| 1 | /* |
| 2 | * Functions related to barrier IO handling |
| 3 | */ |
| 4 | #include <linux/kernel.h> |
| 5 | #include <linux/module.h> |
| 6 | #include <linux/bio.h> |
| 7 | #include <linux/blkdev.h> |
| 8 | #include <linux/gfp.h> |
| 9 | |
| 10 | #include "blk.h" |
| 11 | |
| 12 | /** |
| 13 | * blk_queue_ordered - does this queue support ordered writes |
| 14 | * @q: the request queue |
| 15 | * @ordered: one of QUEUE_ORDERED_* |
| 16 | * @prepare_flush_fn: rq setup helper for cache flush ordered writes |
| 17 | * |
| 18 | * Description: |
| 19 | * For journalled file systems, doing ordered writes on a commit |
| 20 | * block instead of explicitly doing wait_on_buffer (which is bad |
| 21 | * for performance) can be a big win. Block drivers supporting this |
| 22 | * feature should call this function and indicate so. |
| 23 | * |
| 24 | **/ |
| 25 | int blk_queue_ordered(struct request_queue *q, unsigned ordered, |
| 26 | prepare_flush_fn *prepare_flush_fn) |
| 27 | { |
| 28 | if (!prepare_flush_fn && (ordered & (QUEUE_ORDERED_DO_PREFLUSH | |
| 29 | QUEUE_ORDERED_DO_POSTFLUSH))) { |
| 30 | printk(KERN_ERR "%s: prepare_flush_fn required\n", __func__); |
| 31 | return -EINVAL; |
| 32 | } |
| 33 | |
| 34 | if (ordered != QUEUE_ORDERED_NONE && |
| 35 | ordered != QUEUE_ORDERED_DRAIN && |
| 36 | ordered != QUEUE_ORDERED_DRAIN_FLUSH && |
| 37 | ordered != QUEUE_ORDERED_DRAIN_FUA && |
| 38 | ordered != QUEUE_ORDERED_TAG && |
| 39 | ordered != QUEUE_ORDERED_TAG_FLUSH && |
| 40 | ordered != QUEUE_ORDERED_TAG_FUA) { |
| 41 | printk(KERN_ERR "blk_queue_ordered: bad value %d\n", ordered); |
| 42 | return -EINVAL; |
| 43 | } |
| 44 | |
| 45 | q->ordered = ordered; |
| 46 | q->next_ordered = ordered; |
| 47 | q->prepare_flush_fn = prepare_flush_fn; |
| 48 | |
| 49 | return 0; |
| 50 | } |
| 51 | EXPORT_SYMBOL(blk_queue_ordered); |
| 52 | |
| 53 | /* |
| 54 | * Cache flushing for ordered writes handling |
| 55 | */ |
| 56 | unsigned blk_ordered_cur_seq(struct request_queue *q) |
| 57 | { |
| 58 | if (!q->ordseq) |
| 59 | return 0; |
| 60 | return 1 << ffz(q->ordseq); |
| 61 | } |
| 62 | |
| 63 | unsigned blk_ordered_req_seq(struct request *rq) |
| 64 | { |
| 65 | struct request_queue *q = rq->q; |
| 66 | |
| 67 | BUG_ON(q->ordseq == 0); |
| 68 | |
| 69 | if (rq == &q->pre_flush_rq) |
| 70 | return QUEUE_ORDSEQ_PREFLUSH; |
| 71 | if (rq == &q->bar_rq) |
| 72 | return QUEUE_ORDSEQ_BAR; |
| 73 | if (rq == &q->post_flush_rq) |
| 74 | return QUEUE_ORDSEQ_POSTFLUSH; |
| 75 | |
| 76 | /* |
| 77 | * !fs requests don't need to follow barrier ordering. Always |
| 78 | * put them at the front. This fixes the following deadlock. |
| 79 | * |
| 80 | * http://thread.gmane.org/gmane.linux.kernel/537473 |
| 81 | */ |
| 82 | if (!blk_fs_request(rq)) |
| 83 | return QUEUE_ORDSEQ_DRAIN; |
| 84 | |
| 85 | if ((rq->cmd_flags & REQ_ORDERED_COLOR) == |
| 86 | (q->orig_bar_rq->cmd_flags & REQ_ORDERED_COLOR)) |
