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/gfp.h>
  20 #include <linux/timer.h>
  21 #include <linux/string.h>
  22 #include <linux/kernel.h>
  23 #include <linux/freezer.h>
  24 #include <linux/kthread.h>
  25 #include <linux/interrupt.h>
  26 #include <linux/blkdev.h>
  27 #include <linux/delay.h>
  28 #include <linux/jiffies.h>
  29
  30 #include <scsi/scsi.h>
  31 #include <scsi/scsi_cmnd.h>
  32 #include <scsi/scsi_dbg.h>
  33 #include <scsi/scsi_device.h>
  34 #include <scsi/scsi_driver.h>
  35 #include <scsi/scsi_eh.h>
  36 #include <scsi/scsi_transport.h>
  37 #include <scsi/scsi_host.h>
  38 #include <scsi/scsi_ioctl.h>
  39
  40 #include "scsi_priv.h"
  41 #include "scsi_logging.h"
  42 #include "scsi_transport_api.h"
  43
  44 #include <trace/events/scsi.h>
  45
  46 static void scsi_eh_done(struct scsi_cmnd *scmd);
  47
  48 /*
  49  * These should *probably* be handled by the host itself.
  50  * Since it is allowed to sleep, it probably should.
  51  */
  52 #define BUS_RESET_SETTLE_TIME   (10)
  53 #define HOST_RESET_SETTLE_TIME  (10)
  54
  55 static int scsi_eh_try_stu(struct scsi_cmnd *scmd);
  56
  57 /* called with shost->host_lock held */
  58 void scsi_eh_wakeup(struct Scsi_Host *shost)
  59 {
  60         if (shost->host_busy == shost->host_failed) {
  61                 trace_scsi_eh_wakeup(shost);
  62                 wake_up_process(shost->ehandler);
  63                 SCSI_LOG_ERROR_RECOVERY(5,
  64                                 printk("Waking error handler thread\n"));
  65         }
  66 }
  67
  68 /**
  69  * scsi_schedule_eh - schedule EH for SCSI host
  70  * @shost:      SCSI host to invoke error handling on.
  71  *
  72  * Schedule SCSI EH without scmd.
  73  */
  74 void scsi_schedule_eh(struct Scsi_Host *shost)
  75 {
  76         unsigned long flags;
  77
  78         spin_lock_irqsave(shost->host_lock, flags);
  79
  80         if (scsi_host_set_state(shost, SHOST_RECOVERY) == 0 ||
  81             scsi_host_set_state(shost, SHOST_CANCEL_RECOVERY) == 0) {
  82                 shost->host_eh_scheduled++;
  83                 scsi_eh_wakeup(shost);
  84         }
  85
  86         spin_unlock_irqrestore(shost->host_lock, flags);
  87 }
  88 EXPORT_SYMBOL_GPL(scsi_schedule_eh);
  89
  90 /**
  91  * scsi_eh_scmd_add - add scsi cmd to error handling.
  92  * @scmd:       scmd to run eh on.
  93  * @eh_flag:    optional SCSI_EH flag.
  94  *
  95  * Return value:
  96  *      0 on failure.
  97  */
  98 int scsi_eh_scmd_add(struct scsi_cmnd *scmd, int eh_flag)
  99 {
 100         struct Scsi_Host *shost = scmd->device->host;
 101         unsigned long flags;
 102         int ret = 0;
 103
 104         if (!shost->ehandler)
 105                 return 0;
 106
 107         spin_lock_irqsave(shost->host_lock, flags);
 108         if (scsi_host_set_state(shost, SHOST_RECOVERY))
 109                 if (scsi_host_set_state(shost, SHOST_CANCEL_RECOVERY))
 110                         goto out_unlock;
 111
 112         ret = 1;
 113         scmd->eh_eflags |= eh_flag;
 114         list_add_tail(&scmd->eh_entry, &shost->eh_cmd_q);
 115         shost->host_failed++;
 116         scsi_eh_wakeup(shost);
 117  out_unlock:
 118         spin_unlock_irqrestore(shost->host_lock, flags);
 119         return ret;
 120 }
 121
 122 /**
 123  * scsi_times_out - Timeout function for normal scsi commands.
 124  * @req:        request that is timing out.
 125  *
 126  * Notes:
 127  *     We do not need to lock this.  There is the potential for a race
 128  *     only in that the normal completion handling might run, but if the
 129  *     normal completion function determines that the timer has already
 130  *     fired, then it mustn't do anything.
 131  */
 132 enum blk_eh_timer_return scsi_times_out(struct request *req)
 133 {
 134         struct scsi_cmnd *scmd = req->special;
 135         enum blk_eh_timer_return rtn = BLK_EH_NOT_HANDLED;
 136         struct Scsi_Host *host = scmd->device->host;
 137
 138         trace_scsi_dispatch_cmd_timeout(scmd);
 139         scsi_log_completion(scmd, TIMEOUT_ERROR);
 140
 141         if (host->transportt->eh_timed_out)
 142                 rtn = host->transportt->eh_timed_out(scmd);
 143         else if (host->hostt->eh_timed_out)
 144                 rtn = host->hostt->eh_timed_out(scmd);
 145
 146         scmd->result |= DID_TIME_OUT << 16;
 147
 148         if (unlikely(rtn == BLK_EH_NOT_HANDLED &&
 149                      !scsi_eh_scmd_add(scmd, SCSI_EH_CANCEL_CMD)))
 150                 rtn = BLK_EH_HANDLED;
 151
 152         return rtn;
 153 }
 154
 155 /**
 156  * scsi_block_when_processing_errors - Prevent cmds from being queued.
 157  * @sdev:       Device on which we are performing recovery.
 158  *
 159  * Description:
 160  *     We block until the host is out of error recovery, and then check to
 161  *     see whether the host or the device is offline.
 162  *
 163  * Return value:
 164  *     0 when dev was taken offline by error recovery. 1 OK to proceed.
 165  */
 166 int scsi_block_when_processing_errors(struct scsi_device *sdev)
 167 {
 168         int online;
 169
 170         wait_event(sdev->host->host_wait, !scsi_host_in_recovery(sdev->host));
 171
 172         online = scsi_device_online(sdev);
 173
 174         SCSI_LOG_ERROR_RECOVERY(5, printk("%s: rtn: %d\n", __func__,
 175                                           online));
 176
 177         return online;
 178 }
 179 EXPORT_SYMBOL(scsi_block_when_processing_errors);
 180
 181 #ifdef CONFIG_SCSI_LOGGING
 182 /**
 183  * scsi_eh_prt_fail_stats - Log info on failures.
 184  * @shost:      scsi host being recovered.
 185  * @work_q:     Queue of scsi cmds to process.
 186  */
 187 static inline void scsi_eh_prt_fail_stats(struct Scsi_Host *shost,
 188                                           struct list_head *work_q)
 189 {
 190         struct scsi_cmnd *scmd;
 191         struct scsi_device *sdev;
 192         int total_failures = 0;
 193         int cmd_failed = 0;
 194         int cmd_cancel = 0;
 195         int devices_failed = 0;
 196
 197         shost_for_each_device(sdev, shost) {
 198                 list_for_each_entry(scmd, work_q, eh_entry) {
 199                         if (scmd->device == sdev) {
 200                                 ++total_failures;
 201                                 if (scmd->eh_eflags & SCSI_EH_CANCEL_CMD)
 202                                         ++cmd_cancel;
 203                                 else
 204                                         ++cmd_failed;
 205                         }
 206                 }
 207
 208                 if (cmd_cancel || cmd_failed) {
 209                         SCSI_LOG_ERROR_RECOVERY(3,
 210                                 sdev_printk(KERN_INFO, sdev,
 211                                             "%s: cmds failed: %d, cancel: %d\n",
 212                                             __func__, cmd_failed,
 213                                             cmd_cancel));
 214                         cmd_cancel = 0;
 215                         cmd_failed = 0;
 216                         ++devices_failed;
 217                 }
 218         }
 219
 220         SCSI_LOG_ERROR_RECOVERY(2, printk("Total of %d commands on %d"
 221                                           " devices require eh work\n",
 222                                    total_failures, devices_failed));
 223 }
 224 #endif
 225
 226 /**
 227  * scsi_check_sense - Examine scsi cmd sense
 228  * @scmd:       Cmd to have sense checked.
 229  *
 230  * Return value:
 231  *      SUCCESS or FAILED or NEEDS_RETRY or TARGET_ERROR
 232  *
 233  * Notes:
 234  *      When a deferred error is detected the current command has
 235  *      not been executed and needs retrying.
 236  */
 237 static int scsi_check_sense(struct scsi_cmnd *scmd)
 238 {
 239         struct scsi_device *sdev = scmd->device;
 240         struct scsi_sense_hdr sshdr;
 241
 242         if (! scsi_command_normalize_sense(scmd, &sshdr))
 243                 return FAILED;  /* no valid sense data */
 244
 245         if (scmd->cmnd[0] == TEST_UNIT_READY && scmd->scsi_done != scsi_eh_done)
 246                 /*
 247                  * nasty: for mid-layer issued TURs, we need to return the
 248                  * actual sense data without any recovery attempt.  For eh
 249                  * issued ones, we need to try to recover and interpret
 250                  */
 251                 return SUCCESS;
 252
 253         if (scsi_sense_is_deferred(&sshdr))
 254                 return NEEDS_RETRY;
 255
 256         if (sdev->scsi_dh_data && sdev->scsi_dh_data->scsi_dh &&
 257                         sdev->scsi_dh_data->scsi_dh->check_sense) {
 258                 int rc;
 259
 260                 rc = sdev->scsi_dh_data->scsi_dh->check_sense(sdev, &sshdr);
 261                 if (rc != SCSI_RETURN_NOT_HANDLED)
 262                         return rc;
 263                 /* handler does not care. Drop down to default handling */
 264         }
 265
 266         /*
 267          * Previous logic looked for FILEMARK, EOM or ILI which are
 268          * mainly associated with tapes and returned SUCCESS.
 269          */
 270         if (sshdr.response_code == 0x70) {
 271                 /* fixed format */
 272                 if (scmd->sense_buffer[2] & 0xe0)
 273                         return SUCCESS;
 274         } else {
 275                 /*
 276                  * descriptor format: look for "stream commands sense data
 277                  * descriptor" (see SSC-3). Assume single sense data
 278                  * descriptor. Ignore ILI from SBC-2 READ LONG and WRITE LONG.
 279                  */
 280                 if ((sshdr.additional_length > 3) &&
 281                     (scmd->sense_buffer[8] == 0x4) &&
 282                     (scmd->sense_buffer[11] & 0xe0))
 283                         return SUCCESS;
 284         }
 285
 286         switch (sshdr.sense_key) {
 287         case NO_SENSE:
 288                 return SUCCESS;
 289         case RECOVERED_ERROR:
 290                 return /* soft_error */ SUCCESS;
 291
 292         case ABORTED_COMMAND:
 293                 if (sshdr.asc == 0x10) /* DIF */
 294                         return SUCCESS;
 295
 296                 return NEEDS_RETRY;
 297         case NOT_READY:
 298         case UNIT_ATTENTION:
 299                 /*
 300                  * if we are expecting a cc/ua because of a bus reset that we
 301                  * performed, treat this just as a retry.  otherwise this is
 302                  * information that we should pass up to the upper-level driver
 303                  * so that we can deal with it there.
