[SCSI] move the mid-layer printk's over to shost/starget/sdev_printk
[linux-2.6.git] / 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_host.h>
33 #include <scsi/scsi_ioctl.h>
34 #include <scsi/scsi_request.h>
35
36 #include "scsi_priv.h"
37 #include "scsi_logging.h"
38
39 #define SENSE_TIMEOUT           (10*HZ)
40 #define START_UNIT_TIMEOUT      (30*HZ)
41
42 /*
43  * These should *probably* be handled by the host itself.
44  * Since it is allowed to sleep, it probably should.
45  */
46 #define BUS_RESET_SETTLE_TIME   (10)
47 #define HOST_RESET_SETTLE_TIME  (10)
48
49 /* called with shost->host_lock held */
50 void scsi_eh_wakeup(struct Scsi_Host *shost)
51 {
52         if (shost->host_busy == shost->host_failed) {
53                 wake_up_process(shost->ehandler);
54                 SCSI_LOG_ERROR_RECOVERY(5,
55                                 printk("Waking error handler thread\n"));
56         }
57 }
58
59 /**
60  * scsi_eh_scmd_add - add scsi cmd to error handling.
61  * @scmd:       scmd to run eh on.
62  * @eh_flag:    optional SCSI_EH flag.
63  *
64  * Return value:
65  *      0 on failure.
66  **/
67 int scsi_eh_scmd_add(struct scsi_cmnd *scmd, int eh_flag)
68 {
69         struct Scsi_Host *shost = scmd->device->host;
70         unsigned long flags;
71         int ret = 0;
72
73         if (!shost->ehandler)
74                 return 0;
75
76         spin_lock_irqsave(shost->host_lock, flags);
77         if (scsi_host_set_state(shost, SHOST_RECOVERY))
78                 if (scsi_host_set_state(shost, SHOST_CANCEL_RECOVERY))
79                         goto out_unlock;
80
81         ret = 1;
82         scmd->eh_eflags |= eh_flag;
83         list_add_tail(&scmd->eh_entry, &shost->eh_cmd_q);
84         shost->host_failed++;
85         scsi_eh_wakeup(shost);
86  out_unlock:
87         spin_unlock_irqrestore(shost->host_lock, flags);
88         return ret;
89 }
90
91 /**
92  * scsi_add_timer - Start timeout timer for a single scsi command.
93  * @scmd:       scsi command that is about to start running.
94  * @timeout:    amount of time to allow this command to run.
95  * @complete:   timeout function to call if timer isn't canceled.
96  *
97  * Notes:
98  *    This should be turned into an inline function.  Each scsi command
99  *    has its own timer, and as it is added to the queue, we set up the
100  *    timer.  When the command completes, we cancel the timer.
101  **/
102 void scsi_add_timer(struct scsi_cmnd *scmd, int timeout,
103                     void (*complete)(struct scsi_cmnd *))
104 {
105
106         /*
107          * If the clock was already running for this command, then
108          * first delete the timer.  The timer handling code gets rather
109          * confused if we don't do this.
110          */
111         if (scmd->eh_timeout.function)
112                 del_timer(&scmd->eh_timeout);
113
114         scmd->eh_timeout.data = (unsigned long)scmd;
115         scmd->eh_timeout.expires = jiffies + timeout;
116         scmd->eh_timeout.function = (void (*)(unsigned long)) complete;
117
118         SCSI_LOG_ERROR_RECOVERY(5, printk("%s: scmd: %p, time:"
119                                           " %d, (%p)\n", __FUNCTION__,
120                                           scmd, timeout, complete));
121
122         add_timer(&scmd->eh_timeout);
123 }
124
125 /**
126  * scsi_delete_timer - Delete/cancel timer for a given function.
127  * @scmd:       Cmd that we are canceling timer for
128  *
129  * Notes:
130  *     This should be turned into an inline function.
131  *
132  * Return value:
133  *     1 if we were able to detach the timer.  0 if we blew it, and the
134  *     timer function has already started to run.
135  **/
136 int scsi_delete_timer(struct scsi_cmnd *scmd)
137 {
138         int rtn;
139
140         rtn = del_timer(&scmd->eh_timeout);
141
142         SCSI_LOG_ERROR_RECOVERY(5, printk("%s: scmd: %p,"
143                                          " rtn: %d\n", __FUNCTION__,
144                                          scmd, rtn));
145
146         scmd->eh_timeout.data = (unsigned long)NULL;
147         scmd->eh_timeout.function = NULL;
148
149         return rtn;
150 }
151
152 /**
153  * scsi_times_out - Timeout function for normal scsi commands.
154  * @scmd:       Cmd that is timing out.
155  *
156  * Notes:
157  *     We do not need to lock this.  There is the potential for a race
158  *     only in that the normal completion handling might run, but if the
159  *     normal completion function determines that the timer has already
160  *     fired, then it mustn't do anything.
161  **/
162 void scsi_times_out(struct scsi_cmnd *scmd)
163 {
164         scsi_log_completion(scmd, TIMEOUT_ERROR);
165
166         if (scmd->device->host->hostt->eh_timed_out)
167                 switch (scmd->device->host->hostt->eh_timed_out(scmd)) {
168                 case EH_HANDLED:
169                         __scsi_done(scmd);
170                         return;
171                 case EH_RESET_TIMER:
172                         /* This allows a single retry even of a command
173                          * with allowed == 0 */
174                         if (scmd->retries++ > scmd->allowed)
175                                 break;
176                         scsi_add_timer(scmd, scmd->timeout_per_command,
177                                        scsi_times_out);
178                         return;
179                 case EH_NOT_HANDLED:
180                         break;
181                 }
182
183         if (unlikely(!scsi_eh_scmd_add(scmd, SCSI_EH_CANCEL_CMD))) {
184                 scmd->result |= DID_TIME_OUT << 16;
185                 __scsi_done(scmd);
186         }
187 }
188
189 /**
190  * scsi_block_when_processing_errors - Prevent cmds from being queued.
191  * @sdev:       Device on which we are performing recovery.
192  *
193  * Description:
194  *     We block until the host is out of error recovery, and then check to
195  *     see whether the host or the device is offline.
196  *
197  * Return value:
198  *     0 when dev was taken offline by error recovery. 1 OK to proceed.
199  **/
200 int scsi_block_when_processing_errors(struct scsi_device *sdev)
201 {
202         int online;
203
204         wait_event(sdev->host->host_wait, !scsi_host_in_recovery(sdev->host));
205
206         online = scsi_device_online(sdev);
207
208         SCSI_LOG_ERROR_RECOVERY(5, printk("%s: rtn: %d\n", __FUNCTION__,
209                                           online));
210
211         return online;
212 }
213 EXPORT_SYMBOL(scsi_block_when_processing_errors);
214
215 #ifdef CONFIG_SCSI_LOGGING
216 /**
217  * scsi_eh_prt_fail_stats - Log info on failures.
218  * @shost:      scsi host being recovered.
219  * @work_q:     Queue of scsi cmds to process.
220  **/
221 static inline void scsi_eh_prt_fail_stats(struct Scsi_Host *shost,
222                                           struct list_head *work_q)
223 {
224         struct scsi_cmnd *scmd;
225         struct scsi_device *sdev;
226         int total_failures = 0;
227         int cmd_failed = 0;
228         int cmd_cancel = 0;
229         int devices_failed = 0;
230
231         shost_for_each_device(sdev, shost) {
232                 list_for_each_entry(scmd, work_q, eh_entry) {
233                         if (scmd->device == sdev) {
234                                 ++total_failures;
235                                 if (scmd->eh_eflags & SCSI_EH_CANCEL_CMD)
236                                         ++cmd_cancel;
237                                 else 
238                                         ++cmd_failed;
239                         }
240                 }
241
242                 if (cmd_cancel || cmd_failed) {
243                         SCSI_LOG_ERROR_RECOVERY(3,
244                                 sdev_printk(KERN_INFO, sdev,
245                                             "%s: cmds failed: %d, cancel: %d\n",
246                                             __FUNCTION__, cmd_failed,
247                                             cmd_cancel));
248                         cmd_cancel = 0;
249                         cmd_failed = 0;
250                         ++devices_failed;
251                 }
252         }
253
254         SCSI_LOG_ERROR_RECOVERY(2, printk("Total of %d commands on %d"
255                                           " devices require eh work\n",
256                                   total_failures, devices_failed));
257 }
258 #endif
259
260 /**
261  * scsi_check_sense - Examine scsi cmd sense
262  * @scmd:       Cmd to have sense checked.
263  *
264  * Return value:
265  *      SUCCESS or FAILED or NEEDS_RETRY
266  *
267  * Notes:
268  *      When a deferred error is detected the current command has
269  *      not been executed and needs retrying.
