| 1 | // SPDX-License-Identifier: GPL-2.0-only |
| 2 | /* |
| 3 | * fs/fs-writeback.c |
| 4 | * |
| 5 | * Copyright (C) 2002, Linus Torvalds. |
| 6 | * |
| 7 | * Contains all the functions related to writing back and waiting |
| 8 | * upon dirty inodes against superblocks, and writing back dirty |
| 9 | * pages against inodes. ie: data writeback. Writeout of the |
| 10 | * inode itself is not handled here. |
| 11 | * |
| 12 | * 10Apr2002 Andrew Morton |
| 13 | * Split out of fs/inode.c |
| 14 | * Additions for address_space-based writeback |
| 15 | */ |
| 16 | |
| 17 | #include <linux/kernel.h> |
| 18 | #include <linux/export.h> |
| 19 | #include <linux/spinlock.h> |
| 20 | #include <linux/slab.h> |
| 21 | #include <linux/sched.h> |
| 22 | #include <linux/fs.h> |
| 23 | #include <linux/mm.h> |
| 24 | #include <linux/pagemap.h> |
| 25 | #include <linux/kthread.h> |
| 26 | #include <linux/writeback.h> |
| 27 | #include <linux/blkdev.h> |
| 28 | #include <linux/backing-dev.h> |
| 29 | #include <linux/tracepoint.h> |
| 30 | #include <linux/device.h> |
| 31 | #include <linux/memcontrol.h> |
| 32 | #include "internal.h" |
| 33 | |
| 34 | /* |
| 35 | * 4MB minimal write chunk size |
| 36 | */ |
| 37 | #define MIN_WRITEBACK_PAGES (4096UL >> (PAGE_SHIFT - 10)) |
| 38 | |
| 39 | /* |
| 40 | * Passed into wb_writeback(), essentially a subset of writeback_control |
| 41 | */ |
| 42 | struct wb_writeback_work { |
| 43 | long nr_pages; |
| 44 | struct super_block *sb; |
| 45 | enum writeback_sync_modes sync_mode; |
| 46 | unsigned int tagged_writepages:1; |
| 47 | unsigned int for_kupdate:1; |
| 48 | unsigned int range_cyclic:1; |
| 49 | unsigned int for_background:1; |
| 50 | unsigned int for_sync:1; /* sync(2) WB_SYNC_ALL writeback */ |
| 51 | unsigned int auto_free:1; /* free on completion */ |
| 52 | enum wb_reason reason; /* why was writeback initiated? */ |
| 53 | |
| 54 | struct list_head list; /* pending work list */ |
| 55 | struct wb_completion *done; /* set if the caller waits */ |
| 56 | }; |
| 57 | |
| 58 | /* |
| 59 | * If an inode is constantly having its pages dirtied, but then the |
| 60 | * updates stop dirtytime_expire_interval seconds in the past, it's |
| 61 | * possible for the worst case time between when an inode has its |
| 62 | * timestamps updated and when they finally get written out to be two |
| 63 | * dirtytime_expire_intervals. We set the default to 12 hours (in |
| 64 | * seconds), which means most of the time inodes will have their |
| 65 | * timestamps written to disk after 12 hours, but in the worst case a |
| 66 | * few inodes might not their timestamps updated for 24 hours. |
| 67 | */ |
| 68 | unsigned int dirtytime_expire_interval = 12 * 60 * 60; |
| 69 | |
| 70 | static inline struct inode *wb_inode(struct list_head *head) |
| 71 | { |
| 72 | return list_entry(head, struct inode, i_io_list); |
| 73 | } |
| 74 | |
| 75 | /* |
| 76 | * Include the creation of the trace points after defining the |
| 77 | * wb_writeback_work structure and inline functions so that the definition |
| 78 | * remains local to this file. |
| 79 | */ |
| 80 | #define CREATE_TRACE_POINTS |
| 81 | #include <trace/events/writeback.h> |
| 82 | |
| 83 | EXPORT_TRACEPOINT_SYMBOL_GPL(wbc_writepage); |
| 84 | |
| 85 | static bool wb_io_lists_populated(struct bdi_writeback *wb) |
| 86 | { |
| 87 | if (wb_has_dirty_io(wb)) { |
| 88 | return false; |
| 89 | } else { |
| 90 | set_bit(WB_has_dirty_io, &wb->state); |
| 91 | WARN_ON_ONCE(!wb->avg_write_bandwidth); |
| 92 | atomic_long_add(wb->avg_write_bandwidth, |
| 93 | &wb->bdi->tot_write_bandwidth); |
| 94 | return true; |
| 95 | } |
| 96 | } |
| 97 | |
| 98 | static void wb_io_lists_depopulated(struct bdi_writeback *wb) |
| 99 | { |
| 100 | if (wb_has_dirty_io(wb) && list_empty(&wb->b_dirty) && |
| 101 | list_empty(&wb->b_io) && list_empty(&wb->b_more_io)) { |
| 102 | clear_bit(WB_has_dirty_io, &wb->state); |
| 103 | WARN_ON_ONCE(atomic_long_sub_return(wb->avg_write_bandwidth, |
| 104 | &wb->bdi->tot_write_bandwidth) < 0); |
| 105 | } |
| 106 | } |
| 107 | |
| 108 | /** |
| 109 | * inode_io_list_move_locked - move an inode onto a bdi_writeback IO list |
| 110 | * @inode: inode to be moved |
| 111 | * @wb: target bdi_writeback |
| 112 | * @head: one of @wb->b_{dirty|io|more_io|dirty_time} |
| 113 | * |
| 114 | * Move @inode->i_io_list to @list of @wb and set %WB_has_dirty_io. |
| 115 | * Returns %true if @inode is the first occupant of the !dirty_time IO |
| 116 | * lists; otherwise, %false. |
| 117 | */ |
| 118 | static bool inode_io_list_move_locked(struct inode *inode, |
| 119 | struct bdi_writeback *wb, |
| 120 | struct list_head *head) |
| 121 | { |
| 122 | assert_spin_locked(&wb->list_lock); |
| 123 | assert_spin_locked(&inode->i_lock); |
| 124 | WARN_ON_ONCE(inode->i_state & I_FREEING); |
| 125 | |
| 126 | list_move(&inode->i_io_list, head); |
| 127 | |
| 128 | /* dirty_time doesn't count as dirty_io until expiration */ |
| 129 | if (head != &wb->b_dirty_time) |
| 130 | return wb_io_lists_populated(wb); |
| 131 | |
| 132 | wb_io_lists_depopulated(wb); |
| 133 | return false; |
| 134 | } |
| 135 | |
| 136 | static void wb_wakeup(struct bdi_writeback *wb) |
| 137 | { |
| 138 | spin_lock_irq(&wb->work_lock); |
| 139 | if (test_bit(WB_registered, &wb->state)) |
| 140 | mod_delayed_work(bdi_wq, &wb->dwork, 0); |
| 141 | spin_unlock_irq(&wb->work_lock); |
| 142 | } |
| 143 | |
| 144 | static void finish_writeback_work(struct bdi_writeback *wb, |
| 145 | struct wb_writeback_work *work) |
| 146 | { |
| 147 | struct wb_completion *done = work->done; |
| 148 | |
| 149 | if (work->auto_free) |
| 150 | kfree(work); |
| 151 | if (done) { |
| 152 | wait_queue_head_t *waitq = done->waitq; |
| 153 | |
| 154 | /* @done can't be accessed after the following dec */ |
| 155 | if (atomic_dec_and_test(&done->cnt)) |
| 156 | wake_up_all(waitq); |
| 157 | } |
| 158 | } |
| 159 | |
| 160 | static void wb_queue_work(struct bdi_writeback *wb, |
| 161 | struct wb_writeback_work *work) |
| 162 | { |
| 163 | trace_writeback_queue(wb, work); |
| 164 | |
| 165 | if (work->done) |
| 166 | atomic_inc(&work->done->cnt); |
| 167 | |
| 168 | spin_lock_irq(&wb->work_lock); |
| 169 | |
| 170 | if (test_bit(WB_registered, &wb->state)) { |
| 171 | list_add_tail(&work->list, &wb->work_list); |
| 172 | mod_delayed_work(bdi_wq, &wb->dwork, 0); |
| 173 | } else |
| 174 | finish_writeback_work(wb, work); |
| 175 | |
| 176 | spin_unlock_irq(&wb->work_lock); |
| 177 | } |
| 178 | |
| 179 | /** |
| 180 | * wb_wait_for_completion - wait for completion of bdi_writeback_works |
| 181 | * @done: target wb_completion |
| 182 | * |
| 183 | * Wait for one or more work items issued to @bdi with their ->done field |
| 184 | * set to @done, which should have been initialized with |
| 185 | * DEFINE_WB_COMPLETION(). This function returns after all such work items |
| 186 | * are completed. Work items which are waited upon aren't freed |
| 187 | * automatically on completion. |
| 188 | */ |
| 189 | void wb_wait_for_completion(struct wb_completion *done) |
| 190 | { |
| 191 | atomic_dec(&done->cnt); /* put down the initial count */ |
| 192 | wait_event(*done->waitq, !atomic_read(&done->cnt)); |
| 193 | } |
| 194 | |
| 195 | #ifdef CONFIG_CGROUP_WRITEBACK |
| 196 | |
| 197 | /* |
| 198 | * Parameters for foreign inode detection, see wbc_detach_inode() to see |
| 199 | * how they're used. |
| 200 | * |
| 201 | * These paramters are inherently heuristical as the detection target |
| 202 | * itself is fuzzy. All we want to do is detaching an inode from the |
| 203 | * current owner if it's being written to by some other cgroups too much. |
| 204 | * |
| 205 | * The current cgroup writeback is built on the assumption that multiple |
| 206 | * cgroups writing to the same inode concurrently is very rare and a mode |
| 207 | * of operation which isn't well supported. As such, the goal is not |
| 208 | * taking too long when a different cgroup takes over an inode while |
| 209 | * avoiding too aggressive flip-flops from occasional foreign writes. |
| 210 | * |
| 211 | * We record, very roughly, 2s worth of IO time history and if more than |
| 212 | * half of that is foreign, trigger the switch. The recording is quantized |
| 213 | * to 16 slots. To avoid tiny writes from swinging the decision too much, |
| 214 | * writes smaller than 1/8 of avg size are ignored. |
| 215 | */ |
| 216 | #define WB_FRN_TIME_SHIFT 13 /* 1s = 2^13, upto 8 secs w/ 16bit */ |
| 217 | #define WB_FRN_TIME_AVG_SHIFT 3 /* avg = avg * 7/8 + new * 1/8 */ |
| 218 | #define WB_FRN_TIME_CUT_DIV 8 /* ignore rounds < avg / 8 */ |
| 219 | #define WB_FRN_TIME_PERIOD (2 * (1 << WB_FRN_TIME_SHIFT)) /* 2s */ |
| 220 | |
| 221 | #define WB_FRN_HIST_SLOTS 16 /* inode->i_wb_frn_history is 16bit */ |
| 222 | #define WB_FRN_HIST_UNIT (WB_FRN_TIME_PERIOD / WB_FRN_HIST_SLOTS) |
| 223 | /* each slot's duration is 2s / 16 */ |
| 224 | #define WB_FRN_HIST_THR_SLOTS (WB_FRN_HIST_SLOTS / 2) |
| 225 | /* if foreign slots >= 8, switch */ |
| 226 | #define WB_FRN_HIST_MAX_SLOTS (WB_FRN_HIST_THR_SLOTS / 2 + 1) |
| 227 | /* one round can affect upto 5 slots */ |
| 228 | #define WB_FRN_MAX_IN_FLIGHT 1024 /* don't queue too many concurrently */ |
| 229 | |
| 230 | /* |
| 231 | * Maximum inodes per isw. A specific value has been chosen to make |
| 232 | * struct inode_switch_wbs_context fit into 1024 bytes kmalloc. |
| 233 | */ |
| 234 | #define WB_MAX_INODES_PER_ISW ((1024UL - sizeof(struct inode_switch_wbs_context)) \ |
| 235 | / sizeof(struct inode *)) |
| 236 | |
| 237 | static atomic_t isw_nr_in_flight = ATOMIC_INIT(0); |
| 238 | static struct workqueue_struct *isw_wq; |
| 239 | |
| 240 | void __inode_attach_wb(struct inode *inode, struct folio *folio) |
| 241 | { |
| 242 | struct backing_dev_info *bdi = inode_to_bdi(inode); |
| 243 | struct bdi_writeback *wb = NULL; |
| 244 | |
| 245 | if (inode_cgwb_enabled(inode)) { |
| 246 | struct cgroup_subsys_state *memcg_css; |
| 247 | |
| 248 | if (folio) { |
| 249 | memcg_css = mem_cgroup_css_from_folio(folio); |
| 250 | wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC); |
| 251 | } else { |
| 252 | /* must pin memcg_css, see wb_get_create() */ |
| 253 | memcg_css = task_get_css(current, memory_cgrp_id); |
| 254 | wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC); |
| 255 | css_put(memcg_css); |
| 256 | } |
| 257 | } |
| 258 | |
| 259 | if (!wb) |
| 260 | wb = &bdi->wb; |
| 261 | |
| 262 | /* |
| 263 | * There may be multiple instances of this function racing to |
| 264 | * update the same inode. Use cmpxchg() to tell the winner. |
| 265 | */ |
| 266 | if (unlikely(cmpxchg(&inode->i_wb, NULL, wb))) |
| 267 | wb_put(wb); |
| 268 | } |
| 269 | EXPORT_SYMBOL_GPL(__inode_attach_wb); |
| 270 | |
| 271 | /** |
| 272 | * inode_cgwb_move_to_attached - put the inode onto wb->b_attached list |
| 273 | * @inode: inode of interest with i_lock held |
| 274 | * @wb: target bdi_writeback |
| 275 | * |
| 276 | * Remove the inode from wb's io lists and if necessarily put onto b_attached |
| 277 | * list. Only inodes attached to cgwb's are kept on this list. |
| 278 | */ |
| 279 | static void inode_cgwb_move_to_attached(struct inode *inode, |
| 280 | struct bdi_writeback *wb) |
| 281 | { |
| 282 | assert_spin_locked(&wb->list_lock); |
| 283 | assert_spin_locked(&inode->i_lock); |
| 284 | WARN_ON_ONCE(inode->i_state & I_FREEING); |
| 285 | |
| 286 | inode->i_state &= ~I_SYNC_QUEUED; |
| 287 | if (wb != &wb->bdi->wb) |
| 288 | list_move(&inode->i_io_list, &wb->b_attached); |
| 289 | else |
| 290 | list_del_init(&inode->i_io_list); |
| 291 | wb_io_lists_depopulated(wb); |
| 292 | } |
| 293 | |
| 294 | /** |
| 295 | * locked_inode_to_wb_and_lock_list - determine a locked inode's wb and lock it |
| 296 | * @inode: inode of interest with i_lock held |
| 297 | * |
| 298 | * Returns @inode's wb with its list_lock held. @inode->i_lock must be |
| 299 | * held on entry and is released on return. The returned wb is guaranteed |
| 300 | * to stay @inode's associated wb until its list_lock is released. |
| 301 | */ |
| 302 | static struct bdi_writeback * |
| 303 | locked_inode_to_wb_and_lock_list(struct inode *inode) |
| 304 | __releases(&inode->i_lock) |
| 305 | __acquires(&wb->list_lock) |
| 306 | { |
| 307 | while (true) { |
| 308 | struct bdi_writeback *wb = inode_to_wb(inode); |
| 309 | |
| 310 | /* |
| 311 | * inode_to_wb() association is protected by both |
| 312 | * @inode->i_lock and @wb->list_lock but list_lock nests |
| 313 | * outside i_lock. Drop i_lock and verify that the |
| 314 | * association hasn't changed after acquiring list_lock. |
| 315 | */ |
| 316 | wb_get(wb); |
| 317 | spin_unlock(&inode->i_lock); |
| 318 | spin_lock(&wb->list_lock); |
| 319 | |
| 320 | /* i_wb may have changed inbetween, can't use inode_to_wb() */ |
| 321 | if (likely(wb == inode->i_wb)) { |
| 322 | wb_put(wb); /* @inode already has ref */ |
| 323 | return wb; |
| 324 | } |
| 325 | |
| 326 | spin_unlock(&wb->list_lock); |
| 327 | wb_put(wb); |
| 328 | cpu_relax(); |
| 329 | spin_lock(&inode->i_lock); |
| 330 | } |
| 331 | } |
| 332 | |
| 333 | /** |
| 334 | * inode_to_wb_and_lock_list - determine an inode's wb and lock it |
| 335 | * @inode: inode of interest |
| 336 | * |
| 337 | * Same as locked_inode_to_wb_and_lock_list() but @inode->i_lock isn't held |
