| 1 | // SPDX-License-Identifier: GPL-2.0-only |
| 2 | |
| 3 | #include <linux/wait.h> |
| 4 | #include <linux/rbtree.h> |
| 5 | #include <linux/backing-dev.h> |
| 6 | #include <linux/kthread.h> |
| 7 | #include <linux/freezer.h> |
| 8 | #include <linux/fs.h> |
| 9 | #include <linux/pagemap.h> |
| 10 | #include <linux/mm.h> |
| 11 | #include <linux/sched/mm.h> |
| 12 | #include <linux/sched.h> |
| 13 | #include <linux/module.h> |
| 14 | #include <linux/writeback.h> |
| 15 | #include <linux/device.h> |
| 16 | #include <trace/events/writeback.h> |
| 17 | |
| 18 | struct backing_dev_info noop_backing_dev_info; |
| 19 | EXPORT_SYMBOL_GPL(noop_backing_dev_info); |
| 20 | |
| 21 | static struct class *bdi_class; |
| 22 | static const char *bdi_unknown_name = "(unknown)"; |
| 23 | |
| 24 | /* |
| 25 | * bdi_lock protects bdi_tree and updates to bdi_list. bdi_list has RCU |
| 26 | * reader side locking. |
| 27 | */ |
| 28 | DEFINE_SPINLOCK(bdi_lock); |
| 29 | static u64 bdi_id_cursor; |
| 30 | static struct rb_root bdi_tree = RB_ROOT; |
| 31 | LIST_HEAD(bdi_list); |
| 32 | |
| 33 | /* bdi_wq serves all asynchronous writeback tasks */ |
| 34 | struct workqueue_struct *bdi_wq; |
| 35 | |
| 36 | #define K(x) ((x) << (PAGE_SHIFT - 10)) |
| 37 | |
| 38 | #ifdef CONFIG_DEBUG_FS |
| 39 | #include <linux/debugfs.h> |
| 40 | #include <linux/seq_file.h> |
| 41 | |
| 42 | static struct dentry *bdi_debug_root; |
| 43 | |
| 44 | static void bdi_debug_init(void) |
| 45 | { |
| 46 | bdi_debug_root = debugfs_create_dir("bdi", NULL); |
| 47 | } |
| 48 | |
| 49 | static int bdi_debug_stats_show(struct seq_file *m, void *v) |
| 50 | { |
| 51 | struct backing_dev_info *bdi = m->private; |
| 52 | struct bdi_writeback *wb = &bdi->wb; |
| 53 | unsigned long background_thresh; |
| 54 | unsigned long dirty_thresh; |
| 55 | unsigned long wb_thresh; |
| 56 | unsigned long nr_dirty, nr_io, nr_more_io, nr_dirty_time; |
| 57 | struct inode *inode; |
| 58 | |
| 59 | nr_dirty = nr_io = nr_more_io = nr_dirty_time = 0; |
| 60 | spin_lock(&wb->list_lock); |
| 61 | list_for_each_entry(inode, &wb->b_dirty, i_io_list) |
| 62 | nr_dirty++; |
| 63 | list_for_each_entry(inode, &wb->b_io, i_io_list) |
| 64 | nr_io++; |
| 65 | list_for_each_entry(inode, &wb->b_more_io, i_io_list) |
| 66 | nr_more_io++; |
| 67 | list_for_each_entry(inode, &wb->b_dirty_time, i_io_list) |
| 68 | if (inode->i_state & I_DIRTY_TIME) |
| 69 | nr_dirty_time++; |
| 70 | spin_unlock(&wb->list_lock); |
| 71 | |
| 72 | global_dirty_limits(&background_thresh, &dirty_thresh); |
| 73 | wb_thresh = wb_calc_thresh(wb, dirty_thresh); |
| 74 | |
| 75 | seq_printf(m, |
| 76 | "BdiWriteback: %10lu kB\n" |
| 77 | "BdiReclaimable: %10lu kB\n" |
| 78 | "BdiDirtyThresh: %10lu kB\n" |
| 79 | "DirtyThresh: %10lu kB\n" |
| 80 | "BackgroundThresh: %10lu kB\n" |
| 81 | "BdiDirtied: %10lu kB\n" |
| 82 | "BdiWritten: %10lu kB\n" |
| 83 | "BdiWriteBandwidth: %10lu kBps\n" |
| 84 | "b_dirty: %10lu\n" |
| 85 | "b_io: %10lu\n" |
| 86 | "b_more_io: %10lu\n" |
| 87 | "b_dirty_time: %10lu\n" |
| 88 | "bdi_list: %10u\n" |
| 89 | "state: %10lx\n", |
| 90 | (unsigned long) K(wb_stat(wb, WB_WRITEBACK)), |
| 91 | (unsigned long) K(wb_stat(wb, WB_RECLAIMABLE)), |
| 92 | K(wb_thresh), |
| 93 | K(dirty_thresh), |
| 94 | K(background_thresh), |
| 95 | (unsigned long) K(wb_stat(wb, WB_DIRTIED)), |
| 96 | (unsigned long) K(wb_stat(wb, WB_WRITTEN)), |
| 97 | (unsigned long) K(wb->write_bandwidth), |
| 98 | nr_dirty, |
| 99 | nr_io, |
| 100 | nr_more_io, |
| 101 | nr_dirty_time, |
| 102 | !list_empty(&bdi->bdi_list), bdi->wb.state); |
| 103 | |
| 104 | return 0; |
| 105 | } |
| 106 | DEFINE_SHOW_ATTRIBUTE(bdi_debug_stats); |
| 107 | |
| 108 | static void bdi_debug_register(struct backing_dev_info *bdi, const char *name) |
| 109 | { |
| 110 | bdi->debug_dir = debugfs_create_dir(name, bdi_debug_root); |
| 111 | |
| 112 | debugfs_create_file("stats", 0444, bdi->debug_dir, bdi, |
| 113 | &bdi_debug_stats_fops); |
| 114 | } |
| 115 | |
| 116 | static void bdi_debug_unregister(struct backing_dev_info *bdi) |
| 117 | { |
| 118 | debugfs_remove_recursive(bdi->debug_dir); |
| 119 | } |
| 120 | #else |
| 121 | static inline void bdi_debug_init(void) |
| 122 | { |
| 123 | } |
| 124 | static inline void bdi_debug_register(struct backing_dev_info *bdi, |
| 125 | const char *name) |
