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