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         32
  35
  36 struct scsi_host_sg_pool {
  37         size_t          size;
  38         char            *name;
  39         kmem_cache_t    *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->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 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->flags |= flags | REQ_BLOCK_PC | REQ_SPECIAL | REQ_QUIET;
 200
 201         /*
 202          * head injection *required* here otherwise quiesce won't work
 203          */
 204         blk_execute_rq(req->q, NULL, req, 1);
 205
 206         ret = req->errors;
 207  out:
 208         blk_put_request(req);
 209
 210         return ret;
 211 }
 212 EXPORT_SYMBOL(scsi_execute);
 213
 214
 215 int scsi_execute_req(struct scsi_device *sdev, const unsigned char *cmd,
 216                      int data_direction, void *buffer, unsigned bufflen,
 217                      struct scsi_sense_hdr *sshdr, int timeout, int retries)
 218 {
 219         char *sense = NULL;
 220         int result;
 221
 222         if (sshdr) {
 223                 sense = kzalloc(SCSI_SENSE_BUFFERSIZE, GFP_NOIO);
 224                 if (!sense)
 225                         return DRIVER_ERROR << 24;
 226         }
 227         result = scsi_execute(sdev, cmd, data_direction, buffer, bufflen,
 228                               sense, timeout, retries, 0);
 229         if (sshdr)
 230                 scsi_normalize_sense(sense, SCSI_SENSE_BUFFERSIZE, sshdr);
 231
 232         kfree(sense);
 233         return result;
 234 }
 235 EXPORT_SYMBOL(scsi_execute_req);
 236
 237 struct scsi_io_context {
 238         void *data;
 239         void (*done)(void *data, char *sense, int result, int resid);
 240         char sense[SCSI_SENSE_BUFFERSIZE];
 241 };
 242
 243 static kmem_cache_t *scsi_io_context_cache;
 244
 245 static void scsi_end_async(struct request *req, int uptodate)
 246 {
 247         struct scsi_io_context *sioc = req->end_io_data;
 248
 249         if (sioc->done)
 250                 sioc->done(sioc->data, sioc->sense, req->errors, req->data_len);
 251
 252         kmem_cache_free(scsi_io_context_cache, sioc);
 253         __blk_put_request(req->q, req);
 254 }
 255
 256 static int scsi_merge_bio(struct request *rq, struct bio *bio)
 257 {
 258         struct request_queue *q = rq->q;
 259
 260         bio->bi_flags &= ~(1 << BIO_SEG_VALID);
 261         if (rq_data_dir(rq) == WRITE)
 262                 bio->bi_rw |= (1 << BIO_RW);
 263         blk_queue_bounce(q, &bio);
 264
 265         if (!rq->bio)
 266                 blk_rq_bio_prep(q, rq, bio);
 267         else if (!q->back_merge_fn(q, rq, bio))
 268                 return -EINVAL;
 269         else {
 270                 rq->biotail->bi_next = bio;
 271                 rq->biotail = bio;
 272                 rq->hard_nr_sectors += bio_sectors(bio);
 273                 rq->nr_sectors = rq->hard_nr_sectors;
 274         }
 275
 276         return 0;
 277 }
 278
 279 static int scsi_bi_endio(struct bio *bio, unsigned int bytes_done, int error)
 280 {
 281         if (bio->bi_size)
 282                 return 1;
 283
 284         bio_put(bio);
 285         return 0;
 286 }
 287
 288 /**
 289  * scsi_req_map_sg - map a scatterlist into a request
 290  * @rq:         request to fill
 291  * @sg:         scatterlist
 292  * @nsegs:      number of elements
 293  * @bufflen:    len of buffer
 294  * @gfp:        memory allocation flags
 295  *
 296  * scsi_req_map_sg maps a scatterlist into a request so that the
 297  * request can be sent to the block layer. We do not trust the scatterlist
 298  * sent to use, as some ULDs use that struct to only organize the pages.
 299  */
 300 static int scsi_req_map_sg(struct request *rq, struct scatterlist *sgl,
 301                            int nsegs, unsigned bufflen, gfp_t gfp)
 302 {
 303         struct request_queue *q = rq->q;
 304         int nr_pages = (bufflen + sgl[0].offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
 305         unsigned int data_len = 0, len, bytes, off;
 306         struct page *page;
 307         struct bio *bio = NULL;
 308         int i, err, nr_vecs = 0;
 309
 310         for (i = 0; i < nsegs; i++) {
 311                 page = sgl[i].page;
 312                 off = sgl[i].offset;
 313                 len = sgl[i].length;
 314                 data_len += len;
 315
 316                 while (len > 0) {
 317                         bytes = min_t(unsigned int, len, PAGE_SIZE - off);
 318
 319                         if (!bio) {
 320                                 nr_vecs = min_t(int, BIO_MAX_PAGES, nr_pages);
 321                                 nr_pages -= nr_vecs;
 322
 323                                 bio = bio_alloc(gfp, nr_vecs);
 324                                 if (!bio) {
 325                                         err = -ENOMEM;
 326                                         goto free_bios;
 327                                 }
 328                                 bio->bi_end_io = scsi_bi_endio;
 329                         }
 330
 331                         if (bio_add_pc_page(q, bio, page, bytes, off) !=
 332                             bytes) {
 333                                 bio_put(bio);
 334                                 err = -EINVAL;
 335                                 goto free_bios;
 336                         }
 337
 338                         if (bio->bi_vcnt >= nr_vecs) {
 339                                 err = scsi_merge_bio(rq, bio);
 340                                 if (err) {
 341                                         bio_endio(bio, bio->bi_size, 0);
 342                                         goto free_bios;
 343                                 }
 344                                 bio = NULL;
 345                         }
 346
 347                         page++;
 348                         len -= bytes;
 349                         off = 0;
 350                 }
 351         }
 352
 353         rq->buffer = rq->data = NULL;
 354         rq->data_len = data_len;
 355         return 0;
 356
 357 free_bios:
 358         while ((bio = rq->bio) != NULL) {
 359                 rq->bio = bio->bi_next;
 360                 /*
 361                  * call endio instead of bio_put incase it was bounced
 362                  */
 363                 bio_endio(bio, bio->bi_size, 0);
 364         }
 365
 366         return err;
 367 }
 368
 369 /**
 370  * scsi_execute_async - insert request
 371  * @sdev:       scsi device
 372  * @cmd:        scsi command
 373  * @cmd_len:    length of scsi cdb
 374  * @data_direction: data direction
 375  * @buffer:     data buffer (this can be a kernel buffer or scatterlist)
 376  * @bufflen:    len of buffer
 377  * @use_sg:     if buffer is a scatterlist this is the number of elements
 378  * @timeout:    request timeout in seconds
 379  * @retries:    number of times to retry request
 380  * @flags:      or into request flags
 381  **/
 382 int scsi_execute_async(struct scsi_device *sdev, const unsigned char *cmd,
 383                        int cmd_len, int data_direction, void *buffer, unsigned bufflen,
 384                        int use_sg, int timeout, int retries, void *privdata,
 385                        void (*done)(void *, char *, int, int), gfp_t gfp)
 386 {
 387         struct request *req;
 388         struct scsi_io_context *sioc;
 389         int err = 0;
 390         int write = (data_direction == DMA_TO_DEVICE);
 391
 392         sioc = kmem_cache_alloc(scsi_io_context_cache, gfp);
 393         if (!sioc)
 394                 return DRIVER_ERROR << 24;
 395         memset(sioc, 0, sizeof(*sioc));
 396
 397         req = blk_get_request(sdev->request_queue, write, gfp);
 398         if (!req)
 399                 goto free_sense;
 400         req->flags |= REQ_BLOCK_PC | REQ_QUIET;
 401
 402         if (use_sg)
 403                 err = scsi_req_map_sg(req, buffer, use_sg, bufflen, gfp);
 404         else if (bufflen)
 405                 err = blk_rq_map_kern(req->q, req, buffer, bufflen, gfp);
 406
 407         if (err)
 408                 goto free_req;
 409
 410         req->cmd_len = cmd_len;
 411         memcpy(req->cmd, cmd, req->cmd_len);
 412         req->sense = sioc->sense;
 413         req->sense_len = 0;
 414         req->timeout = timeout;
 415         req->retries = retries;
 416         req->end_io_data = sioc;
 417
 418         sioc->data = privdata;
 419         sioc->done = done;
 420
 421         blk_execute_rq_nowait(req->q, NULL, req, 1, scsi_end_async);
 422         return 0;
 423
 424 free_req:
 425         blk_put_request(req);
 426 free_sense:
 427         kfree(sioc);
 428         return DRIVER_ERROR << 24;
 429 }
 430 EXPORT_SYMBOL_GPL(scsi_execute_async);
 431
 432 /*
 433  * Function:    scsi_init_cmd_errh()
 434  *
 435  * Purpose:     Initialize cmd fields related to error handling.
 436  *
 437  * Arguments:   cmd     - command that is ready to be queued.
 438  *
 439  * Notes:       This function has the job of initializing a number of
 440  *              fields related to error handling.   Typically this will
 441  *              be called once for each command, as required.
