Merge HEAD from ../scsi-misc-2.6-tmp
[linux-2.6] / 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         scmd->eh_eflags |= eh_flag;
77         list_add_tail(&scmd->eh_entry, &shost->eh_cmd_q);
78         scsi_host_set_state(shost, SHOST_RECOVERY);
79         shost->host_failed++;
80         scsi_eh_wakeup(shost);
81         spin_unlock_irqrestore(shost->host_lock, flags);
82         return 1;
83 }
84
85 /**
86  * scsi_add_timer - Start timeout timer for a single scsi command.
87  * @scmd:       scsi command that is about to start running.
88  * @timeout:    amount of time to allow this command to run.
89  * @complete:   timeout function to call if timer isn't canceled.
90  *
91  * Notes:
92  *    This should be turned into an inline function.  Each scsi command
93  *    has its own timer, and as it is added to the queue, we set up the
94  *    timer.  When the command completes, we cancel the timer.
95  **/
96 void scsi_add_timer(struct scsi_cmnd *scmd, int timeout,
97                     void (*complete)(struct scsi_cmnd *))
98 {
99
100         /*
101          * If the clock was already running for this command, then
102          * first delete the timer.  The timer handling code gets rather
103          * confused if we don't do this.
104          */
105         if (scmd->eh_timeout.function)
106                 del_timer(&scmd->eh_timeout);
107
108         scmd->eh_timeout.data = (unsigned long)scmd;
109         scmd->eh_timeout.expires = jiffies + timeout;
110         scmd->eh_timeout.function = (void (*)(unsigned long)) complete;
111
112         SCSI_LOG_ERROR_RECOVERY(5, printk("%s: scmd: %p, time:"
113                                           " %d, (%p)\n", __FUNCTION__,
114                                           scmd, timeout, complete));
115
116         add_timer(&scmd->eh_timeout);
117 }
118 EXPORT_SYMBOL(scsi_add_timer);
119
120 /**
121  * scsi_delete_timer - Delete/cancel timer for a given function.
122  * @scmd:       Cmd that we are canceling timer for
123  *
124  * Notes:
125  *     This should be turned into an inline function.
126  *
127  * Return value:
128  *     1 if we were able to detach the timer.  0 if we blew it, and the
129  *     timer function has already started to run.
130  **/
131 int scsi_delete_timer(struct scsi_cmnd *scmd)
132 {
133         int rtn;
134
135         rtn = del_timer(&scmd->eh_timeout);
136
137         SCSI_LOG_ERROR_RECOVERY(5, printk("%s: scmd: %p,"
138                                          " rtn: %d\n", __FUNCTION__,
139                                          scmd, rtn));
140
141         scmd->eh_timeout.data = (unsigned long)NULL;
142         scmd->eh_timeout.function = NULL;
143
144         return rtn;
145 }
146 EXPORT_SYMBOL(scsi_delete_timer);
147
148 /**
149  * scsi_times_out - Timeout function for normal scsi commands.
150  * @scmd:       Cmd that is timing out.
151  *
152  * Notes:
153  *     We do not need to lock this.  There is the potential for a race
154  *     only in that the normal completion handling might run, but if the
155  *     normal completion function determines that the timer has already
156  *     fired, then it mustn't do anything.
157  **/
158 void scsi_times_out(struct scsi_cmnd *scmd)
159 {
160         scsi_log_completion(scmd, TIMEOUT_ERROR);
161
162         if (scmd->device->host->hostt->eh_timed_out)
163                 switch (scmd->device->host->hostt->eh_timed_out(scmd)) {
164                 case EH_HANDLED:
165                         __scsi_done(scmd);
166                         return;
167                 case EH_RESET_TIMER:
168                         /* This allows a single retry even of a command
169                          * with allowed == 0 */
170                         if (scmd->retries++ > scmd->allowed)
171                                 break;
172                         scsi_add_timer(scmd, scmd->timeout_per_command,
173                                        scsi_times_out);
174                         return;
175                 case EH_NOT_HANDLED:
176                         break;
177                 }
178
179         if (unlikely(!scsi_eh_scmd_add(scmd, SCSI_EH_CANCEL_CMD))) {
180                 panic("Error handler thread not present at %p %p %s %d",
181                       scmd, scmd->device->host, __FILE__, __LINE__);
182         }
183 }
184
185 /**
186  * scsi_block_when_processing_errors - Prevent cmds from being queued.
187  * @sdev:       Device on which we are performing recovery.
188  *
189  * Description:
190  *     We block until the host is out of error recovery, and then check to
191  *     see whether the host or the device is offline.
192  *
193  * Return value:
194  *     0 when dev was taken offline by error recovery. 1 OK to proceed.
195  **/
196 int scsi_block_when_processing_errors(struct scsi_device *sdev)
197 {
198         int online;
199
200         wait_event(sdev->host->host_wait, (sdev->host->shost_state !=
201                                            SHOST_RECOVERY));
202
203         online = scsi_device_online(sdev);
204
205         SCSI_LOG_ERROR_RECOVERY(5, printk("%s: rtn: %d\n", __FUNCTION__,
206                                           online));
207
208         return online;
209 }
210 EXPORT_SYMBOL(scsi_block_when_processing_errors);
211
212 #ifdef CONFIG_SCSI_LOGGING
213 /**
214  * scsi_eh_prt_fail_stats - Log info on failures.
215  * @shost:      scsi host being recovered.
216  * @work_q:     Queue of scsi cmds to process.
217  **/
218 static inline void scsi_eh_prt_fail_stats(struct Scsi_Host *shost,
219                                           struct list_head *work_q)
220 {
221         struct scsi_cmnd *scmd;
222         struct scsi_device *sdev;
223         int total_failures = 0;
224         int cmd_failed = 0;
225         int cmd_cancel = 0;
226         int devices_failed = 0;
227
228         shost_for_each_device(sdev, shost) {
229                 list_for_each_entry(scmd, work_q, eh_entry) {
230                         if (scmd->device == sdev) {
231                                 ++total_failures;
232                                 if (scmd->eh_eflags & SCSI_EH_CANCEL_CMD)
233                                         ++cmd_cancel;
234                                 else 
235                                         ++cmd_failed;
236                         }
237                 }
238
239                 if (cmd_cancel || cmd_failed) {
240                         SCSI_LOG_ERROR_RECOVERY(3,
241                                 printk("%s: %d:%d:%d:%d cmds failed: %d,"
242                                        " cancel: %d\n",
243                                        __FUNCTION__, shost->host_no,
244                                        sdev->channel, sdev->id, sdev->lun,
245                                        cmd_failed, cmd_cancel));
246                         cmd_cancel = 0;
247                         cmd_failed = 0;
248                         ++devices_failed;
249                 }
250         }
251
252         SCSI_LOG_ERROR_RECOVERY(2, printk("Total of %d commands on %d"
253                                           " devices require eh work\n",
254                                   total_failures, devices_failed));
255 }
256 #endif
257
258 /**
259  * scsi_check_sense - Examine scsi cmd sense
260  * @scmd:       Cmd to have sense checked.