| 126 | { |
| 127 | } |
| 128 | static inline void bdi_debug_unregister(struct backing_dev_info *bdi) |
| 129 | { |
| 130 | } |
| 131 | #endif |
| 132 | |
| 133 | static ssize_t read_ahead_kb_store(struct device *dev, |
| 134 | struct device_attribute *attr, |
| 135 | const char *buf, size_t count) |
| 136 | { |
| 137 | struct backing_dev_info *bdi = dev_get_drvdata(dev); |
| 138 | unsigned long read_ahead_kb; |
| 139 | ssize_t ret; |
| 140 | |
| 141 | ret = kstrtoul(buf, 10, &read_ahead_kb); |
| 142 | if (ret < 0) |
| 143 | return ret; |
| 144 | |
| 145 | bdi->ra_pages = read_ahead_kb >> (PAGE_SHIFT - 10); |
| 146 | |
| 147 | return count; |
| 148 | } |
| 149 | |
| 150 | #define BDI_SHOW(name, expr) \ |
| 151 | static ssize_t name##_show(struct device *dev, \ |
| 152 | struct device_attribute *attr, char *buf) \ |
| 153 | { \ |
| 154 | struct backing_dev_info *bdi = dev_get_drvdata(dev); \ |
| 155 | \ |
| 156 | return sysfs_emit(buf, "%lld\n", (long long)expr); \ |
| 157 | } \ |
| 158 | static DEVICE_ATTR_RW(name); |
| 159 | |
| 160 | BDI_SHOW(read_ahead_kb, K(bdi->ra_pages)) |
| 161 | |
| 162 | static ssize_t min_ratio_store(struct device *dev, |
| 163 | struct device_attribute *attr, const char *buf, size_t count) |
| 164 | { |
| 165 | struct backing_dev_info *bdi = dev_get_drvdata(dev); |
| 166 | unsigned int ratio; |
| 167 | ssize_t ret; |
| 168 | |
| 169 | ret = kstrtouint(buf, 10, &ratio); |
| 170 | if (ret < 0) |
| 171 | return ret; |
| 172 | |
| 173 | ret = bdi_set_min_ratio(bdi, ratio); |
| 174 | if (!ret) |
| 175 | ret = count; |
| 176 | |
| 177 | return ret; |
| 178 | } |
| 179 | BDI_SHOW(min_ratio, bdi->min_ratio) |
| 180 | |
| 181 | static ssize_t max_ratio_store(struct device *dev, |
| 182 | struct device_attribute *attr, const char *buf, size_t count) |
| 183 | { |
| 184 | struct backing_dev_info *bdi = dev_get_drvdata(dev); |
| 185 | unsigned int ratio; |
| 186 | ssize_t ret; |
| 187 | |
| 188 | ret = kstrtouint(buf, 10, &ratio); |
| 189 | if (ret < 0) |
| 190 | return ret; |
| 191 | |
| 192 | ret = bdi_set_max_ratio(bdi, ratio); |
| 193 | if (!ret) |
| 194 | ret = count; |
| 195 | |
| 196 | return ret; |
| 197 | } |
| 198 | BDI_SHOW(max_ratio, bdi->max_ratio) |
| 199 | |
| 200 | static ssize_t stable_pages_required_show(struct device *dev, |
| 201 | struct device_attribute *attr, |
| 202 | char *buf) |
| 203 | { |
| 204 | dev_warn_once(dev, |
| 205 | "the stable_pages_required attribute has been removed. Use the stable_writes queue attribute instead.\n"); |
| 206 | return sysfs_emit(buf, "%d\n", 0); |
| 207 | } |
| 208 | static DEVICE_ATTR_RO(stable_pages_required); |
| 209 | |
| 210 | static struct attribute *bdi_dev_attrs[] = { |
| 211 | &dev_attr_read_ahead_kb.attr, |
| 212 | &dev_attr_min_ratio.attr, |
| 213 | &dev_attr_max_ratio.attr, |
| 214 | &dev_attr_stable_pages_required.attr, |
| 215 | NULL, |
| 216 | }; |
| 217 | ATTRIBUTE_GROUPS(bdi_dev); |
| 218 | |
| 219 | static __init int bdi_class_init(void) |
| 220 | { |
| 221 | bdi_class = class_create(THIS_MODULE, "bdi"); |
| 222 | if (IS_ERR(bdi_class)) |
| 223 | return PTR_ERR(bdi_class); |
| 224 | |
| 225 | bdi_class->dev_groups = bdi_dev_groups; |
| 226 | bdi_debug_init(); |
| 227 | |
| 228 | return 0; |
| 229 | } |
| 230 | postcore_initcall(bdi_class_init); |
| 231 | |
| 232 | static int bdi_init(struct backing_dev_info *bdi); |
| 233 | |
| 234 | static int __init default_bdi_init(void) |
| 235 | { |
| 236 | int err; |
| 237 | |
| 238 | bdi_wq = alloc_workqueue("writeback", WQ_MEM_RECLAIM | WQ_UNBOUND | |
| 239 | WQ_SYSFS, 0); |
| 240 | if (!bdi_wq) |
| 241 | return -ENOMEM; |
| 242 | |
| 243 | err = bdi_init(&noop_backing_dev_info); |
| 244 | |
| 245 | return err; |
| 246 | } |
| 247 | subsys_initcall(default_bdi_init); |
| 248 | |
| 249 | /* |
| 250 | * This function is used when the first inode for this wb is marked dirty. It |
| 251 | * wakes-up the corresponding bdi thread which should then take care of the |
| 252 | * periodic background write-out of dirty inodes. Since the write-out would |
| 253 | * starts only 'dirty_writeback_interval' centisecs from now anyway, we just |
| 254 | * set up a timer which wakes the bdi thread up later. |
| 255 | * |
| 256 | * Note, we wouldn't bother setting up the timer, but this function is on the |
