state); 120 WARN_ON_ONCE(atomic_long_sub_return(wb->avg_write_bandwidth, 121 &wb->bdi->tot_write_bandwidth) < 0); 122 } 123 } 124 125 /** 126 * inode_io_list_move_locked - move an inode onto a bdi_writeback IO list 127 * @inode: inode to be moved 128 * @wb: target bdi_writeback 129 * @head: one of @wb->b_{dirty|io|more_io} 130 * 131 * Move @inode->i_io_list to @list of @wb and set %WB_has_dirty_io. 132 * Returns %true if @inode is the first occupant of the !dirty_time IO 133 * lists; otherwise, %false. 134 */ 135 static bool inode_io_list_move_locked(struct inode *inode, 136 struct bdi_writeback *wb, 137 struct list_head *head) 138 { 139 assert_spin_locked(&wb->list_lock); 140 141 list_move(&inode->i_io_list, head); 142 143 /* dirty_time doesn't count as dirty_io until expiration */ 144 if (head != &wb->b_dirty_time) 145 return wb_io_lists_populated(wb); 146 147 wb_io_lists_depopulated(wb); 148 return false; 149 } 150 151 /** 152 * inode_io_list_del_locked - remove an inode from its bdi_writeback IO list 153 * @inode: inode to be removed 154 * @wb: bdi_writeback @inode is being removed from 155 * 156 * Remove @inode which may be on one of @wb->b_{dirty|io|more_io} lists and 157 * clear %WB_has_dirty_io if all are empty afterwards. 158 */ 159 static void inode_io_list_del_locked(struct inode *inode, 160 struct bdi_writeback *wb) 161 { 162 assert_spin_locked(&wb->list_lock); 163 164 list_del_init(&inode->i_io_list); 165 wb_io_lists_depopulated(wb); 166 } 167 168 static void wb_wakeup(struct bdi_writeback *wb) 169 { 170 spin_lock_bh(&wb->work_lock); 171 if (test_bit(WB_registered, &wb->state)) 172 mod_delayed_work(bdi_wq, &wb->dwork, 0); 173 spin_unlock_bh(&wb->work_lock); 174 } 175 176 static void finish_writeback_work(struct bdi_writeback *wb, 177 struct wb_writeback_work *work) 178 { 179 struct wb_completion *done = work->done; 180 181 if (work->auto_free) 182 kfree(work); 183 if (done && atomic_dec_and_test(&done->cnt)) 184 wake_up_all(&wb->bdi->wb_waitq); 185 } 186 187 static void wb_queue_work(struct bdi_writeback *wb, 188 struct wb_writeback_work *work) 189 { 190 trace_writeback_queue(wb, work); 191 192 if (work->done) 193 atomic_inc(&work->done->cnt); 194 195 spin_lock_bh(&wb->work_lock); 196 197 if (test_bit(WB_registered, &wb->state)) { 198 list_add_tail(&work->list, &wb->work_list); 199 mod_delayed_work(bdi_wq, &wb->dwork, 0); 200 } else 201 finish_writeback_work(wb, work); 202 203 spin_unlock_bh(&wb->work_lock); 204 } 205 206 /** 207 * wb_wait_for_completion - wait for completion of bdi_writeback_works 208 * @bdi: bdi work items were issued to 209 * @done: target wb_completion 210 * 211 * Wait for one or more work items issued to @bdi with their ->done field 212 * set to @done, which should have been defined with 213 * DEFINE_WB_COMPLETION_ONSTACK(). This function returns after all such 214 * work items are completed. Work items which are waited upon aren't freed 215 * automatically on completion. 216 */ 217 static void wb_wait_for_completion(struct backing_dev_info *bdi, 218 struct wb_completion *done) 219 { 220 atomic_dec(&done->cnt); /* put down the initial count */ 221 wait_event(bdi->wb_waitq, !atomic_read(&done->cnt)); 222 } 223 224 #ifdef CONFIG_CGROUP_WRITEBACK 225 226 /* parameters for foreign inode detection, see wb_detach_inode() */ 227 #define WB_FRN_TIME_SHIFT 13 /* 1s = 2^13, upto 8 secs w/ 16bit */ 228 #define WB_FRN_TIME_AVG_SHIFT 3 /* avg = avg * 7/8 + new * 1/8 */ 229 #define WB_FRN_TIME_CUT_DIV 2 /* ignore rounds < avg / 2 */ 230 #define WB_FRN_TIME_PERIOD (2 * (1 << WB_FRN_TIME_SHIFT)) /* 2s */ 231 232 #define WB_FRN_HIST_SLOTS 16 /* inode->i_wb_frn_history is 16bit */ 233 #define WB_FRN_HIST_UNIT (WB_FRN_TIME_PERIOD / WB_FRN_HIST_SLOTS) 234 /* each slot's duration is 2s / 16 */ 235 #define WB_FRN_HIST_THR_SLOTS (WB_FRN_HIST_SLOTS / 2) 236 /* if foreign slots >= 8, switch */ 237 #define WB_FRN_HIST_MAX_SLOTS (WB_FRN_HIST_THR_SLOTS / 2 + 1) 238 /* one round can affect upto 5 slots */ 239 240 static atomic_t isw_nr_in_flight = ATOMIC_INIT(0); 241 static struct workqueue_struct *isw_wq; 242 243 void __inode_attach_wb(struct inode *inode, struct page *page) 244 { 245 struct backing_dev_info *bdi = inode_to_bdi(inode); 246 struct bdi_writeback *wb = NULL; 247 248 if (inode_cgwb_enabled(inode)) { 249 struct cgroup_subsys_state *memcg_css; 250 251 if (page) { 252 memcg_css = mem_cgroup_css_from_page(page); 253 wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC); 254 } else { 255 /* must pin memcg_css, see wb_get_create() */ 256 memcg_css = task_get_css(current, memory_cgrp_id); 257 wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC); 258 css_put(memcg_css); 259 } 260 } 261 262 if (!wb) 263 wb = &bdi->wb; 264 265 /* 266 * There may be multiple instances of this function racing to 267 * update the same inode. Use cmpxchg() to tell the winner. 268 */ 269 if (unlikely(cmpxchg(&inode->i_wb, NULL, wb))) 270 wb_put(wb); 271 } 272 273 /** 274 * locked_inode_to_wb_and_lock_list - determine a locked inode's wb and lock it 275 * @inode: inode of interest with i_lock held 276 * 277 * Returns @inode's wb with its list_lock held. @inode->i_lock must be 278 * held on entry and is released on return. The returned wb is guaranteed 279 * to stay @inode's associated wb until its list_lock is released. 280 */ 281 static struct bdi_writeback * 282 locked_inode_to_wb_and_lock_list(struct inode *inode) 283 __releases(&inode->i_lock) 284 __acquires(&wb->list_lock) 285 { 286 while (true) { 287 struct bdi_writeback *wb = inode_to_wb(inode); 288 289 /* 290 * inode_to_wb() association is protected by both 291 * @inode->i_lock and @wb->list_lock but list_lock nests 292 * outside i_lock. Drop i_lock and verify that the 293 * association hasn't changed after acquiring list_lock. 294 */ 295 wb_get(wb); 296 spin_unlock(&inode->i_lock); 297 spin_lock(&wb->list_lock); 298 299 /* i_wb may have changed inbetween, can't use inode_to_wb() */ 300 if (likely(wb == inode->i_wb)) { 301 wb_put(wb); /* @inode already has ref */ 302 return wb; 303 } 304 305 spin_unlock(&wb->list_lock); 306 wb_put(wb); 307 cpu_relax(); 308 spin_lock(&inode->i_lock); 309 } 310 } 311 312 /** 313 * inode_to_wb_and_lock_list - determine an inode's wb and lock it 314 * @inode: inode of interest 315 * 316 * Same as locked_inode_to_wb_and_lock_list() but @inode->i_lock isn't held 317 * on entry. 318 */ 319 static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode) 320 __acquires(&wb->list_lock) 321 { 322 spin_lock(&inode->i_lock); 323 return locked_inode_to_wb_and_lock_list(inode); 324 } 325 326 struct inode_switch_wbs_context { 327 struct inode *inode; 328 struct bdi_writeback *new_wb; 329 330 struct rcu_head rcu_head; 331 struct work_struct work; 332 }; 333 334 static void inode_switch_wbs_work_fn(struct work_struct *work) 335 { 336 struct inode_switch_wbs_context *isw = 337 container_of(work, struct inode_switch_wbs_context, work); 338 struct inode *inode = isw->inode; 339 struct address_space *mapping = inode->i_mapping; 340 struct bdi_writeback *old_wb = inode->i_wb; 341 struct bdi_writeback *new_wb = isw->new_wb; 342 struct radix_tree_iter iter; 343 bool switched = false; 344 void **slot; 345 346 /* 347 * By the time control reaches here, RCU grace period has passed 348 * since I_WB_SWITCH assertion and all wb stat update transactions 349 * between unlocked_inode_to_wb_begin/end() are guaranteed to be 350 * synchronizing against mapping->tree_lock. 351 * 352 * Grabbing old_wb->list_lock, inode->i_lock and mapping->tree_lock 353 * gives us exclusion against all wb related operations on @inode 354 * including IO list manipulations and stat updates. 355 */ 356 if (old_wb < new_wb) { 357 spin_lock(&old_wb->list_lock); 358 spin_lock_nested(&new_wb->list_lock, SINGLE_DEPTH_NESTING); 359 } else { 360 spin_lock(&new_wb->list_lock); 361 spin_lock_nested(&old_wb->list_lock, SINGLE_DEPTH_NESTING); 362 } 363 spin_lock(&inode->i_lock); 364 spin_lock_irq(&mapping->tree_lock); 365 366 /* 367 * Once I_FREEING is visible under i_lock, the eviction path owns 368 * the inode and we shouldn't modify ->i_io_list. 369 */ 370 if (unlikely(inode->i_state & I_FREEING)) 371 goto skip_switch; 372 373 /* 374 * Count and transfer stats. Note that PAGECACHE_TAG_DIRTY points 375 * to possibly dirty pages while PAGECACHE_TAG_WRITEBACK points to 376 * pages actually under underwriteback. 377 */ 378 radix_tree_for_each_tagged(slot, &mapping->page_tree, &iter, 0, 379 PAGECACHE_TAG_DIRTY) { 380 struct page *page = radix_tree_deref_slot_protected(slot, 381 &mapping->tree_lock); 382 if (likely(page) && PageDirty(page)) { 383 dec_wb_stat(old_wb, WB_RECLAIMABLE); 384 inc_wb_stat(new_wb, WB_RECLAIMABLE); 385 } 386 } 387 388 radix_tree_for_each_tagged(slot, &mapping->page_tree, &iter, 0, 389 PAGECACHE_TAG_WRITEBACK) { 390 struct page *page = radix_tree_deref_slot_protected(slot, 391 &mapping->tree_lock); 392 if (likely(page)) { 393 WARN_ON_ONCE(!PageWriteback(page)); 394 dec_wb_stat(old_wb, WB_WRITEBACK); 395 inc_wb_stat(new_wb, WB_WRITEBACK); 396 } 397 } 398 399 wb_get(new_wb); 400 401 /* 402 * Transfer to @new_wb's IO list if necessary. The specific list 403 * @inode was on is ignored and the inode is put on ->b_dirty which 404 * is always correct including from ->b_dirty_time. The transfer 405 * preserves @inode->dirtied_when ordering. 