drivers/scsi/scsi_error.c

   1 /*
   2  *  scsi_error.c Copyright (C) 1997 Eric Youngdale
   3  *
   4  *  SCSI error/timeout handling
   5  *      Initial versions: Eric Youngdale.  Based upon conversations with
   6  *                        Leonard Zubkoff and David Miller at Linux Expo,
   7  *                        ideas originating from all over the place.
   8  *
   9  *      Restructured scsi_unjam_host and associated functions.
  10  *      September 04, 2002 Mike Anderson (andmike@us.ibm.com)
  11  *
  12  *      Forward port of Russell King's (rmk@arm.linux.org.uk) changes and
  13  *      minor  cleanups.
  14  *      September 30, 2002 Mike Anderson (andmike@us.ibm.com)
  15  */
  16
  17 #include <linux/module.h>
  18 #include <linux/sched.h>
  19 #include <linux/timer.h>
  20 #include <linux/string.h>
  21 #include <linux/slab.h>
  22 #include <linux/kernel.h>
  23 #include <linux/kthread.h>
  24 #include <linux/interrupt.h>
  25 #include <linux/blkdev.h>
  26 #include <linux/delay.h>
  27
  28 #include <scsi/scsi.h>
  29 #include <scsi/scsi_dbg.h>
  30 #include <scsi/scsi_device.h>
  31 #include <scsi/scsi_eh.h>
  32 #include <scsi/scsi_transport.h>
  33 #include <scsi/scsi_host.h>
  34 #include <scsi/scsi_ioctl.h>
  35 #include <scsi/scsi_request.h>
  36
  37 #include "scsi_priv.h"
  38 #include "scsi_logging.h"
  39
  40 #define SENSE_TIMEOUT           (10*HZ)
  41 #define START_UNIT_TIMEOUT      (30*HZ)
  42
  43 /*
  44  * These should *probably* be handled by the host itself.
  45  * Since it is allowed to sleep, it probably should.
  46  */
  47 #define BUS_RESET_SETTLE_TIME   (10)
  48 #define HOST_RESET_SETTLE_TIME  (10)
  49
  50 /* called with shost->host_lock held */
  51 void scsi_eh_wakeup(struct Scsi_Host *shost)
  52 {
  53         if (shost->host_busy == shost->host_failed) {
  54                 wake_up_process(shost->ehandler);
  55                 SCSI_LOG_ERROR_RECOVERY(5,
  56                                 printk("Waking error handler thread\n"));
  57         }
  58 }
  59
  60 /**
  61  * scsi_schedule_eh - schedule EH for SCSI host
  62  * @shost:      SCSI host to invoke error handling on.
  63  *
  64  * Schedule SCSI EH without scmd.
  65  **/
  66 void scsi_schedule_eh(struct Scsi_Host *shost)
  67 {
  68         unsigned long flags;
  69
  70         spin_lock_irqsave(shost->host_lock, flags);
  71
  72         if (scsi_host_set_state(shost, SHOST_RECOVERY) == 0 ||
  73             scsi_host_set_state(shost, SHOST_CANCEL_RECOVERY) == 0) {
  74                 shost->host_eh_scheduled++;
  75                 scsi_eh_wakeup(shost);
  76         }
  77
  78         spin_unlock_irqrestore(shost->host_lock, flags);
  79 }
  80 EXPORT_SYMBOL_GPL(scsi_schedule_eh);
  81
  82 /**
  83  * scsi_eh_scmd_add - add scsi cmd to error handling.
  84  * @scmd:       scmd to run eh on.
  85  * @eh_flag:    optional SCSI_EH flag.
  86  *
  87  * Return value:
  88  *      0 on failure.
  89  **/
  90 int scsi_eh_scmd_add(struct scsi_cmnd *scmd, int eh_flag)
  91 {
  92         struct Scsi_Host *shost = scmd->device->host;
  93         unsigned long flags;
  94         int ret = 0;
  95
  96         if (!shost->ehandler)
  97                 return 0;
  98
  99         spin_lock_irqsave(shost->host_lock, flags);
 100         if (scsi_host_set_state(shost, SHOST_RECOVERY))
 101                 if (scsi_host_set_state(shost, SHOST_CANCEL_RECOVERY))
 102                         goto out_unlock;
 103
 104         ret = 1;
 105         scmd->eh_eflags |= eh_flag;
 106         list_add_tail(&scmd->eh_entry, &shost->eh_cmd_q);
 107         shost->host_failed++;
 108         scsi_eh_wakeup(shost);
 109  out_unlock:
 110         spin_unlock_irqrestore(shost->host_lock, flags);
 111         return ret;
 112 }
 113
 114 /**
 115  * scsi_add_timer - Start timeout timer for a single scsi command.
 116  * @scmd:       scsi command that is about to start running.
 117  * @timeout:    amount of time to allow this command to run.
 118  * @complete:   timeout function to call if timer isn't canceled.
 119  *
 120  * Notes:
 121  *    This should be turned into an inline function.  Each scsi command
 122  *    has its own timer, and as it is added to the queue, we set up the
 123  *    timer.  When the command completes, we cancel the timer.
 124  **/
 125 void scsi_add_timer(struct scsi_cmnd *scmd, int timeout,
 126                     void (*complete)(struct scsi_cmnd *))
 127 {
 128
 129         /*
 130          * If the clock was already running for this command, then
 131          * first delete the timer.  The timer handling code gets rather
 132          * confused if we don't do this.
 133          */
 134         if (scmd->eh_timeout.function)
 135                 del_timer(&scmd->eh_timeout);
 136
 137         scmd->eh_timeout.data = (unsigned long)scmd;
 138         scmd->eh_timeout.expires = jiffies + timeout;
 139         scmd->eh_timeout.function = (void (*)(unsigned long)) complete;
 140
 141         SCSI_LOG_ERROR_RECOVERY(5, printk("%s: scmd: %p, time:"
 142                                           " %d, (%p)\n", __FUNCTION__,
 143                                           scmd, timeout, complete));
 144
 145         add_timer(&scmd->eh_timeout);
 146 }
 147
 148 /**
 149  * scsi_delete_timer - Delete/cancel timer for a given function.
 150  * @scmd:       Cmd that we are canceling timer for
 151  *
 152  * Notes:
 153  *     This should be turned into an inline function.
 154  *
 155  * Return value:
 156  *     1 if we were able to detach the timer.  0 if we blew it, and the
 157  *     timer function has already started to run.
 158  **/
 159 int scsi_delete_timer(struct scsi_cmnd *scmd)
 160 {
 161         int rtn;
 162
 163         rtn = del_timer(&scmd->eh_timeout);
 164
 165         SCSI_LOG_ERROR_RECOVERY(5, printk("%s: scmd: %p,"
 166                                          " rtn: %d\n", __FUNCTION__,
 167                                          scmd, rtn));
 168
 169         scmd->eh_timeout.data = (unsigned long)NULL;
 170         scmd->eh_timeout.function = NULL;
 171
 172         return rtn;
 173 }
 174
 175 /**
 176  * scsi_times_out - Timeout function for normal scsi commands.
 177  * @scmd:       Cmd that is timing out.
 178  *
 179  * Notes:
 180  *     We do not need to lock this.  There is the potential for a race
 181  *     only in that the normal completion handling might run, but if the
 182  *     normal completion function determines that the timer has already
 183  *     fired, then it mustn't do anything.
 184  **/
 185 void scsi_times_out(struct scsi_cmnd *scmd)
 186 {
 187         scsi_log_completion(scmd, TIMEOUT_ERROR);
 188
 189         if (scmd->device->host->transportt->eh_timed_out)
 190                 switch (scmd->device->host->transportt->eh_timed_out(scmd)) {
 191                 case EH_HANDLED:
 192                         __scsi_done(scmd);
 193                         return;
 194                 case EH_RESET_TIMER:
 195                         scsi_add_timer(scmd, scmd->timeout_per_command,
 196                                        scsi_times_out);
 197                         return;
 198                 case EH_NOT_HANDLED:
 199                         break;
 200                 }
 201
 202         if (unlikely(!scsi_eh_scmd_add(scmd, SCSI_EH_CANCEL_CMD))) {
 203                 scmd->result |= DID_TIME_OUT << 16;
 204                 __scsi_done(scmd);
 205         }
 206 }
 207
 208 /**
 209  * scsi_block_when_processing_errors - Prevent cmds from being queued.
 210  * @sdev:       Device on which we are performing recovery.
 211  *
 212  * Description:
 213  *     We block until the host is out of error recovery, and then check to
 214  *     see whether the host or the device is offline.
 215  *
 216  * Return value:
 217  *     0 when dev was taken offline by error recovery. 1 OK to proceed.
 218  **/
 219 int scsi_block_when_processing_errors(struct scsi_device *sdev)
 220 {
 221         int online;
 222
 223         wait_event(sdev->host->host_wait, !scsi_host_in_recovery(sdev->host));
 224
 225         online = scsi_device_online(sdev);
 226
 227         SCSI_LOG_ERROR_RECOVERY(5, printk("%s: rtn: %d\n", __FUNCTION__,
 228                                           online));
 229
 230         return online;
 231 }
 232 EXPORT_SYMBOL(scsi_block_when_processing_errors);
 233
 234 #ifdef CONFIG_SCSI_LOGGING
 235 /**
 236  * scsi_eh_prt_fail_stats - Log info on failures.
