drivers/md/dm-delay.c

   1 /*
   2  * Copyright (C) 2005-2007 Red Hat GmbH
   3  *
   4  * A target that delays reads and/or writes and can send
   5  * them to different devices.
   6  *
   7  * This file is released under the GPL.
   8  */
   9
  10 #include <linux/module.h>
  11 #include <linux/init.h>
  12 #include <linux/blkdev.h>
  13 #include <linux/bio.h>
  14 #include <linux/slab.h>
  15
  16 #include <linux/device-mapper.h>
  17
  18 #define DM_MSG_PREFIX "delay"
  19
  20 struct delay_c {
  21         struct timer_list delay_timer;
  22         struct mutex timer_lock;
  23         struct workqueue_struct *kdelayd_wq;
  24         struct work_struct flush_expired_bios;
  25         struct list_head delayed_bios;
  26         atomic_t may_delay;
  27
  28         struct dm_dev *dev_read;
  29         sector_t start_read;
  30         unsigned read_delay;
  31         unsigned reads;
  32
  33         struct dm_dev *dev_write;
  34         sector_t start_write;
  35         unsigned write_delay;
  36         unsigned writes;
  37 };
  38
  39 struct dm_delay_info {
  40         struct delay_c *context;
  41         struct list_head list;
  42         unsigned long expires;
  43 };
  44
  45 static DEFINE_MUTEX(delayed_bios_lock);
  46
  47 static void handle_delayed_timer(unsigned long data)
  48 {
  49         struct delay_c *dc = (struct delay_c *)data;
  50
  51         queue_work(dc->kdelayd_wq, &dc->flush_expired_bios);
  52 }
  53
  54 static void queue_timeout(struct delay_c *dc, unsigned long expires)
  55 {
  56         mutex_lock(&dc->timer_lock);
  57
  58         if (!timer_pending(&dc->delay_timer) || expires < dc->delay_timer.expires)
  59                 mod_timer(&dc->delay_timer, expires);
  60
  61         mutex_unlock(&dc->timer_lock);
  62 }
  63
  64 static void flush_bios(struct bio *bio)
  65 {
  66         struct bio *n;
  67
  68         while (bio) {
  69                 n = bio->bi_next;
  70                 bio->bi_next = NULL;
  71                 generic_make_request(bio);
  72                 bio = n;
  73         }
  74 }
  75
  76 static struct bio *flush_delayed_bios(struct delay_c *dc, int flush_all)
  77 {
  78         struct dm_delay_info *delayed, *next;
  79         unsigned long next_expires = 0;
  80         int start_timer = 0;
  81         struct bio_list flush_bios = { };
  82
  83         mutex_lock(&delayed_bios_lock);
  84         list_for_each_entry_safe(delayed, next, &dc->delayed_bios, list) {
  85                 if (flush_all || time_after_eq(jiffies, delayed->expires)) {
  86                         struct bio *bio = dm_bio_from_per_bio_data(delayed,
  87                                                 sizeof(struct dm_delay_info));
  88                         list_del(&delayed->list);
  89                         bio_list_add(&flush_bios, bio);
  90                         if ((bio_data_dir(bio) == WRITE))
  91                                 delayed->context->writes--;
  92                         else
  93                                 delayed->context->reads--;
  94                         continue;
  95                 }
  96
  97                 if (!start_timer) {
  98                         start_timer = 1;
  99                         next_expires = delayed->expires;
 100                 } else
 101                         next_expires = min(next_expires, delayed->expires);
 102         }
 103
 104         mutex_unlock(&delayed_bios_lock);
 105
 106         if (start_timer)
 107                 queue_timeout(dc, next_expires);
 108
 109         return bio_list_get(&flush_bios);
 110 }
 111
 112 static void flush_expired_bios(struct work_struct *work)
 113 {
 114         struct delay_c *dc;
 115
 116         dc = container_of(work, struct delay_c, flush_expired_bios);
 117         flush_bios(flush_delayed_bios(dc, 0));
 118 }
 119
 120 /*
 121  * Mapping parameters:
 122  *    <device> <offset> <delay> [<write_device> <write_offset> <write_delay>]
 123  *
 124  * With separate write parameters, the first set is only used for reads.
 125  * Delays are specified in milliseconds.