 304                  */
 305                 if (scmd->device->expecting_cc_ua) {
 306                         /*
 307                          * Because some device does not queue unit
 308                          * attentions correctly, we carefully check
 309                          * additional sense code and qualifier so as
 310                          * not to squash media change unit attention.
 311                          */
 312                         if (sshdr.asc != 0x28 || sshdr.ascq != 0x00) {
 313                                 scmd->device->expecting_cc_ua = 0;
 314                                 return NEEDS_RETRY;
 315                         }
 316                 }
 317                 /*
 318                  * if the device is in the process of becoming ready, we
 319                  * should retry.
 320                  */
 321                 if ((sshdr.asc == 0x04) && (sshdr.ascq == 0x01))
 322                         return NEEDS_RETRY;
 323                 /*
 324                  * if the device is not started, we need to wake
 325                  * the error handler to start the motor
 326                  */
 327                 if (scmd->device->allow_restart &&
 328                     (sshdr.asc == 0x04) && (sshdr.ascq == 0x02))
 329                         return FAILED;
 330
 331                 if (sshdr.asc == 0x3f && sshdr.ascq == 0x0e)
 332                         scmd_printk(KERN_WARNING, scmd,
 333                                     "Warning! Received an indication that the "
 334                                     "LUN assignments on this target have "
 335                                     "changed. The Linux SCSI layer does not "
 336                                     "automatically remap LUN assignments.\n");
 337                 else if (sshdr.asc == 0x3f)
 338                         scmd_printk(KERN_WARNING, scmd,
 339                                     "Warning! Received an indication that the "
 340                                     "operating parameters on this target have "
 341                                     "changed. The Linux SCSI layer does not "
 342                                     "automatically adjust these parameters.\n");
 343
 344                 if (sshdr.asc == 0x38 && sshdr.ascq == 0x07)
 345                         scmd_printk(KERN_WARNING, scmd,
 346                                     "Warning! Received an indication that the "
 347                                     "LUN reached a thin provisioning soft "
 348                                     "threshold.\n");
 349
 350                 /*
 351                  * Pass the UA upwards for a determination in the completion
 352                  * functions.
 353                  */
 354                 return SUCCESS;
 355
 356                 /* these are not supported */
 357         case COPY_ABORTED:
 358         case VOLUME_OVERFLOW:
 359         case MISCOMPARE:
 360         case BLANK_CHECK:
 361         case DATA_PROTECT:
 362                 return TARGET_ERROR;
 363
 364         case MEDIUM_ERROR:
 365                 if (sshdr.asc == 0x11 || /* UNRECOVERED READ ERR */
 366                     sshdr.asc == 0x13 || /* AMNF DATA FIELD */
 367                     sshdr.asc == 0x14) { /* RECORD NOT FOUND */
 368                         return TARGET_ERROR;
 369                 }
 370                 return NEEDS_RETRY;
 371
 372         case HARDWARE_ERROR:
 373                 if (scmd->device->retry_hwerror)
 374                         return ADD_TO_MLQUEUE;
 375                 else
 376                         return TARGET_ERROR;
 377
 378         case ILLEGAL_REQUEST:
 379                 if (sshdr.asc == 0x20 || /* Invalid command operation code */
 380                     sshdr.asc == 0x21 || /* Logical block address out of range */
 381                     sshdr.asc == 0x24 || /* Invalid field in cdb */
 382                     sshdr.asc == 0x26) { /* Parameter value invalid */
 383                         return TARGET_ERROR;
 384                 }
 385                 return SUCCESS;
 386
 387         default:
 388                 return SUCCESS;
 389         }
 390 }
 391
 392 static void scsi_handle_queue_ramp_up(struct scsi_device *sdev)
 393 {
 394         struct scsi_host_template *sht = sdev->host->hostt;
 395         struct scsi_device *tmp_sdev;
 396
 397         if (!sht->change_queue_depth ||
 398             sdev->queue_depth >= sdev->max_queue_depth)
 399                 return;
 400
 401         if (time_before(jiffies,
 402             sdev->last_queue_ramp_up + sdev->queue_ramp_up_period))
 403                 return;
 404
 405         if (time_before(jiffies,
 406             sdev->last_queue_full_time + sdev->queue_ramp_up_period))
 407                 return;
 408
 409         /*
 410          * Walk all devices of a target and do
 411          * ramp up on them.
 412          */
 413         shost_for_each_device(tmp_sdev, sdev->host) {
 414                 if (tmp_sdev->channel != sdev->channel ||
 415                     tmp_sdev->id != sdev->id ||
 416                     tmp_sdev->queue_depth == sdev->max_queue_depth)
 417                         continue;
 418                 /*
 419                  * call back into LLD to increase queue_depth by one
 420                  * with ramp up reason code.
 421                  */
 422                 sht->change_queue_depth(tmp_sdev, tmp_sdev->queue_depth + 1,
 423                                         SCSI_QDEPTH_RAMP_UP);
 424                 sdev->last_queue_ramp_up = jiffies;
 425         }
 426 }
 427
 428 static void scsi_handle_queue_full(struct scsi_device *sdev)
 429 {
 430         struct scsi_host_template *sht = sdev->host->hostt;
 431         struct scsi_device *tmp_sdev;
 432
 433         if (!sht->change_queue_depth)
 434                 return;
 435
 436         shost_for_each_device(tmp_sdev, sdev->host) {
 437                 if (tmp_sdev->channel != sdev->channel ||
 438                     tmp_sdev->id != sdev->id)
 439                         continue;
 440                 /*
 441                  * We do not know the number of commands that were at
 442                  * the device when we got the queue full so we start
 443                  * from the highest possible value and work our way down.
 444                  */
 445                 sht->change_queue_depth(tmp_sdev, tmp_sdev->queue_depth - 1,
 446                                         SCSI_QDEPTH_QFULL);
 447         }
 448 }
 449
 450 /**
 451  * scsi_eh_completed_normally - Disposition a eh cmd on return from LLD.
 452  * @scmd:       SCSI cmd to examine.
 453  *
 454  * Notes:
 455  *    This is *only* called when we are examining the status of commands
 456  *    queued during error recovery.  the main difference here is that we
 457  *    don't allow for the possibility of retries here, and we are a lot
 458  *    more restrictive about what we consider acceptable.
 459  */
 460 static int scsi_eh_completed_normally(struct scsi_cmnd *scmd)
 461 {
 462         /*
 463          * first check the host byte, to see if there is anything in there
 464          * that would indicate what we need to do.
 465          */
 466         if (host_byte(scmd->result) == DID_RESET) {
 467                 /*
 468                  * rats.  we are already in the error handler, so we now
 469                  * get to try and figure out what to do next.  if the sense
 470                  * is valid, we have a pretty good idea of what to do.
 471                  * if not, we mark it as FAILED.
 472                  */
 473                 return scsi_check_sense(scmd);
 474         }
 475         if (host_byte(scmd->result) != DID_OK)
 476                 return FAILED;
 477
 478         /*
 479          * next, check the message byte.
 480          */
 481         if (msg_byte(scmd->result) != COMMAND_COMPLETE)
 482                 return FAILED;
 483
 484         /*
 485          * now, check the status byte to see if this indicates
 486          * anything special.
 487          */
 488         switch (status_byte(scmd->result)) {
 489         case GOOD:
 490                 scsi_handle_queue_ramp_up(scmd->device);
 491         case COMMAND_TERMINATED:
 492                 return SUCCESS;
 493         case CHECK_CONDITION:
 494                 return scsi_check_sense(scmd);
 495         case CONDITION_GOOD:
 496         case INTERMEDIATE_GOOD:
 497         case INTERMEDIATE_C_GOOD:
 498                 /*
 499                  * who knows?  FIXME(eric)
 500                  */
 501                 return SUCCESS;
 502         case RESERVATION_CONFLICT:
 503                 if (scmd->cmnd[0] == TEST_UNIT_READY)
 504                         /* it is a success, we probed the device and
 505                          * found it */
 506                         return SUCCESS;
 507                 /* otherwise, we failed to send the command */
 508                 return FAILED;
 509         case QUEUE_FULL:
 510                 scsi_handle_queue_full(scmd->device);
 511                 /* fall through */
 512         case BUSY:
 513                 return NEEDS_RETRY;
 514         default:
 515                 return FAILED;
 516         }
 517         return FAILED;
 518 }
 519
 520 /**
 521  * scsi_eh_done - Completion function for error handling.
 522  * @scmd:       Cmd that is done.
 523  */
 524 static void scsi_eh_done(struct scsi_cmnd *scmd)
 525 {
 526         struct completion *eh_action;
 527
 528         SCSI_LOG_ERROR_RECOVERY(3,
 529                 printk("%s scmd: %p result: %x\n",
 530                         __func__, scmd, scmd->result));
 531
 532         eh_action = scmd->device->host->eh_action;
 533         if (eh_action)
 534                 complete(eh_action);
 535 }
 536
 537 /**
 538  * scsi_try_host_reset - ask host adapter to reset itself
 539  * @scmd:       SCSI cmd to send host reset.
 540  */
 541 static int scsi_try_host_reset(struct scsi_cmnd *scmd)
 542 {
 543         unsigned long flags;
 544         int rtn;
 545         struct Scsi_Host *host = scmd->device->host;
 546         struct scsi_host_template *hostt = host->hostt;
 547
 548         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Snd Host RST\n",
 549                                           __func__));
 550
 551         if (!hostt->eh_host_reset_handler)
 552                 return FAILED;
 553
 554         rtn = hostt->eh_host_reset_handler(scmd);
 555
 556         if (rtn == SUCCESS) {
 557                 if (!hostt->skip_settle_delay)
 558                         ssleep(HOST_RESET_SETTLE_TIME);
 559                 spin_lock_irqsave(host->host_lock, flags);
 560                 scsi_report_bus_reset(host, scmd_channel(scmd));
 561                 spin_unlock_irqrestore(host->host_lock, flags);
 562         }
 563
 564         return rtn;
 565 }
 566
 567 /**
 568  * scsi_try_bus_reset - ask host to perform a bus reset
 569  * @scmd:       SCSI cmd to send bus reset.