| 257 | * fast-path (used by '__mark_inode_dirty()'), so we save few context switches |
| 258 | * by delaying the wake-up. |
| 259 | * |
| 260 | * We have to be careful not to postpone flush work if it is scheduled for |
| 261 | * earlier. Thus we use queue_delayed_work(). |
| 262 | */ |
| 263 | void wb_wakeup_delayed(struct bdi_writeback *wb) |
| 264 | { |
| 265 | unsigned long timeout; |
| 266 | |
| 267 | timeout = msecs_to_jiffies(dirty_writeback_interval * 10); |
| 268 | spin_lock_bh(&wb->work_lock); |
| 269 | if (test_bit(WB_registered, &wb->state)) |
| 270 | queue_delayed_work(bdi_wq, &wb->dwork, timeout); |
| 271 | spin_unlock_bh(&wb->work_lock); |
| 272 | } |
| 273 | |
| 274 | /* |
| 275 | * Initial write bandwidth: 100 MB/s |
| 276 | */ |
| 277 | #define INIT_BW (100 << (20 - PAGE_SHIFT)) |
| 278 | |
| 279 | static int wb_init(struct bdi_writeback *wb, struct backing_dev_info *bdi, |
| 280 | gfp_t gfp) |
| 281 | { |
| 282 | int i, err; |
| 283 | |
| 284 | memset(wb, 0, sizeof(*wb)); |
| 285 | |
| 286 | if (wb != &bdi->wb) |
| 287 | bdi_get(bdi); |
| 288 | wb->bdi = bdi; |
| 289 | wb->last_old_flush = jiffies; |
| 290 | INIT_LIST_HEAD(&wb->b_dirty); |
| 291 | INIT_LIST_HEAD(&wb->b_io); |
| 292 | INIT_LIST_HEAD(&wb->b_more_io); |
| 293 | INIT_LIST_HEAD(&wb->b_dirty_time); |
| 294 | spin_lock_init(&wb->list_lock); |
| 295 | |
| 296 | wb->bw_time_stamp = jiffies; |
| 297 | wb->balanced_dirty_ratelimit = INIT_BW; |
| 298 | wb->dirty_ratelimit = INIT_BW; |
| 299 | wb->write_bandwidth = INIT_BW; |
| 300 | wb->avg_write_bandwidth = INIT_BW; |
| 301 | |
| 302 | spin_lock_init(&wb->work_lock); |
| 303 | INIT_LIST_HEAD(&wb->work_list); |
| 304 | INIT_DELAYED_WORK(&wb->dwork, wb_workfn); |
| 305 | wb->dirty_sleep = jiffies; |
| 306 | |
| 307 | err = fprop_local_init_percpu(&wb->completions, gfp); |
| 308 | if (err) |
| 309 | goto out_put_bdi; |
| 310 | |
| 311 | for (i = 0; i < NR_WB_STAT_ITEMS; i++) { |
| 312 | err = percpu_counter_init(&wb->stat[i], 0, gfp); |
| 313 | if (err) |
| 314 | goto out_destroy_stat; |
| 315 | } |
| 316 | |
| 317 | return 0; |
| 318 | |
| 319 | out_destroy_stat: |
| 320 | while (i--) |
| 321 | percpu_counter_destroy(&wb->stat[i]); |
| 322 | fprop_local_destroy_percpu(&wb->completions); |
| 323 | out_put_bdi: |
| 324 | if (wb != &bdi->wb) |
| 325 | bdi_put(bdi); |
| 326 | return err; |
| 327 | } |
| 328 | |
| 329 | static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb); |
| 330 | |
| 331 | /* |
| 332 | * Remove bdi from the global list and shutdown any threads we have running |
| 333 | */ |
| 334 | static void wb_shutdown(struct bdi_writeback *wb) |
| 335 | { |
| 336 | /* Make sure nobody queues further work */ |
| 337 | spin_lock_bh(&wb->work_lock); |
| 338 | if (!test_and_clear_bit(WB_registered, &wb->state)) { |
| 339 | spin_unlock_bh(&wb->work_lock); |
| 340 | return; |
| 341 | } |
| 342 | spin_unlock_bh(&wb->work_lock); |
| 343 | |
| 344 | cgwb_remove_from_bdi_list(wb); |
| 345 | /* |
| 346 | * Drain work list and shutdown the delayed_work. !WB_registered |
| 347 | * tells wb_workfn() that @wb is dying and its work_list needs to |
| 348 | * be drained no matter what. |
| 349 | */ |
| 350 | mod_delayed_work(bdi_wq, &wb->dwork, 0); |
| 351 | flush_delayed_work(&wb->dwork); |
| 352 | WARN_ON(!list_empty(&wb->work_list)); |
| 353 | } |
| 354 | |
| 355 | static void wb_exit(struct bdi_writeback *wb) |
| 356 | { |
| 357 | int i; |
| 358 | |
| 359 | WARN_ON(delayed_work_pending(&wb->dwork)); |
| 360 | |
| 361 | for (i = 0; i < NR_WB_STAT_ITEMS; i++) |
| 362 | percpu_counter_destroy(&wb->stat[i]); |
| 363 | |
| 364 | fprop_local_destroy_percpu(&wb->completions); |
| 365 | if (wb != &wb->bdi->wb) |
| 366 | bdi_put(wb->bdi); |
| 367 | } |
| 368 | |
| 369 | #ifdef CONFIG_CGROUP_WRITEBACK |
| 370 | |
| 371 | #include <linux/memcontrol.h> |
| 372 | |
| 373 | /* |
| 374 | * cgwb_lock protects bdi->cgwb_tree, blkcg->cgwb_list, and memcg->cgwb_list. |
| 375 | * bdi->cgwb_tree is also RCU protected. |
| 376 | */ |
| 377 | static DEFINE_SPINLOCK(cgwb_lock); |
| 378 | static struct workqueue_struct *cgwb_release_wq; |
| 379 | |
| 380 | static void cgwb_release_workfn(struct work_struct *work) |
| 381 | { |
| 382 | struct bdi_writeback *wb = container_of(work, struct bdi_writeback, |
| 383 | release_work); |