406 */ 407 if (!list_empty(&inode->i_io_list)) { 408 struct inode *pos; 409 410 inode_io_list_del_locked(inode, old_wb); 411 inode->i_wb = new_wb; 412 list_for_each_entry(pos, &new_wb->b_dirty, i_io_list) 413 if (time_after_eq(inode->dirtied_when, 414 pos->dirtied_when)) 415 break; 416 inode_io_list_move_locked(inode, new_wb, pos->i_io_list.prev); 417 } else { 418 inode->i_wb = new_wb; 419 } 420 421 /* ->i_wb_frn updates may race wbc_detach_inode() but doesn't matter */ 422 inode->i_wb_frn_winner = 0; 423 inode->i_wb_frn_avg_time = 0; 424 inode->i_wb_frn_history = 0; 425 switched = true; 426 skip_switch: 427 /* 428 * Paired with load_acquire in unlocked_inode_to_wb_begin() and 429 * ensures that the new wb is visible if they see !I_WB_SWITCH. 430 */ 431 smp_store_release(&inode->i_state, inode->i_state & ~I_WB_SWITCH); 432 433 spin_unlock_irq(&mapping->tree_lock); 434 spin_unlock(&inode->i_lock); 435 spin_unlock(&new_wb->list_lock); 436 spin_unlock(&old_wb->list_lock); 437 438 if (switched) { 439 wb_wakeup(new_wb); 440 wb_put(old_wb); 441 } 442 wb_put(new_wb); 443 444 iput(inode); 445 kfree(isw); 446 447 atomic_dec(&isw_nr_in_flight); 448 } 449 450 static void inode_switch_wbs_rcu_fn(struct rcu_head *rcu_head) 451 { 452 struct inode_switch_wbs_context *isw = container_of(rcu_head, 453 struct inode_switch_wbs_context, rcu_head); 454 455 /* needs to grab bh-unsafe locks, bounce to work item */ 456 INIT_WORK(&isw->work, inode_switch_wbs_work_fn); 457 queue_work(isw_wq, &isw->work); 458 } 459 460 /** 461 * inode_switch_wbs - change the wb association of an inode 462 * @inode: target inode 463 * @new_wb_id: ID of the new wb 464 * 465 * Switch @inode's wb association to the wb identified by @new_wb_id. The 466 * switching is performed asynchronously and may fail silently. 467 */ 468 static void inode_switch_wbs(struct inode *inode, int new_wb_id) 469 { 470 struct backing_dev_info *bdi = inode_to_bdi(inode); 471 struct cgroup_subsys_state *memcg_css; 472 struct inode_switch_wbs_context *isw; 473 474 /* noop if seems to be already in progress */ 475 if (inode->i_state & I_WB_SWITCH) 476 return; 477 478 isw = kzalloc(sizeof(*isw), GFP_ATOMIC); 479 if (!isw) 480 return; 481 482 /* find and pin the new wb */ 483 rcu_read_lock(); 484 memcg_css = css_from_id(new_wb_id, &memory_cgrp_subsys); 485 if (memcg_css) 486 isw->new_wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC); 487 rcu_read_unlock(); 488 if (!isw->new_wb) 489 goto out_free; 490 491 /* while holding I_WB_SWITCH, no one else can update the association */ 492 spin_lock(&inode->i_lock); 493 if (!(inode->i_sb->s_flags & MS_ACTIVE) || 494 inode->i_state & (I_WB_SWITCH | I_FREEING) || 495 inode_to_wb(inode) == isw->new_wb) { 496 spin_unlock(&inode->i_lock); 497 goto out_free; 498 } 499 inode->i_state |= I_WB_SWITCH; 500 __iget(inode); 501 spin_unlock(&inode->i_lock); 502 503 isw->inode = inode; 504 505 atomic_inc(&isw_nr_in_flight); 506 507 /* 508 * In addition to synchronizing among switchers, I_WB_SWITCH tells 509 * the RCU protected stat update paths to grab the mapping's 510 * tree_lock so that stat transfer can synchronize against them. 511 * Let's continue after I_WB_SWITCH is guaranteed to be visible. 512 */ 513 call_rcu(&isw->rcu_head, inode_switch_wbs_rcu_fn); 514 return; 515 516 out_free: 517 if (isw->new_wb) 518 wb_put(isw->new_wb); 519 kfree(isw); 520 } 521 522 /** 523 * wbc_attach_and_unlock_inode - associate wbc with target inode and unlock it 524 * @wbc: writeback_control of interest 525 * @inode: target inode 526 * 527 * @inode is locked and about to be written back under the control of @wbc. 528 * Record @inode's writeback context into @wbc and unlock the i_lock. On 529 * writeback completion, wbc_detach_inode() should be called. This is used 530 * to track the cgroup writeback context. 531 */ 532 void wbc_attach_and_unlock_inode(struct writeback_control *wbc, 533 struct inode *inode) 534 { 535 if (!inode_cgwb_enabled(inode)) { 536 spin_unlock(&inode->i_lock); 537 return; 538 } 539 540 wbc->wb = inode_to_wb(inode); 541 wbc->inode = inode; 542 543 wbc->wb_id = wbc->wb->memcg_css->id; 544 wbc->wb_lcand_id = inode->i_wb_frn_winner; 545 wbc->wb_tcand_id = 0; 546 wbc->wb_bytes = 0; 547 wbc->wb_lcand_bytes = 0; 548 wbc->wb_tcand_bytes = 0; 549 550 wb_get(wbc->wb); 551 spin_unlock(&inode->i_lock); 552 553 /* 554 * A dying wb indicates that the memcg-blkcg mapping has changed 555 * and a new wb is already serving the memcg. Switch immediately. 556 */ 557 if (unlikely(wb_dying(wbc->wb))) 558 inode_switch_wbs(inode, wbc->wb_id); 559 } 560 561 /** 562 * wbc_detach_inode - disassociate wbc from inode and perform foreign detection 563 * @wbc: writeback_control of the just finished writeback 564 * 565 * To be called after a writeback attempt of an inode finishes and undoes 566 * wbc_attach_and_unlock_inode(). Can be called under any context. 567 * 568 * As concurrent write sharing of an inode is expected to be very rare and 569 * memcg only tracks page ownership on first-use basis severely confining 570 * the usefulness of such sharing, cgroup writeback tracks ownership 571 * per-inode. While the support for concurrent write sharing of an inode 572 * is deemed unnecessary, an inode being written to by different cgroups at 573 * different points in time is a lot more common, and, more importantly, 574 * charging only by first-use can too readily lead to grossly incorrect 575 * behaviors (single foreign page can lead to gigabytes of writeback to be 576 * incorrectly attributed). 577 * 578 * To resolve this issue, cgroup writeback detects the majority dirtier of 579 * an inode and transfers the ownership to it. To avoid unnnecessary 580 * oscillation, the detection mechanism keeps track of history and gives 581 * out the switch verdict only if the foreign usage pattern is stable over 582 * a certain amount of time and/or writeback attempts. 583 * 584 * On each writeback attempt, @wbc tries to detect the majority writer 585 * using Boyer-Moore majority vote algorithm. In addition to the byte 586 * count from the majority voting, it also counts the bytes written for the 587 * current wb and the last round's winner wb (max of last round's current 588 * wb, the winner from two rounds ago, and the last round's majority 589 * candidate). Keeping track of the historical winner helps the algorithm 590 * to semi-reliably detect the most active writer even when it's not the 591 * absolute majority. 592 * 593 * Once the winner of the round is determined, whether the winner is 594 * foreign or not and how much IO time the round consumed is recorded in 595 * inode->i_wb_frn_history. If the amount of recorded foreign IO time is 596 * over a certain threshold, the switch verdict is given. 597 */ 598 void wbc_detach_inode(struct writeback_control *wbc) 599 { 600 struct bdi_writeback *wb = wbc->wb; 601 struct inode *inode = wbc->inode; 602 unsigned long avg_time, max_bytes, max_time; 603 u16 history; 604 int max_id; 605 606 if (!wb) 607 return; 608 609 history = inode->i_wb_frn_history; 610 avg_time = inode->i_wb_frn_avg_time; 611 612 /* pick the winner of this round */ 613 if (wbc->wb_bytes >= wbc->wb_lcand_bytes && 614 wbc->wb_bytes >= wbc->wb_tcand_bytes) { 615 max_id = wbc->wb_id; 616 max_bytes = wbc->wb_bytes; 617 } else if (wbc->wb_lcand_bytes >= wbc->wb_tcand_bytes) { 618 max_id = wbc->wb_lcand_id; 619 max_bytes = wbc->wb_lcand_bytes; 620 } else { 621 max_id = wbc->wb_tcand_id; 622 max_bytes = wbc->wb_tcand_bytes; 623 } 624 625 /* 626 * Calculate the amount of IO time the winner consumed and fold it 627 * into the running average kept per inode. If the consumed IO 628 * time is lower than avag / WB_FRN_TIME_CUT_DIV, ignore it for 629 * deciding whether to switch or not. This is to prevent one-off 630 * small dirtiers from skewing the verdict. 631 */ 632 max_time = DIV_ROUND_UP((max_bytes >> PAGE_SHIFT) << WB_FRN_TIME_SHIFT, 633 wb->avg_write_bandwidth); 634 if (avg_time) 635 avg_time += (max_time >> WB_FRN_TIME_AVG_SHIFT) - 636 (avg_time >> WB_FRN_TIME_AVG_SHIFT); 637 else 638 avg_time = max_time; /* immediate catch up on first run */ 639 640 if (max_time >= avg_time / WB_FRN_TIME_CUT_DIV) { 641 int slots; 642 643 /* 644 * The switch verdict is reached if foreign wb's consume 645 * more than a certain proportion of IO time in a 646 * WB_FRN_TIME_PERIOD. This is loosely tracked by 16 slot 647 * history mask where each bit represents one sixteenth of 648 * the period. Determine the number of slots to shift into 649 * history from @max_time. 650 */ 651 slots = min(DIV_ROUND_UP(max_time, WB_FRN_HIST_UNIT), 652 (unsigned long)WB_FRN_HIST_MAX_SLOTS); 653 history <<= slots; 654 if (wbc->wb_id != max_id) 655 history |= (1U << slots) - 1; 656 657 /* 658 * Switch if the current wb isn't the consistent winner. 659 * If there are multiple closely competing dirtiers, the 660 * inode may switch across them repeatedly over time, which 661 * is okay. The main goal is avoiding keeping an inode on 662 * the wrong wb for an extended period of time. 663 */ 664 if (hweight32(history) > WB_FRN_HIST_THR_SLOTS) 665 inode_switch_wbs(inode, max_id); 666 } 667 668 /* 669 * Multiple instances of this function may race to update the 670 * following fields but we don't mind occassional inaccuracies. 