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