| 384 | struct blkcg *blkcg = css_to_blkcg(wb->blkcg_css); |
| 385 | |
| 386 | mutex_lock(&wb->bdi->cgwb_release_mutex); |
| 387 | wb_shutdown(wb); |
| 388 | |
| 389 | css_put(wb->memcg_css); |
| 390 | css_put(wb->blkcg_css); |
| 391 | mutex_unlock(&wb->bdi->cgwb_release_mutex); |
| 392 | |
| 393 | /* triggers blkg destruction if no online users left */ |
| 394 | blkcg_unpin_online(blkcg); |
| 395 | |
| 396 | fprop_local_destroy_percpu(&wb->memcg_completions); |
| 397 | percpu_ref_exit(&wb->refcnt); |
| 398 | wb_exit(wb); |
| 399 | kfree_rcu(wb, rcu); |
| 400 | } |
| 401 | |
| 402 | static void cgwb_release(struct percpu_ref *refcnt) |
| 403 | { |
| 404 | struct bdi_writeback *wb = container_of(refcnt, struct bdi_writeback, |
| 405 | refcnt); |
| 406 | queue_work(cgwb_release_wq, &wb->release_work); |
| 407 | } |
| 408 | |
| 409 | static void cgwb_kill(struct bdi_writeback *wb) |
| 410 | { |
| 411 | lockdep_assert_held(&cgwb_lock); |
| 412 | |
| 413 | WARN_ON(!radix_tree_delete(&wb->bdi->cgwb_tree, wb->memcg_css->id)); |
| 414 | list_del(&wb->memcg_node); |
| 415 | list_del(&wb->blkcg_node); |
| 416 | percpu_ref_kill(&wb->refcnt); |
| 417 | } |
| 418 | |
| 419 | static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb) |
| 420 | { |
| 421 | spin_lock_irq(&cgwb_lock); |
| 422 | list_del_rcu(&wb->bdi_node); |
| 423 | spin_unlock_irq(&cgwb_lock); |
| 424 | } |
| 425 | |
| 426 | static int cgwb_create(struct backing_dev_info *bdi, |
| 427 | struct cgroup_subsys_state *memcg_css, gfp_t gfp) |
| 428 | { |
| 429 | struct mem_cgroup *memcg; |
| 430 | struct cgroup_subsys_state *blkcg_css; |
| 431 | struct blkcg *blkcg; |
| 432 | struct list_head *memcg_cgwb_list, *blkcg_cgwb_list; |
| 433 | struct bdi_writeback *wb; |
| 434 | unsigned long flags; |
| 435 | int ret = 0; |
| 436 | |
| 437 | memcg = mem_cgroup_from_css(memcg_css); |
| 438 | blkcg_css = cgroup_get_e_css(memcg_css->cgroup, &io_cgrp_subsys); |
| 439 | blkcg = css_to_blkcg(blkcg_css); |
| 440 | memcg_cgwb_list = &memcg->cgwb_list; |
| 441 | blkcg_cgwb_list = &blkcg->cgwb_list; |
| 442 | |
| 443 | /* look up again under lock and discard on blkcg mismatch */ |
| 444 | spin_lock_irqsave(&cgwb_lock, flags); |
| 445 | wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id); |
| 446 | if (wb && wb->blkcg_css != blkcg_css) { |
| 447 | cgwb_kill(wb); |
| 448 | wb = NULL; |
| 449 | } |
| 450 | spin_unlock_irqrestore(&cgwb_lock, flags); |
| 451 | if (wb) |
| 452 | goto out_put; |
| 453 | |
| 454 | /* need to create a new one */ |
| 455 | wb = kmalloc(sizeof(*wb), gfp); |
| 456 | if (!wb) { |
| 457 | ret = -ENOMEM; |
| 458 | goto out_put; |
| 459 | } |
| 460 | |
| 461 | ret = wb_init(wb, bdi, gfp); |
| 462 | if (ret) |
| 463 | goto err_free; |
| 464 | |
| 465 | ret = percpu_ref_init(&wb->refcnt, cgwb_release, 0, gfp); |
| 466 | if (ret) |
| 467 | goto err_wb_exit; |
| 468 | |
| 469 | ret = fprop_local_init_percpu(&wb->memcg_completions, gfp); |
| 470 | if (ret) |
| 471 | goto err_ref_exit; |
| 472 | |
| 473 | wb->memcg_css = memcg_css; |
| 474 | wb->blkcg_css = blkcg_css; |
| 475 | INIT_WORK(&wb->release_work, cgwb_release_workfn); |
| 476 | set_bit(WB_registered, &wb->state); |
| 477 | |
| 478 | /* |
| 479 | * The root wb determines the registered state of the whole bdi and |
| 480 | * memcg_cgwb_list and blkcg_cgwb_list's next pointers indicate |
| 481 | * whether they're still online. Don't link @wb if any is dead. |
| 482 | * See wb_memcg_offline() and wb_blkcg_offline(). |
| 483 | */ |
| 484 | ret = -ENODEV; |
| 485 | spin_lock_irqsave(&cgwb_lock, flags); |
| 486 | if (test_bit(WB_registered, &bdi->wb.state) && |
| 487 | blkcg_cgwb_list->next && memcg_cgwb_list->next) { |
| 488 | /* we might have raced another instance of this function */ |
| 489 | ret = radix_tree_insert(&bdi->cgwb_tree, memcg_css->id, wb); |
| 490 | if (!ret) { |
| 491 | list_add_tail_rcu(&wb->bdi_node, &bdi->wb_list); |
| 492 | list_add(&wb->memcg_node, memcg_cgwb_list); |
| 493 | list_add(&wb->blkcg_node, blkcg_cgwb_list); |
| 494 | blkcg_pin_online(blkcg); |
| 495 | css_get(memcg_css); |
| 496 | css_get(blkcg_css); |
| 497 | } |
| 498 | } |
| 499 | spin_unlock_irqrestore(&cgwb_lock, flags); |
| 500 | if (ret) { |
| 501 | if (ret == -EEXIST) |
| 502 | ret = 0; |
| 503 | goto err_fprop_exit; |
| 504 | } |
| 505 | goto out_put; |
| 506 | |