 237  * @shost:      scsi host being recovered.
 238  * @work_q:     Queue of scsi cmds to process.
 239  **/
 240 static inline void scsi_eh_prt_fail_stats(struct Scsi_Host *shost,
 241                                           struct list_head *work_q)
 242 {
 243         struct scsi_cmnd *scmd;
 244         struct scsi_device *sdev;
 245         int total_failures = 0;
 246         int cmd_failed = 0;
 247         int cmd_cancel = 0;
 248         int devices_failed = 0;
 249
 250         shost_for_each_device(sdev, shost) {
 251                 list_for_each_entry(scmd, work_q, eh_entry) {
 252                         if (scmd->device == sdev) {
 253                                 ++total_failures;
 254                                 if (scmd->eh_eflags & SCSI_EH_CANCEL_CMD)
 255                                         ++cmd_cancel;
 256                                 else
 257                                         ++cmd_failed;
 258                         }
 259                 }
 260
 261                 if (cmd_cancel || cmd_failed) {
 262                         SCSI_LOG_ERROR_RECOVERY(3,
 263                                 sdev_printk(KERN_INFO, sdev,
 264                                             "%s: cmds failed: %d, cancel: %d\n",
 265                                             __FUNCTION__, cmd_failed,
 266                                             cmd_cancel));
 267                         cmd_cancel = 0;
 268                         cmd_failed = 0;
 269                         ++devices_failed;
 270                 }
 271         }
 272
 273         SCSI_LOG_ERROR_RECOVERY(2, printk("Total of %d commands on %d"
 274                                           " devices require eh work\n",
 275                                   total_failures, devices_failed));
 276 }
 277 #endif
 278
 279 /**
 280  * scsi_check_sense - Examine scsi cmd sense
 281  * @scmd:       Cmd to have sense checked.
 282  *
 283  * Return value:
 284  *      SUCCESS or FAILED or NEEDS_RETRY
 285  *
 286  * Notes:
 287  *      When a deferred error is detected the current command has
 288  *      not been executed and needs retrying.
 289  **/
 290 static int scsi_check_sense(struct scsi_cmnd *scmd)
 291 {
 292         struct scsi_sense_hdr sshdr;
 293
 294         if (! scsi_command_normalize_sense(scmd, &sshdr))
 295                 return FAILED;  /* no valid sense data */
 296
 297         if (scsi_sense_is_deferred(&sshdr))
 298                 return NEEDS_RETRY;
 299
 300         /*
 301          * Previous logic looked for FILEMARK, EOM or ILI which are
 302          * mainly associated with tapes and returned SUCCESS.
 303          */
 304         if (sshdr.response_code == 0x70) {
 305                 /* fixed format */
 306                 if (scmd->sense_buffer[2] & 0xe0)
 307                         return SUCCESS;
 308         } else {
 309                 /*
 310                  * descriptor format: look for "stream commands sense data
 311                  * descriptor" (see SSC-3). Assume single sense data
 312                  * descriptor. Ignore ILI from SBC-2 READ LONG and WRITE LONG.
 313                  */
 314                 if ((sshdr.additional_length > 3) &&
 315                     (scmd->sense_buffer[8] == 0x4) &&
 316                     (scmd->sense_buffer[11] & 0xe0))
 317                         return SUCCESS;
 318         }
 319
 320         switch (sshdr.sense_key) {
 321         case NO_SENSE:
 322                 return SUCCESS;
 323         case RECOVERED_ERROR:
 324                 return /* soft_error */ SUCCESS;
 325
 326         case ABORTED_COMMAND:
 327                 return NEEDS_RETRY;
 328         case NOT_READY:
 329         case UNIT_ATTENTION:
 330                 /*
 331                  * if we are expecting a cc/ua because of a bus reset that we
 332                  * performed, treat this just as a retry.  otherwise this is
 333                  * information that we should pass up to the upper-level driver
 334                  * so that we can deal with it there.
 335                  */
 336                 if (scmd->device->expecting_cc_ua) {
 337                         scmd->device->expecting_cc_ua = 0;
 338                         return NEEDS_RETRY;
 339                 }
 340                 /*
 341                  * if the device is in the process of becoming ready, we
 342                  * should retry.
 343                  */
 344                 if ((sshdr.asc == 0x04) && (sshdr.ascq == 0x01))
 345                         return NEEDS_RETRY;
 346                 /*
 347                  * if the device is not started, we need to wake
 348                  * the error handler to start the motor
 349                  */
 350                 if (scmd->device->allow_restart &&
 351                     (sshdr.asc == 0x04) && (sshdr.ascq == 0x02))
 352                         return FAILED;
 353                 return SUCCESS;
 354
 355                 /* these three are not supported */
 356         case COPY_ABORTED:
 357         case VOLUME_OVERFLOW:
 358         case MISCOMPARE:
 359                 return SUCCESS;
 360
 361         case MEDIUM_ERROR:
 362                 return NEEDS_RETRY;
 363
 364         case HARDWARE_ERROR:
 365                 if (scmd->device->retry_hwerror)
 366                         return NEEDS_RETRY;
 367                 else
 368                         return SUCCESS;
 369
 370         case ILLEGAL_REQUEST:
 371         case BLANK_CHECK:
 372         case DATA_PROTECT:
 373         default:
 374                 return SUCCESS;
 375         }
 376 }
 377
 378 /**
 379  * scsi_eh_completed_normally - Disposition a eh cmd on return from LLD.
 380  * @scmd:       SCSI cmd to examine.
 381  *
 382  * Notes:
 383  *    This is *only* called when we are examining the status of commands
 384  *    queued during error recovery.  the main difference here is that we
 385  *    don't allow for the possibility of retries here, and we are a lot
 386  *    more restrictive about what we consider acceptable.
 387  **/
 388 static int scsi_eh_completed_normally(struct scsi_cmnd *scmd)
 389 {
 390         /*
 391          * first check the host byte, to see if there is anything in there
 392          * that would indicate what we need to do.
 393          */
 394         if (host_byte(scmd->result) == DID_RESET) {
 395                 /*
 396                  * rats.  we are already in the error handler, so we now
 397                  * get to try and figure out what to do next.  if the sense
 398                  * is valid, we have a pretty good idea of what to do.
 399                  * if not, we mark it as FAILED.
 400                  */
 401                 return scsi_check_sense(scmd);
 402         }
 403         if (host_byte(scmd->result) != DID_OK)
 404                 return FAILED;
 405
 406         /*
 407          * next, check the message byte.
 408          */
 409         if (msg_byte(scmd->result) != COMMAND_COMPLETE)
 410                 return FAILED;
 411
 412         /*
 413          * now, check the status byte to see if this indicates
 414          * anything special.
 415          */
 416         switch (status_byte(scmd->result)) {
 417         case GOOD:
 418         case COMMAND_TERMINATED:
 419                 return SUCCESS;
 420         case CHECK_CONDITION:
 421                 return scsi_check_sense(scmd);
 422         case CONDITION_GOOD:
 423         case INTERMEDIATE_GOOD:
 424         case INTERMEDIATE_C_GOOD:
 425                 /*
 426                  * who knows?  FIXME(eric)
 427                  */
 428                 return SUCCESS;
 429         case BUSY:
 430         case QUEUE_FULL:
 431         case RESERVATION_CONFLICT:
 432         default:
 433                 return FAILED;
 434         }
 435         return FAILED;
 436 }
 437
 438 /**
 439  * scsi_eh_done - Completion function for error handling.
 440  * @scmd:       Cmd that is done.
 441  **/
 442 static void scsi_eh_done(struct scsi_cmnd *scmd)
 443 {
 444         struct completion     *eh_action;
 445
 446         SCSI_LOG_ERROR_RECOVERY(3,
 447                 printk("%s scmd: %p result: %x\n",
 448                         __FUNCTION__, scmd, scmd->result));
 449
 450         eh_action = scmd->device->host->eh_action;
 451         if (eh_action)
 452                 complete(eh_action);
 453 }
 454
 455 /**
 456  * scsi_send_eh_cmnd  - send a cmd to a device as part of error recovery.
 457  * @scmd:       SCSI Cmd to send.
 458  * @timeout:    Timeout for cmd.