 126  */
 127 static int delay_ctr(struct dm_target *ti, unsigned int argc, char **argv)
 128 {
 129         struct delay_c *dc;
 130         unsigned long long tmpll;
 131         char dummy;
 132         int ret;
 133
 134         if (argc != 3 && argc != 6) {
 135                 ti->error = "requires exactly 3 or 6 arguments";
 136                 return -EINVAL;
 137         }
 138
 139         dc = kmalloc(sizeof(*dc), GFP_KERNEL);
 140         if (!dc) {
 141                 ti->error = "Cannot allocate context";
 142                 return -ENOMEM;
 143         }
 144
 145         dc->reads = dc->writes = 0;
 146
 147         ret = -EINVAL;
 148         if (sscanf(argv[1], "%llu%c", &tmpll, &dummy) != 1) {
 149                 ti->error = "Invalid device sector";
 150                 goto bad;
 151         }
 152         dc->start_read = tmpll;
 153
 154         if (sscanf(argv[2], "%u%c", &dc->read_delay, &dummy) != 1) {
 155                 ti->error = "Invalid delay";
 156                 goto bad;
 157         }
 158
 159         ret = dm_get_device(ti, argv[0], dm_table_get_mode(ti->table),
 160                             &dc->dev_read);
 161         if (ret) {
 162                 ti->error = "Device lookup failed";
 163                 goto bad;
 164         }
 165
 166         ret = -EINVAL;
 167         dc->dev_write = NULL;
 168         if (argc == 3)
 169                 goto out;
 170
 171         if (sscanf(argv[4], "%llu%c", &tmpll, &dummy) != 1) {
 172                 ti->error = "Invalid write device sector";
 173                 goto bad_dev_read;
 174         }
 175         dc->start_write = tmpll;
 176
 177         if (sscanf(argv[5], "%u%c", &dc->write_delay, &dummy) != 1) {
 178                 ti->error = "Invalid write delay";
 179                 goto bad_dev_read;
 180         }
 181
 182         ret = dm_get_device(ti, argv[3], dm_table_get_mode(ti->table),
 183                             &dc->dev_write);
 184         if (ret) {
 185                 ti->error = "Write device lookup failed";
 186                 goto bad_dev_read;
 187         }
 188
 189 out:
 190         ret = -EINVAL;
 191         dc->kdelayd_wq = alloc_workqueue("kdelayd", WQ_MEM_RECLAIM, 0);
 192         if (!dc->kdelayd_wq) {
 193                 DMERR("Couldn't start kdelayd");
 194                 goto bad_queue;
 195         }
 196
 197         setup_timer(&dc->delay_timer, handle_delayed_timer, (unsigned long)dc);
 198
 199         INIT_WORK(&dc->flush_expired_bios, flush_expired_bios);
 200         INIT_LIST_HEAD(&dc->delayed_bios);
 201         mutex_init(&dc->timer_lock);
 202         atomic_set(&dc->may_delay, 1);
 203
 204         ti->num_flush_bios = 1;
 205         ti->num_discard_bios = 1;
 206         ti->per_bio_data_size = sizeof(struct dm_delay_info);
 207         ti->private = dc;
 208         return 0;
 209
 210 bad_queue:
 211         if (dc->dev_write)
 212                 dm_put_device(ti, dc->dev_write);
 213 bad_dev_read:
 214         dm_put_device(ti, dc->dev_read);
 215 bad:
 216         kfree(dc);
 217         return ret;
 218 }
 219
 220 static void delay_dtr(struct dm_target *ti)
 221 {
 222         struct delay_c *dc = ti->private;
 223
 224         destroy_workqueue(dc->kdelayd_wq);
 225
 226         dm_put_device(ti, dc->dev_read);
 227
 228         if (dc->dev_write)
 229                 dm_put_device(ti, dc->dev_write);
 230
 231         kfree(dc);
 232 }
 233
 234 static int delay_bio(struct delay_c *dc, int delay, struct bio *bio)
 235 {
 236         struct dm_delay_info *delayed;
 237         unsigned long expires = 0;
 238
 239         if (!delay || !atomic_read(&dc->may_delay))
 240                 return 1;
 241
 242         delayed = dm_per_bio_data(bio, sizeof(struct dm_delay_info));
 243
 244         delayed->context = dc;
 245         delayed->expires = expires = jiffies + msecs_to_jiffies(delay);
 246
 247         mutex_lock(&delayed_bios_lock);
 248
 249         if (bio_data_dir(bio) == WRITE)
 250                 dc->writes++;
 251         else