| 507 | err_fprop_exit: |
| 508 | fprop_local_destroy_percpu(&wb->memcg_completions); |
| 509 | err_ref_exit: |
| 510 | percpu_ref_exit(&wb->refcnt); |
| 511 | err_wb_exit: |
| 512 | wb_exit(wb); |
| 513 | err_free: |
| 514 | kfree(wb); |
| 515 | out_put: |
| 516 | css_put(blkcg_css); |
| 517 | return ret; |
| 518 | } |
| 519 | |
| 520 | /** |
| 521 | * wb_get_lookup - get wb for a given memcg |
| 522 | * @bdi: target bdi |
| 523 | * @memcg_css: cgroup_subsys_state of the target memcg (must have positive ref) |
| 524 | * |
| 525 | * Try to get the wb for @memcg_css on @bdi. The returned wb has its |
| 526 | * refcount incremented. |
| 527 | * |
| 528 | * This function uses css_get() on @memcg_css and thus expects its refcnt |
| 529 | * to be positive on invocation. IOW, rcu_read_lock() protection on |
| 530 | * @memcg_css isn't enough. try_get it before calling this function. |
| 531 | * |
| 532 | * A wb is keyed by its associated memcg. As blkcg implicitly enables |
| 533 | * memcg on the default hierarchy, memcg association is guaranteed to be |
| 534 | * more specific (equal or descendant to the associated blkcg) and thus can |
| 535 | * identify both the memcg and blkcg associations. |
| 536 | * |
| 537 | * Because the blkcg associated with a memcg may change as blkcg is enabled |
| 538 | * and disabled closer to root in the hierarchy, each wb keeps track of |
| 539 | * both the memcg and blkcg associated with it and verifies the blkcg on |
| 540 | * each lookup. On mismatch, the existing wb is discarded and a new one is |
| 541 | * created. |
| 542 | */ |
| 543 | struct bdi_writeback *wb_get_lookup(struct backing_dev_info *bdi, |
| 544 | struct cgroup_subsys_state *memcg_css) |
| 545 | { |
| 546 | struct bdi_writeback *wb; |
| 547 | |
| 548 | if (!memcg_css->parent) |
| 549 | return &bdi->wb; |
| 550 | |
| 551 | rcu_read_lock(); |
| 552 | wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id); |
| 553 | if (wb) { |
| 554 | struct cgroup_subsys_state *blkcg_css; |
| 555 | |
| 556 | /* see whether the blkcg association has changed */ |
| 557 | blkcg_css = cgroup_get_e_css(memcg_css->cgroup, &io_cgrp_subsys); |
| 558 | if (unlikely(wb->blkcg_css != blkcg_css || !wb_tryget(wb))) |
| 559 | wb = NULL; |
| 560 | css_put(blkcg_css); |
| 561 | } |
| 562 | rcu_read_unlock(); |
| 563 | |
| 564 | return wb; |
| 565 | } |
| 566 | |
| 567 | /** |
| 568 | * wb_get_create - get wb for a given memcg, create if necessary |
| 569 | * @bdi: target bdi |
| 570 | * @memcg_css: cgroup_subsys_state of the target memcg (must have positive ref) |
| 571 | * @gfp: allocation mask to use |
| 572 | * |
| 573 | * Try to get the wb for @memcg_css on @bdi. If it doesn't exist, try to |
| 574 | * create one. See wb_get_lookup() for more details. |
| 575 | */ |
| 576 | struct bdi_writeback *wb_get_create(struct backing_dev_info *bdi, |
| 577 | struct cgroup_subsys_state *memcg_css, |
| 578 | gfp_t gfp) |
| 579 | { |
| 580 | struct bdi_writeback *wb; |
| 581 | |
| 582 | might_alloc(gfp); |
| 583 | |
| 584 | if (!memcg_css->parent) |
| 585 | return &bdi->wb; |
| 586 | |
| 587 | do { |
| 588 | wb = wb_get_lookup(bdi, memcg_css); |
| 589 | } while (!wb && !cgwb_create(bdi, memcg_css, gfp)); |
| 590 | |
| 591 | return wb; |
| 592 | } |
| 593 | |
| 594 | static int cgwb_bdi_init(struct backing_dev_info *bdi) |
| 595 | { |
| 596 | int ret; |
| 597 | |
| 598 | INIT_RADIX_TREE(&bdi->cgwb_tree, GFP_ATOMIC); |
| 599 | mutex_init(&bdi->cgwb_release_mutex); |
| 600 | init_rwsem(&bdi->wb_switch_rwsem); |
| 601 | |
| 602 | ret = wb_init(&bdi->wb, bdi, GFP_KERNEL); |
| 603 | if (!ret) { |
| 604 | bdi->wb.memcg_css = &root_mem_cgroup->css; |
| 605 | bdi->wb.blkcg_css = blkcg_root_css; |
| 606 | } |
| 607 | return ret; |
| 608 | } |
| 609 | |
| 610 | static void cgwb_bdi_unregister(struct backing_dev_info *bdi) |
| 611 | { |
| 612 | struct radix_tree_iter iter; |
| 613 | void **slot; |
| 614 | struct bdi_writeback *wb; |
| 615 | |
| 616 | WARN_ON(test_bit(WB_registered, &bdi->wb.state)); |
| 617 | |
| 618 | spin_lock_irq(&cgwb_lock); |
| 619 | radix_tree_for_each_slot(slot, &bdi->cgwb_tree, &iter, 0) |
| 620 | cgwb_kill(*slot); |
| 621 | spin_unlock_irq(&cgwb_lock); |
| 622 | |
| 623 | mutex_lock(&bdi->cgwb_release_mutex); |
| 624 | spin_lock_irq(&cgwb_lock); |
| 625 | while (!list_empty(&bdi->wb_list)) { |