261  *
262  * Return value:
263  *      SUCCESS or FAILED or NEEDS_RETRY
264  *
265  * Notes:
266  *      When a deferred error is detected the current command has
267  *      not been executed and needs retrying.
268  **/
269 static int scsi_check_sense(struct scsi_cmnd *scmd)
270 {
271         struct scsi_sense_hdr sshdr;
272
273         if (! scsi_command_normalize_sense(scmd, &sshdr))
274                 return FAILED;  /* no valid sense data */
275
276         if (scsi_sense_is_deferred(&sshdr))
277                 return NEEDS_RETRY;
278
279         /*
280          * Previous logic looked for FILEMARK, EOM or ILI which are
281          * mainly associated with tapes and returned SUCCESS.
282          */
283         if (sshdr.response_code == 0x70) {
284                 /* fixed format */
285                 if (scmd->sense_buffer[2] & 0xe0)
286                         return SUCCESS;
287         } else {
288                 /*
289                  * descriptor format: look for "stream commands sense data
290                  * descriptor" (see SSC-3). Assume single sense data
291                  * descriptor. Ignore ILI from SBC-2 READ LONG and WRITE LONG.
292                  */
293                 if ((sshdr.additional_length > 3) &&
294                     (scmd->sense_buffer[8] == 0x4) &&
295                     (scmd->sense_buffer[11] & 0xe0))
296                         return SUCCESS;
297         }
298
299         switch (sshdr.sense_key) {
300         case NO_SENSE:
301                 return SUCCESS;
302         case RECOVERED_ERROR:
303                 return /* soft_error */ SUCCESS;
304
305         case ABORTED_COMMAND:
306                 return NEEDS_RETRY;
307         case NOT_READY:
308         case UNIT_ATTENTION:
309                 /*
310                  * if we are expecting a cc/ua because of a bus reset that we
311                  * performed, treat this just as a retry.  otherwise this is
312                  * information that we should pass up to the upper-level driver
313                  * so that we can deal with it there.
314                  */
315                 if (scmd->device->expecting_cc_ua) {
316                         scmd->device->expecting_cc_ua = 0;
317                         return NEEDS_RETRY;
318                 }
319                 /*
320                  * if the device is in the process of becoming ready, we 
321                  * should retry.
322                  */
323                 if ((sshdr.asc == 0x04) && (sshdr.ascq == 0x01))
324                         return NEEDS_RETRY;
325                 /*
326                  * if the device is not started, we need to wake
327                  * the error handler to start the motor
328                  */
329                 if (scmd->device->allow_restart &&
330                     (sshdr.asc == 0x04) && (sshdr.ascq == 0x02))
331                         return FAILED;
332                 return SUCCESS;
333
334                 /* these three are not supported */
335         case COPY_ABORTED:
336         case VOLUME_OVERFLOW:
337         case MISCOMPARE:
338                 return SUCCESS;
339
340         case MEDIUM_ERROR:
341                 return NEEDS_RETRY;
342
343         case HARDWARE_ERROR:
344                 if (scmd->device->retry_hwerror)
345                         return NEEDS_RETRY;
346                 else
347                         return SUCCESS;
348
349         case ILLEGAL_REQUEST:
350         case BLANK_CHECK:
351         case DATA_PROTECT:
352         default:
353                 return SUCCESS;
354         }
355 }
356
357 /**
358  * scsi_eh_completed_normally - Disposition a eh cmd on return from LLD.
359  * @scmd:       SCSI cmd to examine.
360  *
361  * Notes:
362  *    This is *only* called when we are examining the status of commands
363  *    queued during error recovery.  the main difference here is that we
364  *    don't allow for the possibility of retries here, and we are a lot
365  *    more restrictive about what we consider acceptable.
366  **/
367 static int scsi_eh_completed_normally(struct scsi_cmnd *scmd)
368 {
369         /*
370          * first check the host byte, to see if there is anything in there
371          * that would indicate what we need to do.
372          */
373         if (host_byte(scmd->result) == DID_RESET) {
374                 /*
375                  * rats.  we are already in the error handler, so we now
376                  * get to try and figure out what to do next.  if the sense
377                  * is valid, we have a pretty good idea of what to do.
378                  * if not, we mark it as FAILED.
379                  */
380                 return scsi_check_sense(scmd);
381         }
382         if (host_byte(scmd->result) != DID_OK)
383                 return FAILED;
384
385         /*
386          * next, check the message byte.
387          */
388         if (msg_byte(scmd->result) != COMMAND_COMPLETE)
389                 return FAILED;
390
391         /*
392          * now, check the status byte to see if this indicates
393          * anything special.
394          */
395         switch (status_byte(scmd->result)) {
396         case GOOD:
397         case COMMAND_TERMINATED:
398                 return SUCCESS;
399         case CHECK_CONDITION:
400                 return scsi_check_sense(scmd);
401         case CONDITION_GOOD:
402         case INTERMEDIATE_GOOD:
403         case INTERMEDIATE_C_GOOD:
404                 /*
405                  * who knows?  FIXME(eric)
406                  */
407                 return SUCCESS;
408         case BUSY:
409         case QUEUE_FULL:
410         case RESERVATION_CONFLICT:
411         default:
412                 return FAILED;
413         }
414         return FAILED;
415 }
416
417 /**
418  * scsi_eh_times_out - timeout function for error handling.
419  * @scmd:       Cmd that is timing out.
420  *
421  * Notes:
422  *    During error handling, the kernel thread will be sleeping waiting
423  *    for some action to complete on the device.  our only job is to
424  *    record that it timed out, and to wake up the thread.
425  **/
426 static void scsi_eh_times_out(struct scsi_cmnd *scmd)
427 {
428         scmd->eh_eflags |= SCSI_EH_REC_TIMEOUT;
429         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: scmd:%p\n", __FUNCTION__,
430                                           scmd));
431
432         up(scmd->device->host->eh_action);
433 }
434
435 /**
436  * scsi_eh_done - Completion function for error handling.
437  * @scmd:       Cmd that is done.
438  **/
439 static void scsi_eh_done(struct scsi_cmnd *scmd)
440 {
441         /*
442          * if the timeout handler is already running, then just set the
443          * flag which says we finished late, and return.  we have no
444          * way of stopping the timeout handler from running, so we must
445          * always defer to it.
446          */
447         if (del_timer(&scmd->eh_timeout)) {
448                 scmd->request->rq_status = RQ_SCSI_DONE;
449
450                 SCSI_LOG_ERROR_RECOVERY(3, printk("%s scmd: %p result: %x\n",
451                                            __FUNCTION__, scmd, scmd->result));
452
453                 up(scmd->device->host->eh_action);
454         }
455 }
456
457 /**
458  * scsi_send_eh_cmnd  - send a cmd to a device as part of error recovery.
459  * @scmd:       SCSI Cmd to send.
460  * @timeout:    Timeout for cmd.
461  *
462  * Notes:
463  *    The initialization of the structures is quite a bit different in
464  *    this case, and furthermore, there is a different completion handler
465  *    vs scsi_dispatch_cmd.