671 */ 672 inode->i_wb_frn_winner = max_id; 673 inode->i_wb_frn_avg_time = min(avg_time, (unsigned long)U16_MAX); 674 inode->i_wb_frn_history = history; 675 676 wb_put(wbc->wb); 677 wbc->wb = NULL; 678 } 679 680 /** 681 * wbc_account_io - account IO issued during writeback 682 * @wbc: writeback_control of the writeback in progress 683 * @page: page being written out 684 * @bytes: number of bytes being written out 685 * 686 * @bytes from @page are about to written out during the writeback 687 * controlled by @wbc. Keep the book for foreign inode detection. See 688 * wbc_detach_inode(). 689 */ 690 void wbc_account_io(struct writeback_control *wbc, struct page *page, 691 size_t bytes) 692 { 693 int id; 694 695 /* 696 * pageout() path doesn't attach @wbc to the inode being written 697 * out. This is intentional as we don't want the function to block 698 * behind a slow cgroup. Ultimately, we want pageout() to kick off 699 * regular writeback instead of writing things out itself. 700 */ 701 if (!wbc->wb) 702 return; 703 704 id = mem_cgroup_css_from_page(page)->id; 705 706 if (id == wbc->wb_id) { 707 wbc->wb_bytes += bytes; 708 return; 709 } 710 711 if (id == wbc->wb_lcand_id) 712 wbc->wb_lcand_bytes += bytes; 713 714 /* Boyer-Moore majority vote algorithm */ 715 if (!wbc->wb_tcand_bytes) 716 wbc->wb_tcand_id = id; 717 if (id == wbc->wb_tcand_id) 718 wbc->wb_tcand_bytes += bytes; 719 else 720 wbc->wb_tcand_bytes -= min(bytes, wbc->wb_tcand_bytes); 721 } 722 EXPORT_SYMBOL_GPL(wbc_account_io); 723 724 /** 725 * inode_congested - test whether an inode is congested 726 * @inode: inode to test for congestion (may be NULL) 727 * @cong_bits: mask of WB_[a]sync_congested bits to test 728 * 729 * Tests whether @inode is congested. @cong_bits is the mask of congestion 730 * bits to test and the return value is the mask of set bits. 731 * 732 * If cgroup writeback is enabled for @inode, the congestion state is 733 * determined by whether the cgwb (cgroup bdi_writeback) for the blkcg 734 * associated with @inode is congested; otherwise, the root wb's congestion 735 * state is used. 736 * 737 * @inode is allowed to be NULL as this function is often called on 738 * mapping->host which is NULL for the swapper space. 739 */ 740 int inode_congested(struct inode *inode, int cong_bits) 741 { 742 /* 743 * Once set, ->i_wb never becomes NULL while the inode is alive. 744 * Start transaction iff ->i_wb is visible. 745 */ 746 if (inode && inode_to_wb_is_valid(inode)) { 747 struct bdi_writeback *wb; 748 bool locked, congested; 749 750 wb = unlocked_inode_to_wb_begin(inode, &locked); 751 congested = wb_congested(wb, cong_bits); 752 unlocked_inode_to_wb_end(inode, locked); 753 return congested; 754 } 755 756 return wb_congested(&inode_to_bdi(inode)->wb, cong_bits); 757 } 758 EXPORT_SYMBOL_GPL(inode_congested); 759 760 /** 761 * wb_split_bdi_pages - split nr_pages to write according to bandwidth 762 * @wb: target bdi_writeback to split @nr_pages to 763 * @nr_pages: number of pages to write for the whole bdi 764 * 765 * Split @wb's portion of @nr_pages according to @wb's write bandwidth in 766 * relation to the total write bandwidth of all wb's w/ dirty inodes on 767 * @wb->bdi. 768 */ 769 static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages) 770 { 771 unsigned long this_bw = wb->avg_write_bandwidth; 772 unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth); 773 774 if (nr_pages == LONG_MAX) 775 return LONG_MAX; 776 777 /* 778 * This may be called on clean wb's and proportional distribution 779 * may not make sense, just use the original @nr_pages in those 780 * cases. In general, we wanna err on the side of writing more. 781 */ 782 if (!tot_bw || this_bw >= tot_bw) 783 return nr_pages; 784 else 785 return DIV_ROUND_UP_ULL((u64)nr_pages * this_bw, tot_bw); 786 } 787 788 /** 789 * bdi_split_work_to_wbs - split a wb_writeback_work to all wb's of a bdi 790 * @bdi: target backing_dev_info 791 * @base_work: wb_writeback_work to issue 792 * @skip_if_busy: skip wb's which already have writeback in progress 793 * 794 * Split and issue @base_work to all wb's (bdi_writeback's) of @bdi which 795 * have dirty inodes. If @base_work->nr_page isn't %LONG_MAX, it's 796 * distributed to the busy wbs according to each wb's proportion in the 797 * total active write bandwidth of @bdi. 798 */ 799 static void bdi_split_work_to_wbs(struct backing_dev_info *bdi, 800 struct wb_writeback_work *base_work, 801 bool skip_if_busy) 802 { 803 struct bdi_writeback *last_wb = NULL; 804 struct bdi_writeback *wb = list_entry(&bdi->wb_list, 805 struct bdi_writeback, bdi_node); 806 807 might_sleep(); 808 restart: 809 rcu_read_lock(); 810 list_for_each_entry_continue_rcu(wb, &bdi->wb_list, bdi_node) { 811 DEFINE_WB_COMPLETION_ONSTACK(fallback_work_done); 812 struct wb_writeback_work fallback_work; 813 struct wb_writeback_work *work; 814 long nr_pages; 815 816 if (last_wb) { 817 wb_put(last_wb); 818 last_wb = NULL; 819 } 820 821 /* SYNC_ALL writes out I_DIRTY_TIME too */ 822 if (!wb_has_dirty_io(wb) && 823 (base_work->sync_mode == WB_SYNC_NONE || 824 list_empty(&wb->b_dirty_time))) 825 continue; 826 if (skip_if_busy && writeback_in_progress(wb)) 827 continue; 828 829 nr_pages = wb_split_bdi_pages(wb, base_work->nr_pages); 830 831 work = kmalloc(sizeof(*work), GFP_ATOMIC); 832 if (work) { 833 *work = *base_work; 834 work->nr_pages = nr_pages; 835 work->auto_free = 1; 836 wb_queue_work(wb, work); 837 continue; 838 } 839 840 /* alloc failed, execute synchronously using on-stack fallback */ 841 work = &fallback_work; 842 *work = *base_work; 843 work->nr_pages = nr_pages; 844 work->auto_free = 0; 845 work->done = &fallback_work_done; 846 847 wb_queue_work(wb, work); 848 849 /* 850 * Pin @wb so that it stays on @bdi->wb_list. This allows 851 * continuing iteration from @wb after dropping and 852 * regrabbing rcu read lock. 853 */ 854 wb_get(wb); 855 last_wb = wb; 856 857 rcu_read_unlock(); 858 wb_wait_for_completion(bdi, &fallback_work_done); 859 goto restart; 860 } 861 rcu_read_unlock(); 862 863 if (last_wb) 864 wb_put(last_wb); 865 } 866 867 /** 868 * cgroup_writeback_umount - flush inode wb switches for umount 869 * 870 * This function is called when a super_block is about to be destroyed and 871 * flushes in-flight inode wb switches. An inode wb switch goes through 872 * RCU and then workqueue, so the two need to be flushed in order to ensure 873 * that all previously scheduled switches are finished. As wb switches are 874 * rare occurrences and synchronize_rcu() can take a while, perform 875 * flushing iff wb switches are in flight. 876 */ 877 void cgroup_writeback_umount(void) 878 { 879 if (atomic_read(&isw_nr_in_flight)) { 880 synchronize_rcu(); 881 flush_workqueue(isw_wq); 882 } 883 } 884 885 static int __init cgroup_writeback_init(void) 886 { 887 isw_wq = alloc_workqueue("inode_switch_wbs", 0, 0); 888 if (!isw_wq) 889 return -ENOMEM; 890 return 0; 891 } 892 fs_initcall(cgroup_writeback_init); 893 894 #else /* CONFIG_CGROUP_WRITEBACK */ 895 896 static struct bdi_writeback * 897 locked_inode_to_wb_and_lock_list(struct inode *inode) 898 __releases(&inode->i_lock) 899 __acquires(&wb->list_lock) 900 { 901 struct bdi_writeback *wb = inode_to_wb(inode); 902 903 spin_unlock(&inode->i_lock); 904 spin_lock(&wb->list_lock); 905 return wb; 906 } 907 908 static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode) 909 __acquires(&wb->list_lock) 910 { 911 struct bdi_writeback *wb = inode_to_wb(inode); 912 913 spin_lock(&wb->list_lock); 914 return wb; 915 } 916 917 static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages) 918 { 919 return nr_pages; 920 } 921 922 static void bdi_split_work_to_wbs(struct backing_dev_info *bdi, 923 struct wb_writeback_work *base_work, 924 bool skip_if_busy) 925 { 926 might_sleep(); 927 928 if (!skip_if_busy || !writeback_in_progress(&bdi->wb)) { 929 base_work->auto_free = 0; 930 wb_queue_work(&bdi->wb, base_work); 931 } 932 } 933 934 #endif /* CONFIG_CGROUP_WRITEBACK */ 935 936 void wb_start_writeback(struct bdi_writeback *wb, long nr_pages, 937 bool range_cyclic, enum wb_reason reason) 938 { 939 struct wb_writeback_work *work; 940 941 if (!wb_has_dirty_io(wb)) 942 return; 943 944 /* 945 * This is WB_SYNC_NONE writeback, so if allocation fails just 946 * wakeup the thread for old dirty data writeback 947 */ 948 work = kzalloc(sizeof(*work), 949 GFP_NOWAIT | __GFP_NOMEMALLOC | __GFP_NOWARN); 950 if (!work) { 951 trace_writeback_nowork(wb); 952 wb_wakeup(wb); 953 return; 954 } 955 956 work->sync_mode = WB_SYNC_NONE; 957 work->nr_pages = nr_pages; 958 work->range_cyclic = range_cyclic; 959 work->reason = reason; 960 work->auto_free = 1; 961 962 wb_queue_work(wb, work); 963 } 964 965 /** 966 * wb_start_background_writeback - start background writeback 967 * @wb: bdi_writback to write from 968 * 969 * Description: 970 * This makes sure WB_SYNC_NONE background writeback happens. When 971 * this function returns, it is only guaranteed that for given wb 972 * some IO is happening if we are over background dirty threshold. 973 * Caller need not hold sb s_umount semaphore. 