| 87 | return QUEUE_ORDSEQ_DRAIN; |
| 88 | else |
| 89 | return QUEUE_ORDSEQ_DONE; |
| 90 | } |
| 91 | |
| 92 | bool blk_ordered_complete_seq(struct request_queue *q, unsigned seq, int error) |
| 93 | { |
| 94 | struct request *rq; |
| 95 | |
| 96 | if (error && !q->orderr) |
| 97 | q->orderr = error; |
| 98 | |
| 99 | BUG_ON(q->ordseq & seq); |
| 100 | q->ordseq |= seq; |
| 101 | |
| 102 | if (blk_ordered_cur_seq(q) != QUEUE_ORDSEQ_DONE) |
| 103 | return false; |
| 104 | |
| 105 | /* |
| 106 | * Okay, sequence complete. |
| 107 | */ |
| 108 | q->ordseq = 0; |
| 109 | rq = q->orig_bar_rq; |
| 110 | __blk_end_request_all(rq, q->orderr); |
| 111 | return true; |
| 112 | } |
| 113 | |
| 114 | static void pre_flush_end_io(struct request *rq, int error) |
| 115 | { |
| 116 | elv_completed_request(rq->q, rq); |
| 117 | blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_PREFLUSH, error); |
| 118 | } |
| 119 | |
| 120 | static void bar_end_io(struct request *rq, int error) |
| 121 | { |
| 122 | elv_completed_request(rq->q, rq); |
| 123 | blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_BAR, error); |
| 124 | } |
| 125 | |
| 126 | static void post_flush_end_io(struct request *rq, int error) |
| 127 | { |
| 128 | elv_completed_request(rq->q, rq); |
| 129 | blk_ordered_complete_seq(rq->q, QUEUE_ORDSEQ_POSTFLUSH, error); |
| 130 | } |
| 131 | |
| 132 | static void queue_flush(struct request_queue *q, unsigned which) |
| 133 | { |
| 134 | struct request *rq; |
| 135 | rq_end_io_fn *end_io; |
| 136 | |
| 137 | if (which == QUEUE_ORDERED_DO_PREFLUSH) { |
| 138 | rq = &q->pre_flush_rq; |
| 139 | end_io = pre_flush_end_io; |
| 140 | } else { |
| 141 | rq = &q->post_flush_rq; |
| 142 | end_io = post_flush_end_io; |
| 143 | } |
| 144 | |
| 145 | blk_rq_init(q, rq); |
| 146 | rq->cmd_flags = REQ_HARDBARRIER; |
| 147 | rq->rq_disk = q->bar_rq.rq_disk; |
| 148 | rq->end_io = end_io; |
| 149 | q->prepare_flush_fn(q, rq); |
| 150 | |
| 151 | elv_insert(q, rq, ELEVATOR_INSERT_FRONT); |
| 152 | } |
| 153 | |
| 154 | static inline bool start_ordered(struct request_queue *q, struct request **rqp) |
| 155 | { |
| 156 | struct request *rq = *rqp; |
| 157 | unsigned skip = 0; |
| 158 | |
| 159 | q->orderr = 0; |
| 160 | q->ordered = q->next_ordered; |
| 161 | q->ordseq |= QUEUE_ORDSEQ_STARTED; |
| 162 | |
| 163 | /* |
| 164 | * For an empty barrier, there's no actual BAR request, which |
| 165 | * in turn makes POSTFLUSH unnecessary. Mask them off. |
| 166 | */ |
| 167 | if (!blk_rq_sectors(rq)) { |
| 168 | q->ordered &= ~(QUEUE_ORDERED_DO_BAR | |
| 169 | QUEUE_ORDERED_DO_POSTFLUSH); |
| 170 | /* |
| 171 | * Empty barrier on a write-through device w/ ordered |
| 172 | * tag has no command to issue and without any command |
| 173 | * to issue, ordering by tag can't be used. Drain |
| 174 | * instead. |
| 175 | */ |
| 176 | if ((q->ordered & QUEUE_ORDERED_BY_TAG) && |