466  * Return value:
467  *    SUCCESS or FAILED or NEEDS_RETRY
468  **/
469 static int scsi_send_eh_cmnd(struct scsi_cmnd *scmd, int timeout)
470 {
471         struct scsi_device *sdev = scmd->device;
472         struct Scsi_Host *shost = sdev->host;
473         DECLARE_MUTEX_LOCKED(sem);
474         unsigned long flags;
475         int rtn = SUCCESS;
476
477         /*
478          * we will use a queued command if possible, otherwise we will
479          * emulate the queuing and calling of completion function ourselves.
480          */
481         if (sdev->scsi_level <= SCSI_2)
482                 scmd->cmnd[1] = (scmd->cmnd[1] & 0x1f) |
483                         (sdev->lun << 5 & 0xe0);
484
485         scsi_add_timer(scmd, timeout, scsi_eh_times_out);
486
487         /*
488          * set up the semaphore so we wait for the command to complete.
489          */
490         shost->eh_action = &sem;
491         scmd->request->rq_status = RQ_SCSI_BUSY;
492
493         spin_lock_irqsave(shost->host_lock, flags);
494         scsi_log_send(scmd);
495         shost->hostt->queuecommand(scmd, scsi_eh_done);
496         spin_unlock_irqrestore(shost->host_lock, flags);
497
498         down(&sem);
499         scsi_log_completion(scmd, SUCCESS);
500
501         shost->eh_action = NULL;
502
503         /*
504          * see if timeout.  if so, tell the host to forget about it.
505          * in other words, we don't want a callback any more.
506          */
507         if (scmd->eh_eflags & SCSI_EH_REC_TIMEOUT) {
508                 scmd->eh_eflags &= ~SCSI_EH_REC_TIMEOUT;
509
510                 /*
511                  * as far as the low level driver is
512                  * concerned, this command is still active, so
513                  * we must give the low level driver a chance
514                  * to abort it. (db) 
515                  *
516                  * FIXME(eric) - we are not tracking whether we could
517                  * abort a timed out command or not.  not sure how
518                  * we should treat them differently anyways.
519                  */
520                 if (shost->hostt->eh_abort_handler)
521                         shost->hostt->eh_abort_handler(scmd);
522                         
523                 scmd->request->rq_status = RQ_SCSI_DONE;
524                 rtn = FAILED;
525         }
526
527         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: scmd: %p, rtn:%x\n",
528                                           __FUNCTION__, scmd, rtn));
529
530         /*
531          * now examine the actual status codes to see whether the command
532          * actually did complete normally.
533          */
534         if (rtn == SUCCESS) {
535                 rtn = scsi_eh_completed_normally(scmd);
536                 SCSI_LOG_ERROR_RECOVERY(3,
537                         printk("%s: scsi_eh_completed_normally %x\n",
538                                __FUNCTION__, rtn));
539                 switch (rtn) {
540                 case SUCCESS:
541                 case NEEDS_RETRY:
542                 case FAILED:
543                         break;
544                 default:
545                         rtn = FAILED;
546                         break;
547                 }
548         }
549
550         return rtn;
551 }
552
553 /**
554  * scsi_request_sense - Request sense data from a particular target.
555  * @scmd:       SCSI cmd for request sense.
556  *
557  * Notes:
558  *    Some hosts automatically obtain this information, others require
559  *    that we obtain it on our own. This function will *not* return until
560  *    the command either times out, or it completes.
561  **/
562 static int scsi_request_sense(struct scsi_cmnd *scmd)
563 {
564         static unsigned char generic_sense[6] =
565         {REQUEST_SENSE, 0, 0, 0, 252, 0};
566         unsigned char *scsi_result;
567         int saved_result;
568         int rtn;
569
570         memcpy(scmd->cmnd, generic_sense, sizeof(generic_sense));
571
572         scsi_result = kmalloc(252, GFP_ATOMIC | ((scmd->device->host->hostt->unchecked_isa_dma) ? __GFP_DMA : 0));
573
574
575         if (unlikely(!scsi_result)) {
576                 printk(KERN_ERR "%s: cannot allocate scsi_result.\n",
577                        __FUNCTION__);
578                 return FAILED;
579         }
580
581         /*
582          * zero the sense buffer.  some host adapters automatically always
583          * request sense, so it is not a good idea that
584          * scmd->request_buffer and scmd->sense_buffer point to the same
585          * address (db).  0 is not a valid sense code. 
586          */
587         memset(scmd->sense_buffer, 0, sizeof(scmd->sense_buffer));
588         memset(scsi_result, 0, 252);
589
590         saved_result = scmd->result;
591         scmd->request_buffer = scsi_result;
592         scmd->request_bufflen = 252;
593         scmd->use_sg = 0;
594         scmd->cmd_len = COMMAND_SIZE(scmd->cmnd[0]);
595         scmd->sc_data_direction = DMA_FROM_DEVICE;
596         scmd->underflow = 0;
597
598         rtn = scsi_send_eh_cmnd(scmd, SENSE_TIMEOUT);
599
600         /* last chance to have valid sense data */
601         if(!SCSI_SENSE_VALID(scmd)) {
602                 memcpy(scmd->sense_buffer, scmd->request_buffer,
603                        sizeof(scmd->sense_buffer));
604         }
605
606         kfree(scsi_result);
607
608         /*
609          * when we eventually call scsi_finish, we really wish to complete
610          * the original request, so let's restore the original data. (db)
611          */
612         scsi_setup_cmd_retry(scmd);
613         scmd->result = saved_result;
614         return rtn;
615 }
616
617 /**
618  * scsi_eh_finish_cmd - Handle a cmd that eh is finished with.
619  * @scmd:       Original SCSI cmd that eh has finished.
620  * @done_q:     Queue for processed commands.
621  *
622  * Notes:
623  *    We don't want to use the normal command completion while we are are
624  *    still handling errors - it may cause other commands to be queued,
625  *    and that would disturb what we are doing.  thus we really want to
626  *    keep a list of pending commands for final completion, and once we
627  *    are ready to leave error handling we handle completion for real.
628  **/
629 static void scsi_eh_finish_cmd(struct scsi_cmnd *scmd,
630                                struct list_head *done_q)
631 {
632         scmd->device->host->host_failed--;
633         scmd->eh_eflags = 0;
634
635         /*
636          * set this back so that the upper level can correctly free up
637          * things.
638          */
639         scsi_setup_cmd_retry(scmd);
640         list_move_tail(&scmd->eh_entry, done_q);
641 }
642
643 /**
644  * scsi_eh_get_sense - Get device sense data.
645  * @work_q:     Queue of commands to process.
646  * @done_q:     Queue of proccessed commands..
647  *
648  * Description:
649  *    See if we need to request sense information.  if so, then get it
650  *    now, so we have a better idea of what to do.  
651  *
652  * Notes:
653  *    This has the unfortunate side effect that if a shost adapter does
654  *    not automatically request sense information, that we end up shutting
655  *    it down before we request it.
656  *
657  *    All drivers should request sense information internally these days,
658  *    so for now all I have to say is tough noogies if you end up in here.
659  *
660  *    XXX: Long term this code should go away, but that needs an audit of
661  *         all LLDDs first.
