| 1 | /* |
| 2 | * Functions related to tagged command queuing |
| 3 | */ |
| 4 | #include <linux/kernel.h> |
| 5 | #include <linux/module.h> |
| 6 | #include <linux/bio.h> |
| 7 | #include <linux/blkdev.h> |
| 8 | |
| 9 | #include "blk.h" |
| 10 | |
| 11 | /** |
| 12 | * blk_queue_find_tag - find a request by its tag and queue |
| 13 | * @q: The request queue for the device |
| 14 | * @tag: The tag of the request |
| 15 | * |
| 16 | * Notes: |
| 17 | * Should be used when a device returns a tag and you want to match |
| 18 | * it with a request. |
| 19 | * |
| 20 | * no locks need be held. |
| 21 | **/ |
| 22 | struct request *blk_queue_find_tag(struct request_queue *q, int tag) |
| 23 | { |
| 24 | return blk_map_queue_find_tag(q->queue_tags, tag); |
| 25 | } |
| 26 | EXPORT_SYMBOL(blk_queue_find_tag); |
| 27 | |
| 28 | /** |
| 29 | * __blk_free_tags - release a given set of tag maintenance info |
| 30 | * @bqt: the tag map to free |
| 31 | * |
| 32 | * Tries to free the specified @bqt. Returns true if it was |
| 33 | * actually freed and false if there are still references using it |
| 34 | */ |
| 35 | static int __blk_free_tags(struct blk_queue_tag *bqt) |
| 36 | { |
| 37 | int retval; |
| 38 | |
| 39 | retval = atomic_dec_and_test(&bqt->refcnt); |
| 40 | if (retval) { |
| 41 | BUG_ON(find_first_bit(bqt->tag_map, bqt->max_depth) < |
| 42 | bqt->max_depth); |
| 43 | |
| 44 | kfree(bqt->tag_index); |
| 45 | bqt->tag_index = NULL; |
| 46 | |
| 47 | kfree(bqt->tag_map); |
| 48 | bqt->tag_map = NULL; |
| 49 | |
| 50 | kfree(bqt); |
| 51 | } |
| 52 | |
| 53 | return retval; |
| 54 | } |
| 55 | |
| 56 | /** |
| 57 | * __blk_queue_free_tags - release tag maintenance info |
| 58 | * @q: the request queue for the device |
| 59 | * |
| 60 | * Notes: |
| 61 | * blk_cleanup_queue() will take care of calling this function, if tagging |
| 62 | * has been used. So there's no need to call this directly. |
| 63 | **/ |
| 64 | void __blk_queue_free_tags(struct request_queue *q) |
| 65 | { |
| 66 | struct blk_queue_tag *bqt = q->queue_tags; |
| 67 | |
| 68 | if (!bqt) |
| 69 | return; |
| 70 | |
| 71 | __blk_free_tags(bqt); |
| 72 | |
| 73 | q->queue_tags = NULL; |
| 74 | queue_flag_clear_unlocked(QUEUE_FLAG_QUEUED, q); |
| 75 | } |
| 76 | |
| 77 | /** |
| 78 | * blk_free_tags - release a given set of tag maintenance info |
| 79 | * @bqt: the tag map to free |
| 80 | * |
| 81 | * For externally managed @bqt frees the map. Callers of this |
| 82 | * function must guarantee to have released all the queues that |
| 83 | * might have been using this tag map. |
| 84 | */ |
| 85 | void blk_free_tags(struct blk_queue_tag *bqt) |
| 86 | { |
| 87 | if (unlikely(!__blk_free_tags(bqt))) |
| 88 | BUG(); |
| 89 | } |
| 90 | EXPORT_SYMBOL(blk_free_tags); |
| 91 | |
| 92 | /** |
| 93 | * blk_queue_free_tags - release tag maintenance info |
| 94 | * @q: the request queue for the device |
| 95 | * |
| 96 | * Notes: |
| 97 | * This is used to disable tagged queuing to a device, yet leave |
| 98 | * queue in function. |
| 99 | **/ |
| 100 | void blk_queue_free_tags(struct request_queue *q) |
| 101 | { |
| 102 | queue_flag_clear_unlocked(QUEUE_FLAG_QUEUED, q); |
| 103 | } |
| 104 | EXPORT_SYMBOL(blk_queue_free_tags); |
| 105 | |
| 106 | static int |
| 107 | init_tag_map(struct request_queue *q, struct blk_queue_tag *tags, int depth) |
| 108 | { |