974 */ 975 void wb_start_background_writeback(struct bdi_writeback *wb) 976 { 977 /* 978 * We just wake up the flusher thread. It will perform background 979 * writeback as soon as there is no other work to do. 980 */ 981 trace_writeback_wake_background(wb); 982 wb_wakeup(wb); 983 } 984 985 /* 986 * Remove the inode from the writeback list it is on. 987 */ 988 void inode_io_list_del(struct inode *inode) 989 { 990 struct bdi_writeback *wb; 991 992 wb = inode_to_wb_and_lock_list(inode); 993 inode_io_list_del_locked(inode, wb); 994 spin_unlock(&wb->list_lock); 995 } 996 997 /* 998 * mark an inode as under writeback on the sb 999 */ 1000 void sb_mark_inode_writeback(struct inode *inode) 1001 { 1002 struct super_block *sb = inode->i_sb; 1003 unsigned long flags; 1004 1005 if (list_empty(&inode->i_wb_list)) { 1006 spin_lock_irqsave(&sb->s_inode_wblist_lock, flags); 1007 if (list_empty(&inode->i_wb_list)) { 1008 list_add_tail(&inode->i_wb_list, &sb->s_inodes_wb); 1009 trace_sb_mark_inode_writeback(inode); 1010 } 1011 spin_unlock_irqrestore(&sb->s_inode_wblist_lock, flags); 1012 } 1013 } 1014 1015 /* 1016 * clear an inode as under writeback on the sb 1017 */ 1018 void sb_clear_inode_writeback(struct inode *inode) 1019 { 1020 struct super_block *sb = inode->i_sb; 1021 unsigned long flags; 1022 1023 if (!list_empty(&inode->i_wb_list)) { 1024 spin_lock_irqsave(&sb->s_inode_wblist_lock, flags); 1025 if (!list_empty(&inode->i_wb_list)) { 1026 list_del_init(&inode->i_wb_list); 1027 trace_sb_clear_inode_writeback(inode); 1028 } 1029 spin_unlock_irqrestore(&sb->s_inode_wblist_lock, flags); 1030 } 1031 } 1032 1033 /* 1034 * Redirty an inode: set its when-it-was dirtied timestamp and move it to the 1035 * furthest end of its superblock's dirty-inode list. 1036 * 1037 * Before stamping the inode's ->dirtied_when, we check to see whether it is 1038 * already the most-recently-dirtied inode on the b_dirty list. If that is 1039 * the case then the inode must have been redirtied while it was being written 1040 * out and we don't reset its dirtied_when. 1041 */ 1042 static void redirty_tail(struct inode *inode, struct bdi_writeback *wb) 1043 { 1044 if (!list_empty(&wb->b_dirty)) { 1045 struct inode *tail; 1046 1047 tail = wb_inode(wb->b_dirty.next); 1048 if (time_before(inode->dirtied_when, tail->dirtied_when)) 1049 inode->dirtied_when = jiffies; 1050 } 1051 inode_io_list_move_locked(inode, wb, &wb->b_dirty); 1052 } 1053 1054 /* 1055 * requeue inode for re-scanning after bdi->b_io list is exhausted. 1056 */ 1057 static void requeue_io(struct inode *inode, struct bdi_writeback *wb) 1058 { 1059 inode_io_list_move_locked(inode, wb, &wb->b_more_io); 1060 } 1061 1062 static void inode_sync_complete(struct inode *inode) 1063 { 1064 inode->i_state &= ~I_SYNC; 1065 /* If inode is clean an unused, put it into LRU now... */ 1066 inode_add_lru(inode); 1067 /* Waiters must see I_SYNC cleared before being woken up */ 1068 smp_mb(); 1069 wake_up_bit(&inode->i_state, __I_SYNC); 1070 } 1071 1072 static bool inode_dirtied_after(struct inode *inode, unsigned long t) 1073 { 1074 bool ret = time_after(inode->dirtied_when, t); 1075 #ifndef CONFIG_64BIT 1076 /* 1077 * For inodes being constantly redirtied, dirtied_when can get stuck. 1078 * It _appears_ to be in the future, but is actually in distant past. 1079 * This test is necessary to prevent such wrapped-around relative times 1080 * from permanently stopping the whole bdi writeback. 1081 */ 1082 ret = ret && time_before_eq(inode->dirtied_when, jiffies); 1083 #endif 1084 return ret; 1085 } 1086 1087 #define EXPIRE_DIRTY_ATIME 0x0001 1088 1089 /* 1090 * Move expired (dirtied before work->older_than_this) dirty inodes from 1091 * @delaying_queue to @dispatch_queue. 1092 */ 1093 static int move_expired_inodes(struct list_head *delaying_queue, 1094 struct list_head *dispatch_queue, 1095 int flags, 1096 struct wb_writeback_work *work) 1097 { 1098 unsigned long *older_than_this = NULL; 1099 unsigned long expire_time; 1100 LIST_HEAD(tmp); 1101 struct list_head *pos, *node; 1102 struct super_block *sb = NULL; 1103 struct inode *inode; 1104 int do_sb_sort = 0; 1105 int moved = 0; 1106 1107 if ((flags & EXPIRE_DIRTY_ATIME) == 0) 1108 older_than_this = work->older_than_this; 1109 else if (!work->for_sync) { 1110 expire_time = jiffies - (dirtytime_expire_interval * HZ); 1111 older_than_this = &expire_time; 1112 } 1113 while (!list_empty(delaying_queue)) { 1114 inode = wb_inode(delaying_queue->prev); 1115 if (older_than_this && 1116 inode_dirtied_after(inode, *older_than_this)) 1117 break; 1118 list_move(&inode->i_io_list, &tmp); 1119 moved++; 1120 if (flags & EXPIRE_DIRTY_ATIME) 1121 set_bit(__I_DIRTY_TIME_EXPIRED, &inode->i_state); 1122 if (sb_is_blkdev_sb(inode->i_sb)) 1123 continue; 1124 if (sb && sb != inode->i_sb) 1125 do_sb_sort = 1; 1126 sb = inode->i_sb; 1127 } 1128 1129 /* just one sb in list, splice to dispatch_queue and we're done */ 1130 if (!do_sb_sort) { 1131 list_splice(&tmp, dispatch_queue); 1132 goto out; 1133 } 1134 1135 /* Move inodes from one superblock together */ 1136 while (!list_empty(&tmp)) { 1137 sb = wb_inode(tmp.prev)->i_sb; 1138 list_for_each_prev_safe(pos, node, &tmp) { 1139 inode = wb_inode(pos); 1140 if (inode->i_sb == sb) 1141 list_move(&inode->i_io_list, dispatch_queue); 1142 } 1143 } 1144 out: 1145 return moved; 1146 } 1147 1148 /* 1149 * Queue all expired dirty inodes for io, eldest first. 1150 * Before 1151 * newly dirtied b_dirty b_io b_more_io 1152 * =============> gf edc BA 1153 * After 1154 * newly dirtied b_dirty b_io b_more_io 1155 * =============> g fBAedc 1156 * | 1157 * +--> dequeue for IO 1158 */ 1159 static void queue_io(struct bdi_writeback *wb, struct wb_writeback_work *work) 1160 { 1161 int moved; 1162 1163 assert_spin_locked(&wb->list_lock); 1164 list_splice_init(&wb->b_more_io, &wb->b_io); 1165 moved = move_expired_inodes(&wb->b_dirty, &wb->b_io, 0, work); 1166 moved += move_expired_inodes(&wb->b_dirty_time, &wb->b_io, 1167 EXPIRE_DIRTY_ATIME, work); 1168 if (moved) 1169 wb_io_lists_populated(wb); 1170 trace_writeback_queue_io(wb, work, moved); 1171 } 1172 1173 static int write_inode(struct inode *inode, struct writeback_control *wbc) 1174 { 1175 int ret; 1176 1177 if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode)) { 1178 trace_writeback_write_inode_start(inode, wbc); 1179 ret = inode->i_sb->s_op->write_inode(inode, wbc); 1180 trace_writeback_write_inode(inode, wbc); 1181 return ret; 1182 } 1183 return 0; 1184 } 1185 1186 /* 1187 * Wait for writeback on an inode to complete. Called with i_lock held. 1188 * Caller must make sure inode cannot go away when we drop i_lock. 1189 */ 1190 static void __inode_wait_for_writeback(struct inode *inode) 1191 __releases(inode->i_lock) 1192 __acquires(inode->i_lock) 1193 { 1194 DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC); 1195 wait_queue_head_t *wqh; 1196 1197 wqh = bit_waitqueue(&inode->i_state, __I_SYNC); 1198 while (inode->i_state & I_SYNC) { 1199 spin_unlock(&inode->i_lock); 1200 __wait_on_bit(wqh, &wq, bit_wait, 1201 TASK_UNINTERRUPTIBLE); 1202 spin_lock(&inode->i_lock); 1203 } 1204 } 1205 1206 /* 1207 * Wait for writeback on an inode to complete. Caller must have inode pinned. 1208 */ 1209 void inode_wait_for_writeback(struct inode *inode) 1210 { 1211 spin_lock(&inode->i_lock); 1212 __inode_wait_for_writeback(inode); 1213 spin_unlock(&inode->i_lock); 1214 } 1215 1216 /* 1217 * Sleep until I_SYNC is cleared. This function must be called with i_lock 1218 * held and drops it. It is aimed for callers not holding any inode reference 1219 * so once i_lock is dropped, inode can go away. 1220 */ 1221 static void inode_sleep_on_writeback(struct inode *inode) 1222 __releases(inode->i_lock) 1223 { 1224 DEFINE_WAIT(wait); 1225 wait_queue_head_t *wqh = bit_waitqueue(&inode->i_state, __I_SYNC); 1226 int sleep; 1227 1228 prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE); 1229 sleep = inode->i_state & I_SYNC; 1230 spin_unlock(&inode->i_lock); 1231 if (sleep) 1232 schedule(); 1233 finish_wait(wqh, &wait); 1234 } 1235 1236 /* 1237 * Find proper writeback list for the inode depending on its current state and 1238 * possibly also change of its state while we were doing writeback. Here we 1239 * handle things such as livelock prevention or fairness of writeback among 1240 * inodes. This function can be called only by flusher thread - noone else 1241 * processes all inodes in writeback lists and requeueing inodes behind flusher 1242 * thread's back can have unexpected consequences. 1243 */ 1244 static void requeue_inode(struct inode *inode, struct bdi_writeback *wb, 1245 struct writeback_control *wbc) 1246 { 1247 if (inode->i_state & I_FREEING) 1248 return; 1249 1250 /* 1251 * Sync livelock prevention. Each inode is tagged and synced in one 1252 * shot. If still dirty, it will be redirty_tail()'ed below. Update 1253 * the dirty time to prevent enqueue and sync it again. 1254 */ 1255 if ((inode->i_state & I_DIRTY) && 1256 (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)) 1257 inode->dirtied_when = jiffies; 1258 1259 if (wbc->pages_skipped) { 1260 /* 1261 * writeback is not making progress due to locked 1262 * buffers. Skip this inode for now. 1263 */ 1264 redirty_tail(inode, wb); 1265 return; 1266 } 1267 1268 if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_DIRTY)) { 1269 /* 1270 * We didn't write back all the pages. nfs_writepages() 1271 * sometimes bales out without doing anything. 1272 */ 1273 if (wbc->nr_to_write <= 0) { 1274 /* Slice used up. Queue for next turn. */ 1275 requeue_io(inode, wb); 1276 } else { 1277 /* 1278 * Writeback blocked by something other than 1279 * congestion. Delay the inode for some time to 1280 * avoid spinning on the CPU (100% iowait) 1281 * retrying writeback of the dirty page/inode 1282 * that cannot be performed immediately. 1283 */ 1284 redirty_tail(inode, wb); 1285 } 1286 } else if (inode->i_state & I_DIRTY) { 1287 /* 1288 * Filesystems can dirty the inode during writeback operations, 1289 * such as delayed allocation during submission or metadata 1290 * updates after data IO completion. 