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