| 177 | !(q->ordered & QUEUE_ORDERED_DO_PREFLUSH)) { |
| 178 | q->ordered &= ~QUEUE_ORDERED_BY_TAG; |
| 179 | q->ordered |= QUEUE_ORDERED_BY_DRAIN; |
| 180 | } |
| 181 | } |
| 182 | |
| 183 | /* stash away the original request */ |
| 184 | blk_dequeue_request(rq); |
| 185 | q->orig_bar_rq = rq; |
| 186 | rq = NULL; |
| 187 | |
| 188 | /* |
| 189 | * Queue ordered sequence. As we stack them at the head, we |
| 190 | * need to queue in reverse order. Note that we rely on that |
| 191 | * no fs request uses ELEVATOR_INSERT_FRONT and thus no fs |
| 192 | * request gets inbetween ordered sequence. |
| 193 | */ |
| 194 | if (q->ordered & QUEUE_ORDERED_DO_POSTFLUSH) { |
| 195 | queue_flush(q, QUEUE_ORDERED_DO_POSTFLUSH); |
| 196 | rq = &q->post_flush_rq; |
| 197 | } else |
| 198 | skip |= QUEUE_ORDSEQ_POSTFLUSH; |
| 199 | |
| 200 | if (q->ordered & QUEUE_ORDERED_DO_BAR) { |
| 201 | rq = &q->bar_rq; |
| 202 | |
| 203 | /* initialize proxy request and queue it */ |
| 204 | blk_rq_init(q, rq); |
| 205 | if (bio_data_dir(q->orig_bar_rq->bio) == WRITE) |
| 206 | rq->cmd_flags |= REQ_RW; |
| 207 | if (q->ordered & QUEUE_ORDERED_DO_FUA) |
| 208 | rq->cmd_flags |= REQ_FUA; |
| 209 | init_request_from_bio(rq, q->orig_bar_rq->bio); |
| 210 | rq->end_io = bar_end_io; |
| 211 | |
| 212 | elv_insert(q, rq, ELEVATOR_INSERT_FRONT); |
| 213 | } else |
| 214 | skip |= QUEUE_ORDSEQ_BAR; |
| 215 | |
| 216 | if (q->ordered & QUEUE_ORDERED_DO_PREFLUSH) { |
| 217 | queue_flush(q, QUEUE_ORDERED_DO_PREFLUSH); |
| 218 | rq = &q->pre_flush_rq; |
| 219 | } else |
| 220 | skip |= QUEUE_ORDSEQ_PREFLUSH; |
| 221 | |
| 222 | if ((q->ordered & QUEUE_ORDERED_BY_DRAIN) && queue_in_flight(q)) |
| 223 | rq = NULL; |
| 224 | else |
| 225 | skip |= QUEUE_ORDSEQ_DRAIN; |
| 226 | |
| 227 | *rqp = rq; |
| 228 | |
| 229 | /* |
| 230 | * Complete skipped sequences. If whole sequence is complete, |
| 231 | * return false to tell elevator that this request is gone. |
| 232 | */ |
| 233 | return !blk_ordered_complete_seq(q, skip, 0); |
| 234 | } |
| 235 | |
| 236 | bool blk_do_ordered(struct request_queue *q, struct request **rqp) |
| 237 | { |
| 238 | struct request *rq = *rqp; |
| 239 | const int is_barrier = blk_fs_request(rq) && blk_barrier_rq(rq); |
| 240 | |
| 241 | if (!q->ordseq) { |
| 242 | if (!is_barrier) |
| 243 | return true; |
| 244 | |
| 245 | if (q->next_ordered != QUEUE_ORDERED_NONE) |
| 246 | return start_ordered(q, rqp); |
| 247 | else { |
| 248 | /* |
| 249 | * Queue ordering not supported. Terminate |
| 250 | * with prejudice. |
| 251 | */ |
| 252 | blk_dequeue_request(rq); |
| 253 | __blk_end_request_all(rq, -EOPNOTSUPP); |
| 254 | *rqp = NULL; |
| 255 | return false; |
| 256 | } |
| 257 | } |
| 258 | |
| 259 | /* |
| 260 | * Ordered sequence in progress |
| 261 | */ |
| 262 | |
| 263 | /* Special requests are not subject to ordering rules. */ |
| 264 | if (!blk_fs_request(rq) && |
| 265 | rq != &q->pre_flush_rq && rq != &q->post_flush_rq) |