1291 */ 1292 redirty_tail(inode, wb); 1293 } else if (inode->i_state & I_DIRTY_TIME) { 1294 inode->dirtied_when = jiffies; 1295 inode_io_list_move_locked(inode, wb, &wb->b_dirty_time); 1296 } else { 1297 /* The inode is clean. Remove from writeback lists. */ 1298 inode_io_list_del_locked(inode, wb); 1299 } 1300 } 1301 1302 /* 1303 * Write out an inode and its dirty pages. Do not update the writeback list 1304 * linkage. That is left to the caller. The caller is also responsible for 1305 * setting I_SYNC flag and calling inode_sync_complete() to clear it. 1306 */ 1307 static int 1308 __writeback_single_inode(struct inode *inode, struct writeback_control *wbc) 1309 { 1310 struct address_space *mapping = inode->i_mapping; 1311 long nr_to_write = wbc->nr_to_write; 1312 unsigned dirty; 1313 int ret; 1314 1315 WARN_ON(!(inode->i_state & I_SYNC)); 1316 1317 trace_writeback_single_inode_start(inode, wbc, nr_to_write); 1318 1319 ret = do_writepages(mapping, wbc); 1320 1321 /* 1322 * Make sure to wait on the data before writing out the metadata. 1323 * This is important for filesystems that modify metadata on data 1324 * I/O completion. We don't do it for sync(2) writeback because it has a 1325 * separate, external IO completion path and ->sync_fs for guaranteeing 1326 * inode metadata is written back correctly. 1327 */ 1328 if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync) { 1329 int err = filemap_fdatawait(mapping); 1330 if (ret == 0) 1331 ret = err; 1332 } 1333 1334 /* 1335 * Some filesystems may redirty the inode during the writeback 1336 * due to delalloc, clear dirty metadata flags right before 1337 * write_inode() 1338 */ 1339 spin_lock(&inode->i_lock); 1340 1341 dirty = inode->i_state & I_DIRTY; 1342 if (inode->i_state & I_DIRTY_TIME) { 1343 if ((dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) || 1344 wbc->sync_mode == WB_SYNC_ALL || 1345 unlikely(inode->i_state & I_DIRTY_TIME_EXPIRED) || 1346 unlikely(time_after(jiffies, 1347 (inode->dirtied_time_when + 1348 dirtytime_expire_interval * HZ)))) { 1349 dirty |= I_DIRTY_TIME | I_DIRTY_TIME_EXPIRED; 1350 trace_writeback_lazytime(inode); 1351 } 1352 } else 1353 inode->i_state &= ~I_DIRTY_TIME_EXPIRED; 1354 inode->i_state &= ~dirty; 1355 1356 /* 1357 * Paired with smp_mb() in __mark_inode_dirty(). This allows 1358 * __mark_inode_dirty() to test i_state without grabbing i_lock - 1359 * either they see the I_DIRTY bits cleared or we see the dirtied 1360 * inode. 1361 * 1362 * I_DIRTY_PAGES is always cleared together above even if @mapping 1363 * still has dirty pages. The flag is reinstated after smp_mb() if 1364 * necessary. This guarantees that either __mark_inode_dirty() 1365 * sees clear I_DIRTY_PAGES or we see PAGECACHE_TAG_DIRTY. 1366 */ 1367 smp_mb(); 1368 1369 if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) 1370 inode->i_state |= I_DIRTY_PAGES; 1371 1372 spin_unlock(&inode->i_lock); 1373 1374 if (dirty & I_DIRTY_TIME) 1375 mark_inode_dirty_sync(inode); 1376 /* Don't write the inode if only I_DIRTY_PAGES was set */ 1377 if (dirty & ~I_DIRTY_PAGES) { 1378 int err = write_inode(inode, wbc); 1379 if (ret == 0) 1380 ret = err; 1381 } 1382 trace_writeback_single_inode(inode, wbc, nr_to_write); 1383 return ret; 1384 } 1385 1386 /* 1387 * Write out an inode's dirty pages. Either the caller has an active reference 1388 * on the inode or the inode has I_WILL_FREE set. 1389 * 1390 * This function is designed to be called for writing back one inode which 1391 * we go e.g. from filesystem. Flusher thread uses __writeback_single_inode() 1392 * and does more profound writeback list handling in writeback_sb_inodes(). 1393 */ 1394 static int writeback_single_inode(struct inode *inode, 1395 struct writeback_control *wbc) 1396 { 1397 struct bdi_writeback *wb; 1398 int ret = 0; 1399 1400 spin_lock(&inode->i_lock); 1401 if (!atomic_read(&inode->i_count)) 1402 WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING))); 1403 else 1404 WARN_ON(inode->i_state & I_WILL_FREE); 1405 1406 if (inode->i_state & I_SYNC) { 1407 if (wbc->sync_mode != WB_SYNC_ALL) 1408 goto out; 1409 /* 1410 * It's a data-integrity sync. We must wait. Since callers hold 1411 * inode reference or inode has I_WILL_FREE set, it cannot go 1412 * away under us. 1413 */ 1414 __inode_wait_for_writeback(inode); 1415 } 1416 WARN_ON(inode->i_state & I_SYNC); 1417 /* 1418 * Skip inode if it is clean and we have no outstanding writeback in 1419 * WB_SYNC_ALL mode. We don't want to mess with writeback lists in this 1420 * function since flusher thread may be doing for example sync in 1421 * parallel and if we move the inode, it could get skipped. So here we 1422 * make sure inode is on some writeback list and leave it there unless 1423 * we have completely cleaned the inode. 1424 */ 1425 if (!(inode->i_state & I_DIRTY_ALL) && 1426 (wbc->sync_mode != WB_SYNC_ALL || 1427 !mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK))) 1428 goto out; 1429 inode->i_state |= I_SYNC; 1430 wbc_attach_and_unlock_inode(wbc, inode); 1431 1432 ret = __writeback_single_inode(inode, wbc); 1433 1434 wbc_detach_inode(wbc); 1435 1436 wb = inode_to_wb_and_lock_list(inode); 1437 spin_lock(&inode->i_lock); 1438 /* 1439 * If inode is clean, remove it from writeback lists. Otherwise don't 1440 * touch it. See comment above for explanation. 1441 */ 1442 if (!(inode->i_state & I_DIRTY_ALL)) 1443 inode_io_list_del_locked(inode, wb); 1444 spin_unlock(&wb->list_lock); 1445 inode_sync_complete(inode); 1446 out: 1447 spin_unlock(&inode->i_lock); 1448 return ret; 1449 } 1450 1451 static long writeback_chunk_size(struct bdi_writeback *wb, 1452 struct wb_writeback_work *work) 1453 { 1454 long pages; 1455 1456 /* 1457 * WB_SYNC_ALL mode does livelock avoidance by syncing dirty 1458 * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX 1459 * here avoids calling into writeback_inodes_wb() more than once. 1460 * 1461 * The intended call sequence for WB_SYNC_ALL writeback is: 1462 * 1463 * wb_writeback() 1464 * writeback_sb_inodes() <== called only once 1465 * write_cache_pages() <== called once for each inode 1466 * (quickly) tag currently dirty pages 1467 * (maybe slowly) sync all tagged pages 1468 */ 1469 if (work->sync_mode == WB_SYNC_ALL || work->tagged_writepages) 1470 pages = LONG_MAX; 1471 else { 1472 pages = min(wb->avg_write_bandwidth / 2, 1473 global_wb_domain.dirty_limit / DIRTY_SCOPE); 1474 pages = min(pages, work->nr_pages); 1475 pages = round_down(pages + MIN_WRITEBACK_PAGES, 1476 MIN_WRITEBACK_PAGES); 1477 } 1478 1479 return pages; 1480 } 1481 1482 /* 1483 * Write a portion of b_io inodes which belong to @sb. 1484 * 1485 * Return the number of pages and/or inodes written. 1486 * 1487 * NOTE! This is called with wb->list_lock held, and will 1488 * unlock and relock that for each inode it ends up doing 1489 * IO for. 1490 */ 1491 static long writeback_sb_inodes(struct super_block *sb, 1492 struct bdi_writeback *wb, 1493 struct wb_writeback_work *work) 1494 { 1495 struct writeback_control wbc = { 1496 .sync_mode = work->sync_mode, 1497 .tagged_writepages = work->tagged_writepages, 1498 .for_kupdate = work->for_kupdate, 1499 .for_background = work->for_background, 1500 .for_sync = work->for_sync, 1501 .range_cyclic = work->range_cyclic, 1502 .range_start = 0, 1503 .range_end = LLONG_MAX, 1504 }; 1505 unsigned long start_time = jiffies; 1506 long write_chunk; 1507 long wrote = 0; /* count both pages and inodes */ 1508 1509 while (!list_empty(&wb->b_io)) { 1510 struct inode *inode = wb_inode(wb->b_io.prev); 1511 struct bdi_writeback *tmp_wb; 1512 1513 if (inode->i_sb != sb) { 1514 if (work->sb) { 1515 /* 1516 * We only want to write back data for this 1517 * superblock, move all inodes not belonging 1518 * to it back onto the dirty list. 1519 */ 1520 redirty_tail(inode, wb); 1521 continue; 1522 } 1523 1524 /* 1525 * The inode belongs to a different superblock. 1526 * Bounce back to the caller to unpin this and 1527 * pin the next superblock. 1528 */ 1529 break; 1530 } 1531 1532 /* 1533 * Don't bother with new inodes or inodes being freed, first 1534 * kind does not need periodic writeout yet, and for the latter 1535 * kind writeout is handled by the freer. 1536 */ 1537 spin_lock(&inode->i_lock); 1538 if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) { 1539 spin_unlock(&inode->i_lock); 1540 redirty_tail(inode, wb); 1541 continue; 1542 } 1543 if ((inode->i_state & I_SYNC) && wbc.sync_mode != WB_SYNC_ALL) { 1544 /* 1545 * If this inode is locked for writeback and we are not 1546 * doing writeback-for-data-integrity, move it to 1547 * b_more_io so that writeback can proceed with the 1548 * other inodes on s_io. 1549 * 1550 * We'll have another go at writing back this inode 1551 * when we completed a full scan of b_io. 1552 */ 1553 spin_unlock(&inode->i_lock); 1554 requeue_io(inode, wb); 1555 trace_writeback_sb_inodes_requeue(inode); 1556 continue; 1557 } 1558 spin_unlock(&wb->list_lock); 1559 1560 /* 1561 * We already requeued the inode if it had I_SYNC set and we 1562 * are doing WB_SYNC_NONE writeback. So this catches only the 1563 * WB_SYNC_ALL case. 1564 */ 1565 if (inode->i_state & I_SYNC) { 1566 /* Wait for I_SYNC. This function drops i_lock... */ 1567 inode_sleep_on_writeback(inode); 1568 /* Inode may be gone, start again */ 1569 spin_lock(&wb->list_lock); 1570 continue; 1571 } 1572 inode->i_state |= I_SYNC; 1573 wbc_attach_and_unlock_inode(&wbc, inode); 1574 1575 write_chunk = writeback_chunk_size(wb, work); 1576 wbc.nr_to_write = write_chunk; 1577 wbc.pages_skipped = 0; 1578 1579 /* 1580 * We use I_SYNC to pin the inode in memory. While it is set 1581 * evict_inode() will wait so the inode cannot be freed. 