 459  *
 460  * Return value:
 461  *    SUCCESS or FAILED or NEEDS_RETRY
 462  **/
 463 static int scsi_send_eh_cmnd(struct scsi_cmnd *scmd, int timeout)
 464 {
 465         struct scsi_device *sdev = scmd->device;
 466         struct Scsi_Host *shost = sdev->host;
 467         DECLARE_COMPLETION(done);
 468         unsigned long timeleft;
 469         unsigned long flags;
 470         int rtn;
 471
 472         if (sdev->scsi_level <= SCSI_2)
 473                 scmd->cmnd[1] = (scmd->cmnd[1] & 0x1f) |
 474                         (sdev->lun << 5 & 0xe0);
 475
 476         shost->eh_action = &done;
 477         scmd->request->rq_status = RQ_SCSI_BUSY;
 478
 479         spin_lock_irqsave(shost->host_lock, flags);
 480         scsi_log_send(scmd);
 481         shost->hostt->queuecommand(scmd, scsi_eh_done);
 482         spin_unlock_irqrestore(shost->host_lock, flags);
 483
 484         timeleft = wait_for_completion_timeout(&done, timeout);
 485
 486         scmd->request->rq_status = RQ_SCSI_DONE;
 487         shost->eh_action = NULL;
 488
 489         scsi_log_completion(scmd, SUCCESS);
 490
 491         SCSI_LOG_ERROR_RECOVERY(3,
 492                 printk("%s: scmd: %p, timeleft: %ld\n",
 493                         __FUNCTION__, scmd, timeleft));
 494
 495         /*
 496          * If there is time left scsi_eh_done got called, and we will
 497          * examine the actual status codes to see whether the command
 498          * actually did complete normally, else tell the host to forget
 499          * about this command.
 500          */
 501         if (timeleft) {
 502                 rtn = scsi_eh_completed_normally(scmd);
 503                 SCSI_LOG_ERROR_RECOVERY(3,
 504                         printk("%s: scsi_eh_completed_normally %x\n",
 505                                __FUNCTION__, rtn));
 506
 507                 switch (rtn) {
 508                 case SUCCESS:
 509                 case NEEDS_RETRY:
 510                 case FAILED:
 511                         break;
 512                 default:
 513                         rtn = FAILED;
 514                         break;
 515                 }
 516         } else {
 517                 /*
 518                  * FIXME(eric) - we are not tracking whether we could
 519                  * abort a timed out command or not.  not sure how
 520                  * we should treat them differently anyways.
 521                  */
 522                 if (shost->hostt->eh_abort_handler)
 523                         shost->hostt->eh_abort_handler(scmd);
 524                 rtn = FAILED;
 525         }
 526
 527         return rtn;
 528 }
 529
 530 /**
 531  * scsi_request_sense - Request sense data from a particular target.
 532  * @scmd:       SCSI cmd for request sense.
 533  *
 534  * Notes:
 535  *    Some hosts automatically obtain this information, others require
 536  *    that we obtain it on our own. This function will *not* return until
 537  *    the command either times out, or it completes.
 538  **/
 539 static int scsi_request_sense(struct scsi_cmnd *scmd)
 540 {
 541         static unsigned char generic_sense[6] =
 542         {REQUEST_SENSE, 0, 0, 0, 252, 0};
 543         unsigned char *scsi_result;
 544         int saved_result;
 545         int rtn;
 546
 547         memcpy(scmd->cmnd, generic_sense, sizeof(generic_sense));
 548
 549         scsi_result = kmalloc(252, GFP_ATOMIC | ((scmd->device->host->hostt->unchecked_isa_dma) ? __GFP_DMA : 0));
 550
 551
 552         if (unlikely(!scsi_result)) {
 553                 printk(KERN_ERR "%s: cannot allocate scsi_result.\n",
 554                        __FUNCTION__);
 555                 return FAILED;
 556         }
 557
 558         /*
 559          * zero the sense buffer.  some host adapters automatically always
 560          * request sense, so it is not a good idea that
 561          * scmd->request_buffer and scmd->sense_buffer point to the same
 562          * address (db).  0 is not a valid sense code.
 563          */
 564         memset(scmd->sense_buffer, 0, sizeof(scmd->sense_buffer));
 565         memset(scsi_result, 0, 252);
 566
 567         saved_result = scmd->result;
 568         scmd->request_buffer = scsi_result;
 569         scmd->request_bufflen = 252;
 570         scmd->use_sg = 0;
 571         scmd->cmd_len = COMMAND_SIZE(scmd->cmnd[0]);
 572         scmd->sc_data_direction = DMA_FROM_DEVICE;
 573         scmd->underflow = 0;
 574
 575         rtn = scsi_send_eh_cmnd(scmd, SENSE_TIMEOUT);
 576
 577         /* last chance to have valid sense data */
 578         if(!SCSI_SENSE_VALID(scmd)) {
 579                 memcpy(scmd->sense_buffer, scmd->request_buffer,
 580                        sizeof(scmd->sense_buffer));
 581         }
 582
 583         kfree(scsi_result);
 584
 585         /*
 586          * when we eventually call scsi_finish, we really wish to complete
 587          * the original request, so let's restore the original data. (db)
 588          */
 589         scsi_setup_cmd_retry(scmd);
 590         scmd->result = saved_result;
 591         return rtn;
 592 }
 593
 594 /**
 595  * scsi_eh_finish_cmd - Handle a cmd that eh is finished with.
 596  * @scmd:       Original SCSI cmd that eh has finished.
 597  * @done_q:     Queue for processed commands.
 598  *
 599  * Notes:
 600  *    We don't want to use the normal command completion while we are are
 601  *    still handling errors - it may cause other commands to be queued,
 602  *    and that would disturb what we are doing.  thus we really want to
 603  *    keep a list of pending commands for final completion, and once we
 604  *    are ready to leave error handling we handle completion for real.
 605  **/
 606 void scsi_eh_finish_cmd(struct scsi_cmnd *scmd, struct list_head *done_q)
 607 {
 608         scmd->device->host->host_failed--;
 609         scmd->eh_eflags = 0;
 610
 611         /*
 612          * set this back so that the upper level can correctly free up
 613          * things.
 614          */
 615         scsi_setup_cmd_retry(scmd);
 616         list_move_tail(&scmd->eh_entry, done_q);
 617 }
 618 EXPORT_SYMBOL(scsi_eh_finish_cmd);
 619
 620 /**
 621  * scsi_eh_get_sense - Get device sense data.
 622  * @work_q:     Queue of commands to process.
 623  * @done_q:     Queue of proccessed commands..
 624  *
 625  * Description:
 626  *    See if we need to request sense information.  if so, then get it
 627  *    now, so we have a better idea of what to do.
 628  *
 629  * Notes:
 630  *    This has the unfortunate side effect that if a shost adapter does
 631  *    not automatically request sense information, that we end up shutting
 632  *    it down before we request it.
 633  *
 634  *    All drivers should request sense information internally these days,
 635  *    so for now all I have to say is tough noogies if you end up in here.
 636  *
 637  *    XXX: Long term this code should go away, but that needs an audit of
 638  *         all LLDDs first.
 639  **/
 640 static int scsi_eh_get_sense(struct list_head *work_q,
 641                              struct list_head *done_q)
 642 {
 643         struct scsi_cmnd *scmd, *next;
 644         int rtn;
 645
 646         list_for_each_entry_safe(scmd, next, work_q, eh_entry) {
 647                 if ((scmd->eh_eflags & SCSI_EH_CANCEL_CMD) ||
 648                     SCSI_SENSE_VALID(scmd))
 649                         continue;
 650
 651                 SCSI_LOG_ERROR_RECOVERY(2, scmd_printk(KERN_INFO, scmd,
 652                                                   "%s: requesting sense\n",
 653                                                   current->comm));
 654                 rtn = scsi_request_sense(scmd);
 655                 if (rtn != SUCCESS)
 656                         continue;
 657
 658                 SCSI_LOG_ERROR_RECOVERY(3, printk("sense requested for %p"
 659                                                   " result %x\n", scmd,
 660                                                   scmd->result));
 661                 SCSI_LOG_ERROR_RECOVERY(3, scsi_print_sense("bh", scmd));
 662
 663                 rtn = scsi_decide_disposition(scmd);
 664
 665                 /*
 666                  * if the result was normal, then just pass it along to the
 667                  * upper level.
 668                  */
 669                 if (rtn == SUCCESS)
 670                         /* we don't want this command reissued, just
 671                          * finished with the sense data, so set
 672                          * retries to the max allowed to ensure it
 673                          * won't get reissued */
 674                         scmd->retries = scmd->allowed;
 675                 else if (rtn != NEEDS_RETRY)
 676                         continue;
 677
 678                 scsi_eh_finish_cmd(scmd, done_q);
 679         }
 680
 681         return list_empty(work_q);
 682 }
 683
 684 /**
 685  * scsi_try_to_abort_cmd - Ask host to abort a running command.
 686  * @scmd:       SCSI cmd to abort from Lower Level.
 687  *
 688  * Notes:
 689  *    This function will not return until the user's completion function
 690  *    has been called.  there is no timeout on this operation.  if the
 691  *    author of the low-level driver wishes this operation to be timed,
 692  *    they can provide this facility themselves.  helper functions in
 693  *    scsi_error.c can be supplied to make this easier to do.
 694  **/
 695 static int scsi_try_to_abort_cmd(struct scsi_cmnd *scmd)
 696 {
 697         if (!scmd->device->host->hostt->eh_abort_handler)
 698                 return FAILED;
 699
 700         /*
 701          * scsi_done was called just after the command timed out and before
 702          * we had a chance to process it. (db)
 703          */
 704         if (scmd->serial_number == 0)
 705                 return SUCCESS;
 706         return scmd->device->host->hostt->eh_abort_handler(scmd);
 707 }
 708
 709 /**
 710  * scsi_eh_tur - Send TUR to device.