 442  */
 443 static void scsi_init_cmd_errh(struct scsi_cmnd *cmd)
 444 {
 445         cmd->serial_number = 0;
 446         memset(cmd->sense_buffer, 0, sizeof cmd->sense_buffer);
 447         if (cmd->cmd_len == 0)
 448                 cmd->cmd_len = COMMAND_SIZE(cmd->cmnd[0]);
 449 }
 450
 451 void scsi_device_unbusy(struct scsi_device *sdev)
 452 {
 453         struct Scsi_Host *shost = sdev->host;
 454         unsigned long flags;
 455
 456         spin_lock_irqsave(shost->host_lock, flags);
 457         shost->host_busy--;
 458         if (unlikely(scsi_host_in_recovery(shost) &&
 459                      (shost->host_failed || shost->host_eh_scheduled)))
 460                 scsi_eh_wakeup(shost);
 461         spin_unlock(shost->host_lock);
 462         spin_lock(sdev->request_queue->queue_lock);
 463         sdev->device_busy--;
 464         spin_unlock_irqrestore(sdev->request_queue->queue_lock, flags);
 465 }
 466
 467 /*
 468  * Called for single_lun devices on IO completion. Clear starget_sdev_user,
 469  * and call blk_run_queue for all the scsi_devices on the target -
 470  * including current_sdev first.
 471  *
 472  * Called with *no* scsi locks held.
 473  */
 474 static void scsi_single_lun_run(struct scsi_device *current_sdev)
 475 {
 476         struct Scsi_Host *shost = current_sdev->host;
 477         struct scsi_device *sdev, *tmp;
 478         struct scsi_target *starget = scsi_target(current_sdev);
 479         unsigned long flags;
 480
 481         spin_lock_irqsave(shost->host_lock, flags);
 482         starget->starget_sdev_user = NULL;
 483         spin_unlock_irqrestore(shost->host_lock, flags);
 484
 485         /*
 486          * Call blk_run_queue for all LUNs on the target, starting with
 487          * current_sdev. We race with others (to set starget_sdev_user),
 488          * but in most cases, we will be first. Ideally, each LU on the
 489          * target would get some limited time or requests on the target.
 490          */
 491         blk_run_queue(current_sdev->request_queue);
 492
 493         spin_lock_irqsave(shost->host_lock, flags);
 494         if (starget->starget_sdev_user)
 495                 goto out;
 496         list_for_each_entry_safe(sdev, tmp, &starget->devices,
 497                         same_target_siblings) {
 498                 if (sdev == current_sdev)
 499                         continue;
 500                 if (scsi_device_get(sdev))
 501                         continue;
 502
 503                 spin_unlock_irqrestore(shost->host_lock, flags);
 504                 blk_run_queue(sdev->request_queue);
 505                 spin_lock_irqsave(shost->host_lock, flags);
 506
 507                 scsi_device_put(sdev);
 508         }
 509  out:
 510         spin_unlock_irqrestore(shost->host_lock, flags);
 511 }
 512
 513 /*
 514  * Function:    scsi_run_queue()
 515  *
 516  * Purpose:     Select a proper request queue to serve next
 517  *
 518  * Arguments:   q       - last request's queue
 519  *
 520  * Returns:     Nothing
 521  *
 522  * Notes:       The previous command was completely finished, start
 523  *              a new one if possible.
 524  */
 525 static void scsi_run_queue(struct request_queue *q)
 526 {
 527         struct scsi_device *sdev = q->queuedata;
 528         struct Scsi_Host *shost = sdev->host;
 529         unsigned long flags;
 530
 531         if (sdev->single_lun)
 532                 scsi_single_lun_run(sdev);
 533
 534         spin_lock_irqsave(shost->host_lock, flags);
 535         while (!list_empty(&shost->starved_list) &&
 536                !shost->host_blocked && !shost->host_self_blocked &&
 537                 !((shost->can_queue > 0) &&
 538                   (shost->host_busy >= shost->can_queue))) {
 539                 /*
 540                  * As long as shost is accepting commands and we have
 541                  * starved queues, call blk_run_queue. scsi_request_fn
 542                  * drops the queue_lock and can add us back to the
 543                  * starved_list.
 544                  *
 545                  * host_lock protects the starved_list and starved_entry.
 546                  * scsi_request_fn must get the host_lock before checking
 547                  * or modifying starved_list or starved_entry.
 548                  */
 549                 sdev = list_entry(shost->starved_list.next,
 550                                           struct scsi_device, starved_entry);
 551                 list_del_init(&sdev->starved_entry);
 552                 spin_unlock_irqrestore(shost->host_lock, flags);
 553
 554                 blk_run_queue(sdev->request_queue);
 555
 556                 spin_lock_irqsave(shost->host_lock, flags);
 557                 if (unlikely(!list_empty(&sdev->starved_entry)))
 558                         /*
 559                          * sdev lost a race, and was put back on the
 560                          * starved list. This is unlikely but without this
 561                          * in theory we could loop forever.
 562                          */
 563                         break;
 564         }
 565         spin_unlock_irqrestore(shost->host_lock, flags);
 566
 567         blk_run_queue(q);
 568 }
 569
 570 /*
 571  * Function:    scsi_requeue_command()
 572  *
 573  * Purpose:     Handle post-processing of completed commands.
 574  *
 575  * Arguments:   q       - queue to operate on
 576  *              cmd     - command that may need to be requeued.
 577  *
 578  * Returns:     Nothing
 579  *
 580  * Notes:       After command completion, there may be blocks left
 581  *              over which weren't finished by the previous command
 582  *              this can be for a number of reasons - the main one is
 583  *              I/O errors in the middle of the request, in which case
 584  *              we need to request the blocks that come after the bad
 585  *              sector.
 586  * Notes:       Upon return, cmd is a stale pointer.
 587  */
 588 static void scsi_requeue_command(struct request_queue *q, struct scsi_cmnd *cmd)
 589 {
 590         struct request *req = cmd->request;
 591         unsigned long flags;
 592
 593         scsi_unprep_request(req);
 594         spin_lock_irqsave(q->queue_lock, flags);
 595         blk_requeue_request(q, req);
 596         spin_unlock_irqrestore(q->queue_lock, flags);
 597
 598         scsi_run_queue(q);
 599 }
 600
 601 void scsi_next_command(struct scsi_cmnd *cmd)
 602 {
 603         struct scsi_device *sdev = cmd->device;
 604         struct request_queue *q = sdev->request_queue;
 605
 606         /* need to hold a reference on the device before we let go of the cmd */
 607         get_device(&sdev->sdev_gendev);
 608
 609         scsi_put_command(cmd);
 610         scsi_run_queue(q);
 611
 612         /* ok to remove device now */
 613         put_device(&sdev->sdev_gendev);
 614 }
 615
 616 void scsi_run_host_queues(struct Scsi_Host *shost)
 617 {
 618         struct scsi_device *sdev;
 619
 620         shost_for_each_device(sdev, shost)
 621                 scsi_run_queue(sdev->request_queue);
 622 }
 623
 624 /*
 625  * Function:    scsi_end_request()
 626  *
 627  * Purpose:     Post-processing of completed commands (usually invoked at end
 628  *              of upper level post-processing and scsi_io_completion).
 629  *
 630  * Arguments:   cmd      - command that is complete.
 631  *              uptodate - 1 if I/O indicates success, <= 0 for I/O error.
 632  *              bytes    - number of bytes of completed I/O
 633  *              requeue  - indicates whether we should requeue leftovers.
 634  *
 635  * Lock status: Assumed that lock is not held upon entry.
 636  *
 637  * Returns:     cmd if requeue required, NULL otherwise.
 638  *
 639  * Notes:       This is called for block device requests in order to
 640  *              mark some number of sectors as complete.
 641  *
 642  *              We are guaranteeing that the request queue will be goosed
 643  *              at some point during this call.
 644  * Notes:       If cmd was requeued, upon return it will be a stale pointer.
 645  */
 646 static struct scsi_cmnd *scsi_end_request(struct scsi_cmnd *cmd, int uptodate,
 647                                           int bytes, int requeue)
 648 {
 649         request_queue_t *q = cmd->device->request_queue;
 650         struct request *req = cmd->request;
 651         unsigned long flags;
 652
 653         /*
 654          * If there are blocks left over at the end, set up the command
 655          * to queue the remainder of them.
 656          */
 657         if (end_that_request_chunk(req, uptodate, bytes)) {
 658                 int leftover = (req->hard_nr_sectors << 9);
 659
 660                 if (blk_pc_request(req))
 661                         leftover = req->data_len;
 662
 663                 /* kill remainder if no retrys */
 664                 if (!uptodate && blk_noretry_request(req))
 665                         end_that_request_chunk(req, 0, leftover);
 666                 else {
 667                         if (requeue) {
 668                                 /*
 669                                  * Bleah.  Leftovers again.  Stick the
 670                                  * leftovers in the front of the
 671                                  * queue, and goose the queue again.
 672                                  */
 673                                 scsi_requeue_command(q, cmd);
 674                                 cmd = NULL;
 675                         }
 676                         return cmd;
 677                 }
 678         }
 679
 680         add_disk_randomness(req->rq_disk);
 681
 682         spin_lock_irqsave(q->queue_lock, flags);
 683         if (blk_rq_tagged(req))
 684                 blk_queue_end_tag(q, req);
 685         end_that_request_last(req, uptodate);
 686         spin_unlock_irqrestore(q->queue_lock, flags);
 687
 688         /*
 689          * This will goose the queue request function at the end, so we don't
 690          * need to worry about launching another command.