662  **/
663 static int scsi_eh_get_sense(struct list_head *work_q,
664                              struct list_head *done_q)
665 {
666         struct scsi_cmnd *scmd, *next;
667         int rtn;
668
669         list_for_each_entry_safe(scmd, next, work_q, eh_entry) {
670                 if ((scmd->eh_eflags & SCSI_EH_CANCEL_CMD) ||
671                     SCSI_SENSE_VALID(scmd))
672                         continue;
673
674                 SCSI_LOG_ERROR_RECOVERY(2, printk("%s: requesting sense"
675                                                   " for id: %d\n",
676                                                   current->comm,
677                                                   scmd->device->id));
678                 rtn = scsi_request_sense(scmd);
679                 if (rtn != SUCCESS)
680                         continue;
681
682                 SCSI_LOG_ERROR_RECOVERY(3, printk("sense requested for %p"
683                                                   " result %x\n", scmd,
684                                                   scmd->result));
685                 SCSI_LOG_ERROR_RECOVERY(3, scsi_print_sense("bh", scmd));
686
687                 rtn = scsi_decide_disposition(scmd);
688
689                 /*
690                  * if the result was normal, then just pass it along to the
691                  * upper level.
692                  */
693                 if (rtn == SUCCESS)
694                         /* we don't want this command reissued, just
695                          * finished with the sense data, so set
696                          * retries to the max allowed to ensure it
697                          * won't get reissued */
698                         scmd->retries = scmd->allowed;
699                 else if (rtn != NEEDS_RETRY)
700                         continue;
701
702                 scsi_eh_finish_cmd(scmd, done_q);
703         }
704
705         return list_empty(work_q);
706 }
707
708 /**
709  * scsi_try_to_abort_cmd - Ask host to abort a running command.
710  * @scmd:       SCSI cmd to abort from Lower Level.
711  *
712  * Notes:
713  *    This function will not return until the user's completion function
714  *    has been called.  there is no timeout on this operation.  if the
715  *    author of the low-level driver wishes this operation to be timed,
716  *    they can provide this facility themselves.  helper functions in
717  *    scsi_error.c can be supplied to make this easier to do.
718  **/
719 static int scsi_try_to_abort_cmd(struct scsi_cmnd *scmd)
720 {
721         if (!scmd->device->host->hostt->eh_abort_handler)
722                 return FAILED;
723
724         /*
725          * scsi_done was called just after the command timed out and before
726          * we had a chance to process it. (db)
727          */
728         if (scmd->serial_number == 0)
729                 return SUCCESS;
730         return scmd->device->host->hostt->eh_abort_handler(scmd);
731 }
732
733 /**
734  * scsi_eh_tur - Send TUR to device.
735  * @scmd:       Scsi cmd to send TUR
736  *
737  * Return value:
738  *    0 - Device is ready. 1 - Device NOT ready.
739  **/
740 static int scsi_eh_tur(struct scsi_cmnd *scmd)
741 {
742         static unsigned char tur_command[6] = {TEST_UNIT_READY, 0, 0, 0, 0, 0};
743         int retry_cnt = 1, rtn;
744         int saved_result;
745
746 retry_tur:
747         memcpy(scmd->cmnd, tur_command, sizeof(tur_command));
748
749         /*
750          * zero the sense buffer.  the scsi spec mandates that any
751          * untransferred sense data should be interpreted as being zero.
752          */
753         memset(scmd->sense_buffer, 0, sizeof(scmd->sense_buffer));
754
755         saved_result = scmd->result;
756         scmd->request_buffer = NULL;
757         scmd->request_bufflen = 0;
758         scmd->use_sg = 0;
759         scmd->cmd_len = COMMAND_SIZE(scmd->cmnd[0]);
760         scmd->underflow = 0;
761         scmd->sc_data_direction = DMA_NONE;
762
763         rtn = scsi_send_eh_cmnd(scmd, SENSE_TIMEOUT);
764
765         /*
766          * when we eventually call scsi_finish, we really wish to complete
767          * the original request, so let's restore the original data. (db)
768          */
769         scsi_setup_cmd_retry(scmd);
770         scmd->result = saved_result;
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 scsi_cmnd *scmd, *next;
801         int rtn;
802
803         list_for_each_entry_safe(scmd, next, work_q, eh_entry) {
804                 if (!(scmd->eh_eflags & SCSI_EH_CANCEL_CMD))
805                         continue;
806                 SCSI_LOG_ERROR_RECOVERY(3, printk("%s: aborting cmd:"
807                                                   "0x%p\n", current->comm,
808                                                   scmd));
809                 rtn = scsi_try_to_abort_cmd(scmd);
810                 if (rtn == SUCCESS) {
811                         scmd->eh_eflags &= ~SCSI_EH_CANCEL_CMD;
812                         if (!scsi_device_online(scmd->device) ||
813                             !scsi_eh_tur(scmd)) {
814                                 scsi_eh_finish_cmd(scmd, done_q);
815                         }
816                                 
817                 } else
818                         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: aborting"
819                                                           " cmd failed:"
820                                                           "0x%p\n",
821                                                           current->comm,
822                                                           scmd));
823         }
824
825         return list_empty(work_q);
826 }
827
828 /**
829  * scsi_try_bus_device_reset - Ask host to perform a BDR on a dev
830  * @scmd:       SCSI cmd used to send BDR       
831  *
832  * Notes:
833  *    There is no timeout for this operation.  if this operation is
834  *    unreliable for a given host, then the host itself needs to put a
835  *    timer on it, and set the host back to a consistent state prior to
836  *    returning.
837  **/
838 static int scsi_try_bus_device_reset(struct scsi_cmnd *scmd)
839 {
840         int rtn;
841
842         if (!scmd->device->host->hostt->eh_device_reset_handler)
843                 return FAILED;
844
845         rtn = scmd->device->host->hostt->eh_device_reset_handler(scmd);
846         if (rtn == SUCCESS) {
847                 scmd->device->was_reset = 1;
848                 scmd->device->expecting_cc_ua = 1;
849         }
850
851         return rtn;
852 }
853
854 /**
855  * scsi_eh_try_stu - Send START_UNIT to device.
856  * @scmd:       Scsi cmd to send START_UNIT
857  *
858  * Return value:
859  *    0 - Device is ready. 1 - Device NOT ready.
860  **/
861 static int scsi_eh_try_stu(struct scsi_cmnd *scmd)
862 {
863         static unsigned char stu_command[6] = {START_STOP, 0, 0, 0, 1, 0};
864         int rtn;
865         int saved_result;
866
867         if (!scmd->device->allow_restart)
868                 return 1;
869
870         memcpy(scmd->cmnd, stu_command, sizeof(stu_command));
871
872         /*
873          * zero the sense buffer.  the scsi spec mandates that any
874          * untransferred sense data should be interpreted as being zero.