| 266 | return true; |
| 267 | |
| 268 | if (q->ordered & QUEUE_ORDERED_BY_TAG) { |
| 269 | /* Ordered by tag. Blocking the next barrier is enough. */ |
| 270 | if (is_barrier && rq != &q->bar_rq) |
| 271 | *rqp = NULL; |
| 272 | } else { |
| 273 | /* Ordered by draining. Wait for turn. */ |
| 274 | WARN_ON(blk_ordered_req_seq(rq) < blk_ordered_cur_seq(q)); |
| 275 | if (blk_ordered_req_seq(rq) > blk_ordered_cur_seq(q)) |
| 276 | *rqp = NULL; |
| 277 | } |
| 278 | |
| 279 | return true; |
| 280 | } |
| 281 | |
| 282 | static void bio_end_empty_barrier(struct bio *bio, int err) |
| 283 | { |
| 284 | if (err) { |
| 285 | if (err == -EOPNOTSUPP) |
| 286 | set_bit(BIO_EOPNOTSUPP, &bio->bi_flags); |
| 287 | clear_bit(BIO_UPTODATE, &bio->bi_flags); |
| 288 | } |
| 289 | |
| 290 | complete(bio->bi_private); |
| 291 | } |
| 292 | |
| 293 | /** |
| 294 | * blkdev_issue_flush - queue a flush |
| 295 | * @bdev: blockdev to issue flush for |
| 296 | * @error_sector: error sector |
| 297 | * |
| 298 | * Description: |
| 299 | * Issue a flush for the block device in question. Caller can supply |
| 300 | * room for storing the error offset in case of a flush error, if they |
| 301 | * wish to. |
| 302 | */ |
| 303 | int blkdev_issue_flush(struct block_device *bdev, sector_t *error_sector) |
| 304 | { |
| 305 | DECLARE_COMPLETION_ONSTACK(wait); |
| 306 | struct request_queue *q; |
| 307 | struct bio *bio; |
| 308 | int ret; |
| 309 | |
| 310 | if (bdev->bd_disk == NULL) |
| 311 | return -ENXIO; |
| 312 | |
| 313 | q = bdev_get_queue(bdev); |
| 314 | if (!q) |
| 315 | return -ENXIO; |
| 316 | |
| 317 | bio = bio_alloc(GFP_KERNEL, 0); |
| 318 | bio->bi_end_io = bio_end_empty_barrier; |
| 319 | bio->bi_private = &wait; |
| 320 | bio->bi_bdev = bdev; |
| 321 | submit_bio(WRITE_BARRIER, bio); |
| 322 | |
| 323 | wait_for_completion(&wait); |
| 324 | |
| 325 | /* |
| 326 | * The driver must store the error location in ->bi_sector, if |
| 327 | * it supports it. For non-stacked drivers, this should be copied |
| 328 | * from blk_rq_pos(rq). |
| 329 | */ |
| 330 | if (error_sector) |
| 331 | *error_sector = bio->bi_sector; |
| 332 | |
| 333 | ret = 0; |
| 334 | if (bio_flagged(bio, BIO_EOPNOTSUPP)) |
| 335 | ret = -EOPNOTSUPP; |
| 336 | else if (!bio_flagged(bio, BIO_UPTODATE)) |
| 337 | ret = -EIO; |
| 338 | |
| 339 | bio_put(bio); |
| 340 | return ret; |
| 341 | } |
| 342 | EXPORT_SYMBOL(blkdev_issue_flush); |
| 343 | |
| 344 | static void blkdev_discard_end_io(struct bio *bio, int err) |
| 345 | { |
| 346 | if (err) { |
| 347 | if (err == -EOPNOTSUPP) |
| 348 | set_bit(BIO_EOPNOTSUPP, &bio->bi_flags); |
| 349 | clear_bit(BIO_UPTODATE, &bio->bi_flags); |
| 350 | } |
| 351 | |
| 352 | if (bio->bi_private) |
| 353 | complete(bio->bi_private); |
| 354 | __free_page(bio_page(bio)); |
| 355 | |
| 356 | bio_put(bio); |
| 357 | } |
| 358 | |
| 359 | /** |