 570  */
 571 static int scsi_try_bus_reset(struct scsi_cmnd *scmd)
 572 {
 573         unsigned long flags;
 574         int rtn;
 575         struct Scsi_Host *host = scmd->device->host;
 576         struct scsi_host_template *hostt = host->hostt;
 577
 578         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Snd Bus RST\n",
 579                                           __func__));
 580
 581         if (!hostt->eh_bus_reset_handler)
 582                 return FAILED;
 583
 584         rtn = hostt->eh_bus_reset_handler(scmd);
 585
 586         if (rtn == SUCCESS) {
 587                 if (!hostt->skip_settle_delay)
 588                         ssleep(BUS_RESET_SETTLE_TIME);
 589                 spin_lock_irqsave(host->host_lock, flags);
 590                 scsi_report_bus_reset(host, scmd_channel(scmd));
 591                 spin_unlock_irqrestore(host->host_lock, flags);
 592         }
 593
 594         return rtn;
 595 }
 596
 597 static void __scsi_report_device_reset(struct scsi_device *sdev, void *data)
 598 {
 599         sdev->was_reset = 1;
 600         sdev->expecting_cc_ua = 1;
 601 }
 602
 603 /**
 604  * scsi_try_target_reset - Ask host to perform a target reset
 605  * @scmd:       SCSI cmd used to send a target reset
 606  *
 607  * Notes:
 608  *    There is no timeout for this operation.  if this operation is
 609  *    unreliable for a given host, then the host itself needs to put a
 610  *    timer on it, and set the host back to a consistent state prior to
 611  *    returning.
 612  */
 613 static int scsi_try_target_reset(struct scsi_cmnd *scmd)
 614 {
 615         unsigned long flags;
 616         int rtn;
 617         struct Scsi_Host *host = scmd->device->host;
 618         struct scsi_host_template *hostt = host->hostt;
 619
 620         if (!hostt->eh_target_reset_handler)
 621                 return FAILED;
 622
 623         rtn = hostt->eh_target_reset_handler(scmd);
 624         if (rtn == SUCCESS) {
 625                 spin_lock_irqsave(host->host_lock, flags);
 626                 __starget_for_each_device(scsi_target(scmd->device), NULL,
 627                                           __scsi_report_device_reset);
 628                 spin_unlock_irqrestore(host->host_lock, flags);
 629         }
 630
 631         return rtn;
 632 }
 633
 634 /**
 635  * scsi_try_bus_device_reset - Ask host to perform a BDR on a dev
 636  * @scmd:       SCSI cmd used to send BDR
 637  *
 638  * Notes:
 639  *    There is no timeout for this operation.  if this operation is
 640  *    unreliable for a given host, then the host itself needs to put a
 641  *    timer on it, and set the host back to a consistent state prior to
 642  *    returning.
 643  */
 644 static int scsi_try_bus_device_reset(struct scsi_cmnd *scmd)
 645 {
 646         int rtn;
 647         struct scsi_host_template *hostt = scmd->device->host->hostt;
 648
 649         if (!hostt->eh_device_reset_handler)
 650                 return FAILED;
 651
 652         rtn = hostt->eh_device_reset_handler(scmd);
 653         if (rtn == SUCCESS)
 654                 __scsi_report_device_reset(scmd->device, NULL);
 655         return rtn;
 656 }
 657
 658 static int scsi_try_to_abort_cmd(struct scsi_host_template *hostt, struct scsi_cmnd *scmd)
 659 {
 660         if (!hostt->eh_abort_handler)
 661                 return FAILED;
 662
 663         return hostt->eh_abort_handler(scmd);
 664 }
 665
 666 static void scsi_abort_eh_cmnd(struct scsi_cmnd *scmd)
 667 {
 668         if (scsi_try_to_abort_cmd(scmd->device->host->hostt, scmd) != SUCCESS)
 669                 if (scsi_try_bus_device_reset(scmd) != SUCCESS)
 670                         if (scsi_try_target_reset(scmd) != SUCCESS)
 671                                 if (scsi_try_bus_reset(scmd) != SUCCESS)
 672                                         scsi_try_host_reset(scmd);
 673 }
 674
 675 /**
 676  * scsi_eh_prep_cmnd  - Save a scsi command info as part of error recovery
 677  * @scmd:       SCSI command structure to hijack
 678  * @ses:        structure to save restore information
 679  * @cmnd:       CDB to send. Can be NULL if no new cmnd is needed
 680  * @cmnd_size:  size in bytes of @cmnd (must be <= BLK_MAX_CDB)
 681  * @sense_bytes: size of sense data to copy. or 0 (if != 0 @cmnd is ignored)
 682  *
 683  * This function is used to save a scsi command information before re-execution
 684  * as part of the error recovery process.  If @sense_bytes is 0 the command
 685  * sent must be one that does not transfer any data.  If @sense_bytes != 0
 686  * @cmnd is ignored and this functions sets up a REQUEST_SENSE command
 687  * and cmnd buffers to read @sense_bytes into @scmd->sense_buffer.
 688  */
 689 void scsi_eh_prep_cmnd(struct scsi_cmnd *scmd, struct scsi_eh_save *ses,
 690                         unsigned char *cmnd, int cmnd_size, unsigned sense_bytes)
 691 {
 692         struct scsi_device *sdev = scmd->device;
 693
 694         /*
 695          * We need saved copies of a number of fields - this is because
 696          * error handling may need to overwrite these with different values
 697          * to run different commands, and once error handling is complete,
 698          * we will need to restore these values prior to running the actual
 699          * command.
 700          */
 701         ses->cmd_len = scmd->cmd_len;
 702         ses->cmnd = scmd->cmnd;
 703         ses->data_direction = scmd->sc_data_direction;
 704         ses->sdb = scmd->sdb;
 705         ses->next_rq = scmd->request->next_rq;
 706         ses->result = scmd->result;
 707         ses->underflow = scmd->underflow;
 708         ses->prot_op = scmd->prot_op;
 709
 710         scmd->prot_op = SCSI_PROT_NORMAL;
 711         scmd->cmnd = ses->eh_cmnd;
 712         memset(scmd->cmnd, 0, BLK_MAX_CDB);
 713         memset(&scmd->sdb, 0, sizeof(scmd->sdb));
 714         scmd->request->next_rq = NULL;
 715
 716         if (sense_bytes) {
 717                 scmd->sdb.length = min_t(unsigned, SCSI_SENSE_BUFFERSIZE,
 718                                          sense_bytes);
 719                 sg_init_one(&ses->sense_sgl, scmd->sense_buffer,
 720                             scmd->sdb.length);
 721                 scmd->sdb.table.sgl = &ses->sense_sgl;
 722                 scmd->sc_data_direction = DMA_FROM_DEVICE;
 723                 scmd->sdb.table.nents = 1;
 724                 scmd->cmnd[0] = REQUEST_SENSE;
 725                 scmd->cmnd[4] = scmd->sdb.length;
 726                 scmd->cmd_len = COMMAND_SIZE(scmd->cmnd[0]);
 727         } else {
 728                 scmd->sc_data_direction = DMA_NONE;
 729                 if (cmnd) {
 730                         BUG_ON(cmnd_size > BLK_MAX_CDB);
 731                         memcpy(scmd->cmnd, cmnd, cmnd_size);
 732                         scmd->cmd_len = COMMAND_SIZE(scmd->cmnd[0]);
 733                 }
 734         }
 735
 736         scmd->underflow = 0;
 737
 738         if (sdev->scsi_level <= SCSI_2 && sdev->scsi_level != SCSI_UNKNOWN)
 739                 scmd->cmnd[1] = (scmd->cmnd[1] & 0x1f) |
 740                         (sdev->lun << 5 & 0xe0);
 741
 742         /*
 743          * Zero the sense buffer.  The scsi spec mandates that any
 744          * untransferred sense data should be interpreted as being zero.
 745          */
 746         memset(scmd->sense_buffer, 0, SCSI_SENSE_BUFFERSIZE);
 747 }
 748 EXPORT_SYMBOL(scsi_eh_prep_cmnd);
 749
 750 /**
 751  * scsi_eh_restore_cmnd  - Restore a scsi command info as part of error recovery
 752  * @scmd:       SCSI command structure to restore
 753  * @ses:        saved information from a coresponding call to scsi_eh_prep_cmnd
 754  *
 755  * Undo any damage done by above scsi_eh_prep_cmnd().
 756  */
 757 void scsi_eh_restore_cmnd(struct scsi_cmnd* scmd, struct scsi_eh_save *ses)
 758 {
 759         /*
 760          * Restore original data
 761          */
 762         scmd->cmd_len = ses->cmd_len;
 763         scmd->cmnd = ses->cmnd;
 764         scmd->sc_data_direction = ses->data_direction;
 765         scmd->sdb = ses->sdb;
 766         scmd->request->next_rq = ses->next_rq;
 767         scmd->result = ses->result;
 768         scmd->underflow = ses->underflow;
 769         scmd->prot_op = ses->prot_op;
 770 }
 771 EXPORT_SYMBOL(scsi_eh_restore_cmnd);
 772
 773 /**
 774  * scsi_send_eh_cmnd  - submit a scsi command as part of error recovery
 775  * @scmd:       SCSI command structure to hijack
 776  * @cmnd:       CDB to send
 777  * @cmnd_size:  size in bytes of @cmnd
 778  * @timeout:    timeout for this request
 779  * @sense_bytes: size of sense data to copy or 0
 780  *
 781  * This function is used to send a scsi command down to a target device
 782  * as part of the error recovery process. See also scsi_eh_prep_cmnd() above.
 783  *
 784  * Return value:
 785  *    SUCCESS or FAILED or NEEDS_RETRY
 786  */
 787 static int scsi_send_eh_cmnd(struct scsi_cmnd *scmd, unsigned char *cmnd,
 788                              int cmnd_size, int timeout, unsigned sense_bytes)
 789 {
 790         struct scsi_device *sdev = scmd->device;
 791         struct Scsi_Host *shost = sdev->host;
 792         DECLARE_COMPLETION_ONSTACK(done);
 793         unsigned long timeleft = timeout;
 794         struct scsi_eh_save ses;
 795         const unsigned long stall_for = msecs_to_jiffies(100);
 796         int rtn;
 797
 798 retry:
 799         scsi_eh_prep_cmnd(scmd, &ses, cmnd, cmnd_size, sense_bytes);
 800         shost->eh_action = &done;
 801
 802         scsi_log_send(scmd);
 803         scmd->scsi_done = scsi_eh_done;
 804         rtn = shost->hostt->queuecommand(shost, scmd);
 805         if (rtn) {
 806                 if (timeleft > stall_for) {
 807                         scsi_eh_restore_cmnd(scmd, &ses);
 808                         timeleft -= stall_for;
 809                         msleep(jiffies_to_msecs(stall_for));
 810                         goto retry;
 811                 }
 812                 /* signal not to enter either branch of the if () below */
 813                 timeleft = 0;
 814                 rtn = NEEDS_RETRY;
 815         } else {
 816                 timeleft = wait_for_completion_timeout(&done, timeout);
 817         }
 818
 819         shost->eh_action = NULL;
 820
 821         scsi_log_completion(scmd, rtn);
 822
 823         SCSI_LOG_ERROR_RECOVERY(3,
 824                 printk("%s: scmd: %p, timeleft: %ld\n",
 825                         __func__, scmd, timeleft));
 826
 827         /*
 828          * If there is time left scsi_eh_done got called, and we will examine
 829          * the actual status codes to see whether the command actually did
 830          * complete normally, else if we have a zero return and no time left,
 831          * the command must still be pending, so abort it and return FAILED.