875          */
876         memset(scmd->sense_buffer, 0, sizeof(scmd->sense_buffer));
877
878         saved_result = scmd->result;
879         scmd->request_buffer = NULL;
880         scmd->request_bufflen = 0;
881         scmd->use_sg = 0;
882         scmd->cmd_len = COMMAND_SIZE(scmd->cmnd[0]);
883         scmd->underflow = 0;
884         scmd->sc_data_direction = DMA_NONE;
885
886         rtn = scsi_send_eh_cmnd(scmd, START_UNIT_TIMEOUT);
887
888         /*
889          * when we eventually call scsi_finish, we really wish to complete
890          * the original request, so let's restore the original data. (db)
891          */
892         scsi_setup_cmd_retry(scmd);
893         scmd->result = saved_result;
894
895         /*
896          * hey, we are done.  let's look to see what happened.
897          */
898         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: scmd %p rtn %x\n",
899                 __FUNCTION__, scmd, rtn));
900         if (rtn == SUCCESS)
901                 return 0;
902         return 1;
903 }
904
905  /**
906  * scsi_eh_stu - send START_UNIT if needed
907  * @shost:      scsi host being recovered.
908  * @eh_done_q:  list_head for processed commands.
909  *
910  * Notes:
911  *    If commands are failing due to not ready, initializing command required,
912  *      try revalidating the device, which will end up sending a start unit. 
913  **/
914 static int scsi_eh_stu(struct Scsi_Host *shost,
915                               struct list_head *work_q,
916                               struct list_head *done_q)
917 {
918         struct scsi_cmnd *scmd, *stu_scmd, *next;
919         struct scsi_device *sdev;
920
921         shost_for_each_device(sdev, shost) {
922                 stu_scmd = NULL;
923                 list_for_each_entry(scmd, work_q, eh_entry)
924                         if (scmd->device == sdev && SCSI_SENSE_VALID(scmd) &&
925                             scsi_check_sense(scmd) == FAILED ) {
926                                 stu_scmd = scmd;
927                                 break;
928                         }
929
930                 if (!stu_scmd)
931                         continue;
932
933                 SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Sending START_UNIT to sdev:"
934                                                   " 0x%p\n", current->comm, sdev));
935
936                 if (!scsi_eh_try_stu(stu_scmd)) {
937                         if (!scsi_device_online(sdev) ||
938                             !scsi_eh_tur(stu_scmd)) {
939                                 list_for_each_entry_safe(scmd, next,
940                                                           work_q, eh_entry) {
941                                         if (scmd->device == sdev)
942                                                 scsi_eh_finish_cmd(scmd, done_q);
943                                 }
944                         }
945                 } else {
946                         SCSI_LOG_ERROR_RECOVERY(3,
947                                                 printk("%s: START_UNIT failed to sdev:"
948                                                        " 0x%p\n", current->comm, sdev));
949                 }
950         }
951
952         return list_empty(work_q);
953 }
954
955
956 /**
957  * scsi_eh_bus_device_reset - send bdr if needed
958  * @shost:      scsi host being recovered.
959  * @eh_done_q:  list_head for processed commands.
960  *
961  * Notes:
962  *    Try a bus device reset.  still, look to see whether we have multiple
963  *    devices that are jammed or not - if we have multiple devices, it
964  *    makes no sense to try bus_device_reset - we really would need to try
965  *    a bus_reset instead. 
966  **/
967 static int scsi_eh_bus_device_reset(struct Scsi_Host *shost,
968                                     struct list_head *work_q,
969                                     struct list_head *done_q)
970 {
971         struct scsi_cmnd *scmd, *bdr_scmd, *next;
972         struct scsi_device *sdev;
973         int rtn;
974
975         shost_for_each_device(sdev, shost) {
976                 bdr_scmd = NULL;
977                 list_for_each_entry(scmd, work_q, eh_entry)
978                         if (scmd->device == sdev) {
979                                 bdr_scmd = scmd;
980                                 break;
981                         }
982
983                 if (!bdr_scmd)
984                         continue;
985
986                 SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Sending BDR sdev:"
987                                                   " 0x%p\n", current->comm,
988                                                   sdev));
989                 rtn = scsi_try_bus_device_reset(bdr_scmd);
990                 if (rtn == SUCCESS) {
991                         if (!scsi_device_online(sdev) ||
992                             !scsi_eh_tur(bdr_scmd)) {
993                                 list_for_each_entry_safe(scmd, next,
994                                                          work_q, eh_entry) {
995                                         if (scmd->device == sdev)
996                                                 scsi_eh_finish_cmd(scmd,
997                                                                    done_q);
998                                 }
999                         }
1000                 } else {
1001                         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: BDR"
1002                                                           " failed sdev:"
1003                                                           "0x%p\n",
1004                                                           current->comm,
1005                                                            sdev));
1006                 }
1007         }
1008
1009         return list_empty(work_q);
1010 }
1011
1012 /**
1013  * scsi_try_bus_reset - ask host to perform a bus reset
1014  * @scmd:       SCSI cmd to send bus reset.
1015  **/
1016 static int scsi_try_bus_reset(struct scsi_cmnd *scmd)
1017 {
1018         unsigned long flags;
1019         int rtn;
1020
1021         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Snd Bus RST\n",
1022                                           __FUNCTION__));
1023
1024         if (!scmd->device->host->hostt->eh_bus_reset_handler)
1025                 return FAILED;
1026
1027         rtn = scmd->device->host->hostt->eh_bus_reset_handler(scmd);
1028
1029         if (rtn == SUCCESS) {
1030                 if (!scmd->device->host->hostt->skip_settle_delay)
1031                         ssleep(BUS_RESET_SETTLE_TIME);
1032                 spin_lock_irqsave(scmd->device->host->host_lock, flags);
1033                 scsi_report_bus_reset(scmd->device->host, scmd->device->channel);
1034                 spin_unlock_irqrestore(scmd->device->host->host_lock, flags);
1035         }
1036
1037         return rtn;
1038 }
1039
1040 /**
1041  * scsi_try_host_reset - ask host adapter to reset itself
1042  * @scmd:       SCSI cmd to send hsot reset.
1043  **/
1044 static int scsi_try_host_reset(struct scsi_cmnd *scmd)
1045 {
1046         unsigned long flags;
1047         int rtn;
1048
1049         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Snd Host RST\n",
1050                                           __FUNCTION__));
1051
1052         if (!scmd->device->host->hostt->eh_host_reset_handler)
1053                 return FAILED;
1054
1055         rtn = scmd->device->host->hostt->eh_host_reset_handler(scmd);
1056
1057         if (rtn == SUCCESS) {
1058                 if (!scmd->device->host->hostt->skip_settle_delay)
1059                         ssleep(HOST_RESET_SETTLE_TIME);
1060                 spin_lock_irqsave(scmd->device->host->host_lock, flags);
1061                 scsi_report_bus_reset(scmd->device->host, scmd->device->channel);
1062                 spin_unlock_irqrestore(scmd->device->host->host_lock, flags);
1063         }
1064
1065         return rtn;
1066 }
1067
1068 /**
1069  * scsi_eh_bus_reset - send a bus reset 
1070  * @shost:      scsi host being recovered.
1071  * @eh_done_q:  list_head for processed commands.