 691          */
 692         scsi_next_command(cmd);
 693         return NULL;
 694 }
 695
 696 static struct scatterlist *scsi_alloc_sgtable(struct scsi_cmnd *cmd, gfp_t gfp_mask)
 697 {
 698         struct scsi_host_sg_pool *sgp;
 699         struct scatterlist *sgl;
 700
 701         BUG_ON(!cmd->use_sg);
 702
 703         switch (cmd->use_sg) {
 704         case 1 ... 8:
 705                 cmd->sglist_len = 0;
 706                 break;
 707         case 9 ... 16:
 708                 cmd->sglist_len = 1;
 709                 break;
 710         case 17 ... 32:
 711                 cmd->sglist_len = 2;
 712                 break;
 713 #if (SCSI_MAX_PHYS_SEGMENTS > 32)
 714         case 33 ... 64:
 715                 cmd->sglist_len = 3;
 716                 break;
 717 #if (SCSI_MAX_PHYS_SEGMENTS > 64)
 718         case 65 ... 128:
 719                 cmd->sglist_len = 4;
 720                 break;
 721 #if (SCSI_MAX_PHYS_SEGMENTS  > 128)
 722         case 129 ... 256:
 723                 cmd->sglist_len = 5;
 724                 break;
 725 #endif
 726 #endif
 727 #endif
 728         default:
 729                 return NULL;
 730         }
 731
 732         sgp = scsi_sg_pools + cmd->sglist_len;
 733         sgl = mempool_alloc(sgp->pool, gfp_mask);
 734         return sgl;
 735 }
 736
 737 static void scsi_free_sgtable(struct scatterlist *sgl, int index)
 738 {
 739         struct scsi_host_sg_pool *sgp;
 740
 741         BUG_ON(index >= SG_MEMPOOL_NR);
 742
 743         sgp = scsi_sg_pools + index;
 744         mempool_free(sgl, sgp->pool);
 745 }
 746
 747 /*
 748  * Function:    scsi_release_buffers()
 749  *
 750  * Purpose:     Completion processing for block device I/O requests.
 751  *
 752  * Arguments:   cmd     - command that we are bailing.
 753  *
 754  * Lock status: Assumed that no lock is held upon entry.
 755  *
 756  * Returns:     Nothing
 757  *
 758  * Notes:       In the event that an upper level driver rejects a
 759  *              command, we must release resources allocated during
 760  *              the __init_io() function.  Primarily this would involve
 761  *              the scatter-gather table, and potentially any bounce
 762  *              buffers.
 763  */
 764 static void scsi_release_buffers(struct scsi_cmnd *cmd)
 765 {
 766         if (cmd->use_sg)
 767                 scsi_free_sgtable(cmd->request_buffer, cmd->sglist_len);
 768
 769         /*
 770          * Zero these out.  They now point to freed memory, and it is
 771          * dangerous to hang onto the pointers.
 772          */
 773         cmd->request_buffer = NULL;
 774         cmd->request_bufflen = 0;
 775 }
 776
 777 /*
 778  * Function:    scsi_io_completion()
 779  *
 780  * Purpose:     Completion processing for block device I/O requests.
 781  *
 782  * Arguments:   cmd   - command that is finished.
 783  *
 784  * Lock status: Assumed that no lock is held upon entry.
 785  *
 786  * Returns:     Nothing
 787  *
 788  * Notes:       This function is matched in terms of capabilities to
 789  *              the function that created the scatter-gather list.
 790  *              In other words, if there are no bounce buffers
 791  *              (the normal case for most drivers), we don't need
 792  *              the logic to deal with cleaning up afterwards.
 793  *
 794  *              We must do one of several things here:
 795  *
 796  *              a) Call scsi_end_request.  This will finish off the
 797  *                 specified number of sectors.  If we are done, the
 798  *                 command block will be released, and the queue
 799  *                 function will be goosed.  If we are not done, then
 800  *                 scsi_end_request will directly goose the queue.
 801  *
 802  *              b) We can just use scsi_requeue_command() here.  This would
 803  *                 be used if we just wanted to retry, for example.
 804  */
 805 void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes)
 806 {
 807         int result = cmd->result;
 808         int this_count = cmd->request_bufflen;
 809         request_queue_t *q = cmd->device->request_queue;
 810         struct request *req = cmd->request;
 811         int clear_errors = 1;
 812         struct scsi_sense_hdr sshdr;
 813         int sense_valid = 0;
 814         int sense_deferred = 0;
 815
 816         scsi_release_buffers(cmd);
 817
 818         if (result) {
 819                 sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
 820                 if (sense_valid)
 821                         sense_deferred = scsi_sense_is_deferred(&sshdr);
 822         }
 823
 824         if (blk_pc_request(req)) { /* SG_IO ioctl from block level */
 825                 req->errors = result;
 826                 if (result) {
 827                         clear_errors = 0;
 828                         if (sense_valid && req->sense) {
 829                                 /*
 830                                  * SG_IO wants current and deferred errors
 831                                  */
 832                                 int len = 8 + cmd->sense_buffer[7];
 833
 834                                 if (len > SCSI_SENSE_BUFFERSIZE)
 835                                         len = SCSI_SENSE_BUFFERSIZE;
 836                                 memcpy(req->sense, cmd->sense_buffer,  len);
 837                                 req->sense_len = len;
 838                         }
 839                 } else
 840                         req->data_len = cmd->resid;
 841         }
 842
 843         /*
 844          * Next deal with any sectors which we were able to correctly
 845          * handle.
 846          */
 847         SCSI_LOG_HLCOMPLETE(1, printk("%ld sectors total, "
 848                                       "%d bytes done.\n",
 849                                       req->nr_sectors, good_bytes));
 850         SCSI_LOG_HLCOMPLETE(1, printk("use_sg is %d\n", cmd->use_sg));
 851
 852         if (clear_errors)
 853                 req->errors = 0;
 854
 855         /* A number of bytes were successfully read.  If there
 856          * are leftovers and there is some kind of error
 857          * (result != 0), retry the rest.
 858          */
 859         if (scsi_end_request(cmd, 1, good_bytes, result == 0) == NULL)
 860                 return;
 861
 862         /* good_bytes = 0, or (inclusive) there were leftovers and
 863          * result = 0, so scsi_end_request couldn't retry.
 864          */
 865         if (sense_valid && !sense_deferred) {
 866                 switch (sshdr.sense_key) {
 867                 case UNIT_ATTENTION:
 868                         if (cmd->device->removable) {
 869                                 /* Detected disc change.  Set a bit
 870                                  * and quietly refuse further access.
 871                                  */
 872                                 cmd->device->changed = 1;
 873                                 scsi_end_request(cmd, 0, this_count, 1);
 874                                 return;
 875                         } else {
 876                                 /* Must have been a power glitch, or a
 877                                  * bus reset.  Could not have been a
 878                                  * media change, so we just retry the
 879                                  * request and see what happens.
 880                                  */
 881                                 scsi_requeue_command(q, cmd);
 882                                 return;
 883                         }
 884                         break;
 885                 case ILLEGAL_REQUEST:
 886                         /* If we had an ILLEGAL REQUEST returned, then
 887                          * we may have performed an unsupported
 888                          * command.  The only thing this should be
 889                          * would be a ten byte read where only a six
 890                          * byte read was supported.  Also, on a system
 891                          * where READ CAPACITY failed, we may have
 892                          * read past the end of the disk.
 893                          */
 894                         if ((cmd->device->use_10_for_rw &&
 895                             sshdr.asc == 0x20 && sshdr.ascq == 0x00) &&
 896                             (cmd->cmnd[0] == READ_10 ||
 897                              cmd->cmnd[0] == WRITE_10)) {
 898                                 cmd->device->use_10_for_rw = 0;
 899                                 /* This will cause a retry with a
 900                                  * 6-byte command.
 901                                  */
 902                                 scsi_requeue_command(q, cmd);
 903                                 return;
 904                         } else {
 905                                 scsi_end_request(cmd, 0, this_count, 1);
 906                                 return;
 907                         }
 908                         break;
 909                 case NOT_READY:
 910                         /* If the device is in the process of becoming
 911                          * ready, or has a temporary blockage, retry.
 912                          */
 913                         if (sshdr.asc == 0x04) {
 914                                 switch (sshdr.ascq) {
 915                                 case 0x01: /* becoming ready */
 916                                 case 0x04: /* format in progress */
 917                                 case 0x05: /* rebuild in progress */
 918                                 case 0x06: /* recalculation in progress */
 919                                 case 0x07: /* operation in progress */
 920                                 case 0x08: /* Long write in progress */
 921                                 case 0x09: /* self test in progress */
 922                                         scsi_requeue_command(q, cmd);
 923                                         return;
 924                                 default:
 925                                         break;
 926                                 }
 927                         }
 928                         if (!(req->flags & REQ_QUIET)) {
 929                                 scmd_printk(KERN_INFO, cmd,
 930                                             "Device not ready: ");
 931                                 scsi_print_sense_hdr("", &sshdr);
 932                         }
 933                         scsi_end_request(cmd, 0, this_count, 1);
 934                         return;
 935                 case VOLUME_OVERFLOW:
 936                         if (!(req->flags & REQ_QUIET)) {
 937                                 scmd_printk(KERN_INFO, cmd,
 938                                             "Volume overflow, CDB: ");
 939                                 __scsi_print_command(cmd->cmnd);
 940                                 scsi_print_sense("", cmd);
 941                         }
 942                         /* See SSC3rXX or current. */
 943                         scsi_end_request(cmd, 0, this_count, 1);
 944                         return;
 945                 default:
 946                         break;
 947                 }
 948         }
 949         if (host_byte(result) == DID_RESET) {
 950                 /* Third party bus reset or reset for error recovery
 951                  * reasons.  Just retry the request and see what
 952                  * happens.