| 338 | * on entry. |
| 339 | */ |
| 340 | static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode) |
| 341 | __acquires(&wb->list_lock) |
| 342 | { |
| 343 | spin_lock(&inode->i_lock); |
| 344 | return locked_inode_to_wb_and_lock_list(inode); |
| 345 | } |
| 346 | |
| 347 | struct inode_switch_wbs_context { |
| 348 | struct rcu_work work; |
| 349 | |
| 350 | /* |
| 351 | * Multiple inodes can be switched at once. The switching procedure |
| 352 | * consists of two parts, separated by a RCU grace period. To make |
| 353 | * sure that the second part is executed for each inode gone through |
| 354 | * the first part, all inode pointers are placed into a NULL-terminated |
| 355 | * array embedded into struct inode_switch_wbs_context. Otherwise |
| 356 | * an inode could be left in a non-consistent state. |
| 357 | */ |
| 358 | struct bdi_writeback *new_wb; |
| 359 | struct inode *inodes[]; |
| 360 | }; |
| 361 | |
| 362 | static void bdi_down_write_wb_switch_rwsem(struct backing_dev_info *bdi) |
| 363 | { |
| 364 | down_write(&bdi->wb_switch_rwsem); |
| 365 | } |
| 366 | |
| 367 | static void bdi_up_write_wb_switch_rwsem(struct backing_dev_info *bdi) |
| 368 | { |
| 369 | up_write(&bdi->wb_switch_rwsem); |
| 370 | } |
| 371 | |
| 372 | static bool inode_do_switch_wbs(struct inode *inode, |
| 373 | struct bdi_writeback *old_wb, |
| 374 | struct bdi_writeback *new_wb) |
| 375 | { |
| 376 | struct address_space *mapping = inode->i_mapping; |
| 377 | XA_STATE(xas, &mapping->i_pages, 0); |
| 378 | struct folio *folio; |
| 379 | bool switched = false; |
| 380 | |
| 381 | spin_lock(&inode->i_lock); |
| 382 | xa_lock_irq(&mapping->i_pages); |
| 383 | |
| 384 | /* |
| 385 | * Once I_FREEING or I_WILL_FREE are visible under i_lock, the eviction |
| 386 | * path owns the inode and we shouldn't modify ->i_io_list. |
| 387 | */ |
| 388 | if (unlikely(inode->i_state & (I_FREEING | I_WILL_FREE))) |
| 389 | goto skip_switch; |
| 390 | |
| 391 | trace_inode_switch_wbs(inode, old_wb, new_wb); |
| 392 | |
| 393 | /* |
| 394 | * Count and transfer stats. Note that PAGECACHE_TAG_DIRTY points |
| 395 | * to possibly dirty folios while PAGECACHE_TAG_WRITEBACK points to |
| 396 | * folios actually under writeback. |
| 397 | */ |
| 398 | xas_for_each_marked(&xas, folio, ULONG_MAX, PAGECACHE_TAG_DIRTY) { |
| 399 | if (folio_test_dirty(folio)) { |
| 400 | long nr = folio_nr_pages(folio); |
| 401 | wb_stat_mod(old_wb, WB_RECLAIMABLE, -nr); |
| 402 | wb_stat_mod(new_wb, WB_RECLAIMABLE, nr); |
| 403 | } |
| 404 | } |
| 405 | |
| 406 | xas_set(&xas, 0); |
| 407 | xas_for_each_marked(&xas, folio, ULONG_MAX, PAGECACHE_TAG_WRITEBACK) { |
| 408 | long nr = folio_nr_pages(folio); |
| 409 | WARN_ON_ONCE(!folio_test_writeback(folio)); |
| 410 | wb_stat_mod(old_wb, WB_WRITEBACK, -nr); |
| 411 | wb_stat_mod(new_wb, WB_WRITEBACK, nr); |
| 412 | } |
| 413 | |
| 414 | if (mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK)) { |
| 415 | atomic_dec(&old_wb->writeback_inodes); |
| 416 | atomic_inc(&new_wb->writeback_inodes); |
| 417 | } |
| 418 | |
| 419 | wb_get(new_wb); |
| 420 | |
| 421 | /* |
| 422 | * Transfer to @new_wb's IO list if necessary. If the @inode is dirty, |
| 423 | * the specific list @inode was on is ignored and the @inode is put on |
| 424 | * ->b_dirty which is always correct including from ->b_dirty_time. |
| 425 | * The transfer preserves @inode->dirtied_when ordering. If the @inode |
| 426 | * was clean, it means it was on the b_attached list, so move it onto |
| 427 | * the b_attached list of @new_wb. |
| 428 | */ |
| 429 | if (!list_empty(&inode->i_io_list)) { |
| 430 | inode->i_wb = new_wb; |
| 431 | |
| 432 | if (inode->i_state & I_DIRTY_ALL) { |
| 433 | struct inode *pos; |
| 434 | |
| 435 | list_for_each_entry(pos, &new_wb->b_dirty, i_io_list) |
| 436 | if (time_after_eq(inode->dirtied_when, |
| 437 | pos->dirtied_when)) |
| 438 | break; |
| 439 | inode_io_list_move_locked(inode, new_wb, |
| 440 | pos->i_io_list.prev); |
| 441 | } else { |
| 442 | inode_cgwb_move_to_attached(inode, new_wb); |
| 443 | } |
| 444 | } else { |
| 445 | inode->i_wb = new_wb; |
| 446 | } |
| 447 | |
| 448 | /* ->i_wb_frn updates may race wbc_detach_inode() but doesn't matter */ |
| 449 | inode->i_wb_frn_winner = 0; |
| 450 | inode->i_wb_frn_avg_time = 0; |
| 451 | inode->i_wb_frn_history = 0; |
| 452 | switched = true; |
| 453 | skip_switch: |
| 454 | /* |
| 455 | * Paired with load_acquire in unlocked_inode_to_wb_begin() and |
| 456 | * ensures that the new wb is visible if they see !I_WB_SWITCH. |
| 457 | */ |
| 458 | smp_store_release(&inode->i_state, inode->i_state & ~I_WB_SWITCH); |
| 459 | |
| 460 | xa_unlock_irq(&mapping->i_pages); |
| 461 | spin_unlock(&inode->i_lock); |
| 462 | |
| 463 | return switched; |
| 464 | } |
| 465 | |
| 466 | static void inode_switch_wbs_work_fn(struct work_struct *work) |
| 467 | { |
| 468 | struct inode_switch_wbs_context *isw = |
| 469 | container_of(to_rcu_work(work), struct inode_switch_wbs_context, work); |
| 470 | struct backing_dev_info *bdi = inode_to_bdi(isw->inodes[0]); |
| 471 | struct bdi_writeback *old_wb = isw->inodes[0]->i_wb; |
| 472 | struct bdi_writeback *new_wb = isw->new_wb; |
| 473 | unsigned long nr_switched = 0; |
| 474 | struct inode **inodep; |
| 475 | |
| 476 | /* |
| 477 | * If @inode switches cgwb membership while sync_inodes_sb() is |
| 478 | * being issued, sync_inodes_sb() might miss it. Synchronize. |
| 479 | */ |
| 480 | down_read(&bdi->wb_switch_rwsem); |
| 481 | |
| 482 | /* |
| 483 | * By the time control reaches here, RCU grace period has passed |
| 484 | * since I_WB_SWITCH assertion and all wb stat update transactions |
| 485 | * between unlocked_inode_to_wb_begin/end() are guaranteed to be |
| 486 | * synchronizing against the i_pages lock. |
| 487 | * |
| 488 | * Grabbing old_wb->list_lock, inode->i_lock and the i_pages lock |
| 489 | * gives us exclusion against all wb related operations on @inode |
| 490 | * including IO list manipulations and stat updates. |
| 491 | */ |
| 492 | if (old_wb < new_wb) { |
| 493 | spin_lock(&old_wb->list_lock); |
| 494 | spin_lock_nested(&new_wb->list_lock, SINGLE_DEPTH_NESTING); |
| 495 | } else { |
| 496 | spin_lock(&new_wb->list_lock); |
| 497 | spin_lock_nested(&old_wb->list_lock, SINGLE_DEPTH_NESTING); |
| 498 | } |
| 499 | |
| 500 | for (inodep = isw->inodes; *inodep; inodep++) { |
| 501 | WARN_ON_ONCE((*inodep)->i_wb != old_wb); |
| 502 | if (inode_do_switch_wbs(*inodep, old_wb, new_wb)) |
| 503 | nr_switched++; |
| 504 | } |
| 505 | |
| 506 | spin_unlock(&new_wb->list_lock); |
| 507 | spin_unlock(&old_wb->list_lock); |
| 508 | |
| 509 | up_read(&bdi->wb_switch_rwsem); |
| 510 | |
| 511 | if (nr_switched) { |
| 512 | wb_wakeup(new_wb); |
| 513 | wb_put_many(old_wb, nr_switched); |
| 514 | } |
| 515 | |
| 516 | for (inodep = isw->inodes; *inodep; inodep++) |
| 517 | iput(*inodep); |
| 518 | wb_put(new_wb); |
| 519 | kfree(isw); |
| 520 | atomic_dec(&isw_nr_in_flight); |
| 521 | } |
| 522 | |
| 523 | static bool inode_prepare_wbs_switch(struct inode *inode, |
| 524 | struct bdi_writeback *new_wb) |
| 525 | { |
| 526 | /* |
| 527 | * Paired with smp_mb() in cgroup_writeback_umount(). |
| 528 | * isw_nr_in_flight must be increased before checking SB_ACTIVE and |
| 529 | * grabbing an inode, otherwise isw_nr_in_flight can be observed as 0 |
| 530 | * in cgroup_writeback_umount() and the isw_wq will be not flushed. |
| 531 | */ |
| 532 | smp_mb(); |
| 533 | |
| 534 | if (IS_DAX(inode)) |
| 535 | return false; |
| 536 | |
| 537 | /* while holding I_WB_SWITCH, no one else can update the association */ |
| 538 | spin_lock(&inode->i_lock); |
| 539 | if (!(inode->i_sb->s_flags & SB_ACTIVE) || |
| 540 | inode->i_state & (I_WB_SWITCH | I_FREEING | I_WILL_FREE) || |
| 541 | inode_to_wb(inode) == new_wb) { |
| 542 | spin_unlock(&inode->i_lock); |
| 543 | return false; |
| 544 | } |
| 545 | inode->i_state |= I_WB_SWITCH; |
| 546 | __iget(inode); |
| 547 | spin_unlock(&inode->i_lock); |
| 548 | |
| 549 | return true; |
| 550 | } |
| 551 | |
| 552 | /** |
| 553 | * inode_switch_wbs - change the wb association of an inode |
| 554 | * @inode: target inode |
| 555 | * @new_wb_id: ID of the new wb |
| 556 | * |
| 557 | * Switch @inode's wb association to the wb identified by @new_wb_id. The |
| 558 | * switching is performed asynchronously and may fail silently. |
| 559 | */ |
| 560 | static void inode_switch_wbs(struct inode *inode, int new_wb_id) |
| 561 | { |
| 562 | struct backing_dev_info *bdi = inode_to_bdi(inode); |
| 563 | struct cgroup_subsys_state *memcg_css; |
| 564 | struct inode_switch_wbs_context *isw; |
| 565 | |
| 566 | /* noop if seems to be already in progress */ |
| 567 | if (inode->i_state & I_WB_SWITCH) |
| 568 | return; |
| 569 | |
| 570 | /* avoid queueing a new switch if too many are already in flight */ |
| 571 | if (atomic_read(&isw_nr_in_flight) > WB_FRN_MAX_IN_FLIGHT) |
| 572 | return; |
| 573 | |
| 574 | isw = kzalloc(struct_size(isw, inodes, 2), GFP_ATOMIC); |
| 575 | if (!isw) |
| 576 | return; |
| 577 | |
| 578 | atomic_inc(&isw_nr_in_flight); |
| 579 | |
| 580 | /* find and pin the new wb */ |
| 581 | rcu_read_lock(); |
| 582 | memcg_css = css_from_id(new_wb_id, &memory_cgrp_subsys); |
| 583 | if (memcg_css && !css_tryget(memcg_css)) |
| 584 | memcg_css = NULL; |
| 585 | rcu_read_unlock(); |
| 586 | if (!memcg_css) |
| 587 | goto out_free; |
| 588 | |
| 589 | isw->new_wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC); |
| 590 | css_put(memcg_css); |
| 591 | if (!isw->new_wb) |
| 592 | goto out_free; |
| 593 | |
| 594 | if (!inode_prepare_wbs_switch(inode, isw->new_wb)) |
| 595 | goto out_free; |
| 596 | |
| 597 | isw->inodes[0] = inode; |
| 598 | |
| 599 | /* |
| 600 | * In addition to synchronizing among switchers, I_WB_SWITCH tells |
| 601 | * the RCU protected stat update paths to grab the i_page |
| 602 | * lock so that stat transfer can synchronize against them. |
| 603 | * Let's continue after I_WB_SWITCH is guaranteed to be visible. |
| 604 | */ |
| 605 | INIT_RCU_WORK(&isw->work, inode_switch_wbs_work_fn); |
| 606 | queue_rcu_work(isw_wq, &isw->work); |
| 607 | return; |
| 608 | |
| 609 | out_free: |
| 610 | atomic_dec(&isw_nr_in_flight); |
| 611 | if (isw->new_wb) |
| 612 | wb_put(isw->new_wb); |
| 613 | kfree(isw); |
| 614 | } |
| 615 | |
| 616 | /** |
| 617 | * cleanup_offline_cgwb - detach associated inodes |
| 618 | * @wb: target wb |
| 619 | * |
| 620 | * Switch all inodes attached to @wb to a nearest living ancestor's wb in order |
| 621 | * to eventually release the dying @wb. Returns %true if not all inodes were |
| 622 | * switched and the function has to be restarted. |
| 623 | */ |
| 624 | bool cleanup_offline_cgwb(struct bdi_writeback *wb) |
| 625 | { |
| 626 | struct cgroup_subsys_state *memcg_css; |
| 627 | struct inode_switch_wbs_context *isw; |
| 628 | struct inode *inode; |
| 629 | int nr; |
| 630 | bool restart = false; |
| 631 | |
| 632 | isw = kzalloc(struct_size(isw, inodes, WB_MAX_INODES_PER_ISW), |
| 633 | GFP_KERNEL); |
| 634 | if (!isw) |
| 635 | return restart; |
| 636 | |
| 637 | atomic_inc(&isw_nr_in_flight); |
| 638 | |
| 639 | for (memcg_css = wb->memcg_css->parent; memcg_css; |
| 640 | memcg_css = memcg_css->parent) { |
| 641 | isw->new_wb = wb_get_create(wb->bdi, memcg_css, GFP_KERNEL); |
| 642 | if (isw->new_wb) |
| 643 | break; |
| 644 | } |
| 645 | if (unlikely(!isw->new_wb)) |
| 646 | isw->new_wb = &wb->bdi->wb; /* wb_get() is noop for bdi's wb */ |
| 647 | |
| 648 | nr = 0; |
| 649 | spin_lock(&wb->list_lock); |
| 650 | list_for_each_entry(inode, &wb->b_attached, i_io_list) { |
| 651 | if (!inode_prepare_wbs_switch(inode, isw->new_wb)) |
| 652 | continue; |
| 653 | |
| 654 | isw->inodes[nr++] = inode; |
| 655 | |
| 656 | if (nr >= WB_MAX_INODES_PER_ISW - 1) { |
| 657 | restart = true; |
| 658 | break; |
| 659 | } |
| 660 | } |
| 661 | spin_unlock(&wb->list_lock); |
| 662 | |
| 663 | /* no attached inodes? bail out */ |
| 664 | if (nr == 0) { |
| 665 | atomic_dec(&isw_nr_in_flight); |
| 666 | wb_put(isw->new_wb); |
| 667 | kfree(isw); |
| 668 | return restart; |
| 669 | } |
| 670 | |
| 671 | /* |
| 672 | * In addition to synchronizing among switchers, I_WB_SWITCH tells |
| 673 | * the RCU protected stat update paths to grab the i_page |
| 674 | * lock so that stat transfer can synchronize against them. |
| 675 | * Let's continue after I_WB_SWITCH is guaranteed to be visible. |
| 676 | */ |
| 677 | INIT_RCU_WORK(&isw->work, inode_switch_wbs_work_fn); |
| 678 | queue_rcu_work(isw_wq, &isw->work); |
| 679 | |
| 680 | return restart; |
| 681 | } |
| 682 | |
| 683 | /** |
| 684 | * wbc_attach_and_unlock_inode - associate wbc with target inode and unlock it |
| 685 | * @wbc: writeback_control of interest |
| 686 | * @inode: target inode |
| 687 | * |
| 688 | * @inode is locked and about to be written back under the control of @wbc. |
| 689 | * Record @inode's writeback context into @wbc and unlock the i_lock. On |
| 690 | * writeback completion, wbc_detach_inode() should be called. This is used |
| 691 | * to track the cgroup writeback context. |
| 692 | */ |
| 693 | void wbc_attach_and_unlock_inode(struct writeback_control *wbc, |
| 694 | struct inode *inode) |
| 695 | { |
| 696 | if (!inode_cgwb_enabled(inode)) { |