| 626 | wb = list_first_entry(&bdi->wb_list, struct bdi_writeback, |
| 627 | bdi_node); |
| 628 | spin_unlock_irq(&cgwb_lock); |
| 629 | wb_shutdown(wb); |
| 630 | spin_lock_irq(&cgwb_lock); |
| 631 | } |
| 632 | spin_unlock_irq(&cgwb_lock); |
| 633 | mutex_unlock(&bdi->cgwb_release_mutex); |
| 634 | } |
| 635 | |
| 636 | /** |
| 637 | * wb_memcg_offline - kill all wb's associated with a memcg being offlined |
| 638 | * @memcg: memcg being offlined |
| 639 | * |
| 640 | * Also prevents creation of any new wb's associated with @memcg. |
| 641 | */ |
| 642 | void wb_memcg_offline(struct mem_cgroup *memcg) |
| 643 | { |
| 644 | struct list_head *memcg_cgwb_list = &memcg->cgwb_list; |
| 645 | struct bdi_writeback *wb, *next; |
| 646 | |
| 647 | spin_lock_irq(&cgwb_lock); |
| 648 | list_for_each_entry_safe(wb, next, memcg_cgwb_list, memcg_node) |
| 649 | cgwb_kill(wb); |
| 650 | memcg_cgwb_list->next = NULL; /* prevent new wb's */ |
| 651 | spin_unlock_irq(&cgwb_lock); |
| 652 | } |
| 653 | |
| 654 | /** |
| 655 | * wb_blkcg_offline - kill all wb's associated with a blkcg being offlined |
| 656 | * @blkcg: blkcg being offlined |
| 657 | * |
| 658 | * Also prevents creation of any new wb's associated with @blkcg. |
| 659 | */ |
| 660 | void wb_blkcg_offline(struct blkcg *blkcg) |
| 661 | { |
| 662 | struct bdi_writeback *wb, *next; |
| 663 | |
| 664 | spin_lock_irq(&cgwb_lock); |
| 665 | list_for_each_entry_safe(wb, next, &blkcg->cgwb_list, blkcg_node) |
| 666 | cgwb_kill(wb); |
| 667 | blkcg->cgwb_list.next = NULL; /* prevent new wb's */ |
| 668 | spin_unlock_irq(&cgwb_lock); |
| 669 | } |
| 670 | |
| 671 | static void cgwb_bdi_register(struct backing_dev_info *bdi) |
| 672 | { |
| 673 | spin_lock_irq(&cgwb_lock); |
| 674 | list_add_tail_rcu(&bdi->wb.bdi_node, &bdi->wb_list); |
| 675 | spin_unlock_irq(&cgwb_lock); |
| 676 | } |
| 677 | |
| 678 | static int __init cgwb_init(void) |
| 679 | { |
| 680 | /* |
| 681 | * There can be many concurrent release work items overwhelming |
| 682 | * system_wq. Put them in a separate wq and limit concurrency. |
| 683 | * There's no point in executing many of these in parallel. |
| 684 | */ |
| 685 | cgwb_release_wq = alloc_workqueue("cgwb_release", 0, 1); |
| 686 | if (!cgwb_release_wq) |
| 687 | return -ENOMEM; |
| 688 | |
| 689 | return 0; |
| 690 | } |
| 691 | subsys_initcall(cgwb_init); |
| 692 | |
| 693 | #else /* CONFIG_CGROUP_WRITEBACK */ |
| 694 | |
| 695 | static int cgwb_bdi_init(struct backing_dev_info *bdi) |
| 696 | { |
| 697 | return wb_init(&bdi->wb, bdi, GFP_KERNEL); |
| 698 | } |
| 699 | |
| 700 | static void cgwb_bdi_unregister(struct backing_dev_info *bdi) { } |
| 701 | |
| 702 | static void cgwb_bdi_register(struct backing_dev_info *bdi) |
| 703 | { |
| 704 | list_add_tail_rcu(&bdi->wb.bdi_node, &bdi->wb_list); |
| 705 | } |
| 706 | |
| 707 | static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb) |
| 708 | { |
| 709 | list_del_rcu(&wb->bdi_node); |
| 710 | } |
| 711 | |
| 712 | #endif /* CONFIG_CGROUP_WRITEBACK */ |
| 713 | |
| 714 | static int bdi_init(struct backing_dev_info *bdi) |
| 715 | { |
| 716 | int ret; |
| 717 | |
| 718 | bdi->dev = NULL; |
| 719 | |
| 720 | kref_init(&bdi->refcnt); |
| 721 | bdi->min_ratio = 0; |
| 722 | bdi->max_ratio = 100; |
| 723 | bdi->max_prop_frac = FPROP_FRAC_BASE; |
| 724 | INIT_LIST_HEAD(&bdi->bdi_list); |
| 725 | INIT_LIST_HEAD(&bdi->wb_list); |
| 726 | init_waitqueue_head(&bdi->wb_waitq); |
| 727 | |
| 728 | ret = cgwb_bdi_init(bdi); |
| 729 | |
| 730 | return ret; |
| 731 | } |
| 732 | |
| 733 | struct backing_dev_info *bdi_alloc(int node_id) |
| 734 | { |
| 735 | struct backing_dev_info *bdi; |
| 736 | |
| 737 | bdi = kzalloc_node(sizeof(*bdi), GFP_KERNEL, node_id); |
| 738 | if (!bdi) |
| 739 | return NULL; |
| 740 | |
| 741 | if (bdi_init(bdi)) { |
| 742 | kfree(bdi); |
| 743 | return NULL; |
| 744 | } |
| 745 | bdi->capabilities = BDI_CAP_WRITEBACK | BDI_CAP_WRITEBACK_ACCT; |
| 746 | bdi->ra_pages = VM_READAHEAD_PAGES; |
| 747 | bdi->io_pages = VM_READAHEAD_PAGES; |
| 748 | return bdi; |
| 749 | } |
| 750 | EXPORT_SYMBOL(bdi_alloc); |
| 751 | |
| 752 | static struct rb_node **bdi_lookup_rb_node(u64 id, struct rb_node **parentp) |
| 753 | { |
| 754 | struct rb_node **p = &bdi_tree.rb_node; |
| 755 | struct rb_node *parent = NULL; |