270  **/
271 static int scsi_check_sense(struct scsi_cmnd *scmd)
272 {
273         struct scsi_sense_hdr sshdr;
274
275         if (! scsi_command_normalize_sense(scmd, &sshdr))
276                 return FAILED;  /* no valid sense data */
277
278         if (scsi_sense_is_deferred(&sshdr))
279                 return NEEDS_RETRY;
280
281         /*
282          * Previous logic looked for FILEMARK, EOM or ILI which are
283          * mainly associated with tapes and returned SUCCESS.
284          */
285         if (sshdr.response_code == 0x70) {
286                 /* fixed format */
287                 if (scmd->sense_buffer[2] & 0xe0)
288                         return SUCCESS;
289         } else {
290                 /*
291                  * descriptor format: look for "stream commands sense data
292                  * descriptor" (see SSC-3). Assume single sense data
293                  * descriptor. Ignore ILI from SBC-2 READ LONG and WRITE LONG.
294                  */
295                 if ((sshdr.additional_length > 3) &&
296                     (scmd->sense_buffer[8] == 0x4) &&
297                     (scmd->sense_buffer[11] & 0xe0))
298                         return SUCCESS;
299         }
300
301         switch (sshdr.sense_key) {
302         case NO_SENSE:
303                 return SUCCESS;
304         case RECOVERED_ERROR:
305                 return /* soft_error */ SUCCESS;
306
307         case ABORTED_COMMAND:
308                 return NEEDS_RETRY;
309         case NOT_READY:
310         case UNIT_ATTENTION:
311                 /*
312                  * if we are expecting a cc/ua because of a bus reset that we
313                  * performed, treat this just as a retry.  otherwise this is
314                  * information that we should pass up to the upper-level driver
315                  * so that we can deal with it there.
316                  */
317                 if (scmd->device->expecting_cc_ua) {
318                         scmd->device->expecting_cc_ua = 0;
319                         return NEEDS_RETRY;
320                 }
321                 /*
322                  * if the device is in the process of becoming ready, we 
323                  * should retry.
324                  */
325                 if ((sshdr.asc == 0x04) && (sshdr.ascq == 0x01))
326                         return NEEDS_RETRY;
327                 /*
328                  * if the device is not started, we need to wake
329                  * the error handler to start the motor
330                  */
331                 if (scmd->device->allow_restart &&
332                     (sshdr.asc == 0x04) && (sshdr.ascq == 0x02))
333                         return FAILED;
334                 return SUCCESS;
335
336                 /* these three are not supported */
337         case COPY_ABORTED:
338         case VOLUME_OVERFLOW:
339         case MISCOMPARE:
340                 return SUCCESS;
341
342         case MEDIUM_ERROR:
343                 return NEEDS_RETRY;
344
345         case HARDWARE_ERROR:
346                 if (scmd->device->retry_hwerror)
347                         return NEEDS_RETRY;
348                 else
349                         return SUCCESS;
350
351         case ILLEGAL_REQUEST:
352         case BLANK_CHECK:
353         case DATA_PROTECT:
354         default:
355                 return SUCCESS;
356         }
357 }
358
359 /**
360  * scsi_eh_completed_normally - Disposition a eh cmd on return from LLD.
361  * @scmd:       SCSI cmd to examine.
362  *
363  * Notes:
364  *    This is *only* called when we are examining the status of commands
365  *    queued during error recovery.  the main difference here is that we
366  *    don't allow for the possibility of retries here, and we are a lot
367  *    more restrictive about what we consider acceptable.
368  **/
369 static int scsi_eh_completed_normally(struct scsi_cmnd *scmd)
370 {
371         /*
372          * first check the host byte, to see if there is anything in there
373          * that would indicate what we need to do.
374          */
375         if (host_byte(scmd->result) == DID_RESET) {
376                 /*
377                  * rats.  we are already in the error handler, so we now
378                  * get to try and figure out what to do next.  if the sense
379                  * is valid, we have a pretty good idea of what to do.
380                  * if not, we mark it as FAILED.
381                  */
382                 return scsi_check_sense(scmd);
383         }
384         if (host_byte(scmd->result) != DID_OK)
385                 return FAILED;
386
387         /*
388          * next, check the message byte.
389          */
390         if (msg_byte(scmd->result) != COMMAND_COMPLETE)
391                 return FAILED;
392
393         /*
394          * now, check the status byte to see if this indicates
395          * anything special.
396          */
397         switch (status_byte(scmd->result)) {
398         case GOOD:
399         case COMMAND_TERMINATED:
400                 return SUCCESS;
401         case CHECK_CONDITION:
402                 return scsi_check_sense(scmd);
403         case CONDITION_GOOD:
404         case INTERMEDIATE_GOOD:
405         case INTERMEDIATE_C_GOOD:
406                 /*
407                  * who knows?  FIXME(eric)
408                  */
409                 return SUCCESS;
410         case BUSY:
411         case QUEUE_FULL:
412         case RESERVATION_CONFLICT:
413         default:
414                 return FAILED;
415         }
416         return FAILED;
417 }
418
419 /**
420  * scsi_eh_times_out - timeout function for error handling.
421  * @scmd:       Cmd that is timing out.
422  *
423  * Notes:
424  *    During error handling, the kernel thread will be sleeping waiting
425  *    for some action to complete on the device.  our only job is to
426  *    record that it timed out, and to wake up the thread.
427  **/
428 static void scsi_eh_times_out(struct scsi_cmnd *scmd)
429 {
430         scmd->eh_eflags |= SCSI_EH_REC_TIMEOUT;
431         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: scmd:%p\n", __FUNCTION__,
432                                           scmd));
433
434         up(scmd->device->host->eh_action);
435 }
436
437 /**
438  * scsi_eh_done - Completion function for error handling.
439  * @scmd:       Cmd that is done.
440  **/
441 static void scsi_eh_done(struct scsi_cmnd *scmd)
442 {
443         /*
444          * if the timeout handler is already running, then just set the
445          * flag which says we finished late, and return.  we have no
446          * way of stopping the timeout handler from running, so we must
447          * always defer to it.
448          */
449         if (del_timer(&scmd->eh_timeout)) {
450                 scmd->request->rq_status = RQ_SCSI_DONE;
451
452                 SCSI_LOG_ERROR_RECOVERY(3, printk("%s scmd: %p result: %x\n",
453                                            __FUNCTION__, scmd, scmd->result));
454
455                 up(scmd->device->host->eh_action);
456         }
457 }
458
459 /**
460  * scsi_send_eh_cmnd  - send a cmd to a device as part of error recovery.
461  * @scmd:       SCSI Cmd to send.
462  * @timeout:    Timeout for cmd.
463  *
464  * Notes:
465  *    The initialization of the structures is quite a bit different in
466  *    this case, and furthermore, there is a different completion handler
467  *    vs scsi_dispatch_cmd.
468  * Return value:
469  *    SUCCESS or FAILED or NEEDS_RETRY
470  **/
471 static int scsi_send_eh_cmnd(struct scsi_cmnd *scmd, int timeout)
472 {
473         struct scsi_device *sdev = scmd->device;
474         struct Scsi_Host *shost = sdev->host;
475         DECLARE_MUTEX_LOCKED(sem);
476         unsigned long flags;
477         int rtn = SUCCESS;
478
479         /*
480          * we will use a queued command if possible, otherwise we will
481          * emulate the queuing and calling of completion function ourselves.
482          */
483         if (sdev->scsi_level <= SCSI_2)
484                 scmd->cmnd[1] = (scmd->cmnd[1] & 0x1f) |
485                         (sdev->lun << 5 & 0xe0);
486
487         scsi_add_timer(scmd, timeout, scsi_eh_times_out);
488
489         /*
490          * set up the semaphore so we wait for the command to complete.
491          */
492         shost->eh_action = &sem;
493         scmd->request->rq_status = RQ_SCSI_BUSY;
494
495         spin_lock_irqsave(shost->host_lock, flags);
496         scsi_log_send(scmd);
497         shost->hostt->queuecommand(scmd, scsi_eh_done);
498         spin_unlock_irqrestore(shost->host_lock, flags);
499
500         down(&sem);
501         scsi_log_completion(scmd, SUCCESS);
502
503         shost->eh_action = NULL;
504
505         /*
506          * see if timeout.  if so, tell the host to forget about it.
507          * in other words, we don't want a callback any more.
508          */
509         if (scmd->eh_eflags & SCSI_EH_REC_TIMEOUT) {
510                 scmd->eh_eflags &= ~SCSI_EH_REC_TIMEOUT;
511
512                 /*
513                  * as far as the low level driver is
514                  * concerned, this command is still active, so
515                  * we must give the low level driver a chance
516                  * to abort it. (db) 
517                  *
518                  * FIXME(eric) - we are not tracking whether we could
519                  * abort a timed out command or not.  not sure how
520                  * we should treat them differently anyways.
521                  */
522                 if (shost->hostt->eh_abort_handler)
523                         shost->hostt->eh_abort_handler(scmd);
524                         
525                 scmd->request->rq_status = RQ_SCSI_DONE;
526                 rtn = FAILED;
527         }
528
529         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: scmd: %p, rtn:%x\n",
530                                           __FUNCTION__, scmd, rtn));
531
532         /*
533          * now examine the actual status codes to see whether the command
534          * actually did complete normally.