| 697 | spin_unlock(&inode->i_lock); |
| 698 | return; |
| 699 | } |
| 700 | |
| 701 | wbc->wb = inode_to_wb(inode); |
| 702 | wbc->inode = inode; |
| 703 | |
| 704 | wbc->wb_id = wbc->wb->memcg_css->id; |
| 705 | wbc->wb_lcand_id = inode->i_wb_frn_winner; |
| 706 | wbc->wb_tcand_id = 0; |
| 707 | wbc->wb_bytes = 0; |
| 708 | wbc->wb_lcand_bytes = 0; |
| 709 | wbc->wb_tcand_bytes = 0; |
| 710 | |
| 711 | wb_get(wbc->wb); |
| 712 | spin_unlock(&inode->i_lock); |
| 713 | |
| 714 | /* |
| 715 | * A dying wb indicates that either the blkcg associated with the |
| 716 | * memcg changed or the associated memcg is dying. In the first |
| 717 | * case, a replacement wb should already be available and we should |
| 718 | * refresh the wb immediately. In the second case, trying to |
| 719 | * refresh will keep failing. |
| 720 | */ |
| 721 | if (unlikely(wb_dying(wbc->wb) && !css_is_dying(wbc->wb->memcg_css))) |
| 722 | inode_switch_wbs(inode, wbc->wb_id); |
| 723 | } |
| 724 | EXPORT_SYMBOL_GPL(wbc_attach_and_unlock_inode); |
| 725 | |
| 726 | /** |
| 727 | * wbc_detach_inode - disassociate wbc from inode and perform foreign detection |
| 728 | * @wbc: writeback_control of the just finished writeback |
| 729 | * |
| 730 | * To be called after a writeback attempt of an inode finishes and undoes |
| 731 | * wbc_attach_and_unlock_inode(). Can be called under any context. |
| 732 | * |
| 733 | * As concurrent write sharing of an inode is expected to be very rare and |
| 734 | * memcg only tracks page ownership on first-use basis severely confining |
| 735 | * the usefulness of such sharing, cgroup writeback tracks ownership |
| 736 | * per-inode. While the support for concurrent write sharing of an inode |
| 737 | * is deemed unnecessary, an inode being written to by different cgroups at |
| 738 | * different points in time is a lot more common, and, more importantly, |
| 739 | * charging only by first-use can too readily lead to grossly incorrect |
| 740 | * behaviors (single foreign page can lead to gigabytes of writeback to be |
| 741 | * incorrectly attributed). |
| 742 | * |
| 743 | * To resolve this issue, cgroup writeback detects the majority dirtier of |
| 744 | * an inode and transfers the ownership to it. To avoid unnecessary |
| 745 | * oscillation, the detection mechanism keeps track of history and gives |
| 746 | * out the switch verdict only if the foreign usage pattern is stable over |
| 747 | * a certain amount of time and/or writeback attempts. |
| 748 | * |
| 749 | * On each writeback attempt, @wbc tries to detect the majority writer |
| 750 | * using Boyer-Moore majority vote algorithm. In addition to the byte |
| 751 | * count from the majority voting, it also counts the bytes written for the |
| 752 | * current wb and the last round's winner wb (max of last round's current |
| 753 | * wb, the winner from two rounds ago, and the last round's majority |
| 754 | * candidate). Keeping track of the historical winner helps the algorithm |
| 755 | * to semi-reliably detect the most active writer even when it's not the |
| 756 | * absolute majority. |
| 757 | * |
| 758 | * Once the winner of the round is determined, whether the winner is |
| 759 | * foreign or not and how much IO time the round consumed is recorded in |
| 760 | * inode->i_wb_frn_history. If the amount of recorded foreign IO time is |
| 761 | * over a certain threshold, the switch verdict is given. |
| 762 | */ |
| 763 | void wbc_detach_inode(struct writeback_control *wbc) |
| 764 | { |
| 765 | struct bdi_writeback *wb = wbc->wb; |
| 766 | struct inode *inode = wbc->inode; |
| 767 | unsigned long avg_time, max_bytes, max_time; |
| 768 | u16 history; |
| 769 | int max_id; |
| 770 | |
| 771 | if (!wb) |
| 772 | return; |
| 773 | |
| 774 | history = inode->i_wb_frn_history; |
| 775 | avg_time = inode->i_wb_frn_avg_time; |
| 776 | |
| 777 | /* pick the winner of this round */ |
| 778 | if (wbc->wb_bytes >= wbc->wb_lcand_bytes && |
| 779 | wbc->wb_bytes >= wbc->wb_tcand_bytes) { |
| 780 | max_id = wbc->wb_id; |
| 781 | max_bytes = wbc->wb_bytes; |
| 782 | } else if (wbc->wb_lcand_bytes >= wbc->wb_tcand_bytes) { |
| 783 | max_id = wbc->wb_lcand_id; |
| 784 | max_bytes = wbc->wb_lcand_bytes; |
| 785 | } else { |
| 786 | max_id = wbc->wb_tcand_id; |
| 787 | max_bytes = wbc->wb_tcand_bytes; |
| 788 | } |
| 789 | |
| 790 | /* |
| 791 | * Calculate the amount of IO time the winner consumed and fold it |
| 792 | * into the running average kept per inode. If the consumed IO |
| 793 | * time is lower than avag / WB_FRN_TIME_CUT_DIV, ignore it for |
| 794 | * deciding whether to switch or not. This is to prevent one-off |
| 795 | * small dirtiers from skewing the verdict. |
| 796 | */ |
| 797 | max_time = DIV_ROUND_UP((max_bytes >> PAGE_SHIFT) << WB_FRN_TIME_SHIFT, |
| 798 | wb->avg_write_bandwidth); |
| 799 | if (avg_time) |
| 800 | avg_time += (max_time >> WB_FRN_TIME_AVG_SHIFT) - |
| 801 | (avg_time >> WB_FRN_TIME_AVG_SHIFT); |
| 802 | else |
| 803 | avg_time = max_time; /* immediate catch up on first run */ |
| 804 | |
| 805 | if (max_time >= avg_time / WB_FRN_TIME_CUT_DIV) { |
| 806 | int slots; |
| 807 | |
| 808 | /* |
| 809 | * The switch verdict is reached if foreign wb's consume |
| 810 | * more than a certain proportion of IO time in a |
| 811 | * WB_FRN_TIME_PERIOD. This is loosely tracked by 16 slot |
| 812 | * history mask where each bit represents one sixteenth of |
| 813 | * the period. Determine the number of slots to shift into |
| 814 | * history from @max_time. |
| 815 | */ |
| 816 | slots = min(DIV_ROUND_UP(max_time, WB_FRN_HIST_UNIT), |
| 817 | (unsigned long)WB_FRN_HIST_MAX_SLOTS); |
| 818 | history <<= slots; |
| 819 | if (wbc->wb_id != max_id) |
| 820 | history |= (1U << slots) - 1; |
| 821 | |
| 822 | if (history) |
| 823 | trace_inode_foreign_history(inode, wbc, history); |
| 824 | |
| 825 | /* |
| 826 | * Switch if the current wb isn't the consistent winner. |
| 827 | * If there are multiple closely competing dirtiers, the |
| 828 | * inode may switch across them repeatedly over time, which |
| 829 | * is okay. The main goal is avoiding keeping an inode on |
| 830 | * the wrong wb for an extended period of time. |
| 831 | */ |
| 832 | if (hweight16(history) > WB_FRN_HIST_THR_SLOTS) |
| 833 | inode_switch_wbs(inode, max_id); |
| 834 | } |
| 835 | |
| 836 | /* |
| 837 | * Multiple instances of this function may race to update the |
| 838 | * following fields but we don't mind occassional inaccuracies. |
| 839 | */ |
| 840 | inode->i_wb_frn_winner = max_id; |
| 841 | inode->i_wb_frn_avg_time = min(avg_time, (unsigned long)U16_MAX); |
| 842 | inode->i_wb_frn_history = history; |
| 843 | |
| 844 | wb_put(wbc->wb); |
| 845 | wbc->wb = NULL; |
| 846 | } |
| 847 | EXPORT_SYMBOL_GPL(wbc_detach_inode); |
| 848 | |
| 849 | /** |
| 850 | * wbc_account_cgroup_owner - account writeback to update inode cgroup ownership |
| 851 | * @wbc: writeback_control of the writeback in progress |
| 852 | * @page: page being written out |
| 853 | * @bytes: number of bytes being written out |
| 854 | * |
| 855 | * @bytes from @page are about to written out during the writeback |
| 856 | * controlled by @wbc. Keep the book for foreign inode detection. See |
| 857 | * wbc_detach_inode(). |
| 858 | */ |
| 859 | void wbc_account_cgroup_owner(struct writeback_control *wbc, struct page *page, |
| 860 | size_t bytes) |
| 861 | { |
| 862 | struct folio *folio; |
| 863 | struct cgroup_subsys_state *css; |
| 864 | int id; |
| 865 | |
| 866 | /* |
| 867 | * pageout() path doesn't attach @wbc to the inode being written |
| 868 | * out. This is intentional as we don't want the function to block |
| 869 | * behind a slow cgroup. Ultimately, we want pageout() to kick off |
| 870 | * regular writeback instead of writing things out itself. |
| 871 | */ |
| 872 | if (!wbc->wb || wbc->no_cgroup_owner) |
| 873 | return; |
| 874 | |
| 875 | folio = page_folio(page); |
| 876 | css = mem_cgroup_css_from_folio(folio); |
| 877 | /* dead cgroups shouldn't contribute to inode ownership arbitration */ |
| 878 | if (!(css->flags & CSS_ONLINE)) |
| 879 | return; |
| 880 | |
| 881 | id = css->id; |
| 882 | |
| 883 | if (id == wbc->wb_id) { |
| 884 | wbc->wb_bytes += bytes; |
| 885 | return; |
| 886 | } |
| 887 | |
| 888 | if (id == wbc->wb_lcand_id) |
| 889 | wbc->wb_lcand_bytes += bytes; |
| 890 | |
| 891 | /* Boyer-Moore majority vote algorithm */ |
| 892 | if (!wbc->wb_tcand_bytes) |
| 893 | wbc->wb_tcand_id = id; |
| 894 | if (id == wbc->wb_tcand_id) |
| 895 | wbc->wb_tcand_bytes += bytes; |
| 896 | else |
| 897 | wbc->wb_tcand_bytes -= min(bytes, wbc->wb_tcand_bytes); |
| 898 | } |
| 899 | EXPORT_SYMBOL_GPL(wbc_account_cgroup_owner); |
| 900 | |
| 901 | /** |
| 902 | * wb_split_bdi_pages - split nr_pages to write according to bandwidth |
| 903 | * @wb: target bdi_writeback to split @nr_pages to |
| 904 | * @nr_pages: number of pages to write for the whole bdi |
| 905 | * |
| 906 | * Split @wb's portion of @nr_pages according to @wb's write bandwidth in |
| 907 | * relation to the total write bandwidth of all wb's w/ dirty inodes on |
| 908 | * @wb->bdi. |
| 909 | */ |
| 910 | static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages) |
| 911 | { |
| 912 | unsigned long this_bw = wb->avg_write_bandwidth; |
| 913 | unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth); |
| 914 | |
| 915 | if (nr_pages == LONG_MAX) |
| 916 | return LONG_MAX; |
| 917 | |
| 918 | /* |
| 919 | * This may be called on clean wb's and proportional distribution |
| 920 | * may not make sense, just use the original @nr_pages in those |
| 921 | * cases. In general, we wanna err on the side of writing more. |
| 922 | */ |
| 923 | if (!tot_bw || this_bw >= tot_bw) |
| 924 | return nr_pages; |
| 925 | else |
| 926 | return DIV_ROUND_UP_ULL((u64)nr_pages * this_bw, tot_bw); |
| 927 | } |
| 928 | |
| 929 | /** |
| 930 | * bdi_split_work_to_wbs - split a wb_writeback_work to all wb's of a bdi |
| 931 | * @bdi: target backing_dev_info |
| 932 | * @base_work: wb_writeback_work to issue |
| 933 | * @skip_if_busy: skip wb's which already have writeback in progress |
| 934 | * |
| 935 | * Split and issue @base_work to all wb's (bdi_writeback's) of @bdi which |
| 936 | * have dirty inodes. If @base_work->nr_page isn't %LONG_MAX, it's |
| 937 | * distributed to the busy wbs according to each wb's proportion in the |
| 938 | * total active write bandwidth of @bdi. |
| 939 | */ |
| 940 | static void bdi_split_work_to_wbs(struct backing_dev_info *bdi, |
| 941 | struct wb_writeback_work *base_work, |
| 942 | bool skip_if_busy) |
| 943 | { |
| 944 | struct bdi_writeback *last_wb = NULL; |
| 945 | struct bdi_writeback *wb = list_entry(&bdi->wb_list, |
| 946 | struct bdi_writeback, bdi_node); |
| 947 | |
| 948 | might_sleep(); |
| 949 | restart: |
| 950 | rcu_read_lock(); |
| 951 | list_for_each_entry_continue_rcu(wb, &bdi->wb_list, bdi_node) { |
| 952 | DEFINE_WB_COMPLETION(fallback_work_done, bdi); |
| 953 | struct wb_writeback_work fallback_work; |
| 954 | struct wb_writeback_work *work; |
| 955 | long nr_pages; |
| 956 | |
| 957 | if (last_wb) { |
| 958 | wb_put(last_wb); |
| 959 | last_wb = NULL; |
| 960 | } |
| 961 | |
| 962 | /* SYNC_ALL writes out I_DIRTY_TIME too */ |
| 963 | if (!wb_has_dirty_io(wb) && |
| 964 | (base_work->sync_mode == WB_SYNC_NONE || |
| 965 | list_empty(&wb->b_dirty_time))) |
| 966 | continue; |
| 967 | if (skip_if_busy && writeback_in_progress(wb)) |
| 968 | continue; |
| 969 | |
| 970 | nr_pages = wb_split_bdi_pages(wb, base_work->nr_pages); |
| 971 | |
| 972 | work = kmalloc(sizeof(*work), GFP_ATOMIC); |
| 973 | if (work) { |
| 974 | *work = *base_work; |
| 975 | work->nr_pages = nr_pages; |
| 976 | work->auto_free = 1; |
| 977 | wb_queue_work(wb, work); |
| 978 | continue; |
| 979 | } |
| 980 | |
| 981 | /* |
| 982 | * If wb_tryget fails, the wb has been shutdown, skip it. |
| 983 | * |
| 984 | * Pin @wb so that it stays on @bdi->wb_list. This allows |
| 985 | * continuing iteration from @wb after dropping and |
| 986 | * regrabbing rcu read lock. |
| 987 | */ |
| 988 | if (!wb_tryget(wb)) |
| 989 | continue; |
| 990 | |
| 991 | /* alloc failed, execute synchronously using on-stack fallback */ |
| 992 | work = &fallback_work; |
| 993 | *work = *base_work; |
| 994 | work->nr_pages = nr_pages; |
| 995 | work->auto_free = 0; |
| 996 | work->done = &fallback_work_done; |
| 997 | |
| 998 | wb_queue_work(wb, work); |
| 999 | last_wb = wb; |
| 1000 | |
| 1001 | rcu_read_unlock(); |
| 1002 | wb_wait_for_completion(&fallback_work_done); |
| 1003 | goto restart; |
| 1004 | } |
| 1005 | rcu_read_unlock(); |
| 1006 | |
| 1007 | if (last_wb) |
| 1008 | wb_put(last_wb); |
| 1009 | } |
| 1010 | |
| 1011 | /** |
| 1012 | * cgroup_writeback_by_id - initiate cgroup writeback from bdi and memcg IDs |
| 1013 | * @bdi_id: target bdi id |
| 1014 | * @memcg_id: target memcg css id |
| 1015 | * @reason: reason why some writeback work initiated |
| 1016 | * @done: target wb_completion |
| 1017 | * |
| 1018 | * Initiate flush of the bdi_writeback identified by @bdi_id and @memcg_id |
| 1019 | * with the specified parameters. |
| 1020 | */ |
| 1021 | int cgroup_writeback_by_id(u64 bdi_id, int memcg_id, |
| 1022 | enum wb_reason reason, struct wb_completion *done) |