 953                  */
 954                 scsi_requeue_command(q, cmd);
 955                 return;
 956         }
 957         if (result) {
 958                 if (!(req->flags & REQ_QUIET)) {
 959                         scmd_printk(KERN_INFO, cmd,
 960                                     "SCSI error: return code = 0x%08x\n",
 961                                     result);
 962                         if (driver_byte(result) & DRIVER_SENSE)
 963                                 scsi_print_sense("", cmd);
 964                 }
 965         }
 966         scsi_end_request(cmd, 0, this_count, !result);
 967 }
 968 EXPORT_SYMBOL(scsi_io_completion);
 969
 970 /*
 971  * Function:    scsi_init_io()
 972  *
 973  * Purpose:     SCSI I/O initialize function.
 974  *
 975  * Arguments:   cmd   - Command descriptor we wish to initialize
 976  *
 977  * Returns:     0 on success
 978  *              BLKPREP_DEFER if the failure is retryable
 979  *              BLKPREP_KILL if the failure is fatal
 980  */
 981 static int scsi_init_io(struct scsi_cmnd *cmd)
 982 {
 983         struct request     *req = cmd->request;
 984         struct scatterlist *sgpnt;
 985         int                count;
 986
 987         /*
 988          * if this is a rq->data based REQ_BLOCK_PC, setup for a non-sg xfer
 989          */
 990         if ((req->flags & REQ_BLOCK_PC) && !req->bio) {
 991                 cmd->request_bufflen = req->data_len;
 992                 cmd->request_buffer = req->data;
 993                 req->buffer = req->data;
 994                 cmd->use_sg = 0;
 995                 return 0;
 996         }
 997
 998         /*
 999          * we used to not use scatter-gather for single segment request,
1000          * but now we do (it makes highmem I/O easier to support without
1001          * kmapping pages)
1002          */
1003         cmd->use_sg = req->nr_phys_segments;
1004
1005         /*
1006          * if sg table allocation fails, requeue request later.
1007          */
1008         sgpnt = scsi_alloc_sgtable(cmd, GFP_ATOMIC);
1009         if (unlikely(!sgpnt)) {
1010                 scsi_unprep_request(req);
1011                 return BLKPREP_DEFER;
1012         }
1013
1014         cmd->request_buffer = (char *) sgpnt;
1015         cmd->request_bufflen = req->nr_sectors << 9;
1016         if (blk_pc_request(req))
1017                 cmd->request_bufflen = req->data_len;
1018         req->buffer = NULL;
1019
1020         /*
1021          * Next, walk the list, and fill in the addresses and sizes of
1022          * each segment.
1023          */
1024         count = blk_rq_map_sg(req->q, req, cmd->request_buffer);
1025
1026         /*
1027          * mapped well, send it off
1028          */
1029         if (likely(count <= cmd->use_sg)) {
1030                 cmd->use_sg = count;
1031                 return 0;
1032         }
1033
1034         printk(KERN_ERR "Incorrect number of segments after building list\n");
1035         printk(KERN_ERR "counted %d, received %d\n", count, cmd->use_sg);
1036         printk(KERN_ERR "req nr_sec %lu, cur_nr_sec %u\n", req->nr_sectors,
1037                         req->current_nr_sectors);
1038
1039         /* release the command and kill it */
1040         scsi_release_buffers(cmd);
1041         scsi_put_command(cmd);
1042         return BLKPREP_KILL;
1043 }
1044
1045 static int scsi_issue_flush_fn(request_queue_t *q, struct gendisk *disk,
1046                                sector_t *error_sector)
1047 {
1048         struct scsi_device *sdev = q->queuedata;
1049         struct scsi_driver *drv;
1050
1051         if (sdev->sdev_state != SDEV_RUNNING)
1052                 return -ENXIO;
1053
1054         drv = *(struct scsi_driver **) disk->private_data;
1055         if (drv->issue_flush)
1056                 return drv->issue_flush(&sdev->sdev_gendev, error_sector);
1057
1058         return -EOPNOTSUPP;
1059 }
1060
1061 static void scsi_blk_pc_done(struct scsi_cmnd *cmd)
1062 {
1063         BUG_ON(!blk_pc_request(cmd->request));
1064         /*
1065          * This will complete the whole command with uptodate=1 so
1066          * as far as the block layer is concerned the command completed
1067          * successfully. Since this is a REQ_BLOCK_PC command the
1068          * caller should check the request's errors value
1069          */
1070         scsi_io_completion(cmd, cmd->request_bufflen);
1071 }
1072
1073 static void scsi_setup_blk_pc_cmnd(struct scsi_cmnd *cmd)
1074 {
1075         struct request *req = cmd->request;
1076
1077         BUG_ON(sizeof(req->cmd) > sizeof(cmd->cmnd));
1078         memcpy(cmd->cmnd, req->cmd, sizeof(cmd->cmnd));
1079         cmd->cmd_len = req->cmd_len;
1080         if (!req->data_len)
1081                 cmd->sc_data_direction = DMA_NONE;
1082         else if (rq_data_dir(req) == WRITE)
1083                 cmd->sc_data_direction = DMA_TO_DEVICE;
1084         else
1085                 cmd->sc_data_direction = DMA_FROM_DEVICE;
1086
1087         cmd->transfersize = req->data_len;
1088         cmd->allowed = req->retries;
1089         cmd->timeout_per_command = req->timeout;
1090         cmd->done = scsi_blk_pc_done;
1091 }
1092
1093 static int scsi_prep_fn(struct request_queue *q, struct request *req)
1094 {
1095         struct scsi_device *sdev = q->queuedata;
1096         struct scsi_cmnd *cmd;
1097         int specials_only = 0;
1098
1099         /*
1100          * Just check to see if the device is online.  If it isn't, we
1101          * refuse to process any commands.  The device must be brought
1102          * online before trying any recovery commands
1103          */
1104         if (unlikely(!scsi_device_online(sdev))) {
1105                 sdev_printk(KERN_ERR, sdev,
1106                             "rejecting I/O to offline device\n");
1107                 goto kill;
1108         }
1109         if (unlikely(sdev->sdev_state != SDEV_RUNNING)) {
1110                 /* OK, we're not in a running state don't prep
1111                  * user commands */
1112                 if (sdev->sdev_state == SDEV_DEL) {
1113                         /* Device is fully deleted, no commands
1114                          * at all allowed down */
1115                         sdev_printk(KERN_ERR, sdev,
1116                                     "rejecting I/O to dead device\n");
1117                         goto kill;
1118                 }
1119                 /* OK, we only allow special commands (i.e. not
1120                  * user initiated ones */
1121                 specials_only = sdev->sdev_state;
1122         }
1123
1124         /*
1125          * Find the actual device driver associated with this command.
1126          * The SPECIAL requests are things like character device or
1127          * ioctls, which did not originate from ll_rw_blk.  Note that
1128          * the special field is also used to indicate the cmd for
1129          * the remainder of a partially fulfilled request that can
1130          * come up when there is a medium error.  We have to treat
1131          * these two cases differently.  We differentiate by looking
1132          * at request->cmd, as this tells us the real story.
1133          */
1134         if (req->flags & REQ_SPECIAL && req->special) {
1135                 cmd = req->special;
1136         } else if (req->flags & (REQ_CMD | REQ_BLOCK_PC)) {
1137
1138                 if(unlikely(specials_only) && !(req->flags & REQ_SPECIAL)) {
1139                         if(specials_only == SDEV_QUIESCE ||
1140                                         specials_only == SDEV_BLOCK)
1141                                 goto defer;
1142
1143                         sdev_printk(KERN_ERR, sdev,
1144                                     "rejecting I/O to device being removed\n");
1145                         goto kill;
1146                 }
1147
1148
1149                 /*
1150                  * Now try and find a command block that we can use.
1151                  */
1152                 if (!req->special) {
1153                         cmd = scsi_get_command(sdev, GFP_ATOMIC);
1154                         if (unlikely(!cmd))
1155                                 goto defer;
1156                 } else
1157                         cmd = req->special;
1158
1159                 /* pull a tag out of the request if we have one */
1160                 cmd->tag = req->tag;
1161         } else {
1162                 blk_dump_rq_flags(req, "SCSI bad req");
1163                 goto kill;
1164         }
1165
1166         /* note the overloading of req->special.  When the tag
1167          * is active it always means cmd.  If the tag goes
1168          * back for re-queueing, it may be reset */
1169         req->special = cmd;
1170         cmd->request = req;
1171
1172         /*
1173          * FIXME: drop the lock here because the functions below
1174          * expect to be called without the queue lock held.  Also,
1175          * previously, we dequeued the request before dropping the
1176          * lock.  We hope REQ_STARTED prevents anything untoward from
1177          * happening now.
1178          */
1179         if (req->flags & (REQ_CMD | REQ_BLOCK_PC)) {
1180                 int ret;
1181
1182                 /*
1183                  * This will do a couple of things:
1184                  *  1) Fill in the actual SCSI command.
1185                  *  2) Fill in any other upper-level specific fields
1186                  * (timeout).
1187                  *
1188                  * If this returns 0, it means that the request failed
1189                  * (reading past end of disk, reading offline device,
1190                  * etc).   This won't actually talk to the device, but
1191                  * some kinds of consistency checking may cause the
1192                  * request to be rejected immediately.
1193                  */
1194
1195                 /*
1196                  * This sets up the scatter-gather table (allocating if
1197                  * required).