1582 */ 1583 __writeback_single_inode(inode, &wbc); 1584 1585 wbc_detach_inode(&wbc); 1586 work->nr_pages -= write_chunk - wbc.nr_to_write; 1587 wrote += write_chunk - wbc.nr_to_write; 1588 1589 if (need_resched()) { 1590 /* 1591 * We're trying to balance between building up a nice 1592 * long list of IOs to improve our merge rate, and 1593 * getting those IOs out quickly for anyone throttling 1594 * in balance_dirty_pages(). cond_resched() doesn't 1595 * unplug, so get our IOs out the door before we 1596 * give up the CPU. 1597 */ 1598 blk_flush_plug(current); 1599 cond_resched(); 1600 } 1601 1602 /* 1603 * Requeue @inode if still dirty. Be careful as @inode may 1604 * have been switched to another wb in the meantime. 1605 */ 1606 tmp_wb = inode_to_wb_and_lock_list(inode); 1607 spin_lock(&inode->i_lock); 1608 if (!(inode->i_state & I_DIRTY_ALL)) 1609 wrote++; 1610 requeue_inode(inode, tmp_wb, &wbc); 1611 inode_sync_complete(inode); 1612 spin_unlock(&inode->i_lock); 1613 1614 if (unlikely(tmp_wb != wb)) { 1615 spin_unlock(&tmp_wb->list_lock); 1616 spin_lock(&wb->list_lock); 1617 } 1618 1619 /* 1620 * bail out to wb_writeback() often enough to check 1621 * background threshold and other termination conditions. 1622 */ 1623 if (wrote) { 1624 if (time_is_before_jiffies(start_time + HZ / 10UL)) 1625 break; 1626 if (work->nr_pages <= 0) 1627 break; 1628 } 1629 } 1630 return wrote; 1631 } 1632 1633 static long __writeback_inodes_wb(struct bdi_writeback *wb, 1634 struct wb_writeback_work *work) 1635 { 1636 unsigned long start_time = jiffies; 1637 long wrote = 0; 1638 1639 while (!list_empty(&wb->b_io)) { 1640 struct inode *inode = wb_inode(wb->b_io.prev); 1641 struct super_block *sb = inode->i_sb; 1642 1643 if (!trylock_super(sb)) { 1644 /* 1645 * trylock_super() may fail consistently due to 1646 * s_umount being grabbed by someone else. Don't use 1647 * requeue_io() to avoid busy retrying the inode/sb. 1648 */ 1649 redirty_tail(inode, wb); 1650 continue; 1651 } 1652 wrote += writeback_sb_inodes(sb, wb, work); 1653 up_read(&sb->s_umount); 1654 1655 /* refer to the same tests at the end of writeback_sb_inodes */ 1656 if (wrote) { 1657 if (time_is_before_jiffies(start_time + HZ / 10UL)) 1658 break; 1659 if (work->nr_pages <= 0) 1660 break; 1661 } 1662 } 1663 /* Leave any unwritten inodes on b_io */ 1664 return wrote; 1665 } 1666 1667 static long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages, 1668 enum wb_reason reason) 1669 { 1670 struct wb_writeback_work work = { 1671 .nr_pages = nr_pages, 1672 .sync_mode = WB_SYNC_NONE, 1673 .range_cyclic = 1, 1674 .reason = reason, 1675 }; 1676 struct blk_plug plug; 1677 1678 blk_start_plug(&plug); 1679 spin_lock(&wb->list_lock); 1680 if (list_empty(&wb->b_io)) 1681 queue_io(wb, &work); 1682 __writeback_inodes_wb(wb, &work); 1683 spin_unlock(&wb->list_lock); 1684 blk_finish_plug(&plug); 1685 1686 return nr_pages - work.nr_pages; 1687 } 1688 1689 /* 1690 * Explicit flushing or periodic writeback of "old" data. 1691 * 1692 * Define "old": the first time one of an inode's pages is dirtied, we mark the 1693 * dirtying-time in the inode's address_space. So this periodic writeback code 1694 * just walks the superblock inode list, writing back any inodes which are 1695 * older than a specific point in time. 1696 * 1697 * Try to run once per dirty_writeback_interval. But if a writeback event 1698 * takes longer than a dirty_writeback_interval interval, then leave a 1699 * one-second gap. 1700 * 1701 * older_than_this takes precedence over nr_to_write. So we'll only write back 1702 * all dirty pages if they are all attached to "old" mappings. 1703 */ 1704 static long wb_writeback(struct bdi_writeback *wb, 1705 struct wb_writeback_work *work) 1706 { 1707 unsigned long wb_start = jiffies; 1708 long nr_pages = work->nr_pages; 1709 unsigned long oldest_jif; 1710 struct inode *inode; 1711 long progress; 1712 struct blk_plug plug; 1713 1714 oldest_jif = jiffies; 1715 work->older_than_this = &oldest_jif; 1716 1717 blk_start_plug(&plug); 1718 spin_lock(&wb->list_lock); 1719 for (;;) { 1720 /* 1721 * Stop writeback when nr_pages has been consumed 1722 */ 1723 if (work->nr_pages <= 0) 1724 break; 1725 1726 /* 1727 * Background writeout and kupdate-style writeback may 1728 * run forever. Stop them if there is other work to do 1729 * so that e.g. sync can proceed. They'll be restarted 1730 * after the other works are all done. 1731 */ 1732 if ((work->for_background || work->for_kupdate) && 1733 !list_empty(&wb->work_list)) 1734 break; 1735 1736 /* 1737 * For background writeout, stop when we are below the 1738 * background dirty threshold 1739 */ 1740 if (work->for_background && !wb_over_bg_thresh(wb)) 1741 break; 1742 1743 /* 1744 * Kupdate and background works are special and we want to 1745 * include all inodes that need writing. Livelock avoidance is 1746 * handled by these works yielding to any other work so we are 1747 * safe. 1748 */ 1749 if (work->for_kupdate) { 1750 oldest_jif = jiffies - 1751 msecs_to_jiffies(dirty_expire_interval * 10); 1752 } else if (work->for_background) 1753 oldest_jif = jiffies; 1754 1755 trace_writeback_start(wb, work); 1756 if (list_empty(&wb->b_io)) 1757 queue_io(wb, work); 1758 if (work->sb) 1759 progress = writeback_sb_inodes(work->sb, wb, work); 1760 else 1761 progress = __writeback_inodes_wb(wb, work); 1762 trace_writeback_written(wb, work); 1763 1764 wb_update_bandwidth(wb, wb_start); 1765 1766 /* 1767 * Did we write something? Try for more 1768 * 1769 * Dirty inodes are moved to b_io for writeback in batches. 1770 * The completion of the current batch does not necessarily 1771 * mean the overall work is done. So we keep looping as long 1772 * as made some progress on cleaning pages or inodes. 1773 */ 1774 if (progress) 1775 continue; 1776 /* 1777 * No more inodes for IO, bail 1778 */ 1779 if (list_empty(&wb->b_more_io)) 1780 break; 1781 /* 1782 * Nothing written. Wait for some inode to 1783 * become available for writeback. Otherwise 1784 * we'll just busyloop. 1785 */ 1786 trace_writeback_wait(wb, work); 1787 inode = wb_inode(wb->b_more_io.prev); 1788 spin_lock(&inode->i_lock); 1789 spin_unlock(&wb->list_lock); 1790 /* This function drops i_lock... */ 1791 inode_sleep_on_writeback(inode); 1792 spin_lock(&wb->list_lock); 1793 } 1794 spin_unlock(&wb->list_lock); 1795 blk_finish_plug(&plug); 1796 1797 return nr_pages - work->nr_pages; 1798 } 1799 1800 /* 1801 * Return the next wb_writeback_work struct that hasn't been processed yet. 1802 */ 1803 static struct wb_writeback_work *get_next_work_item(struct bdi_writeback *wb) 1804 { 1805 struct wb_writeback_work *work = NULL; 1806 1807 spin_lock_bh(&wb->work_lock); 1808 if (!list_empty(&wb->work_list)) { 1809 work = list_entry(wb->work_list.next, 1810 struct wb_writeback_work, list); 1811 list_del_init(&work->list); 1812 } 1813 spin_unlock_bh(&wb->work_lock); 1814 return work; 1815 } 1816 1817 /* 1818 * Add in the number of potentially dirty inodes, because each inode 1819 * write can dirty pagecache in the underlying blockdev. 1820 */ 1821 static unsigned long get_nr_dirty_pages(void) 1822 { 1823 return global_node_page_state(NR_FILE_DIRTY) + 1824 global_node_page_state(NR_UNSTABLE_NFS) + 1825 get_nr_dirty_inodes(); 1826 } 1827 1828 static long wb_check_background_flush(struct bdi_writeback *wb) 1829 { 1830 if (wb_over_bg_thresh(wb)) { 1831 1832 struct wb_writeback_work work = { 1833 .nr_pages = LONG_MAX, 1834 .sync_mode = WB_SYNC_NONE, 1835 .for_background = 1, 1836 .range_cyclic = 1, 1837 .reason = WB_REASON_BACKGROUND, 1838 }; 1839 1840 return wb_writeback(wb, &work); 1841 } 1842 1843 return 0; 1844 } 1845 1846 static long wb_check_old_data_flush(struct bdi_writeback *wb) 1847 { 1848 unsigned long expired; 1849 long nr_pages; 1850 1851 /* 1852 * When set to zero, disable periodic writeback 1853 */ 1854 if (!dirty_writeback_interval) 1855 return 0; 1856 1857 expired = wb->last_old_flush + 1858 msecs_to_jiffies(dirty_writeback_interval * 10); 1859 if (time_before(jiffies, expired)) 1860 return 0; 1861 1862 wb->last_old_flush = jiffies; 1863 nr_pages = get_nr_dirty_pages(); 1864 1865 if (nr_pages) { 1866 struct wb_writeback_work work = { 1867 .nr_pages = nr_pages, 1868 .sync_mode = WB_SYNC_NONE, 1869 .for_kupdate = 1, 1870 .range_cyclic = 1, 1871 .reason = WB_REASON_PERIODIC, 1872 }; 1873 1874 return wb_writeback(wb, &work); 1875 } 1876 1877 return 0; 1878 } 1879 1880 /* 1881 * Retrieve work items and do the writeback they describe 1882 */ 1883 static long wb_do_writeback(struct bdi_writeback *wb) 1884 { 1885 struct wb_writeback_work *work; 1886 long wrote = 0; 1887 1888 set_bit(WB_writeback_running, &wb->state); 1889 while ((work = get_next_work_item(wb)) != NULL) { 1890 trace_writeback_exec(wb, work); 1891 wrote += wb_writeback(wb, work); 1892 finish_writeback_work(wb, work); 1893 } 1894 1895 /* 1896 * Check for periodic writeback, kupdated() style 1897 */ 1898 wrote += wb_check_old_data_flush(wb); 1899 wrote += wb_check_background_flush(wb); 1900 clear_bit(WB_writeback_running, &wb->state); 1901 1902 return wrote; 1903 } 1904 1905 /* 1906 * Handle writeback of dirty data for the device backed by this bdi. Also 1907 * reschedules periodically and does kupdated style flushing. 