| 1023 | { |
| 1024 | struct backing_dev_info *bdi; |
| 1025 | struct cgroup_subsys_state *memcg_css; |
| 1026 | struct bdi_writeback *wb; |
| 1027 | struct wb_writeback_work *work; |
| 1028 | unsigned long dirty; |
| 1029 | int ret; |
| 1030 | |
| 1031 | /* lookup bdi and memcg */ |
| 1032 | bdi = bdi_get_by_id(bdi_id); |
| 1033 | if (!bdi) |
| 1034 | return -ENOENT; |
| 1035 | |
| 1036 | rcu_read_lock(); |
| 1037 | memcg_css = css_from_id(memcg_id, &memory_cgrp_subsys); |
| 1038 | if (memcg_css && !css_tryget(memcg_css)) |
| 1039 | memcg_css = NULL; |
| 1040 | rcu_read_unlock(); |
| 1041 | if (!memcg_css) { |
| 1042 | ret = -ENOENT; |
| 1043 | goto out_bdi_put; |
| 1044 | } |
| 1045 | |
| 1046 | /* |
| 1047 | * And find the associated wb. If the wb isn't there already |
| 1048 | * there's nothing to flush, don't create one. |
| 1049 | */ |
| 1050 | wb = wb_get_lookup(bdi, memcg_css); |
| 1051 | if (!wb) { |
| 1052 | ret = -ENOENT; |
| 1053 | goto out_css_put; |
| 1054 | } |
| 1055 | |
| 1056 | /* |
| 1057 | * The caller is attempting to write out most of |
| 1058 | * the currently dirty pages. Let's take the current dirty page |
| 1059 | * count and inflate it by 25% which should be large enough to |
| 1060 | * flush out most dirty pages while avoiding getting livelocked by |
| 1061 | * concurrent dirtiers. |
| 1062 | * |
| 1063 | * BTW the memcg stats are flushed periodically and this is best-effort |
| 1064 | * estimation, so some potential error is ok. |
| 1065 | */ |
| 1066 | dirty = memcg_page_state(mem_cgroup_from_css(memcg_css), NR_FILE_DIRTY); |
| 1067 | dirty = dirty * 10 / 8; |
| 1068 | |
| 1069 | /* issue the writeback work */ |
| 1070 | work = kzalloc(sizeof(*work), GFP_NOWAIT | __GFP_NOWARN); |
| 1071 | if (work) { |
| 1072 | work->nr_pages = dirty; |
| 1073 | work->sync_mode = WB_SYNC_NONE; |
| 1074 | work->range_cyclic = 1; |
| 1075 | work->reason = reason; |
| 1076 | work->done = done; |
| 1077 | work->auto_free = 1; |
| 1078 | wb_queue_work(wb, work); |
| 1079 | ret = 0; |
| 1080 | } else { |
| 1081 | ret = -ENOMEM; |
| 1082 | } |
| 1083 | |
| 1084 | wb_put(wb); |
| 1085 | out_css_put: |
| 1086 | css_put(memcg_css); |
| 1087 | out_bdi_put: |
| 1088 | bdi_put(bdi); |
| 1089 | return ret; |
| 1090 | } |
| 1091 | |
| 1092 | /** |
| 1093 | * cgroup_writeback_umount - flush inode wb switches for umount |
| 1094 | * |
| 1095 | * This function is called when a super_block is about to be destroyed and |
| 1096 | * flushes in-flight inode wb switches. An inode wb switch goes through |
| 1097 | * RCU and then workqueue, so the two need to be flushed in order to ensure |
| 1098 | * that all previously scheduled switches are finished. As wb switches are |
| 1099 | * rare occurrences and synchronize_rcu() can take a while, perform |
| 1100 | * flushing iff wb switches are in flight. |
| 1101 | */ |
| 1102 | void cgroup_writeback_umount(void) |
| 1103 | { |
| 1104 | /* |
| 1105 | * SB_ACTIVE should be reliably cleared before checking |
| 1106 | * isw_nr_in_flight, see generic_shutdown_super(). |
| 1107 | */ |
| 1108 | smp_mb(); |
| 1109 | |
| 1110 | if (atomic_read(&isw_nr_in_flight)) { |
| 1111 | /* |
| 1112 | * Use rcu_barrier() to wait for all pending callbacks to |
| 1113 | * ensure that all in-flight wb switches are in the workqueue. |
| 1114 | */ |
| 1115 | rcu_barrier(); |
| 1116 | flush_workqueue(isw_wq); |
| 1117 | } |
| 1118 | } |
| 1119 | |
| 1120 | static int __init cgroup_writeback_init(void) |
| 1121 | { |
| 1122 | isw_wq = alloc_workqueue("inode_switch_wbs", 0, 0); |
| 1123 | if (!isw_wq) |
| 1124 | return -ENOMEM; |
| 1125 | return 0; |
| 1126 | } |
| 1127 | fs_initcall(cgroup_writeback_init); |
| 1128 | |
| 1129 | #else /* CONFIG_CGROUP_WRITEBACK */ |
| 1130 | |
| 1131 | static void bdi_down_write_wb_switch_rwsem(struct backing_dev_info *bdi) { } |
| 1132 | static void bdi_up_write_wb_switch_rwsem(struct backing_dev_info *bdi) { } |
| 1133 | |
| 1134 | static void inode_cgwb_move_to_attached(struct inode *inode, |
| 1135 | struct bdi_writeback *wb) |
| 1136 | { |
| 1137 | assert_spin_locked(&wb->list_lock); |
| 1138 | assert_spin_locked(&inode->i_lock); |
| 1139 | WARN_ON_ONCE(inode->i_state & I_FREEING); |
| 1140 | |
| 1141 | inode->i_state &= ~I_SYNC_QUEUED; |
| 1142 | list_del_init(&inode->i_io_list); |
| 1143 | wb_io_lists_depopulated(wb); |
| 1144 | } |
| 1145 | |
| 1146 | static struct bdi_writeback * |
| 1147 | locked_inode_to_wb_and_lock_list(struct inode *inode) |
| 1148 | __releases(&inode->i_lock) |
| 1149 | __acquires(&wb->list_lock) |
| 1150 | { |
| 1151 | struct bdi_writeback *wb = inode_to_wb(inode); |
| 1152 | |
| 1153 | spin_unlock(&inode->i_lock); |
| 1154 | spin_lock(&wb->list_lock); |
| 1155 | return wb; |
| 1156 | } |
| 1157 | |
| 1158 | static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode) |
| 1159 | __acquires(&wb->list_lock) |
| 1160 | { |
| 1161 | struct bdi_writeback *wb = inode_to_wb(inode); |
| 1162 | |
| 1163 | spin_lock(&wb->list_lock); |
| 1164 | return wb; |
| 1165 | } |
| 1166 | |
| 1167 | static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages) |
| 1168 | { |
| 1169 | return nr_pages; |
| 1170 | } |
| 1171 | |
| 1172 | static void bdi_split_work_to_wbs(struct backing_dev_info *bdi, |
| 1173 | struct wb_writeback_work *base_work, |
| 1174 | bool skip_if_busy) |
| 1175 | { |
| 1176 | might_sleep(); |
| 1177 | |
| 1178 | if (!skip_if_busy || !writeback_in_progress(&bdi->wb)) { |
| 1179 | base_work->auto_free = 0; |
| 1180 | wb_queue_work(&bdi->wb, base_work); |
| 1181 | } |
| 1182 | } |
| 1183 | |
| 1184 | #endif /* CONFIG_CGROUP_WRITEBACK */ |
| 1185 | |
| 1186 | /* |
| 1187 | * Add in the number of potentially dirty inodes, because each inode |
| 1188 | * write can dirty pagecache in the underlying blockdev. |
| 1189 | */ |
| 1190 | static unsigned long get_nr_dirty_pages(void) |
| 1191 | { |
| 1192 | return global_node_page_state(NR_FILE_DIRTY) + |
| 1193 | get_nr_dirty_inodes(); |
| 1194 | } |
| 1195 | |
| 1196 | static void wb_start_writeback(struct bdi_writeback *wb, enum wb_reason reason) |
| 1197 | { |
| 1198 | if (!wb_has_dirty_io(wb)) |
| 1199 | return; |
| 1200 | |
| 1201 | /* |
| 1202 | * All callers of this function want to start writeback of all |
| 1203 | * dirty pages. Places like vmscan can call this at a very |
| 1204 | * high frequency, causing pointless allocations of tons of |
| 1205 | * work items and keeping the flusher threads busy retrieving |
| 1206 | * that work. Ensure that we only allow one of them pending and |
| 1207 | * inflight at the time. |
| 1208 | */ |
| 1209 | if (test_bit(WB_start_all, &wb->state) || |
| 1210 | test_and_set_bit(WB_start_all, &wb->state)) |
| 1211 | return; |
| 1212 | |
| 1213 | wb->start_all_reason = reason; |
| 1214 | wb_wakeup(wb); |
| 1215 | } |
| 1216 | |
| 1217 | /** |
| 1218 | * wb_start_background_writeback - start background writeback |
| 1219 | * @wb: bdi_writback to write from |
| 1220 | * |
| 1221 | * Description: |
| 1222 | * This makes sure WB_SYNC_NONE background writeback happens. When |
| 1223 | * this function returns, it is only guaranteed that for given wb |
| 1224 | * some IO is happening if we are over background dirty threshold. |
| 1225 | * Caller need not hold sb s_umount semaphore. |
| 1226 | */ |
| 1227 | void wb_start_background_writeback(struct bdi_writeback *wb) |
| 1228 | { |
| 1229 | /* |
| 1230 | * We just wake up the flusher thread. It will perform background |
| 1231 | * writeback as soon as there is no other work to do. |
| 1232 | */ |
| 1233 | trace_writeback_wake_background(wb); |
| 1234 | wb_wakeup(wb); |
| 1235 | } |
| 1236 | |
| 1237 | /* |
| 1238 | * Remove the inode from the writeback list it is on. |
| 1239 | */ |
| 1240 | void inode_io_list_del(struct inode *inode) |
| 1241 | { |
| 1242 | struct bdi_writeback *wb; |
| 1243 | |
| 1244 | wb = inode_to_wb_and_lock_list(inode); |
| 1245 | spin_lock(&inode->i_lock); |
| 1246 | |
| 1247 | inode->i_state &= ~I_SYNC_QUEUED; |
| 1248 | list_del_init(&inode->i_io_list); |
| 1249 | wb_io_lists_depopulated(wb); |
| 1250 | |
| 1251 | spin_unlock(&inode->i_lock); |
| 1252 | spin_unlock(&wb->list_lock); |
| 1253 | } |
| 1254 | EXPORT_SYMBOL(inode_io_list_del); |
| 1255 | |
| 1256 | /* |
| 1257 | * mark an inode as under writeback on the sb |
| 1258 | */ |
| 1259 | void sb_mark_inode_writeback(struct inode *inode) |
| 1260 | { |
| 1261 | struct super_block *sb = inode->i_sb; |
| 1262 | unsigned long flags; |
| 1263 | |
| 1264 | if (list_empty(&inode->i_wb_list)) { |
| 1265 | spin_lock_irqsave(&sb->s_inode_wblist_lock, flags); |
| 1266 | if (list_empty(&inode->i_wb_list)) { |
| 1267 | list_add_tail(&inode->i_wb_list, &sb->s_inodes_wb); |
| 1268 | trace_sb_mark_inode_writeback(inode); |
| 1269 | } |
| 1270 | spin_unlock_irqrestore(&sb->s_inode_wblist_lock, flags); |
| 1271 | } |
| 1272 | } |
| 1273 | |
| 1274 | /* |
| 1275 | * clear an inode as under writeback on the sb |
| 1276 | */ |
| 1277 | void sb_clear_inode_writeback(struct inode *inode) |
| 1278 | { |
| 1279 | struct super_block *sb = inode->i_sb; |
| 1280 | unsigned long flags; |
| 1281 | |
| 1282 | if (!list_empty(&inode->i_wb_list)) { |
| 1283 | spin_lock_irqsave(&sb->s_inode_wblist_lock, flags); |
| 1284 | if (!list_empty(&inode->i_wb_list)) { |
| 1285 | list_del_init(&inode->i_wb_list); |
| 1286 | trace_sb_clear_inode_writeback(inode); |
| 1287 | } |
| 1288 | spin_unlock_irqrestore(&sb->s_inode_wblist_lock, flags); |
| 1289 | } |
| 1290 | } |
| 1291 | |
| 1292 | /* |
| 1293 | * Redirty an inode: set its when-it-was dirtied timestamp and move it to the |
| 1294 | * furthest end of its superblock's dirty-inode list. |
| 1295 | * |
| 1296 | * Before stamping the inode's ->dirtied_when, we check to see whether it is |
| 1297 | * already the most-recently-dirtied inode on the b_dirty list. If that is |
| 1298 | * the case then the inode must have been redirtied while it was being written |
| 1299 | * out and we don't reset its dirtied_when. |
| 1300 | */ |
| 1301 | static void redirty_tail_locked(struct inode *inode, struct bdi_writeback *wb) |
| 1302 | { |
| 1303 | assert_spin_locked(&inode->i_lock); |
| 1304 | |
| 1305 | inode->i_state &= ~I_SYNC_QUEUED; |
| 1306 | /* |
| 1307 | * When the inode is being freed just don't bother with dirty list |
| 1308 | * tracking. Flush worker will ignore this inode anyway and it will |
| 1309 | * trigger assertions in inode_io_list_move_locked(). |
| 1310 | */ |
| 1311 | if (inode->i_state & I_FREEING) { |
| 1312 | list_del_init(&inode->i_io_list); |
| 1313 | wb_io_lists_depopulated(wb); |
| 1314 | return; |
| 1315 | } |
| 1316 | if (!list_empty(&wb->b_dirty)) { |
| 1317 | struct inode *tail; |
| 1318 | |
| 1319 | tail = wb_inode(wb->b_dirty.next); |
| 1320 | if (time_before(inode->dirtied_when, tail->dirtied_when)) |
| 1321 | inode->dirtied_when = jiffies; |
| 1322 | } |
| 1323 | inode_io_list_move_locked(inode, wb, &wb->b_dirty); |
| 1324 | } |
| 1325 | |
| 1326 | static void redirty_tail(struct inode *inode, struct bdi_writeback *wb) |
| 1327 | { |
| 1328 | spin_lock(&inode->i_lock); |
| 1329 | redirty_tail_locked(inode, wb); |
| 1330 | spin_unlock(&inode->i_lock); |
| 1331 | } |
| 1332 | |
| 1333 | /* |
| 1334 | * requeue inode for re-scanning after bdi->b_io list is exhausted. |
| 1335 | */ |
| 1336 | static void requeue_io(struct inode *inode, struct bdi_writeback *wb) |
| 1337 | { |
| 1338 | inode_io_list_move_locked(inode, wb, &wb->b_more_io); |
| 1339 | } |
| 1340 | |
| 1341 | static void inode_sync_complete(struct inode *inode) |
| 1342 | { |
| 1343 | inode->i_state &= ~I_SYNC; |
| 1344 | /* If inode is clean an unused, put it into LRU now... */ |
| 1345 | inode_add_lru(inode); |
| 1346 | /* Waiters must see I_SYNC cleared before being woken up */ |
| 1347 | smp_mb(); |
| 1348 | wake_up_bit(&inode->i_state, __I_SYNC); |
| 1349 | } |
| 1350 | |
| 1351 | static bool inode_dirtied_after(struct inode *inode, unsigned long t) |
| 1352 | { |
| 1353 | bool ret = time_after(inode->dirtied_when, t); |
| 1354 | #ifndef CONFIG_64BIT |
| 1355 | /* |
| 1356 | * For inodes being constantly redirtied, dirtied_when can get stuck. |
| 1357 | * It _appears_ to be in the future, but is actually in distant past. |
| 1358 | * This test is necessary to prevent such wrapped-around relative times |
| 1359 | * from permanently stopping the whole bdi writeback. |
| 1360 | */ |
| 1361 | ret = ret && time_before_eq(inode->dirtied_when, jiffies); |
| 1362 | #endif |
| 1363 | return ret; |
| 1364 | } |
| 1365 | |
| 1366 | /* |
| 1367 | * Move expired (dirtied before dirtied_before) dirty inodes from |
| 1368 | * @delaying_queue to @dispatch_queue. |
| 1369 | */ |
| 1370 | static int move_expired_inodes(struct list_head *delaying_queue, |
| 1371 | struct list_head *dispatch_queue, |
| 1372 | unsigned long dirtied_before) |
| 1373 | { |
| 1374 | LIST_HEAD(tmp); |
| 1375 | struct list_head *pos, *node; |
| 1376 | struct super_block *sb = NULL; |
| 1377 | struct inode *inode; |
| 1378 | int do_sb_sort = 0; |
| 1379 | int moved = 0; |
| 1380 | |
| 1381 | while (!list_empty(delaying_queue)) { |
| 1382 | inode = wb_inode(delaying_queue->prev); |
| 1383 | if (inode_dirtied_after(inode, dirtied_before)) |
| 1384 | break; |
| 1385 | spin_lock(&inode->i_lock); |
| 1386 | list_move(&inode->i_io_list, &tmp); |