 711  * @scmd:       Scsi cmd to send TUR
 712  *
 713  * Return value:
 714  *    0 - Device is ready. 1 - Device NOT ready.
 715  **/
 716 static int scsi_eh_tur(struct scsi_cmnd *scmd)
 717 {
 718         static unsigned char tur_command[6] = {TEST_UNIT_READY, 0, 0, 0, 0, 0};
 719         int retry_cnt = 1, rtn;
 720         int saved_result;
 721
 722 retry_tur:
 723         memcpy(scmd->cmnd, tur_command, sizeof(tur_command));
 724
 725         /*
 726          * zero the sense buffer.  the scsi spec mandates that any
 727          * untransferred sense data should be interpreted as being zero.
 728          */
 729         memset(scmd->sense_buffer, 0, sizeof(scmd->sense_buffer));
 730
 731         saved_result = scmd->result;
 732         scmd->request_buffer = NULL;
 733         scmd->request_bufflen = 0;
 734         scmd->use_sg = 0;
 735         scmd->cmd_len = COMMAND_SIZE(scmd->cmnd[0]);
 736         scmd->underflow = 0;
 737         scmd->sc_data_direction = DMA_NONE;
 738
 739         rtn = scsi_send_eh_cmnd(scmd, SENSE_TIMEOUT);
 740
 741         /*
 742          * when we eventually call scsi_finish, we really wish to complete
 743          * the original request, so let's restore the original data. (db)
 744          */
 745         scsi_setup_cmd_retry(scmd);
 746         scmd->result = saved_result;
 747
 748         /*
 749          * hey, we are done.  let's look to see what happened.
 750          */
 751         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: scmd %p rtn %x\n",
 752                 __FUNCTION__, scmd, rtn));
 753         if (rtn == SUCCESS)
 754                 return 0;
 755         else if (rtn == NEEDS_RETRY) {
 756                 if (retry_cnt--)
 757                         goto retry_tur;
 758                 return 0;
 759         }
 760         return 1;
 761 }
 762
 763 /**
 764  * scsi_eh_abort_cmds - abort canceled commands.
 765  * @shost:      scsi host being recovered.
 766  * @eh_done_q:  list_head for processed commands.
 767  *
 768  * Decription:
 769  *    Try and see whether or not it makes sense to try and abort the
 770  *    running command.  this only works out to be the case if we have one
 771  *    command that has timed out.  if the command simply failed, it makes
 772  *    no sense to try and abort the command, since as far as the shost
 773  *    adapter is concerned, it isn't running.
 774  **/
 775 static int scsi_eh_abort_cmds(struct list_head *work_q,
 776                               struct list_head *done_q)
 777 {
 778         struct scsi_cmnd *scmd, *next;
 779         int rtn;
 780
 781         list_for_each_entry_safe(scmd, next, work_q, eh_entry) {
 782                 if (!(scmd->eh_eflags & SCSI_EH_CANCEL_CMD))
 783                         continue;
 784                 SCSI_LOG_ERROR_RECOVERY(3, printk("%s: aborting cmd:"
 785                                                   "0x%p\n", current->comm,
 786                                                   scmd));
 787                 rtn = scsi_try_to_abort_cmd(scmd);
 788                 if (rtn == SUCCESS) {
 789                         scmd->eh_eflags &= ~SCSI_EH_CANCEL_CMD;
 790                         if (!scsi_device_online(scmd->device) ||
 791                             !scsi_eh_tur(scmd)) {
 792                                 scsi_eh_finish_cmd(scmd, done_q);
 793                         }
 794
 795                 } else
 796                         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: aborting"
 797                                                           " cmd failed:"
 798                                                           "0x%p\n",
 799                                                           current->comm,
 800                                                           scmd));
 801         }
 802
 803         return list_empty(work_q);
 804 }
 805
 806 /**
 807  * scsi_try_bus_device_reset - Ask host to perform a BDR on a dev
 808  * @scmd:       SCSI cmd used to send BDR
 809  *
 810  * Notes:
 811  *    There is no timeout for this operation.  if this operation is
 812  *    unreliable for a given host, then the host itself needs to put a
 813  *    timer on it, and set the host back to a consistent state prior to
 814  *    returning.
 815  **/
 816 static int scsi_try_bus_device_reset(struct scsi_cmnd *scmd)
 817 {
 818         int rtn;
 819
 820         if (!scmd->device->host->hostt->eh_device_reset_handler)
 821                 return FAILED;
 822
 823         rtn = scmd->device->host->hostt->eh_device_reset_handler(scmd);
 824         if (rtn == SUCCESS) {
 825                 scmd->device->was_reset = 1;
 826                 scmd->device->expecting_cc_ua = 1;
 827         }
 828
 829         return rtn;
 830 }
 831
 832 /**
 833  * scsi_eh_try_stu - Send START_UNIT to device.
 834  * @scmd:       Scsi cmd to send START_UNIT
 835  *
 836  * Return value:
 837  *    0 - Device is ready. 1 - Device NOT ready.
 838  **/
 839 static int scsi_eh_try_stu(struct scsi_cmnd *scmd)
 840 {
 841         static unsigned char stu_command[6] = {START_STOP, 0, 0, 0, 1, 0};
 842         int rtn;
 843         int saved_result;
 844
 845         if (!scmd->device->allow_restart)
 846                 return 1;
 847
 848         memcpy(scmd->cmnd, stu_command, sizeof(stu_command));
 849
 850         /*
 851          * zero the sense buffer.  the scsi spec mandates that any
 852          * untransferred sense data should be interpreted as being zero.
 853          */
 854         memset(scmd->sense_buffer, 0, sizeof(scmd->sense_buffer));
 855
 856         saved_result = scmd->result;
 857         scmd->request_buffer = NULL;
 858         scmd->request_bufflen = 0;
 859         scmd->use_sg = 0;
 860         scmd->cmd_len = COMMAND_SIZE(scmd->cmnd[0]);
 861         scmd->underflow = 0;
 862         scmd->sc_data_direction = DMA_NONE;
 863
 864         rtn = scsi_send_eh_cmnd(scmd, START_UNIT_TIMEOUT);
 865
 866         /*
 867          * when we eventually call scsi_finish, we really wish to complete
 868          * the original request, so let's restore the original data. (db)
 869          */
 870         scsi_setup_cmd_retry(scmd);
 871         scmd->result = saved_result;
 872
 873         /*
 874          * hey, we are done.  let's look to see what happened.
 875          */
 876         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: scmd %p rtn %x\n",
 877                 __FUNCTION__, scmd, rtn));
 878         if (rtn == SUCCESS)
 879                 return 0;
 880         return 1;
 881 }
 882
 883  /**
 884  * scsi_eh_stu - send START_UNIT if needed
 885  * @shost:      scsi host being recovered.
 886  * @eh_done_q:  list_head for processed commands.
 887  *
 888  * Notes:
 889  *    If commands are failing due to not ready, initializing command required,
 890  *      try revalidating the device, which will end up sending a start unit.
 891  **/
 892 static int scsi_eh_stu(struct Scsi_Host *shost,
 893                               struct list_head *work_q,
 894                               struct list_head *done_q)
 895 {
 896         struct scsi_cmnd *scmd, *stu_scmd, *next;
 897         struct scsi_device *sdev;
 898
 899         shost_for_each_device(sdev, shost) {
 900                 stu_scmd = NULL;
 901                 list_for_each_entry(scmd, work_q, eh_entry)
 902                         if (scmd->device == sdev && SCSI_SENSE_VALID(scmd) &&
 903                             scsi_check_sense(scmd) == FAILED ) {
 904                                 stu_scmd = scmd;
 905                                 break;
 906                         }
 907
 908                 if (!stu_scmd)
 909                         continue;
 910
 911                 SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Sending START_UNIT to sdev:"
 912                                                   " 0x%p\n", current->comm, sdev));
 913
 914                 if (!scsi_eh_try_stu(stu_scmd)) {
 915                         if (!scsi_device_online(sdev) ||
 916                             !scsi_eh_tur(stu_scmd)) {
 917                                 list_for_each_entry_safe(scmd, next,
 918                                                           work_q, eh_entry) {
 919                                         if (scmd->device == sdev)
 920                                                 scsi_eh_finish_cmd(scmd, done_q);
 921                                 }
 922                         }
 923                 } else {
 924                         SCSI_LOG_ERROR_RECOVERY(3,
 925                                                 printk("%s: START_UNIT failed to sdev:"
 926                                                        " 0x%p\n", current->comm, sdev));
 927                 }
 928         }
 929
 930         return list_empty(work_q);
 931 }
 932
 933
 934 /**
 935  * scsi_eh_bus_device_reset - send bdr if needed
 936  * @shost:      scsi host being recovered.
 937  * @eh_done_q:  list_head for processed commands.
 938  *
 939  * Notes:
 940  *    Try a bus device reset.  still, look to see whether we have multiple
 941  *    devices that are jammed or not - if we have multiple devices, it
 942  *    makes no sense to try bus_device_reset - we really would need to try
 943  *    a bus_reset instead.