1072  **/
1073 static int scsi_eh_bus_reset(struct Scsi_Host *shost,
1074                              struct list_head *work_q,
1075                              struct list_head *done_q)
1076 {
1077         struct scsi_cmnd *scmd, *chan_scmd, *next;
1078         unsigned int channel;
1079         int rtn;
1080
1081         /*
1082          * we really want to loop over the various channels, and do this on
1083          * a channel by channel basis.  we should also check to see if any
1084          * of the failed commands are on soft_reset devices, and if so, skip
1085          * the reset.  
1086          */
1087
1088         for (channel = 0; channel <= shost->max_channel; channel++) {
1089                 chan_scmd = NULL;
1090                 list_for_each_entry(scmd, work_q, eh_entry) {
1091                         if (channel == scmd->device->channel) {
1092                                 chan_scmd = scmd;
1093                                 break;
1094                                 /*
1095                                  * FIXME add back in some support for
1096                                  * soft_reset devices.
1097                                  */
1098                         }
1099                 }
1100
1101                 if (!chan_scmd)
1102                         continue;
1103                 SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Sending BRST chan:"
1104                                                   " %d\n", current->comm,
1105                                                   channel));
1106                 rtn = scsi_try_bus_reset(chan_scmd);
1107                 if (rtn == SUCCESS) {
1108                         list_for_each_entry_safe(scmd, next, work_q, eh_entry) {
1109                                 if (channel == scmd->device->channel)
1110                                         if (!scsi_device_online(scmd->device) ||
1111                                             !scsi_eh_tur(scmd))
1112                                                 scsi_eh_finish_cmd(scmd,
1113                                                                    done_q);
1114                         }
1115                 } else {
1116                         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: BRST"
1117                                                           " failed chan: %d\n",
1118                                                           current->comm,
1119                                                           channel));
1120                 }
1121         }
1122         return list_empty(work_q);
1123 }
1124
1125 /**
1126  * scsi_eh_host_reset - send a host reset 
1127  * @work_q:     list_head for processed commands.
1128  * @done_q:     list_head for processed commands.
1129  **/
1130 static int scsi_eh_host_reset(struct list_head *work_q,
1131                               struct list_head *done_q)
1132 {
1133         struct scsi_cmnd *scmd, *next;
1134         int rtn;
1135
1136         if (!list_empty(work_q)) {
1137                 scmd = list_entry(work_q->next,
1138                                   struct scsi_cmnd, eh_entry);
1139
1140                 SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Sending HRST\n"
1141                                                   , current->comm));
1142
1143                 rtn = scsi_try_host_reset(scmd);
1144                 if (rtn == SUCCESS) {
1145                         list_for_each_entry_safe(scmd, next, work_q, eh_entry) {
1146                                 if (!scsi_device_online(scmd->device) ||
1147                                     (!scsi_eh_try_stu(scmd) && !scsi_eh_tur(scmd)) ||
1148                                     !scsi_eh_tur(scmd))
1149                                         scsi_eh_finish_cmd(scmd, done_q);
1150                         }
1151                 } else {
1152                         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: HRST"
1153                                                           " failed\n",
1154                                                           current->comm));
1155                 }
1156         }
1157         return list_empty(work_q);
1158 }
1159
1160 /**
1161  * scsi_eh_offline_sdevs - offline scsi devices that fail to recover
1162  * @work_q:     list_head for processed commands.
1163  * @done_q:     list_head for processed commands.
1164  *
1165  **/
1166 static void scsi_eh_offline_sdevs(struct list_head *work_q,
1167                                   struct list_head *done_q)
1168 {
1169         struct scsi_cmnd *scmd, *next;
1170
1171         list_for_each_entry_safe(scmd, next, work_q, eh_entry) {
1172                 printk(KERN_INFO "scsi: Device offlined - not"
1173                                 " ready after error recovery: host"
1174                                 " %d channel %d id %d lun %d\n",
1175                                 scmd->device->host->host_no,
1176                                 scmd->device->channel,
1177                                 scmd->device->id,
1178                                 scmd->device->lun);
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                 printk(KERN_INFO "scsi: reservation conflict: host"
1341                                 " %d channel %d id %d lun %d\n",
1342                        scmd->device->host->host_no, scmd->device->channel,
1343                        scmd->device->id, scmd->device->lun);
1344                 return SUCCESS; /* causes immediate i/o error */
1345         default:
1346                 return FAILED;
1347         }
1348         return FAILED;
1349
1350       maybe_retry:
1351
1352         /* we requeue for retry because the error was retryable, and
1353          * the request was not marked fast fail.  Note that above,
1354          * even if the request is marked fast fail, we still requeue
1355          * for queue congestion conditions (QUEUE_FULL or BUSY) */
1356         if ((++scmd->retries) < scmd->allowed 
1357             && !blk_noretry_request(scmd->request)) {
1358                 return NEEDS_RETRY;
1359         } else {
1360                 /*
1361                  * no more retries - report this one back to upper level.
1362                  */
1363                 return SUCCESS;
1364         }
1365 }
1366
1367 /**
1368  * scsi_eh_lock_done - done function for eh door lock request
1369  * @scmd:       SCSI command block for the door lock request
1370  *
1371  * Notes:
1372  *      We completed the asynchronous door lock request, and it has either
1373  *      locked the door or failed.  We must free the command structures
1374  *      associated with this request.
1375  **/
1376 static void scsi_eh_lock_done(struct scsi_cmnd *scmd)
1377 {
1378         struct scsi_request *sreq = scmd->sc_request;
1379
1380         scsi_release_request(sreq);
1381 }
1382
1383
1384 /**
1385  * scsi_eh_lock_door - Prevent medium removal for the specified device
1386  * @sdev:       SCSI device to prevent medium removal
1387  *
1388  * Locking:
1389  *      We must be called from process context; scsi_allocate_request()
1390  *      may sleep.
1391  *
1392  * Notes:
1393  *      We queue up an asynchronous "ALLOW MEDIUM REMOVAL" request on the
1394  *      head of the devices request queue, and continue.
1395  *
1396  * Bugs:
1397  *      scsi_allocate_request() may sleep waiting for existing requests to
1398  *      be processed.  However, since we haven't kicked off any request
1399  *      processing for this host, this may deadlock.
1400  *
1401  *      If scsi_allocate_request() fails for what ever reason, we
1402  *      completely forget to lock the door.
1403  **/
1404 static void scsi_eh_lock_door(struct scsi_device *sdev)
1405 {
1406         struct scsi_request *sreq = scsi_allocate_request(sdev, GFP_KERNEL);
1407
1408         if (unlikely(!sreq)) {
1409                 printk(KERN_ERR "%s: request allocate failed,"
1410                        "prevent media removal cmd not sent\n", __FUNCTION__);
1411                 return;
1412         }
1413
1414         sreq->sr_cmnd[0] = ALLOW_MEDIUM_REMOVAL;
1415         sreq->sr_cmnd[1] = 0;
1416         sreq->sr_cmnd[2] = 0;
1417         sreq->sr_cmnd[3] = 0;
1418         sreq->sr_cmnd[4] = SCSI_REMOVAL_PREVENT;
1419         sreq->sr_cmnd[5] = 0;
1420         sreq->sr_data_direction = DMA_NONE;
1421         sreq->sr_bufflen = 0;
1422         sreq->sr_buffer = NULL;
1423         sreq->sr_allowed = 5;
1424         sreq->sr_done = scsi_eh_lock_done;
1425         sreq->sr_timeout_per_command = 10 * HZ;
1426         sreq->sr_cmd_len = COMMAND_SIZE(sreq->sr_cmnd[0]);
1427
1428         scsi_insert_special_req(sreq, 1);
1429 }
1430
1431
1432 /**
1433  * scsi_restart_operations - restart io operations to the specified host.
