1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * Copyright (C) 2005-2007 Red Hat GmbH 4 * 5 * A target that delays reads and/or writes and can send 6 * them to different devices. 7 * 8 * This file is released under the GPL. 9 */ 10 11 #include <linux/module.h> 12 #include <linux/init.h> 13 #include <linux/blkdev.h> 14 #include <linux/bio.h> 15 #include <linux/slab.h> 16 #include <linux/kthread.h> 17 18 #include <linux/device-mapper.h> 19 20 #define DM_MSG_PREFIX "delay" 21 22 struct delay_class { 23 struct dm_dev *dev; 24 sector_t start; 25 unsigned int delay; 26 unsigned int ops; 27 }; 28 29 struct delay_c { 30 struct timer_list delay_timer; 31 struct mutex timer_lock; 32 struct workqueue_struct *kdelayd_wq; 33 struct work_struct flush_expired_bios; 34 struct list_head delayed_bios; 35 struct task_struct *worker; 36 bool may_delay; 37 38 struct delay_class read; 39 struct delay_class write; 40 struct delay_class flush; 41 42 int argc; 43 }; 44 45 struct dm_delay_info { 46 struct delay_c *context; 47 struct delay_class *class; 48 struct list_head list; 49 unsigned long expires; 50 }; 51 52 static DEFINE_MUTEX(delayed_bios_lock); 53 54 static void handle_delayed_timer(struct timer_list *t) 55 { 56 struct delay_c *dc = from_timer(dc, t, delay_timer); 57 58 queue_work(dc->kdelayd_wq, &dc->flush_expired_bios); 59 } 60 61 static void queue_timeout(struct delay_c *dc, unsigned long expires) 62 { 63 mutex_lock(&dc->timer_lock); 64 65 if (!timer_pending(&dc->delay_timer) || expires < dc->delay_timer.expires) 66 mod_timer(&dc->delay_timer, expires); 67 68 mutex_unlock(&dc->timer_lock); 69 } 70 71 static inline bool delay_is_fast(struct delay_c *dc) 72 { 73 return !!dc->worker; 74 } 75 76 static void flush_bios(struct bio *bio) 77 { 78 struct bio *n; 79 80 while (bio) { 81 n = bio->bi_next; 82 bio->bi_next = NULL; 83 dm_submit_bio_remap(bio, NULL); 84 bio = n; 85 } 86 } 87 88 static void flush_delayed_bios_fast(struct delay_c *dc, bool flush_all) 89 { 90 struct dm_delay_info *delayed, *next; 91 struct bio_list flush_bio_list; 92 bio_list_init(&flush_bio_list); 93 94 mutex_lock(&delayed_bios_lock); 95 list_for_each_entry_safe(delayed, next, &dc->delayed_bios, list) { 96 if (flush_all || time_after_eq(jiffies, delayed->expires)) { 97 struct bio *bio = dm_bio_from_per_bio_data(delayed, 98 sizeof(struct dm_delay_info)); 99 list_del(&delayed->list); 100 bio_list_add(&flush_bio_list, bio); 101 delayed->class->ops--; 102 } 103 } 104 mutex_unlock(&delayed_bios_lock); 105 106 flush_bios(bio_list_get(&flush_bio_list)); 107 } 108 109 static int flush_worker_fn(void *data) 110 { 111 struct delay_c *dc = data; 112 113 while (!kthread_should_stop()) { 114 flush_delayed_bios_fast(dc, false); 115 mutex_lock(&delayed_bios_lock); 116 if (unlikely(list_empty(&dc->delayed_bios))) { 117 set_current_state(TASK_INTERRUPTIBLE); 118 mutex_unlock(&delayed_bios_lock); 119 schedule(); 120 } else { 121 mutex_unlock(&delayed_bios_lock); 122 cond_resched(); 123 } 124 } 125 126 return 0; 127 } 128 129 static void flush_delayed_bios(struct delay_c *dc, bool flush_all) 130 { 131 struct dm_delay_info *delayed, *next; 132 unsigned long next_expires = 0; 133 unsigned long start_timer = 0; 134 struct bio_list flush_bio_list; 135 bio_list_init(&flush_bio_list); 136 137 mutex_lock(&delayed_bios_lock); 138 list_for_each_entry_safe(delayed, next, &dc->delayed_bios, list) { 139 if (flush_all || time_after_eq(jiffies, delayed->expires)) { 140 struct bio *bio = dm_bio_from_per_bio_data(delayed, 141 sizeof(struct dm_delay_info)); 142 list_del(&delayed->list); 143 bio_list_add(&flush_bio_list, bio); 144 delayed->class->ops--; 145 continue; 146 } 147 148 if (!start_timer) { 149 start_timer = 1; 150 next_expires = delayed->expires; 151 } else 152 next_expires = min(next_expires, delayed->expires); 153 } 154 mutex_unlock(&delayed_bios_lock); 155 