535          */
536         if (rtn == SUCCESS) {
537                 rtn = scsi_eh_completed_normally(scmd);
538                 SCSI_LOG_ERROR_RECOVERY(3,
539                         printk("%s: scsi_eh_completed_normally %x\n",
540                                __FUNCTION__, rtn));
541                 switch (rtn) {
542                 case SUCCESS:
543                 case NEEDS_RETRY:
544                 case FAILED:
545                         break;
546                 default:
547                         rtn = FAILED;
548                         break;
549                 }
550         }
551
552         return rtn;
553 }
554
555 /**
556  * scsi_request_sense - Request sense data from a particular target.
557  * @scmd:       SCSI cmd for request sense.
558  *
559  * Notes:
560  *    Some hosts automatically obtain this information, others require
561  *    that we obtain it on our own. This function will *not* return until
562  *    the command either times out, or it completes.
563  **/
564 static int scsi_request_sense(struct scsi_cmnd *scmd)
565 {
566         static unsigned char generic_sense[6] =
567         {REQUEST_SENSE, 0, 0, 0, 252, 0};
568         unsigned char *scsi_result;
569         int saved_result;
570         int rtn;
571
572         memcpy(scmd->cmnd, generic_sense, sizeof(generic_sense));
573
574         scsi_result = kmalloc(252, GFP_ATOMIC | ((scmd->device->host->hostt->unchecked_isa_dma) ? __GFP_DMA : 0));
575
576
577         if (unlikely(!scsi_result)) {
578                 printk(KERN_ERR "%s: cannot allocate scsi_result.\n",
579                        __FUNCTION__);
580                 return FAILED;
581         }
582
583         /*
584          * zero the sense buffer.  some host adapters automatically always
585          * request sense, so it is not a good idea that
586          * scmd->request_buffer and scmd->sense_buffer point to the same
587          * address (db).  0 is not a valid sense code. 
588          */
589         memset(scmd->sense_buffer, 0, sizeof(scmd->sense_buffer));
590         memset(scsi_result, 0, 252);
591
592         saved_result = scmd->result;
593         scmd->request_buffer = scsi_result;
594         scmd->request_bufflen = 252;
595         scmd->use_sg = 0;
596         scmd->cmd_len = COMMAND_SIZE(scmd->cmnd[0]);
597         scmd->sc_data_direction = DMA_FROM_DEVICE;
598         scmd->underflow = 0;
599
600         rtn = scsi_send_eh_cmnd(scmd, SENSE_TIMEOUT);
601
602         /* last chance to have valid sense data */
603         if(!SCSI_SENSE_VALID(scmd)) {
604                 memcpy(scmd->sense_buffer, scmd->request_buffer,
605                        sizeof(scmd->sense_buffer));
606         }
607
608         kfree(scsi_result);
609
610         /*
611          * when we eventually call scsi_finish, we really wish to complete
612          * the original request, so let's restore the original data. (db)
613          */
614         scsi_setup_cmd_retry(scmd);
615         scmd->result = saved_result;
616         return rtn;
617 }
618
619 /**
620  * scsi_eh_finish_cmd - Handle a cmd that eh is finished with.
621  * @scmd:       Original SCSI cmd that eh has finished.
622  * @done_q:     Queue for processed commands.
623  *
624  * Notes:
625  *    We don't want to use the normal command completion while we are are
626  *    still handling errors - it may cause other commands to be queued,
627  *    and that would disturb what we are doing.  thus we really want to
628  *    keep a list of pending commands for final completion, and once we
629  *    are ready to leave error handling we handle completion for real.
630  **/
631 static void scsi_eh_finish_cmd(struct scsi_cmnd *scmd,
632                                struct list_head *done_q)
633 {
634         scmd->device->host->host_failed--;
635         scmd->eh_eflags = 0;
636
637         /*
638          * set this back so that the upper level can correctly free up
639          * things.
640          */
641         scsi_setup_cmd_retry(scmd);
642         list_move_tail(&scmd->eh_entry, done_q);
643 }
644
645 /**
646  * scsi_eh_get_sense - Get device sense data.
647  * @work_q:     Queue of commands to process.
648  * @done_q:     Queue of proccessed commands..
649  *
650  * Description:
651  *    See if we need to request sense information.  if so, then get it
652  *    now, so we have a better idea of what to do.  
653  *
654  * Notes:
655  *    This has the unfortunate side effect that if a shost adapter does
656  *    not automatically request sense information, that we end up shutting
657  *    it down before we request it.
658  *
659  *    All drivers should request sense information internally these days,
660  *    so for now all I have to say is tough noogies if you end up in here.
661  *
662  *    XXX: Long term this code should go away, but that needs an audit of
663  *         all LLDDs first.
664  **/
665 static int scsi_eh_get_sense(struct list_head *work_q,
666                              struct list_head *done_q)
667 {
668         struct scsi_cmnd *scmd, *next;
669         int rtn;
670
671         list_for_each_entry_safe(scmd, next, work_q, eh_entry) {
672                 if ((scmd->eh_eflags & SCSI_EH_CANCEL_CMD) ||
673                     SCSI_SENSE_VALID(scmd))
674                         continue;
675
676                 SCSI_LOG_ERROR_RECOVERY(2, printk("%s: requesting sense"
677                                                   " for id: %d\n",
678                                                   current->comm,
679                                                   scmd->device->id));
680                 rtn = scsi_request_sense(scmd);
681                 if (rtn != SUCCESS)
682                         continue;
683
684                 SCSI_LOG_ERROR_RECOVERY(3, printk("sense requested for %p"
685                                                   " result %x\n", scmd,
686                                                   scmd->result));
687                 SCSI_LOG_ERROR_RECOVERY(3, scsi_print_sense("bh", scmd));
688
689                 rtn = scsi_decide_disposition(scmd);
690
691                 /*
692                  * if the result was normal, then just pass it along to the
693                  * upper level.
694                  */
695                 if (rtn == SUCCESS)
696                         /* we don't want this command reissued, just
697                          * finished with the sense data, so set
698                          * retries to the max allowed to ensure it
699                          * won't get reissued */
700                         scmd->retries = scmd->allowed;
701                 else if (rtn != NEEDS_RETRY)
702                         continue;
703
704                 scsi_eh_finish_cmd(scmd, done_q);
705         }
706
707         return list_empty(work_q);
708 }
709
710 /**
711  * scsi_try_to_abort_cmd - Ask host to abort a running command.
712  * @scmd:       SCSI cmd to abort from Lower Level.
713  *
714  * Notes:
715  *    This function will not return until the user's completion function
716  *    has been called.  there is no timeout on this operation.  if the
717  *    author of the low-level driver wishes this operation to be timed,
718  *    they can provide this facility themselves.  helper functions in
719  *    scsi_error.c can be supplied to make this easier to do.
720  **/
721 static int scsi_try_to_abort_cmd(struct scsi_cmnd *scmd)
722 {
723         if (!scmd->device->host->hostt->eh_abort_handler)
724                 return FAILED;
725
726         /*
727          * scsi_done was called just after the command timed out and before
728          * we had a chance to process it. (db)
729          */
730         if (scmd->serial_number == 0)
731                 return SUCCESS;
732         return scmd->device->host->hostt->eh_abort_handler(scmd);
733 }
734
735 /**
736  * scsi_eh_tur - Send TUR to device.
737  * @scmd:       Scsi cmd to send TUR
738  *
739  * Return value:
740  *    0 - Device is ready. 1 - Device NOT ready.
741  **/
742 static int scsi_eh_tur(struct scsi_cmnd *scmd)
743 {
744         static unsigned char tur_command[6] = {TEST_UNIT_READY, 0, 0, 0, 0, 0};
745         int retry_cnt = 1, rtn;
746         int saved_result;
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         saved_result = scmd->result;
758         scmd->request_buffer = NULL;
759         scmd->request_bufflen = 0;
760         scmd->use_sg = 0;
761         scmd->cmd_len = COMMAND_SIZE(scmd->cmnd[0]);
762         scmd->underflow = 0;
763         scmd->sc_data_direction = DMA_NONE;
764
765         rtn = scsi_send_eh_cmnd(scmd, SENSE_TIMEOUT);
766
767         /*
768          * when we eventually call scsi_finish, we really wish to complete
769          * the original request, so let's restore the original data. (db)
770          */
771         scsi_setup_cmd_retry(scmd);
772         scmd->result = saved_result;
773
774         /*
775          * hey, we are done.  let's look to see what happened.
776          */
777         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: scmd %p rtn %x\n",
778                 __FUNCTION__, scmd, rtn));
779         if (rtn == SUCCESS)
780                 return 0;
781         else if (rtn == NEEDS_RETRY) {
782                 if (retry_cnt--)
783                         goto retry_tur;
784                 return 0;
785         }
786         return 1;
787 }
788
789 /**
790  * scsi_eh_abort_cmds - abort canceled commands.