 832          * If we never actually managed to issue the command, because
 833          * ->queuecommand() kept returning non zero, use the rtn = FAILED
 834          * value above (so don't execute either branch of the if)
 835          */
 836         if (timeleft) {
 837                 rtn = scsi_eh_completed_normally(scmd);
 838                 SCSI_LOG_ERROR_RECOVERY(3,
 839                         printk("%s: scsi_eh_completed_normally %x\n",
 840                                __func__, rtn));
 841
 842                 switch (rtn) {
 843                 case SUCCESS:
 844                 case NEEDS_RETRY:
 845                 case FAILED:
 846                 case TARGET_ERROR:
 847                         break;
 848                 case ADD_TO_MLQUEUE:
 849                         rtn = NEEDS_RETRY;
 850                         break;
 851                 default:
 852                         rtn = FAILED;
 853                         break;
 854                 }
 855         } else if (!rtn) {
 856                 scsi_abort_eh_cmnd(scmd);
 857                 rtn = FAILED;
 858         }
 859
 860         scsi_eh_restore_cmnd(scmd, &ses);
 861
 862         if (scmd->request->cmd_type != REQ_TYPE_BLOCK_PC) {
 863                 struct scsi_driver *sdrv = scsi_cmd_to_driver(scmd);
 864                 if (sdrv->eh_action)
 865                         rtn = sdrv->eh_action(scmd, cmnd, cmnd_size, rtn);
 866         }
 867
 868         return rtn;
 869 }
 870
 871 /**
 872  * scsi_request_sense - Request sense data from a particular target.
 873  * @scmd:       SCSI cmd for request sense.
 874  *
 875  * Notes:
 876  *    Some hosts automatically obtain this information, others require
 877  *    that we obtain it on our own. This function will *not* return until
 878  *    the command either times out, or it completes.
 879  */
 880 static int scsi_request_sense(struct scsi_cmnd *scmd)
 881 {
 882         return scsi_send_eh_cmnd(scmd, NULL, 0, scmd->device->eh_timeout, ~0);
 883 }
 884
 885 /**
 886  * scsi_eh_finish_cmd - Handle a cmd that eh is finished with.
 887  * @scmd:       Original SCSI cmd that eh has finished.
 888  * @done_q:     Queue for processed commands.
 889  *
 890  * Notes:
 891  *    We don't want to use the normal command completion while we are are
 892  *    still handling errors - it may cause other commands to be queued,
 893  *    and that would disturb what we are doing.  Thus we really want to
 894  *    keep a list of pending commands for final completion, and once we
 895  *    are ready to leave error handling we handle completion for real.
 896  */
 897 void scsi_eh_finish_cmd(struct scsi_cmnd *scmd, struct list_head *done_q)
 898 {
 899         scmd->device->host->host_failed--;
 900         scmd->eh_eflags = 0;
 901         list_move_tail(&scmd->eh_entry, done_q);
 902 }
 903 EXPORT_SYMBOL(scsi_eh_finish_cmd);
 904
 905 /**
 906  * scsi_eh_get_sense - Get device sense data.
 907  * @work_q:     Queue of commands to process.
 908  * @done_q:     Queue of processed commands.
 909  *
 910  * Description:
 911  *    See if we need to request sense information.  if so, then get it
 912  *    now, so we have a better idea of what to do.
 913  *
 914  * Notes:
 915  *    This has the unfortunate side effect that if a shost adapter does
 916  *    not automatically request sense information, we end up shutting
 917  *    it down before we request it.
 918  *
 919  *    All drivers should request sense information internally these days,
 920  *    so for now all I have to say is tough noogies if you end up in here.
 921  *
 922  *    XXX: Long term this code should go away, but that needs an audit of
 923  *         all LLDDs first.
 924  */
 925 int scsi_eh_get_sense(struct list_head *work_q,
 926                       struct list_head *done_q)
 927 {
 928         struct scsi_cmnd *scmd, *next;
 929         int rtn;
 930
 931         list_for_each_entry_safe(scmd, next, work_q, eh_entry) {
 932                 if ((scmd->eh_eflags & SCSI_EH_CANCEL_CMD) ||
 933                     SCSI_SENSE_VALID(scmd))
 934                         continue;
 935
 936                 SCSI_LOG_ERROR_RECOVERY(2, scmd_printk(KERN_INFO, scmd,
 937                                                   "%s: requesting sense\n",
 938                                                   current->comm));
 939                 rtn = scsi_request_sense(scmd);
 940                 if (rtn != SUCCESS)
 941                         continue;
 942
 943                 SCSI_LOG_ERROR_RECOVERY(3, printk("sense requested for %p"
 944                                                   " result %x\n", scmd,
 945                                                   scmd->result));
 946                 SCSI_LOG_ERROR_RECOVERY(3, scsi_print_sense("bh", scmd));
 947
 948                 rtn = scsi_decide_disposition(scmd);
 949
 950                 /*
 951                  * if the result was normal, then just pass it along to the
 952                  * upper level.
 953                  */
 954                 if (rtn == SUCCESS)
 955                         /* we don't want this command reissued, just
 956                          * finished with the sense data, so set
 957                          * retries to the max allowed to ensure it
 958                          * won't get reissued */
 959                         scmd->retries = scmd->allowed;
 960                 else if (rtn != NEEDS_RETRY)
 961                         continue;
 962
 963                 scsi_eh_finish_cmd(scmd, done_q);
 964         }
 965
 966         return list_empty(work_q);
 967 }
 968 EXPORT_SYMBOL_GPL(scsi_eh_get_sense);
 969
 970 /**
 971  * scsi_eh_tur - Send TUR to device.
 972  * @scmd:       &scsi_cmnd to send TUR
 973  *
 974  * Return value:
 975  *    0 - Device is ready. 1 - Device NOT ready.
 976  */
 977 static int scsi_eh_tur(struct scsi_cmnd *scmd)
 978 {
 979         static unsigned char tur_command[6] = {TEST_UNIT_READY, 0, 0, 0, 0, 0};
 980         int retry_cnt = 1, rtn;
 981
 982 retry_tur:
 983         rtn = scsi_send_eh_cmnd(scmd, tur_command, 6,
 984                                 scmd->device->eh_timeout, 0);
 985
 986         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: scmd %p rtn %x\n",
 987                 __func__, scmd, rtn));
 988
 989         switch (rtn) {
 990         case NEEDS_RETRY:
 991                 if (retry_cnt--)
 992                         goto retry_tur;
 993                 /*FALLTHRU*/
 994         case SUCCESS:
 995                 return 0;
 996         default:
 997                 return 1;
 998         }
 999 }
1000
1001 /**
1002  * scsi_eh_test_devices - check if devices are responding from error recovery.
1003  * @cmd_list:   scsi commands in error recovery.
1004  * @work_q:     queue for commands which still need more error recovery
1005  * @done_q:     queue for commands which are finished
1006  * @try_stu:    boolean on if a STU command should be tried in addition to TUR.
1007  *
1008  * Decription:
1009  *    Tests if devices are in a working state.  Commands to devices now in
1010  *    a working state are sent to the done_q while commands to devices which
1011  *    are still failing to respond are returned to the work_q for more
1012  *    processing.
1013  **/
1014 static int scsi_eh_test_devices(struct list_head *cmd_list,
1015                                 struct list_head *work_q,
1016                                 struct list_head *done_q, int try_stu)
1017 {
1018         struct scsi_cmnd *scmd, *next;
1019         struct scsi_device *sdev;
1020         int finish_cmds;
1021
1022         while (!list_empty(cmd_list)) {
1023                 scmd = list_entry(cmd_list->next, struct scsi_cmnd, eh_entry);
1024                 sdev = scmd->device;
1025
1026                 finish_cmds = !scsi_device_online(scmd->device) ||
1027                         (try_stu && !scsi_eh_try_stu(scmd) &&
1028                          !scsi_eh_tur(scmd)) ||
1029                         !scsi_eh_tur(scmd);
1030
1031                 list_for_each_entry_safe(scmd, next, cmd_list, eh_entry)
1032                         if (scmd->device == sdev) {
1033                                 if (finish_cmds)
1034                                         scsi_eh_finish_cmd(scmd, done_q);
1035                                 else
1036                                         list_move_tail(&scmd->eh_entry, work_q);
1037                         }
1038         }
1039         return list_empty(work_q);
1040 }
1041
1042
1043 /**
1044  * scsi_eh_abort_cmds - abort pending commands.
1045  * @work_q:     &list_head for pending commands.
1046  * @done_q:     &list_head for processed commands.
1047  *
1048  * Decription:
1049  *    Try and see whether or not it makes sense to try and abort the
1050  *    running command.  This only works out to be the case if we have one
1051  *    command that has timed out.  If the command simply failed, it makes
1052  *    no sense to try and abort the command, since as far as the shost
1053  *    adapter is concerned, it isn't running.
1054  */
1055 static int scsi_eh_abort_cmds(struct list_head *work_q,
1056                               struct list_head *done_q)
1057 {
1058         struct scsi_cmnd *scmd, *next;
1059         LIST_HEAD(check_list);
1060         int rtn;
1061
1062         list_for_each_entry_safe(scmd, next, work_q, eh_entry) {
1063                 if (!(scmd->eh_eflags & SCSI_EH_CANCEL_CMD))
1064                         continue;
1065                 SCSI_LOG_ERROR_RECOVERY(3, printk("%s: aborting cmd:"
1066                                                   "0x%p\n", current->comm,
1067                                                   scmd));
1068                 rtn = scsi_try_to_abort_cmd(scmd->device->host->hostt, scmd);
1069                 if (rtn == SUCCESS || rtn == FAST_IO_FAIL) {
1070                         scmd->eh_eflags &= ~SCSI_EH_CANCEL_CMD;
1071                         if (rtn == FAST_IO_FAIL)
1072                                 scsi_eh_finish_cmd(scmd, done_q);
1073                         else
1074                                 list_move_tail(&scmd->eh_entry, &check_list);
1075                 } else
1076                         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: aborting"
1077                                                           " cmd failed:"
1078                                                           "0x%p\n",
1079                                                           current->comm,
1080                                                           scmd));
1081         }
1082
1083         return scsi_eh_test_devices(&check_list, work_q, done_q, 0);
1084 }
1085
1086 /**
1087  * scsi_eh_try_stu - Send START_UNIT to device.
1088  * @scmd:       &scsi_cmnd to send START_UNIT
1089  *
1090  * Return value:
1091  *    0 - Device is ready. 1 - Device NOT ready.
1092  */
1093 static int scsi_eh_try_stu(struct scsi_cmnd *scmd)
1094 {
1095         static unsigned char stu_command[6] = {START_STOP, 0, 0, 0, 1, 0};
1096
1097         if (scmd->device->allow_restart) {
1098                 int i, rtn = NEEDS_RETRY;
1099
1100                 for (i = 0; rtn == NEEDS_RETRY && i < 2; i++)
1101                         rtn = scsi_send_eh_cmnd(scmd, stu_command, 6, scmd->device->request_queue->rq_timeout, 0);
1102
1103                 if (rtn == SUCCESS)
1104                         return 0;
1105         }
1106
1107         return 1;
1108 }
1109
1110  /**
1111  * scsi_eh_stu - send START_UNIT if needed
1112  * @shost:      &scsi host being recovered.