| 1387 | moved++; |
| 1388 | inode->i_state |= I_SYNC_QUEUED; |
| 1389 | spin_unlock(&inode->i_lock); |
| 1390 | if (sb_is_blkdev_sb(inode->i_sb)) |
| 1391 | continue; |
| 1392 | if (sb && sb != inode->i_sb) |
| 1393 | do_sb_sort = 1; |
| 1394 | sb = inode->i_sb; |
| 1395 | } |
| 1396 | |
| 1397 | /* just one sb in list, splice to dispatch_queue and we're done */ |
| 1398 | if (!do_sb_sort) { |
| 1399 | list_splice(&tmp, dispatch_queue); |
| 1400 | goto out; |
| 1401 | } |
| 1402 | |
| 1403 | /* |
| 1404 | * Although inode's i_io_list is moved from 'tmp' to 'dispatch_queue', |
| 1405 | * we don't take inode->i_lock here because it is just a pointless overhead. |
| 1406 | * Inode is already marked as I_SYNC_QUEUED so writeback list handling is |
| 1407 | * fully under our control. |
| 1408 | */ |
| 1409 | while (!list_empty(&tmp)) { |
| 1410 | sb = wb_inode(tmp.prev)->i_sb; |
| 1411 | list_for_each_prev_safe(pos, node, &tmp) { |
| 1412 | inode = wb_inode(pos); |
| 1413 | if (inode->i_sb == sb) |
| 1414 | list_move(&inode->i_io_list, dispatch_queue); |
| 1415 | } |
| 1416 | } |
| 1417 | out: |
| 1418 | return moved; |
| 1419 | } |
| 1420 | |
| 1421 | /* |
| 1422 | * Queue all expired dirty inodes for io, eldest first. |
| 1423 | * Before |
| 1424 | * newly dirtied b_dirty b_io b_more_io |
| 1425 | * =============> gf edc BA |
| 1426 | * After |
| 1427 | * newly dirtied b_dirty b_io b_more_io |
| 1428 | * =============> g fBAedc |
| 1429 | * | |
| 1430 | * +--> dequeue for IO |
| 1431 | */ |
| 1432 | static void queue_io(struct bdi_writeback *wb, struct wb_writeback_work *work, |
| 1433 | unsigned long dirtied_before) |
| 1434 | { |
| 1435 | int moved; |
| 1436 | unsigned long time_expire_jif = dirtied_before; |
| 1437 | |
| 1438 | assert_spin_locked(&wb->list_lock); |
| 1439 | list_splice_init(&wb->b_more_io, &wb->b_io); |
| 1440 | moved = move_expired_inodes(&wb->b_dirty, &wb->b_io, dirtied_before); |
| 1441 | if (!work->for_sync) |
| 1442 | time_expire_jif = jiffies - dirtytime_expire_interval * HZ; |
| 1443 | moved += move_expired_inodes(&wb->b_dirty_time, &wb->b_io, |
| 1444 | time_expire_jif); |
| 1445 | if (moved) |
| 1446 | wb_io_lists_populated(wb); |
| 1447 | trace_writeback_queue_io(wb, work, dirtied_before, moved); |
| 1448 | } |
| 1449 | |
| 1450 | static int write_inode(struct inode *inode, struct writeback_control *wbc) |
| 1451 | { |
| 1452 | int ret; |
| 1453 | |
| 1454 | if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode)) { |
| 1455 | trace_writeback_write_inode_start(inode, wbc); |
| 1456 | ret = inode->i_sb->s_op->write_inode(inode, wbc); |
| 1457 | trace_writeback_write_inode(inode, wbc); |
| 1458 | return ret; |
| 1459 | } |
| 1460 | return 0; |
| 1461 | } |
| 1462 | |
| 1463 | /* |
| 1464 | * Wait for writeback on an inode to complete. Called with i_lock held. |
| 1465 | * Caller must make sure inode cannot go away when we drop i_lock. |
| 1466 | */ |
| 1467 | static void __inode_wait_for_writeback(struct inode *inode) |
| 1468 | __releases(inode->i_lock) |
| 1469 | __acquires(inode->i_lock) |
| 1470 | { |
| 1471 | DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC); |
| 1472 | wait_queue_head_t *wqh; |
| 1473 | |
| 1474 | wqh = bit_waitqueue(&inode->i_state, __I_SYNC); |
| 1475 | while (inode->i_state & I_SYNC) { |
| 1476 | spin_unlock(&inode->i_lock); |
| 1477 | __wait_on_bit(wqh, &wq, bit_wait, |
| 1478 | TASK_UNINTERRUPTIBLE); |
| 1479 | spin_lock(&inode->i_lock); |
| 1480 | } |
| 1481 | } |
| 1482 | |
| 1483 | /* |
| 1484 | * Wait for writeback on an inode to complete. Caller must have inode pinned. |
| 1485 | */ |
| 1486 | void inode_wait_for_writeback(struct inode *inode) |
| 1487 | { |
| 1488 | spin_lock(&inode->i_lock); |
| 1489 | __inode_wait_for_writeback(inode); |
| 1490 | spin_unlock(&inode->i_lock); |
| 1491 | } |
| 1492 | |
| 1493 | /* |
| 1494 | * Sleep until I_SYNC is cleared. This function must be called with i_lock |
| 1495 | * held and drops it. It is aimed for callers not holding any inode reference |
| 1496 | * so once i_lock is dropped, inode can go away. |
| 1497 | */ |
| 1498 | static void inode_sleep_on_writeback(struct inode *inode) |
| 1499 | __releases(inode->i_lock) |
| 1500 | { |
| 1501 | DEFINE_WAIT(wait); |
| 1502 | wait_queue_head_t *wqh = bit_waitqueue(&inode->i_state, __I_SYNC); |
| 1503 | int sleep; |
| 1504 | |
| 1505 | prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE); |
| 1506 | sleep = inode->i_state & I_SYNC; |
| 1507 | spin_unlock(&inode->i_lock); |
| 1508 | if (sleep) |
| 1509 | schedule(); |
| 1510 | finish_wait(wqh, &wait); |
| 1511 | } |
| 1512 | |
| 1513 | /* |
| 1514 | * Find proper writeback list for the inode depending on its current state and |
| 1515 | * possibly also change of its state while we were doing writeback. Here we |
| 1516 | * handle things such as livelock prevention or fairness of writeback among |
| 1517 | * inodes. This function can be called only by flusher thread - noone else |
| 1518 | * processes all inodes in writeback lists and requeueing inodes behind flusher |
| 1519 | * thread's back can have unexpected consequences. |
| 1520 | */ |
| 1521 | static void requeue_inode(struct inode *inode, struct bdi_writeback *wb, |
| 1522 | struct writeback_control *wbc) |
| 1523 | { |
| 1524 | if (inode->i_state & I_FREEING) |
| 1525 | return; |
| 1526 | |
| 1527 | /* |
| 1528 | * Sync livelock prevention. Each inode is tagged and synced in one |
| 1529 | * shot. If still dirty, it will be redirty_tail()'ed below. Update |
| 1530 | * the dirty time to prevent enqueue and sync it again. |
| 1531 | */ |
| 1532 | if ((inode->i_state & I_DIRTY) && |
| 1533 | (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)) |
| 1534 | inode->dirtied_when = jiffies; |
| 1535 | |
| 1536 | if (wbc->pages_skipped) { |
| 1537 | /* |
| 1538 | * writeback is not making progress due to locked |
| 1539 | * buffers. Skip this inode for now. |
| 1540 | */ |
| 1541 | redirty_tail_locked(inode, wb); |
| 1542 | return; |
| 1543 | } |
| 1544 | |
| 1545 | if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_DIRTY)) { |
| 1546 | /* |
| 1547 | * We didn't write back all the pages. nfs_writepages() |
| 1548 | * sometimes bales out without doing anything. |
| 1549 | */ |
| 1550 | if (wbc->nr_to_write <= 0) { |
| 1551 | /* Slice used up. Queue for next turn. */ |
| 1552 | requeue_io(inode, wb); |
| 1553 | } else { |
| 1554 | /* |
| 1555 | * Writeback blocked by something other than |
| 1556 | * congestion. Delay the inode for some time to |
| 1557 | * avoid spinning on the CPU (100% iowait) |
| 1558 | * retrying writeback of the dirty page/inode |
| 1559 | * that cannot be performed immediately. |
| 1560 | */ |
| 1561 | redirty_tail_locked(inode, wb); |
| 1562 | } |
| 1563 | } else if (inode->i_state & I_DIRTY) { |
| 1564 | /* |
| 1565 | * Filesystems can dirty the inode during writeback operations, |
| 1566 | * such as delayed allocation during submission or metadata |
| 1567 | * updates after data IO completion. |
| 1568 | */ |
| 1569 | redirty_tail_locked(inode, wb); |
| 1570 | } else if (inode->i_state & I_DIRTY_TIME) { |
| 1571 | inode->dirtied_when = jiffies; |
| 1572 | inode_io_list_move_locked(inode, wb, &wb->b_dirty_time); |
| 1573 | inode->i_state &= ~I_SYNC_QUEUED; |
| 1574 | } else { |
| 1575 | /* The inode is clean. Remove from writeback lists. */ |
| 1576 | inode_cgwb_move_to_attached(inode, wb); |
| 1577 | } |
| 1578 | } |
| 1579 | |
| 1580 | /* |
| 1581 | * Write out an inode and its dirty pages (or some of its dirty pages, depending |
| 1582 | * on @wbc->nr_to_write), and clear the relevant dirty flags from i_state. |
| 1583 | * |
| 1584 | * This doesn't remove the inode from the writeback list it is on, except |
| 1585 | * potentially to move it from b_dirty_time to b_dirty due to timestamp |
| 1586 | * expiration. The caller is otherwise responsible for writeback list handling. |
| 1587 | * |
| 1588 | * The caller is also responsible for setting the I_SYNC flag beforehand and |
| 1589 | * calling inode_sync_complete() to clear it afterwards. |
| 1590 | */ |
| 1591 | static int |
| 1592 | __writeback_single_inode(struct inode *inode, struct writeback_control *wbc) |
| 1593 | { |
| 1594 | struct address_space *mapping = inode->i_mapping; |
| 1595 | long nr_to_write = wbc->nr_to_write; |
| 1596 | unsigned dirty; |
| 1597 | int ret; |
| 1598 | |
| 1599 | WARN_ON(!(inode->i_state & I_SYNC)); |
| 1600 | |
| 1601 | trace_writeback_single_inode_start(inode, wbc, nr_to_write); |
| 1602 | |
| 1603 | ret = do_writepages(mapping, wbc); |
| 1604 | |
| 1605 | /* |
| 1606 | * Make sure to wait on the data before writing out the metadata. |
| 1607 | * This is important for filesystems that modify metadata on data |
| 1608 | * I/O completion. We don't do it for sync(2) writeback because it has a |
| 1609 | * separate, external IO completion path and ->sync_fs for guaranteeing |
| 1610 | * inode metadata is written back correctly. |
| 1611 | */ |
| 1612 | if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync) { |
| 1613 | int err = filemap_fdatawait(mapping); |
| 1614 | if (ret == 0) |
| 1615 | ret = err; |
| 1616 | } |
| 1617 | |
| 1618 | /* |
| 1619 | * If the inode has dirty timestamps and we need to write them, call |
| 1620 | * mark_inode_dirty_sync() to notify the filesystem about it and to |
| 1621 | * change I_DIRTY_TIME into I_DIRTY_SYNC. |
| 1622 | */ |
| 1623 | if ((inode->i_state & I_DIRTY_TIME) && |
| 1624 | (wbc->sync_mode == WB_SYNC_ALL || |
| 1625 | time_after(jiffies, inode->dirtied_time_when + |
| 1626 | dirtytime_expire_interval * HZ))) { |
| 1627 | trace_writeback_lazytime(inode); |
| 1628 | mark_inode_dirty_sync(inode); |
| 1629 | } |
| 1630 | |
| 1631 | /* |
| 1632 | * Get and clear the dirty flags from i_state. This needs to be done |
| 1633 | * after calling writepages because some filesystems may redirty the |
| 1634 | * inode during writepages due to delalloc. It also needs to be done |
| 1635 | * after handling timestamp expiration, as that may dirty the inode too. |
| 1636 | */ |
| 1637 | spin_lock(&inode->i_lock); |
| 1638 | dirty = inode->i_state & I_DIRTY; |
| 1639 | inode->i_state &= ~dirty; |
| 1640 | |
| 1641 | /* |
| 1642 | * Paired with smp_mb() in __mark_inode_dirty(). This allows |
| 1643 | * __mark_inode_dirty() to test i_state without grabbing i_lock - |
| 1644 | * either they see the I_DIRTY bits cleared or we see the dirtied |
| 1645 | * inode. |
| 1646 | * |
| 1647 | * I_DIRTY_PAGES is always cleared together above even if @mapping |
| 1648 | * still has dirty pages. The flag is reinstated after smp_mb() if |
| 1649 | * necessary. This guarantees that either __mark_inode_dirty() |
| 1650 | * sees clear I_DIRTY_PAGES or we see PAGECACHE_TAG_DIRTY. |
| 1651 | */ |
| 1652 | smp_mb(); |
| 1653 | |
| 1654 | if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) |
| 1655 | inode->i_state |= I_DIRTY_PAGES; |
| 1656 | else if (unlikely(inode->i_state & I_PINNING_FSCACHE_WB)) { |
| 1657 | if (!(inode->i_state & I_DIRTY_PAGES)) { |
| 1658 | inode->i_state &= ~I_PINNING_FSCACHE_WB; |
| 1659 | wbc->unpinned_fscache_wb = true; |
| 1660 | dirty |= I_PINNING_FSCACHE_WB; /* Cause write_inode */ |
| 1661 | } |
| 1662 | } |
| 1663 | |
| 1664 | spin_unlock(&inode->i_lock); |
| 1665 | |
| 1666 | /* Don't write the inode if only I_DIRTY_PAGES was set */ |
| 1667 | if (dirty & ~I_DIRTY_PAGES) { |
| 1668 | int err = write_inode(inode, wbc); |
| 1669 | if (ret == 0) |
| 1670 | ret = err; |
| 1671 | } |
| 1672 | wbc->unpinned_fscache_wb = false; |
| 1673 | trace_writeback_single_inode(inode, wbc, nr_to_write); |
| 1674 | return ret; |
| 1675 | } |
| 1676 | |
| 1677 | /* |
| 1678 | * Write out an inode's dirty data and metadata on-demand, i.e. separately from |
| 1679 | * the regular batched writeback done by the flusher threads in |
| 1680 | * writeback_sb_inodes(). @wbc controls various aspects of the write, such as |
| 1681 | * whether it is a data-integrity sync (%WB_SYNC_ALL) or not (%WB_SYNC_NONE). |
| 1682 | * |
| 1683 | * To prevent the inode from going away, either the caller must have a reference |
| 1684 | * to the inode, or the inode must have I_WILL_FREE or I_FREEING set. |
| 1685 | */ |
| 1686 | static int writeback_single_inode(struct inode *inode, |
| 1687 | struct writeback_control *wbc) |
| 1688 | { |
| 1689 | struct bdi_writeback *wb; |
| 1690 | int ret = 0; |
| 1691 | |
| 1692 | spin_lock(&inode->i_lock); |
| 1693 | if (!atomic_read(&inode->i_count)) |
| 1694 | WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING))); |
| 1695 | else |
| 1696 | WARN_ON(inode->i_state & I_WILL_FREE); |
| 1697 | |
| 1698 | if (inode->i_state & I_SYNC) { |
| 1699 | /* |
| 1700 | * Writeback is already running on the inode. For WB_SYNC_NONE, |
| 1701 | * that's enough and we can just return. For WB_SYNC_ALL, we |
| 1702 | * must wait for the existing writeback to complete, then do |
| 1703 | * writeback again if there's anything left. |
| 1704 | */ |
| 1705 | if (wbc->sync_mode != WB_SYNC_ALL) |
| 1706 | goto out; |
| 1707 | __inode_wait_for_writeback(inode); |
| 1708 | } |
| 1709 | WARN_ON(inode->i_state & I_SYNC); |
| 1710 | /* |
| 1711 | * If the inode is already fully clean, then there's nothing to do. |
| 1712 | * |
| 1713 | * For data-integrity syncs we also need to check whether any pages are |