791  * @shost:      scsi host being recovered.
792  * @eh_done_q:  list_head for processed commands.
793  *
794  * Decription:
795  *    Try and see whether or not it makes sense to try and abort the
796  *    running command.  this only works out to be the case if we have one
797  *    command that has timed out.  if the command simply failed, it makes
798  *    no sense to try and abort the command, since as far as the shost
799  *    adapter is concerned, it isn't running.
800  **/
801 static int scsi_eh_abort_cmds(struct list_head *work_q,
802                               struct list_head *done_q)
803 {
804         struct scsi_cmnd *scmd, *next;
805         int rtn;
806
807         list_for_each_entry_safe(scmd, next, work_q, eh_entry) {
808                 if (!(scmd->eh_eflags & SCSI_EH_CANCEL_CMD))
809                         continue;
810                 SCSI_LOG_ERROR_RECOVERY(3, printk("%s: aborting cmd:"
811                                                   "0x%p\n", current->comm,
812                                                   scmd));
813                 rtn = scsi_try_to_abort_cmd(scmd);
814                 if (rtn == SUCCESS) {
815                         scmd->eh_eflags &= ~SCSI_EH_CANCEL_CMD;
816                         if (!scsi_device_online(scmd->device) ||
817                             !scsi_eh_tur(scmd)) {
818                                 scsi_eh_finish_cmd(scmd, done_q);
819                         }
820                                 
821                 } else
822                         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: aborting"
823                                                           " cmd failed:"
824                                                           "0x%p\n",
825                                                           current->comm,
826                                                           scmd));
827         }
828
829         return list_empty(work_q);
830 }
831
832 /**
833  * scsi_try_bus_device_reset - Ask host to perform a BDR on a dev
834  * @scmd:       SCSI cmd used to send BDR       
835  *
836  * Notes:
837  *    There is no timeout for this operation.  if this operation is
838  *    unreliable for a given host, then the host itself needs to put a
839  *    timer on it, and set the host back to a consistent state prior to
840  *    returning.
841  **/
842 static int scsi_try_bus_device_reset(struct scsi_cmnd *scmd)
843 {
844         int rtn;
845
846         if (!scmd->device->host->hostt->eh_device_reset_handler)
847                 return FAILED;
848
849         rtn = scmd->device->host->hostt->eh_device_reset_handler(scmd);
850         if (rtn == SUCCESS) {
851                 scmd->device->was_reset = 1;
852                 scmd->device->expecting_cc_ua = 1;
853         }
854
855         return rtn;
856 }
857
858 /**
859  * scsi_eh_try_stu - Send START_UNIT to device.
860  * @scmd:       Scsi cmd to send START_UNIT
861  *
862  * Return value:
863  *    0 - Device is ready. 1 - Device NOT ready.
864  **/
865 static int scsi_eh_try_stu(struct scsi_cmnd *scmd)
866 {
867         static unsigned char stu_command[6] = {START_STOP, 0, 0, 0, 1, 0};
868         int rtn;
869         int saved_result;
870
871         if (!scmd->device->allow_restart)
872                 return 1;
873
874         memcpy(scmd->cmnd, stu_command, sizeof(stu_command));
875
876         /*
877          * zero the sense buffer.  the scsi spec mandates that any
878          * untransferred sense data should be interpreted as being zero.
879          */
880         memset(scmd->sense_buffer, 0, sizeof(scmd->sense_buffer));
881
882         saved_result = scmd->result;
883         scmd->request_buffer = NULL;
884         scmd->request_bufflen = 0;
885         scmd->use_sg = 0;
886         scmd->cmd_len = COMMAND_SIZE(scmd->cmnd[0]);
887         scmd->underflow = 0;
888         scmd->sc_data_direction = DMA_NONE;
889
890         rtn = scsi_send_eh_cmnd(scmd, START_UNIT_TIMEOUT);
891
892         /*
893          * when we eventually call scsi_finish, we really wish to complete
894          * the original request, so let's restore the original data. (db)
895          */
896         scsi_setup_cmd_retry(scmd);
897         scmd->result = saved_result;
898
899         /*
900          * hey, we are done.  let's look to see what happened.
901          */
902         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: scmd %p rtn %x\n",
903                 __FUNCTION__, scmd, rtn));
904         if (rtn == SUCCESS)
905                 return 0;
906         return 1;
907 }
908
909  /**
910  * scsi_eh_stu - send START_UNIT if needed
911  * @shost:      scsi host being recovered.
912  * @eh_done_q:  list_head for processed commands.
913  *
914  * Notes:
915  *    If commands are failing due to not ready, initializing command required,
916  *      try revalidating the device, which will end up sending a start unit. 
917  **/
918 static int scsi_eh_stu(struct Scsi_Host *shost,
919                               struct list_head *work_q,
920                               struct list_head *done_q)
921 {
922         struct scsi_cmnd *scmd, *stu_scmd, *next;
923         struct scsi_device *sdev;
924
925         shost_for_each_device(sdev, shost) {
926                 stu_scmd = NULL;
927                 list_for_each_entry(scmd, work_q, eh_entry)
928                         if (scmd->device == sdev && SCSI_SENSE_VALID(scmd) &&
929                             scsi_check_sense(scmd) == FAILED ) {
930                                 stu_scmd = scmd;
931                                 break;
932                         }
933
934                 if (!stu_scmd)
935                         continue;
936
937                 SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Sending START_UNIT to sdev:"
938                                                   " 0x%p\n", current->comm, sdev));
939
940                 if (!scsi_eh_try_stu(stu_scmd)) {
941                         if (!scsi_device_online(sdev) ||
942                             !scsi_eh_tur(stu_scmd)) {
943                                 list_for_each_entry_safe(scmd, next,
944                                                           work_q, eh_entry) {
945                                         if (scmd->device == sdev)
946                                                 scsi_eh_finish_cmd(scmd, done_q);
947                                 }
948                         }
949                 } else {
950                         SCSI_LOG_ERROR_RECOVERY(3,
951                                                 printk("%s: START_UNIT failed to sdev:"
952                                                        " 0x%p\n", current->comm, sdev));
953                 }
954         }
955
956         return list_empty(work_q);
957 }
958
959
960 /**
961  * scsi_eh_bus_device_reset - send bdr if needed
962  * @shost:      scsi host being recovered.
963  * @eh_done_q:  list_head for processed commands.
964  *
965  * Notes:
966  *    Try a bus device reset.  still, look to see whether we have multiple
967  *    devices that are jammed or not - if we have multiple devices, it
968  *    makes no sense to try bus_device_reset - we really would need to try
969  *    a bus_reset instead. 
970  **/
971 static int scsi_eh_bus_device_reset(struct Scsi_Host *shost,
972                                     struct list_head *work_q,
973                                     struct list_head *done_q)
974 {
975         struct scsi_cmnd *scmd, *bdr_scmd, *next;
976         struct scsi_device *sdev;
977         int rtn;
978
979         shost_for_each_device(sdev, shost) {
980                 bdr_scmd = NULL;
981                 list_for_each_entry(scmd, work_q, eh_entry)
982                         if (scmd->device == sdev) {
983                                 bdr_scmd = scmd;
984                                 break;
985                         }
986
987                 if (!bdr_scmd)
988                         continue;
989
990                 SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Sending BDR sdev:"
991                                                   " 0x%p\n", current->comm,
992                                                   sdev));
993                 rtn = scsi_try_bus_device_reset(bdr_scmd);
994                 if (rtn == SUCCESS) {
995                         if (!scsi_device_online(sdev) ||
996                             !scsi_eh_tur(bdr_scmd)) {
997                                 list_for_each_entry_safe(scmd, next,
998                                                          work_q, eh_entry) {
999                                         if (scmd->device == sdev)
1000                                                 scsi_eh_finish_cmd(scmd,
1001                                                                    done_q);
1002                                 }
1003                         }
1004                 } else {
1005                         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: BDR"
1006                                                           " failed sdev:"
1007                                                           "0x%p\n",
1008                                                           current->comm,
1009                                                            sdev));
1010                 }
1011         }
1012
1013         return list_empty(work_q);
1014 }
1015
1016 /**
1017  * scsi_try_bus_reset - ask host to perform a bus reset
1018  * @scmd:       SCSI cmd to send bus reset.
1019  **/
1020 static int scsi_try_bus_reset(struct scsi_cmnd *scmd)
1021 {
1022         unsigned long flags;
1023         int rtn;
1024
1025         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Snd Bus RST\n",
1026                                           __FUNCTION__));
1027
1028         if (!scmd->device->host->hostt->eh_bus_reset_handler)
1029                 return FAILED;
1030
1031         rtn = scmd->device->host->hostt->eh_bus_reset_handler(scmd);
1032
1033         if (rtn == SUCCESS) {
1034                 if (!scmd->device->host->hostt->skip_settle_delay)
1035                         ssleep(BUS_RESET_SETTLE_TIME);
1036                 spin_lock_irqsave(scmd->device->host->host_lock, flags);
1037                 scsi_report_bus_reset(scmd->device->host, scmd->device->channel);
1038                 spin_unlock_irqrestore(scmd->device->host->host_lock, flags);
1039         }
1040
1041         return rtn;
1042 }
1043
1044 /**
1045  * scsi_try_host_reset - ask host adapter to reset itself
1046  * @scmd:       SCSI cmd to send hsot reset.