1908 */ 1909 void wb_workfn(struct work_struct *work) 1910 { 1911 struct bdi_writeback *wb = container_of(to_delayed_work(work), 1912 struct bdi_writeback, dwork); 1913 long pages_written; 1914 1915 set_worker_desc("flush-%s", dev_name(wb->bdi->dev)); 1916 current->flags |= PF_SWAPWRITE; 1917 1918 if (likely(!current_is_workqueue_rescuer() || 1919 !test_bit(WB_registered, &wb->state))) { 1920 /* 1921 * The normal path. Keep writing back @wb until its 1922 * work_list is empty. Note that this path is also taken 1923 * if @wb is shutting down even when we're running off the 1924 * rescuer as work_list needs to be drained. 1925 */ 1926 do { 1927 pages_written = wb_do_writeback(wb); 1928 trace_writeback_pages_written(pages_written); 1929 } while (!list_empty(&wb->work_list)); 1930 } else { 1931 /* 1932 * bdi_wq can't get enough workers and we're running off 1933 * the emergency worker. Don't hog it. Hopefully, 1024 is 1934 * enough for efficient IO. 1935 */ 1936 pages_written = writeback_inodes_wb(wb, 1024, 1937 WB_REASON_FORKER_THREAD); 1938 trace_writeback_pages_written(pages_written); 1939 } 1940 1941 if (!list_empty(&wb->work_list)) 1942 mod_delayed_work(bdi_wq, &wb->dwork, 0); 1943 else if (wb_has_dirty_io(wb) && dirty_writeback_interval) 1944 wb_wakeup_delayed(wb); 1945 1946 current->flags &= ~PF_SWAPWRITE; 1947 } 1948 1949 /* 1950 * Start writeback of `nr_pages' pages. If `nr_pages' is zero, write back 1951 * the whole world. 1952 */ 1953 void wakeup_flusher_threads(long nr_pages, enum wb_reason reason) 1954 { 1955 struct backing_dev_info *bdi; 1956 1957 /* 1958 * If we are expecting writeback progress we must submit plugged IO. 1959 */ 1960 if (blk_needs_flush_plug(current)) 1961 blk_schedule_flush_plug(current); 1962 1963 if (!nr_pages) 1964 nr_pages = get_nr_dirty_pages(); 1965 1966 rcu_read_lock(); 1967 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) { 1968 struct bdi_writeback *wb; 1969 1970 if (!bdi_has_dirty_io(bdi)) 1971 continue; 1972 1973 list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node) 1974 wb_start_writeback(wb, wb_split_bdi_pages(wb, nr_pages), 1975 false, reason); 1976 } 1977 rcu_read_unlock(); 1978 } 1979 1980 /* 1981 * Wake up bdi's periodically to make sure dirtytime inodes gets 1982 * written back periodically. We deliberately do *not* check the 1983 * b_dirtytime list in wb_has_dirty_io(), since this would cause the 1984 * kernel to be constantly waking up once there are any dirtytime 1985 * inodes on the system. So instead we define a separate delayed work 1986 * function which gets called much more rarely. (By default, only 1987 * once every 12 hours.) 1988 * 1989 * If there is any other write activity going on in the file system, 1990 * this function won't be necessary. But if the only thing that has 1991 * happened on the file system is a dirtytime inode caused by an atime 1992 * update, we need this infrastructure below to make sure that inode 1993 * eventually gets pushed out to disk. 1994 */ 1995 static void wakeup_dirtytime_writeback(struct work_struct *w); 1996 static DECLARE_DELAYED_WORK(dirtytime_work, wakeup_dirtytime_writeback); 1997 1998 static void wakeup_dirtytime_writeback(struct work_struct *w) 1999 { 2000 struct backing_dev_info *bdi; 2001 2002 rcu_read_lock(); 2003 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) { 2004 struct bdi_writeback *wb; 2005 2006 list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node) 2007 if (!list_empty(&wb->b_dirty_time)) 2008 wb_wakeup(wb); 2009 } 2010 rcu_read_unlock(); 2011 schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ); 2012 } 2013 2014 static int __init start_dirtytime_writeback(void) 2015 { 2016 schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ); 2017 return 0; 2018 } 2019 __initcall(start_dirtytime_writeback); 2020 2021 int dirtytime_interval_handler(struct ctl_table *table, int write, 2022 void __user *buffer, size_t *lenp, loff_t *ppos) 2023 { 2024 int ret; 2025 2026 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); 2027 if (ret == 0 && write) 2028 mod_delayed_work(system_wq, &dirtytime_work, 0); 2029 return ret; 2030 } 2031 2032 static noinline void block_dump___mark_inode_dirty(struct inode *inode) 2033 { 2034 if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) { 2035 struct dentry *dentry; 2036 const char *name = "?"; 2037 2038 dentry = d_find_alias(inode); 2039 if (dentry) { 2040 spin_lock(&dentry->d_lock); 2041 name = (const char *) dentry->d_name.name; 2042 } 2043 printk(KERN_DEBUG 2044 "%s(%d): dirtied inode %lu (%s) on %s\n", 2045 current->comm, task_pid_nr(current), inode->i_ino, 2046 name, inode->i_sb->s_id); 2047 if (dentry) { 2048 spin_unlock(&dentry->d_lock); 2049 dput(dentry); 2050 } 2051 } 2052 } 2053 2054 /** 2055 * __mark_inode_dirty - internal function 2056 * 2057 * @inode: inode to mark 2058 * @flags: what kind of dirty (i.e. I_DIRTY_SYNC) 2059 * 2060 * Mark an inode as dirty. Callers should use mark_inode_dirty or 2061 * mark_inode_dirty_sync. 2062 * 2063 * Put the inode on the super block's dirty list. 2064 * 2065 * CAREFUL! We mark it dirty unconditionally, but move it onto the 2066 * dirty list only if it is hashed or if it refers to a blockdev. 2067 * If it was not hashed, it will never be added to the dirty list 2068 * even if it is later hashed, as it will have been marked dirty already. 2069 * 2070 * In short, make sure you hash any inodes _before_ you start marking 2071 * them dirty. 2072 * 2073 * Note that for blockdevs, inode->dirtied_when represents the dirtying time of 2074 * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of 2075 * the kernel-internal blockdev inode represents the dirtying time of the 2076 * blockdev's pages. This is why for I_DIRTY_PAGES we always use 2077 * page->mapping->host, so the page-dirtying time is recorded in the internal 2078 * blockdev inode. 2079 */ 2080 void __mark_inode_dirty(struct inode *inode, int flags) 2081 { 2082 #define I_DIRTY_INODE (I_DIRTY_SYNC | I_DIRTY_DATASYNC) 2083 struct super_block *sb = inode->i_sb; 2084 int dirtytime; 2085 2086 trace_writeback_mark_inode_dirty(inode, flags); 2087 2088 /* 2089 * Don't do this for I_DIRTY_PAGES - that doesn't actually 2090 * dirty the inode itself 2091 */ 2092 if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC | I_DIRTY_TIME)) { 2093 trace_writeback_dirty_inode_start(inode, flags); 2094 2095 if (sb->s_op->dirty_inode) 2096 sb->s_op->dirty_inode(inode, flags); 2097 2098 trace_writeback_dirty_inode(inode, flags); 2099 } 2100 if (flags & I_DIRTY_INODE) 2101 flags &= ~I_DIRTY_TIME; 2102 dirtytime = flags & I_DIRTY_TIME; 2103 2104 /* 2105 * Paired with smp_mb() in __writeback_single_inode() for the 2106 * following lockless i_state test. See there for details. 2107 */ 2108 smp_mb(); 2109 2110 if (((inode->i_state & flags) == flags) || 2111 (dirtytime && (inode->i_state & I_DIRTY_INODE))) 2112 return; 2113 2114 if (unlikely(block_dump)) 2115 block_dump___mark_inode_dirty(inode); 2116 2117 spin_lock(&inode->i_lock); 2118 if (dirtytime && (inode->i_state & I_DIRTY_INODE)) 2119 goto out_unlock_inode; 2120 if ((inode->i_state & flags) != flags) { 2121 const int was_dirty = inode->i_state & I_DIRTY; 2122 2123 inode_attach_wb(inode, NULL); 2124 2125 if (flags & I_DIRTY_INODE) 2126 inode->i_state &= ~I_DIRTY_TIME; 2127 inode->i_state |= flags; 2128 2129 /* 2130 * If the inode is being synced, just update its dirty state. 2131 * The unlocker will place the inode on the appropriate 2132 * superblock list, based upon its state. 2133 */ 2134 if (inode->i_state & I_SYNC) 2135 goto out_unlock_inode; 2136 2137 /* 2138 * Only add valid (hashed) inodes to the superblock's 2139 * dirty list. Add blockdev inodes as well. 2140 */ 2141 if (!S_ISBLK(inode->i_mode)) { 2142 if (inode_unhashed(inode)) 2143 goto out_unlock_inode; 2144 } 2145 if (inode->i_state & I_FREEING) 2146 goto out_unlock_inode; 2147 2148 /* 2149 * If the inode was already on b_dirty/b_io/b_more_io, don't 2150 * reposition it (that would break b_dirty time-ordering). 2151 */ 2152 if (!was_dirty) { 2153 struct bdi_writeback *wb; 2154 struct list_head *dirty_list; 2155 bool wakeup_bdi = false; 2156 2157 wb = locked_inode_to_wb_and_lock_list(inode); 2158 2159 WARN(bdi_cap_writeback_dirty(wb->bdi) && 2160 !test_bit(WB_registered, &wb->state), 2161 "bdi-%s not registered\n", wb->bdi->name); 2162 2163 inode->dirtied_when = jiffies; 2164 if (dirtytime) 2165 inode->dirtied_time_when = jiffies; 2166 2167 if (inode->i_state & (I_DIRTY_INODE | I_DIRTY_PAGES)) 2168 dirty_list = &wb->b_dirty; 2169 else 2170 dirty_list = &wb->b_dirty_time; 2171 2172 wakeup_bdi = inode_io_list_move_locked(inode, wb, 2173 dirty_list); 2174 2175 spin_unlock(&wb->list_lock); 2176 trace_writeback_dirty_inode_enqueue(inode); 2177 2178 /* 2179 * If this is the first dirty inode for this bdi, 2180 * we have to wake-up the corresponding bdi thread 2181 * to make sure background write-back happens 2182 * later. 2183 */ 2184 if (bdi_cap_writeback_dirty(wb->bdi) && wakeup_bdi) 2185 wb_wakeup_delayed(wb); 2186 return; 2187 } 2188 } 2189 out_unlock_inode: 2190 spin_unlock(&inode->i_lock); 2191 2192 #undef I_DIRTY_INODE 2193 } 2194 EXPORT_SYMBOL(__mark_inode_dirty); 2195 2196 /* 2197 * The @s_sync_lock is used to serialise concurrent sync operations 2198 * to avoid lock contention problems with concurrent wait_sb_inodes() calls. 