| 109 | struct request **tag_index; |
| 110 | unsigned long *tag_map; |
| 111 | int nr_ulongs; |
| 112 | |
| 113 | if (q && depth > q->nr_requests * 2) { |
| 114 | depth = q->nr_requests * 2; |
| 115 | printk(KERN_ERR "%s: adjusted depth to %d\n", |
| 116 | __func__, depth); |
| 117 | } |
| 118 | |
| 119 | tag_index = kzalloc(depth * sizeof(struct request *), GFP_ATOMIC); |
| 120 | if (!tag_index) |
| 121 | goto fail; |
| 122 | |
| 123 | nr_ulongs = ALIGN(depth, BITS_PER_LONG) / BITS_PER_LONG; |
| 124 | tag_map = kzalloc(nr_ulongs * sizeof(unsigned long), GFP_ATOMIC); |
| 125 | if (!tag_map) |
| 126 | goto fail; |
| 127 | |
| 128 | tags->real_max_depth = depth; |
| 129 | tags->max_depth = depth; |
| 130 | tags->tag_index = tag_index; |
| 131 | tags->tag_map = tag_map; |
| 132 | |
| 133 | return 0; |
| 134 | fail: |
| 135 | kfree(tag_index); |
| 136 | return -ENOMEM; |
| 137 | } |
| 138 | |
| 139 | static struct blk_queue_tag *__blk_queue_init_tags(struct request_queue *q, |
| 140 | int depth) |
| 141 | { |
| 142 | struct blk_queue_tag *tags; |
| 143 | |
| 144 | tags = kmalloc(sizeof(struct blk_queue_tag), GFP_ATOMIC); |
| 145 | if (!tags) |
| 146 | goto fail; |
| 147 | |
| 148 | if (init_tag_map(q, tags, depth)) |
| 149 | goto fail; |
| 150 | |
| 151 | atomic_set(&tags->refcnt, 1); |
| 152 | return tags; |
| 153 | fail: |
| 154 | kfree(tags); |
| 155 | return NULL; |
| 156 | } |
| 157 | |
| 158 | /** |
| 159 | * blk_init_tags - initialize the tag info for an external tag map |
| 160 | * @depth: the maximum queue depth supported |
| 161 | **/ |
| 162 | struct blk_queue_tag *blk_init_tags(int depth) |
| 163 | { |
| 164 | return __blk_queue_init_tags(NULL, depth); |
| 165 | } |
| 166 | EXPORT_SYMBOL(blk_init_tags); |
| 167 | |
| 168 | /** |
| 169 | * blk_queue_init_tags - initialize the queue tag info |
| 170 | * @q: the request queue for the device |
| 171 | * @depth: the maximum queue depth supported |
| 172 | * @tags: the tag to use |
| 173 | * |
| 174 | * Queue lock must be held here if the function is called to resize an |
| 175 | * existing map. |
| 176 | **/ |
| 177 | int blk_queue_init_tags(struct request_queue *q, int depth, |
| 178 | struct blk_queue_tag *tags) |
| 179 | { |
| 180 | int rc; |
| 181 | |
| 182 | BUG_ON(tags && q->queue_tags && tags != q->queue_tags); |
| 183 | |
| 184 | if (!tags && !q->queue_tags) { |
| 185 | tags = __blk_queue_init_tags(q, depth); |
| 186 | |
| 187 | if (!tags) |
| 188 | goto fail; |
| 189 | } else if (q->queue_tags) { |
| 190 | rc = blk_queue_resize_tags(q, depth); |
| 191 | if (rc) |
| 192 | return rc; |
| 193 | queue_flag_set(QUEUE_FLAG_QUEUED, q); |
| 194 | return 0; |
| 195 | } else |
| 196 | atomic_inc(&tags->refcnt); |
| 197 | |
| 198 | /* |
| 199 | * assign it, all done |
| 200 | */ |
| 201 | q->queue_tags = tags; |
| 202 | queue_flag_set_unlocked(QUEUE_FLAG_QUEUED, q); |
| 203 | INIT_LIST_HEAD(&q->tag_busy_list); |
| 204 | return 0; |
| 205 | fail: |
| 206 | kfree(tags); |
| 207 | return -ENOMEM; |
| 208 | } |
| 209 | EXPORT_SYMBOL(blk_queue_init_tags); |
| 210 | |
| 211 | /** |
| 212 | * blk_queue_resize_tags - change the queueing depth |
| 213 | * @q: the request queue for the device |
| 214 | * @new_depth: the new max command queueing depth |
| 215 | * |
| 216 | * Notes: |
| 217 | * Must be called with the queue lock held. |
| 218 | **/ |