1047  **/
1048 static int scsi_try_host_reset(struct scsi_cmnd *scmd)
1049 {
1050         unsigned long flags;
1051         int rtn;
1052
1053         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Snd Host RST\n",
1054                                           __FUNCTION__));
1055
1056         if (!scmd->device->host->hostt->eh_host_reset_handler)
1057                 return FAILED;
1058
1059         rtn = scmd->device->host->hostt->eh_host_reset_handler(scmd);
1060
1061         if (rtn == SUCCESS) {
1062                 if (!scmd->device->host->hostt->skip_settle_delay)
1063                         ssleep(HOST_RESET_SETTLE_TIME);
1064                 spin_lock_irqsave(scmd->device->host->host_lock, flags);
1065                 scsi_report_bus_reset(scmd->device->host, scmd->device->channel);
1066                 spin_unlock_irqrestore(scmd->device->host->host_lock, flags);
1067         }
1068
1069         return rtn;
1070 }
1071
1072 /**
1073  * scsi_eh_bus_reset - send a bus reset 
1074  * @shost:      scsi host being recovered.
1075  * @eh_done_q:  list_head for processed commands.
1076  **/
1077 static int scsi_eh_bus_reset(struct Scsi_Host *shost,
1078                              struct list_head *work_q,
1079                              struct list_head *done_q)
1080 {
1081         struct scsi_cmnd *scmd, *chan_scmd, *next;
1082         unsigned int channel;
1083         int rtn;
1084
1085         /*
1086          * we really want to loop over the various channels, and do this on
1087          * a channel by channel basis.  we should also check to see if any
1088          * of the failed commands are on soft_reset devices, and if so, skip
1089          * the reset.  
1090          */
1091
1092         for (channel = 0; channel <= shost->max_channel; channel++) {
1093                 chan_scmd = NULL;
1094                 list_for_each_entry(scmd, work_q, eh_entry) {
1095                         if (channel == scmd->device->channel) {
1096                                 chan_scmd = scmd;
1097                                 break;
1098                                 /*
1099                                  * FIXME add back in some support for
1100                                  * soft_reset devices.
1101                                  */
1102                         }
1103                 }
1104
1105                 if (!chan_scmd)
1106                         continue;
1107                 SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Sending BRST chan:"
1108                                                   " %d\n", current->comm,
1109                                                   channel));
1110                 rtn = scsi_try_bus_reset(chan_scmd);
1111                 if (rtn == SUCCESS) {
1112                         list_for_each_entry_safe(scmd, next, work_q, eh_entry) {
1113                                 if (channel == scmd->device->channel)
1114                                         if (!scsi_device_online(scmd->device) ||
1115                                             !scsi_eh_tur(scmd))
1116                                                 scsi_eh_finish_cmd(scmd,
1117                                                                    done_q);
1118                         }
1119                 } else {
1120                         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: BRST"
1121                                                           " failed chan: %d\n",
1122                                                           current->comm,
1123                                                           channel));
1124                 }
1125         }
1126         return list_empty(work_q);
1127 }
1128
1129 /**
1130  * scsi_eh_host_reset - send a host reset 
1131  * @work_q:     list_head for processed commands.
1132  * @done_q:     list_head for processed commands.
1133  **/
1134 static int scsi_eh_host_reset(struct list_head *work_q,
1135                               struct list_head *done_q)
1136 {
1137         struct scsi_cmnd *scmd, *next;
1138         int rtn;
1139
1140         if (!list_empty(work_q)) {
1141                 scmd = list_entry(work_q->next,
1142                                   struct scsi_cmnd, eh_entry);
1143
1144                 SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Sending HRST\n"
1145                                                   , current->comm));
1146
1147                 rtn = scsi_try_host_reset(scmd);
1148                 if (rtn == SUCCESS) {
1149                         list_for_each_entry_safe(scmd, next, work_q, eh_entry) {
1150                                 if (!scsi_device_online(scmd->device) ||
1151                                     (!scsi_eh_try_stu(scmd) && !scsi_eh_tur(scmd)) ||
1152                                     !scsi_eh_tur(scmd))
1153                                         scsi_eh_finish_cmd(scmd, done_q);
1154                         }
1155                 } else {
1156                         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: HRST"
1157                                                           " failed\n",
1158                                                           current->comm));
1159                 }
1160         }
1161         return list_empty(work_q);
1162 }
1163
1164 /**
1165  * scsi_eh_offline_sdevs - offline scsi devices that fail to recover
1166  * @work_q:     list_head for processed commands.
1167  * @done_q:     list_head for processed commands.
1168  *
1169  **/
1170 static void scsi_eh_offline_sdevs(struct list_head *work_q,
1171                                   struct list_head *done_q)
1172 {
1173         struct scsi_cmnd *scmd, *next;
1174
1175         list_for_each_entry_safe(scmd, next, work_q, eh_entry) {
1176                 sdev_printk(KERN_INFO, scmd->device,
1177                             "scsi: Device offlined - not"
1178                             " ready after error recovery\n");
1179                 scsi_device_set_state(scmd->device, SDEV_OFFLINE);
1180                 if (scmd->eh_eflags & SCSI_EH_CANCEL_CMD) {
1181                         /*
1182                          * FIXME: Handle lost cmds.
1183                          */
1184                 }
1185                 scsi_eh_finish_cmd(scmd, done_q);
1186         }
1187         return;
1188 }
1189
1190 /**
1191  * scsi_decide_disposition - Disposition a cmd on return from LLD.
1192  * @scmd:       SCSI cmd to examine.
1193  *
1194  * Notes:
1195  *    This is *only* called when we are examining the status after sending
1196  *    out the actual data command.  any commands that are queued for error
1197  *    recovery (e.g. test_unit_ready) do *not* come through here.
1198  *
1199  *    When this routine returns failed, it means the error handler thread
1200  *    is woken.  In cases where the error code indicates an error that
1201  *    doesn't require the error handler read (i.e. we don't need to
1202  *    abort/reset), this function should return SUCCESS.
1203  **/
1204 int scsi_decide_disposition(struct scsi_cmnd *scmd)
1205 {
1206         int rtn;
1207
1208         /*
1209          * if the device is offline, then we clearly just pass the result back
1210          * up to the top level.
1211          */
1212         if (!scsi_device_online(scmd->device)) {
1213                 SCSI_LOG_ERROR_RECOVERY(5, printk("%s: device offline - report"
1214                                                   " as SUCCESS\n",
1215                                                   __FUNCTION__));
1216                 return SUCCESS;
1217         }
1218
1219         /*
1220          * first check the host byte, to see if there is anything in there
1221          * that would indicate what we need to do.
1222          */
1223         switch (host_byte(scmd->result)) {
1224         case DID_PASSTHROUGH:
1225                 /*
1226                  * no matter what, pass this through to the upper layer.
1227                  * nuke this special code so that it looks like we are saying
1228                  * did_ok.
1229                  */
1230                 scmd->result &= 0xff00ffff;
1231                 return SUCCESS;
1232         case DID_OK:
1233                 /*
1234                  * looks good.  drop through, and check the next byte.
1235                  */
1236                 break;
1237         case DID_NO_CONNECT:
1238         case DID_BAD_TARGET:
1239         case DID_ABORT:
1240                 /*
1241                  * note - this means that we just report the status back
1242                  * to the top level driver, not that we actually think
1243                  * that it indicates SUCCESS.
1244                  */
1245                 return SUCCESS;
1246                 /*
1247                  * when the low level driver returns did_soft_error,
1248                  * it is responsible for keeping an internal retry counter 
1249                  * in order to avoid endless loops (db)
1250                  *
1251                  * actually this is a bug in this function here.  we should
1252                  * be mindful of the maximum number of retries specified
1253                  * and not get stuck in a loop.
1254                  */
1255         case DID_SOFT_ERROR:
1256                 goto maybe_retry;
1257         case DID_IMM_RETRY:
1258                 return NEEDS_RETRY;
1259
1260         case DID_REQUEUE:
1261                 return ADD_TO_MLQUEUE;
1262
1263         case DID_ERROR:
1264                 if (msg_byte(scmd->result) == COMMAND_COMPLETE &&
1265                     status_byte(scmd->result) == RESERVATION_CONFLICT)
1266                         /*
1267                          * execute reservation conflict processing code
1268                          * lower down
1269                          */
1270                         break;
1271                 /* fallthrough */
1272
1273         case DID_BUS_BUSY:
1274         case DID_PARITY:
1275                 goto maybe_retry;
1276         case DID_TIME_OUT:
1277                 /*
1278                  * when we scan the bus, we get timeout messages for
1279                  * these commands if there is no device available.