1198                  */
1199                 ret = scsi_init_io(cmd);
1200                 switch(ret) {
1201                         /* For BLKPREP_KILL/DEFER the cmd was released */
1202                 case BLKPREP_KILL:
1203                         goto kill;
1204                 case BLKPREP_DEFER:
1205                         goto defer;
1206                 }
1207
1208                 /*
1209                  * Initialize the actual SCSI command for this request.
1210                  */
1211                 if (req->flags & REQ_BLOCK_PC) {
1212                         scsi_setup_blk_pc_cmnd(cmd);
1213                 } else if (req->rq_disk) {
1214                         struct scsi_driver *drv;
1215
1216                         drv = *(struct scsi_driver **)req->rq_disk->private_data;
1217                         if (unlikely(!drv->init_command(cmd))) {
1218                                 scsi_release_buffers(cmd);
1219                                 scsi_put_command(cmd);
1220                                 goto kill;
1221                         }
1222                 }
1223         }
1224
1225         /*
1226          * The request is now prepped, no need to come back here
1227          */
1228         req->flags |= REQ_DONTPREP;
1229         return BLKPREP_OK;
1230
1231  defer:
1232         /* If we defer, the elv_next_request() returns NULL, but the
1233          * queue must be restarted, so we plug here if no returning
1234          * command will automatically do that. */
1235         if (sdev->device_busy == 0)
1236                 blk_plug_device(q);
1237         return BLKPREP_DEFER;
1238  kill:
1239         req->errors = DID_NO_CONNECT << 16;
1240         return BLKPREP_KILL;
1241 }
1242
1243 /*
1244  * scsi_dev_queue_ready: if we can send requests to sdev, return 1 else
1245  * return 0.
1246  *
1247  * Called with the queue_lock held.
1248  */
1249 static inline int scsi_dev_queue_ready(struct request_queue *q,
1250                                   struct scsi_device *sdev)
1251 {
1252         if (sdev->device_busy >= sdev->queue_depth)
1253                 return 0;
1254         if (sdev->device_busy == 0 && sdev->device_blocked) {
1255                 /*
1256                  * unblock after device_blocked iterates to zero
1257                  */
1258                 if (--sdev->device_blocked == 0) {
1259                         SCSI_LOG_MLQUEUE(3,
1260                                    sdev_printk(KERN_INFO, sdev,
1261                                    "unblocking device at zero depth\n"));
1262                 } else {
1263                         blk_plug_device(q);
1264                         return 0;
1265                 }
1266         }
1267         if (sdev->device_blocked)
1268                 return 0;
1269
1270         return 1;
1271 }
1272
1273 /*
1274  * scsi_host_queue_ready: if we can send requests to shost, return 1 else
1275  * return 0. We must end up running the queue again whenever 0 is
1276  * returned, else IO can hang.
1277  *
1278  * Called with host_lock held.
1279  */
1280 static inline int scsi_host_queue_ready(struct request_queue *q,
1281                                    struct Scsi_Host *shost,
1282                                    struct scsi_device *sdev)
1283 {
1284         if (scsi_host_in_recovery(shost))
1285                 return 0;
1286         if (shost->host_busy == 0 && shost->host_blocked) {
1287                 /*
1288                  * unblock after host_blocked iterates to zero
1289                  */
1290                 if (--shost->host_blocked == 0) {
1291                         SCSI_LOG_MLQUEUE(3,
1292                                 printk("scsi%d unblocking host at zero depth\n",
1293                                         shost->host_no));
1294                 } else {
1295                         blk_plug_device(q);
1296                         return 0;
1297                 }
1298         }
1299         if ((shost->can_queue > 0 && shost->host_busy >= shost->can_queue) ||
1300             shost->host_blocked || shost->host_self_blocked) {
1301                 if (list_empty(&sdev->starved_entry))
1302                         list_add_tail(&sdev->starved_entry, &shost->starved_list);
1303                 return 0;
1304         }
1305
1306         /* We're OK to process the command, so we can't be starved */
1307         if (!list_empty(&sdev->starved_entry))
1308                 list_del_init(&sdev->starved_entry);
1309
1310         return 1;
1311 }
1312
1313 /*
1314  * Kill a request for a dead device
1315  */
1316 static void scsi_kill_request(struct request *req, request_queue_t *q)
1317 {
1318         struct scsi_cmnd *cmd = req->special;
1319         struct scsi_device *sdev = cmd->device;
1320         struct Scsi_Host *shost = sdev->host;
1321
1322         blkdev_dequeue_request(req);
1323
1324         if (unlikely(cmd == NULL)) {
1325                 printk(KERN_CRIT "impossible request in %s.\n",
1326                                  __FUNCTION__);
1327                 BUG();
1328         }
1329
1330         scsi_init_cmd_errh(cmd);
1331         cmd->result = DID_NO_CONNECT << 16;
1332         atomic_inc(&cmd->device->iorequest_cnt);
1333
1334         /*
1335          * SCSI request completion path will do scsi_device_unbusy(),
1336          * bump busy counts.  To bump the counters, we need to dance
1337          * with the locks as normal issue path does.
1338          */
1339         sdev->device_busy++;
1340         spin_unlock(sdev->request_queue->queue_lock);
1341         spin_lock(shost->host_lock);
1342         shost->host_busy++;
1343         spin_unlock(shost->host_lock);
1344         spin_lock(sdev->request_queue->queue_lock);
1345
1346         __scsi_done(cmd);
1347 }
1348
1349 static void scsi_softirq_done(struct request *rq)
1350 {
1351         struct scsi_cmnd *cmd = rq->completion_data;
1352         unsigned long wait_for = (cmd->allowed + 1) * cmd->timeout_per_command;
1353         int disposition;
1354
1355         INIT_LIST_HEAD(&cmd->eh_entry);
1356
1357         disposition = scsi_decide_disposition(cmd);
1358         if (disposition != SUCCESS &&
1359             time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) {
1360                 sdev_printk(KERN_ERR, cmd->device,
1361                             "timing out command, waited %lus\n",
1362                             wait_for/HZ);
1363                 disposition = SUCCESS;
1364         }
1365
1366         scsi_log_completion(cmd, disposition);
1367
1368         switch (disposition) {
1369                 case SUCCESS:
1370                         scsi_finish_command(cmd);
1371                         break;
1372                 case NEEDS_RETRY:
1373                         scsi_retry_command(cmd);
1374                         break;
1375                 case ADD_TO_MLQUEUE:
1376                         scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY);
1377                         break;
1378                 default:
1379                         if (!scsi_eh_scmd_add(cmd, 0))
1380                                 scsi_finish_command(cmd);
1381         }
1382 }
1383
1384 /*
1385  * Function:    scsi_request_fn()
1386  *
1387  * Purpose:     Main strategy routine for SCSI.
1388  *
1389  * Arguments:   q       - Pointer to actual queue.
1390  *
1391  * Returns:     Nothing
1392  *
1393  * Lock status: IO request lock assumed to be held when called.
1394  */
1395 static void scsi_request_fn(struct request_queue *q)
1396 {
1397         struct scsi_device *sdev = q->queuedata;
1398         struct Scsi_Host *shost;
1399         struct scsi_cmnd *cmd;
1400         struct request *req;
1401
1402         if (!sdev) {
1403                 printk("scsi: killing requests for dead queue\n");
1404                 while ((req = elv_next_request(q)) != NULL)
1405                         scsi_kill_request(req, q);
1406                 return;
1407         }
1408
1409         if(!get_device(&sdev->sdev_gendev))
1410                 /* We must be tearing the block queue down already */
1411                 return;
1412
1413         /*
1414          * To start with, we keep looping until the queue is empty, or until
1415          * the host is no longer able to accept any more requests.
1416          */
1417         shost = sdev->host;
1418         while (!blk_queue_plugged(q)) {
1419                 int rtn;
1420                 /*
1421                  * get next queueable request.  We do this early to make sure
1422                  * that the request is fully prepared even if we cannot
1423                  * accept it.
1424                  */
1425                 req = elv_next_request(q);
1426                 if (!req || !scsi_dev_queue_ready(q, sdev))
1427                         break;
1428
1429                 if (unlikely(!scsi_device_online(sdev))) {
1430                         sdev_printk(KERN_ERR, sdev,
1431                                     "rejecting I/O to offline device\n");
1432                         scsi_kill_request(req, q);
1433                         continue;
1434                 }
1435
1436
1437                 /*
1438                  * Remove the request from the request list.
1439                  */
1440                 if (!(blk_queue_tagged(q) && !blk_queue_start_tag(q, req)))
1441                         blkdev_dequeue_request(req);
1442                 sdev->device_busy++;
1443
1444                 spin_unlock(q->queue_lock);
1445                 cmd = req->special;
1446                 if (unlikely(cmd == NULL)) {
1447                         printk(KERN_CRIT "impossible request in %s.\n"
1448                                          "please mail a stack trace to "
1449                                          "linux-scsi@vger.kernel.org",
1450                                          __FUNCTION__);
1451                         BUG();
1452                 }
1453                 spin_lock(shost->host_lock);
1454
1455                 if (!scsi_host_queue_ready(q, shost, sdev))
1456                         goto not_ready;
1457                 if (sdev->single_lun) {
1458                         if (scsi_target(sdev)->starget_sdev_user &&
1459                             scsi_target(sdev)->starget_sdev_user != sdev)
1460                                 goto not_ready;
1461                         scsi_target(sdev)->starget_sdev_user = sdev;
1462                 }
1463                 shost->host_busy++;
1464
1465                 /*
1466                  * XXX(hch): This is rather suboptimal, scsi_dispatch_cmd will
1467                  *              take the lock again.