1434  * @shost:      Host we are restarting.
1435  *
1436  * Notes:
1437  *    When we entered the error handler, we blocked all further i/o to
1438  *    this device.  we need to 'reverse' this process.
1439  **/
1440 static void scsi_restart_operations(struct Scsi_Host *shost)
1441 {
1442         struct scsi_device *sdev;
1443
1444         /*
1445          * If the door was locked, we need to insert a door lock request
1446          * onto the head of the SCSI request queue for the device.  There
1447          * is no point trying to lock the door of an off-line device.
1448          */
1449         shost_for_each_device(sdev, shost) {
1450                 if (scsi_device_online(sdev) && sdev->locked)
1451                         scsi_eh_lock_door(sdev);
1452         }
1453
1454         /*
1455          * next free up anything directly waiting upon the host.  this
1456          * will be requests for character device operations, and also for
1457          * ioctls to queued block devices.
1458          */
1459         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: waking up host to restart\n",
1460                                           __FUNCTION__));
1461
1462         scsi_host_set_state(shost, SHOST_RUNNING);
1463
1464         wake_up(&shost->host_wait);
1465
1466         /*
1467          * finally we need to re-initiate requests that may be pending.  we will
1468          * have had everything blocked while error handling is taking place, and
1469          * now that error recovery is done, we will need to ensure that these
1470          * requests are started.
1471          */
1472         scsi_run_host_queues(shost);
1473 }
1474
1475 /**
1476  * scsi_eh_ready_devs - check device ready state and recover if not.
1477  * @shost:      host to be recovered.
1478  * @eh_done_q:  list_head for processed commands.
1479  *
1480  **/
1481 static void scsi_eh_ready_devs(struct Scsi_Host *shost,
1482                                struct list_head *work_q,
1483                                struct list_head *done_q)
1484 {
1485         if (!scsi_eh_stu(shost, work_q, done_q))
1486                 if (!scsi_eh_bus_device_reset(shost, work_q, done_q))
1487                         if (!scsi_eh_bus_reset(shost, work_q, done_q))
1488                                 if (!scsi_eh_host_reset(work_q, done_q))
1489                                         scsi_eh_offline_sdevs(work_q, done_q);
1490 }
1491
1492 /**
1493  * scsi_eh_flush_done_q - finish processed commands or retry them.
1494  * @done_q:     list_head of processed commands.
1495  *
1496  **/
1497 static void scsi_eh_flush_done_q(struct list_head *done_q)
1498 {
1499         struct scsi_cmnd *scmd, *next;
1500
1501         list_for_each_entry_safe(scmd, next, done_q, eh_entry) {
1502                 list_del_init(&scmd->eh_entry);
1503                 if (scsi_device_online(scmd->device) &&
1504                     !blk_noretry_request(scmd->request) &&
1505                     (++scmd->retries < scmd->allowed)) {
1506                         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: flush"
1507                                                           " retry cmd: %p\n",
1508                                                           current->comm,
1509                                                           scmd));
1510                                 scsi_queue_insert(scmd, SCSI_MLQUEUE_EH_RETRY);
1511                 } else {
1512                         /*
1513                          * If just we got sense for the device (called
1514                          * scsi_eh_get_sense), scmd->result is already
1515                          * set, do not set DRIVER_TIMEOUT.
1516                          */
1517                         if (!scmd->result)
1518                                 scmd->result |= (DRIVER_TIMEOUT << 24);
1519                         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: flush finish"
1520                                                         " cmd: %p\n",
1521                                                         current->comm, scmd));
1522                         scsi_finish_command(scmd);
1523                 }
1524         }
1525 }
1526
1527 /**
1528  * scsi_unjam_host - Attempt to fix a host which has a cmd that failed.
1529  * @shost:      Host to unjam.
1530  *
1531  * Notes:
1532  *    When we come in here, we *know* that all commands on the bus have
1533  *    either completed, failed or timed out.  we also know that no further
1534  *    commands are being sent to the host, so things are relatively quiet
1535  *    and we have freedom to fiddle with things as we wish.
1536  *
1537  *    This is only the *default* implementation.  it is possible for
1538  *    individual drivers to supply their own version of this function, and
1539  *    if the maintainer wishes to do this, it is strongly suggested that
1540  *    this function be taken as a template and modified.  this function
1541  *    was designed to correctly handle problems for about 95% of the
1542  *    different cases out there, and it should always provide at least a
1543  *    reasonable amount of error recovery.
1544  *
1545  *    Any command marked 'failed' or 'timeout' must eventually have
1546  *    scsi_finish_cmd() called for it.  we do all of the retry stuff
1547  *    here, so when we restart the host after we return it should have an
1548  *    empty queue.
1549  **/
1550 static void scsi_unjam_host(struct Scsi_Host *shost)
1551 {
1552         unsigned long flags;
1553         LIST_HEAD(eh_work_q);
1554         LIST_HEAD(eh_done_q);
1555
1556         spin_lock_irqsave(shost->host_lock, flags);
1557         list_splice_init(&shost->eh_cmd_q, &eh_work_q);
1558         spin_unlock_irqrestore(shost->host_lock, flags);
1559
1560         SCSI_LOG_ERROR_RECOVERY(1, scsi_eh_prt_fail_stats(shost, &eh_work_q));
1561
1562         if (!scsi_eh_get_sense(&eh_work_q, &eh_done_q))
1563                 if (!scsi_eh_abort_cmds(&eh_work_q, &eh_done_q))
1564                         scsi_eh_ready_devs(shost, &eh_work_q, &eh_done_q);
1565
1566         scsi_eh_flush_done_q(&eh_done_q);
1567 }
1568
1569 /**
1570  * scsi_error_handler - Handle errors/timeouts of SCSI cmds.
1571  * @data:       Host for which we are running.
1572  *
1573  * Notes:
1574  *    This is always run in the context of a kernel thread.  The idea is
1575  *    that we start this thing up when the kernel starts up (one per host
1576  *    that we detect), and it immediately goes to sleep and waits for some
1577  *    event (i.e. failure).  When this takes place, we have the job of
1578  *    trying to unjam the bus and restarting things.
1579  **/
1580 int scsi_error_handler(void *data)
1581 {
1582         struct Scsi_Host *shost = (struct Scsi_Host *) data;
1583         int rtn;
1584         DECLARE_MUTEX_LOCKED(sem);
1585
1586         /*
1587          *    Flush resources
1588          */
1589
1590         daemonize("scsi_eh_%d", shost->host_no);
1591
1592         current->flags |= PF_NOFREEZE;
1593
1594         shost->eh_wait = &sem;
1595         shost->ehandler = current;
1596
1597         /*
1598          * Wake up the thread that created us.