| 756 | struct backing_dev_info *bdi; |
| 757 | |
| 758 | lockdep_assert_held(&bdi_lock); |
| 759 | |
| 760 | while (*p) { |
| 761 | parent = *p; |
| 762 | bdi = rb_entry(parent, struct backing_dev_info, rb_node); |
| 763 | |
| 764 | if (bdi->id > id) |
| 765 | p = &(*p)->rb_left; |
| 766 | else if (bdi->id < id) |
| 767 | p = &(*p)->rb_right; |
| 768 | else |
| 769 | break; |
| 770 | } |
| 771 | |
| 772 | if (parentp) |
| 773 | *parentp = parent; |
| 774 | return p; |
| 775 | } |
| 776 | |
| 777 | /** |
| 778 | * bdi_get_by_id - lookup and get bdi from its id |
| 779 | * @id: bdi id to lookup |
| 780 | * |
| 781 | * Find bdi matching @id and get it. Returns NULL if the matching bdi |
| 782 | * doesn't exist or is already unregistered. |
| 783 | */ |
| 784 | struct backing_dev_info *bdi_get_by_id(u64 id) |
| 785 | { |
| 786 | struct backing_dev_info *bdi = NULL; |
| 787 | struct rb_node **p; |
| 788 | |
| 789 | spin_lock_bh(&bdi_lock); |
| 790 | p = bdi_lookup_rb_node(id, NULL); |
| 791 | if (*p) { |
| 792 | bdi = rb_entry(*p, struct backing_dev_info, rb_node); |
| 793 | bdi_get(bdi); |
| 794 | } |
| 795 | spin_unlock_bh(&bdi_lock); |
| 796 | |
| 797 | return bdi; |
| 798 | } |
| 799 | |
| 800 | int bdi_register_va(struct backing_dev_info *bdi, const char *fmt, va_list args) |
| 801 | { |
| 802 | struct device *dev; |
| 803 | struct rb_node *parent, **p; |
| 804 | |
| 805 | if (bdi->dev) /* The driver needs to use separate queues per device */ |
| 806 | return 0; |
| 807 | |
| 808 | vsnprintf(bdi->dev_name, sizeof(bdi->dev_name), fmt, args); |
| 809 | dev = device_create(bdi_class, NULL, MKDEV(0, 0), bdi, bdi->dev_name); |
| 810 | if (IS_ERR(dev)) |
| 811 | return PTR_ERR(dev); |
| 812 | |
| 813 | cgwb_bdi_register(bdi); |
| 814 | bdi->dev = dev; |
| 815 | |
| 816 | bdi_debug_register(bdi, dev_name(dev)); |
| 817 | set_bit(WB_registered, &bdi->wb.state); |
| 818 | |
| 819 | spin_lock_bh(&bdi_lock); |
| 820 | |
| 821 | bdi->id = ++bdi_id_cursor; |
| 822 | |
| 823 | p = bdi_lookup_rb_node(bdi->id, &parent); |
| 824 | rb_link_node(&bdi->rb_node, parent, p); |
| 825 | rb_insert_color(&bdi->rb_node, &bdi_tree); |
| 826 | |
| 827 | list_add_tail_rcu(&bdi->bdi_list, &bdi_list); |
| 828 | |
| 829 | spin_unlock_bh(&bdi_lock); |
| 830 | |
| 831 | trace_writeback_bdi_register(bdi); |
| 832 | return 0; |
| 833 | } |
| 834 | |
| 835 | int bdi_register(struct backing_dev_info *bdi, const char *fmt, ...) |
| 836 | { |
| 837 | va_list args; |
| 838 | int ret; |
| 839 | |
| 840 | va_start(args, fmt); |
| 841 | ret = bdi_register_va(bdi, fmt, args); |
| 842 | va_end(args); |
| 843 | return ret; |
| 844 | } |
| 845 | EXPORT_SYMBOL(bdi_register); |
| 846 | |
| 847 | void bdi_set_owner(struct backing_dev_info *bdi, struct device *owner) |
| 848 | { |
| 849 | WARN_ON_ONCE(bdi->owner); |
| 850 | bdi->owner = owner; |
| 851 | get_device(owner); |
| 852 | } |
| 853 | |
| 854 | /* |
| 855 | * Remove bdi from bdi_list, and ensure that it is no longer visible |
| 856 | */ |
| 857 | static void bdi_remove_from_list(struct backing_dev_info *bdi) |
| 858 | { |
| 859 | spin_lock_bh(&bdi_lock); |
| 860 | rb_erase(&bdi->rb_node, &bdi_tree); |
| 861 | list_del_rcu(&bdi->bdi_list); |
| 862 | spin_unlock_bh(&bdi_lock); |
| 863 | |
| 864 | synchronize_rcu_expedited(); |
| 865 | } |
| 866 | |
| 867 | void bdi_unregister(struct backing_dev_info *bdi) |
| 868 | { |
| 869 | /* make sure nobody finds us on the bdi_list anymore */ |
| 870 | bdi_remove_from_list(bdi); |
| 871 | wb_shutdown(&bdi->wb); |
| 872 | cgwb_bdi_unregister(bdi); |
| 873 | |
| 874 | if (bdi->dev) { |
| 875 | bdi_debug_unregister(bdi); |
| 876 | device_unregister(bdi->dev); |
| 877 | bdi->dev = NULL; |
| 878 | } |
| 879 | |
| 880 | if (bdi->owner) { |
| 881 | put_device(bdi->owner); |
| 882 | bdi->owner = NULL; |
| 883 | } |
| 884 | } |
| 885 | |
| 886 | static void release_bdi(struct kref *ref) |
| 887 | { |
| 888 | struct backing_dev_info *bdi = |
| 889 | container_of(ref, struct backing_dev_info, refcnt); |
| 890 | |
| 891 | if (test_bit(WB_registered, &bdi->wb.state)) |
| 892 | bdi_unregister(bdi); |
| 893 | WARN_ON_ONCE(bdi->dev); |
| 894 | wb_exit(&bdi->wb); |
| 895 | kfree(bdi); |
| 896 | } |
| 897 | |
| 898 | void bdi_put(struct backing_dev_info *bdi) |
| 899 | { |
