drivers/scsi/scsi_lib.c

   1 /*
   2  *  scsi_lib.c Copyright (C) 1999 Eric Youngdale
   3  *
   4  *  SCSI queueing library.
   5  *      Initial versions: Eric Youngdale (eric@andante.org).
   6  *                        Based upon conversations with large numbers
   7  *                        of people at Linux Expo.
   8  */
   9
  10 #include <linux/bio.h>
  11 #include <linux/blkdev.h>
  12 #include <linux/completion.h>
  13 #include <linux/kernel.h>
  14 #include <linux/mempool.h>
  15 #include <linux/slab.h>
  16 #include <linux/init.h>
  17 #include <linux/pci.h>
  18 #include <linux/delay.h>
  19 #include <linux/hardirq.h>
  20
  21 #include <scsi/scsi.h>
  22 #include <scsi/scsi_cmnd.h>
  23 #include <scsi/scsi_dbg.h>
  24 #include <scsi/scsi_device.h>
  25 #include <scsi/scsi_driver.h>
  26 #include <scsi/scsi_eh.h>
  27 #include <scsi/scsi_host.h>
  28
  29 #include "scsi_priv.h"
  30 #include "scsi_logging.h"
  31
  32
  33 #define SG_MEMPOOL_NR           ARRAY_SIZE(scsi_sg_pools)
  34 #define SG_MEMPOOL_SIZE         2
  35
  36 struct scsi_host_sg_pool {
  37         size_t          size;
  38         char            *name;
  39         struct kmem_cache       *slab;
  40         mempool_t       *pool;
  41 };
  42
  43 #if (SCSI_MAX_PHYS_SEGMENTS < 32)
  44 #error SCSI_MAX_PHYS_SEGMENTS is too small
  45 #endif
  46
  47 #define SP(x) { x, "sgpool-" #x }
  48 static struct scsi_host_sg_pool scsi_sg_pools[] = {
  49         SP(8),
  50         SP(16),
  51         SP(32),
  52 #if (SCSI_MAX_PHYS_SEGMENTS > 32)
  53         SP(64),
  54 #if (SCSI_MAX_PHYS_SEGMENTS > 64)
  55         SP(128),
  56 #if (SCSI_MAX_PHYS_SEGMENTS > 128)
  57         SP(256),
  58 #if (SCSI_MAX_PHYS_SEGMENTS > 256)
  59 #error SCSI_MAX_PHYS_SEGMENTS is too large
  60 #endif
  61 #endif
  62 #endif
  63 #endif
  64 };
  65 #undef SP
  66
  67 static void scsi_run_queue(struct request_queue *q);
  68
  69 /*
  70  * Function:    scsi_unprep_request()
  71  *
  72  * Purpose:     Remove all preparation done for a request, including its
  73  *              associated scsi_cmnd, so that it can be requeued.
  74  *
  75  * Arguments:   req     - request to unprepare
  76  *
  77  * Lock status: Assumed that no locks are held upon entry.
  78  *
  79  * Returns:     Nothing.
  80  */
  81 static void scsi_unprep_request(struct request *req)
  82 {
  83         struct scsi_cmnd *cmd = req->special;
  84
  85         req->cmd_flags &= ~REQ_DONTPREP;
  86         req->special = NULL;
  87
  88         scsi_put_command(cmd);
  89 }
  90
  91 /*
  92  * Function:    scsi_queue_insert()
  93  *
  94  * Purpose:     Insert a command in the midlevel queue.
  95  *
  96  * Arguments:   cmd    - command that we are adding to queue.
  97  *              reason - why we are inserting command to queue.
  98  *
  99  * Lock status: Assumed that lock is not held upon entry.
 100  *
 101  * Returns:     Nothing.
 102  *
 103  * Notes:       We do this for one of two cases.  Either the host is busy
 104  *              and it cannot accept any more commands for the time being,
 105  *              or the device returned QUEUE_FULL and can accept no more
 106  *              commands.
 107  * Notes:       This could be called either from an interrupt context or a
 108  *              normal process context.
 109  */
 110 int scsi_queue_insert(struct scsi_cmnd *cmd, int reason)
 111 {
 112         struct Scsi_Host *host = cmd->device->host;
 113         struct scsi_device *device = cmd->device;
 114         struct request_queue *q = device->request_queue;
 115         unsigned long flags;
 116
 117         SCSI_LOG_MLQUEUE(1,
 118                  printk("Inserting command %p into mlqueue\n", cmd));
 119
 120         /*
 121          * Set the appropriate busy bit for the device/host.
 122          *
 123          * If the host/device isn't busy, assume that something actually
 124          * completed, and that we should be able to queue a command now.
 125          *
 126          * Note that the prior mid-layer assumption that any host could
 127          * always queue at least one command is now broken.  The mid-layer
 128          * will implement a user specifiable stall (see
 129          * scsi_host.max_host_blocked and scsi_device.max_device_blocked)
 130          * if a command is requeued with no other commands outstanding
 131          * either for the device or for the host.
 132          */
 133         if (reason == SCSI_MLQUEUE_HOST_BUSY)
 134                 host->host_blocked = host->max_host_blocked;
 135         else if (reason == SCSI_MLQUEUE_DEVICE_BUSY)
 136                 device->device_blocked = device->max_device_blocked;
 137
 138         /*
 139          * Decrement the counters, since these commands are no longer
 140          * active on the host/device.
 141          */
 142         scsi_device_unbusy(device);
 143
 144         /*
 145          * Requeue this command.  It will go before all other commands
 146          * that are already in the queue.
 147          *
 148          * NOTE: there is magic here about the way the queue is plugged if
 149          * we have no outstanding commands.
 150          *
 151          * Although we *don't* plug the queue, we call the request
 152          * function.  The SCSI request function detects the blocked condition
 153          * and plugs the queue appropriately.
 154          */
 155         spin_lock_irqsave(q->queue_lock, flags);
 156         blk_requeue_request(q, cmd->request);
 157         spin_unlock_irqrestore(q->queue_lock, flags);
 158
 159         scsi_run_queue(q);
 160
 161         return 0;
 162 }
 163
 164 /**
 165  * scsi_execute - insert request and wait for the result
 166  * @sdev:       scsi device
 167  * @cmd:        scsi command
 168  * @data_direction: data direction
 169  * @buffer:     data buffer
 170  * @bufflen:    len of buffer
 171  * @sense:      optional sense buffer
 172  * @timeout:    request timeout in seconds
 173  * @retries:    number of times to retry request
 174  * @flags:      or into request flags;
 175  *
 176  * returns the req->errors value which is the scsi_cmnd result
 177  * field.
 178  **/
 179 int scsi_execute(struct scsi_device *sdev, const unsigned char *cmd,
 180                  int data_direction, void *buffer, unsigned bufflen,
 181                  unsigned char *sense, int timeout, int retries, int flags)
 182 {
 183         struct request *req;
 184         int write = (data_direction == DMA_TO_DEVICE);
 185         int ret = DRIVER_ERROR << 24;
 186
 187         req = blk_get_request(sdev->request_queue, write, __GFP_WAIT);
 188
 189         if (bufflen &&  blk_rq_map_kern(sdev->request_queue, req,
 190                                         buffer, bufflen, __GFP_WAIT))
 191                 goto out;
 192
 193         req->cmd_len = COMMAND_SIZE(cmd[0]);
 194         memcpy(req->cmd, cmd, req->cmd_len);
 195         req->sense = sense;
 196         req->sense_len = 0;
 197         req->retries = retries;
 198         req->timeout = timeout;
 199         req->cmd_type = REQ_TYPE_BLOCK_PC;
 200         req->cmd_flags |= flags | REQ_QUIET | REQ_PREEMPT;
 201
 202         /*
 203          * head injection *required* here otherwise quiesce won't work
 204          */
 205         blk_execute_rq(req->q, NULL, req, 1);
 206
 207         ret = req->errors;
 208  out:
 209         blk_put_request(req);
 210
 211         return ret;
 212 }
 213 EXPORT_SYMBOL(scsi_execute);
 214
 215
 216 int scsi_execute_req(struct scsi_device *sdev, const unsigned char *cmd,
 217                      int data_direction, void *buffer, unsigned bufflen,
 218                      struct scsi_sense_hdr *sshdr, int timeout, int retries)
 219 {
 220         char *sense = NULL;
 221         int result;
 222
 223         if (sshdr) {
 224                 sense = kzalloc(SCSI_SENSE_BUFFERSIZE, GFP_NOIO);
 225                 if (!sense)
 226                         return DRIVER_ERROR << 24;
 227         }
 228         result = scsi_execute(sdev, cmd, data_direction, buffer, bufflen,
 229                               sense, timeout, retries, 0);
 230         if (sshdr)
 231                 scsi_normalize_sense(sense, SCSI_SENSE_BUFFERSIZE, sshdr);
 232
 233         kfree(sense);
 234         return result;
 235 }
 236 EXPORT_SYMBOL(scsi_execute_req);
 237
 238 struct scsi_io_context {
 239         void *data;
 240         void (*done)(void *data, char *sense, int result, int resid);
 241         char sense[SCSI_SENSE_BUFFERSIZE];
 242 };
 243
 244 static struct kmem_cache *scsi_io_context_cache;
 245
 246 static void scsi_end_async(struct request *req, int uptodate)
 247 {
 248         struct scsi_io_context *sioc = req->end_io_data;
 249
 250         if (sioc->done)
 251                 sioc->done(sioc->data, sioc->sense, req->errors, req->data_len);
 252
 253         kmem_cache_free(scsi_io_context_cache, sioc);
 254         __blk_put_request(req->q, req);
 255 }
 256
 257 static int scsi_merge_bio(struct request *rq, struct bio *bio)
 258 {
 259         struct request_queue *q = rq->q;
 260
 261         bio->bi_flags &= ~(1 << BIO_SEG_VALID);
 262         if (rq_data_dir(rq) == WRITE)
 263                 bio->bi_rw |= (1 << BIO_RW);
 264         blk_queue_bounce(q, &bio);
 265
 266         return blk_rq_append_bio(q, rq, bio);
 267 }
 268
 269 static void scsi_bi_endio(struct bio *bio, int error)
 270 {
 271         bio_put(bio);
 272 }
 273
 274 /**
 275  * scsi_req_map_sg - map a scatterlist into a request
 276  * @rq:         request to fill
 277  * @sg:         scatterlist
 278  * @nsegs:      number of elements
 279  * @bufflen:    len of buffer
 280  * @gfp:        memory allocation flags
 281  *
 282  * scsi_req_map_sg maps a scatterlist into a request so that the
 283  * request can be sent to the block layer. We do not trust the scatterlist
 284  * sent to use, as some ULDs use that struct to only organize the pages.