2199 * Concurrent callers will block on the s_sync_lock rather than doing contending 2200 * walks. The queueing maintains sync(2) required behaviour as all the IO that 2201 * has been issued up to the time this function is enter is guaranteed to be 2202 * completed by the time we have gained the lock and waited for all IO that is 2203 * in progress regardless of the order callers are granted the lock. 2204 */ 2205 static void wait_sb_inodes(struct super_block *sb) 2206 { 2207 LIST_HEAD(sync_list); 2208 2209 /* 2210 * We need to be protected against the filesystem going from 2211 * r/o to r/w or vice versa. 2212 */ 2213 WARN_ON(!rwsem_is_locked(&sb->s_umount)); 2214 2215 mutex_lock(&sb->s_sync_lock); 2216 2217 /* 2218 * Splice the writeback list onto a temporary list to avoid waiting on 2219 * inodes that have started writeback after this point. 2220 * 2221 * Use rcu_read_lock() to keep the inodes around until we have a 2222 * reference. s_inode_wblist_lock protects sb->s_inodes_wb as well as 2223 * the local list because inodes can be dropped from either by writeback 2224 * completion. 2225 */ 2226 rcu_read_lock(); 2227 spin_lock_irq(&sb->s_inode_wblist_lock); 2228 list_splice_init(&sb->s_inodes_wb, &sync_list); 2229 2230 /* 2231 * Data integrity sync. Must wait for all pages under writeback, because 2232 * there may have been pages dirtied before our sync call, but which had 2233 * writeout started before we write it out. In which case, the inode 2234 * may not be on the dirty list, but we still have to wait for that 2235 * writeout. 2236 */ 2237 while (!list_empty(&sync_list)) { 2238 struct inode *inode = list_first_entry(&sync_list, struct inode, 2239 i_wb_list); 2240 struct address_space *mapping = inode->i_mapping; 2241 2242 /* 2243 * Move each inode back to the wb list before we drop the lock 2244 * to preserve consistency between i_wb_list and the mapping 2245 * writeback tag. Writeback completion is responsible to remove 2246 * the inode from either list once the writeback tag is cleared. 2247 */ 2248 list_move_tail(&inode->i_wb_list, &sb->s_inodes_wb); 2249 2250 /* 2251 * The mapping can appear untagged while still on-list since we 2252 * do not have the mapping lock. Skip it here, wb completion 2253 * will remove it. 2254 */ 2255 if (!mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK)) 2256 continue; 2257 2258 spin_unlock_irq(&sb->s_inode_wblist_lock); 2259 2260 spin_lock(&inode->i_lock); 2261 if (inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) { 2262 spin_unlock(&inode->i_lock); 2263 2264 spin_lock_irq(&sb->s_inode_wblist_lock); 2265 continue; 2266 } 2267 __iget(inode); 2268 spin_unlock(&inode->i_lock); 2269 rcu_read_unlock(); 2270 2271 /* 2272 * We keep the error status of individual mapping so that 2273 * applications can catch the writeback error using fsync(2). 2274 * See filemap_fdatawait_keep_errors() for details. 2275 */ 2276 filemap_fdatawait_keep_errors(mapping); 2277 2278 cond_resched(); 2279 2280 iput(inode); 2281 2282 rcu_read_lock(); 2283 spin_lock_irq(&sb->s_inode_wblist_lock); 2284 } 2285 spin_unlock_irq(&sb->s_inode_wblist_lock); 2286 rcu_read_unlock(); 2287 mutex_unlock(&sb->s_sync_lock); 2288 } 2289 2290 static void __writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr, 2291 enum wb_reason reason, bool skip_if_busy) 2292 { 2293 DEFINE_WB_COMPLETION_ONSTACK(done); 2294 struct wb_writeback_work work = { 2295 .sb = sb, 2296 .sync_mode = WB_SYNC_NONE, 2297 .tagged_writepages = 1, 2298 .done = &done, 2299 .nr_pages = nr, 2300 .reason = reason, 2301 }; 2302 struct backing_dev_info *bdi = sb->s_bdi; 2303 2304 if (!bdi_has_dirty_io(bdi) || bdi == &noop_backing_dev_info) 2305 return; 2306 WARN_ON(!rwsem_is_locked(&sb->s_umount)); 2307 2308 bdi_split_work_to_wbs(sb->s_bdi, &work, skip_if_busy); 2309 wb_wait_for_completion(bdi, &done); 2310 } 2311 2312 /** 2313 * writeback_inodes_sb_nr - writeback dirty inodes from given super_block 2314 * @sb: the superblock 2315 * @nr: the number of pages to write 2316 * @reason: reason why some writeback work initiated 2317 * 2318 * Start writeback on some inodes on this super_block. No guarantees are made 2319 * on how many (if any) will be written, and this function does not wait 2320 * for IO completion of submitted IO. 2321 */ 2322 void writeback_inodes_sb_nr(struct super_block *sb, 2323 unsigned long nr, 2324 enum wb_reason reason) 2325 { 2326 __writeback_inodes_sb_nr(sb, nr, reason, false); 2327 } 2328 EXPORT_SYMBOL(writeback_inodes_sb_nr); 2329 2330 /** 2331 * writeback_inodes_sb - writeback dirty inodes from given super_block 2332 * @sb: the superblock 2333 * @reason: reason why some writeback work was initiated 2334 * 2335 * Start writeback on some inodes on this super_block. No guarantees are made 2336 * on how many (if any) will be written, and this function does not wait 2337 * for IO completion of submitted IO. 2338 */ 2339 void writeback_inodes_sb(struct super_block *sb, enum wb_reason reason) 2340 { 2341 return writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason); 2342 } 2343 EXPORT_SYMBOL(writeback_inodes_sb); 2344 2345 /** 2346 * try_to_writeback_inodes_sb_nr - try to start writeback if none underway 2347 * @sb: the superblock 2348 * @nr: the number of pages to write 2349 * @reason: the reason of writeback 2350 * 2351 * Invoke writeback_inodes_sb_nr if no writeback is currently underway. 2352 * Returns 1 if writeback was started, 0 if not. 2353 */ 2354 bool try_to_writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr, 2355 enum wb_reason reason) 2356 { 2357 if (!down_read_trylock(&sb->s_umount)) 2358 return false; 2359 2360 __writeback_inodes_sb_nr(sb, nr, reason, true); 2361 up_read(&sb->s_umount); 2362 return true; 2363 } 2364 EXPORT_SYMBOL(try_to_writeback_inodes_sb_nr); 2365 2366 /** 2367 * try_to_writeback_inodes_sb - try to start writeback if none underway 2368 * @sb: the superblock 2369 * @reason: reason why some writeback work was initiated 2370 * 2371 * Implement by try_to_writeback_inodes_sb_nr() 2372 * Returns 1 if writeback was started, 0 if not. 2373 */ 2374 bool try_to_writeback_inodes_sb(struct super_block *sb, enum wb_reason reason) 2375 { 2376 return try_to_writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason); 2377 } 2378 EXPORT_SYMBOL(try_to_writeback_inodes_sb); 2379 2380 /** 2381 * sync_inodes_sb - sync sb inode pages 2382 * @sb: the superblock 2383 * 2384 * This function writes and waits on any dirty inode belonging to this 2385 * super_block. 2386 */ 2387 void sync_inodes_sb(struct super_block *sb) 2388 { 2389 DEFINE_WB_COMPLETION_ONSTACK(done); 2390 struct wb_writeback_work work = { 2391 .sb = sb, 2392 .sync_mode = WB_SYNC_ALL, 2393 .nr_pages = LONG_MAX, 2394 .range_cyclic = 0, 2395 .done = &done, 2396 .reason = WB_REASON_SYNC, 2397 .for_sync = 1, 2398 }; 2399 struct backing_dev_info *bdi = sb->s_bdi; 2400 2401 /* 2402 * Can't skip on !bdi_has_dirty() because we should wait for !dirty 2403 * inodes under writeback and I_DIRTY_TIME inodes ignored by 2404 * bdi_has_dirty() need to be written out too. 2405 */ 2406 if (bdi == &noop_backing_dev_info) 2407 return; 2408 WARN_ON(!rwsem_is_locked(&sb->s_umount)); 2409 2410 bdi_split_work_to_wbs(bdi, &work, false); 2411 wb_wait_for_completion(bdi, &done); 2412 2413 wait_sb_inodes(sb); 2414 } 2415 EXPORT_SYMBOL(sync_inodes_sb); 2416 2417 /** 2418 * write_inode_now - write an inode to disk 2419 * @inode: inode to write to disk 2420 * @sync: whether the write should be synchronous or not 2421 * 2422 * This function commits an inode to disk immediately if it is dirty. This is 2423 * primarily needed by knfsd. 2424 * 2425 * The caller must either have a ref on the inode or must have set I_WILL_FREE. 2426 */ 2427 int write_inode_now(struct inode *inode, int sync) 2428 { 2429 struct writeback_control wbc = { 2430 .nr_to_write = LONG_MAX, 2431 .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE, 2432 .range_start = 0, 2433 .range_end = LLONG_MAX, 2434 }; 2435 2436 if (!mapping_cap_writeback_dirty(inode->i_mapping)) 2437 wbc.nr_to_write = 0; 2438 2439 might_sleep(); 2440 return writeback_single_inode(inode, &wbc); 2441 } 2442 EXPORT_SYMBOL(write_inode_now); 2443 2444 /** 2445 * sync_inode - write an inode and its pages to disk. 2446 * @inode: the inode to sync 2447 * @wbc: controls the writeback mode 2448 * 2449 * sync_inode() will write an inode and its pages to disk. It will also 2450 * correctly update the inode on its superblock's dirty inode lists and will 2451 * update inode->i_state. 2452 * 2453 * The caller must have a ref on the inode. 2454 */ 2455 int sync_inode(struct inode *inode, struct writeback_control *wbc) 2456 { 2457 return writeback_single_inode(inode, wbc); 2458 } 2459 EXPORT_SYMBOL(sync_inode); 2460 2461 /** 2462 * sync_inode_metadata - write an inode to disk 2463 * @inode: the inode to sync 2464 * @wait: wait for I/O to complete. 2465 * 2466 * Write an inode to disk and adjust its dirty state after completion. 2467 * 2468 * Note: only writes the actual inode, no associated data or other metadata. 2469 */ 2470 int sync_inode_metadata(struct inode *inode, int wait) 2471 { 2472 struct writeback_control wbc = { 2473 .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE, 2474 .nr_to_write = 0, /* metadata-only */ 2475 }; 2476 2477 return sync_inode(inode, &wbc); 2478 } 2479 EXPORT_SYMBOL(sync_inode_metadata); 2480