1468                  */
1469                 spin_unlock_irq(shost->host_lock);
1470
1471                 /*
1472                  * Finally, initialize any error handling parameters, and set up
1473                  * the timers for timeouts.
1474                  */
1475                 scsi_init_cmd_errh(cmd);
1476
1477                 /*
1478                  * Dispatch the command to the low-level driver.
1479                  */
1480                 rtn = scsi_dispatch_cmd(cmd);
1481                 spin_lock_irq(q->queue_lock);
1482                 if(rtn) {
1483                         /* we're refusing the command; because of
1484                          * the way locks get dropped, we need to
1485                          * check here if plugging is required */
1486                         if(sdev->device_busy == 0)
1487                                 blk_plug_device(q);
1488
1489                         break;
1490                 }
1491         }
1492
1493         goto out;
1494
1495  not_ready:
1496         spin_unlock_irq(shost->host_lock);
1497
1498         /*
1499          * lock q, handle tag, requeue req, and decrement device_busy. We
1500          * must return with queue_lock held.
1501          *
1502          * Decrementing device_busy without checking it is OK, as all such
1503          * cases (host limits or settings) should run the queue at some
1504          * later time.
1505          */
1506         spin_lock_irq(q->queue_lock);
1507         blk_requeue_request(q, req);
1508         sdev->device_busy--;
1509         if(sdev->device_busy == 0)
1510                 blk_plug_device(q);
1511  out:
1512         /* must be careful here...if we trigger the ->remove() function
1513          * we cannot be holding the q lock */
1514         spin_unlock_irq(q->queue_lock);
1515         put_device(&sdev->sdev_gendev);
1516         spin_lock_irq(q->queue_lock);
1517 }
1518
1519 u64 scsi_calculate_bounce_limit(struct Scsi_Host *shost)
1520 {
1521         struct device *host_dev;
1522         u64 bounce_limit = 0xffffffff;
1523
1524         if (shost->unchecked_isa_dma)
1525                 return BLK_BOUNCE_ISA;
1526         /*
1527          * Platforms with virtual-DMA translation
1528          * hardware have no practical limit.
1529          */
1530         if (!PCI_DMA_BUS_IS_PHYS)
1531                 return BLK_BOUNCE_ANY;
1532
1533         host_dev = scsi_get_device(shost);
1534         if (host_dev && host_dev->dma_mask)
1535                 bounce_limit = *host_dev->dma_mask;
1536
1537         return bounce_limit;
1538 }
1539 EXPORT_SYMBOL(scsi_calculate_bounce_limit);
1540
1541 struct request_queue *scsi_alloc_queue(struct scsi_device *sdev)
1542 {
1543         struct Scsi_Host *shost = sdev->host;
1544         struct request_queue *q;
1545
1546         q = blk_init_queue(scsi_request_fn, NULL);
1547         if (!q)
1548                 return NULL;
1549
1550         blk_queue_prep_rq(q, scsi_prep_fn);
1551
1552         blk_queue_max_hw_segments(q, shost->sg_tablesize);
1553         blk_queue_max_phys_segments(q, SCSI_MAX_PHYS_SEGMENTS);
1554         blk_queue_max_sectors(q, shost->max_sectors);
1555         blk_queue_bounce_limit(q, scsi_calculate_bounce_limit(shost));
1556         blk_queue_segment_boundary(q, shost->dma_boundary);
1557         blk_queue_issue_flush_fn(q, scsi_issue_flush_fn);
1558         blk_queue_softirq_done(q, scsi_softirq_done);
1559
1560         if (!shost->use_clustering)
1561                 clear_bit(QUEUE_FLAG_CLUSTER, &q->queue_flags);
1562         return q;
1563 }
1564
1565 void scsi_free_queue(struct request_queue *q)
1566 {
1567         blk_cleanup_queue(q);
1568 }
1569
1570 /*
1571  * Function:    scsi_block_requests()
1572  *
1573  * Purpose:     Utility function used by low-level drivers to prevent further
1574  *              commands from being queued to the device.
1575  *
1576  * Arguments:   shost       - Host in question
1577  *
1578  * Returns:     Nothing
1579  *
1580  * Lock status: No locks are assumed held.
1581  *
1582  * Notes:       There is no timer nor any other means by which the requests
1583  *              get unblocked other than the low-level driver calling
1584  *              scsi_unblock_requests().
1585  */
1586 void scsi_block_requests(struct Scsi_Host *shost)
1587 {
1588         shost->host_self_blocked = 1;
1589 }
1590 EXPORT_SYMBOL(scsi_block_requests);
1591
1592 /*
1593  * Function:    scsi_unblock_requests()
1594  *
1595  * Purpose:     Utility function used by low-level drivers to allow further
1596  *              commands from being queued to the device.
1597  *
1598  * Arguments:   shost       - Host in question
1599  *
1600  * Returns:     Nothing
1601  *
1602  * Lock status: No locks are assumed held.
1603  *
1604  * Notes:       There is no timer nor any other means by which the requests
1605  *              get unblocked other than the low-level driver calling
1606  *              scsi_unblock_requests().
1607  *
1608  *              This is done as an API function so that changes to the
1609  *              internals of the scsi mid-layer won't require wholesale
1610  *              changes to drivers that use this feature.
1611  */
1612 void scsi_unblock_requests(struct Scsi_Host *shost)
1613 {
1614         shost->host_self_blocked = 0;
1615         scsi_run_host_queues(shost);
1616 }
1617 EXPORT_SYMBOL(scsi_unblock_requests);
1618
1619 int __init scsi_init_queue(void)
1620 {
1621         int i;
1622
1623         scsi_io_context_cache = kmem_cache_create("scsi_io_context",
1624                                         sizeof(struct scsi_io_context),
1625                                         0, 0, NULL, NULL);
1626         if (!scsi_io_context_cache) {
1627                 printk(KERN_ERR "SCSI: can't init scsi io context cache\n");
1628                 return -ENOMEM;
1629         }
1630
1631         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1632                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1633                 int size = sgp->size * sizeof(struct scatterlist);
1634
1635                 sgp->slab = kmem_cache_create(sgp->name, size, 0,
1636                                 SLAB_HWCACHE_ALIGN, NULL, NULL);
1637                 if (!sgp->slab) {
1638                         printk(KERN_ERR "SCSI: can't init sg slab %s\n",
1639                                         sgp->name);
1640                 }
1641
1642                 sgp->pool = mempool_create_slab_pool(SG_MEMPOOL_SIZE,
1643                                                      sgp->slab);
1644                 if (!sgp->pool) {
1645                         printk(KERN_ERR "SCSI: can't init sg mempool %s\n",
1646                                         sgp->name);
1647                 }
1648         }
1649
1650         return 0;
1651 }
1652
1653 void scsi_exit_queue(void)
1654 {
1655         int i;
1656
1657         kmem_cache_destroy(scsi_io_context_cache);
1658
1659         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1660                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1661                 mempool_destroy(sgp->pool);
1662                 kmem_cache_destroy(sgp->slab);
1663         }
1664 }
1665
1666 /**
1667  *      scsi_mode_select - issue a mode select
1668  *      @sdev:  SCSI device to be queried
1669  *      @pf:    Page format bit (1 == standard, 0 == vendor specific)
1670  *      @sp:    Save page bit (0 == don't save, 1 == save)
1671  *      @modepage: mode page being requested
1672  *      @buffer: request buffer (may not be smaller than eight bytes)
1673  *      @len:   length of request buffer.
1674  *      @timeout: command timeout
1675  *      @retries: number of retries before failing
1676  *      @data: returns a structure abstracting the mode header data
1677  *      @sense: place to put sense data (or NULL if no sense to be collected).
1678  *              must be SCSI_SENSE_BUFFERSIZE big.
1679  *
1680  *      Returns zero if successful; negative error number or scsi
1681  *      status on error
1682  *
1683  */
1684 int
1685 scsi_mode_select(struct scsi_device *sdev, int pf, int sp, int modepage,
1686                  unsigned char *buffer, int len, int timeout, int retries,
1687                  struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1688 {
1689         unsigned char cmd[10];
1690         unsigned char *real_buffer;
1691         int ret;
1692
1693         memset(cmd, 0, sizeof(cmd));
1694         cmd[1] = (pf ? 0x10 : 0) | (sp ? 0x01 : 0);
1695
1696         if (sdev->use_10_for_ms) {
1697                 if (len > 65535)
1698                         return -EINVAL;
1699                 real_buffer = kmalloc(8 + len, GFP_KERNEL);
1700                 if (!real_buffer)
1701                         return -ENOMEM;
1702                 memcpy(real_buffer + 8, buffer, len);
1703                 len += 8;
1704                 real_buffer[0] = 0;
1705                 real_buffer[1] = 0;
1706                 real_buffer[2] = data->medium_type;
1707                 real_buffer[3] = data->device_specific;
1708                 real_buffer[4] = data->longlba ? 0x01 : 0;
1709                 real_buffer[5] = 0;
1710                 real_buffer[6] = data->block_descriptor_length >> 8;
1711                 real_buffer[7] = data->block_descriptor_length;
1712
1713                 cmd[0] = MODE_SELECT_10;
1714                 cmd[7] = len >> 8;
1715                 cmd[8] = len;
1716         } else {
1717                 if (len > 255 || data->block_descriptor_length > 255 ||
1718                     data->longlba)
1719                         return -EINVAL;
1720
1721                 real_buffer = kmalloc(4 + len, GFP_KERNEL);
1722                 if (!real_buffer)
1723                         return -ENOMEM;
1724                 memcpy(real_buffer + 4, buffer, len);
1725                 len += 4;
1726                 real_buffer[0] = 0;
1727                 real_buffer[1] = data->medium_type;
1728                 real_buffer[2] = data->device_specific;
1729                 real_buffer[3] = data->block_descriptor_length;
1730
1731
1732                 cmd[0] = MODE_SELECT;
1733                 cmd[4] = len;
1734         }
1735
1736         ret = scsi_execute_req(sdev, cmd, DMA_TO_DEVICE, real_buffer, len,
1737                                sshdr, timeout, retries);
1738         kfree(real_buffer);
1739         return ret;
1740 }
1741 EXPORT_SYMBOL_GPL(scsi_mode_select);
1742
1743 /**
1744  *      scsi_mode_sense - issue a mode sense, falling back from 10 to
1745  *              six bytes if necessary.