 285  */
 286 static int scsi_req_map_sg(struct request *rq, struct scatterlist *sgl,
 287                            int nsegs, unsigned bufflen, gfp_t gfp)
 288 {
 289         struct request_queue *q = rq->q;
 290         int nr_pages = (bufflen + sgl[0].offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
 291         unsigned int data_len = 0, len, bytes, off;
 292         struct page *page;
 293         struct bio *bio = NULL;
 294         int i, err, nr_vecs = 0;
 295
 296         for (i = 0; i < nsegs; i++) {
 297                 page = sgl[i].page;
 298                 off = sgl[i].offset;
 299                 len = sgl[i].length;
 300                 data_len += len;
 301
 302                 while (len > 0) {
 303                         bytes = min_t(unsigned int, len, PAGE_SIZE - off);
 304
 305                         if (!bio) {
 306                                 nr_vecs = min_t(int, BIO_MAX_PAGES, nr_pages);
 307                                 nr_pages -= nr_vecs;
 308
 309                                 bio = bio_alloc(gfp, nr_vecs);
 310                                 if (!bio) {
 311                                         err = -ENOMEM;
 312                                         goto free_bios;
 313                                 }
 314                                 bio->bi_end_io = scsi_bi_endio;
 315                         }
 316
 317                         if (bio_add_pc_page(q, bio, page, bytes, off) !=
 318                             bytes) {
 319                                 bio_put(bio);
 320                                 err = -EINVAL;
 321                                 goto free_bios;
 322                         }
 323
 324                         if (bio->bi_vcnt >= nr_vecs) {
 325                                 err = scsi_merge_bio(rq, bio);
 326                                 if (err) {
 327                                         bio_endio(bio, 0);
 328                                         goto free_bios;
 329                                 }
 330                                 bio = NULL;
 331                         }
 332
 333                         page++;
 334                         len -= bytes;
 335                         off = 0;
 336                 }
 337         }
 338
 339         rq->buffer = rq->data = NULL;
 340         rq->data_len = data_len;
 341         return 0;
 342
 343 free_bios:
 344         while ((bio = rq->bio) != NULL) {
 345                 rq->bio = bio->bi_next;
 346                 /*
 347                  * call endio instead of bio_put incase it was bounced
 348                  */
 349                 bio_endio(bio, 0);
 350         }
 351
 352         return err;
 353 }
 354
 355 /**
 356  * scsi_execute_async - insert request
 357  * @sdev:       scsi device
 358  * @cmd:        scsi command
 359  * @cmd_len:    length of scsi cdb
 360  * @data_direction: data direction
 361  * @buffer:     data buffer (this can be a kernel buffer or scatterlist)
 362  * @bufflen:    len of buffer
 363  * @use_sg:     if buffer is a scatterlist this is the number of elements
 364  * @timeout:    request timeout in seconds
 365  * @retries:    number of times to retry request
 366  * @flags:      or into request flags
 367  **/
 368 int scsi_execute_async(struct scsi_device *sdev, const unsigned char *cmd,
 369                        int cmd_len, int data_direction, void *buffer, unsigned bufflen,
 370                        int use_sg, int timeout, int retries, void *privdata,
 371                        void (*done)(void *, char *, int, int), gfp_t gfp)
 372 {
 373         struct request *req;
 374         struct scsi_io_context *sioc;
 375         int err = 0;
 376         int write = (data_direction == DMA_TO_DEVICE);
 377
 378         sioc = kmem_cache_zalloc(scsi_io_context_cache, gfp);
 379         if (!sioc)
 380                 return DRIVER_ERROR << 24;
 381
 382         req = blk_get_request(sdev->request_queue, write, gfp);
 383         if (!req)
 384                 goto free_sense;
 385         req->cmd_type = REQ_TYPE_BLOCK_PC;
 386         req->cmd_flags |= REQ_QUIET;
 387
 388         if (use_sg)
 389                 err = scsi_req_map_sg(req, buffer, use_sg, bufflen, gfp);
 390         else if (bufflen)
 391                 err = blk_rq_map_kern(req->q, req, buffer, bufflen, gfp);
 392
 393         if (err)
 394                 goto free_req;
 395
 396         req->cmd_len = cmd_len;
 397         memset(req->cmd, 0, BLK_MAX_CDB); /* ATAPI hates garbage after CDB */
 398         memcpy(req->cmd, cmd, req->cmd_len);
 399         req->sense = sioc->sense;
 400         req->sense_len = 0;
 401         req->timeout = timeout;
 402         req->retries = retries;
 403         req->end_io_data = sioc;
 404
 405         sioc->data = privdata;
 406         sioc->done = done;
 407
 408         blk_execute_rq_nowait(req->q, NULL, req, 1, scsi_end_async);
 409         return 0;
 410
 411 free_req:
 412         blk_put_request(req);
 413 free_sense:
 414         kmem_cache_free(scsi_io_context_cache, sioc);
 415         return DRIVER_ERROR << 24;
 416 }
 417 EXPORT_SYMBOL_GPL(scsi_execute_async);
 418
 419 /*
 420  * Function:    scsi_init_cmd_errh()
 421  *
 422  * Purpose:     Initialize cmd fields related to error handling.
 423  *
 424  * Arguments:   cmd     - command that is ready to be queued.
 425  *
 426  * Notes:       This function has the job of initializing a number of
 427  *              fields related to error handling.   Typically this will
 428  *              be called once for each command, as required.
 429  */
 430 static void scsi_init_cmd_errh(struct scsi_cmnd *cmd)
 431 {
 432         cmd->serial_number = 0;
 433         memset(cmd->sense_buffer, 0, sizeof cmd->sense_buffer);
 434         if (cmd->cmd_len == 0)
 435                 cmd->cmd_len = COMMAND_SIZE(cmd->cmnd[0]);
 436 }
 437
 438 void scsi_device_unbusy(struct scsi_device *sdev)
 439 {
 440         struct Scsi_Host *shost = sdev->host;
 441         unsigned long flags;
 442
 443         spin_lock_irqsave(shost->host_lock, flags);
 444         shost->host_busy--;
 445         if (unlikely(scsi_host_in_recovery(shost) &&
 446                      (shost->host_failed || shost->host_eh_scheduled)))
 447                 scsi_eh_wakeup(shost);
 448         spin_unlock(shost->host_lock);
 449         spin_lock(sdev->request_queue->queue_lock);
 450         sdev->device_busy--;
 451         spin_unlock_irqrestore(sdev->request_queue->queue_lock, flags);
 452 }
 453
 454 /*
 455  * Called for single_lun devices on IO completion. Clear starget_sdev_user,
 456  * and call blk_run_queue for all the scsi_devices on the target -
 457  * including current_sdev first.
 458  *
 459  * Called with *no* scsi locks held.
 460  */
 461 static void scsi_single_lun_run(struct scsi_device *current_sdev)
 462 {
 463         struct Scsi_Host *shost = current_sdev->host;
 464         struct scsi_device *sdev, *tmp;
 465         struct scsi_target *starget = scsi_target(current_sdev);
 466         unsigned long flags;
 467
 468         spin_lock_irqsave(shost->host_lock, flags);
 469         starget->starget_sdev_user = NULL;
 470         spin_unlock_irqrestore(shost->host_lock, flags);
 471
 472         /*
 473          * Call blk_run_queue for all LUNs on the target, starting with
 474          * current_sdev. We race with others (to set starget_sdev_user),
 475          * but in most cases, we will be first. Ideally, each LU on the
 476          * target would get some limited time or requests on the target.
 477          */
 478         blk_run_queue(current_sdev->request_queue);
 479
 480         spin_lock_irqsave(shost->host_lock, flags);
 481         if (starget->starget_sdev_user)
 482                 goto out;
 483         list_for_each_entry_safe(sdev, tmp, &starget->devices,
 484                         same_target_siblings) {
 485                 if (sdev == current_sdev)
 486                         continue;
 487                 if (scsi_device_get(sdev))
 488                         continue;
 489
 490                 spin_unlock_irqrestore(shost->host_lock, flags);
 491                 blk_run_queue(sdev->request_queue);
 492                 spin_lock_irqsave(shost->host_lock, flags);
 493
 494                 scsi_device_put(sdev);
 495         }
 496  out:
 497         spin_unlock_irqrestore(shost->host_lock, flags);
 498 }
 499
 500 /*
 501  * Function:    scsi_run_queue()
 502  *
 503  * Purpose:     Select a proper request queue to serve next
 504  *
 505  * Arguments:   q       - last request's queue
 506  *
 507  * Returns:     Nothing
 508  *
 509  * Notes:       The previous command was completely finished, start
 510  *              a new one if possible.
 511  */
 512 static void scsi_run_queue(struct request_queue *q)
 513 {
 514         struct scsi_device *sdev = q->queuedata;
 515         struct Scsi_Host *shost = sdev->host;
 516         unsigned long flags;
 517
 518         if (sdev->single_lun)
 519                 scsi_single_lun_run(sdev);
 520
 521         spin_lock_irqsave(shost->host_lock, flags);
 522         while (!list_empty(&shost->starved_list) &&
 523                !shost->host_blocked && !shost->host_self_blocked &&
 524                 !((shost->can_queue > 0) &&
 525                   (shost->host_busy >= shost->can_queue))) {
 526                 /*
 527                  * As long as shost is accepting commands and we have
 528                  * starved queues, call blk_run_queue. scsi_request_fn
 529                  * drops the queue_lock and can add us back to the
 530                  * starved_list.
 531                  *
 532                  * host_lock protects the starved_list and starved_entry.
 533                  * scsi_request_fn must get the host_lock before checking
 534                  * or modifying starved_list or starved_entry.