1280                  * other hosts report did_no_connect for the same thing.
1281                  */
1282                 if ((scmd->cmnd[0] == TEST_UNIT_READY ||
1283                      scmd->cmnd[0] == INQUIRY)) {
1284                         return SUCCESS;
1285                 } else {
1286                         return FAILED;
1287                 }
1288         case DID_RESET:
1289                 return SUCCESS;
1290         default:
1291                 return FAILED;
1292         }
1293
1294         /*
1295          * next, check the message byte.
1296          */
1297         if (msg_byte(scmd->result) != COMMAND_COMPLETE)
1298                 return FAILED;
1299
1300         /*
1301          * check the status byte to see if this indicates anything special.
1302          */
1303         switch (status_byte(scmd->result)) {
1304         case QUEUE_FULL:
1305                 /*
1306                  * the case of trying to send too many commands to a
1307                  * tagged queueing device.
1308                  */
1309         case BUSY:
1310                 /*
1311                  * device can't talk to us at the moment.  Should only
1312                  * occur (SAM-3) when the task queue is empty, so will cause
1313                  * the empty queue handling to trigger a stall in the
1314                  * device.
1315                  */
1316                 return ADD_TO_MLQUEUE;
1317         case GOOD:
1318         case COMMAND_TERMINATED:
1319         case TASK_ABORTED:
1320                 return SUCCESS;
1321         case CHECK_CONDITION:
1322                 rtn = scsi_check_sense(scmd);
1323                 if (rtn == NEEDS_RETRY)
1324                         goto maybe_retry;
1325                 /* if rtn == FAILED, we have no sense information;
1326                  * returning FAILED will wake the error handler thread
1327                  * to collect the sense and redo the decide
1328                  * disposition */
1329                 return rtn;
1330         case CONDITION_GOOD:
1331         case INTERMEDIATE_GOOD:
1332         case INTERMEDIATE_C_GOOD:
1333         case ACA_ACTIVE:
1334                 /*
1335                  * who knows?  FIXME(eric)
1336                  */
1337                 return SUCCESS;
1338
1339         case RESERVATION_CONFLICT:
1340                 sdev_printk(KERN_INFO, scmd->device,
1341                             "reservation conflict\n");
1342                 return SUCCESS; /* causes immediate i/o error */
1343         default:
1344                 return FAILED;
1345         }
1346         return FAILED;
1347
1348       maybe_retry:
1349
1350         /* we requeue for retry because the error was retryable, and
1351          * the request was not marked fast fail.  Note that above,
1352          * even if the request is marked fast fail, we still requeue
1353          * for queue congestion conditions (QUEUE_FULL or BUSY) */
1354         if ((++scmd->retries) < scmd->allowed 
1355             && !blk_noretry_request(scmd->request)) {
1356                 return NEEDS_RETRY;
1357         } else {
1358                 /*
1359                  * no more retries - report this one back to upper level.
1360                  */
1361                 return SUCCESS;
1362         }
1363 }
1364
1365 /**
1366  * scsi_eh_lock_done - done function for eh door lock request
1367  * @scmd:       SCSI command block for the door lock request
1368  *
1369  * Notes:
1370  *      We completed the asynchronous door lock request, and it has either
1371  *      locked the door or failed.  We must free the command structures
1372  *      associated with this request.
1373  **/
1374 static void scsi_eh_lock_done(struct scsi_cmnd *scmd)
1375 {
1376         struct scsi_request *sreq = scmd->sc_request;
1377
1378         scsi_release_request(sreq);
1379 }
1380
1381
1382 /**
1383  * scsi_eh_lock_door - Prevent medium removal for the specified device
1384  * @sdev:       SCSI device to prevent medium removal
1385  *
1386  * Locking:
1387  *      We must be called from process context; scsi_allocate_request()
1388  *      may sleep.
1389  *
1390  * Notes:
1391  *      We queue up an asynchronous "ALLOW MEDIUM REMOVAL" request on the
1392  *      head of the devices request queue, and continue.
1393  *
1394  * Bugs:
1395  *      scsi_allocate_request() may sleep waiting for existing requests to
1396  *      be processed.  However, since we haven't kicked off any request
1397  *      processing for this host, this may deadlock.
1398  *
1399  *      If scsi_allocate_request() fails for what ever reason, we
1400  *      completely forget to lock the door.
1401  **/
1402 static void scsi_eh_lock_door(struct scsi_device *sdev)
1403 {
1404         struct scsi_request *sreq = scsi_allocate_request(sdev, GFP_KERNEL);
1405
1406         if (unlikely(!sreq)) {
1407                 printk(KERN_ERR "%s: request allocate failed,"
1408                        "prevent media removal cmd not sent\n", __FUNCTION__);
1409                 return;
1410         }
1411
1412         sreq->sr_cmnd[0] = ALLOW_MEDIUM_REMOVAL;
1413         sreq->sr_cmnd[1] = 0;
1414         sreq->sr_cmnd[2] = 0;
1415         sreq->sr_cmnd[3] = 0;
1416         sreq->sr_cmnd[4] = SCSI_REMOVAL_PREVENT;
1417         sreq->sr_cmnd[5] = 0;
1418         sreq->sr_data_direction = DMA_NONE;
1419         sreq->sr_bufflen = 0;
1420         sreq->sr_buffer = NULL;
1421         sreq->sr_allowed = 5;
1422         sreq->sr_done = scsi_eh_lock_done;
1423         sreq->sr_timeout_per_command = 10 * HZ;
1424         sreq->sr_cmd_len = COMMAND_SIZE(sreq->sr_cmnd[0]);
1425
1426         scsi_insert_special_req(sreq, 1);
1427 }
1428
1429
1430 /**
1431  * scsi_restart_operations - restart io operations to the specified host.
1432  * @shost:      Host we are restarting.
1433  *
1434  * Notes:
1435  *    When we entered the error handler, we blocked all further i/o to
1436  *    this device.  we need to 'reverse' this process.
1437  **/
1438 static void scsi_restart_operations(struct Scsi_Host *shost)
1439 {
1440         struct scsi_device *sdev;
1441         unsigned long flags;
1442
1443         /*
1444          * If the door was locked, we need to insert a door lock request
1445          * onto the head of the SCSI request queue for the device.  There
1446          * is no point trying to lock the door of an off-line device.
1447          */
1448         shost_for_each_device(sdev, shost) {
1449                 if (scsi_device_online(sdev) && sdev->locked)
1450                         scsi_eh_lock_door(sdev);
1451         }
1452
1453         /*
1454          * next free up anything directly waiting upon the host.  this
1455          * will be requests for character device operations, and also for
1456          * ioctls to queued block devices.
1457          */
1458         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: waking up host to restart\n",
1459                                           __FUNCTION__));
1460
1461         spin_lock_irqsave(shost->host_lock, flags);
1462         if (scsi_host_set_state(shost, SHOST_RUNNING))
1463                 if (scsi_host_set_state(shost, SHOST_CANCEL))
1464                         BUG_ON(scsi_host_set_state(shost, SHOST_DEL));
1465         spin_unlock_irqrestore(shost->host_lock, flags);
1466
1467         wake_up(&shost->host_wait);
1468
1469         /*
1470          * finally we need to re-initiate requests that may be pending.  we will
1471          * have had everything blocked while error handling is taking place, and
1472          * now that error recovery is done, we will need to ensure that these
1473          * requests are started.
1474          */
1475         scsi_run_host_queues(shost);
1476 }
1477
1478 /**
1479  * scsi_eh_ready_devs - check device ready state and recover if not.
1480  * @shost:      host to be recovered.
1481  * @eh_done_q:  list_head for processed commands.
1482  *
1483  **/
1484 static void scsi_eh_ready_devs(struct Scsi_Host *shost,
1485                                struct list_head *work_q,
1486                                struct list_head *done_q)
1487 {
1488         if (!scsi_eh_stu(shost, work_q, done_q))
1489                 if (!scsi_eh_bus_device_reset(shost, work_q, done_q))
1490                         if (!scsi_eh_bus_reset(shost, work_q, done_q))
1491                                 if (!scsi_eh_host_reset(work_q, done_q))
1492                                         scsi_eh_offline_sdevs(work_q, done_q);
1493 }
1494
1495 /**
1496  * scsi_eh_flush_done_q - finish processed commands or retry them.
1497  * @done_q:     list_head of processed commands.
1498  *
1499  **/
1500 static void scsi_eh_flush_done_q(struct list_head *done_q)
1501 {
1502         struct scsi_cmnd *scmd, *next;
1503
1504         list_for_each_entry_safe(scmd, next, done_q, eh_entry) {
1505                 list_del_init(&scmd->eh_entry);
1506                 if (scsi_device_online(scmd->device) &&
1507                     !blk_noretry_request(scmd->request) &&
1508                     (++scmd->retries < scmd->allowed)) {
1509                         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: flush"
1510                                                           " retry cmd: %p\n",
1511                                                           current->comm,
1512                                                           scmd));
1513                                 scsi_queue_insert(scmd, SCSI_MLQUEUE_EH_RETRY);
1514                 } else {
1515                         /*
1516                          * If just we got sense for the device (called
1517                          * scsi_eh_get_sense), scmd->result is already
1518                          * set, do not set DRIVER_TIMEOUT.