| 1714 | * still under writeback, e.g. due to prior WB_SYNC_NONE writeback. If |
| 1715 | * there are any such pages, we'll need to wait for them. |
| 1716 | */ |
| 1717 | if (!(inode->i_state & I_DIRTY_ALL) && |
| 1718 | (wbc->sync_mode != WB_SYNC_ALL || |
| 1719 | !mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK))) |
| 1720 | goto out; |
| 1721 | inode->i_state |= I_SYNC; |
| 1722 | wbc_attach_and_unlock_inode(wbc, inode); |
| 1723 | |
| 1724 | ret = __writeback_single_inode(inode, wbc); |
| 1725 | |
| 1726 | wbc_detach_inode(wbc); |
| 1727 | |
| 1728 | wb = inode_to_wb_and_lock_list(inode); |
| 1729 | spin_lock(&inode->i_lock); |
| 1730 | /* |
| 1731 | * If the inode is freeing, its i_io_list shoudn't be updated |
| 1732 | * as it can be finally deleted at this moment. |
| 1733 | */ |
| 1734 | if (!(inode->i_state & I_FREEING)) { |
| 1735 | /* |
| 1736 | * If the inode is now fully clean, then it can be safely |
| 1737 | * removed from its writeback list (if any). Otherwise the |
| 1738 | * flusher threads are responsible for the writeback lists. |
| 1739 | */ |
| 1740 | if (!(inode->i_state & I_DIRTY_ALL)) |
| 1741 | inode_cgwb_move_to_attached(inode, wb); |
| 1742 | else if (!(inode->i_state & I_SYNC_QUEUED)) { |
| 1743 | if ((inode->i_state & I_DIRTY)) |
| 1744 | redirty_tail_locked(inode, wb); |
| 1745 | else if (inode->i_state & I_DIRTY_TIME) { |
| 1746 | inode->dirtied_when = jiffies; |
| 1747 | inode_io_list_move_locked(inode, |
| 1748 | wb, |
| 1749 | &wb->b_dirty_time); |
| 1750 | } |
| 1751 | } |
| 1752 | } |
| 1753 | |
| 1754 | spin_unlock(&wb->list_lock); |
| 1755 | inode_sync_complete(inode); |
| 1756 | out: |
| 1757 | spin_unlock(&inode->i_lock); |
| 1758 | return ret; |
| 1759 | } |
| 1760 | |
| 1761 | static long writeback_chunk_size(struct bdi_writeback *wb, |
| 1762 | struct wb_writeback_work *work) |
| 1763 | { |
| 1764 | long pages; |
| 1765 | |
| 1766 | /* |
| 1767 | * WB_SYNC_ALL mode does livelock avoidance by syncing dirty |
| 1768 | * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX |
| 1769 | * here avoids calling into writeback_inodes_wb() more than once. |
| 1770 | * |
| 1771 | * The intended call sequence for WB_SYNC_ALL writeback is: |
| 1772 | * |
| 1773 | * wb_writeback() |
| 1774 | * writeback_sb_inodes() <== called only once |
| 1775 | * write_cache_pages() <== called once for each inode |
| 1776 | * (quickly) tag currently dirty pages |
| 1777 | * (maybe slowly) sync all tagged pages |
| 1778 | */ |
| 1779 | if (work->sync_mode == WB_SYNC_ALL || work->tagged_writepages) |
| 1780 | pages = LONG_MAX; |
| 1781 | else { |
| 1782 | pages = min(wb->avg_write_bandwidth / 2, |
| 1783 | global_wb_domain.dirty_limit / DIRTY_SCOPE); |
| 1784 | pages = min(pages, work->nr_pages); |
| 1785 | pages = round_down(pages + MIN_WRITEBACK_PAGES, |
| 1786 | MIN_WRITEBACK_PAGES); |
| 1787 | } |
| 1788 | |
| 1789 | return pages; |
| 1790 | } |
| 1791 | |
| 1792 | /* |
| 1793 | * Write a portion of b_io inodes which belong to @sb. |
| 1794 | * |
| 1795 | * Return the number of pages and/or inodes written. |
| 1796 | * |
| 1797 | * NOTE! This is called with wb->list_lock held, and will |
| 1798 | * unlock and relock that for each inode it ends up doing |
| 1799 | * IO for. |
| 1800 | */ |
| 1801 | static long writeback_sb_inodes(struct super_block *sb, |
| 1802 | struct bdi_writeback *wb, |
| 1803 | struct wb_writeback_work *work) |
| 1804 | { |
| 1805 | struct writeback_control wbc = { |
| 1806 | .sync_mode = work->sync_mode, |
| 1807 | .tagged_writepages = work->tagged_writepages, |
| 1808 | .for_kupdate = work->for_kupdate, |
| 1809 | .for_background = work->for_background, |
| 1810 | .for_sync = work->for_sync, |
| 1811 | .range_cyclic = work->range_cyclic, |
| 1812 | .range_start = 0, |
| 1813 | .range_end = LLONG_MAX, |
| 1814 | }; |
| 1815 | unsigned long start_time = jiffies; |
| 1816 | long write_chunk; |
| 1817 | long total_wrote = 0; /* count both pages and inodes */ |
| 1818 | |
| 1819 | while (!list_empty(&wb->b_io)) { |
| 1820 | struct inode *inode = wb_inode(wb->b_io.prev); |
| 1821 | struct bdi_writeback *tmp_wb; |
| 1822 | long wrote; |
| 1823 | |
| 1824 | if (inode->i_sb != sb) { |
| 1825 | if (work->sb) { |
| 1826 | /* |
| 1827 | * We only want to write back data for this |
| 1828 | * superblock, move all inodes not belonging |
| 1829 | * to it back onto the dirty list. |
| 1830 | */ |
| 1831 | redirty_tail(inode, wb); |
| 1832 | continue; |
| 1833 | } |
| 1834 | |
| 1835 | /* |
| 1836 | * The inode belongs to a different superblock. |
| 1837 | * Bounce back to the caller to unpin this and |
| 1838 | * pin the next superblock. |
| 1839 | */ |
| 1840 | break; |
| 1841 | } |
| 1842 | |
| 1843 | /* |
| 1844 | * Don't bother with new inodes or inodes being freed, first |
| 1845 | * kind does not need periodic writeout yet, and for the latter |
| 1846 | * kind writeout is handled by the freer. |
| 1847 | */ |
| 1848 | spin_lock(&inode->i_lock); |
| 1849 | if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) { |
| 1850 | redirty_tail_locked(inode, wb); |
| 1851 | spin_unlock(&inode->i_lock); |
| 1852 | continue; |
| 1853 | } |
| 1854 | if ((inode->i_state & I_SYNC) && wbc.sync_mode != WB_SYNC_ALL) { |
| 1855 | /* |
| 1856 | * If this inode is locked for writeback and we are not |
| 1857 | * doing writeback-for-data-integrity, move it to |
| 1858 | * b_more_io so that writeback can proceed with the |
| 1859 | * other inodes on s_io. |
| 1860 | * |
| 1861 | * We'll have another go at writing back this inode |
| 1862 | * when we completed a full scan of b_io. |
| 1863 | */ |
| 1864 | requeue_io(inode, wb); |
| 1865 | spin_unlock(&inode->i_lock); |
| 1866 | trace_writeback_sb_inodes_requeue(inode); |
| 1867 | continue; |
| 1868 | } |
| 1869 | spin_unlock(&wb->list_lock); |
| 1870 | |
| 1871 | /* |
| 1872 | * We already requeued the inode if it had I_SYNC set and we |
| 1873 | * are doing WB_SYNC_NONE writeback. So this catches only the |
| 1874 | * WB_SYNC_ALL case. |
| 1875 | */ |
| 1876 | if (inode->i_state & I_SYNC) { |
| 1877 | /* Wait for I_SYNC. This function drops i_lock... */ |
| 1878 | inode_sleep_on_writeback(inode); |
| 1879 | /* Inode may be gone, start again */ |
| 1880 | spin_lock(&wb->list_lock); |
| 1881 | continue; |
| 1882 | } |
| 1883 | inode->i_state |= I_SYNC; |
| 1884 | wbc_attach_and_unlock_inode(&wbc, inode); |
| 1885 | |
| 1886 | write_chunk = writeback_chunk_size(wb, work); |
| 1887 | wbc.nr_to_write = write_chunk; |
| 1888 | wbc.pages_skipped = 0; |
| 1889 | |
| 1890 | /* |
| 1891 | * We use I_SYNC to pin the inode in memory. While it is set |
| 1892 | * evict_inode() will wait so the inode cannot be freed. |
| 1893 | */ |
| 1894 | __writeback_single_inode(inode, &wbc); |
| 1895 | |
| 1896 | wbc_detach_inode(&wbc); |
| 1897 | work->nr_pages -= write_chunk - wbc.nr_to_write; |
| 1898 | wrote = write_chunk - wbc.nr_to_write - wbc.pages_skipped; |
| 1899 | wrote = wrote < 0 ? 0 : wrote; |
| 1900 | total_wrote += wrote; |
| 1901 | |
| 1902 | if (need_resched()) { |
| 1903 | /* |
| 1904 | * We're trying to balance between building up a nice |
| 1905 | * long list of IOs to improve our merge rate, and |
| 1906 | * getting those IOs out quickly for anyone throttling |
| 1907 | * in balance_dirty_pages(). cond_resched() doesn't |
| 1908 | * unplug, so get our IOs out the door before we |
| 1909 | * give up the CPU. |
| 1910 | */ |
| 1911 | blk_flush_plug(current->plug, false); |
| 1912 | cond_resched(); |
| 1913 | } |
| 1914 | |
| 1915 | /* |
| 1916 | * Requeue @inode if still dirty. Be careful as @inode may |
| 1917 | * have been switched to another wb in the meantime. |
| 1918 | */ |
| 1919 | tmp_wb = inode_to_wb_and_lock_list(inode); |
| 1920 | spin_lock(&inode->i_lock); |
| 1921 | if (!(inode->i_state & I_DIRTY_ALL)) |
| 1922 | total_wrote++; |
| 1923 | requeue_inode(inode, tmp_wb, &wbc); |
| 1924 | inode_sync_complete(inode); |
| 1925 | spin_unlock(&inode->i_lock); |
| 1926 | |
| 1927 | if (unlikely(tmp_wb != wb)) { |
| 1928 | spin_unlock(&tmp_wb->list_lock); |
| 1929 | spin_lock(&wb->list_lock); |
| 1930 | } |
| 1931 | |
| 1932 | /* |
| 1933 | * bail out to wb_writeback() often enough to check |
| 1934 | * background threshold and other termination conditions. |
| 1935 | */ |
| 1936 | if (total_wrote) { |
| 1937 | if (time_is_before_jiffies(start_time + HZ / 10UL)) |
| 1938 | break; |
| 1939 | if (work->nr_pages <= 0) |
| 1940 | break; |
| 1941 | } |
| 1942 | } |
| 1943 | return total_wrote; |
| 1944 | } |
| 1945 | |
| 1946 | static long __writeback_inodes_wb(struct bdi_writeback *wb, |
| 1947 | struct wb_writeback_work *work) |
| 1948 | { |
| 1949 | unsigned long start_time = jiffies; |
| 1950 | long wrote = 0; |
| 1951 | |
| 1952 | while (!list_empty(&wb->b_io)) { |
| 1953 | struct inode *inode = wb_inode(wb->b_io.prev); |
| 1954 | struct super_block *sb = inode->i_sb; |
| 1955 | |
| 1956 | if (!trylock_super(sb)) { |
| 1957 | /* |
| 1958 | * trylock_super() may fail consistently due to |
| 1959 | * s_umount being grabbed by someone else. Don't use |
| 1960 | * requeue_io() to avoid busy retrying the inode/sb. |
| 1961 | */ |
| 1962 | redirty_tail(inode, wb); |
| 1963 | continue; |
| 1964 | } |
| 1965 | wrote += writeback_sb_inodes(sb, wb, work); |
| 1966 | up_read(&sb->s_umount); |
| 1967 | |
| 1968 | /* refer to the same tests at the end of writeback_sb_inodes */ |
| 1969 | if (wrote) { |
| 1970 | if (time_is_before_jiffies(start_time + HZ / 10UL)) |
| 1971 | break; |
| 1972 | if (work->nr_pages <= 0) |
| 1973 | break; |
| 1974 | } |
| 1975 | } |
| 1976 | /* Leave any unwritten inodes on b_io */ |
| 1977 | return wrote; |
| 1978 | } |
| 1979 | |
| 1980 | static long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages, |
| 1981 | enum wb_reason reason) |
| 1982 | { |
| 1983 | struct wb_writeback_work work = { |
| 1984 | .nr_pages = nr_pages, |
| 1985 | .sync_mode = WB_SYNC_NONE, |
| 1986 | .range_cyclic = 1, |
| 1987 | .reason = reason, |
| 1988 | }; |
| 1989 | struct blk_plug plug; |
| 1990 | |
| 1991 | blk_start_plug(&plug); |
| 1992 | spin_lock(&wb->list_lock); |
| 1993 | if (list_empty(&wb->b_io)) |
| 1994 | queue_io(wb, &work, jiffies); |
| 1995 | __writeback_inodes_wb(wb, &work); |
| 1996 | spin_unlock(&wb->list_lock); |
| 1997 | blk_finish_plug(&plug); |
| 1998 | |
| 1999 | return nr_pages - work.nr_pages; |
| 2000 | } |
| 2001 | |
| 2002 | /* |
| 2003 | * Explicit flushing or periodic writeback of "old" data. |
| 2004 | * |
| 2005 | * Define "old": the first time one of an inode's pages is dirtied, we mark the |
| 2006 | * dirtying-time in the inode's address_space. So this periodic writeback code |
| 2007 | * just walks the superblock inode list, writing back any inodes which are |
| 2008 | * older than a specific point in time. |
| 2009 | * |
| 2010 | * Try to run once per dirty_writeback_interval. But if a writeback event |
| 2011 | * takes longer than a dirty_writeback_interval interval, then leave a |
| 2012 | * one-second gap. |
| 2013 | * |
| 2014 | * dirtied_before takes precedence over nr_to_write. So we'll only write back |
| 2015 | * all dirty pages if they are all attached to "old" mappings. |
| 2016 | */ |
| 2017 | static long wb_writeback(struct bdi_writeback *wb, |
| 2018 | struct wb_writeback_work *work) |
| 2019 | { |
| 2020 | long nr_pages = work->nr_pages; |
| 2021 | unsigned long dirtied_before = jiffies; |
| 2022 | struct inode *inode; |
| 2023 | long progress; |
| 2024 | struct blk_plug plug; |
| 2025 | |
| 2026 | blk_start_plug(&plug); |
| 2027 | spin_lock(&wb->list_lock); |
| 2028 | for (;;) { |
| 2029 | /* |
| 2030 | * Stop writeback when nr_pages has been consumed |
| 2031 | */ |
| 2032 | if (work->nr_pages <= 0) |
| 2033 | break; |
| 2034 | |
| 2035 | /* |
| 2036 | * Background writeout and kupdate-style writeback may |
| 2037 | * run forever. Stop them if there is other work to do |
| 2038 | * so that e.g. sync can proceed. They'll be restarted |
| 2039 | * after the other works are all done. |
| 2040 | */ |
| 2041 | if ((work->for_background || work->for_kupdate) && |
| 2042 | !list_empty(&wb->work_list)) |
| 2043 | break; |
| 2044 | |
| 2045 | /* |
| 2046 | * For background writeout, stop when we are below the |
| 2047 | * background dirty threshold |
| 2048 | */ |
| 2049 | if (work->for_background && !wb_over_bg_thresh(wb)) |
| 2050 | break; |
| 2051 | |
| 2052 | /* |
| 2053 | * Kupdate and background works are special and we want to |
| 2054 | * include all inodes that need writing. Livelock avoidance is |
| 2055 | * handled by these works yielding to any other work so we are |
| 2056 | * safe. |
| 2057 | */ |
| 2058 | if (work->for_kupdate) { |
| 2059 | dirtied_before = jiffies - |
| 2060 | msecs_to_jiffies(dirty_expire_interval * 10); |
| 2061 | } else if (work->for_background) |
| 2062 | dirtied_before = jiffies; |
| 2063 | |
| 2064 | trace_writeback_start(wb, work); |
| 2065 | if (list_empty(&wb->b_io)) |
| 2066 | queue_io(wb, work, dirtied_before); |
| 2067 | if (work->sb) |
| 2068 | progress = writeback_sb_inodes(work->sb, wb, work); |
| 2069 | else |
| 2070 | progress = __writeback_inodes_wb(wb, work); |
| 2071 | trace_writeback_written(wb, work); |
| 2072 | |
| 2073 | /* |
| 2074 | * Did we write something? Try for more |