 944  **/
 945 static int scsi_eh_bus_device_reset(struct Scsi_Host *shost,
 946                                     struct list_head *work_q,
 947                                     struct list_head *done_q)
 948 {
 949         struct scsi_cmnd *scmd, *bdr_scmd, *next;
 950         struct scsi_device *sdev;
 951         int rtn;
 952
 953         shost_for_each_device(sdev, shost) {
 954                 bdr_scmd = NULL;
 955                 list_for_each_entry(scmd, work_q, eh_entry)
 956                         if (scmd->device == sdev) {
 957                                 bdr_scmd = scmd;
 958                                 break;
 959                         }
 960
 961                 if (!bdr_scmd)
 962                         continue;
 963
 964                 SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Sending BDR sdev:"
 965                                                   " 0x%p\n", current->comm,
 966                                                   sdev));
 967                 rtn = scsi_try_bus_device_reset(bdr_scmd);
 968                 if (rtn == SUCCESS) {
 969                         if (!scsi_device_online(sdev) ||
 970                             !scsi_eh_tur(bdr_scmd)) {
 971                                 list_for_each_entry_safe(scmd, next,
 972                                                          work_q, eh_entry) {
 973                                         if (scmd->device == sdev)
 974                                                 scsi_eh_finish_cmd(scmd,
 975                                                                    done_q);
 976                                 }
 977                         }
 978                 } else {
 979                         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: BDR"
 980                                                           " failed sdev:"
 981                                                           "0x%p\n",
 982                                                           current->comm,
 983                                                            sdev));
 984                 }
 985         }
 986
 987         return list_empty(work_q);
 988 }
 989
 990 /**
 991  * scsi_try_bus_reset - ask host to perform a bus reset
 992  * @scmd:       SCSI cmd to send bus reset.
 993  **/
 994 static int scsi_try_bus_reset(struct scsi_cmnd *scmd)
 995 {
 996         unsigned long flags;
 997         int rtn;
 998
 999         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Snd Bus RST\n",
1000                                           __FUNCTION__));
1001
1002         if (!scmd->device->host->hostt->eh_bus_reset_handler)
1003                 return FAILED;
1004
1005         rtn = scmd->device->host->hostt->eh_bus_reset_handler(scmd);
1006
1007         if (rtn == SUCCESS) {
1008                 if (!scmd->device->host->hostt->skip_settle_delay)
1009                         ssleep(BUS_RESET_SETTLE_TIME);
1010                 spin_lock_irqsave(scmd->device->host->host_lock, flags);
1011                 scsi_report_bus_reset(scmd->device->host,
1012                                       scmd_channel(scmd));
1013                 spin_unlock_irqrestore(scmd->device->host->host_lock, flags);
1014         }
1015
1016         return rtn;
1017 }
1018
1019 /**
1020  * scsi_try_host_reset - ask host adapter to reset itself
1021  * @scmd:       SCSI cmd to send hsot reset.
1022  **/
1023 static int scsi_try_host_reset(struct scsi_cmnd *scmd)
1024 {
1025         unsigned long flags;
1026         int rtn;
1027
1028         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Snd Host RST\n",
1029                                           __FUNCTION__));
1030
1031         if (!scmd->device->host->hostt->eh_host_reset_handler)
1032                 return FAILED;
1033
1034         rtn = scmd->device->host->hostt->eh_host_reset_handler(scmd);
1035
1036         if (rtn == SUCCESS) {
1037                 if (!scmd->device->host->hostt->skip_settle_delay)
1038                         ssleep(HOST_RESET_SETTLE_TIME);
1039                 spin_lock_irqsave(scmd->device->host->host_lock, flags);
1040                 scsi_report_bus_reset(scmd->device->host,
1041                                       scmd_channel(scmd));
1042                 spin_unlock_irqrestore(scmd->device->host->host_lock, flags);
1043         }
1044
1045         return rtn;
1046 }
1047
1048 /**
1049  * scsi_eh_bus_reset - send a bus reset
1050  * @shost:      scsi host being recovered.
1051  * @eh_done_q:  list_head for processed commands.
1052  **/
1053 static int scsi_eh_bus_reset(struct Scsi_Host *shost,
1054                              struct list_head *work_q,
1055                              struct list_head *done_q)
1056 {
1057         struct scsi_cmnd *scmd, *chan_scmd, *next;
1058         unsigned int channel;
1059         int rtn;
1060
1061         /*
1062          * we really want to loop over the various channels, and do this on
1063          * a channel by channel basis.  we should also check to see if any
1064          * of the failed commands are on soft_reset devices, and if so, skip
1065          * the reset.
1066          */
1067
1068         for (channel = 0; channel <= shost->max_channel; channel++) {
1069                 chan_scmd = NULL;
1070                 list_for_each_entry(scmd, work_q, eh_entry) {
1071                         if (channel == scmd_channel(scmd)) {
1072                                 chan_scmd = scmd;
1073                                 break;
1074                                 /*
1075                                  * FIXME add back in some support for
1076                                  * soft_reset devices.
1077                                  */
1078                         }
1079                 }
1080
1081                 if (!chan_scmd)
1082                         continue;
1083                 SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Sending BRST chan:"
1084                                                   " %d\n", current->comm,
1085                                                   channel));
1086                 rtn = scsi_try_bus_reset(chan_scmd);
1087                 if (rtn == SUCCESS) {
1088                         list_for_each_entry_safe(scmd, next, work_q, eh_entry) {
1089                                 if (channel == scmd_channel(scmd))
1090                                         if (!scsi_device_online(scmd->device) ||
1091                                             !scsi_eh_tur(scmd))
1092                                                 scsi_eh_finish_cmd(scmd,
1093                                                                    done_q);
1094                         }
1095                 } else {
1096                         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: BRST"
1097                                                           " failed chan: %d\n",
1098                                                           current->comm,
1099                                                           channel));
1100                 }
1101         }
1102         return list_empty(work_q);
1103 }
1104
1105 /**
1106  * scsi_eh_host_reset - send a host reset
1107  * @work_q:     list_head for processed commands.
1108  * @done_q:     list_head for processed commands.
1109  **/
1110 static int scsi_eh_host_reset(struct list_head *work_q,
1111                               struct list_head *done_q)
1112 {
1113         struct scsi_cmnd *scmd, *next;
1114         int rtn;
1115
1116         if (!list_empty(work_q)) {
1117                 scmd = list_entry(work_q->next,
1118                                   struct scsi_cmnd, eh_entry);
1119
1120                 SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Sending HRST\n"
1121                                                   , current->comm));
1122
1123                 rtn = scsi_try_host_reset(scmd);
1124                 if (rtn == SUCCESS) {
1125                         list_for_each_entry_safe(scmd, next, work_q, eh_entry) {
1126                                 if (!scsi_device_online(scmd->device) ||
1127                                     (!scsi_eh_try_stu(scmd) && !scsi_eh_tur(scmd)) ||
1128                                     !scsi_eh_tur(scmd))
1129                                         scsi_eh_finish_cmd(scmd, done_q);
1130                         }
1131                 } else {
1132                         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: HRST"
1133                                                           " failed\n",
1134                                                           current->comm));
1135                 }
1136         }
1137         return list_empty(work_q);
1138 }
1139
1140 /**
1141  * scsi_eh_offline_sdevs - offline scsi devices that fail to recover
1142  * @work_q:     list_head for processed commands.
1143  * @done_q:     list_head for processed commands.
1144  *
1145  **/
1146 static void scsi_eh_offline_sdevs(struct list_head *work_q,
1147                                   struct list_head *done_q)
1148 {
1149         struct scsi_cmnd *scmd, *next;
1150
1151         list_for_each_entry_safe(scmd, next, work_q, eh_entry) {
1152                 sdev_printk(KERN_INFO, scmd->device,
1153                             "scsi: Device offlined - not"
1154                             " ready after error recovery\n");
1155                 scsi_device_set_state(scmd->device, SDEV_OFFLINE);
1156                 if (scmd->eh_eflags & SCSI_EH_CANCEL_CMD) {
1157                         /*
1158                          * FIXME: Handle lost cmds.
1159                          */
1160                 }
1161                 scsi_eh_finish_cmd(scmd, done_q);
1162         }
1163         return;
1164 }
1165
1166 /**
1167  * scsi_decide_disposition - Disposition a cmd on return from LLD.
1168  * @scmd:       SCSI cmd to examine.
1169  *
1170  * Notes:
1171  *    This is *only* called when we are examining the status after sending
1172  *    out the actual data command.  any commands that are queued for error
1173  *    recovery (e.g. test_unit_ready) do *not* come through here.
1174  *
1175  *    When this routine returns failed, it means the error handler thread
1176  *    is woken.  In cases where the error code indicates an error that
1177  *    doesn't require the error handler read (i.e. we don't need to
1178  *    abort/reset), this function should return SUCCESS.
1179  **/
1180 int scsi_decide_disposition(struct scsi_cmnd *scmd)
1181 {
1182         int rtn;
1183
1184         /*
1185          * if the device is offline, then we clearly just pass the result back
1186          * up to the top level.
