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