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                         return -ENOMEM;
 190
 191         } else if (q->queue_tags) {
 192                 rc = blk_queue_resize_tags(q, depth);
 193                 if (rc)
 194                         return rc;
 195                 queue_flag_set(QUEUE_FLAG_QUEUED, q);
 196                 return 0;
 197         } else
 198                 atomic_inc(&tags->refcnt);
 199
 200         /*
 201          * assign it, all done
 202          */
 203         q->queue_tags = tags;
 204         queue_flag_set_unlocked(QUEUE_FLAG_QUEUED, q);
 205         INIT_LIST_HEAD(&q->tag_busy_list);
 206         return 0;
 207 }
 208 EXPORT_SYMBOL(blk_queue_init_tags);
 209
 210 /**
 211  * blk_queue_resize_tags - change the queueing depth
 212  * @q:  the request queue for the device
 213  * @new_depth: the new max command queueing depth
 214  *
 215  *  Notes:
 216  *    Must be called with the queue lock held.
 217  **/
 218 int blk_queue_resize_tags(struct request_queue *q, int new_depth)
 219 {
 220         struct blk_queue_tag *bqt = q->queue_tags;
 221         struct request **tag_index;
 222         unsigned long *tag_map;
 223         int max_depth, nr_ulongs;
 224
 225         if (!bqt)
 226                 return -ENXIO;
 227
 228         /*
 229          * if we already have large enough real_max_depth.  just
 230          * adjust max_depth.  *NOTE* as requests with tag value
 231          * between new_depth and real_max_depth can be in-flight, tag
 232          * map can not be shrunk blindly here.
 233          */
 234         if (new_depth <= bqt->real_max_depth) {
 235                 bqt->max_depth = new_depth;
 236                 return 0;
 237         }
 238
 239         /*
 240          * Currently cannot replace a shared tag map with a new
 241          * one, so error out if this is the case
 242          */
 243         if (atomic_read(&bqt->refcnt) != 1)
 244                 return -EBUSY;
 245
 246         /*
 247          * save the old state info, so we can copy it back
 248          */
 249         tag_index = bqt->tag_index;
 250         tag_map = bqt->tag_map;
 251         max_depth = bqt->real_max_depth;
 252
 253         if (init_tag_map(q, bqt, new_depth))
 254                 return -ENOMEM;
 255
 256         memcpy(bqt->tag_index, tag_index, max_depth * sizeof(struct request *));
 257         nr_ulongs = ALIGN(max_depth, BITS_PER_LONG) / BITS_PER_LONG;
 258         memcpy(bqt->tag_map, tag_map, nr_ulongs * sizeof(unsigned long));
 259
 260         kfree(tag_index);
 261         kfree(tag_map);
 262         return 0;
 263 }
 264 EXPORT_SYMBOL(blk_queue_resize_tags);
 265
 266 /**
 267  * blk_queue_end_tag - end tag operations for a request
 268  * @q:  the request queue for the device
 269  * @rq: the request that has completed
 270  *
 271  *  Description:
 272  *    Typically called when end_that_request_first() returns %0, meaning
 273  *    all transfers have been done for a request. It's important to call
 274  *    this function before end_that_request_last(), as that will put the
 275  *    request back on the free list thus corrupting the internal tag list.
 276  *
 277  *  Notes:
 278  *   queue lock must be held.
 279  **/
 280 void blk_queue_end_tag(struct request_queue *q, struct request *rq)
 281 {
 282         struct blk_queue_tag *bqt = q->queue_tags;
 283         unsigned tag = rq->tag; /* negative tags invalid */
 284
 285         BUG_ON(tag >= bqt->real_max_depth);
 286
 287         list_del_init(&rq->queuelist);
 288         rq->cmd_flags &= ~REQ_QUEUED;
 289         rq->tag = -1;
 290
 291         if (unlikely(bqt->tag_index[tag] == NULL))
 292                 printk(KERN_ERR "%s: tag %d is missing\n",
 293                        __func__, tag);
 294
 295         bqt->tag_index[tag] = NULL;
 296
 297         if (unlikely(!test_bit(tag, bqt->tag_map))) {
 298                 printk(KERN_ERR "%s: attempt to clear non-busy tag (%d)\n",
 299                        __func__, tag);
 300                 return;
 301         }
 302         /*
 303          * The tag_map bit acts as a lock for tag_index[bit], so we need
 304          * unlock memory barrier semantics.
