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