1113  * @work_q:     &list_head for pending commands.
1114  * @done_q:     &list_head for processed commands.
1115  *
1116  * Notes:
1117  *    If commands are failing due to not ready, initializing command required,
1118  *      try revalidating the device, which will end up sending a start unit.
1119  */
1120 static int scsi_eh_stu(struct Scsi_Host *shost,
1121                               struct list_head *work_q,
1122                               struct list_head *done_q)
1123 {
1124         struct scsi_cmnd *scmd, *stu_scmd, *next;
1125         struct scsi_device *sdev;
1126
1127         shost_for_each_device(sdev, shost) {
1128                 stu_scmd = NULL;
1129                 list_for_each_entry(scmd, work_q, eh_entry)
1130                         if (scmd->device == sdev && SCSI_SENSE_VALID(scmd) &&
1131                             scsi_check_sense(scmd) == FAILED ) {
1132                                 stu_scmd = scmd;
1133                                 break;
1134                         }
1135
1136                 if (!stu_scmd)
1137                         continue;
1138
1139                 SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Sending START_UNIT to sdev:"
1140                                                   " 0x%p\n", current->comm, sdev));
1141
1142                 if (!scsi_eh_try_stu(stu_scmd)) {
1143                         if (!scsi_device_online(sdev) ||
1144                             !scsi_eh_tur(stu_scmd)) {
1145                                 list_for_each_entry_safe(scmd, next,
1146                                                           work_q, eh_entry) {
1147                                         if (scmd->device == sdev)
1148                                                 scsi_eh_finish_cmd(scmd, done_q);
1149                                 }
1150                         }
1151                 } else {
1152                         SCSI_LOG_ERROR_RECOVERY(3,
1153                                                 printk("%s: START_UNIT failed to sdev:"
1154                                                        " 0x%p\n", current->comm, sdev));
1155                 }
1156         }
1157
1158         return list_empty(work_q);
1159 }
1160
1161
1162 /**
1163  * scsi_eh_bus_device_reset - send bdr if needed
1164  * @shost:      scsi host being recovered.
1165  * @work_q:     &list_head for pending commands.
1166  * @done_q:     &list_head for processed commands.
1167  *
1168  * Notes:
1169  *    Try a bus device reset.  Still, look to see whether we have multiple
1170  *    devices that are jammed or not - if we have multiple devices, it
1171  *    makes no sense to try bus_device_reset - we really would need to try
1172  *    a bus_reset instead.
1173  */
1174 static int scsi_eh_bus_device_reset(struct Scsi_Host *shost,
1175                                     struct list_head *work_q,
1176                                     struct list_head *done_q)
1177 {
1178         struct scsi_cmnd *scmd, *bdr_scmd, *next;
1179         struct scsi_device *sdev;
1180         int rtn;
1181
1182         shost_for_each_device(sdev, shost) {
1183                 bdr_scmd = NULL;
1184                 list_for_each_entry(scmd, work_q, eh_entry)
1185                         if (scmd->device == sdev) {
1186                                 bdr_scmd = scmd;
1187                                 break;
1188                         }
1189
1190                 if (!bdr_scmd)
1191                         continue;
1192
1193                 SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Sending BDR sdev:"
1194                                                   " 0x%p\n", current->comm,
1195                                                   sdev));
1196                 rtn = scsi_try_bus_device_reset(bdr_scmd);
1197                 if (rtn == SUCCESS || rtn == FAST_IO_FAIL) {
1198                         if (!scsi_device_online(sdev) ||
1199                             rtn == FAST_IO_FAIL ||
1200                             !scsi_eh_tur(bdr_scmd)) {
1201                                 list_for_each_entry_safe(scmd, next,
1202                                                          work_q, eh_entry) {
1203                                         if (scmd->device == sdev)
1204                                                 scsi_eh_finish_cmd(scmd,
1205                                                                    done_q);
1206                                 }
1207                         }
1208                 } else {
1209                         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: BDR"
1210                                                           " failed sdev:"
1211                                                           "0x%p\n",
1212                                                           current->comm,
1213                                                            sdev));
1214                 }
1215         }
1216
1217         return list_empty(work_q);
1218 }
1219
1220 /**
1221  * scsi_eh_target_reset - send target reset if needed
1222  * @shost:      scsi host being recovered.
1223  * @work_q:     &list_head for pending commands.
1224  * @done_q:     &list_head for processed commands.
1225  *
1226  * Notes:
1227  *    Try a target reset.
1228  */
1229 static int scsi_eh_target_reset(struct Scsi_Host *shost,
1230                                 struct list_head *work_q,
1231                                 struct list_head *done_q)
1232 {
1233         LIST_HEAD(tmp_list);
1234         LIST_HEAD(check_list);
1235
1236         list_splice_init(work_q, &tmp_list);
1237
1238         while (!list_empty(&tmp_list)) {
1239                 struct scsi_cmnd *next, *scmd;
1240                 int rtn;
1241                 unsigned int id;
1242
1243                 scmd = list_entry(tmp_list.next, struct scsi_cmnd, eh_entry);
1244                 id = scmd_id(scmd);
1245
1246                 SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Sending target reset "
1247                                                   "to target %d\n",
1248                                                   current->comm, id));
1249                 rtn = scsi_try_target_reset(scmd);
1250                 if (rtn != SUCCESS && rtn != FAST_IO_FAIL)
1251                         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Target reset"
1252                                                           " failed target: "
1253                                                           "%d\n",
1254                                                           current->comm, id));
1255                 list_for_each_entry_safe(scmd, next, &tmp_list, eh_entry) {
1256                         if (scmd_id(scmd) != id)
1257                                 continue;
1258
1259                         if (rtn == SUCCESS)
1260                                 list_move_tail(&scmd->eh_entry, &check_list);
1261                         else if (rtn == FAST_IO_FAIL)
1262                                 scsi_eh_finish_cmd(scmd, done_q);
1263                         else
1264                                 /* push back on work queue for further processing */
1265                                 list_move(&scmd->eh_entry, work_q);
1266                 }
1267         }
1268
1269         return scsi_eh_test_devices(&check_list, work_q, done_q, 0);
1270 }
1271
1272 /**
1273  * scsi_eh_bus_reset - send a bus reset
1274  * @shost:      &scsi host being recovered.
1275  * @work_q:     &list_head for pending commands.
1276  * @done_q:     &list_head for processed commands.
1277  */
1278 static int scsi_eh_bus_reset(struct Scsi_Host *shost,
1279                              struct list_head *work_q,
1280                              struct list_head *done_q)
1281 {
1282         struct scsi_cmnd *scmd, *chan_scmd, *next;
1283         LIST_HEAD(check_list);
1284         unsigned int channel;
1285         int rtn;
1286
1287         /*
1288          * we really want to loop over the various channels, and do this on
1289          * a channel by channel basis.  we should also check to see if any
1290          * of the failed commands are on soft_reset devices, and if so, skip
1291          * the reset.
1292          */
1293
1294         for (channel = 0; channel <= shost->max_channel; channel++) {
1295                 chan_scmd = NULL;
1296                 list_for_each_entry(scmd, work_q, eh_entry) {
1297                         if (channel == scmd_channel(scmd)) {
1298                                 chan_scmd = scmd;
1299                                 break;
1300                                 /*
1301                                  * FIXME add back in some support for
1302                                  * soft_reset devices.
1303                                  */
1304                         }
1305                 }
1306
1307                 if (!chan_scmd)
1308                         continue;
1309                 SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Sending BRST chan:"
1310                                                   " %d\n", current->comm,
1311                                                   channel));
1312                 rtn = scsi_try_bus_reset(chan_scmd);
1313                 if (rtn == SUCCESS || rtn == FAST_IO_FAIL) {
1314                         list_for_each_entry_safe(scmd, next, work_q, eh_entry) {
1315                                 if (channel == scmd_channel(scmd)) {
1316                                         if (rtn == FAST_IO_FAIL)
1317                                                 scsi_eh_finish_cmd(scmd,
1318                                                                    done_q);
1319                                         else
1320                                                 list_move_tail(&scmd->eh_entry,
1321                                                                &check_list);
1322                                 }
1323                         }
1324                 } else {
1325                         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: BRST"
1326                                                           " failed chan: %d\n",
1327                                                           current->comm,
1328                                                           channel));
1329                 }
1330         }
1331         return scsi_eh_test_devices(&check_list, work_q, done_q, 0);
1332 }
1333
1334 /**
1335  * scsi_eh_host_reset - send a host reset
1336  * @work_q:     list_head for processed commands.
1337  * @done_q:     list_head for processed commands.
1338  */
1339 static int scsi_eh_host_reset(struct list_head *work_q,
1340                               struct list_head *done_q)
1341 {
1342         struct scsi_cmnd *scmd, *next;
1343         LIST_HEAD(check_list);
1344         int rtn;
1345
1346         if (!list_empty(work_q)) {
1347                 scmd = list_entry(work_q->next,
1348                                   struct scsi_cmnd, eh_entry);
1349
1350                 SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Sending HRST\n"
1351                                                   , current->comm));
1352
1353                 rtn = scsi_try_host_reset(scmd);
1354                 if (rtn == SUCCESS) {
1355                         list_splice_init(work_q, &check_list);
1356                 } else if (rtn == FAST_IO_FAIL) {
1357                         list_for_each_entry_safe(scmd, next, work_q, eh_entry) {
1358                                         scsi_eh_finish_cmd(scmd, done_q);
1359                         }
1360                 } else {
1361                         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: HRST"
1362                                                           " failed\n",
1363                                                           current->comm));
1364                 }
1365         }
1366         return scsi_eh_test_devices(&check_list, work_q, done_q, 1);
1367 }
1368
1369 /**
1370  * scsi_eh_offline_sdevs - offline scsi devices that fail to recover
1371  * @work_q:     list_head for processed commands.
1372  * @done_q:     list_head for processed commands.
1373  */
1374 static void scsi_eh_offline_sdevs(struct list_head *work_q,
1375                                   struct list_head *done_q)
1376 {
1377         struct scsi_cmnd *scmd, *next;
1378
1379         list_for_each_entry_safe(scmd, next, work_q, eh_entry) {
1380                 sdev_printk(KERN_INFO, scmd->device, "Device offlined - "
1381                             "not ready after error recovery\n");
1382                 scsi_device_set_state(scmd->device, SDEV_OFFLINE);
1383                 if (scmd->eh_eflags & SCSI_EH_CANCEL_CMD) {
1384                         /*
1385                          * FIXME: Handle lost cmds.
1386                          */
1387                 }
1388                 scsi_eh_finish_cmd(scmd, done_q);
1389         }
1390         return;
1391 }
1392
1393 /**
1394  * scsi_noretry_cmd - determinte if command should be failed fast
1395  * @scmd:       SCSI cmd to examine.
