block/blk-tag.c

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