| 360 | * blkdev_issue_discard - queue a discard |
| 361 | * @bdev: blockdev to issue discard for |
| 362 | * @sector: start sector |
| 363 | * @nr_sects: number of sectors to discard |
| 364 | * @gfp_mask: memory allocation flags (for bio_alloc) |
| 365 | * @flags: DISCARD_FL_* flags to control behaviour |
| 366 | * |
| 367 | * Description: |
| 368 | * Issue a discard request for the sectors in question. |
| 369 | */ |
| 370 | int blkdev_issue_discard(struct block_device *bdev, sector_t sector, |
| 371 | sector_t nr_sects, gfp_t gfp_mask, int flags) |
| 372 | { |
| 373 | DECLARE_COMPLETION_ONSTACK(wait); |
| 374 | struct request_queue *q = bdev_get_queue(bdev); |
| 375 | int type = flags & DISCARD_FL_BARRIER ? |
| 376 | DISCARD_BARRIER : DISCARD_NOBARRIER; |
| 377 | struct bio *bio; |
| 378 | struct page *page; |
| 379 | int ret = 0; |
| 380 | |
| 381 | if (!q) |
| 382 | return -ENXIO; |
| 383 | |
| 384 | if (!blk_queue_discard(q)) |
| 385 | return -EOPNOTSUPP; |
| 386 | |
| 387 | while (nr_sects && !ret) { |
| 388 | unsigned int sector_size = q->limits.logical_block_size; |
| 389 | unsigned int max_discard_sectors = |
| 390 | min(q->limits.max_discard_sectors, UINT_MAX >> 9); |
| 391 | |
| 392 | bio = bio_alloc(gfp_mask, 1); |
| 393 | if (!bio) |
| 394 | goto out; |
| 395 | bio->bi_sector = sector; |
| 396 | bio->bi_end_io = blkdev_discard_end_io; |
| 397 | bio->bi_bdev = bdev; |
| 398 | if (flags & DISCARD_FL_WAIT) |
| 399 | bio->bi_private = &wait; |
| 400 | |
| 401 | /* |
| 402 | * Add a zeroed one-sector payload as that's what |
| 403 | * our current implementations need. If we'll ever need |
| 404 | * more the interface will need revisiting. |
| 405 | */ |
| 406 | page = alloc_page(gfp_mask | __GFP_ZERO); |
| 407 | if (!page) |
| 408 | goto out_free_bio; |
| 409 | if (bio_add_pc_page(q, bio, page, sector_size, 0) < sector_size) |
| 410 | goto out_free_page; |
| 411 | |
| 412 | /* |
| 413 | * And override the bio size - the way discard works we |
| 414 | * touch many more blocks on disk than the actual payload |
| 415 | * length. |
| 416 | */ |
| 417 | if (nr_sects > max_discard_sectors) { |
| 418 | bio->bi_size = max_discard_sectors << 9; |
| 419 | nr_sects -= max_discard_sectors; |
| 420 | sector += max_discard_sectors; |
| 421 | } else { |
| 422 | bio->bi_size = nr_sects << 9; |
| 423 | nr_sects = 0; |
| 424 | } |
| 425 | |
| 426 | bio_get(bio); |
| 427 | submit_bio(type, bio); |
| 428 | |
| 429 | if (flags & DISCARD_FL_WAIT) |
| 430 | wait_for_completion(&wait); |
| 431 | |
| 432 | if (bio_flagged(bio, BIO_EOPNOTSUPP)) |
| 433 | ret = -EOPNOTSUPP; |
| 434 | else if (!bio_flagged(bio, BIO_UPTODATE)) |
| 435 | ret = -EIO; |
| 436 | bio_put(bio); |
| 437 | } |
| 438 | return ret; |
| 439 | out_free_page: |
| 440 | __free_page(page); |
| 441 | out_free_bio: |
| 442 | bio_put(bio); |
| 443 | out: |
| 444 | return -ENOMEM; |
| 445 | } |
| 446 | EXPORT_SYMBOL(blkdev_issue_discard); |