1519                          */
1520                         if (!scmd->result)
1521                                 scmd->result |= (DRIVER_TIMEOUT << 24);
1522                         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: flush finish"
1523                                                         " cmd: %p\n",
1524                                                         current->comm, scmd));
1525                         scsi_finish_command(scmd);
1526                 }
1527         }
1528 }
1529
1530 /**
1531  * scsi_unjam_host - Attempt to fix a host which has a cmd that failed.
1532  * @shost:      Host to unjam.
1533  *
1534  * Notes:
1535  *    When we come in here, we *know* that all commands on the bus have
1536  *    either completed, failed or timed out.  we also know that no further
1537  *    commands are being sent to the host, so things are relatively quiet
1538  *    and we have freedom to fiddle with things as we wish.
1539  *
1540  *    This is only the *default* implementation.  it is possible for
1541  *    individual drivers to supply their own version of this function, and
1542  *    if the maintainer wishes to do this, it is strongly suggested that
1543  *    this function be taken as a template and modified.  this function
1544  *    was designed to correctly handle problems for about 95% of the
1545  *    different cases out there, and it should always provide at least a
1546  *    reasonable amount of error recovery.
1547  *
1548  *    Any command marked 'failed' or 'timeout' must eventually have
1549  *    scsi_finish_cmd() called for it.  we do all of the retry stuff
1550  *    here, so when we restart the host after we return it should have an
1551  *    empty queue.
1552  **/
1553 static void scsi_unjam_host(struct Scsi_Host *shost)
1554 {
1555         unsigned long flags;
1556         LIST_HEAD(eh_work_q);
1557         LIST_HEAD(eh_done_q);
1558
1559         spin_lock_irqsave(shost->host_lock, flags);
1560         list_splice_init(&shost->eh_cmd_q, &eh_work_q);
1561         spin_unlock_irqrestore(shost->host_lock, flags);
1562
1563         SCSI_LOG_ERROR_RECOVERY(1, scsi_eh_prt_fail_stats(shost, &eh_work_q));
1564
1565         if (!scsi_eh_get_sense(&eh_work_q, &eh_done_q))
1566                 if (!scsi_eh_abort_cmds(&eh_work_q, &eh_done_q))
1567                         scsi_eh_ready_devs(shost, &eh_work_q, &eh_done_q);
1568
1569         scsi_eh_flush_done_q(&eh_done_q);
1570 }
1571
1572 /**
1573  * scsi_error_handler - Handle errors/timeouts of SCSI cmds.
1574  * @data:       Host for which we are running.
1575  *
1576  * Notes:
1577  *    This is always run in the context of a kernel thread.  The idea is
1578  *    that we start this thing up when the kernel starts up (one per host
1579  *    that we detect), and it immediately goes to sleep and waits for some
1580  *    event (i.e. failure).  When this takes place, we have the job of
1581  *    trying to unjam the bus and restarting things.
1582  **/
1583 int scsi_error_handler(void *data)
1584 {
1585         struct Scsi_Host *shost = (struct Scsi_Host *) data;
1586         int rtn;
1587
1588         current->flags |= PF_NOFREEZE;
1589
1590         
1591         /*
1592          * Note - we always use TASK_INTERRUPTIBLE even if the module
1593          * was loaded as part of the kernel.  The reason is that
1594          * UNINTERRUPTIBLE would cause this thread to be counted in
1595          * the load average as a running process, and an interruptible
1596          * wait doesn't.
1597          */
1598         set_current_state(TASK_INTERRUPTIBLE);
1599         while (!kthread_should_stop()) {
1600                 if (shost->host_failed == 0 ||
1601                     shost->host_failed != shost->host_busy) {
1602                         SCSI_LOG_ERROR_RECOVERY(1, printk("Error handler"
1603                                                           " scsi_eh_%d"
1604                                                           " sleeping\n",
1605                                                           shost->host_no));
1606                         schedule();
1607                         set_current_state(TASK_INTERRUPTIBLE);
1608                         continue;
1609                 }
1610
1611                 __set_current_state(TASK_RUNNING);
1612                 SCSI_LOG_ERROR_RECOVERY(1, printk("Error handler"
1613                                                   " scsi_eh_%d waking"
1614                                                   " up\n",shost->host_no));
1615
1616                 shost->eh_active = 1;
1617
1618                 /*
1619                  * We have a host that is failing for some reason.  Figure out
1620                  * what we need to do to get it up and online again (if we can).
1621                  * If we fail, we end up taking the thing offline.
1622                  */
1623                 if (shost->hostt->eh_strategy_handler) 
1624                         rtn = shost->hostt->eh_strategy_handler(shost);
1625                 else
1626                         scsi_unjam_host(shost);
1627
1628                 shost->eh_active = 0;
1629
1630                 /*
1631                  * Note - if the above fails completely, the action is to take
1632                  * individual devices offline and flush the queue of any
1633                  * outstanding requests that may have been pending.  When we
1634                  * restart, we restart any I/O to any other devices on the bus
1635                  * which are still online.
1636                  */
1637                 scsi_restart_operations(shost);
1638                 set_current_state(TASK_INTERRUPTIBLE);
1639         }
1640
1641         __set_current_state(TASK_RUNNING);
1642
1643         SCSI_LOG_ERROR_RECOVERY(1, printk("Error handler scsi_eh_%d"
1644                                           " exiting\n",shost->host_no));
1645
1646         /*
1647          * Make sure that nobody tries to wake us up again.
1648          */
1649         shost->ehandler = NULL;
1650         return 0;
1651 }
1652
1653 /*
1654  * Function:    scsi_report_bus_reset()
1655  *
1656  * Purpose:     Utility function used by low-level drivers to report that
1657  *              they have observed a bus reset on the bus being handled.
1658  *
1659  * Arguments:   shost       - Host in question
1660  *              channel     - channel on which reset was observed.
1661  *
1662  * Returns:     Nothing
1663  *
1664  * Lock status: Host lock must be held.
1665  *
1666  * Notes:       This only needs to be called if the reset is one which
1667  *              originates from an unknown location.  Resets originated
1668  *              by the mid-level itself don't need to call this, but there
1669  *              should be no harm.
1670  *
1671  *              The main purpose of this is to make sure that a CHECK_CONDITION
1672  *              is properly treated.
1673  */
1674 void scsi_report_bus_reset(struct Scsi_Host *shost, int channel)
1675 {
1676         struct scsi_device *sdev;
1677
1678         __shost_for_each_device(sdev, shost) {
1679                 if (channel == sdev->channel) {
1680                         sdev->was_reset = 1;
1681                         sdev->expecting_cc_ua = 1;
1682                 }
1683         }
1684 }
1685 EXPORT_SYMBOL(scsi_report_bus_reset);
1686
1687 /*
1688  * Function:    scsi_report_device_reset()
1689  *
1690  * Purpose:     Utility function used by low-level drivers to report that
1691  *              they have observed a device reset on the device being handled.
1692  *
1693  * Arguments:   shost       - Host in question
1694  *              channel     - channel on which reset was observed
1695  *              target      - target on which reset was observed
1696  *
1697  * Returns:     Nothing
1698  *
1699  * Lock status: Host lock must be held
1700  *
1701  * Notes:       This only needs to be called if the reset is one which
1702  *              originates from an unknown location.  Resets originated
1703  *              by the mid-level itself don't need to call this, but there
1704  *              should be no harm.
1705  *
1706  *              The main purpose of this is to make sure that a CHECK_CONDITION
1707  *              is properly treated.
1708  */
1709 void scsi_report_device_reset(struct Scsi_Host *shost, int channel, int target)
1710 {
1711         struct scsi_device *sdev;
1712
1713         __shost_for_each_device(sdev, shost) {
1714                 if (channel == sdev->channel &&
1715                     target == sdev->id) {
1716                         sdev->was_reset = 1;
1717                         sdev->expecting_cc_ua = 1;
1718                 }
1719         }
1720 }
1721 EXPORT_SYMBOL(scsi_report_device_reset);
1722
1723 static void
1724 scsi_reset_provider_done_command(struct scsi_cmnd *scmd)
1725 {
1726 }
1727
1728 /*
1729  * Function:    scsi_reset_provider
1730  *
1731  * Purpose:     Send requested reset to a bus or device at any phase.
1732  *
1733  * Arguments:   device  - device to send reset to
1734  *              flag - reset type (see scsi.h)
1735  *
1736  * Returns:     SUCCESS/FAILURE.
1737  *
1738  * Notes:       This is used by the SCSI Generic driver to provide
1739  *              Bus/Device reset capability.