 535                  */
 536                 sdev = list_entry(shost->starved_list.next,
 537                                           struct scsi_device, starved_entry);
 538                 list_del_init(&sdev->starved_entry);
 539                 spin_unlock_irqrestore(shost->host_lock, flags);
 540
 541
 542                 if (test_bit(QUEUE_FLAG_REENTER, &q->queue_flags) &&
 543                     !test_and_set_bit(QUEUE_FLAG_REENTER,
 544                                       &sdev->request_queue->queue_flags)) {
 545                         blk_run_queue(sdev->request_queue);
 546                         clear_bit(QUEUE_FLAG_REENTER,
 547                                   &sdev->request_queue->queue_flags);
 548                 } else
 549                         blk_run_queue(sdev->request_queue);
 550
 551                 spin_lock_irqsave(shost->host_lock, flags);
 552                 if (unlikely(!list_empty(&sdev->starved_entry)))
 553                         /*
 554                          * sdev lost a race, and was put back on the
 555                          * starved list. This is unlikely but without this
 556                          * in theory we could loop forever.
 557                          */
 558                         break;
 559         }
 560         spin_unlock_irqrestore(shost->host_lock, flags);
 561
 562         blk_run_queue(q);
 563 }
 564
 565 /*
 566  * Function:    scsi_requeue_command()
 567  *
 568  * Purpose:     Handle post-processing of completed commands.
 569  *
 570  * Arguments:   q       - queue to operate on
 571  *              cmd     - command that may need to be requeued.
 572  *
 573  * Returns:     Nothing
 574  *
 575  * Notes:       After command completion, there may be blocks left
 576  *              over which weren't finished by the previous command
 577  *              this can be for a number of reasons - the main one is
 578  *              I/O errors in the middle of the request, in which case
 579  *              we need to request the blocks that come after the bad
 580  *              sector.
 581  * Notes:       Upon return, cmd is a stale pointer.
 582  */
 583 static void scsi_requeue_command(struct request_queue *q, struct scsi_cmnd *cmd)
 584 {
 585         struct request *req = cmd->request;
 586         unsigned long flags;
 587
 588         scsi_unprep_request(req);
 589         spin_lock_irqsave(q->queue_lock, flags);
 590         blk_requeue_request(q, req);
 591         spin_unlock_irqrestore(q->queue_lock, flags);
 592
 593         scsi_run_queue(q);
 594 }
 595
 596 void scsi_next_command(struct scsi_cmnd *cmd)
 597 {
 598         struct scsi_device *sdev = cmd->device;
 599         struct request_queue *q = sdev->request_queue;
 600
 601         /* need to hold a reference on the device before we let go of the cmd */
 602         get_device(&sdev->sdev_gendev);
 603
 604         scsi_put_command(cmd);
 605         scsi_run_queue(q);
 606
 607         /* ok to remove device now */
 608         put_device(&sdev->sdev_gendev);
 609 }
 610
 611 void scsi_run_host_queues(struct Scsi_Host *shost)
 612 {
 613         struct scsi_device *sdev;
 614
 615         shost_for_each_device(sdev, shost)
 616                 scsi_run_queue(sdev->request_queue);
 617 }
 618
 619 /*
 620  * Function:    scsi_end_request()
 621  *
 622  * Purpose:     Post-processing of completed commands (usually invoked at end
 623  *              of upper level post-processing and scsi_io_completion).
 624  *
 625  * Arguments:   cmd      - command that is complete.
 626  *              uptodate - 1 if I/O indicates success, <= 0 for I/O error.
 627  *              bytes    - number of bytes of completed I/O
 628  *              requeue  - indicates whether we should requeue leftovers.
 629  *
 630  * Lock status: Assumed that lock is not held upon entry.
 631  *
 632  * Returns:     cmd if requeue required, NULL otherwise.
 633  *
 634  * Notes:       This is called for block device requests in order to
 635  *              mark some number of sectors as complete.
 636  *
 637  *              We are guaranteeing that the request queue will be goosed
 638  *              at some point during this call.
 639  * Notes:       If cmd was requeued, upon return it will be a stale pointer.
 640  */
 641 static struct scsi_cmnd *scsi_end_request(struct scsi_cmnd *cmd, int uptodate,
 642                                           int bytes, int requeue)
 643 {
 644         struct request_queue *q = cmd->device->request_queue;
 645         struct request *req = cmd->request;
 646         unsigned long flags;
 647
 648         /*
 649          * If there are blocks left over at the end, set up the command
 650          * to queue the remainder of them.
 651          */
 652         if (end_that_request_chunk(req, uptodate, bytes)) {
 653                 int leftover = (req->hard_nr_sectors << 9);
 654
 655                 if (blk_pc_request(req))
 656                         leftover = req->data_len;
 657
 658                 /* kill remainder if no retrys */
 659                 if (!uptodate && blk_noretry_request(req))
 660                         end_that_request_chunk(req, 0, leftover);
 661                 else {
 662                         if (requeue) {
 663                                 /*
 664                                  * Bleah.  Leftovers again.  Stick the
 665                                  * leftovers in the front of the
 666                                  * queue, and goose the queue again.
 667                                  */
 668                                 scsi_requeue_command(q, cmd);
 669                                 cmd = NULL;
 670                         }
 671                         return cmd;
 672                 }
 673         }
 674
 675         add_disk_randomness(req->rq_disk);
 676
 677         spin_lock_irqsave(q->queue_lock, flags);
 678         if (blk_rq_tagged(req))
 679                 blk_queue_end_tag(q, req);
 680         end_that_request_last(req, uptodate);
 681         spin_unlock_irqrestore(q->queue_lock, flags);
 682
 683         /*
 684          * This will goose the queue request function at the end, so we don't
 685          * need to worry about launching another command.
 686          */
 687         scsi_next_command(cmd);
 688         return NULL;
 689 }
 690
 691 struct scatterlist *scsi_alloc_sgtable(struct scsi_cmnd *cmd, gfp_t gfp_mask)
 692 {
 693         struct scsi_host_sg_pool *sgp;
 694         struct scatterlist *sgl;
 695
 696         BUG_ON(!cmd->use_sg);
 697
 698         switch (cmd->use_sg) {
 699         case 1 ... 8:
 700                 cmd->sglist_len = 0;
 701                 break;
 702         case 9 ... 16:
 703                 cmd->sglist_len = 1;
 704                 break;
 705         case 17 ... 32:
 706                 cmd->sglist_len = 2;
 707                 break;
 708 #if (SCSI_MAX_PHYS_SEGMENTS > 32)
 709         case 33 ... 64:
 710                 cmd->sglist_len = 3;
 711                 break;
 712 #if (SCSI_MAX_PHYS_SEGMENTS > 64)
 713         case 65 ... 128:
 714                 cmd->sglist_len = 4;
 715                 break;
 716 #if (SCSI_MAX_PHYS_SEGMENTS  > 128)
 717         case 129 ... 256:
 718                 cmd->sglist_len = 5;
 719                 break;
 720 #endif
 721 #endif
 722 #endif
 723         default:
 724                 return NULL;
 725         }
 726
 727         sgp = scsi_sg_pools + cmd->sglist_len;
 728         sgl = mempool_alloc(sgp->pool, gfp_mask);
 729         return sgl;
 730 }
 731
 732 EXPORT_SYMBOL(scsi_alloc_sgtable);
 733
 734 void scsi_free_sgtable(struct scatterlist *sgl, int index)
 735 {
 736         struct scsi_host_sg_pool *sgp;
 737
 738         BUG_ON(index >= SG_MEMPOOL_NR);
 739
 740         sgp = scsi_sg_pools + index;
 741         mempool_free(sgl, sgp->pool);
 742 }
 743
 744 EXPORT_SYMBOL(scsi_free_sgtable);
 745
 746 /*
 747  * Function:    scsi_release_buffers()
 748  *
 749  * Purpose:     Completion processing for block device I/O requests.
 750  *
 751  * Arguments:   cmd     - command that we are bailing.
 752  *
 753  * Lock status: Assumed that no lock is held upon entry.
 754  *
 755  * Returns:     Nothing
 756  *
 757  * Notes:       In the event that an upper level driver rejects a
 758  *              command, we must release resources allocated during
 759  *              the __init_io() function.  Primarily this would involve
 760  *              the scatter-gather table, and potentially any bounce
 761  *              buffers.
 762  */
 763 static void scsi_release_buffers(struct scsi_cmnd *cmd)
 764 {
 765         if (cmd->use_sg)
 766                 scsi_free_sgtable(cmd->request_buffer, cmd->sglist_len);
 767
 768         /*
 769          * Zero these out.  They now point to freed memory, and it is
 770          * dangerous to hang onto the pointers.
 771          */
 772         cmd->request_buffer = NULL;
 773         cmd->request_bufflen = 0;
 774 }
 775
 776 /*
 777  * Function:    scsi_io_completion()
 778  *
 779  * Purpose:     Completion processing for block device I/O requests.
 780  *
 781  * Arguments:   cmd   - command that is finished.
 782  *
 783  * Lock status: Assumed that no lock is held upon entry.
 784  *
 785  * Returns:     Nothing
 786  *
 787  * Notes:       This function is matched in terms of capabilities to
 788  *              the function that created the scatter-gather list.
 789  *              In other words, if there are no bounce buffers
 790  *              (the normal case for most drivers), we don't need
 791  *              the logic to deal with cleaning up afterwards.
 792  *
 793  *              We must do one of several things here:
 794  *
 795  *              a) Call scsi_end_request.  This will finish off the
 796  *                 specified number of sectors.  If we are done, the
 797  *                 command block will be released, and the queue
 798  *                 function will be goosed.  If we are not done, then
 799  *                 scsi_end_request will directly goose the queue.
 800  *
 801  *              b) We can just use scsi_requeue_command() here.  This would
 802  *                 be used if we just wanted to retry, for example.
 803  */
 804 void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes)
 805 {
 806         int result = cmd->result;
 807         int this_count = cmd->request_bufflen;
 808         struct request_queue *q = cmd->device->request_queue;
 809         struct request *req = cmd->request;
 810         int clear_errors = 1;
 811         struct scsi_sense_hdr sshdr;
 812         int sense_valid = 0;
 813         int sense_deferred = 0;
 814
 815         scsi_release_buffers(cmd);
 816
 817         if (result) {
 818                 sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
 819                 if (sense_valid)
 820                         sense_deferred = scsi_sense_is_deferred(&sshdr);
 821         }
 822
 823         if (blk_pc_request(req)) { /* SG_IO ioctl from block level */
 824                 req->errors = result;
 825                 if (result) {
 826                         clear_errors = 0;
 827                         if (sense_valid && req->sense) {
 828                                 /*
 829                                  * SG_IO wants current and deferred errors
 830                                  */
 831                                 int len = 8 + cmd->sense_buffer[7];
 832
 833                                 if (len > SCSI_SENSE_BUFFERSIZE)
 834                                         len = SCSI_SENSE_BUFFERSIZE;
 835                                 memcpy(req->sense, cmd->sense_buffer,  len);
 836                                 req->sense_len = len;
 837                         }
 838                 }
 839                 req->data_len = cmd->resid;
 840         }
 841
 842         /*
 843          * Next deal with any sectors which we were able to correctly
 844          * handle.