156 if (start_timer) 157 queue_timeout(dc, next_expires); 158 159 flush_bios(bio_list_get(&flush_bio_list)); 160 } 161 162 static void flush_expired_bios(struct work_struct *work) 163 { 164 struct delay_c *dc; 165 166 dc = container_of(work, struct delay_c, flush_expired_bios); 167 if (delay_is_fast(dc)) 168 flush_delayed_bios_fast(dc, false); 169 else 170 flush_delayed_bios(dc, false); 171 } 172 173 static void delay_dtr(struct dm_target *ti) 174 { 175 struct delay_c *dc = ti->private; 176 177 if (dc->kdelayd_wq) 178 destroy_workqueue(dc->kdelayd_wq); 179 180 if (dc->read.dev) 181 dm_put_device(ti, dc->read.dev); 182 if (dc->write.dev) 183 dm_put_device(ti, dc->write.dev); 184 if (dc->flush.dev) 185 dm_put_device(ti, dc->flush.dev); 186 if (dc->worker) 187 kthread_stop(dc->worker); 188 189 mutex_destroy(&dc->timer_lock); 190 191 kfree(dc); 192 } 193 194 static int delay_class_ctr(struct dm_target *ti, struct delay_class *c, char **argv) 195 { 196 int ret; 197 unsigned long long tmpll; 198 char dummy; 199 200 if (sscanf(argv[1], "%llu%c", &tmpll, &dummy) != 1 || tmpll != (sector_t)tmpll) { 201 ti->error = "Invalid device sector"; 202 return -EINVAL; 203 } 204 c->start = tmpll; 205 206 if (sscanf(argv[2], "%u%c", &c->delay, &dummy) != 1) { 207 ti->error = "Invalid delay"; 208 return -EINVAL; 209 } 210 211 ret = dm_get_device(ti, argv[0], dm_table_get_mode(ti->table), &c->dev); 212 if (ret) { 213 ti->error = "Device lookup failed"; 214 return ret; 215 } 216 217 return 0; 218 } 219 220 /* 221 * Mapping parameters: 222 * <device> <offset> <delay> [<write_device> <write_offset> <write_delay>] 223 * 224 * With separate write parameters, the first set is only used for reads. 225 * Offsets are specified in sectors. 226 * Delays are specified in milliseconds. 227 */ 228 static int delay_ctr(struct dm_target *ti, unsigned int argc, char **argv) 229 { 230 struct delay_c *dc; 231 int ret; 232 unsigned int max_delay; 233 234 if (argc != 3 && argc != 6 && argc != 9) { 235 ti->error = "Requires exactly 3, 6 or 9 arguments"; 236 return -EINVAL; 237 } 238 239 dc = kzalloc(sizeof(*dc), GFP_KERNEL); 240 if (!dc) { 241 ti->error = "Cannot allocate context"; 242 return -ENOMEM; 243 } 244 245 ti->private = dc; 246 INIT_LIST_HEAD(&dc->delayed_bios); 247 mutex_init(&dc->timer_lock); 248 dc->may_delay = true; 249 dc->argc = argc; 250 251 ret = delay_class_ctr(ti, &dc->read, argv); 252 if (ret) 253 goto bad; 254 max_delay = dc->read.delay; 255 256 if (argc == 3) { 257 ret = delay_class_ctr(ti, &dc->write, argv); 258 if (ret) 259 goto bad; 260 ret = delay_class_ctr(ti, &dc->flush, argv); 261 if (ret) 262 goto bad; 263 max_delay = max(max_delay, dc->write.delay); 264 max_delay = max(max_delay, dc->flush.delay); 265 goto out; 266 } 267 268 ret = delay_class_ctr(ti, &dc->write, argv + 3); 269 if (ret) 270 goto bad; 271 if (argc == 6) { 272 ret = delay_class_ctr(ti, &dc->flush, argv + 3); 273 if (ret) 274 goto bad; 275 max_delay = max(max_delay, dc->flush.delay); 276 goto out; 277 } 278 279 ret = delay_class_ctr(ti, &dc->flush, argv + 6); 280 if (ret) 281 goto bad; 282 max_delay = max(max_delay, dc->flush.delay); 283 284 out: 285 if (max_delay < 50) { 286 /* 287 * In case of small requested delays, use kthread instead of 288 * timers and workqueue to achieve better latency. 