 252                 dc->reads++;
 253
 254         list_add_tail(&delayed->list, &dc->delayed_bios);
 255
 256         mutex_unlock(&delayed_bios_lock);
 257
 258         queue_timeout(dc, expires);
 259
 260         return 0;
 261 }
 262
 263 static void delay_presuspend(struct dm_target *ti)
 264 {
 265         struct delay_c *dc = ti->private;
 266
 267         atomic_set(&dc->may_delay, 0);
 268         del_timer_sync(&dc->delay_timer);
 269         flush_bios(flush_delayed_bios(dc, 1));
 270 }
 271
 272 static void delay_resume(struct dm_target *ti)
 273 {
 274         struct delay_c *dc = ti->private;
 275
 276         atomic_set(&dc->may_delay, 1);
 277 }
 278
 279 static int delay_map(struct dm_target *ti, struct bio *bio)
 280 {
 281         struct delay_c *dc = ti->private;
 282
 283         if ((bio_data_dir(bio) == WRITE) && (dc->dev_write)) {
 284                 bio->bi_bdev = dc->dev_write->bdev;
 285                 if (bio_sectors(bio))
 286                         bio->bi_iter.bi_sector = dc->start_write +
 287                                 dm_target_offset(ti, bio->bi_iter.bi_sector);
 288
 289                 return delay_bio(dc, dc->write_delay, bio);
 290         }
 291
 292         bio->bi_bdev = dc->dev_read->bdev;
 293         bio->bi_iter.bi_sector = dc->start_read +
 294                 dm_target_offset(ti, bio->bi_iter.bi_sector);
 295
 296         return delay_bio(dc, dc->read_delay, bio);
 297 }
 298
 299 static void delay_status(struct dm_target *ti, status_type_t type,
 300                          unsigned status_flags, char *result, unsigned maxlen)
 301 {
 302         struct delay_c *dc = ti->private;
 303         int sz = 0;
 304
 305         switch (type) {
 306         case STATUSTYPE_INFO:
 307                 DMEMIT("%u %u", dc->reads, dc->writes);
 308                 break;
 309
 310         case STATUSTYPE_TABLE:
 311                 DMEMIT("%s %llu %u", dc->dev_read->name,
 312                        (unsigned long long) dc->start_read,
 313                        dc->read_delay);
 314                 if (dc->dev_write)
 315                         DMEMIT(" %s %llu %u", dc->dev_write->name,
 316                                (unsigned long long) dc->start_write,
 317                                dc->write_delay);
 318                 break;
 319         }
 320 }
 321
 322 static int delay_iterate_devices(struct dm_target *ti,
 323                                  iterate_devices_callout_fn fn, void *data)
 324 {
 325         struct delay_c *dc = ti->private;
 326         int ret = 0;
 327
 328         ret = fn(ti, dc->dev_read, dc->start_read, ti->len, data);
 329         if (ret)
 330                 goto out;
 331
 332         if (dc->dev_write)
 333                 ret = fn(ti, dc->dev_write, dc->start_write, ti->len, data);
 334
 335 out:
 336         return ret;
 337 }
 338
 339 static struct target_type delay_target = {
 340         .name        = "delay",
 341         .version     = {1, 2, 1},
 342         .module      = THIS_MODULE,
 343         .ctr         = delay_ctr,
 344         .dtr         = delay_dtr,
 345         .map         = delay_map,
 346         .presuspend  = delay_presuspend,
 347         .resume      = delay_resume,
 348         .status      = delay_status,
 349         .iterate_devices = delay_iterate_devices,
 350 };
 351
 352 static int __init dm_delay_init(void)
 353 {
 354         int r;
 355
 356         r = dm_register_target(&delay_target);
 357         if (r < 0) {
 358                 DMERR("register failed %d", r);
 359                 goto bad_register;
 360         }
 361
 362         return 0;
 363
 364 bad_register:
 365         return r;
 366 }
 367
 368 static void __exit dm_delay_exit(void)
 369 {
 370         dm_unregister_target(&delay_target);
 371 }
 372
 373 /* Module hooks */
 374 module_init(dm_delay_init);
 375 module_exit(dm_delay_exit);
 376
 377 MODULE_DESCRIPTION(DM_NAME " delay target");
 378 MODULE_AUTHOR("Heinz Mauelshagen <mauelshagen@redhat.com>");
 379 MODULE_LICENSE("GPL");