| 219 | int blk_queue_resize_tags(struct request_queue *q, int new_depth) |
| 220 | { |
| 221 | struct blk_queue_tag *bqt = q->queue_tags; |
| 222 | struct request **tag_index; |
| 223 | unsigned long *tag_map; |
| 224 | int max_depth, nr_ulongs; |
| 225 | |
| 226 | if (!bqt) |
| 227 | return -ENXIO; |
| 228 | |
| 229 | /* |
| 230 | * if we already have large enough real_max_depth. just |
| 231 | * adjust max_depth. *NOTE* as requests with tag value |
| 232 | * between new_depth and real_max_depth can be in-flight, tag |
| 233 | * map can not be shrunk blindly here. |
| 234 | */ |
| 235 | if (new_depth <= bqt->real_max_depth) { |
| 236 | bqt->max_depth = new_depth; |
| 237 | return 0; |
| 238 | } |
| 239 | |
| 240 | /* |
| 241 | * Currently cannot replace a shared tag map with a new |
| 242 | * one, so error out if this is the case |
| 243 | */ |
| 244 | if (atomic_read(&bqt->refcnt) != 1) |
| 245 | return -EBUSY; |
| 246 | |
| 247 | /* |
| 248 | * save the old state info, so we can copy it back |
| 249 | */ |
| 250 | tag_index = bqt->tag_index; |
| 251 | tag_map = bqt->tag_map; |
| 252 | max_depth = bqt->real_max_depth; |
| 253 | |
| 254 | if (init_tag_map(q, bqt, new_depth)) |
| 255 | return -ENOMEM; |
| 256 | |
| 257 | memcpy(bqt->tag_index, tag_index, max_depth * sizeof(struct request *)); |
| 258 | nr_ulongs = ALIGN(max_depth, BITS_PER_LONG) / BITS_PER_LONG; |
| 259 | memcpy(bqt->tag_map, tag_map, nr_ulongs * sizeof(unsigned long)); |
| 260 | |
| 261 | kfree(tag_index); |
| 262 | kfree(tag_map); |
| 263 | return 0; |
| 264 | } |
| 265 | EXPORT_SYMBOL(blk_queue_resize_tags); |
| 266 | |
| 267 | /** |
| 268 | * blk_queue_end_tag - end tag operations for a request |
| 269 | * @q: the request queue for the device |
| 270 | * @rq: the request that has completed |
| 271 | * |
| 272 | * Description: |
| 273 | * Typically called when end_that_request_first() returns %0, meaning |
| 274 | * all transfers have been done for a request. It's important to call |
| 275 | * this function before end_that_request_last(), as that will put the |
| 276 | * request back on the free list thus corrupting the internal tag list. |
| 277 | * |
| 278 | * Notes: |
| 279 | * queue lock must be held. |
| 280 | **/ |
| 281 | void blk_queue_end_tag(struct request_queue *q, struct request *rq) |
| 282 | { |
| 283 | struct blk_queue_tag *bqt = q->queue_tags; |
| 284 | int tag = rq->tag; |
| 285 | |
| 286 | BUG_ON(tag == -1); |
| 287 | |
| 288 | if (unlikely(tag >= bqt->real_max_depth)) |
| 289 | /* |
| 290 | * This can happen after tag depth has been reduced. |
| 291 | * FIXME: how about a warning or info message here? |
| 292 | */ |
| 293 | return; |
| 294 | |
| 295 | list_del_init(&rq->queuelist); |
| 296 | rq->cmd_flags &= ~REQ_QUEUED; |
| 297 | rq->tag = -1; |
| 298 | |
| 299 | if (unlikely(bqt->tag_index[tag] == NULL)) |
| 300 | printk(KERN_ERR "%s: tag %d is missing\n", |
| 301 | __func__, tag); |
| 302 | |
| 303 | bqt->tag_index[tag] = NULL; |
| 304 | |
| 305 | if (unlikely(!test_bit(tag, bqt->tag_map))) { |
| 306 | printk(KERN_ERR "%s: attempt to clear non-busy tag (%d)\n", |
| 307 | __func__, tag); |
| 308 | return; |
| 309 | } |
| 310 | /* |
| 311 | * The tag_map bit acts as a lock for tag_index[bit], so we need |
| 312 | * unlock memory barrier semantics. |
| 313 | */ |
| 314 | clear_bit_unlock(tag, bqt->tag_map); |
| 315 | } |