289 */ 290 dc->worker = kthread_create(&flush_worker_fn, dc, 291 "dm-delay-flush-worker"); 292 if (IS_ERR(dc->worker)) { 293 ret = PTR_ERR(dc->worker); 294 dc->worker = NULL; 295 goto bad; 296 } 297 } else { 298 timer_setup(&dc->delay_timer, handle_delayed_timer, 0); 299 INIT_WORK(&dc->flush_expired_bios, flush_expired_bios); 300 dc->kdelayd_wq = alloc_workqueue("kdelayd", WQ_MEM_RECLAIM, 0); 301 if (!dc->kdelayd_wq) { 302 ret = -EINVAL; 303 DMERR("Couldn't start kdelayd"); 304 goto bad; 305 } 306 } 307 308 ti->num_flush_bios = 1; 309 ti->num_discard_bios = 1; 310 ti->accounts_remapped_io = true; 311 ti->per_io_data_size = sizeof(struct dm_delay_info); 312 return 0; 313 314 bad: 315 delay_dtr(ti); 316 return ret; 317 } 318 319 static int delay_bio(struct delay_c *dc, struct delay_class *c, struct bio *bio) 320 { 321 struct dm_delay_info *delayed; 322 unsigned long expires = 0; 323 324 if (!c->delay) 325 return DM_MAPIO_REMAPPED; 326 327 delayed = dm_per_bio_data(bio, sizeof(struct dm_delay_info)); 328 329 delayed->context = dc; 330 delayed->expires = expires = jiffies + msecs_to_jiffies(c->delay); 331 332 mutex_lock(&delayed_bios_lock); 333 if (unlikely(!dc->may_delay)) { 334 mutex_unlock(&delayed_bios_lock); 335 return DM_MAPIO_REMAPPED; 336 } 337 c->ops++; 338 list_add_tail(&delayed->list, &dc->delayed_bios); 339 mutex_unlock(&delayed_bios_lock); 340 341 if (delay_is_fast(dc)) 342 wake_up_process(dc->worker); 343 else 344 queue_timeout(dc, expires); 345 346 return DM_MAPIO_SUBMITTED; 347 } 348 349 static void delay_presuspend(struct dm_target *ti) 350 { 351 struct delay_c *dc = ti->private; 352 353 mutex_lock(&delayed_bios_lock); 354 dc->may_delay = false; 355 mutex_unlock(&delayed_bios_lock); 356 357 if (delay_is_fast(dc)) { 358 flush_delayed_bios_fast(dc, true); 359 } else { 360 del_timer_sync(&dc->delay_timer); 361 flush_delayed_bios(dc, true); 362 } 363 } 364 365 static void delay_resume(struct dm_target *ti) 366 { 367 struct delay_c *dc = ti->private; 368 369 dc->may_delay = true; 370 } 371 372 static int delay_map(struct dm_target *ti, struct bio *bio) 373 { 374 struct delay_c *dc = ti->private; 375 struct delay_class *c; 376 struct dm_delay_info *delayed = dm_per_bio_data(bio, sizeof(struct dm_delay_info)); 377 378 if (bio_data_dir(bio) == WRITE) { 379 if (unlikely(bio->bi_opf & REQ_PREFLUSH)) 380 c = &dc->flush; 381 else 382 c = &dc->write; 383 } else { 384 c = &dc->read; 385 } 386 delayed->class = c; 387 bio_set_dev(bio, c->dev->bdev); 388 bio->bi_iter.bi_sector = c->start + dm_target_offset(ti, bio->bi_iter.bi_sector); 389 390 return delay_bio(dc, c, bio); 391 } 392 393 #define DMEMIT_DELAY_CLASS(c) \ 394 DMEMIT("%s %llu %u", (c)->dev->name, (unsigned long long)(c)->start, (c)->delay) 395 396 static void delay_status(struct dm_target *ti, status_type_t type, 397 unsigned int status_flags, char *result, unsigned int maxlen) 398 { 399 struct delay_c *dc = ti->private; 400 int sz = 0; 401 402 switch (type) { 403 case STATUSTYPE_INFO: 404 DMEMIT("%u %u %u", dc->read.ops, dc->write.ops, dc->flush.ops); 405 break; 406 407 case STATUSTYPE_TABLE: 408 DMEMIT_DELAY_CLASS(&dc->read); 409 if (dc->argc >= 6) { 410 DMEMIT(" "); 411 DMEMIT_DELAY_CLASS(&dc->write); 412 } 413 if (dc->argc >= 9) { 414 DMEMIT(" "); 415 DMEMIT_DELAY_CLASS(&dc->flush); 416 } 417 break; 418 419 case STATUSTYPE_IMA: 420 *result = '\0'; 421 break; 422 } 423 } 424 425 static int delay_iterate_devices(struct dm_target *ti, 426 iterate_devices_callout_fn fn, void *data) 427 { 428 struct delay_c *dc = ti->private; 429 int ret = 0; 430 431 ret = fn(ti, dc->read.dev, dc->read.start, ti->len, data); 432 if (ret) 433 goto out; 434 ret = fn(ti, dc->write.dev, dc->write.start, ti->len, data); 435 if (ret) 436 goto out; 437 ret = fn(ti, dc->flush.dev, dc->flush.start, ti->len, data); 438 if (ret) 439 goto out; 440 441 out: 442 return ret; 443 } 444 445 static struct target_type delay_target = { 446 .name = "delay", 447 .version = {1, 4, 0}, 448 .features = DM_TARGET_PASSES_INTEGRITY, 449 .module = THIS_MODULE, 450 .ctr = delay_ctr, 451 .dtr = delay_dtr, 452 .map = delay_map, 453 .presuspend = delay_presuspend, 454 .resume = delay_resume, 455 .status = delay_status, 456 .iterate_devices = delay_iterate_devices, 457 }; 458 module_dm(delay); 459 460 MODULE_DESCRIPTION(DM_NAME " delay target"); 461 MODULE_AUTHOR("Heinz Mauelshagen <[email protected]>"); 462 MODULE_LICENSE("GPL"); 463