 845          */
 846         SCSI_LOG_HLCOMPLETE(1, printk("%ld sectors total, "
 847                                       "%d bytes done.\n",
 848                                       req->nr_sectors, good_bytes));
 849         SCSI_LOG_HLCOMPLETE(1, printk("use_sg is %d\n", cmd->use_sg));
 850
 851         if (clear_errors)
 852                 req->errors = 0;
 853
 854         /* A number of bytes were successfully read.  If there
 855          * are leftovers and there is some kind of error
 856          * (result != 0), retry the rest.
 857          */
 858         if (scsi_end_request(cmd, 1, good_bytes, result == 0) == NULL)
 859                 return;
 860
 861         /* good_bytes = 0, or (inclusive) there were leftovers and
 862          * result = 0, so scsi_end_request couldn't retry.
 863          */
 864         if (sense_valid && !sense_deferred) {
 865                 switch (sshdr.sense_key) {
 866                 case UNIT_ATTENTION:
 867                         if (cmd->device->removable) {
 868                                 /* Detected disc change.  Set a bit
 869                                  * and quietly refuse further access.
 870                                  */
 871                                 cmd->device->changed = 1;
 872                                 scsi_end_request(cmd, 0, this_count, 1);
 873                                 return;
 874                         } else {
 875                                 /* Must have been a power glitch, or a
 876                                  * bus reset.  Could not have been a
 877                                  * media change, so we just retry the
 878                                  * request and see what happens.
 879                                  */
 880                                 scsi_requeue_command(q, cmd);
 881                                 return;
 882                         }
 883                         break;
 884                 case ILLEGAL_REQUEST:
 885                         /* If we had an ILLEGAL REQUEST returned, then
 886                          * we may have performed an unsupported
 887                          * command.  The only thing this should be
 888                          * would be a ten byte read where only a six
 889                          * byte read was supported.  Also, on a system
 890                          * where READ CAPACITY failed, we may have
 891                          * read past the end of the disk.
 892                          */
 893                         if ((cmd->device->use_10_for_rw &&
 894                             sshdr.asc == 0x20 && sshdr.ascq == 0x00) &&
 895                             (cmd->cmnd[0] == READ_10 ||
 896                              cmd->cmnd[0] == WRITE_10)) {
 897                                 cmd->device->use_10_for_rw = 0;
 898                                 /* This will cause a retry with a
 899                                  * 6-byte command.
 900                                  */
 901                                 scsi_requeue_command(q, cmd);
 902                                 return;
 903                         } else {
 904                                 scsi_end_request(cmd, 0, this_count, 1);
 905                                 return;
 906                         }
 907                         break;
 908                 case NOT_READY:
 909                         /* If the device is in the process of becoming
 910                          * ready, or has a temporary blockage, retry.
 911                          */
 912                         if (sshdr.asc == 0x04) {
 913                                 switch (sshdr.ascq) {
 914                                 case 0x01: /* becoming ready */
 915                                 case 0x04: /* format in progress */
 916                                 case 0x05: /* rebuild in progress */
 917                                 case 0x06: /* recalculation in progress */
 918                                 case 0x07: /* operation in progress */
 919                                 case 0x08: /* Long write in progress */
 920                                 case 0x09: /* self test in progress */
 921                                         scsi_requeue_command(q, cmd);
 922                                         return;
 923                                 default:
 924                                         break;
 925                                 }
 926                         }
 927                         if (!(req->cmd_flags & REQ_QUIET)) {
 928                                 scmd_printk(KERN_INFO, cmd,
 929                                             "Device not ready: ");
 930                                 scsi_print_sense_hdr("", &sshdr);
 931                         }
 932                         scsi_end_request(cmd, 0, this_count, 1);
 933                         return;
 934                 case VOLUME_OVERFLOW:
 935                         if (!(req->cmd_flags & REQ_QUIET)) {
 936                                 scmd_printk(KERN_INFO, cmd,
 937                                             "Volume overflow, CDB: ");
 938                                 __scsi_print_command(cmd->cmnd);
 939                                 scsi_print_sense("", cmd);
 940                         }
 941                         /* See SSC3rXX or current. */
 942                         scsi_end_request(cmd, 0, this_count, 1);
 943                         return;
 944                 default:
 945                         break;
 946                 }
 947         }
 948         if (host_byte(result) == DID_RESET) {
 949                 /* Third party bus reset or reset for error recovery
 950                  * reasons.  Just retry the request and see what
 951                  * happens.
 952                  */
 953                 scsi_requeue_command(q, cmd);
 954                 return;
 955         }
 956         if (result) {
 957                 if (!(req->cmd_flags & REQ_QUIET)) {
 958                         scsi_print_result(cmd);
 959                         if (driver_byte(result) & DRIVER_SENSE)
 960                                 scsi_print_sense("", cmd);
 961                 }
 962         }
 963         scsi_end_request(cmd, 0, this_count, !result);
 964 }
 965 EXPORT_SYMBOL(scsi_io_completion);
 966
 967 /*
 968  * Function:    scsi_init_io()
 969  *
 970  * Purpose:     SCSI I/O initialize function.
 971  *
 972  * Arguments:   cmd   - Command descriptor we wish to initialize
 973  *
 974  * Returns:     0 on success
 975  *              BLKPREP_DEFER if the failure is retryable
 976  *              BLKPREP_KILL if the failure is fatal
 977  */
 978 static int scsi_init_io(struct scsi_cmnd *cmd)
 979 {
 980         struct request     *req = cmd->request;
 981         struct scatterlist *sgpnt;
 982         int                count;
 983
 984         /*
 985          * We used to not use scatter-gather for single segment request,
 986          * but now we do (it makes highmem I/O easier to support without
 987          * kmapping pages)
 988          */
 989         cmd->use_sg = req->nr_phys_segments;
 990
 991         /*
 992          * If sg table allocation fails, requeue request later.
 993          */
 994         sgpnt = scsi_alloc_sgtable(cmd, GFP_ATOMIC);
 995         if (unlikely(!sgpnt)) {
 996                 scsi_unprep_request(req);
 997                 return BLKPREP_DEFER;
 998         }
 999
1000         req->buffer = NULL;
1001         cmd->request_buffer = (char *) sgpnt;
1002         if (blk_pc_request(req))
1003                 cmd->request_bufflen = req->data_len;
1004         else
1005                 cmd->request_bufflen = req->nr_sectors << 9;
1006
1007         /*
1008          * Next, walk the list, and fill in the addresses and sizes of
1009          * each segment.
1010          */
1011         count = blk_rq_map_sg(req->q, req, cmd->request_buffer);
1012         if (likely(count <= cmd->use_sg)) {
1013                 cmd->use_sg = count;
1014                 return BLKPREP_OK;
1015         }
1016
1017         printk(KERN_ERR "Incorrect number of segments after building list\n");
1018         printk(KERN_ERR "counted %d, received %d\n", count, cmd->use_sg);
1019         printk(KERN_ERR "req nr_sec %lu, cur_nr_sec %u\n", req->nr_sectors,
1020                         req->current_nr_sectors);
1021
1022         /* release the command and kill it */
1023         scsi_release_buffers(cmd);
1024         scsi_put_command(cmd);
1025         return BLKPREP_KILL;
1026 }
1027
1028 static struct scsi_cmnd *scsi_get_cmd_from_req(struct scsi_device *sdev,
1029                 struct request *req)
1030 {
1031         struct scsi_cmnd *cmd;
1032
1033         if (!req->special) {
1034                 cmd = scsi_get_command(sdev, GFP_ATOMIC);
1035                 if (unlikely(!cmd))
1036                         return NULL;
1037                 req->special = cmd;
1038         } else {
1039                 cmd = req->special;
1040         }
1041
1042         /* pull a tag out of the request if we have one */
1043         cmd->tag = req->tag;
1044         cmd->request = req;
1045
1046         return cmd;
1047 }
1048
1049 static void scsi_blk_pc_done(struct scsi_cmnd *cmd)
1050 {
1051         BUG_ON(!blk_pc_request(cmd->request));
1052         /*
1053          * This will complete the whole command with uptodate=1 so
1054          * as far as the block layer is concerned the command completed
1055          * successfully. Since this is a REQ_BLOCK_PC command the
1056          * caller should check the request's errors value
1057          */
1058         scsi_io_completion(cmd, cmd->request_bufflen);
1059 }
1060
1061 static int scsi_setup_blk_pc_cmnd(struct scsi_device *sdev, struct request *req)
1062 {
1063         struct scsi_cmnd *cmd;
1064
1065         cmd = scsi_get_cmd_from_req(sdev, req);
1066         if (unlikely(!cmd))
1067                 return BLKPREP_DEFER;
1068
1069         /*
1070          * BLOCK_PC requests may transfer data, in which case they must
1071          * a bio attached to them.  Or they might contain a SCSI command
1072          * that does not transfer data, in which case they may optionally
1073          * submit a request without an attached bio.
1074          */
1075         if (req->bio) {
1076                 int ret;
1077
1078                 BUG_ON(!req->nr_phys_segments);
1079
1080                 ret = scsi_init_io(cmd);
1081                 if (unlikely(ret))
1082                         return ret;
1083         } else {
1084                 BUG_ON(req->data_len);
1085                 BUG_ON(req->data);
1086
1087                 cmd->request_bufflen = 0;
1088                 cmd->request_buffer = NULL;
1089                 cmd->use_sg = 0;
1090                 req->buffer = NULL;
1091         }
1092
1093         BUILD_BUG_ON(sizeof(req->cmd) > sizeof(cmd->cmnd));
1094         memcpy(cmd->cmnd, req->cmd, sizeof(cmd->cmnd));
1095         cmd->cmd_len = req->cmd_len;
1096         if (!req->data_len)
1097                 cmd->sc_data_direction = DMA_NONE;
1098         else if (rq_data_dir(req) == WRITE)
1099                 cmd->sc_data_direction = DMA_TO_DEVICE;
1100         else
1101                 cmd->sc_data_direction = DMA_FROM_DEVICE;
1102
1103         cmd->transfersize = req->data_len;
1104         cmd->allowed = req->retries;
1105         cmd->timeout_per_command = req->timeout;
1106         cmd->done = scsi_blk_pc_done;
1107         return BLKPREP_OK;
1108 }
1109
1110 /*
1111  * Setup a REQ_TYPE_FS command.  These are simple read/write request
1112  * from filesystems that still need to be translated to SCSI CDBs from
1113  * the ULD.
