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