1396  */
1397 int scsi_noretry_cmd(struct scsi_cmnd *scmd)
1398 {
1399         switch (host_byte(scmd->result)) {
1400         case DID_OK:
1401                 break;
1402         case DID_BUS_BUSY:
1403                 return (scmd->request->cmd_flags & REQ_FAILFAST_TRANSPORT);
1404         case DID_PARITY:
1405                 return (scmd->request->cmd_flags & REQ_FAILFAST_DEV);
1406         case DID_ERROR:
1407                 if (msg_byte(scmd->result) == COMMAND_COMPLETE &&
1408                     status_byte(scmd->result) == RESERVATION_CONFLICT)
1409                         return 0;
1410                 /* fall through */
1411         case DID_SOFT_ERROR:
1412                 return (scmd->request->cmd_flags & REQ_FAILFAST_DRIVER);
1413         }
1414
1415         switch (status_byte(scmd->result)) {
1416         case CHECK_CONDITION:
1417                 /*
1418                  * assume caller has checked sense and determinted
1419                  * the check condition was retryable.
1420                  */
1421                 if (scmd->request->cmd_flags & REQ_FAILFAST_DEV ||
1422                     scmd->request->cmd_type == REQ_TYPE_BLOCK_PC)
1423                         return 1;
1424         }
1425
1426         return 0;
1427 }
1428
1429 /**
1430  * scsi_decide_disposition - Disposition a cmd on return from LLD.
1431  * @scmd:       SCSI cmd to examine.
1432  *
1433  * Notes:
1434  *    This is *only* called when we are examining the status after sending
1435  *    out the actual data command.  any commands that are queued for error
1436  *    recovery (e.g. test_unit_ready) do *not* come through here.
1437  *
1438  *    When this routine returns failed, it means the error handler thread
1439  *    is woken.  In cases where the error code indicates an error that
1440  *    doesn't require the error handler read (i.e. we don't need to
1441  *    abort/reset), this function should return SUCCESS.
1442  */
1443 int scsi_decide_disposition(struct scsi_cmnd *scmd)
1444 {
1445         int rtn;
1446
1447         /*
1448          * if the device is offline, then we clearly just pass the result back
1449          * up to the top level.
1450          */
1451         if (!scsi_device_online(scmd->device)) {
1452                 SCSI_LOG_ERROR_RECOVERY(5, printk("%s: device offline - report"
1453                                                   " as SUCCESS\n",
1454                                                   __func__));
1455                 return SUCCESS;
1456         }
1457
1458         /*
1459          * first check the host byte, to see if there is anything in there
1460          * that would indicate what we need to do.
1461          */
1462         switch (host_byte(scmd->result)) {
1463         case DID_PASSTHROUGH:
1464                 /*
1465                  * no matter what, pass this through to the upper layer.
1466                  * nuke this special code so that it looks like we are saying
1467                  * did_ok.
1468                  */
1469                 scmd->result &= 0xff00ffff;
1470                 return SUCCESS;
1471         case DID_OK:
1472                 /*
1473                  * looks good.  drop through, and check the next byte.
1474                  */
1475                 break;
1476         case DID_NO_CONNECT:
1477         case DID_BAD_TARGET:
1478         case DID_ABORT:
1479                 /*
1480                  * note - this means that we just report the status back
1481                  * to the top level driver, not that we actually think
1482                  * that it indicates SUCCESS.
1483                  */
1484                 return SUCCESS;
1485                 /*
1486                  * when the low level driver returns did_soft_error,
1487                  * it is responsible for keeping an internal retry counter
1488                  * in order to avoid endless loops (db)
1489                  *
1490                  * actually this is a bug in this function here.  we should
1491                  * be mindful of the maximum number of retries specified
1492                  * and not get stuck in a loop.
1493                  */
1494         case DID_SOFT_ERROR:
1495                 goto maybe_retry;
1496         case DID_IMM_RETRY:
1497                 return NEEDS_RETRY;
1498
1499         case DID_REQUEUE:
1500                 return ADD_TO_MLQUEUE;
1501         case DID_TRANSPORT_DISRUPTED:
1502                 /*
1503                  * LLD/transport was disrupted during processing of the IO.
1504                  * The transport class is now blocked/blocking,
1505                  * and the transport will decide what to do with the IO
1506                  * based on its timers and recovery capablilities if
1507                  * there are enough retries.
1508                  */
1509                 goto maybe_retry;
1510         case DID_TRANSPORT_FAILFAST:
1511                 /*
1512                  * The transport decided to failfast the IO (most likely
1513                  * the fast io fail tmo fired), so send IO directly upwards.
1514                  */
1515                 return SUCCESS;
1516         case DID_ERROR:
1517                 if (msg_byte(scmd->result) == COMMAND_COMPLETE &&
1518                     status_byte(scmd->result) == RESERVATION_CONFLICT)
1519                         /*
1520                          * execute reservation conflict processing code
1521                          * lower down
1522                          */
1523                         break;
1524                 /* fallthrough */
1525         case DID_BUS_BUSY:
1526         case DID_PARITY:
1527                 goto maybe_retry;
1528         case DID_TIME_OUT:
1529                 /*
1530                  * when we scan the bus, we get timeout messages for
1531                  * these commands if there is no device available.
1532                  * other hosts report did_no_connect for the same thing.
1533                  */
1534                 if ((scmd->cmnd[0] == TEST_UNIT_READY ||
1535                      scmd->cmnd[0] == INQUIRY)) {
1536                         return SUCCESS;
1537                 } else {
1538                         return FAILED;
1539                 }
1540         case DID_RESET:
1541                 return SUCCESS;
1542         default:
1543                 return FAILED;
1544         }
1545
1546         /*
1547          * next, check the message byte.
1548          */
1549         if (msg_byte(scmd->result) != COMMAND_COMPLETE)
1550                 return FAILED;
1551
1552         /*
1553          * check the status byte to see if this indicates anything special.
1554          */
1555         switch (status_byte(scmd->result)) {
1556         case QUEUE_FULL:
1557                 scsi_handle_queue_full(scmd->device);
1558                 /*
1559                  * the case of trying to send too many commands to a
1560                  * tagged queueing device.
1561                  */
1562         case BUSY:
1563                 /*
1564                  * device can't talk to us at the moment.  Should only
1565                  * occur (SAM-3) when the task queue is empty, so will cause
1566                  * the empty queue handling to trigger a stall in the
1567                  * device.
1568                  */
1569                 return ADD_TO_MLQUEUE;
1570         case GOOD:
1571                 scsi_handle_queue_ramp_up(scmd->device);
1572         case COMMAND_TERMINATED:
1573                 return SUCCESS;
1574         case TASK_ABORTED:
1575                 goto maybe_retry;
1576         case CHECK_CONDITION:
1577                 rtn = scsi_check_sense(scmd);
1578                 if (rtn == NEEDS_RETRY)
1579                         goto maybe_retry;
1580                 else if (rtn == TARGET_ERROR) {
1581                         /*
1582                          * Need to modify host byte to signal a
1583                          * permanent target failure
1584                          */
1585                         set_host_byte(scmd, DID_TARGET_FAILURE);
1586                         rtn = SUCCESS;
1587                 }
1588                 /* if rtn == FAILED, we have no sense information;
1589                  * returning FAILED will wake the error handler thread
1590                  * to collect the sense and redo the decide
1591                  * disposition */
1592                 return rtn;
1593         case CONDITION_GOOD:
1594         case INTERMEDIATE_GOOD:
1595         case INTERMEDIATE_C_GOOD:
1596         case ACA_ACTIVE:
1597                 /*
1598                  * who knows?  FIXME(eric)
1599                  */
1600                 return SUCCESS;
1601
1602         case RESERVATION_CONFLICT:
1603                 sdev_printk(KERN_INFO, scmd->device,
1604                             "reservation conflict\n");
1605                 set_host_byte(scmd, DID_NEXUS_FAILURE);
1606                 return SUCCESS; /* causes immediate i/o error */
1607         default:
1608                 return FAILED;
1609         }
1610         return FAILED;
1611
1612       maybe_retry:
1613
1614         /* we requeue for retry because the error was retryable, and
1615          * the request was not marked fast fail.  Note that above,
1616          * even if the request is marked fast fail, we still requeue
1617          * for queue congestion conditions (QUEUE_FULL or BUSY) */
1618         if ((++scmd->retries) <= scmd->allowed
1619             && !scsi_noretry_cmd(scmd)) {
1620                 return NEEDS_RETRY;
1621         } else {
1622                 /*
1623                  * no more retries - report this one back to upper level.
1624                  */
1625                 return SUCCESS;
1626         }
1627 }
1628
1629 static void eh_lock_door_done(struct request *req, int uptodate)
1630 {
1631         __blk_put_request(req->q, req);
1632 }
1633
1634 /**
1635  * scsi_eh_lock_door - Prevent medium removal for the specified device
1636  * @sdev:       SCSI device to prevent medium removal
1637  *
1638  * Locking:
1639  *      We must be called from process context.
1640  *
1641  * Notes:
1642  *      We queue up an asynchronous "ALLOW MEDIUM REMOVAL" request on the
1643  *      head of the devices request queue, and continue.
1644  */
1645 static void scsi_eh_lock_door(struct scsi_device *sdev)
1646 {
1647         struct request *req;
1648
1649         /*
1650          * blk_get_request with GFP_KERNEL (__GFP_WAIT) sleeps until a
1651          * request becomes available
1652          */
1653         req = blk_get_request(sdev->request_queue, READ, GFP_KERNEL);
1654
1655         req->cmd[0] = ALLOW_MEDIUM_REMOVAL;
1656         req->cmd[1] = 0;
1657         req->cmd[2] = 0;
1658         req->cmd[3] = 0;
1659         req->cmd[4] = SCSI_REMOVAL_PREVENT;
1660         req->cmd[5] = 0;
1661
1662         req->cmd_len = COMMAND_SIZE(req->cmd[0]);
1663
1664         req->cmd_type = REQ_TYPE_BLOCK_PC;
1665         req->cmd_flags |= REQ_QUIET;
1666         req->timeout = 10 * HZ;
1667         req->retries = 5;
1668
1669         blk_execute_rq_nowait(req->q, NULL, req, 1, eh_lock_door_done);
1670 }
1671
1672 /**
1673  * scsi_restart_operations - restart io operations to the specified host.
1674  * @shost:      Host we are restarting.
1675  *
1676  * Notes:
1677  *    When we entered the error handler, we blocked all further i/o to
1678  *    this device.  we need to 'reverse' this process.
1679  */
1680 static void scsi_restart_operations(struct Scsi_Host *shost)
1681 {
1682         struct scsi_device *sdev;
1683         unsigned long flags;
1684
1685         /*
1686          * If the door was locked, we need to insert a door lock request
1687          * onto the head of the SCSI request queue for the device.  There
1688          * is no point trying to lock the door of an off-line device.
1689          */
1690         shost_for_each_device(sdev, shost) {
1691                 if (scsi_device_online(sdev) && sdev->locked)
1692                         scsi_eh_lock_door(sdev);
1693         }
1694
1695         /*
1696          * next free up anything directly waiting upon the host.  this
1697          * will be requests for character device operations, and also for
1698          * ioctls to queued block devices.