1114  */
1115 static int scsi_setup_fs_cmnd(struct scsi_device *sdev, struct request *req)
1116 {
1117         struct scsi_cmnd *cmd;
1118         struct scsi_driver *drv;
1119         int ret;
1120
1121         /*
1122          * Filesystem requests must transfer data.
1123          */
1124         BUG_ON(!req->nr_phys_segments);
1125
1126         cmd = scsi_get_cmd_from_req(sdev, req);
1127         if (unlikely(!cmd))
1128                 return BLKPREP_DEFER;
1129
1130         ret = scsi_init_io(cmd);
1131         if (unlikely(ret))
1132                 return ret;
1133
1134         /*
1135          * Initialize the actual SCSI command for this request.
1136          */
1137         drv = *(struct scsi_driver **)req->rq_disk->private_data;
1138         if (unlikely(!drv->init_command(cmd))) {
1139                 scsi_release_buffers(cmd);
1140                 scsi_put_command(cmd);
1141                 return BLKPREP_KILL;
1142         }
1143
1144         return BLKPREP_OK;
1145 }
1146
1147 static int scsi_prep_fn(struct request_queue *q, struct request *req)
1148 {
1149         struct scsi_device *sdev = q->queuedata;
1150         int ret = BLKPREP_OK;
1151
1152         /*
1153          * If the device is not in running state we will reject some
1154          * or all commands.
1155          */
1156         if (unlikely(sdev->sdev_state != SDEV_RUNNING)) {
1157                 switch (sdev->sdev_state) {
1158                 case SDEV_OFFLINE:
1159                         /*
1160                          * If the device is offline we refuse to process any
1161                          * commands.  The device must be brought online
1162                          * before trying any recovery commands.
1163                          */
1164                         sdev_printk(KERN_ERR, sdev,
1165                                     "rejecting I/O to offline device\n");
1166                         ret = BLKPREP_KILL;
1167                         break;
1168                 case SDEV_DEL:
1169                         /*
1170                          * If the device is fully deleted, we refuse to
1171                          * process any commands as well.
1172                          */
1173                         sdev_printk(KERN_ERR, sdev,
1174                                     "rejecting I/O to dead device\n");
1175                         ret = BLKPREP_KILL;
1176                         break;
1177                 case SDEV_QUIESCE:
1178                 case SDEV_BLOCK:
1179                         /*
1180                          * If the devices is blocked we defer normal commands.
1181                          */
1182                         if (!(req->cmd_flags & REQ_PREEMPT))
1183                                 ret = BLKPREP_DEFER;
1184                         break;
1185                 default:
1186                         /*
1187                          * For any other not fully online state we only allow
1188                          * special commands.  In particular any user initiated
1189                          * command is not allowed.
1190                          */
1191                         if (!(req->cmd_flags & REQ_PREEMPT))
1192                                 ret = BLKPREP_KILL;
1193                         break;
1194                 }
1195
1196                 if (ret != BLKPREP_OK)
1197                         goto out;
1198         }
1199
1200         switch (req->cmd_type) {
1201         case REQ_TYPE_BLOCK_PC:
1202                 ret = scsi_setup_blk_pc_cmnd(sdev, req);
1203                 break;
1204         case REQ_TYPE_FS:
1205                 ret = scsi_setup_fs_cmnd(sdev, req);
1206                 break;
1207         default:
1208                 /*
1209                  * All other command types are not supported.
1210                  *
1211                  * Note that these days the SCSI subsystem does not use
1212                  * REQ_TYPE_SPECIAL requests anymore.  These are only used
1213                  * (directly or via blk_insert_request) by non-SCSI drivers.
1214                  */
1215                 blk_dump_rq_flags(req, "SCSI bad req");
1216                 ret = BLKPREP_KILL;
1217                 break;
1218         }
1219
1220  out:
1221         switch (ret) {
1222         case BLKPREP_KILL:
1223                 req->errors = DID_NO_CONNECT << 16;
1224                 break;
1225         case BLKPREP_DEFER:
1226                 /*
1227                  * If we defer, the elv_next_request() returns NULL, but the
1228                  * queue must be restarted, so we plug here if no returning
1229                  * command will automatically do that.
1230                  */
1231                 if (sdev->device_busy == 0)
1232                         blk_plug_device(q);
1233                 break;
1234         default:
1235                 req->cmd_flags |= REQ_DONTPREP;
1236         }
1237
1238         return ret;
1239 }
1240
1241 /*
1242  * scsi_dev_queue_ready: if we can send requests to sdev, return 1 else
1243  * return 0.
1244  *
1245  * Called with the queue_lock held.
1246  */
1247 static inline int scsi_dev_queue_ready(struct request_queue *q,
1248                                   struct scsi_device *sdev)
1249 {
1250         if (sdev->device_busy >= sdev->queue_depth)
1251                 return 0;
1252         if (sdev->device_busy == 0 && sdev->device_blocked) {
1253                 /*
1254                  * unblock after device_blocked iterates to zero
1255                  */
1256                 if (--sdev->device_blocked == 0) {
1257                         SCSI_LOG_MLQUEUE(3,
1258                                    sdev_printk(KERN_INFO, sdev,
1259                                    "unblocking device at zero depth\n"));
1260                 } else {
1261                         blk_plug_device(q);
1262                         return 0;
1263                 }
1264         }
1265         if (sdev->device_blocked)
1266                 return 0;
1267
1268         return 1;
1269 }
1270
1271 /*
1272  * scsi_host_queue_ready: if we can send requests to shost, return 1 else
1273  * return 0. We must end up running the queue again whenever 0 is
1274  * returned, else IO can hang.
1275  *
1276  * Called with host_lock held.
1277  */
1278 static inline int scsi_host_queue_ready(struct request_queue *q,
1279                                    struct Scsi_Host *shost,
1280                                    struct scsi_device *sdev)
1281 {
1282         if (scsi_host_in_recovery(shost))
1283                 return 0;
1284         if (shost->host_busy == 0 && shost->host_blocked) {
1285                 /*
1286                  * unblock after host_blocked iterates to zero
1287                  */
1288                 if (--shost->host_blocked == 0) {
1289                         SCSI_LOG_MLQUEUE(3,
1290                                 printk("scsi%d unblocking host at zero depth\n",
1291                                         shost->host_no));
1292                 } else {
1293                         blk_plug_device(q);
1294                         return 0;
1295                 }
1296         }
1297         if ((shost->can_queue > 0 && shost->host_busy >= shost->can_queue) ||
1298             shost->host_blocked || shost->host_self_blocked) {
1299                 if (list_empty(&sdev->starved_entry))
1300                         list_add_tail(&sdev->starved_entry, &shost->starved_list);
1301                 return 0;
1302         }
1303
1304         /* We're OK to process the command, so we can't be starved */
1305         if (!list_empty(&sdev->starved_entry))
1306                 list_del_init(&sdev->starved_entry);
1307
1308         return 1;
1309 }
1310
1311 /*
1312  * Kill a request for a dead device
1313  */
1314 static void scsi_kill_request(struct request *req, struct request_queue *q)
1315 {
1316         struct scsi_cmnd *cmd = req->special;
1317         struct scsi_device *sdev = cmd->device;
1318         struct Scsi_Host *shost = sdev->host;
1319
1320         blkdev_dequeue_request(req);
1321
1322         if (unlikely(cmd == NULL)) {
1323                 printk(KERN_CRIT "impossible request in %s.\n",
1324                                  __FUNCTION__);
1325                 BUG();
1326         }
1327
1328         scsi_init_cmd_errh(cmd);
1329         cmd->result = DID_NO_CONNECT << 16;
1330         atomic_inc(&cmd->device->iorequest_cnt);
1331
1332         /*
1333          * SCSI request completion path will do scsi_device_unbusy(),
1334          * bump busy counts.  To bump the counters, we need to dance
1335          * with the locks as normal issue path does.
1336          */
1337         sdev->device_busy++;
1338         spin_unlock(sdev->request_queue->queue_lock);
1339         spin_lock(shost->host_lock);
1340         shost->host_busy++;
1341         spin_unlock(shost->host_lock);
1342         spin_lock(sdev->request_queue->queue_lock);
1343
1344         __scsi_done(cmd);
1345 }
1346
1347 static void scsi_softirq_done(struct request *rq)
1348 {
1349         struct scsi_cmnd *cmd = rq->completion_data;
1350         unsigned long wait_for = (cmd->allowed + 1) * cmd->timeout_per_command;
1351         int disposition;
1352
1353         INIT_LIST_HEAD(&cmd->eh_entry);
1354
1355         disposition = scsi_decide_disposition(cmd);
1356         if (disposition != SUCCESS &&
1357             time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) {
1358                 sdev_printk(KERN_ERR, cmd->device,
1359                             "timing out command, waited %lus\n",
1360                             wait_for/HZ);
1361                 disposition = SUCCESS;
1362         }
1363
1364         scsi_log_completion(cmd, disposition);
1365
1366         switch (disposition) {
1367                 case SUCCESS:
1368                         scsi_finish_command(cmd);
1369                         break;
1370                 case NEEDS_RETRY:
1371                         scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY);
1372                         break;
1373                 case ADD_TO_MLQUEUE:
1374                         scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY);
1375                         break;
1376                 default:
1377                         if (!scsi_eh_scmd_add(cmd, 0))
1378                                 scsi_finish_command(cmd);
1379         }
1380 }
1381
1382 /*
1383  * Function:    scsi_request_fn()
1384  *
1385  * Purpose:     Main strategy routine for SCSI.
1386  *
1387  * Arguments:   q       - Pointer to actual queue.
1388  *
1389  * Returns:     Nothing
1390  *
1391  * Lock status: IO request lock assumed to be held when called.
