block/blk-tag.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Functions related to tagged command queuing
   4  */
   5 #include <linux/kernel.h>
   6 #include <linux/module.h>
   7 #include <linux/bio.h>
   8 #include <linux/blkdev.h>
   9 #include <linux/slab.h>
  10
  11 #include "blk.h"
  12
  13 /**
  14  * blk_queue_find_tag - find a request by its tag and queue
  15  * @q:   The request queue for the device
  16  * @tag: The tag of the request
  17  *
  18  * Notes:
  19  *    Should be used when a device returns a tag and you want to match
  20  *    it with a request.
  21  *
  22  *    no locks need be held.
  23  **/
  24 struct request *blk_queue_find_tag(struct request_queue *q, int tag)
  25 {
  26         return blk_map_queue_find_tag(q->queue_tags, tag);
  27 }
  28 EXPORT_SYMBOL(blk_queue_find_tag);
  29
  30 /**
  31  * blk_free_tags - release a given set of tag maintenance info
  32  * @bqt:        the tag map to free
  33  *
  34  * Drop the reference count on @bqt and frees it when the last reference
  35  * is dropped.
  36  */
  37 void blk_free_tags(struct blk_queue_tag *bqt)
  38 {
  39         if (atomic_dec_and_test(&bqt->refcnt)) {
  40                 BUG_ON(find_first_bit(bqt->tag_map, bqt->max_depth) <
  41                                                         bqt->max_depth);
  42
  43                 kfree(bqt->tag_index);
  44                 bqt->tag_index = NULL;
  45
  46                 kfree(bqt->tag_map);
  47                 bqt->tag_map = NULL;
  48
  49                 kfree(bqt);
  50         }
  51 }
  52 EXPORT_SYMBOL(blk_free_tags);
  53
  54 /**
  55  * __blk_queue_free_tags - release tag maintenance info
  56  * @q:  the request queue for the device
  57  *
  58  *  Notes:
  59  *    blk_cleanup_queue() will take care of calling this function, if tagging
  60  *    has been used. So there's no need to call this directly.
  61  **/
  62 void __blk_queue_free_tags(struct request_queue *q)
  63 {
  64         struct blk_queue_tag *bqt = q->queue_tags;
  65
  66         if (!bqt)
  67                 return;
  68
  69         blk_free_tags(bqt);
  70
  71         q->queue_tags = NULL;
  72         queue_flag_clear_unlocked(QUEUE_FLAG_QUEUED, q);
  73 }
  74
  75 /**
  76  * blk_queue_free_tags - release tag maintenance info
  77  * @q:  the request queue for the device
  78  *
  79  *  Notes:
  80  *      This is used to disable tagged queuing to a device, yet leave
  81  *      queue in function.
  82  **/
  83 void blk_queue_free_tags(struct request_queue *q)
  84 {
  85         queue_flag_clear_unlocked(QUEUE_FLAG_QUEUED, q);
  86 }
  87 EXPORT_SYMBOL(blk_queue_free_tags);
  88
  89 static int
  90 init_tag_map(struct request_queue *q, struct blk_queue_tag *tags, int depth)
  91 {
  92         struct request **tag_index;
  93         unsigned long *tag_map;
  94         int nr_ulongs;
  95
  96         if (q && depth > q->nr_requests * 2) {
  97                 depth = q->nr_requests * 2;
  98                 printk(KERN_ERR "%s: adjusted depth to %d\n",
  99                        __func__, depth);
 100         }
 101
 102         tag_index = kzalloc(depth * sizeof(struct request *), GFP_ATOMIC);
 103         if (!tag_index)
 104                 goto fail;
 105
 106         nr_ulongs = ALIGN(depth, BITS_PER_LONG) / BITS_PER_LONG;
 107         tag_map = kzalloc(nr_ulongs * sizeof(unsigned long), GFP_ATOMIC);
 108         if (!tag_map)
 109                 goto fail;
 110
 111         tags->real_max_depth = depth;
 112         tags->max_depth = depth;
 113         tags->tag_index = tag_index;
 114         tags->tag_map = tag_map;
 115
 116         return 0;
 117 fail:
 118         kfree(tag_index);
 119         return -ENOMEM;
 120 }
 121
 122 static struct blk_queue_tag *__blk_queue_init_tags(struct request_queue *q,
 123                                                 int depth, int alloc_policy)
 124 {
 125         struct blk_queue_tag *tags;
 126
 127         tags = kmalloc(sizeof(struct blk_queue_tag), GFP_ATOMIC);
 128         if (!tags)
 129                 goto fail;
 130
 131         if (init_tag_map(q, tags, depth))
 132                 goto fail;
 133
 134         atomic_set(&tags->refcnt, 1);
 135         tags->alloc_policy = alloc_policy;
 136         tags->next_tag = 0;
 137         return tags;
 138 fail:
 139         kfree(tags);
 140         return NULL;
 141 }
 142
 143 /**
 144  * blk_init_tags - initialize the tag info for an external tag map
 145  * @depth:      the maximum queue depth supported
 146  * @alloc_policy: tag allocation policy
 147  **/
 148 struct blk_queue_tag *blk_init_tags(int depth, int alloc_policy)
 149 {
 150         return __blk_queue_init_tags(NULL, depth, alloc_policy);
 151 }
 152 EXPORT_SYMBOL(blk_init_tags);
 153
 154 /**
 155  * blk_queue_init_tags - initialize the queue tag info
 156  * @q:  the request queue for the device
 157  * @depth:  the maximum queue depth supported
 158  * @tags: the tag to use
 159  * @alloc_policy: tag allocation policy
 160  *
 161  * Queue lock must be held here if the function is called to resize an
 162  * existing map.