1699          */
1700         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: waking up host to restart\n",
1701                                           __func__));
1702
1703         spin_lock_irqsave(shost->host_lock, flags);
1704         if (scsi_host_set_state(shost, SHOST_RUNNING))
1705                 if (scsi_host_set_state(shost, SHOST_CANCEL))
1706                         BUG_ON(scsi_host_set_state(shost, SHOST_DEL));
1707         spin_unlock_irqrestore(shost->host_lock, flags);
1708
1709         wake_up(&shost->host_wait);
1710
1711         /*
1712          * finally we need to re-initiate requests that may be pending.  we will
1713          * have had everything blocked while error handling is taking place, and
1714          * now that error recovery is done, we will need to ensure that these
1715          * requests are started.
1716          */
1717         scsi_run_host_queues(shost);
1718
1719         /*
1720          * if eh is active and host_eh_scheduled is pending we need to re-run
1721          * recovery.  we do this check after scsi_run_host_queues() to allow
1722          * everything pent up since the last eh run a chance to make forward
1723          * progress before we sync again.  Either we'll immediately re-run
1724          * recovery or scsi_device_unbusy() will wake us again when these
1725          * pending commands complete.
1726          */
1727         spin_lock_irqsave(shost->host_lock, flags);
1728         if (shost->host_eh_scheduled)
1729                 if (scsi_host_set_state(shost, SHOST_RECOVERY))
1730                         WARN_ON(scsi_host_set_state(shost, SHOST_CANCEL_RECOVERY));
1731         spin_unlock_irqrestore(shost->host_lock, flags);
1732 }
1733
1734 /**
1735  * scsi_eh_ready_devs - check device ready state and recover if not.
1736  * @shost:      host to be recovered.
1737  * @work_q:     &list_head for pending commands.
1738  * @done_q:     &list_head for processed commands.
1739  */
1740 void scsi_eh_ready_devs(struct Scsi_Host *shost,
1741                         struct list_head *work_q,
1742                         struct list_head *done_q)
1743 {
1744         if (!scsi_eh_stu(shost, work_q, done_q))
1745                 if (!scsi_eh_bus_device_reset(shost, work_q, done_q))
1746                         if (!scsi_eh_target_reset(shost, work_q, done_q))
1747                                 if (!scsi_eh_bus_reset(shost, work_q, done_q))
1748                                         if (!scsi_eh_host_reset(work_q, done_q))
1749                                                 scsi_eh_offline_sdevs(work_q,
1750                                                                       done_q);
1751 }
1752 EXPORT_SYMBOL_GPL(scsi_eh_ready_devs);
1753
1754 /**
1755  * scsi_eh_flush_done_q - finish processed commands or retry them.
1756  * @done_q:     list_head of processed commands.
1757  */
1758 void scsi_eh_flush_done_q(struct list_head *done_q)
1759 {
1760         struct scsi_cmnd *scmd, *next;
1761
1762         list_for_each_entry_safe(scmd, next, done_q, eh_entry) {
1763                 list_del_init(&scmd->eh_entry);
1764                 if (scsi_device_online(scmd->device) &&
1765                     !scsi_noretry_cmd(scmd) &&
1766                     (++scmd->retries <= scmd->allowed)) {
1767                         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: flush"
1768                                                           " retry cmd: %p\n",
1769                                                           current->comm,
1770                                                           scmd));
1771                                 scsi_queue_insert(scmd, SCSI_MLQUEUE_EH_RETRY);
1772                 } else {
1773                         /*
1774                          * If just we got sense for the device (called
1775                          * scsi_eh_get_sense), scmd->result is already
1776                          * set, do not set DRIVER_TIMEOUT.
1777                          */
1778                         if (!scmd->result)
1779                                 scmd->result |= (DRIVER_TIMEOUT << 24);
1780                         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: flush finish"
1781                                                         " cmd: %p\n",
1782                                                         current->comm, scmd));
1783                         scsi_finish_command(scmd);
1784                 }
1785         }
1786 }
1787 EXPORT_SYMBOL(scsi_eh_flush_done_q);
1788
1789 /**
1790  * scsi_unjam_host - Attempt to fix a host which has a cmd that failed.
1791  * @shost:      Host to unjam.
1792  *
1793  * Notes:
1794  *    When we come in here, we *know* that all commands on the bus have
1795  *    either completed, failed or timed out.  we also know that no further
1796  *    commands are being sent to the host, so things are relatively quiet
1797  *    and we have freedom to fiddle with things as we wish.
1798  *
1799  *    This is only the *default* implementation.  it is possible for
1800  *    individual drivers to supply their own version of this function, and
1801  *    if the maintainer wishes to do this, it is strongly suggested that
1802  *    this function be taken as a template and modified.  this function
1803  *    was designed to correctly handle problems for about 95% of the
1804  *    different cases out there, and it should always provide at least a
1805  *    reasonable amount of error recovery.
1806  *
1807  *    Any command marked 'failed' or 'timeout' must eventually have
1808  *    scsi_finish_cmd() called for it.  we do all of the retry stuff
1809  *    here, so when we restart the host after we return it should have an
1810  *    empty queue.
1811  */
1812 static void scsi_unjam_host(struct Scsi_Host *shost)
1813 {
1814         unsigned long flags;
1815         LIST_HEAD(eh_work_q);
1816         LIST_HEAD(eh_done_q);
1817
1818         spin_lock_irqsave(shost->host_lock, flags);
1819         list_splice_init(&shost->eh_cmd_q, &eh_work_q);
1820         spin_unlock_irqrestore(shost->host_lock, flags);
1821
1822         SCSI_LOG_ERROR_RECOVERY(1, scsi_eh_prt_fail_stats(shost, &eh_work_q));
1823
1824         if (!scsi_eh_get_sense(&eh_work_q, &eh_done_q))
1825                 if (!scsi_eh_abort_cmds(&eh_work_q, &eh_done_q))
1826                         scsi_eh_ready_devs(shost, &eh_work_q, &eh_done_q);
1827
1828         scsi_eh_flush_done_q(&eh_done_q);
1829 }
1830
1831 /**
1832  * scsi_error_handler - SCSI error handler thread
1833  * @data:       Host for which we are running.
1834  *
1835  * Notes:
1836  *    This is the main error handling loop.  This is run as a kernel thread
1837  *    for every SCSI host and handles all error handling activity.
1838  */
1839 int scsi_error_handler(void *data)
1840 {
1841         struct Scsi_Host *shost = data;
1842
1843         /*
1844          * We use TASK_INTERRUPTIBLE so that the thread is not
1845          * counted against the load average as a running process.
1846          * We never actually get interrupted because kthread_run
1847          * disables signal delivery for the created thread.
1848          */
1849         while (!kthread_should_stop()) {
1850                 set_current_state(TASK_INTERRUPTIBLE);
1851                 if ((shost->host_failed == 0 && shost->host_eh_scheduled == 0) ||
1852                     shost->host_failed != shost->host_busy) {
1853                         SCSI_LOG_ERROR_RECOVERY(1,
1854                                 printk("Error handler scsi_eh_%d sleeping\n",
1855                                         shost->host_no));
1856                         schedule();
1857                         continue;
1858                 }
1859
1860                 __set_current_state(TASK_RUNNING);
1861                 SCSI_LOG_ERROR_RECOVERY(1,
1862                         printk("Error handler scsi_eh_%d waking up\n",
1863                                 shost->host_no));
1864
1865                 /*
1866                  * We have a host that is failing for some reason.  Figure out
1867                  * what we need to do to get it up and online again (if we can).
1868                  * If we fail, we end up taking the thing offline.
1869                  */
1870                 if (!shost->eh_noresume && scsi_autopm_get_host(shost) != 0) {
1871                         SCSI_LOG_ERROR_RECOVERY(1,
1872                                 printk(KERN_ERR "Error handler scsi_eh_%d "
1873                                                 "unable to autoresume\n",
1874                                                 shost->host_no));
1875                         continue;
1876                 }
1877
1878                 if (shost->transportt->eh_strategy_handler)
1879                         shost->transportt->eh_strategy_handler(shost);
1880                 else
1881                         scsi_unjam_host(shost);
1882
1883                 /*
1884                  * Note - if the above fails completely, the action is to take
1885                  * individual devices offline and flush the queue of any
1886                  * outstanding requests that may have been pending.  When we
1887                  * restart, we restart any I/O to any other devices on the bus
1888                  * which are still online.
1889                  */
1890                 scsi_restart_operations(shost);
1891                 if (!shost->eh_noresume)
1892                         scsi_autopm_put_host(shost);
1893         }
1894         __set_current_state(TASK_RUNNING);
1895
1896         SCSI_LOG_ERROR_RECOVERY(1,
1897                 printk("Error handler scsi_eh_%d exiting\n", shost->host_no));
1898         shost->ehandler = NULL;
1899         return 0;
1900 }
1901
1902 /*
1903  * Function:    scsi_report_bus_reset()
1904  *
1905  * Purpose:     Utility function used by low-level drivers to report that
1906  *              they have observed a bus reset on the bus being handled.
1907  *
1908  * Arguments:   shost       - Host in question
1909  *              channel     - channel on which reset was observed.
1910  *
1911  * Returns:     Nothing
1912  *
1913  * Lock status: Host lock must be held.
1914  *
1915  * Notes:       This only needs to be called if the reset is one which
1916  *              originates from an unknown location.  Resets originated
1917  *              by the mid-level itself don't need to call this, but there
1918  *              should be no harm.
1919  *
1920  *              The main purpose of this is to make sure that a CHECK_CONDITION
1921  *              is properly treated.
1922  */
1923 void scsi_report_bus_reset(struct Scsi_Host *shost, int channel)
1924 {
1925         struct scsi_device *sdev;
1926
1927         __shost_for_each_device(sdev, shost) {
1928                 if (channel == sdev_channel(sdev))
1929                         __scsi_report_device_reset(sdev, NULL);
1930         }
1931 }
1932 EXPORT_SYMBOL(scsi_report_bus_reset);
1933
1934 /*
1935  * Function:    scsi_report_device_reset()
1936  *
1937  * Purpose:     Utility function used by low-level drivers to report that
1938  *              they have observed a device reset on the device being handled.
1939  *
1940  * Arguments:   shost       - Host in question
1941  *              channel     - channel on which reset was observed
1942  *              target      - target on which reset was observed
1943  *
1944  * Returns:     Nothing
1945  *
1946  * Lock status: Host lock must be held
1947  *
1948  * Notes:       This only needs to be called if the reset is one which
1949  *              originates from an unknown location.  Resets originated
1950  *              by the mid-level itself don't need to call this, but there
1951  *              should be no harm.
1952  *
1953  *              The main purpose of this is to make sure that a CHECK_CONDITION
1954  *              is properly treated.