1187          */
1188         if (!scsi_device_online(scmd->device)) {
1189                 SCSI_LOG_ERROR_RECOVERY(5, printk("%s: device offline - report"
1190                                                   " as SUCCESS\n",
1191                                                   __FUNCTION__));
1192                 return SUCCESS;
1193         }
1194
1195         /*
1196          * first check the host byte, to see if there is anything in there
1197          * that would indicate what we need to do.
1198          */
1199         switch (host_byte(scmd->result)) {
1200         case DID_PASSTHROUGH:
1201                 /*
1202                  * no matter what, pass this through to the upper layer.
1203                  * nuke this special code so that it looks like we are saying
1204                  * did_ok.
1205                  */
1206                 scmd->result &= 0xff00ffff;
1207                 return SUCCESS;
1208         case DID_OK:
1209                 /*
1210                  * looks good.  drop through, and check the next byte.
1211                  */
1212                 break;
1213         case DID_NO_CONNECT:
1214         case DID_BAD_TARGET:
1215         case DID_ABORT:
1216                 /*
1217                  * note - this means that we just report the status back
1218                  * to the top level driver, not that we actually think
1219                  * that it indicates SUCCESS.
1220                  */
1221                 return SUCCESS;
1222                 /*
1223                  * when the low level driver returns did_soft_error,
1224                  * it is responsible for keeping an internal retry counter
1225                  * in order to avoid endless loops (db)
1226                  *
1227                  * actually this is a bug in this function here.  we should
1228                  * be mindful of the maximum number of retries specified
1229                  * and not get stuck in a loop.
1230                  */
1231         case DID_SOFT_ERROR:
1232                 goto maybe_retry;
1233         case DID_IMM_RETRY:
1234                 return NEEDS_RETRY;
1235
1236         case DID_REQUEUE:
1237                 return ADD_TO_MLQUEUE;
1238
1239         case DID_ERROR:
1240                 if (msg_byte(scmd->result) == COMMAND_COMPLETE &&
1241                     status_byte(scmd->result) == RESERVATION_CONFLICT)
1242                         /*
1243                          * execute reservation conflict processing code
1244                          * lower down
1245                          */
1246                         break;
1247                 /* fallthrough */
1248
1249         case DID_BUS_BUSY:
1250         case DID_PARITY:
1251                 goto maybe_retry;
1252         case DID_TIME_OUT:
1253                 /*
1254                  * when we scan the bus, we get timeout messages for
1255                  * these commands if there is no device available.
1256                  * other hosts report did_no_connect for the same thing.
1257                  */
1258                 if ((scmd->cmnd[0] == TEST_UNIT_READY ||
1259                      scmd->cmnd[0] == INQUIRY)) {
1260                         return SUCCESS;
1261                 } else {
1262                         return FAILED;
1263                 }
1264         case DID_RESET:
1265                 return SUCCESS;
1266         default:
1267                 return FAILED;
1268         }
1269
1270         /*
1271          * next, check the message byte.
1272          */
1273         if (msg_byte(scmd->result) != COMMAND_COMPLETE)
1274                 return FAILED;
1275
1276         /*
1277          * check the status byte to see if this indicates anything special.
1278          */
1279         switch (status_byte(scmd->result)) {
1280         case QUEUE_FULL:
1281                 /*
1282                  * the case of trying to send too many commands to a
1283                  * tagged queueing device.
1284                  */
1285         case BUSY:
1286                 /*
1287                  * device can't talk to us at the moment.  Should only
1288                  * occur (SAM-3) when the task queue is empty, so will cause
1289                  * the empty queue handling to trigger a stall in the
1290                  * device.
1291                  */
1292                 return ADD_TO_MLQUEUE;
1293         case GOOD:
1294         case COMMAND_TERMINATED:
1295         case TASK_ABORTED:
1296                 return SUCCESS;
1297         case CHECK_CONDITION:
1298                 rtn = scsi_check_sense(scmd);
1299                 if (rtn == NEEDS_RETRY)
1300                         goto maybe_retry;
1301                 /* if rtn == FAILED, we have no sense information;
1302                  * returning FAILED will wake the error handler thread
1303                  * to collect the sense and redo the decide
1304                  * disposition */
1305                 return rtn;
1306         case CONDITION_GOOD:
1307         case INTERMEDIATE_GOOD:
1308         case INTERMEDIATE_C_GOOD:
1309         case ACA_ACTIVE:
1310                 /*
1311                  * who knows?  FIXME(eric)
1312                  */
1313                 return SUCCESS;
1314
1315         case RESERVATION_CONFLICT:
1316                 sdev_printk(KERN_INFO, scmd->device,
1317                             "reservation conflict\n");
1318                 return SUCCESS; /* causes immediate i/o error */
1319         default:
1320                 return FAILED;
1321         }
1322         return FAILED;
1323
1324       maybe_retry:
1325
1326         /* we requeue for retry because the error was retryable, and
1327          * the request was not marked fast fail.  Note that above,
1328          * even if the request is marked fast fail, we still requeue
1329          * for queue congestion conditions (QUEUE_FULL or BUSY) */
1330         if ((++scmd->retries) <= scmd->allowed
1331             && !blk_noretry_request(scmd->request)) {
1332                 return NEEDS_RETRY;
1333         } else {
1334                 /*
1335                  * no more retries - report this one back to upper level.
1336                  */
1337                 return SUCCESS;
1338         }
1339 }
1340
1341 /**
1342  * scsi_eh_lock_door - Prevent medium removal for the specified device
1343  * @sdev:       SCSI device to prevent medium removal
1344  *
1345  * Locking:
1346  *      We must be called from process context; scsi_allocate_request()
1347  *      may sleep.
1348  *
1349  * Notes:
1350  *      We queue up an asynchronous "ALLOW MEDIUM REMOVAL" request on the
1351  *      head of the devices request queue, and continue.
1352  *
1353  * Bugs:
1354  *      scsi_allocate_request() may sleep waiting for existing requests to
1355  *      be processed.  However, since we haven't kicked off any request
1356  *      processing for this host, this may deadlock.
1357  *
1358  *      If scsi_allocate_request() fails for what ever reason, we
1359  *      completely forget to lock the door.
1360  **/
1361 static void scsi_eh_lock_door(struct scsi_device *sdev)
1362 {
1363         unsigned char cmnd[MAX_COMMAND_SIZE];
1364
1365         cmnd[0] = ALLOW_MEDIUM_REMOVAL;
1366         cmnd[1] = 0;
1367         cmnd[2] = 0;
1368         cmnd[3] = 0;
1369         cmnd[4] = SCSI_REMOVAL_PREVENT;
1370         cmnd[5] = 0;
1371
1372         scsi_execute_async(sdev, cmnd, 6, DMA_NONE, NULL, 0, 0, 10 * HZ,
1373                            5, NULL, NULL, GFP_KERNEL);
1374 }
1375
1376
1377 /**
1378  * scsi_restart_operations - restart io operations to the specified host.
1379  * @shost:      Host we are restarting.
1380  *
1381  * Notes:
1382  *    When we entered the error handler, we blocked all further i/o to
1383  *    this device.  we need to 'reverse' this process.
1384  **/
1385 static void scsi_restart_operations(struct Scsi_Host *shost)
1386 {
1387         struct scsi_device *sdev;
1388         unsigned long flags;
1389
1390         /*
1391          * If the door was locked, we need to insert a door lock request
1392          * onto the head of the SCSI request queue for the device.  There
1393          * is no point trying to lock the door of an off-line device.
1394          */
1395         shost_for_each_device(sdev, shost) {
1396                 if (scsi_device_online(sdev) && sdev->locked)
1397                         scsi_eh_lock_door(sdev);
1398         }
1399
1400         /*
1401          * next free up anything directly waiting upon the host.  this
1402          * will be requests for character device operations, and also for
1403          * ioctls to queued block devices.
1404          */
1405         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: waking up host to restart\n",
1406                                           __FUNCTION__));
1407
1408         spin_lock_irqsave(shost->host_lock, flags);
1409         if (scsi_host_set_state(shost, SHOST_RUNNING))
1410                 if (scsi_host_set_state(shost, SHOST_CANCEL))
1411                         BUG_ON(scsi_host_set_state(shost, SHOST_DEL));
1412         spin_unlock_irqrestore(shost->host_lock, flags);
1413
1414         wake_up(&shost->host_wait);
1415
1416         /*
1417          * finally we need to re-initiate requests that may be pending.  we will
1418          * have had everything blocked while error handling is taking place, and
1419          * now that error recovery is done, we will need to ensure that these
1420          * requests are started.
1421          */
1422         scsi_run_host_queues(shost);
1423 }
1424
1425 /**
1426  * scsi_eh_ready_devs - check device ready state and recover if not.
1427  * @shost:      host to be recovered.
1428  * @eh_done_q:  list_head for processed commands.
1429  *
1430  **/
1431 static void scsi_eh_ready_devs(struct Scsi_Host *shost,
1432                                struct list_head *work_q,
1433                                struct list_head *done_q)
1434 {
1435         if (!scsi_eh_stu(shost, work_q, done_q))
1436                 if (!scsi_eh_bus_device_reset(shost, work_q, done_q))
1437                         if (!scsi_eh_bus_reset(shost, work_q, done_q))
1438                                 if (!scsi_eh_host_reset(work_q, done_q))
1439                                         scsi_eh_offline_sdevs(work_q, done_q);
1440 }
1441
1442 /**
1443  * scsi_eh_flush_done_q - finish processed commands or retry them.
