drivers/scsi/scsi_error.c

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