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
   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  * @tags: the tag to use
 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, offset;
 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))
 360                 offset = 0;
 361         else
 362                 offset = max_depth >> 2;
 363
 364         do {
 365                 tag = find_next_zero_bit(bqt->tag_map, max_depth, offset);
 366                 if (tag >= max_depth)
 367                         return 1;
 368
 369         } while (test_and_set_bit_lock(tag, bqt->tag_map));
 370         /*
 371          * We need lock ordering semantics given by test_and_set_bit_lock.
 372          * See blk_queue_end_tag for details.
 373          */
 374
 375         rq->cmd_flags |= REQ_QUEUED;
 376         rq->tag = tag;
 377         bqt->tag_index[tag] = rq;
 378         blkdev_dequeue_request(rq);
 379         list_add(&rq->queuelist, &q->tag_busy_list);
 380         return 0;
 381 }
 382 EXPORT_SYMBOL(blk_queue_start_tag);
 383
 384 /**
 385  * blk_queue_invalidate_tags - invalidate all pending tags
 386  * @q:  the request queue for the device
 387  *
 388  *  Description:
 389  *   Hardware conditions may dictate a need to stop all pending requests.
 390  *   In this case, we will safely clear the block side of the tag queue and
 391  *   readd all requests to the request queue in the right order.
 392  *
 393  *  Notes:
 394  *   queue lock must be held.
 395  **/
 396 void blk_queue_invalidate_tags(struct request_queue *q)
 397 {
 398         struct list_head *tmp, *n;
 399
 400         list_for_each_safe(tmp, n, &q->tag_busy_list)
 401                 blk_requeue_request(q, list_entry_rq(tmp));
 402 }
 403 EXPORT_SYMBOL(blk_queue_invalidate_tags);