1746  *      @sdev:  SCSI device to be queried
1747  *      @dbd:   set if mode sense will allow block descriptors to be returned
1748  *      @modepage: mode page being requested
1749  *      @buffer: request buffer (may not be smaller than eight bytes)
1750  *      @len:   length of request buffer.
1751  *      @timeout: command timeout
1752  *      @retries: number of retries before failing
1753  *      @data: returns a structure abstracting the mode header data
1754  *      @sense: place to put sense data (or NULL if no sense to be collected).
1755  *              must be SCSI_SENSE_BUFFERSIZE big.
1756  *
1757  *      Returns zero if unsuccessful, or the header offset (either 4
1758  *      or 8 depending on whether a six or ten byte command was
1759  *      issued) if successful.
1760  **/
1761 int
1762 scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
1763                   unsigned char *buffer, int len, int timeout, int retries,
1764                   struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1765 {
1766         unsigned char cmd[12];
1767         int use_10_for_ms;
1768         int header_length;
1769         int result;
1770         struct scsi_sense_hdr my_sshdr;
1771
1772         memset(data, 0, sizeof(*data));
1773         memset(&cmd[0], 0, 12);
1774         cmd[1] = dbd & 0x18;    /* allows DBD and LLBA bits */
1775         cmd[2] = modepage;
1776
1777         /* caller might not be interested in sense, but we need it */
1778         if (!sshdr)
1779                 sshdr = &my_sshdr;
1780
1781  retry:
1782         use_10_for_ms = sdev->use_10_for_ms;
1783
1784         if (use_10_for_ms) {
1785                 if (len < 8)
1786                         len = 8;
1787
1788                 cmd[0] = MODE_SENSE_10;
1789                 cmd[8] = len;
1790                 header_length = 8;
1791         } else {
1792                 if (len < 4)
1793                         len = 4;
1794
1795                 cmd[0] = MODE_SENSE;
1796                 cmd[4] = len;
1797                 header_length = 4;
1798         }
1799
1800         memset(buffer, 0, len);
1801
1802         result = scsi_execute_req(sdev, cmd, DMA_FROM_DEVICE, buffer, len,
1803                                   sshdr, timeout, retries);
1804
1805         /* This code looks awful: what it's doing is making sure an
1806          * ILLEGAL REQUEST sense return identifies the actual command
1807          * byte as the problem.  MODE_SENSE commands can return
1808          * ILLEGAL REQUEST if the code page isn't supported */
1809
1810         if (use_10_for_ms && !scsi_status_is_good(result) &&
1811             (driver_byte(result) & DRIVER_SENSE)) {
1812                 if (scsi_sense_valid(sshdr)) {
1813                         if ((sshdr->sense_key == ILLEGAL_REQUEST) &&
1814                             (sshdr->asc == 0x20) && (sshdr->ascq == 0)) {
1815                                 /*
1816                                  * Invalid command operation code
1817                                  */
1818                                 sdev->use_10_for_ms = 0;
1819                                 goto retry;
1820                         }
1821                 }
1822         }
1823
1824         if(scsi_status_is_good(result)) {
1825                 if (unlikely(buffer[0] == 0x86 && buffer[1] == 0x0b &&
1826                              (modepage == 6 || modepage == 8))) {
1827                         /* Initio breakage? */
1828                         header_length = 0;
1829                         data->length = 13;
1830                         data->medium_type = 0;
1831                         data->device_specific = 0;
1832                         data->longlba = 0;
1833                         data->block_descriptor_length = 0;
1834                 } else if(use_10_for_ms) {
1835                         data->length = buffer[0]*256 + buffer[1] + 2;
1836                         data->medium_type = buffer[2];
1837                         data->device_specific = buffer[3];
1838                         data->longlba = buffer[4] & 0x01;
1839                         data->block_descriptor_length = buffer[6]*256
1840                                 + buffer[7];
1841                 } else {
1842                         data->length = buffer[0] + 1;
1843                         data->medium_type = buffer[1];
1844                         data->device_specific = buffer[2];
1845                         data->block_descriptor_length = buffer[3];
1846                 }
1847                 data->header_length = header_length;
1848         }
1849
1850         return result;
1851 }
1852 EXPORT_SYMBOL(scsi_mode_sense);
1853
1854 int
1855 scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries)
1856 {
1857         char cmd[] = {
1858                 TEST_UNIT_READY, 0, 0, 0, 0, 0,
1859         };
1860         struct scsi_sense_hdr sshdr;
1861         int result;
1862
1863         result = scsi_execute_req(sdev, cmd, DMA_NONE, NULL, 0, &sshdr,
1864                                   timeout, retries);
1865
1866         if ((driver_byte(result) & DRIVER_SENSE) && sdev->removable) {
1867
1868                 if ((scsi_sense_valid(&sshdr)) &&
1869                     ((sshdr.sense_key == UNIT_ATTENTION) ||
1870                      (sshdr.sense_key == NOT_READY))) {
1871                         sdev->changed = 1;
1872                         result = 0;
1873                 }
1874         }
1875         return result;
1876 }
1877 EXPORT_SYMBOL(scsi_test_unit_ready);
1878
1879 /**
1880  *      scsi_device_set_state - Take the given device through the device
1881  *              state model.
1882  *      @sdev:  scsi device to change the state of.
1883  *      @state: state to change to.
1884  *
1885  *      Returns zero if unsuccessful or an error if the requested
1886  *      transition is illegal.
1887  **/
1888 int
1889 scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
1890 {
1891         enum scsi_device_state oldstate = sdev->sdev_state;
1892
1893         if (state == oldstate)
1894                 return 0;
1895
1896         switch (state) {
1897         case SDEV_CREATED:
1898                 /* There are no legal states that come back to
1899                  * created.  This is the manually initialised start
1900                  * state */
1901                 goto illegal;
1902
1903         case SDEV_RUNNING:
1904                 switch (oldstate) {
1905                 case SDEV_CREATED:
1906                 case SDEV_OFFLINE:
1907                 case SDEV_QUIESCE:
1908                 case SDEV_BLOCK:
1909                         break;
1910                 default:
1911                         goto illegal;
1912                 }
1913                 break;
1914
1915         case SDEV_QUIESCE:
1916                 switch (oldstate) {
1917                 case SDEV_RUNNING:
1918                 case SDEV_OFFLINE:
1919                         break;
1920                 default:
1921                         goto illegal;
1922                 }
1923                 break;
1924
1925         case SDEV_OFFLINE:
1926                 switch (oldstate) {
1927                 case SDEV_CREATED:
1928                 case SDEV_RUNNING:
1929                 case SDEV_QUIESCE:
1930                 case SDEV_BLOCK:
1931                         break;
1932                 default:
1933                         goto illegal;
1934                 }
1935                 break;
1936
1937         case SDEV_BLOCK:
1938                 switch (oldstate) {
1939                 case SDEV_CREATED:
1940                 case SDEV_RUNNING:
1941                         break;
1942                 default:
1943                         goto illegal;
1944                 }
1945                 break;
1946
1947         case SDEV_CANCEL:
1948                 switch (oldstate) {
1949                 case SDEV_CREATED:
1950                 case SDEV_RUNNING:
1951                 case SDEV_QUIESCE:
1952                 case SDEV_OFFLINE:
1953                 case SDEV_BLOCK:
1954                         break;
1955                 default:
1956                         goto illegal;
1957                 }
1958                 break;
1959
1960         case SDEV_DEL:
1961                 switch (oldstate) {
1962                 case SDEV_CREATED:
1963                 case SDEV_RUNNING:
1964                 case SDEV_OFFLINE:
1965                 case SDEV_CANCEL:
1966                         break;
1967                 default:
1968                         goto illegal;
1969                 }
1970                 break;
1971
1972         }
1973         sdev->sdev_state = state;
1974         return 0;
1975
1976  illegal:
1977         SCSI_LOG_ERROR_RECOVERY(1,
1978                                 sdev_printk(KERN_ERR, sdev,
1979                                             "Illegal state transition %s->%s\n",
1980                                             scsi_device_state_name(oldstate),
1981                                             scsi_device_state_name(state))
1982                                 );
1983         return -EINVAL;
1984 }
1985 EXPORT_SYMBOL(scsi_device_set_state);
1986
1987 /**
1988  *      scsi_device_quiesce - Block user issued commands.