| 2075 | * |
| 2076 | * Dirty inodes are moved to b_io for writeback in batches. |
| 2077 | * The completion of the current batch does not necessarily |
| 2078 | * mean the overall work is done. So we keep looping as long |
| 2079 | * as made some progress on cleaning pages or inodes. |
| 2080 | */ |
| 2081 | if (progress) |
| 2082 | continue; |
| 2083 | /* |
| 2084 | * No more inodes for IO, bail |
| 2085 | */ |
| 2086 | if (list_empty(&wb->b_more_io)) |
| 2087 | break; |
| 2088 | /* |
| 2089 | * Nothing written. Wait for some inode to |
| 2090 | * become available for writeback. Otherwise |
| 2091 | * we'll just busyloop. |
| 2092 | */ |
| 2093 | trace_writeback_wait(wb, work); |
| 2094 | inode = wb_inode(wb->b_more_io.prev); |
| 2095 | spin_lock(&inode->i_lock); |
| 2096 | spin_unlock(&wb->list_lock); |
| 2097 | /* This function drops i_lock... */ |
| 2098 | inode_sleep_on_writeback(inode); |
| 2099 | spin_lock(&wb->list_lock); |
| 2100 | } |
| 2101 | spin_unlock(&wb->list_lock); |
| 2102 | blk_finish_plug(&plug); |
| 2103 | |
| 2104 | return nr_pages - work->nr_pages; |
| 2105 | } |
| 2106 | |
| 2107 | /* |
| 2108 | * Return the next wb_writeback_work struct that hasn't been processed yet. |
| 2109 | */ |
| 2110 | static struct wb_writeback_work *get_next_work_item(struct bdi_writeback *wb) |
| 2111 | { |
| 2112 | struct wb_writeback_work *work = NULL; |
| 2113 | |
| 2114 | spin_lock_irq(&wb->work_lock); |
| 2115 | if (!list_empty(&wb->work_list)) { |
| 2116 | work = list_entry(wb->work_list.next, |
| 2117 | struct wb_writeback_work, list); |
| 2118 | list_del_init(&work->list); |
| 2119 | } |
| 2120 | spin_unlock_irq(&wb->work_lock); |
| 2121 | return work; |
| 2122 | } |
| 2123 | |
| 2124 | static long wb_check_background_flush(struct bdi_writeback *wb) |
| 2125 | { |
| 2126 | if (wb_over_bg_thresh(wb)) { |
| 2127 | |
| 2128 | struct wb_writeback_work work = { |
| 2129 | .nr_pages = LONG_MAX, |
| 2130 | .sync_mode = WB_SYNC_NONE, |
| 2131 | .for_background = 1, |
| 2132 | .range_cyclic = 1, |
| 2133 | .reason = WB_REASON_BACKGROUND, |
| 2134 | }; |
| 2135 | |
| 2136 | return wb_writeback(wb, &work); |
| 2137 | } |
| 2138 | |
| 2139 | return 0; |
| 2140 | } |
| 2141 | |
| 2142 | static long wb_check_old_data_flush(struct bdi_writeback *wb) |
| 2143 | { |
| 2144 | unsigned long expired; |
| 2145 | long nr_pages; |
| 2146 | |
| 2147 | /* |
| 2148 | * When set to zero, disable periodic writeback |
| 2149 | */ |
| 2150 | if (!dirty_writeback_interval) |
| 2151 | return 0; |
| 2152 | |
| 2153 | expired = wb->last_old_flush + |
| 2154 | msecs_to_jiffies(dirty_writeback_interval * 10); |
| 2155 | if (time_before(jiffies, expired)) |
| 2156 | return 0; |
| 2157 | |
| 2158 | wb->last_old_flush = jiffies; |
| 2159 | nr_pages = get_nr_dirty_pages(); |
| 2160 | |
| 2161 | if (nr_pages) { |
| 2162 | struct wb_writeback_work work = { |
| 2163 | .nr_pages = nr_pages, |
| 2164 | .sync_mode = WB_SYNC_NONE, |
| 2165 | .for_kupdate = 1, |
| 2166 | .range_cyclic = 1, |
| 2167 | .reason = WB_REASON_PERIODIC, |
| 2168 | }; |
| 2169 | |
| 2170 | return wb_writeback(wb, &work); |
| 2171 | } |
| 2172 | |
| 2173 | return 0; |
| 2174 | } |
| 2175 | |
| 2176 | static long wb_check_start_all(struct bdi_writeback *wb) |
| 2177 | { |
| 2178 | long nr_pages; |
| 2179 | |
| 2180 | if (!test_bit(WB_start_all, &wb->state)) |
| 2181 | return 0; |
| 2182 | |
| 2183 | nr_pages = get_nr_dirty_pages(); |
| 2184 | if (nr_pages) { |
| 2185 | struct wb_writeback_work work = { |
| 2186 | .nr_pages = wb_split_bdi_pages(wb, nr_pages), |
| 2187 | .sync_mode = WB_SYNC_NONE, |
| 2188 | .range_cyclic = 1, |
| 2189 | .reason = wb->start_all_reason, |
| 2190 | }; |
| 2191 | |
| 2192 | nr_pages = wb_writeback(wb, &work); |
| 2193 | } |
| 2194 | |
| 2195 | clear_bit(WB_start_all, &wb->state); |
| 2196 | return nr_pages; |
| 2197 | } |
| 2198 | |
| 2199 | |
| 2200 | /* |
| 2201 | * Retrieve work items and do the writeback they describe |
| 2202 | */ |
| 2203 | static long wb_do_writeback(struct bdi_writeback *wb) |
| 2204 | { |
| 2205 | struct wb_writeback_work *work; |
| 2206 | long wrote = 0; |
| 2207 | |
| 2208 | set_bit(WB_writeback_running, &wb->state); |
| 2209 | while ((work = get_next_work_item(wb)) != NULL) { |
| 2210 | trace_writeback_exec(wb, work); |
| 2211 | wrote += wb_writeback(wb, work); |
| 2212 | finish_writeback_work(wb, work); |
| 2213 | } |
| 2214 | |
| 2215 | /* |
| 2216 | * Check for a flush-everything request |
| 2217 | */ |
| 2218 | wrote += wb_check_start_all(wb); |
| 2219 | |
| 2220 | /* |
| 2221 | * Check for periodic writeback, kupdated() style |
| 2222 | */ |
| 2223 | wrote += wb_check_old_data_flush(wb); |
| 2224 | wrote += wb_check_background_flush(wb); |
| 2225 | clear_bit(WB_writeback_running, &wb->state); |
| 2226 | |
| 2227 | return wrote; |
| 2228 | } |
| 2229 | |
| 2230 | /* |
| 2231 | * Handle writeback of dirty data for the device backed by this bdi. Also |
| 2232 | * reschedules periodically and does kupdated style flushing. |
| 2233 | */ |
| 2234 | void wb_workfn(struct work_struct *work) |
| 2235 | { |
| 2236 | struct bdi_writeback *wb = container_of(to_delayed_work(work), |
| 2237 | struct bdi_writeback, dwork); |
| 2238 | long pages_written; |
| 2239 | |
| 2240 | set_worker_desc("flush-%s", bdi_dev_name(wb->bdi)); |
| 2241 | |
| 2242 | if (likely(!current_is_workqueue_rescuer() || |
| 2243 | !test_bit(WB_registered, &wb->state))) { |
| 2244 | /* |
| 2245 | * The normal path. Keep writing back @wb until its |
| 2246 | * work_list is empty. Note that this path is also taken |
| 2247 | * if @wb is shutting down even when we're running off the |
| 2248 | * rescuer as work_list needs to be drained. |
| 2249 | */ |
| 2250 | do { |
| 2251 | pages_written = wb_do_writeback(wb); |
| 2252 | trace_writeback_pages_written(pages_written); |
| 2253 | } while (!list_empty(&wb->work_list)); |
| 2254 | } else { |
| 2255 | /* |
| 2256 | * bdi_wq can't get enough workers and we're running off |
| 2257 | * the emergency worker. Don't hog it. Hopefully, 1024 is |
| 2258 | * enough for efficient IO. |
| 2259 | */ |
| 2260 | pages_written = writeback_inodes_wb(wb, 1024, |
| 2261 | WB_REASON_FORKER_THREAD); |
| 2262 | trace_writeback_pages_written(pages_written); |
| 2263 | } |
| 2264 | |
| 2265 | if (!list_empty(&wb->work_list)) |
| 2266 | wb_wakeup(wb); |
| 2267 | else if (wb_has_dirty_io(wb) && dirty_writeback_interval) |
| 2268 | wb_wakeup_delayed(wb); |
| 2269 | } |
| 2270 | |
| 2271 | /* |
| 2272 | * Start writeback of `nr_pages' pages on this bdi. If `nr_pages' is zero, |
| 2273 | * write back the whole world. |
| 2274 | */ |
| 2275 | static void __wakeup_flusher_threads_bdi(struct backing_dev_info *bdi, |
| 2276 | enum wb_reason reason) |
| 2277 | { |
| 2278 | struct bdi_writeback *wb; |
| 2279 | |
| 2280 | if (!bdi_has_dirty_io(bdi)) |
| 2281 | return; |
| 2282 | |
| 2283 | list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node) |
| 2284 | wb_start_writeback(wb, reason); |
| 2285 | } |
| 2286 | |
| 2287 | void wakeup_flusher_threads_bdi(struct backing_dev_info *bdi, |
| 2288 | enum wb_reason reason) |
| 2289 | { |
| 2290 | rcu_read_lock(); |
| 2291 | __wakeup_flusher_threads_bdi(bdi, reason); |
| 2292 | rcu_read_unlock(); |
| 2293 | } |
| 2294 | |
| 2295 | /* |
| 2296 | * Wakeup the flusher threads to start writeback of all currently dirty pages |
| 2297 | */ |
| 2298 | void wakeup_flusher_threads(enum wb_reason reason) |
| 2299 | { |
| 2300 | struct backing_dev_info *bdi; |
| 2301 | |
| 2302 | /* |
| 2303 | * If we are expecting writeback progress we must submit plugged IO. |
| 2304 | */ |
| 2305 | blk_flush_plug(current->plug, true); |
| 2306 | |
| 2307 | rcu_read_lock(); |
| 2308 | list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) |
| 2309 | __wakeup_flusher_threads_bdi(bdi, reason); |
| 2310 | rcu_read_unlock(); |
| 2311 | } |
| 2312 | |
| 2313 | /* |
| 2314 | * Wake up bdi's periodically to make sure dirtytime inodes gets |
| 2315 | * written back periodically. We deliberately do *not* check the |
| 2316 | * b_dirtytime list in wb_has_dirty_io(), since this would cause the |
| 2317 | * kernel to be constantly waking up once there are any dirtytime |
| 2318 | * inodes on the system. So instead we define a separate delayed work |
| 2319 | * function which gets called much more rarely. (By default, only |
| 2320 | * once every 12 hours.) |
| 2321 | * |
| 2322 | * If there is any other write activity going on in the file system, |
| 2323 | * this function won't be necessary. But if the only thing that has |
| 2324 | * happened on the file system is a dirtytime inode caused by an atime |
| 2325 | * update, we need this infrastructure below to make sure that inode |
| 2326 | * eventually gets pushed out to disk. |
| 2327 | */ |
| 2328 | static void wakeup_dirtytime_writeback(struct work_struct *w); |
| 2329 | static DECLARE_DELAYED_WORK(dirtytime_work, wakeup_dirtytime_writeback); |
| 2330 | |
| 2331 | static void wakeup_dirtytime_writeback(struct work_struct *w) |
| 2332 | { |
| 2333 | struct backing_dev_info *bdi; |
| 2334 | |
| 2335 | rcu_read_lock(); |
| 2336 | list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) { |
| 2337 | struct bdi_writeback *wb; |
| 2338 | |
| 2339 | list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node) |
| 2340 | if (!list_empty(&wb->b_dirty_time)) |
| 2341 | wb_wakeup(wb); |
| 2342 | } |
| 2343 | rcu_read_unlock(); |
| 2344 | schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ); |
| 2345 | } |
| 2346 | |
| 2347 | static int __init start_dirtytime_writeback(void) |
| 2348 | { |
| 2349 | schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ); |
| 2350 | return 0; |
| 2351 | } |
| 2352 | __initcall(start_dirtytime_writeback); |
| 2353 | |
| 2354 | int dirtytime_interval_handler(struct ctl_table *table, int write, |
| 2355 | void *buffer, size_t *lenp, loff_t *ppos) |
| 2356 | { |
| 2357 | int ret; |
| 2358 | |
| 2359 | ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
| 2360 | if (ret == 0 && write) |
| 2361 | mod_delayed_work(system_wq, &dirtytime_work, 0); |
| 2362 | return ret; |
| 2363 | } |
| 2364 | |
| 2365 | /** |
| 2366 | * __mark_inode_dirty - internal function to mark an inode dirty |
| 2367 | * |
| 2368 | * @inode: inode to mark |
| 2369 | * @flags: what kind of dirty, e.g. I_DIRTY_SYNC. This can be a combination of |
| 2370 | * multiple I_DIRTY_* flags, except that I_DIRTY_TIME can't be combined |
| 2371 | * with I_DIRTY_PAGES. |
| 2372 | * |
| 2373 | * Mark an inode as dirty. We notify the filesystem, then update the inode's |
| 2374 | * dirty flags. Then, if needed we add the inode to the appropriate dirty list. |
| 2375 | * |
| 2376 | * Most callers should use mark_inode_dirty() or mark_inode_dirty_sync() |
| 2377 | * instead of calling this directly. |
| 2378 | * |
| 2379 | * CAREFUL! We only add the inode to the dirty list if it is hashed or if it |
| 2380 | * refers to a blockdev. Unhashed inodes will never be added to the dirty list |
| 2381 | * even if they are later hashed, as they will have been marked dirty already. |
| 2382 | * |
| 2383 | * In short, ensure you hash any inodes _before_ you start marking them dirty. |
| 2384 | * |
| 2385 | * Note that for blockdevs, inode->dirtied_when represents the dirtying time of |
| 2386 | * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of |
| 2387 | * the kernel-internal blockdev inode represents the dirtying time of the |
| 2388 | * blockdev's pages. This is why for I_DIRTY_PAGES we always use |
| 2389 | * page->mapping->host, so the page-dirtying time is recorded in the internal |
| 2390 | * blockdev inode. |
| 2391 | */ |
| 2392 | void __mark_inode_dirty(struct inode *inode, int flags) |
| 2393 | { |
| 2394 | struct super_block *sb = inode->i_sb; |
| 2395 | int dirtytime = 0; |
| 2396 | struct bdi_writeback *wb = NULL; |
| 2397 | |
| 2398 | trace_writeback_mark_inode_dirty(inode, flags); |
| 2399 | |
| 2400 | if (flags & I_DIRTY_INODE) { |
| 2401 | /* |
| 2402 | * Inode timestamp update will piggback on this dirtying. |
| 2403 | * We tell ->dirty_inode callback that timestamps need to |
| 2404 | * be updated by setting I_DIRTY_TIME in flags. |
| 2405 | */ |
| 2406 | if (inode->i_state & I_DIRTY_TIME) { |
| 2407 | spin_lock(&inode->i_lock); |
| 2408 | if (inode->i_state & I_DIRTY_TIME) { |
| 2409 | inode->i_state &= ~I_DIRTY_TIME; |
| 2410 | flags |= I_DIRTY_TIME; |
| 2411 | } |
| 2412 | spin_unlock(&inode->i_lock); |
| 2413 | } |
| 2414 | |
| 2415 | /* |
| 2416 | * Notify the filesystem about the inode being dirtied, so that |
| 2417 | * (if needed) it can update on-disk fields and journal the |
| 2418 | * inode. This is only needed when the inode itself is being |
| 2419 | * dirtied now. I.e. it's only needed for I_DIRTY_INODE, not |
| 2420 | * for just I_DIRTY_PAGES or I_DIRTY_TIME. |
| 2421 | */ |
| 2422 | trace_writeback_dirty_inode_start(inode, flags); |
| 2423 | if (sb->s_op->dirty_inode) |
| 2424 | sb->s_op->dirty_inode(inode, |
| 2425 | flags & (I_DIRTY_INODE | I_DIRTY_TIME)); |
| 2426 | trace_writeback_dirty_inode(inode, flags); |
| 2427 | |
| 2428 | /* I_DIRTY_INODE supersedes I_DIRTY_TIME. */ |
| 2429 | flags &= ~I_DIRTY_TIME; |
| 2430 | } else { |
| 2431 | /* |
| 2432 | * Else it's either I_DIRTY_PAGES, I_DIRTY_TIME, or nothing. |