 163  **/
 164 int blk_queue_init_tags(struct request_queue *q, int depth,
 165                         struct blk_queue_tag *tags, int alloc_policy)
 166 {
 167         int rc;
 168
 169         BUG_ON(tags && q->queue_tags && tags != q->queue_tags);
 170
 171         if (!tags && !q->queue_tags) {
 172                 tags = __blk_queue_init_tags(q, depth, alloc_policy);
 173
 174                 if (!tags)
 175                         return -ENOMEM;
 176
 177         } else if (q->queue_tags) {
 178                 rc = blk_queue_resize_tags(q, depth);
 179                 if (rc)
 180                         return rc;
 181                 queue_flag_set(QUEUE_FLAG_QUEUED, q);
 182                 return 0;
 183         } else
 184                 atomic_inc(&tags->refcnt);
 185
 186         /*
 187          * assign it, all done
 188          */
 189         q->queue_tags = tags;
 190         queue_flag_set_unlocked(QUEUE_FLAG_QUEUED, q);
 191         INIT_LIST_HEAD(&q->tag_busy_list);
 192         return 0;
 193 }
 194 EXPORT_SYMBOL(blk_queue_init_tags);
 195
 196 /**
 197  * blk_queue_resize_tags - change the queueing depth
 198  * @q:  the request queue for the device
 199  * @new_depth: the new max command queueing depth
 200  *
 201  *  Notes:
 202  *    Must be called with the queue lock held.
 203  **/
 204 int blk_queue_resize_tags(struct request_queue *q, int new_depth)
 205 {
 206         struct blk_queue_tag *bqt = q->queue_tags;
 207         struct request **tag_index;
 208         unsigned long *tag_map;
 209         int max_depth, nr_ulongs;
 210
 211         if (!bqt)
 212                 return -ENXIO;
 213
 214         /*
 215          * if we already have large enough real_max_depth.  just
 216          * adjust max_depth.  *NOTE* as requests with tag value
 217          * between new_depth and real_max_depth can be in-flight, tag
 218          * map can not be shrunk blindly here.
 219          */
 220         if (new_depth <= bqt->real_max_depth) {
 221                 bqt->max_depth = new_depth;
 222                 return 0;
 223         }
 224
 225         /*
 226          * Currently cannot replace a shared tag map with a new
 227          * one, so error out if this is the case
 228          */
 229         if (atomic_read(&bqt->refcnt) != 1)
 230                 return -EBUSY;
 231
 232         /*
 233          * save the old state info, so we can copy it back
 234          */
 235         tag_index = bqt->tag_index;
 236         tag_map = bqt->tag_map;
 237         max_depth = bqt->real_max_depth;
 238
 239         if (init_tag_map(q, bqt, new_depth))
 240                 return -ENOMEM;
 241
 242         memcpy(bqt->tag_index, tag_index, max_depth * sizeof(struct request *));
 243         nr_ulongs = ALIGN(max_depth, BITS_PER_LONG) / BITS_PER_LONG;
 244         memcpy(bqt->tag_map, tag_map, nr_ulongs * sizeof(unsigned long));
 245
 246         kfree(tag_index);
 247         kfree(tag_map);
 248         return 0;
 249 }
 250 EXPORT_SYMBOL(blk_queue_resize_tags);
 251
 252 /**
 253  * blk_queue_end_tag - end tag operations for a request
 254  * @q:  the request queue for the device
 255  * @rq: the request that has completed
 256  *
 257  *  Description:
 258  *    Typically called when end_that_request_first() returns %0, meaning
 259  *    all transfers have been done for a request. It's important to call
 260  *    this function before end_that_request_last(), as that will put the
 261  *    request back on the free list thus corrupting the internal tag list.
 262  **/
 263 void blk_queue_end_tag(struct request_queue *q, struct request *rq)
 264 {
 265         struct blk_queue_tag *bqt = q->queue_tags;
 266         unsigned tag = rq->tag; /* negative tags invalid */
 267
 268         lockdep_assert_held(q->queue_lock);
 269
 270         BUG_ON(tag >= bqt->real_max_depth);
 271
 272         list_del_init(&rq->queuelist);
 273         rq->rq_flags &= ~RQF_QUEUED;
 274         rq->tag = -1;
 275         rq->internal_tag = -1;
 276
 277         if (unlikely(bqt->tag_index[tag] == NULL))
 278                 printk(KERN_ERR "%s: tag %d is missing\n",
 279                        __func__, tag);
 280
 281         bqt->tag_index[tag] = NULL;
 282
 283         if (unlikely(!test_bit(tag, bqt->tag_map))) {
 284                 printk(KERN_ERR "%s: attempt to clear non-busy tag (%d)\n",
 285                        __func__, tag);
 286                 return;
 287         }
 288         /*
 289          * The tag_map bit acts as a lock for tag_index[bit], so we need
 290          * unlock memory barrier semantics.
 291          */
 292         clear_bit_unlock(tag, bqt->tag_map);
 293 }
 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);