| 316 | EXPORT_SYMBOL(blk_queue_end_tag); |
| 317 | |
| 318 | /** |
| 319 | * blk_queue_start_tag - find a free tag and assign it |
| 320 | * @q: the request queue for the device |
| 321 | * @rq: the block request that needs tagging |
| 322 | * |
| 323 | * Description: |
| 324 | * This can either be used as a stand-alone helper, or possibly be |
| 325 | * assigned as the queue &prep_rq_fn (in which case &struct request |
| 326 | * automagically gets a tag assigned). Note that this function |
| 327 | * assumes that any type of request can be queued! if this is not |
| 328 | * true for your device, you must check the request type before |
| 329 | * calling this function. The request will also be removed from |
| 330 | * the request queue, so it's the drivers responsibility to readd |
| 331 | * it if it should need to be restarted for some reason. |
| 332 | * |
| 333 | * Notes: |
| 334 | * queue lock must be held. |
| 335 | **/ |
| 336 | int blk_queue_start_tag(struct request_queue *q, struct request *rq) |
| 337 | { |
| 338 | struct blk_queue_tag *bqt = q->queue_tags; |
| 339 | unsigned max_depth; |
| 340 | int tag; |
| 341 | |
| 342 | if (unlikely((rq->cmd_flags & REQ_QUEUED))) { |
| 343 | printk(KERN_ERR |
| 344 | "%s: request %p for device [%s] already tagged %d", |
| 345 | __func__, rq, |
| 346 | rq->rq_disk ? rq->rq_disk->disk_name : "?", rq->tag); |
| 347 | BUG(); |
| 348 | } |
| 349 | |
| 350 | /* |
| 351 | * Protect against shared tag maps, as we may not have exclusive |
| 352 | * access to the tag map. |
| 353 | * |
| 354 | * We reserve a few tags just for sync IO, since we don't want |
| 355 | * to starve sync IO on behalf of flooding async IO. |
| 356 | */ |
| 357 | max_depth = bqt->max_depth; |
| 358 | if (!rq_is_sync(rq) && max_depth > 1) { |
| 359 | max_depth -= 2; |
| 360 | if (!max_depth) |
| 361 | max_depth = 1; |
| 362 | if (q->in_flight[BLK_RW_ASYNC] > max_depth) |
| 363 | return 1; |
| 364 | } |
| 365 | |
| 366 | do { |
| 367 | tag = find_first_zero_bit(bqt->tag_map, max_depth); |
| 368 | if (tag >= max_depth) |
| 369 | return 1; |
| 370 | |
| 371 | } while (test_and_set_bit_lock(tag, bqt->tag_map)); |
| 372 | /* |
| 373 | * We need lock ordering semantics given by test_and_set_bit_lock. |
| 374 | * See blk_queue_end_tag for details. |
| 375 | */ |
| 376 | |
| 377 | rq->cmd_flags |= REQ_QUEUED; |
| 378 | rq->tag = tag; |
| 379 | bqt->tag_index[tag] = rq; |
| 380 | blk_start_request(rq); |
| 381 | list_add(&rq->queuelist, &q->tag_busy_list); |
| 382 | return 0; |
| 383 | } |
| 384 | EXPORT_SYMBOL(blk_queue_start_tag); |
| 385 | |
| 386 | /** |
| 387 | * blk_queue_invalidate_tags - invalidate all pending tags |
| 388 | * @q: the request queue for the device |
| 389 | * |
| 390 | * Description: |
| 391 | * Hardware conditions may dictate a need to stop all pending requests. |
| 392 | * In this case, we will safely clear the block side of the tag queue and |
| 393 | * readd all requests to the request queue in the right order. |
| 394 | * |
| 395 | * Notes: |
| 396 | * queue lock must be held. |
| 397 | **/ |
| 398 | void blk_queue_invalidate_tags(struct request_queue *q) |
| 399 | { |
| 400 | struct list_head *tmp, *n; |
| 401 | |
| 402 | list_for_each_safe(tmp, n, &q->tag_busy_list) |
| 403 | blk_requeue_request(q, list_entry_rq(tmp)); |
| 404 | } |
| 405 | EXPORT_SYMBOL(blk_queue_invalidate_tags); |