| 2433 | * (We don't support setting both I_DIRTY_PAGES and I_DIRTY_TIME |
| 2434 | * in one call to __mark_inode_dirty().) |
| 2435 | */ |
| 2436 | dirtytime = flags & I_DIRTY_TIME; |
| 2437 | WARN_ON_ONCE(dirtytime && flags != I_DIRTY_TIME); |
| 2438 | } |
| 2439 | |
| 2440 | /* |
| 2441 | * Paired with smp_mb() in __writeback_single_inode() for the |
| 2442 | * following lockless i_state test. See there for details. |
| 2443 | */ |
| 2444 | smp_mb(); |
| 2445 | |
| 2446 | if ((inode->i_state & flags) == flags) |
| 2447 | return; |
| 2448 | |
| 2449 | spin_lock(&inode->i_lock); |
| 2450 | if ((inode->i_state & flags) != flags) { |
| 2451 | const int was_dirty = inode->i_state & I_DIRTY; |
| 2452 | |
| 2453 | inode_attach_wb(inode, NULL); |
| 2454 | |
| 2455 | inode->i_state |= flags; |
| 2456 | |
| 2457 | /* |
| 2458 | * Grab inode's wb early because it requires dropping i_lock and we |
| 2459 | * need to make sure following checks happen atomically with dirty |
| 2460 | * list handling so that we don't move inodes under flush worker's |
| 2461 | * hands. |
| 2462 | */ |
| 2463 | if (!was_dirty) { |
| 2464 | wb = locked_inode_to_wb_and_lock_list(inode); |
| 2465 | spin_lock(&inode->i_lock); |
| 2466 | } |
| 2467 | |
| 2468 | /* |
| 2469 | * If the inode is queued for writeback by flush worker, just |
| 2470 | * update its dirty state. Once the flush worker is done with |
| 2471 | * the inode it will place it on the appropriate superblock |
| 2472 | * list, based upon its state. |
| 2473 | */ |
| 2474 | if (inode->i_state & I_SYNC_QUEUED) |
| 2475 | goto out_unlock; |
| 2476 | |
| 2477 | /* |
| 2478 | * Only add valid (hashed) inodes to the superblock's |
| 2479 | * dirty list. Add blockdev inodes as well. |
| 2480 | */ |
| 2481 | if (!S_ISBLK(inode->i_mode)) { |
| 2482 | if (inode_unhashed(inode)) |
| 2483 | goto out_unlock; |
| 2484 | } |
| 2485 | if (inode->i_state & I_FREEING) |
| 2486 | goto out_unlock; |
| 2487 | |
| 2488 | /* |
| 2489 | * If the inode was already on b_dirty/b_io/b_more_io, don't |
| 2490 | * reposition it (that would break b_dirty time-ordering). |
| 2491 | */ |
| 2492 | if (!was_dirty) { |
| 2493 | struct list_head *dirty_list; |
| 2494 | bool wakeup_bdi = false; |
| 2495 | |
| 2496 | inode->dirtied_when = jiffies; |
| 2497 | if (dirtytime) |
| 2498 | inode->dirtied_time_when = jiffies; |
| 2499 | |
| 2500 | if (inode->i_state & I_DIRTY) |
| 2501 | dirty_list = &wb->b_dirty; |
| 2502 | else |
| 2503 | dirty_list = &wb->b_dirty_time; |
| 2504 | |
| 2505 | wakeup_bdi = inode_io_list_move_locked(inode, wb, |
| 2506 | dirty_list); |
| 2507 | |
| 2508 | spin_unlock(&wb->list_lock); |
| 2509 | spin_unlock(&inode->i_lock); |
| 2510 | trace_writeback_dirty_inode_enqueue(inode); |
| 2511 | |
| 2512 | /* |
| 2513 | * If this is the first dirty inode for this bdi, |
| 2514 | * we have to wake-up the corresponding bdi thread |
| 2515 | * to make sure background write-back happens |
| 2516 | * later. |
| 2517 | */ |
| 2518 | if (wakeup_bdi && |
| 2519 | (wb->bdi->capabilities & BDI_CAP_WRITEBACK)) |
| 2520 | wb_wakeup_delayed(wb); |
| 2521 | return; |
| 2522 | } |
| 2523 | } |
| 2524 | out_unlock: |
| 2525 | if (wb) |
| 2526 | spin_unlock(&wb->list_lock); |
| 2527 | spin_unlock(&inode->i_lock); |
| 2528 | } |
| 2529 | EXPORT_SYMBOL(__mark_inode_dirty); |
| 2530 | |
| 2531 | /* |
| 2532 | * The @s_sync_lock is used to serialise concurrent sync operations |
| 2533 | * to avoid lock contention problems with concurrent wait_sb_inodes() calls. |
| 2534 | * Concurrent callers will block on the s_sync_lock rather than doing contending |
| 2535 | * walks. The queueing maintains sync(2) required behaviour as all the IO that |
| 2536 | * has been issued up to the time this function is enter is guaranteed to be |
| 2537 | * completed by the time we have gained the lock and waited for all IO that is |
| 2538 | * in progress regardless of the order callers are granted the lock. |
| 2539 | */ |
| 2540 | static void wait_sb_inodes(struct super_block *sb) |
| 2541 | { |
| 2542 | LIST_HEAD(sync_list); |
| 2543 | |
| 2544 | /* |
| 2545 | * We need to be protected against the filesystem going from |
| 2546 | * r/o to r/w or vice versa. |
| 2547 | */ |
| 2548 | WARN_ON(!rwsem_is_locked(&sb->s_umount)); |
| 2549 | |
| 2550 | mutex_lock(&sb->s_sync_lock); |
| 2551 | |
| 2552 | /* |
| 2553 | * Splice the writeback list onto a temporary list to avoid waiting on |
| 2554 | * inodes that have started writeback after this point. |
| 2555 | * |
| 2556 | * Use rcu_read_lock() to keep the inodes around until we have a |
| 2557 | * reference. s_inode_wblist_lock protects sb->s_inodes_wb as well as |
| 2558 | * the local list because inodes can be dropped from either by writeback |
| 2559 | * completion. |
| 2560 | */ |
| 2561 | rcu_read_lock(); |
| 2562 | spin_lock_irq(&sb->s_inode_wblist_lock); |
| 2563 | list_splice_init(&sb->s_inodes_wb, &sync_list); |
| 2564 | |
| 2565 | /* |
| 2566 | * Data integrity sync. Must wait for all pages under writeback, because |
| 2567 | * there may have been pages dirtied before our sync call, but which had |
| 2568 | * writeout started before we write it out. In which case, the inode |
| 2569 | * may not be on the dirty list, but we still have to wait for that |
| 2570 | * writeout. |
| 2571 | */ |
| 2572 | while (!list_empty(&sync_list)) { |
| 2573 | struct inode *inode = list_first_entry(&sync_list, struct inode, |
| 2574 | i_wb_list); |
| 2575 | struct address_space *mapping = inode->i_mapping; |
| 2576 | |
| 2577 | /* |
| 2578 | * Move each inode back to the wb list before we drop the lock |
| 2579 | * to preserve consistency between i_wb_list and the mapping |
| 2580 | * writeback tag. Writeback completion is responsible to remove |
| 2581 | * the inode from either list once the writeback tag is cleared. |
| 2582 | */ |
| 2583 | list_move_tail(&inode->i_wb_list, &sb->s_inodes_wb); |
| 2584 | |
| 2585 | /* |
| 2586 | * The mapping can appear untagged while still on-list since we |
| 2587 | * do not have the mapping lock. Skip it here, wb completion |
| 2588 | * will remove it. |
| 2589 | */ |
| 2590 | if (!mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK)) |
| 2591 | continue; |
| 2592 | |
| 2593 | spin_unlock_irq(&sb->s_inode_wblist_lock); |
| 2594 | |
| 2595 | spin_lock(&inode->i_lock); |
| 2596 | if (inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) { |
| 2597 | spin_unlock(&inode->i_lock); |
| 2598 | |
| 2599 | spin_lock_irq(&sb->s_inode_wblist_lock); |
| 2600 | continue; |
| 2601 | } |
| 2602 | __iget(inode); |
| 2603 | spin_unlock(&inode->i_lock); |
| 2604 | rcu_read_unlock(); |
| 2605 | |
| 2606 | /* |
| 2607 | * We keep the error status of individual mapping so that |
| 2608 | * applications can catch the writeback error using fsync(2). |
| 2609 | * See filemap_fdatawait_keep_errors() for details. |
| 2610 | */ |
| 2611 | filemap_fdatawait_keep_errors(mapping); |
| 2612 | |
| 2613 | cond_resched(); |
| 2614 | |
| 2615 | iput(inode); |
| 2616 | |
| 2617 | rcu_read_lock(); |
| 2618 | spin_lock_irq(&sb->s_inode_wblist_lock); |
| 2619 | } |
| 2620 | spin_unlock_irq(&sb->s_inode_wblist_lock); |
| 2621 | rcu_read_unlock(); |
| 2622 | mutex_unlock(&sb->s_sync_lock); |
| 2623 | } |
| 2624 | |
| 2625 | static void __writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr, |
| 2626 | enum wb_reason reason, bool skip_if_busy) |
| 2627 | { |
| 2628 | struct backing_dev_info *bdi = sb->s_bdi; |
| 2629 | DEFINE_WB_COMPLETION(done, bdi); |
| 2630 | struct wb_writeback_work work = { |
| 2631 | .sb = sb, |
| 2632 | .sync_mode = WB_SYNC_NONE, |
| 2633 | .tagged_writepages = 1, |
| 2634 | .done = &done, |
| 2635 | .nr_pages = nr, |
| 2636 | .reason = reason, |
| 2637 | }; |
| 2638 | |
| 2639 | if (!bdi_has_dirty_io(bdi) || bdi == &noop_backing_dev_info) |
| 2640 | return; |
| 2641 | WARN_ON(!rwsem_is_locked(&sb->s_umount)); |
| 2642 | |
| 2643 | bdi_split_work_to_wbs(sb->s_bdi, &work, skip_if_busy); |
| 2644 | wb_wait_for_completion(&done); |
| 2645 | } |
| 2646 | |
| 2647 | /** |
| 2648 | * writeback_inodes_sb_nr - writeback dirty inodes from given super_block |
| 2649 | * @sb: the superblock |
| 2650 | * @nr: the number of pages to write |
| 2651 | * @reason: reason why some writeback work initiated |
| 2652 | * |
| 2653 | * Start writeback on some inodes on this super_block. No guarantees are made |
| 2654 | * on how many (if any) will be written, and this function does not wait |
| 2655 | * for IO completion of submitted IO. |
| 2656 | */ |
| 2657 | void writeback_inodes_sb_nr(struct super_block *sb, |
| 2658 | unsigned long nr, |
| 2659 | enum wb_reason reason) |
| 2660 | { |
| 2661 | __writeback_inodes_sb_nr(sb, nr, reason, false); |
| 2662 | } |
| 2663 | EXPORT_SYMBOL(writeback_inodes_sb_nr); |
| 2664 | |
| 2665 | /** |
| 2666 | * writeback_inodes_sb - writeback dirty inodes from given super_block |
| 2667 | * @sb: the superblock |
| 2668 | * @reason: reason why some writeback work was initiated |
| 2669 | * |
| 2670 | * Start writeback on some inodes on this super_block. No guarantees are made |
| 2671 | * on how many (if any) will be written, and this function does not wait |
| 2672 | * for IO completion of submitted IO. |
| 2673 | */ |
| 2674 | void writeback_inodes_sb(struct super_block *sb, enum wb_reason reason) |
| 2675 | { |
| 2676 | return writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason); |
| 2677 | } |
| 2678 | EXPORT_SYMBOL(writeback_inodes_sb); |
| 2679 | |
| 2680 | /** |
| 2681 | * try_to_writeback_inodes_sb - try to start writeback if none underway |
| 2682 | * @sb: the superblock |
| 2683 | * @reason: reason why some writeback work was initiated |
| 2684 | * |
| 2685 | * Invoke __writeback_inodes_sb_nr if no writeback is currently underway. |
| 2686 | */ |
| 2687 | void try_to_writeback_inodes_sb(struct super_block *sb, enum wb_reason reason) |
| 2688 | { |
| 2689 | if (!down_read_trylock(&sb->s_umount)) |
| 2690 | return; |
| 2691 | |
| 2692 | __writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason, true); |
| 2693 | up_read(&sb->s_umount); |
| 2694 | } |
| 2695 | EXPORT_SYMBOL(try_to_writeback_inodes_sb); |
| 2696 | |
| 2697 | /** |
| 2698 | * sync_inodes_sb - sync sb inode pages |
| 2699 | * @sb: the superblock |
| 2700 | * |
| 2701 | * This function writes and waits on any dirty inode belonging to this |
| 2702 | * super_block. |
| 2703 | */ |
| 2704 | void sync_inodes_sb(struct super_block *sb) |
| 2705 | { |
| 2706 | struct backing_dev_info *bdi = sb->s_bdi; |
| 2707 | DEFINE_WB_COMPLETION(done, bdi); |
| 2708 | struct wb_writeback_work work = { |
| 2709 | .sb = sb, |
| 2710 | .sync_mode = WB_SYNC_ALL, |
| 2711 | .nr_pages = LONG_MAX, |
| 2712 | .range_cyclic = 0, |
| 2713 | .done = &done, |
| 2714 | .reason = WB_REASON_SYNC, |
| 2715 | .for_sync = 1, |
| 2716 | }; |
| 2717 | |
| 2718 | /* |
| 2719 | * Can't skip on !bdi_has_dirty() because we should wait for !dirty |
| 2720 | * inodes under writeback and I_DIRTY_TIME inodes ignored by |
| 2721 | * bdi_has_dirty() need to be written out too. |
| 2722 | */ |
| 2723 | if (bdi == &noop_backing_dev_info) |
| 2724 | return; |
| 2725 | WARN_ON(!rwsem_is_locked(&sb->s_umount)); |
| 2726 | |
| 2727 | /* protect against inode wb switch, see inode_switch_wbs_work_fn() */ |
| 2728 | bdi_down_write_wb_switch_rwsem(bdi); |
| 2729 | bdi_split_work_to_wbs(bdi, &work, false); |
| 2730 | wb_wait_for_completion(&done); |
| 2731 | bdi_up_write_wb_switch_rwsem(bdi); |
| 2732 | |
| 2733 | wait_sb_inodes(sb); |
| 2734 | } |
| 2735 | EXPORT_SYMBOL(sync_inodes_sb); |
| 2736 | |
| 2737 | /** |
| 2738 | * write_inode_now - write an inode to disk |
| 2739 | * @inode: inode to write to disk |
| 2740 | * @sync: whether the write should be synchronous or not |
| 2741 | * |
| 2742 | * This function commits an inode to disk immediately if it is dirty. This is |
| 2743 | * primarily needed by knfsd. |
| 2744 | * |
| 2745 | * The caller must either have a ref on the inode or must have set I_WILL_FREE. |
| 2746 | */ |
| 2747 | int write_inode_now(struct inode *inode, int sync) |
| 2748 | { |
| 2749 | struct writeback_control wbc = { |
| 2750 | .nr_to_write = LONG_MAX, |
| 2751 | .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE, |
| 2752 | .range_start = 0, |
| 2753 | .range_end = LLONG_MAX, |
| 2754 | }; |
| 2755 | |
| 2756 | if (!mapping_can_writeback(inode->i_mapping)) |
| 2757 | wbc.nr_to_write = 0; |
| 2758 | |
| 2759 | might_sleep(); |
| 2760 | return writeback_single_inode(inode, &wbc); |
| 2761 | } |
| 2762 | EXPORT_SYMBOL(write_inode_now); |
| 2763 | |
| 2764 | /** |
| 2765 | * sync_inode_metadata - write an inode to disk |
| 2766 | * @inode: the inode to sync |
| 2767 | * @wait: wait for I/O to complete. |
| 2768 | * |
| 2769 | * Write an inode to disk and adjust its dirty state after completion. |
| 2770 | * |
| 2771 | * Note: only writes the actual inode, no associated data or other metadata. |
| 2772 | */ |
| 2773 | int sync_inode_metadata(struct inode *inode, int wait) |
| 2774 | { |
| 2775 | struct writeback_control wbc = { |
| 2776 | .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE, |
| 2777 | .nr_to_write = 0, /* metadata-only */ |
| 2778 | }; |
| 2779 | |
| 2780 | return writeback_single_inode(inode, &wbc); |
| 2781 | } |
| 2782 | EXPORT_SYMBOL(sync_inode_metadata); |