1392  */
1393 static void scsi_request_fn(struct request_queue *q)
1394 {
1395         struct scsi_device *sdev = q->queuedata;
1396         struct Scsi_Host *shost;
1397         struct scsi_cmnd *cmd;
1398         struct request *req;
1399
1400         if (!sdev) {
1401                 printk("scsi: killing requests for dead queue\n");
1402                 while ((req = elv_next_request(q)) != NULL)
1403                         scsi_kill_request(req, q);
1404                 return;
1405         }
1406
1407         if(!get_device(&sdev->sdev_gendev))
1408                 /* We must be tearing the block queue down already */
1409                 return;
1410
1411         /*
1412          * To start with, we keep looping until the queue is empty, or until
1413          * the host is no longer able to accept any more requests.
1414          */
1415         shost = sdev->host;
1416         while (!blk_queue_plugged(q)) {
1417                 int rtn;
1418                 /*
1419                  * get next queueable request.  We do this early to make sure
1420                  * that the request is fully prepared even if we cannot
1421                  * accept it.
1422                  */
1423                 req = elv_next_request(q);
1424                 if (!req || !scsi_dev_queue_ready(q, sdev))
1425                         break;
1426
1427                 if (unlikely(!scsi_device_online(sdev))) {
1428                         sdev_printk(KERN_ERR, sdev,
1429                                     "rejecting I/O to offline device\n");
1430                         scsi_kill_request(req, q);
1431                         continue;
1432                 }
1433
1434
1435                 /*
1436                  * Remove the request from the request list.
1437                  */
1438                 if (!(blk_queue_tagged(q) && !blk_queue_start_tag(q, req)))
1439                         blkdev_dequeue_request(req);
1440                 sdev->device_busy++;
1441
1442                 spin_unlock(q->queue_lock);
1443                 cmd = req->special;
1444                 if (unlikely(cmd == NULL)) {
1445                         printk(KERN_CRIT "impossible request in %s.\n"
1446                                          "please mail a stack trace to "
1447                                          "linux-scsi@vger.kernel.org\n",
1448                                          __FUNCTION__);
1449                         blk_dump_rq_flags(req, "foo");
1450                         BUG();
1451                 }
1452                 spin_lock(shost->host_lock);
1453
1454                 if (!scsi_host_queue_ready(q, shost, sdev))
1455                         goto not_ready;
1456                 if (sdev->single_lun) {
1457                         if (scsi_target(sdev)->starget_sdev_user &&
1458                             scsi_target(sdev)->starget_sdev_user != sdev)
1459                                 goto not_ready;
1460                         scsi_target(sdev)->starget_sdev_user = sdev;
1461                 }
1462                 shost->host_busy++;
1463
1464                 /*
1465                  * XXX(hch): This is rather suboptimal, scsi_dispatch_cmd will
1466                  *              take the lock again.
1467                  */
1468                 spin_unlock_irq(shost->host_lock);
1469
1470                 /*
1471                  * Finally, initialize any error handling parameters, and set up
1472                  * the timers for timeouts.
1473                  */
1474                 scsi_init_cmd_errh(cmd);
1475
1476                 /*
1477                  * Dispatch the command to the low-level driver.
1478                  */
1479                 rtn = scsi_dispatch_cmd(cmd);
1480                 spin_lock_irq(q->queue_lock);
1481                 if(rtn) {
1482                         /* we're refusing the command; because of
1483                          * the way locks get dropped, we need to
1484                          * check here if plugging is required */
1485                         if(sdev->device_busy == 0)
1486                                 blk_plug_device(q);
1487
1488                         break;
1489                 }
1490         }
1491
1492         goto out;
1493
1494  not_ready:
1495         spin_unlock_irq(shost->host_lock);
1496
1497         /*
1498          * lock q, handle tag, requeue req, and decrement device_busy. We
1499          * must return with queue_lock held.
1500          *
1501          * Decrementing device_busy without checking it is OK, as all such
1502          * cases (host limits or settings) should run the queue at some
1503          * later time.
1504          */
1505         spin_lock_irq(q->queue_lock);
1506         blk_requeue_request(q, req);
1507         sdev->device_busy--;
1508         if(sdev->device_busy == 0)
1509                 blk_plug_device(q);
1510  out:
1511         /* must be careful here...if we trigger the ->remove() function
1512          * we cannot be holding the q lock */
1513         spin_unlock_irq(q->queue_lock);
1514         put_device(&sdev->sdev_gendev);
1515         spin_lock_irq(q->queue_lock);
1516 }
1517
1518 u64 scsi_calculate_bounce_limit(struct Scsi_Host *shost)
1519 {
1520         struct device *host_dev;
1521         u64 bounce_limit = 0xffffffff;
1522
1523         if (shost->unchecked_isa_dma)
1524                 return BLK_BOUNCE_ISA;
1525         /*
1526          * Platforms with virtual-DMA translation
1527          * hardware have no practical limit.
1528          */
1529         if (!PCI_DMA_BUS_IS_PHYS)
1530                 return BLK_BOUNCE_ANY;
1531
1532         host_dev = scsi_get_device(shost);
1533         if (host_dev && host_dev->dma_mask)
1534                 bounce_limit = *host_dev->dma_mask;
1535
1536         return bounce_limit;
1537 }
1538 EXPORT_SYMBOL(scsi_calculate_bounce_limit);
1539
1540 struct request_queue *__scsi_alloc_queue(struct Scsi_Host *shost,
1541                                          request_fn_proc *request_fn)
1542 {
1543         struct request_queue *q;
1544
1545         q = blk_init_queue(request_fn, NULL);
1546         if (!q)
1547                 return NULL;
1548
1549         blk_queue_max_hw_segments(q, shost->sg_tablesize);
1550         blk_queue_max_phys_segments(q, SCSI_MAX_PHYS_SEGMENTS);
1551         blk_queue_max_sectors(q, shost->max_sectors);
1552         blk_queue_bounce_limit(q, scsi_calculate_bounce_limit(shost));
1553         blk_queue_segment_boundary(q, shost->dma_boundary);
1554
1555         if (!shost->use_clustering)
1556                 clear_bit(QUEUE_FLAG_CLUSTER, &q->queue_flags);
1557         return q;
1558 }
1559 EXPORT_SYMBOL(__scsi_alloc_queue);
1560
1561 struct request_queue *scsi_alloc_queue(struct scsi_device *sdev)
1562 {
1563         struct request_queue *q;
1564
1565         q = __scsi_alloc_queue(sdev->host, scsi_request_fn);
1566         if (!q)
1567                 return NULL;
1568
1569         blk_queue_prep_rq(q, scsi_prep_fn);
1570         blk_queue_softirq_done(q, scsi_softirq_done);
1571         return q;
1572 }
1573
1574 void scsi_free_queue(struct request_queue *q)
1575 {
1576         blk_cleanup_queue(q);
1577 }
1578
1579 /*
1580  * Function:    scsi_block_requests()
1581  *
1582  * Purpose:     Utility function used by low-level drivers to prevent further
1583  *              commands from being queued to the device.
1584  *
1585  * Arguments:   shost       - Host in question
1586  *
1587  * Returns:     Nothing
1588  *
1589  * Lock status: No locks are assumed held.
1590  *
1591  * Notes:       There is no timer nor any other means by which the requests
1592  *              get unblocked other than the low-level driver calling
1593  *              scsi_unblock_requests().
1594  */
1595 void scsi_block_requests(struct Scsi_Host *shost)
1596 {
1597         shost->host_self_blocked = 1;
1598 }
1599 EXPORT_SYMBOL(scsi_block_requests);
1600
1601 /*
1602  * Function:    scsi_unblock_requests()
1603  *
1604  * Purpose:     Utility function used by low-level drivers to allow further
1605  *              commands from being queued to the device.
1606  *
1607  * Arguments:   shost       - Host in question
1608  *
1609  * Returns:     Nothing
1610  *
1611  * Lock status: No locks are assumed held.
1612  *
1613  * Notes:       There is no timer nor any other means by which the requests
1614  *              get unblocked other than the low-level driver calling
1615  *              scsi_unblock_requests().
1616  *
1617  *              This is done as an API function so that changes to the
1618  *              internals of the scsi mid-layer won't require wholesale
1619  *              changes to drivers that use this feature.
1620  */
1621 void scsi_unblock_requests(struct Scsi_Host *shost)
1622 {
1623         shost->host_self_blocked = 0;
1624         scsi_run_host_queues(shost);
1625 }
1626 EXPORT_SYMBOL(scsi_unblock_requests);
1627
1628 int __init scsi_init_queue(void)
1629 {
1630         int i;
1631
1632         scsi_io_context_cache = kmem_cache_create("scsi_io_context",
1633                                         sizeof(struct scsi_io_context),
1634                                         0, 0, NULL);
1635         if (!scsi_io_context_cache) {
1636                 printk(KERN_ERR "SCSI: can't init scsi io context cache\n");
1637                 return -ENOMEM;
1638         }
1639
1640         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1641                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1642                 int size = sgp->size * sizeof(struct scatterlist);
1643
1644                 sgp->slab = kmem_cache_create(sgp->name, size, 0,
1645                                 SLAB_HWCACHE_ALIGN, NULL);
1646                 if (!sgp->slab) {
1647                         printk(KERN_ERR "SCSI: can't init sg slab %s\n",
1648                                         sgp->name);
1649                 }
1650
1651                 sgp->pool = mempool_create_slab_pool(SG_MEMPOOL_SIZE,
1652                                                      sgp->slab);
1653                 if (!sgp->pool) {
1654                         printk(KERN_ERR "SCSI: can't init sg mempool %s\n",
1655                                         sgp->name);
1656                 }
1657         }
1658
1659         return 0;
1660 }
1661
1662 void scsi_exit_queue(void)
1663 {
1664         int i;
1665
1666         kmem_cache_destroy(scsi_io_context_cache);
1667
1668         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1669                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1670                 mempool_destroy(sgp->pool);
1671                 kmem_cache_destroy(sgp->slab);
1672         }
1673 }
1674
1675 /**
1676  *      scsi_mode_select - issue a mode select
1677  *      @sdev:  SCSI device to be queried
1678  *      @pf:    Page format bit (1 == standard, 0 == vendor specific)
1679  *      @sp:    Save page bit (0 == don't save, 1 == save)
1680  *      @modepage: mode page being requested
1681  *      @buffer: request buffer (may not be smaller than eight bytes)
1682  *      @len:   length of request buffer.