1989  *      @sdev:  scsi device to quiesce.
1990  *
1991  *      This works by trying to transition to the SDEV_QUIESCE state
1992  *      (which must be a legal transition).  When the device is in this
1993  *      state, only special requests will be accepted, all others will
1994  *      be deferred.  Since special requests may also be requeued requests,
1995  *      a successful return doesn't guarantee the device will be
1996  *      totally quiescent.
1997  *
1998  *      Must be called with user context, may sleep.
1999  *
2000  *      Returns zero if unsuccessful or an error if not.
2001  **/
2002 int
2003 scsi_device_quiesce(struct scsi_device *sdev)
2004 {
2005         int err = scsi_device_set_state(sdev, SDEV_QUIESCE);
2006         if (err)
2007                 return err;
2008
2009         scsi_run_queue(sdev->request_queue);
2010         while (sdev->device_busy) {
2011                 msleep_interruptible(200);
2012                 scsi_run_queue(sdev->request_queue);
2013         }
2014         return 0;
2015 }
2016 EXPORT_SYMBOL(scsi_device_quiesce);
2017
2018 /**
2019  *      scsi_device_resume - Restart user issued commands to a quiesced device.
2020  *      @sdev:  scsi device to resume.
2021  *
2022  *      Moves the device from quiesced back to running and restarts the
2023  *      queues.
2024  *
2025  *      Must be called with user context, may sleep.
2026  **/
2027 void
2028 scsi_device_resume(struct scsi_device *sdev)
2029 {
2030         if(scsi_device_set_state(sdev, SDEV_RUNNING))
2031                 return;
2032         scsi_run_queue(sdev->request_queue);
2033 }
2034 EXPORT_SYMBOL(scsi_device_resume);
2035
2036 static void
2037 device_quiesce_fn(struct scsi_device *sdev, void *data)
2038 {
2039         scsi_device_quiesce(sdev);
2040 }
2041
2042 void
2043 scsi_target_quiesce(struct scsi_target *starget)
2044 {
2045         starget_for_each_device(starget, NULL, device_quiesce_fn);
2046 }
2047 EXPORT_SYMBOL(scsi_target_quiesce);
2048
2049 static void
2050 device_resume_fn(struct scsi_device *sdev, void *data)
2051 {
2052         scsi_device_resume(sdev);
2053 }
2054
2055 void
2056 scsi_target_resume(struct scsi_target *starget)
2057 {
2058         starget_for_each_device(starget, NULL, device_resume_fn);
2059 }
2060 EXPORT_SYMBOL(scsi_target_resume);
2061
2062 /**
2063  * scsi_internal_device_block - internal function to put a device
2064  *                              temporarily into the SDEV_BLOCK state
2065  * @sdev:       device to block
2066  *
2067  * Block request made by scsi lld's to temporarily stop all
2068  * scsi commands on the specified device.  Called from interrupt
2069  * or normal process context.
2070  *
2071  * Returns zero if successful or error if not
2072  *
2073  * Notes:
2074  *      This routine transitions the device to the SDEV_BLOCK state
2075  *      (which must be a legal transition).  When the device is in this
2076  *      state, all commands are deferred until the scsi lld reenables
2077  *      the device with scsi_device_unblock or device_block_tmo fires.
2078  *      This routine assumes the host_lock is held on entry.
2079  **/
2080 int
2081 scsi_internal_device_block(struct scsi_device *sdev)
2082 {
2083         request_queue_t *q = sdev->request_queue;
2084         unsigned long flags;
2085         int err = 0;
2086
2087         err = scsi_device_set_state(sdev, SDEV_BLOCK);
2088         if (err)
2089                 return err;
2090
2091         /*
2092          * The device has transitioned to SDEV_BLOCK.  Stop the
2093          * block layer from calling the midlayer with this device's
2094          * request queue.
2095          */
2096         spin_lock_irqsave(q->queue_lock, flags);
2097         blk_stop_queue(q);
2098         spin_unlock_irqrestore(q->queue_lock, flags);
2099
2100         return 0;
2101 }
2102 EXPORT_SYMBOL_GPL(scsi_internal_device_block);
2103
2104 /**
2105  * scsi_internal_device_unblock - resume a device after a block request
2106  * @sdev:       device to resume
2107  *
2108  * Called by scsi lld's or the midlayer to restart the device queue
2109  * for the previously suspended scsi device.  Called from interrupt or
2110  * normal process context.
2111  *
2112  * Returns zero if successful or error if not.
2113  *
2114  * Notes:
2115  *      This routine transitions the device to the SDEV_RUNNING state
2116  *      (which must be a legal transition) allowing the midlayer to
2117  *      goose the queue for this device.  This routine assumes the
2118  *      host_lock is held upon entry.
2119  **/
2120 int
2121 scsi_internal_device_unblock(struct scsi_device *sdev)
2122 {
2123         request_queue_t *q = sdev->request_queue;
2124         int err;
2125         unsigned long flags;
2126
2127         /*
2128          * Try to transition the scsi device to SDEV_RUNNING
2129          * and goose the device queue if successful.
2130          */
2131         err = scsi_device_set_state(sdev, SDEV_RUNNING);
2132         if (err)
2133                 return err;
2134
2135         spin_lock_irqsave(q->queue_lock, flags);
2136         blk_start_queue(q);
2137         spin_unlock_irqrestore(q->queue_lock, flags);
2138
2139         return 0;
2140 }
2141 EXPORT_SYMBOL_GPL(scsi_internal_device_unblock);
2142
2143 static void
2144 device_block(struct scsi_device *sdev, void *data)
2145 {
2146         scsi_internal_device_block(sdev);
2147 }
2148
2149 static int
2150 target_block(struct device *dev, void *data)
2151 {
2152         if (scsi_is_target_device(dev))
2153                 starget_for_each_device(to_scsi_target(dev), NULL,
2154                                         device_block);
2155         return 0;
2156 }
2157
2158 void
2159 scsi_target_block(struct device *dev)
2160 {
2161         if (scsi_is_target_device(dev))
2162                 starget_for_each_device(to_scsi_target(dev), NULL,
2163                                         device_block);
2164         else
2165                 device_for_each_child(dev, NULL, target_block);
2166 }
2167 EXPORT_SYMBOL_GPL(scsi_target_block);
2168
2169 static void
2170 device_unblock(struct scsi_device *sdev, void *data)
2171 {
2172         scsi_internal_device_unblock(sdev);
2173 }
2174
2175 static int
2176 target_unblock(struct device *dev, void *data)
2177 {
2178         if (scsi_is_target_device(dev))
2179                 starget_for_each_device(to_scsi_target(dev), NULL,
2180                                         device_unblock);
2181         return 0;
2182 }
2183
2184 void
2185 scsi_target_unblock(struct device *dev)
2186 {
2187         if (scsi_is_target_device(dev))
2188                 starget_for_each_device(to_scsi_target(dev), NULL,
2189                                         device_unblock);
2190         else
2191                 device_for_each_child(dev, NULL, target_unblock);
2192 }
2193 EXPORT_SYMBOL_GPL(scsi_target_unblock);
2194
2195 /**
2196  * scsi_kmap_atomic_sg - find and atomically map an sg-elemnt
2197  * @sg:         scatter-gather list
2198  * @sg_count:   number of segments in sg
2199  * @offset:     offset in bytes into sg, on return offset into the mapped area
2200  * @len:        bytes to map, on return number of bytes mapped
2201  *
2202  * Returns virtual address of the start of the mapped page
2203  */
2204 void *scsi_kmap_atomic_sg(struct scatterlist *sg, int sg_count,
2205                           size_t *offset, size_t *len)
2206 {
2207         int i;
2208         size_t sg_len = 0, len_complete = 0;
2209         struct page *page;
2210
2211         for (i = 0; i < sg_count; i++) {
2212                 len_complete = sg_len; /* Complete sg-entries */
2213                 sg_len += sg[i].length;
2214                 if (sg_len > *offset)
2215                         break;
2216         }
2217
2218         if (unlikely(i == sg_count)) {
2219                 printk(KERN_ERR "%s: Bytes in sg: %zu, requested offset %zu, "
2220                         "elements %d\n",
2221                        __FUNCTION__, sg_len, *offset, sg_count);
2222                 WARN_ON(1);
2223                 return NULL;
2224         }
2225
2226         /* Offset starting from the beginning of first page in this sg-entry */
2227         *offset = *offset - len_complete + sg[i].offset;
2228
2229         /* Assumption: contiguous pages can be accessed as "page + i" */
2230         page = nth_page(sg[i].page, (*offset >> PAGE_SHIFT));
2231         *offset &= ~PAGE_MASK;
2232
2233         /* Bytes in this sg-entry from *offset to the end of the page */
2234         sg_len = PAGE_SIZE - *offset;
2235         if (*len > sg_len)
2236                 *len = sg_len;
2237
2238         return kmap_atomic(page, KM_BIO_SRC_IRQ);
2239 }
2240 EXPORT_SYMBOL(scsi_kmap_atomic_sg);
2241
2242 /**
2243  * scsi_kunmap_atomic_sg - atomically unmap a virtual address, previously
2244  *                         mapped with scsi_kmap_atomic_sg
2245  * @virt:       virtual address to be unmapped
2246  */
2247 void scsi_kunmap_atomic_sg(void *virt)
2248 {
2249         kunmap_atomic(virt, KM_BIO_SRC_IRQ);
2250 }
2251 EXPORT_SYMBOL(scsi_kunmap_atomic_sg);