1444  * @done_q:     list_head of processed commands.
1445  *
1446  **/
1447 void scsi_eh_flush_done_q(struct list_head *done_q)
1448 {
1449         struct scsi_cmnd *scmd, *next;
1450
1451         list_for_each_entry_safe(scmd, next, done_q, eh_entry) {
1452                 list_del_init(&scmd->eh_entry);
1453                 if (scsi_device_online(scmd->device) &&
1454                     !blk_noretry_request(scmd->request) &&
1455                     (++scmd->retries <= scmd->allowed)) {
1456                         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: flush"
1457                                                           " retry cmd: %p\n",
1458                                                           current->comm,
1459                                                           scmd));
1460                                 scsi_queue_insert(scmd, SCSI_MLQUEUE_EH_RETRY);
1461                 } else {
1462                         /*
1463                          * If just we got sense for the device (called
1464                          * scsi_eh_get_sense), scmd->result is already
1465                          * set, do not set DRIVER_TIMEOUT.
1466                          */
1467                         if (!scmd->result)
1468                                 scmd->result |= (DRIVER_TIMEOUT << 24);
1469                         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: flush finish"
1470                                                         " cmd: %p\n",
1471                                                         current->comm, scmd));
1472                         scsi_finish_command(scmd);
1473                 }
1474         }
1475 }
1476 EXPORT_SYMBOL(scsi_eh_flush_done_q);
1477
1478 /**
1479  * scsi_unjam_host - Attempt to fix a host which has a cmd that failed.
1480  * @shost:      Host to unjam.
1481  *
1482  * Notes:
1483  *    When we come in here, we *know* that all commands on the bus have
1484  *    either completed, failed or timed out.  we also know that no further
1485  *    commands are being sent to the host, so things are relatively quiet
1486  *    and we have freedom to fiddle with things as we wish.
1487  *
1488  *    This is only the *default* implementation.  it is possible for
1489  *    individual drivers to supply their own version of this function, and
1490  *    if the maintainer wishes to do this, it is strongly suggested that
1491  *    this function be taken as a template and modified.  this function
1492  *    was designed to correctly handle problems for about 95% of the
1493  *    different cases out there, and it should always provide at least a
1494  *    reasonable amount of error recovery.
1495  *
1496  *    Any command marked 'failed' or 'timeout' must eventually have
1497  *    scsi_finish_cmd() called for it.  we do all of the retry stuff
1498  *    here, so when we restart the host after we return it should have an
1499  *    empty queue.
1500  **/
1501 static void scsi_unjam_host(struct Scsi_Host *shost)
1502 {
1503         unsigned long flags;
1504         LIST_HEAD(eh_work_q);
1505         LIST_HEAD(eh_done_q);
1506
1507         spin_lock_irqsave(shost->host_lock, flags);
1508         list_splice_init(&shost->eh_cmd_q, &eh_work_q);
1509         spin_unlock_irqrestore(shost->host_lock, flags);
1510
1511         SCSI_LOG_ERROR_RECOVERY(1, scsi_eh_prt_fail_stats(shost, &eh_work_q));
1512
1513         if (!scsi_eh_get_sense(&eh_work_q, &eh_done_q))
1514                 if (!scsi_eh_abort_cmds(&eh_work_q, &eh_done_q))
1515                         scsi_eh_ready_devs(shost, &eh_work_q, &eh_done_q);
1516
1517         scsi_eh_flush_done_q(&eh_done_q);
1518 }
1519
1520 /**
1521  * scsi_error_handler - SCSI error handler thread
1522  * @data:       Host for which we are running.
1523  *
1524  * Notes:
1525  *    This is the main error handling loop.  This is run as a kernel thread
1526  *    for every SCSI host and handles all error handling activity.
1527  **/
1528 int scsi_error_handler(void *data)
1529 {
1530         struct Scsi_Host *shost = data;
1531
1532         current->flags |= PF_NOFREEZE;
1533
1534         /*
1535          * We use TASK_INTERRUPTIBLE so that the thread is not
1536          * counted against the load average as a running process.
1537          * We never actually get interrupted because kthread_run
1538          * disables singal delivery for the created thread.
1539          */
1540         set_current_state(TASK_INTERRUPTIBLE);
1541         while (!kthread_should_stop()) {
1542                 if ((shost->host_failed == 0 && shost->host_eh_scheduled == 0) ||
1543                     shost->host_failed != shost->host_busy) {
1544                         SCSI_LOG_ERROR_RECOVERY(1,
1545                                 printk("Error handler scsi_eh_%d sleeping\n",
1546                                         shost->host_no));
1547                         schedule();
1548                         set_current_state(TASK_INTERRUPTIBLE);
1549                         continue;
1550                 }
1551
1552                 __set_current_state(TASK_RUNNING);
1553                 SCSI_LOG_ERROR_RECOVERY(1,
1554                         printk("Error handler scsi_eh_%d waking up\n",
1555                                 shost->host_no));
1556
1557                 /*
1558                  * We have a host that is failing for some reason.  Figure out
1559                  * what we need to do to get it up and online again (if we can).
1560                  * If we fail, we end up taking the thing offline.
1561                  */
1562                 if (shost->transportt->eh_strategy_handler)
1563                         shost->transportt->eh_strategy_handler(shost);
1564                 else
1565                         scsi_unjam_host(shost);
1566
1567                 /*
1568                  * Note - if the above fails completely, the action is to take
1569                  * individual devices offline and flush the queue of any
1570                  * outstanding requests that may have been pending.  When we
1571                  * restart, we restart any I/O to any other devices on the bus
1572                  * which are still online.
1573                  */
1574                 scsi_restart_operations(shost);
1575                 set_current_state(TASK_INTERRUPTIBLE);
1576         }
1577         __set_current_state(TASK_RUNNING);
1578
1579         SCSI_LOG_ERROR_RECOVERY(1,
1580                 printk("Error handler scsi_eh_%d exiting\n", shost->host_no));
1581         shost->ehandler = NULL;
1582         return 0;
1583 }
1584
1585 /*
1586  * Function:    scsi_report_bus_reset()
1587  *
1588  * Purpose:     Utility function used by low-level drivers to report that
1589  *              they have observed a bus reset on the bus being handled.
1590  *
1591  * Arguments:   shost       - Host in question
1592  *              channel     - channel on which reset was observed.
1593  *
1594  * Returns:     Nothing
1595  *
1596  * Lock status: Host lock must be held.
1597  *
1598  * Notes:       This only needs to be called if the reset is one which
1599  *              originates from an unknown location.  Resets originated
1600  *              by the mid-level itself don't need to call this, but there
1601  *              should be no harm.
1602  *
1603  *              The main purpose of this is to make sure that a CHECK_CONDITION
1604  *              is properly treated.
1605  */
1606 void scsi_report_bus_reset(struct Scsi_Host *shost, int channel)
1607 {
1608         struct scsi_device *sdev;
1609
1610         __shost_for_each_device(sdev, shost) {
1611                 if (channel == sdev_channel(sdev)) {
1612                         sdev->was_reset = 1;
1613                         sdev->expecting_cc_ua = 1;
1614                 }
1615         }
1616 }
1617 EXPORT_SYMBOL(scsi_report_bus_reset);
1618
1619 /*
1620  * Function:    scsi_report_device_reset()
1621  *
1622  * Purpose:     Utility function used by low-level drivers to report that
1623  *              they have observed a device reset on the device being handled.
1624  *
1625  * Arguments:   shost       - Host in question
1626  *              channel     - channel on which reset was observed
1627  *              target      - target on which reset was observed
1628  *
1629  * Returns:     Nothing
1630  *
1631  * Lock status: Host lock must be held
1632  *
1633  * Notes:       This only needs to be called if the reset is one which
1634  *              originates from an unknown location.  Resets originated
1635  *              by the mid-level itself don't need to call this, but there
1636  *              should be no harm.
1637  *
1638  *              The main purpose of this is to make sure that a CHECK_CONDITION
1639  *              is properly treated.
1640  */
1641 void scsi_report_device_reset(struct Scsi_Host *shost, int channel, int target)
1642 {
1643         struct scsi_device *sdev;
1644
1645         __shost_for_each_device(sdev, shost) {
1646                 if (channel == sdev_channel(sdev) &&
1647                     target == sdev_id(sdev)) {
1648                         sdev->was_reset = 1;
1649                         sdev->expecting_cc_ua = 1;
1650                 }
1651         }
1652 }
1653 EXPORT_SYMBOL(scsi_report_device_reset);
1654
1655 static void
1656 scsi_reset_provider_done_command(struct scsi_cmnd *scmd)
1657 {
1658 }
1659
1660 /*
1661  * Function:    scsi_reset_provider
1662  *
1663  * Purpose:     Send requested reset to a bus or device at any phase.
1664  *
1665  * Arguments:   device  - device to send reset to
1666  *              flag - reset type (see scsi.h)
1667  *
1668  * Returns:     SUCCESS/FAILURE.