1683  *      @timeout: command timeout
1684  *      @retries: number of retries before failing
1685  *      @data: returns a structure abstracting the mode header data
1686  *      @sense: place to put sense data (or NULL if no sense to be collected).
1687  *              must be SCSI_SENSE_BUFFERSIZE big.
1688  *
1689  *      Returns zero if successful; negative error number or scsi
1690  *      status on error
1691  *
1692  */
1693 int
1694 scsi_mode_select(struct scsi_device *sdev, int pf, int sp, int modepage,
1695                  unsigned char *buffer, int len, int timeout, int retries,
1696                  struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1697 {
1698         unsigned char cmd[10];
1699         unsigned char *real_buffer;
1700         int ret;
1701
1702         memset(cmd, 0, sizeof(cmd));
1703         cmd[1] = (pf ? 0x10 : 0) | (sp ? 0x01 : 0);
1704
1705         if (sdev->use_10_for_ms) {
1706                 if (len > 65535)
1707                         return -EINVAL;
1708                 real_buffer = kmalloc(8 + len, GFP_KERNEL);
1709                 if (!real_buffer)
1710                         return -ENOMEM;
1711                 memcpy(real_buffer + 8, buffer, len);
1712                 len += 8;
1713                 real_buffer[0] = 0;
1714                 real_buffer[1] = 0;
1715                 real_buffer[2] = data->medium_type;
1716                 real_buffer[3] = data->device_specific;
1717                 real_buffer[4] = data->longlba ? 0x01 : 0;
1718                 real_buffer[5] = 0;
1719                 real_buffer[6] = data->block_descriptor_length >> 8;
1720                 real_buffer[7] = data->block_descriptor_length;
1721
1722                 cmd[0] = MODE_SELECT_10;
1723                 cmd[7] = len >> 8;
1724                 cmd[8] = len;
1725         } else {
1726                 if (len > 255 || data->block_descriptor_length > 255 ||
1727                     data->longlba)
1728                         return -EINVAL;
1729
1730                 real_buffer = kmalloc(4 + len, GFP_KERNEL);
1731                 if (!real_buffer)
1732                         return -ENOMEM;
1733                 memcpy(real_buffer + 4, buffer, len);
1734                 len += 4;
1735                 real_buffer[0] = 0;
1736                 real_buffer[1] = data->medium_type;
1737                 real_buffer[2] = data->device_specific;
1738                 real_buffer[3] = data->block_descriptor_length;
1739
1740
1741                 cmd[0] = MODE_SELECT;
1742                 cmd[4] = len;
1743         }
1744
1745         ret = scsi_execute_req(sdev, cmd, DMA_TO_DEVICE, real_buffer, len,
1746                                sshdr, timeout, retries);
1747         kfree(real_buffer);
1748         return ret;
1749 }
1750 EXPORT_SYMBOL_GPL(scsi_mode_select);
1751
1752 /**
1753  *      scsi_mode_sense - issue a mode sense, falling back from 10 to
1754  *              six bytes if necessary.
1755  *      @sdev:  SCSI device to be queried
1756  *      @dbd:   set if mode sense will allow block descriptors to be returned
1757  *      @modepage: mode page being requested
1758  *      @buffer: request buffer (may not be smaller than eight bytes)
1759  *      @len:   length of request buffer.
1760  *      @timeout: command timeout
1761  *      @retries: number of retries before failing
1762  *      @data: returns a structure abstracting the mode header data
1763  *      @sense: place to put sense data (or NULL if no sense to be collected).
1764  *              must be SCSI_SENSE_BUFFERSIZE big.
1765  *
1766  *      Returns zero if unsuccessful, or the header offset (either 4
1767  *      or 8 depending on whether a six or ten byte command was
1768  *      issued) if successful.
1769  **/
1770 int
1771 scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
1772                   unsigned char *buffer, int len, int timeout, int retries,
1773                   struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1774 {
1775         unsigned char cmd[12];
1776         int use_10_for_ms;
1777         int header_length;
1778         int result;
1779         struct scsi_sense_hdr my_sshdr;
1780
1781         memset(data, 0, sizeof(*data));
1782         memset(&cmd[0], 0, 12);
1783         cmd[1] = dbd & 0x18;    /* allows DBD and LLBA bits */
1784         cmd[2] = modepage;
1785
1786         /* caller might not be interested in sense, but we need it */
1787         if (!sshdr)
1788                 sshdr = &my_sshdr;
1789
1790  retry:
1791         use_10_for_ms = sdev->use_10_for_ms;
1792
1793         if (use_10_for_ms) {
1794                 if (len < 8)
1795                         len = 8;
1796
1797                 cmd[0] = MODE_SENSE_10;
1798                 cmd[8] = len;
1799                 header_length = 8;
1800         } else {
1801                 if (len < 4)
1802                         len = 4;
1803
1804                 cmd[0] = MODE_SENSE;
1805                 cmd[4] = len;
1806                 header_length = 4;
1807         }
1808
1809         memset(buffer, 0, len);
1810
1811         result = scsi_execute_req(sdev, cmd, DMA_FROM_DEVICE, buffer, len,
1812                                   sshdr, timeout, retries);
1813
1814         /* This code looks awful: what it's doing is making sure an
1815          * ILLEGAL REQUEST sense return identifies the actual command
1816          * byte as the problem.  MODE_SENSE commands can return
1817          * ILLEGAL REQUEST if the code page isn't supported */
1818
1819         if (use_10_for_ms && !scsi_status_is_good(result) &&
1820             (driver_byte(result) & DRIVER_SENSE)) {
1821                 if (scsi_sense_valid(sshdr)) {
1822                         if ((sshdr->sense_key == ILLEGAL_REQUEST) &&
1823                             (sshdr->asc == 0x20) && (sshdr->ascq == 0)) {
1824                                 /*
1825                                  * Invalid command operation code
1826                                  */
1827                                 sdev->use_10_for_ms = 0;
1828                                 goto retry;
1829                         }
1830                 }
1831         }
1832
1833         if(scsi_status_is_good(result)) {
1834                 if (unlikely(buffer[0] == 0x86 && buffer[1] == 0x0b &&
1835                              (modepage == 6 || modepage == 8))) {
1836                         /* Initio breakage? */
1837                         header_length = 0;
1838                         data->length = 13;
1839                         data->medium_type = 0;
1840                         data->device_specific = 0;
1841                         data->longlba = 0;
1842                         data->block_descriptor_length = 0;
1843                 } else if(use_10_for_ms) {
1844                         data->length = buffer[0]*256 + buffer[1] + 2;
1845                         data->medium_type = buffer[2];
1846                         data->device_specific = buffer[3];
1847                         data->longlba = buffer[4] & 0x01;
1848                         data->block_descriptor_length = buffer[6]*256
1849                                 + buffer[7];
1850                 } else {
1851                         data->length = buffer[0] + 1;
1852                         data->medium_type = buffer[1];
1853                         data->device_specific = buffer[2];
1854                         data->block_descriptor_length = buffer[3];
1855                 }
1856                 data->header_length = header_length;
1857         }
1858
1859         return result;
1860 }
1861 EXPORT_SYMBOL(scsi_mode_sense);
1862
1863 int
1864 scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries)
1865 {
1866         char cmd[] = {
1867                 TEST_UNIT_READY, 0, 0, 0, 0, 0,
1868         };
1869         struct scsi_sense_hdr sshdr;
1870         int result;
1871
1872         result = scsi_execute_req(sdev, cmd, DMA_NONE, NULL, 0, &sshdr,
1873                                   timeout, retries);
1874
1875         if ((driver_byte(result) & DRIVER_SENSE) && sdev->removable) {
1876
1877                 if ((scsi_sense_valid(&sshdr)) &&
1878                     ((sshdr.sense_key == UNIT_ATTENTION) ||
1879                      (sshdr.sense_key == NOT_READY))) {
1880                         sdev->changed = 1;
1881                         result = 0;
1882                 }
1883         }
1884         return result;
1885 }
1886 EXPORT_SYMBOL(scsi_test_unit_ready);
1887
1888 /**
1889  *      scsi_device_set_state - Take the given device through the device
1890  *              state model.
1891  *      @sdev:  scsi device to change the state of.
1892  *      @state: state to change to.
1893  *
1894  *      Returns zero if unsuccessful or an error if the requested
1895  *      transition is illegal.
1896  **/
1897 int
1898 scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
1899 {
1900         enum scsi_device_state oldstate = sdev->sdev_state;
1901
1902         if (state == oldstate)
1903                 return 0;
1904
1905         switch (state) {
1906         case SDEV_CREATED:
1907                 /* There are no legal states that come back to
1908                  * created.  This is the manually initialised start
1909                  * state */
1910                 goto illegal;
1911
1912         case SDEV_RUNNING:
1913                 switch (oldstate) {
1914                 case SDEV_CREATED:
1915                 case SDEV_OFFLINE:
1916                 case SDEV_QUIESCE:
1917                 case SDEV_BLOCK:
1918                         break;
1919                 default:
1920                         goto illegal;
1921                 }
1922                 break;
1923
1924         case SDEV_QUIESCE:
1925                 switch (oldstate) {
1926                 case SDEV_RUNNING:
1927                 case SDEV_OFFLINE:
1928                         break;
1929                 default:
1930                         goto illegal;
1931                 }
1932                 break;
1933
1934         case SDEV_OFFLINE:
1935                 switch (oldstate) {
1936                 case SDEV_CREATED:
1937                 case SDEV_RUNNING:
1938                 case SDEV_QUIESCE:
1939                 case SDEV_BLOCK:
1940                         break;
1941                 default:
1942                         goto illegal;
1943                 }
1944                 break;
1945
1946         case SDEV_BLOCK:
1947                 switch (oldstate) {
1948                 case SDEV_CREATED:
1949                 case SDEV_RUNNING:
1950                         break;
1951                 default:
1952                         goto illegal;
1953                 }
1954                 break;
1955
1956         case SDEV_CANCEL:
1957                 switch (oldstate) {
1958                 case SDEV_CREATED:
1959                 case SDEV_RUNNING:
1960                 case SDEV_QUIESCE:
1961                 case SDEV_OFFLINE:
1962                 case SDEV_BLOCK:
1963                         break;
1964                 default:
1965                         goto illegal;
1966                 }
1967                 break;
1968
1969         case SDEV_DEL:
1970                 switch (oldstate) {
1971                 case SDEV_CREATED:
1972                 case SDEV_RUNNING:
1973                 case SDEV_OFFLINE:
1974                 case SDEV_CANCEL:
1975                         break;
1976                 default:
1977                         goto illegal;
1978                 }
1979                 break;
1980
1981         }
1982         sdev->sdev_state = state;
1983         return 0;
1984
1985  illegal:
1986         SCSI_LOG_ERROR_RECOVERY(1,
1987                                 sdev_printk(KERN_ERR, sdev,
1988                                             "Illegal state transition %s->%s\n",
1989                                             scsi_device_state_name(oldstate),
1990                                             scsi_device_state_name(state))
1991                                 );
1992         return -EINVAL;
1993 }
1994 EXPORT_SYMBOL(scsi_device_set_state);
1995
1996 /**
1997  *      scsi_device_quiesce - Block user issued commands.