| 900 | kref_put(&bdi->refcnt, release_bdi); |
| 901 | } |
| 902 | EXPORT_SYMBOL(bdi_put); |
| 903 | |
| 904 | const char *bdi_dev_name(struct backing_dev_info *bdi) |
| 905 | { |
| 906 | if (!bdi || !bdi->dev) |
| 907 | return bdi_unknown_name; |
| 908 | return bdi->dev_name; |
| 909 | } |
| 910 | EXPORT_SYMBOL_GPL(bdi_dev_name); |
| 911 | |
| 912 | static wait_queue_head_t congestion_wqh[2] = { |
| 913 | __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[0]), |
| 914 | __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[1]) |
| 915 | }; |
| 916 | static atomic_t nr_wb_congested[2]; |
| 917 | |
| 918 | void clear_bdi_congested(struct backing_dev_info *bdi, int sync) |
| 919 | { |
| 920 | wait_queue_head_t *wqh = &congestion_wqh[sync]; |
| 921 | enum wb_congested_state bit; |
| 922 | |
| 923 | bit = sync ? WB_sync_congested : WB_async_congested; |
| 924 | if (test_and_clear_bit(bit, &bdi->wb.congested)) |
| 925 | atomic_dec(&nr_wb_congested[sync]); |
| 926 | smp_mb__after_atomic(); |
| 927 | if (waitqueue_active(wqh)) |
| 928 | wake_up(wqh); |
| 929 | } |
| 930 | EXPORT_SYMBOL(clear_bdi_congested); |
| 931 | |
| 932 | void set_bdi_congested(struct backing_dev_info *bdi, int sync) |
| 933 | { |
| 934 | enum wb_congested_state bit; |
| 935 | |
| 936 | bit = sync ? WB_sync_congested : WB_async_congested; |
| 937 | if (!test_and_set_bit(bit, &bdi->wb.congested)) |
| 938 | atomic_inc(&nr_wb_congested[sync]); |
| 939 | } |
| 940 | EXPORT_SYMBOL(set_bdi_congested); |
| 941 | |
| 942 | /** |
| 943 | * congestion_wait - wait for a backing_dev to become uncongested |
| 944 | * @sync: SYNC or ASYNC IO |
| 945 | * @timeout: timeout in jiffies |
| 946 | * |
| 947 | * Waits for up to @timeout jiffies for a backing_dev (any backing_dev) to exit |
| 948 | * write congestion. If no backing_devs are congested then just wait for the |
| 949 | * next write to be completed. |
| 950 | */ |
| 951 | long congestion_wait(int sync, long timeout) |
| 952 | { |
| 953 | long ret; |
| 954 | unsigned long start = jiffies; |
| 955 | DEFINE_WAIT(wait); |
| 956 | wait_queue_head_t *wqh = &congestion_wqh[sync]; |
| 957 | |
| 958 | prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE); |
| 959 | ret = io_schedule_timeout(timeout); |
| 960 | finish_wait(wqh, &wait); |
| 961 | |
| 962 | trace_writeback_congestion_wait(jiffies_to_usecs(timeout), |
| 963 | jiffies_to_usecs(jiffies - start)); |
| 964 | |
| 965 | return ret; |
| 966 | } |
| 967 | EXPORT_SYMBOL(congestion_wait); |
| 968 | |
| 969 | /** |
| 970 | * wait_iff_congested - Conditionally wait for a backing_dev to become uncongested or a pgdat to complete writes |
| 971 | * @sync: SYNC or ASYNC IO |
| 972 | * @timeout: timeout in jiffies |
| 973 | * |
| 974 | * In the event of a congested backing_dev (any backing_dev) this waits |
| 975 | * for up to @timeout jiffies for either a BDI to exit congestion of the |
| 976 | * given @sync queue or a write to complete. |
| 977 | * |
| 978 | * The return value is 0 if the sleep is for the full timeout. Otherwise, |
| 979 | * it is the number of jiffies that were still remaining when the function |
| 980 | * returned. return_value == timeout implies the function did not sleep. |
| 981 | */ |
| 982 | long wait_iff_congested(int sync, long timeout) |
| 983 | { |
| 984 | long ret; |
| 985 | unsigned long start = jiffies; |
| 986 | DEFINE_WAIT(wait); |
| 987 | wait_queue_head_t *wqh = &congestion_wqh[sync]; |
| 988 | |
| 989 | /* |
| 990 | * If there is no congestion, yield if necessary instead |
| 991 | * of sleeping on the congestion queue |
| 992 | */ |
| 993 | if (atomic_read(&nr_wb_congested[sync]) == 0) { |
| 994 | cond_resched(); |
| 995 | |
| 996 | /* In case we scheduled, work out time remaining */ |
| 997 | ret = timeout - (jiffies - start); |
| 998 | if (ret < 0) |
| 999 | ret = 0; |
| 1000 | |
| 1001 | goto out; |
| 1002 | } |
| 1003 | |
| 1004 | /* Sleep until uncongested or a write happens */ |
| 1005 | prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE); |
| 1006 | ret = io_schedule_timeout(timeout); |
| 1007 | finish_wait(wqh, &wait); |
| 1008 | |
| 1009 | out: |
| 1010 | trace_writeback_wait_iff_congested(jiffies_to_usecs(timeout), |
| 1011 | jiffies_to_usecs(jiffies - start)); |
| 1012 | |
| 1013 | return ret; |
| 1014 | } |
| 1015 | EXPORT_SYMBOL(wait_iff_congested); |