1669  *
1670  * Notes:       This is used by the SCSI Generic driver to provide
1671  *              Bus/Device reset capability.
1672  */
1673 int
1674 scsi_reset_provider(struct scsi_device *dev, int flag)
1675 {
1676         struct scsi_cmnd *scmd = scsi_get_command(dev, GFP_KERNEL);
1677         struct request req;
1678         int rtn;
1679
1680         scmd->request = &req;
1681         memset(&scmd->eh_timeout, 0, sizeof(scmd->eh_timeout));
1682         scmd->request->rq_status        = RQ_SCSI_BUSY;
1683
1684         memset(&scmd->cmnd, '\0', sizeof(scmd->cmnd));
1685
1686         scmd->scsi_done         = scsi_reset_provider_done_command;
1687         scmd->done                      = NULL;
1688         scmd->buffer                    = NULL;
1689         scmd->bufflen                   = 0;
1690         scmd->request_buffer            = NULL;
1691         scmd->request_bufflen           = 0;
1692
1693         scmd->cmd_len                   = 0;
1694
1695         scmd->sc_data_direction         = DMA_BIDIRECTIONAL;
1696         scmd->sc_request                = NULL;
1697         scmd->sc_magic                  = SCSI_CMND_MAGIC;
1698
1699         init_timer(&scmd->eh_timeout);
1700
1701         /*
1702          * Sometimes the command can get back into the timer chain,
1703          * so use the pid as an identifier.
1704          */
1705         scmd->pid                       = 0;
1706
1707         switch (flag) {
1708         case SCSI_TRY_RESET_DEVICE:
1709                 rtn = scsi_try_bus_device_reset(scmd);
1710                 if (rtn == SUCCESS)
1711                         break;
1712                 /* FALLTHROUGH */
1713         case SCSI_TRY_RESET_BUS:
1714                 rtn = scsi_try_bus_reset(scmd);
1715                 if (rtn == SUCCESS)
1716                         break;
1717                 /* FALLTHROUGH */
1718         case SCSI_TRY_RESET_HOST:
1719                 rtn = scsi_try_host_reset(scmd);
1720                 break;
1721         default:
1722                 rtn = FAILED;
1723         }
1724
1725         scsi_next_command(scmd);
1726         return rtn;
1727 }
1728 EXPORT_SYMBOL(scsi_reset_provider);
1729
1730 /**
1731  * scsi_normalize_sense - normalize main elements from either fixed or
1732  *                      descriptor sense data format into a common format.
1733  *
1734  * @sense_buffer:       byte array containing sense data returned by device
1735  * @sb_len:             number of valid bytes in sense_buffer
1736  * @sshdr:              pointer to instance of structure that common
1737  *                      elements are written to.
1738  *
1739  * Notes:
1740  *      The "main elements" from sense data are: response_code, sense_key,
1741  *      asc, ascq and additional_length (only for descriptor format).
1742  *
1743  *      Typically this function can be called after a device has
1744  *      responded to a SCSI command with the CHECK_CONDITION status.
1745  *
1746  * Return value:
1747  *      1 if valid sense data information found, else 0;
1748  **/
1749 int scsi_normalize_sense(const u8 *sense_buffer, int sb_len,
1750                          struct scsi_sense_hdr *sshdr)
1751 {
1752         if (!sense_buffer || !sb_len)
1753                 return 0;
1754
1755         memset(sshdr, 0, sizeof(struct scsi_sense_hdr));
1756
1757         sshdr->response_code = (sense_buffer[0] & 0x7f);
1758
1759         if (!scsi_sense_valid(sshdr))
1760                 return 0;
1761
1762         if (sshdr->response_code >= 0x72) {
1763                 /*
1764                  * descriptor format
1765                  */
1766                 if (sb_len > 1)
1767                         sshdr->sense_key = (sense_buffer[1] & 0xf);
1768                 if (sb_len > 2)
1769                         sshdr->asc = sense_buffer[2];
1770                 if (sb_len > 3)
1771                         sshdr->ascq = sense_buffer[3];
1772                 if (sb_len > 7)
1773                         sshdr->additional_length = sense_buffer[7];
1774         } else {
1775                 /*
1776                  * fixed format
1777                  */
1778                 if (sb_len > 2)
1779                         sshdr->sense_key = (sense_buffer[2] & 0xf);
1780                 if (sb_len > 7) {
1781                         sb_len = (sb_len < (sense_buffer[7] + 8)) ?
1782                                          sb_len : (sense_buffer[7] + 8);
1783                         if (sb_len > 12)
1784                                 sshdr->asc = sense_buffer[12];
1785                         if (sb_len > 13)
1786                                 sshdr->ascq = sense_buffer[13];
1787                 }
1788         }
1789
1790         return 1;
1791 }
1792 EXPORT_SYMBOL(scsi_normalize_sense);
1793
1794 int scsi_request_normalize_sense(struct scsi_request *sreq,
1795                                  struct scsi_sense_hdr *sshdr)
1796 {
1797         return scsi_normalize_sense(sreq->sr_sense_buffer,
1798                         sizeof(sreq->sr_sense_buffer), sshdr);
1799 }
1800 EXPORT_SYMBOL(scsi_request_normalize_sense);
1801
1802 int scsi_command_normalize_sense(struct scsi_cmnd *cmd,
1803                                  struct scsi_sense_hdr *sshdr)
1804 {
1805         return scsi_normalize_sense(cmd->sense_buffer,
1806                         sizeof(cmd->sense_buffer), sshdr);
1807 }
1808 EXPORT_SYMBOL(scsi_command_normalize_sense);
1809
1810 /**
1811  * scsi_sense_desc_find - search for a given descriptor type in
1812  *                      descriptor sense data format.
1813  *
1814  * @sense_buffer:       byte array of descriptor format sense data
1815  * @sb_len:             number of valid bytes in sense_buffer
1816  * @desc_type:          value of descriptor type to find
1817  *                      (e.g. 0 -> information)
1818  *
1819  * Notes:
1820  *      only valid when sense data is in descriptor format
1821  *
1822  * Return value:
1823  *      pointer to start of (first) descriptor if found else NULL
1824  **/
1825 const u8 * scsi_sense_desc_find(const u8 * sense_buffer, int sb_len,
1826                                 int desc_type)
1827 {
1828         int add_sen_len, add_len, desc_len, k;
1829         const u8 * descp;
1830
1831         if ((sb_len < 8) || (0 == (add_sen_len = sense_buffer[7])))
1832                 return NULL;
1833         if ((sense_buffer[0] < 0x72) || (sense_buffer[0] > 0x73))
1834                 return NULL;
1835         add_sen_len = (add_sen_len < (sb_len - 8)) ?
1836                         add_sen_len : (sb_len - 8);
1837         descp = &sense_buffer[8];
1838         for (desc_len = 0, k = 0; k < add_sen_len; k += desc_len) {
1839                 descp += desc_len;
1840                 add_len = (k < (add_sen_len - 1)) ? descp[1]: -1;
1841                 desc_len = add_len + 2;
1842                 if (descp[0] == desc_type)
1843                         return descp;
1844                 if (add_len < 0) // short descriptor ??
1845                         break;
1846         }
1847         return NULL;
1848 }
1849 EXPORT_SYMBOL(scsi_sense_desc_find);
1850
1851 /**
1852  * scsi_get_sense_info_fld - attempts to get information field from
1853  *                      sense data (either fixed or descriptor format)
1854  *
1855  * @sense_buffer:       byte array of sense data
1856  * @sb_len:             number of valid bytes in sense_buffer
1857  * @info_out:           pointer to 64 integer where 8 or 4 byte information
1858  *                      field will be placed if found.
1859  *
1860  * Return value:
1861  *      1 if information field found, 0 if not found.
1862  **/
1863 int scsi_get_sense_info_fld(const u8 * sense_buffer, int sb_len,
1864                             u64 * info_out)
1865 {
1866         int j;
1867         const u8 * ucp;
1868         u64 ull;
1869
1870         if (sb_len < 7)
1871                 return 0;
1872         switch (sense_buffer[0] & 0x7f) {
1873         case 0x70:
1874         case 0x71:
1875                 if (sense_buffer[0] & 0x80) {
1876                         *info_out = (sense_buffer[3] << 24) +
1877                                     (sense_buffer[4] << 16) +
1878                                     (sense_buffer[5] << 8) + sense_buffer[6];
1879                         return 1;
1880                 } else
1881                         return 0;
1882         case 0x72:
1883         case 0x73:
1884                 ucp = scsi_sense_desc_find(sense_buffer, sb_len,
1885                                            0 /* info desc */);
1886                 if (ucp && (0xa == ucp[1])) {
1887                         ull = 0;
1888                         for (j = 0; j < 8; ++j) {
1889                                 if (j > 0)
1890                                         ull <<= 8;
1891                                 ull |= ucp[4 + j];
1892                         }
1893                         *info_out = ull;
1894                         return 1;
1895                 } else
1896                         return 0;
1897         default:
1898                 return 0;
1899         }
1900 }
1901 EXPORT_SYMBOL(scsi_get_sense_info_fld);