1998  *      @sdev:  scsi device to quiesce.
1999  *
2000  *      This works by trying to transition to the SDEV_QUIESCE state
2001  *      (which must be a legal transition).  When the device is in this
2002  *      state, only special requests will be accepted, all others will
2003  *      be deferred.  Since special requests may also be requeued requests,
2004  *      a successful return doesn't guarantee the device will be
2005  *      totally quiescent.
2006  *
2007  *      Must be called with user context, may sleep.
2008  *
2009  *      Returns zero if unsuccessful or an error if not.
2010  **/
2011 int
2012 scsi_device_quiesce(struct scsi_device *sdev)
2013 {
2014         int err = scsi_device_set_state(sdev, SDEV_QUIESCE);
2015         if (err)
2016                 return err;
2017
2018         scsi_run_queue(sdev->request_queue);
2019         while (sdev->device_busy) {
2020                 msleep_interruptible(200);
2021                 scsi_run_queue(sdev->request_queue);
2022         }
2023         return 0;
2024 }
2025 EXPORT_SYMBOL(scsi_device_quiesce);
2026
2027 /**
2028  *      scsi_device_resume - Restart user issued commands to a quiesced device.
2029  *      @sdev:  scsi device to resume.
2030  *
2031  *      Moves the device from quiesced back to running and restarts the
2032  *      queues.
2033  *
2034  *      Must be called with user context, may sleep.
2035  **/
2036 void
2037 scsi_device_resume(struct scsi_device *sdev)
2038 {
2039         if(scsi_device_set_state(sdev, SDEV_RUNNING))
2040                 return;
2041         scsi_run_queue(sdev->request_queue);
2042 }
2043 EXPORT_SYMBOL(scsi_device_resume);
2044
2045 static void
2046 device_quiesce_fn(struct scsi_device *sdev, void *data)
2047 {
2048         scsi_device_quiesce(sdev);
2049 }
2050
2051 void
2052 scsi_target_quiesce(struct scsi_target *starget)
2053 {
2054         starget_for_each_device(starget, NULL, device_quiesce_fn);
2055 }
2056 EXPORT_SYMBOL(scsi_target_quiesce);
2057
2058 static void
2059 device_resume_fn(struct scsi_device *sdev, void *data)
2060 {
2061         scsi_device_resume(sdev);
2062 }
2063
2064 void
2065 scsi_target_resume(struct scsi_target *starget)
2066 {
2067         starget_for_each_device(starget, NULL, device_resume_fn);
2068 }
2069 EXPORT_SYMBOL(scsi_target_resume);
2070
2071 /**
2072  * scsi_internal_device_block - internal function to put a device
2073  *                              temporarily into the SDEV_BLOCK state
2074  * @sdev:       device to block
2075  *
2076  * Block request made by scsi lld's to temporarily stop all
2077  * scsi commands on the specified device.  Called from interrupt
2078  * or normal process context.
2079  *
2080  * Returns zero if successful or error if not
2081  *
2082  * Notes:
2083  *      This routine transitions the device to the SDEV_BLOCK state
2084  *      (which must be a legal transition).  When the device is in this
2085  *      state, all commands are deferred until the scsi lld reenables
2086  *      the device with scsi_device_unblock or device_block_tmo fires.
2087  *      This routine assumes the host_lock is held on entry.
2088  **/
2089 int
2090 scsi_internal_device_block(struct scsi_device *sdev)
2091 {
2092         struct request_queue *q = sdev->request_queue;
2093         unsigned long flags;
2094         int err = 0;
2095
2096         err = scsi_device_set_state(sdev, SDEV_BLOCK);
2097         if (err)
2098                 return err;
2099
2100         /*
2101          * The device has transitioned to SDEV_BLOCK.  Stop the
2102          * block layer from calling the midlayer with this device's
2103          * request queue.
2104          */
2105         spin_lock_irqsave(q->queue_lock, flags);
2106         blk_stop_queue(q);
2107         spin_unlock_irqrestore(q->queue_lock, flags);
2108
2109         return 0;
2110 }
2111 EXPORT_SYMBOL_GPL(scsi_internal_device_block);
2112
2113 /**
2114  * scsi_internal_device_unblock - resume a device after a block request
2115  * @sdev:       device to resume
2116  *
2117  * Called by scsi lld's or the midlayer to restart the device queue
2118  * for the previously suspended scsi device.  Called from interrupt or
2119  * normal process context.
2120  *
2121  * Returns zero if successful or error if not.
2122  *
2123  * Notes:
2124  *      This routine transitions the device to the SDEV_RUNNING state
2125  *      (which must be a legal transition) allowing the midlayer to
2126  *      goose the queue for this device.  This routine assumes the
2127  *      host_lock is held upon entry.
2128  **/
2129 int
2130 scsi_internal_device_unblock(struct scsi_device *sdev)
2131 {
2132         struct request_queue *q = sdev->request_queue;
2133         int err;
2134         unsigned long flags;
2135
2136         /*
2137          * Try to transition the scsi device to SDEV_RUNNING
2138          * and goose the device queue if successful.
2139          */
2140         err = scsi_device_set_state(sdev, SDEV_RUNNING);
2141         if (err)
2142                 return err;
2143
2144         spin_lock_irqsave(q->queue_lock, flags);
2145         blk_start_queue(q);
2146         spin_unlock_irqrestore(q->queue_lock, flags);
2147
2148         return 0;
2149 }
2150 EXPORT_SYMBOL_GPL(scsi_internal_device_unblock);
2151
2152 static void
2153 device_block(struct scsi_device *sdev, void *data)
2154 {
2155         scsi_internal_device_block(sdev);
2156 }
2157
2158 static int
2159 target_block(struct device *dev, void *data)
2160 {
2161         if (scsi_is_target_device(dev))
2162                 starget_for_each_device(to_scsi_target(dev), NULL,
2163                                         device_block);
2164         return 0;
2165 }
2166
2167 void
2168 scsi_target_block(struct device *dev)
2169 {
2170         if (scsi_is_target_device(dev))
2171                 starget_for_each_device(to_scsi_target(dev), NULL,
2172                                         device_block);
2173         else
2174                 device_for_each_child(dev, NULL, target_block);
2175 }
2176 EXPORT_SYMBOL_GPL(scsi_target_block);
2177
2178 static void
2179 device_unblock(struct scsi_device *sdev, void *data)
2180 {
2181         scsi_internal_device_unblock(sdev);
2182 }
2183
2184 static int
2185 target_unblock(struct device *dev, void *data)
2186 {
2187         if (scsi_is_target_device(dev))
2188                 starget_for_each_device(to_scsi_target(dev), NULL,
2189                                         device_unblock);
2190         return 0;
2191 }
2192
2193 void
2194 scsi_target_unblock(struct device *dev)
2195 {
2196         if (scsi_is_target_device(dev))
2197                 starget_for_each_device(to_scsi_target(dev), NULL,
2198                                         device_unblock);
2199         else
2200                 device_for_each_child(dev, NULL, target_unblock);
2201 }
2202 EXPORT_SYMBOL_GPL(scsi_target_unblock);
2203
2204 /**
2205  * scsi_kmap_atomic_sg - find and atomically map an sg-elemnt
2206  * @sg:         scatter-gather list
2207  * @sg_count:   number of segments in sg
2208  * @offset:     offset in bytes into sg, on return offset into the mapped area
2209  * @len:        bytes to map, on return number of bytes mapped
2210  *
2211  * Returns virtual address of the start of the mapped page
2212  */
2213 void *scsi_kmap_atomic_sg(struct scatterlist *sg, int sg_count,
2214                           size_t *offset, size_t *len)
2215 {
2216         int i;
2217         size_t sg_len = 0, len_complete = 0;
2218         struct page *page;
2219
2220         WARN_ON(!irqs_disabled());
2221
2222         for (i = 0; i < sg_count; i++) {
2223                 len_complete = sg_len; /* Complete sg-entries */
2224                 sg_len += sg[i].length;
2225                 if (sg_len > *offset)
2226                         break;
2227         }
2228
2229         if (unlikely(i == sg_count)) {
2230                 printk(KERN_ERR "%s: Bytes in sg: %zu, requested offset %zu, "
2231                         "elements %d\n",
2232                        __FUNCTION__, sg_len, *offset, sg_count);
2233                 WARN_ON(1);
2234                 return NULL;
2235         }
2236
2237         /* Offset starting from the beginning of first page in this sg-entry */
2238         *offset = *offset - len_complete + sg[i].offset;
2239
2240         /* Assumption: contiguous pages can be accessed as "page + i" */
2241         page = nth_page(sg[i].page, (*offset >> PAGE_SHIFT));
2242         *offset &= ~PAGE_MASK;
2243
2244         /* Bytes in this sg-entry from *offset to the end of the page */
2245         sg_len = PAGE_SIZE - *offset;
2246         if (*len > sg_len)
2247                 *len = sg_len;
2248
2249         return kmap_atomic(page, KM_BIO_SRC_IRQ);
2250 }
2251 EXPORT_SYMBOL(scsi_kmap_atomic_sg);
2252
2253 /**
2254  * scsi_kunmap_atomic_sg - atomically unmap a virtual address, previously
2255  *                         mapped with scsi_kmap_atomic_sg
2256  * @virt:       virtual address to be unmapped
2257  */
2258 void scsi_kunmap_atomic_sg(void *virt)
2259 {
2260         kunmap_atomic(virt, KM_BIO_SRC_IRQ);
2261 }
2262 EXPORT_SYMBOL(scsi_kunmap_atomic_sg);