 305          */
 306         clear_bit_unlock(tag, bqt->tag_map);
 307 }
 308 EXPORT_SYMBOL(blk_queue_end_tag);
 309
 310 /**
 311  * blk_queue_start_tag - find a free tag and assign it
 312  * @q:  the request queue for the device
 313  * @rq:  the block request that needs tagging
 314  *
 315  *  Description:
 316  *    This can either be used as a stand-alone helper, or possibly be
 317  *    assigned as the queue &prep_rq_fn (in which case &struct request
 318  *    automagically gets a tag assigned). Note that this function
 319  *    assumes that any type of request can be queued! if this is not
 320  *    true for your device, you must check the request type before
 321  *    calling this function.  The request will also be removed from
 322  *    the request queue, so it's the drivers responsibility to readd
 323  *    it if it should need to be restarted for some reason.
 324  *
 325  *  Notes:
 326  *   queue lock must be held.
 327  **/
 328 int blk_queue_start_tag(struct request_queue *q, struct request *rq)
 329 {
 330         struct blk_queue_tag *bqt = q->queue_tags;
 331         unsigned max_depth;
 332         int tag;
 333
 334         if (unlikely((rq->cmd_flags & REQ_QUEUED))) {
 335                 printk(KERN_ERR
 336                        "%s: request %p for device [%s] already tagged %d",
 337                        __func__, rq,
 338                        rq->rq_disk ? rq->rq_disk->disk_name : "?", rq->tag);
 339                 BUG();
 340         }
 341
 342         /*
 343          * Protect against shared tag maps, as we may not have exclusive
 344          * access to the tag map.
 345          *
 346          * We reserve a few tags just for sync IO, since we don't want
 347          * to starve sync IO on behalf of flooding async IO.
 348          */
 349         max_depth = bqt->max_depth;
 350         if (!rq_is_sync(rq) && max_depth > 1) {
 351                 switch (max_depth) {
 352                 case 2:
 353                         max_depth = 1;
 354                         break;
 355                 case 3:
 356                         max_depth = 2;
 357                         break;
 358                 default:
 359                         max_depth -= 2;
 360                 }
 361                 if (q->in_flight[BLK_RW_ASYNC] > max_depth)
 362                         return 1;
 363         }
 364
 365         do {
 366                 tag = find_first_zero_bit(bqt->tag_map, max_depth);
 367                 if (tag >= max_depth)
 368                         return 1;
 369
 370         } while (test_and_set_bit_lock(tag, bqt->tag_map));
 371         /*
 372          * We need lock ordering semantics given by test_and_set_bit_lock.
 373          * See blk_queue_end_tag for details.
 374          */
 375
 376         rq->cmd_flags |= REQ_QUEUED;
 377         rq->tag = tag;
 378         bqt->tag_index[tag] = rq;
 379         blk_start_request(rq);
 380         list_add(&rq->queuelist, &q->tag_busy_list);
 381         return 0;
 382 }
 383 EXPORT_SYMBOL(blk_queue_start_tag);
 384
 385 /**
 386  * blk_queue_invalidate_tags - invalidate all pending tags
 387  * @q:  the request queue for the device
 388  *
 389  *  Description:
 390  *   Hardware conditions may dictate a need to stop all pending requests.
 391  *   In this case, we will safely clear the block side of the tag queue and
 392  *   readd all requests to the request queue in the right order.
 393  *
 394  *  Notes:
 395  *   queue lock must be held.
 396  **/
 397 void blk_queue_invalidate_tags(struct request_queue *q)
 398 {
 399         struct list_head *tmp, *n;
 400
 401         list_for_each_safe(tmp, n, &q->tag_busy_list)
 402                 blk_requeue_request(q, list_entry_rq(tmp));
 403 }
 404 EXPORT_SYMBOL(blk_queue_invalidate_tags);