1740  */
1741 int
1742 scsi_reset_provider(struct scsi_device *dev, int flag)
1743 {
1744         struct scsi_cmnd *scmd = scsi_get_command(dev, GFP_KERNEL);
1745         struct request req;
1746         int rtn;
1747
1748         scmd->request = &req;
1749         memset(&scmd->eh_timeout, 0, sizeof(scmd->eh_timeout));
1750         scmd->request->rq_status        = RQ_SCSI_BUSY;
1751
1752         memset(&scmd->cmnd, '\0', sizeof(scmd->cmnd));
1753     
1754         scmd->scsi_done         = scsi_reset_provider_done_command;
1755         scmd->done                      = NULL;
1756         scmd->buffer                    = NULL;
1757         scmd->bufflen                   = 0;
1758         scmd->request_buffer            = NULL;
1759         scmd->request_bufflen           = 0;
1760
1761         scmd->cmd_len                   = 0;
1762
1763         scmd->sc_data_direction         = DMA_BIDIRECTIONAL;
1764         scmd->sc_request                = NULL;
1765         scmd->sc_magic                  = SCSI_CMND_MAGIC;
1766
1767         init_timer(&scmd->eh_timeout);
1768
1769         /*
1770          * Sometimes the command can get back into the timer chain,
1771          * so use the pid as an identifier.
1772          */
1773         scmd->pid                       = 0;
1774
1775         switch (flag) {
1776         case SCSI_TRY_RESET_DEVICE:
1777                 rtn = scsi_try_bus_device_reset(scmd);
1778                 if (rtn == SUCCESS)
1779                         break;
1780                 /* FALLTHROUGH */
1781         case SCSI_TRY_RESET_BUS:
1782                 rtn = scsi_try_bus_reset(scmd);
1783                 if (rtn == SUCCESS)
1784                         break;
1785                 /* FALLTHROUGH */
1786         case SCSI_TRY_RESET_HOST:
1787                 rtn = scsi_try_host_reset(scmd);
1788                 break;
1789         default:
1790                 rtn = FAILED;
1791         }
1792
1793         scsi_next_command(scmd);
1794         return rtn;
1795 }
1796 EXPORT_SYMBOL(scsi_reset_provider);
1797
1798 /**
1799  * scsi_normalize_sense - normalize main elements from either fixed or
1800  *                      descriptor sense data format into a common format.
1801  *
1802  * @sense_buffer:       byte array containing sense data returned by device
1803  * @sb_len:             number of valid bytes in sense_buffer
1804  * @sshdr:              pointer to instance of structure that common
1805  *                      elements are written to.
1806  *
1807  * Notes:
1808  *      The "main elements" from sense data are: response_code, sense_key,
1809  *      asc, ascq and additional_length (only for descriptor format).
1810  *
1811  *      Typically this function can be called after a device has
1812  *      responded to a SCSI command with the CHECK_CONDITION status.
1813  *
1814  * Return value:
1815  *      1 if valid sense data information found, else 0;
1816  **/
1817 int scsi_normalize_sense(const u8 *sense_buffer, int sb_len,
1818                          struct scsi_sense_hdr *sshdr)
1819 {
1820         if (!sense_buffer || !sb_len)
1821                 return 0;
1822
1823         memset(sshdr, 0, sizeof(struct scsi_sense_hdr));
1824
1825         sshdr->response_code = (sense_buffer[0] & 0x7f);
1826
1827         if (!scsi_sense_valid(sshdr))
1828                 return 0;
1829
1830         if (sshdr->response_code >= 0x72) {
1831                 /*
1832                  * descriptor format
1833                  */
1834                 if (sb_len > 1)
1835                         sshdr->sense_key = (sense_buffer[1] & 0xf);
1836                 if (sb_len > 2)
1837                         sshdr->asc = sense_buffer[2];
1838                 if (sb_len > 3)
1839                         sshdr->ascq = sense_buffer[3];
1840                 if (sb_len > 7)
1841                         sshdr->additional_length = sense_buffer[7];
1842         } else {
1843                 /* 
1844                  * fixed format
1845                  */
1846                 if (sb_len > 2)
1847                         sshdr->sense_key = (sense_buffer[2] & 0xf);
1848                 if (sb_len > 7) {
1849                         sb_len = (sb_len < (sense_buffer[7] + 8)) ?
1850                                          sb_len : (sense_buffer[7] + 8);
1851                         if (sb_len > 12)
1852                                 sshdr->asc = sense_buffer[12];
1853                         if (sb_len > 13)
1854                                 sshdr->ascq = sense_buffer[13];
1855                 }
1856         }
1857
1858         return 1;
1859 }
1860 EXPORT_SYMBOL(scsi_normalize_sense);
1861
1862 int scsi_request_normalize_sense(struct scsi_request *sreq,
1863                                  struct scsi_sense_hdr *sshdr)
1864 {
1865         return scsi_normalize_sense(sreq->sr_sense_buffer,
1866                         sizeof(sreq->sr_sense_buffer), sshdr);
1867 }
1868 EXPORT_SYMBOL(scsi_request_normalize_sense);
1869
1870 int scsi_command_normalize_sense(struct scsi_cmnd *cmd,
1871                                  struct scsi_sense_hdr *sshdr)
1872 {
1873         return scsi_normalize_sense(cmd->sense_buffer,
1874                         sizeof(cmd->sense_buffer), sshdr);
1875 }
1876 EXPORT_SYMBOL(scsi_command_normalize_sense);
1877
1878 /**
1879  * scsi_sense_desc_find - search for a given descriptor type in
1880  *                      descriptor sense data format.
1881  *
1882  * @sense_buffer:       byte array of descriptor format sense data
1883  * @sb_len:             number of valid bytes in sense_buffer
1884  * @desc_type:          value of descriptor type to find
1885  *                      (e.g. 0 -> information)
1886  *
1887  * Notes:
1888  *      only valid when sense data is in descriptor format
1889  *
1890  * Return value:
1891  *      pointer to start of (first) descriptor if found else NULL
1892  **/
1893 const u8 * scsi_sense_desc_find(const u8 * sense_buffer, int sb_len,
1894                                 int desc_type)
1895 {
1896         int add_sen_len, add_len, desc_len, k;
1897         const u8 * descp;
1898
1899         if ((sb_len < 8) || (0 == (add_sen_len = sense_buffer[7])))
1900                 return NULL;
1901         if ((sense_buffer[0] < 0x72) || (sense_buffer[0] > 0x73))
1902                 return NULL;
1903         add_sen_len = (add_sen_len < (sb_len - 8)) ?
1904                         add_sen_len : (sb_len - 8);
1905         descp = &sense_buffer[8];
1906         for (desc_len = 0, k = 0; k < add_sen_len; k += desc_len) {
1907                 descp += desc_len;
1908                 add_len = (k < (add_sen_len - 1)) ? descp[1]: -1;
1909                 desc_len = add_len + 2;
1910                 if (descp[0] == desc_type)
1911                         return descp;
1912                 if (add_len < 0) // short descriptor ??
1913                         break;
1914         }
1915         return NULL;
1916 }
1917 EXPORT_SYMBOL(scsi_sense_desc_find);
1918
1919 /**
1920  * scsi_get_sense_info_fld - attempts to get information field from
1921  *                      sense data (either fixed or descriptor format)
1922  *
1923  * @sense_buffer:       byte array of sense data
1924  * @sb_len:             number of valid bytes in sense_buffer
1925  * @info_out:           pointer to 64 integer where 8 or 4 byte information
1926  *                      field will be placed if found.
1927  *
1928  * Return value:
1929  *      1 if information field found, 0 if not found.
1930  **/
1931 int scsi_get_sense_info_fld(const u8 * sense_buffer, int sb_len,
1932                             u64 * info_out)
1933 {
1934         int j;
1935         const u8 * ucp;
1936         u64 ull;
1937
1938         if (sb_len < 7)
1939                 return 0;
1940         switch (sense_buffer[0] & 0x7f) {
1941         case 0x70:
1942         case 0x71:
1943                 if (sense_buffer[0] & 0x80) {
1944                         *info_out = (sense_buffer[3] << 24) +
1945                                     (sense_buffer[4] << 16) +
1946                                     (sense_buffer[5] << 8) + sense_buffer[6];
1947                         return 1;
1948                 } else
1949                         return 0;
1950         case 0x72:
1951         case 0x73:
1952                 ucp = scsi_sense_desc_find(sense_buffer, sb_len,
1953                                            0 /* info desc */);
1954                 if (ucp && (0xa == ucp[1])) {
1955                         ull = 0;
1956                         for (j = 0; j < 8; ++j) {
1957                                 if (j > 0)
1958                                         ull <<= 8;
1959                                 ull |= ucp[4 + j];
1960                         }
1961                         *info_out = ull;
1962                         return 1;
1963                 } else
1964                         return 0;
1965         default:
1966                 return 0;
1967         }
1968 }
1969 EXPORT_SYMBOL(scsi_get_sense_info_fld);