1955  */
1956 void scsi_report_device_reset(struct Scsi_Host *shost, int channel, int target)
1957 {
1958         struct scsi_device *sdev;
1959
1960         __shost_for_each_device(sdev, shost) {
1961                 if (channel == sdev_channel(sdev) &&
1962                     target == sdev_id(sdev))
1963                         __scsi_report_device_reset(sdev, NULL);
1964         }
1965 }
1966 EXPORT_SYMBOL(scsi_report_device_reset);
1967
1968 static void
1969 scsi_reset_provider_done_command(struct scsi_cmnd *scmd)
1970 {
1971 }
1972
1973 /*
1974  * Function:    scsi_reset_provider
1975  *
1976  * Purpose:     Send requested reset to a bus or device at any phase.
1977  *
1978  * Arguments:   device  - device to send reset to
1979  *              flag - reset type (see scsi.h)
1980  *
1981  * Returns:     SUCCESS/FAILURE.
1982  *
1983  * Notes:       This is used by the SCSI Generic driver to provide
1984  *              Bus/Device reset capability.
1985  */
1986 int
1987 scsi_reset_provider(struct scsi_device *dev, int flag)
1988 {
1989         struct scsi_cmnd *scmd;
1990         struct Scsi_Host *shost = dev->host;
1991         struct request req;
1992         unsigned long flags;
1993         int rtn;
1994
1995         if (scsi_autopm_get_host(shost) < 0)
1996                 return FAILED;
1997
1998         scmd = scsi_get_command(dev, GFP_KERNEL);
1999         blk_rq_init(NULL, &req);
2000         scmd->request = &req;
2001
2002         scmd->cmnd = req.cmd;
2003
2004         scmd->scsi_done         = scsi_reset_provider_done_command;
2005         memset(&scmd->sdb, 0, sizeof(scmd->sdb));
2006
2007         scmd->cmd_len                   = 0;
2008
2009         scmd->sc_data_direction         = DMA_BIDIRECTIONAL;
2010
2011         spin_lock_irqsave(shost->host_lock, flags);
2012         shost->tmf_in_progress = 1;
2013         spin_unlock_irqrestore(shost->host_lock, flags);
2014
2015         switch (flag) {
2016         case SCSI_TRY_RESET_DEVICE:
2017                 rtn = scsi_try_bus_device_reset(scmd);
2018                 if (rtn == SUCCESS)
2019                         break;
2020                 /* FALLTHROUGH */
2021         case SCSI_TRY_RESET_TARGET:
2022                 rtn = scsi_try_target_reset(scmd);
2023                 if (rtn == SUCCESS)
2024                         break;
2025                 /* FALLTHROUGH */
2026         case SCSI_TRY_RESET_BUS:
2027                 rtn = scsi_try_bus_reset(scmd);
2028                 if (rtn == SUCCESS)
2029                         break;
2030                 /* FALLTHROUGH */
2031         case SCSI_TRY_RESET_HOST:
2032                 rtn = scsi_try_host_reset(scmd);
2033                 break;
2034         default:
2035                 rtn = FAILED;
2036         }
2037
2038         spin_lock_irqsave(shost->host_lock, flags);
2039         shost->tmf_in_progress = 0;
2040         spin_unlock_irqrestore(shost->host_lock, flags);
2041
2042         /*
2043          * be sure to wake up anyone who was sleeping or had their queue
2044          * suspended while we performed the TMF.
2045          */
2046         SCSI_LOG_ERROR_RECOVERY(3,
2047                 printk("%s: waking up host to restart after TMF\n",
2048                 __func__));
2049
2050         wake_up(&shost->host_wait);
2051
2052         scsi_run_host_queues(shost);
2053
2054         scsi_next_command(scmd);
2055         scsi_autopm_put_host(shost);
2056         return rtn;
2057 }
2058 EXPORT_SYMBOL(scsi_reset_provider);
2059
2060 /**
2061  * scsi_normalize_sense - normalize main elements from either fixed or
2062  *                      descriptor sense data format into a common format.
2063  *
2064  * @sense_buffer:       byte array containing sense data returned by device
2065  * @sb_len:             number of valid bytes in sense_buffer
2066  * @sshdr:              pointer to instance of structure that common
2067  *                      elements are written to.
2068  *
2069  * Notes:
2070  *      The "main elements" from sense data are: response_code, sense_key,
2071  *      asc, ascq and additional_length (only for descriptor format).
2072  *
2073  *      Typically this function can be called after a device has
2074  *      responded to a SCSI command with the CHECK_CONDITION status.
2075  *
2076  * Return value:
2077  *      1 if valid sense data information found, else 0;
2078  */
2079 int scsi_normalize_sense(const u8 *sense_buffer, int sb_len,
2080                          struct scsi_sense_hdr *sshdr)
2081 {
2082         if (!sense_buffer || !sb_len)
2083                 return 0;
2084
2085         memset(sshdr, 0, sizeof(struct scsi_sense_hdr));
2086
2087         sshdr->response_code = (sense_buffer[0] & 0x7f);
2088
2089         if (!scsi_sense_valid(sshdr))
2090                 return 0;
2091
2092         if (sshdr->response_code >= 0x72) {
2093                 /*
2094                  * descriptor format
2095                  */
2096                 if (sb_len > 1)
2097                         sshdr->sense_key = (sense_buffer[1] & 0xf);
2098                 if (sb_len > 2)
2099                         sshdr->asc = sense_buffer[2];
2100                 if (sb_len > 3)
2101                         sshdr->ascq = sense_buffer[3];
2102                 if (sb_len > 7)
2103                         sshdr->additional_length = sense_buffer[7];
2104         } else {
2105                 /*
2106                  * fixed format
2107                  */
2108                 if (sb_len > 2)
2109                         sshdr->sense_key = (sense_buffer[2] & 0xf);
2110                 if (sb_len > 7) {
2111                         sb_len = (sb_len < (sense_buffer[7] + 8)) ?
2112                                          sb_len : (sense_buffer[7] + 8);
2113                         if (sb_len > 12)
2114                                 sshdr->asc = sense_buffer[12];
2115                         if (sb_len > 13)
2116                                 sshdr->ascq = sense_buffer[13];
2117                 }
2118         }
2119
2120         return 1;
2121 }
2122 EXPORT_SYMBOL(scsi_normalize_sense);
2123
2124 int scsi_command_normalize_sense(struct scsi_cmnd *cmd,
2125                                  struct scsi_sense_hdr *sshdr)
2126 {
2127         return scsi_normalize_sense(cmd->sense_buffer,
2128                         SCSI_SENSE_BUFFERSIZE, sshdr);
2129 }
2130 EXPORT_SYMBOL(scsi_command_normalize_sense);
2131
2132 /**
2133  * scsi_sense_desc_find - search for a given descriptor type in descriptor sense data format.
2134  * @sense_buffer:       byte array of descriptor format sense data
2135  * @sb_len:             number of valid bytes in sense_buffer
2136  * @desc_type:          value of descriptor type to find
2137  *                      (e.g. 0 -> information)
2138  *
2139  * Notes:
2140  *      only valid when sense data is in descriptor format
2141  *
2142  * Return value:
2143  *      pointer to start of (first) descriptor if found else NULL
2144  */
2145 const u8 * scsi_sense_desc_find(const u8 * sense_buffer, int sb_len,
2146                                 int desc_type)
2147 {
2148         int add_sen_len, add_len, desc_len, k;
2149         const u8 * descp;
2150
2151         if ((sb_len < 8) || (0 == (add_sen_len = sense_buffer[7])))
2152                 return NULL;
2153         if ((sense_buffer[0] < 0x72) || (sense_buffer[0] > 0x73))
2154                 return NULL;
2155         add_sen_len = (add_sen_len < (sb_len - 8)) ?
2156                         add_sen_len : (sb_len - 8);
2157         descp = &sense_buffer[8];
2158         for (desc_len = 0, k = 0; k < add_sen_len; k += desc_len) {
2159                 descp += desc_len;
2160                 add_len = (k < (add_sen_len - 1)) ? descp[1]: -1;
2161                 desc_len = add_len + 2;
2162                 if (descp[0] == desc_type)
2163                         return descp;
2164                 if (add_len < 0) // short descriptor ??
2165                         break;
2166         }
2167         return NULL;
2168 }
2169 EXPORT_SYMBOL(scsi_sense_desc_find);
2170
2171 /**
2172  * scsi_get_sense_info_fld - get information field from sense data (either fixed or descriptor format)
2173  * @sense_buffer:       byte array of sense data
2174  * @sb_len:             number of valid bytes in sense_buffer
2175  * @info_out:           pointer to 64 integer where 8 or 4 byte information
2176  *                      field will be placed if found.
2177  *
2178  * Return value:
2179  *      1 if information field found, 0 if not found.
2180  */
2181 int scsi_get_sense_info_fld(const u8 * sense_buffer, int sb_len,
2182                             u64 * info_out)
2183 {
2184         int j;
2185         const u8 * ucp;
2186         u64 ull;
2187
2188         if (sb_len < 7)
2189                 return 0;
2190         switch (sense_buffer[0] & 0x7f) {
2191         case 0x70:
2192         case 0x71:
2193                 if (sense_buffer[0] & 0x80) {
2194                         *info_out = (sense_buffer[3] << 24) +
2195                                     (sense_buffer[4] << 16) +
2196                                     (sense_buffer[5] << 8) + sense_buffer[6];
2197                         return 1;
2198                 } else
2199                         return 0;
2200         case 0x72:
2201         case 0x73:
2202                 ucp = scsi_sense_desc_find(sense_buffer, sb_len,
2203                                            0 /* info desc */);
2204                 if (ucp && (0xa == ucp[1])) {
2205                         ull = 0;
2206                         for (j = 0; j < 8; ++j) {
2207                                 if (j > 0)
2208                                         ull <<= 8;
2209                                 ull |= ucp[4 + j];
2210                         }
2211                         *info_out = ull;
2212                         return 1;
2213                 } else
2214                         return 0;
2215         default:
2216                 return 0;
2217         }
2218 }
2219 EXPORT_SYMBOL(scsi_get_sense_info_fld);
2220
2221 /**
2222  * scsi_build_sense_buffer - build sense data in a buffer
2223  * @desc:       Sense format (non zero == descriptor format,
2224  *              0 == fixed format)
2225  * @buf:        Where to build sense data
2226  * @key:        Sense key
2227  * @asc:        Additional sense code
2228  * @ascq:       Additional sense code qualifier
2229  *
2230  **/
2231 void scsi_build_sense_buffer(int desc, u8 *buf, u8 key, u8 asc, u8 ascq)
2232 {
2233         if (desc) {
2234                 buf[0] = 0x72;  /* descriptor, current */
2235                 buf[1] = key;
2236                 buf[2] = asc;
2237                 buf[3] = ascq;
2238                 buf[7] = 0;
2239         } else {
2240                 buf[0] = 0x70;  /* fixed, current */
2241                 buf[2] = key;
2242                 buf[7] = 0xa;
2243                 buf[12] = asc;
2244                 buf[13] = ascq;
2245         }
2246 }
2247 EXPORT_SYMBOL(scsi_build_sense_buffer);