1599          */
1600         SCSI_LOG_ERROR_RECOVERY(3, printk("Wake up parent of"
1601                                           " scsi_eh_%d\n",shost->host_no));
1602
1603         complete(shost->eh_notify);
1604
1605         while (1) {
1606                 /*
1607                  * If we get a signal, it means we are supposed to go
1608                  * away and die.  This typically happens if the user is
1609                  * trying to unload a module.
1610                  */
1611                 SCSI_LOG_ERROR_RECOVERY(1, printk("Error handler"
1612                                                   " scsi_eh_%d"
1613                                                   " sleeping\n",shost->host_no));
1614
1615                 /*
1616                  * Note - we always use down_interruptible with the semaphore
1617                  * even if the module was loaded as part of the kernel.  The
1618                  * reason is that down() will cause this thread to be counted
1619                  * in the load average as a running process, and down
1620                  * interruptible doesn't.  Given that we need to allow this
1621                  * thread to die if the driver was loaded as a module, using
1622                  * semaphores isn't unreasonable.
1623                  */
1624                 down_interruptible(&sem);
1625                 if (shost->eh_kill)
1626                         break;
1627
1628                 SCSI_LOG_ERROR_RECOVERY(1, printk("Error handler"
1629                                                   " scsi_eh_%d waking"
1630                                                   " up\n",shost->host_no));
1631
1632                 shost->eh_active = 1;
1633
1634                 /*
1635                  * We have a host that is failing for some reason.  Figure out
1636                  * what we need to do to get it up and online again (if we can).
1637                  * If we fail, we end up taking the thing offline.
1638                  */
1639                 if (shost->hostt->eh_strategy_handler) 
1640                         rtn = shost->hostt->eh_strategy_handler(shost);
1641                 else
1642                         scsi_unjam_host(shost);
1643
1644                 shost->eh_active = 0;
1645
1646                 /*
1647                  * Note - if the above fails completely, the action is to take
1648                  * individual devices offline and flush the queue of any
1649                  * outstanding requests that may have been pending.  When we
1650                  * restart, we restart any I/O to any other devices on the bus
1651                  * which are still online.
1652                  */
1653                 scsi_restart_operations(shost);
1654
1655         }
1656
1657         SCSI_LOG_ERROR_RECOVERY(1, printk("Error handler scsi_eh_%d"
1658                                           " exiting\n",shost->host_no));
1659
1660         /*
1661          * Make sure that nobody tries to wake us up again.
1662          */
1663         shost->eh_wait = NULL;
1664
1665         /*
1666          * Knock this down too.  From this point on, the host is flying
1667          * without a pilot.  If this is because the module is being unloaded,
1668          * that's fine.  If the user sent a signal to this thing, we are
1669          * potentially in real danger.
1670          */
1671         shost->eh_active = 0;
1672         shost->ehandler = NULL;
1673
1674         /*
1675          * If anyone is waiting for us to exit (i.e. someone trying to unload
1676          * a driver), then wake up that process to let them know we are on
1677          * the way out the door.
1678          */
1679         complete_and_exit(shost->eh_notify, 0);
1680         return 0;
1681 }
1682
1683 /*
1684  * Function:    scsi_report_bus_reset()
1685  *
1686  * Purpose:     Utility function used by low-level drivers to report that
1687  *              they have observed a bus reset on the bus being handled.
1688  *
1689  * Arguments:   shost       - Host in question
1690  *              channel     - channel on which reset was observed.
1691  *
1692  * Returns:     Nothing
1693  *
1694  * Lock status: Host lock must be held.
1695  *
1696  * Notes:       This only needs to be called if the reset is one which
1697  *              originates from an unknown location.  Resets originated
1698  *              by the mid-level itself don't need to call this, but there
1699  *              should be no harm.
1700  *
1701  *              The main purpose of this is to make sure that a CHECK_CONDITION
1702  *              is properly treated.
1703  */
1704 void scsi_report_bus_reset(struct Scsi_Host *shost, int channel)
1705 {
1706         struct scsi_device *sdev;
1707
1708         __shost_for_each_device(sdev, shost) {
1709                 if (channel == sdev->channel) {
1710                         sdev->was_reset = 1;
1711                         sdev->expecting_cc_ua = 1;
1712                 }
1713         }
1714 }
1715 EXPORT_SYMBOL(scsi_report_bus_reset);
1716
1717 /*
1718  * Function:    scsi_report_device_reset()
1719  *
1720  * Purpose:     Utility function used by low-level drivers to report that
1721  *              they have observed a device reset on the device being handled.
1722  *
1723  * Arguments:   shost       - Host in question
1724  *              channel     - channel on which reset was observed
1725  *              target      - target on which reset was observed
1726  *
1727  * Returns:     Nothing
1728  *
1729  * Lock status: Host lock must be held
1730  *
1731  * Notes:       This only needs to be called if the reset is one which
1732  *              originates from an unknown location.  Resets originated
1733  *              by the mid-level itself don't need to call this, but there
1734  *              should be no harm.
1735  *
1736  *              The main purpose of this is to make sure that a CHECK_CONDITION
1737  *              is properly treated.
1738  */
1739 void scsi_report_device_reset(struct Scsi_Host *shost, int channel, int target)
1740 {
1741         struct scsi_device *sdev;
1742
1743         __shost_for_each_device(sdev, shost) {
1744                 if (channel == sdev->channel &&
1745                     target == sdev->id) {
1746                         sdev->was_reset = 1;
1747                         sdev->expecting_cc_ua = 1;
1748                 }
1749         }
1750 }
1751 EXPORT_SYMBOL(scsi_report_device_reset);
1752
1753 static void
1754 scsi_reset_provider_done_command(struct scsi_cmnd *scmd)
1755 {
1756 }
1757
1758 /*
1759  * Function:    scsi_reset_provider
1760  *
1761  * Purpose:     Send requested reset to a bus or device at any phase.
1762  *
1763  * Arguments:   device  - device to send reset to
1764  *              flag - reset type (see scsi.h)
1765  *
1766  * Returns:     SUCCESS/FAILURE.
1767  *
1768  * Notes:       This is used by the SCSI Generic driver to provide
1769  *              Bus/Device reset capability.
1770  */
1771 int
1772 scsi_reset_provider(struct scsi_device *dev, int flag)
1773 {
1774         struct scsi_cmnd *scmd = scsi_get_command(dev, GFP_KERNEL);
1775         struct request req;
1776         int rtn;
1777
1778         scmd->request = &req;
1779         memset(&scmd->eh_timeout, 0, sizeof(scmd->eh_timeout));
1780         scmd->request->rq_status        = RQ_SCSI_BUSY;
1781
1782         memset(&scmd->cmnd, '\0', sizeof(scmd->cmnd));
1783     
1784         scmd->scsi_done         = scsi_reset_provider_done_command;
1785         scmd->done                      = NULL;
1786         scmd->buffer                    = NULL;
1787         scmd->bufflen                   = 0;
1788         scmd->request_buffer            = NULL;
1789         scmd->request_bufflen           = 0;
1790
1791         scmd->cmd_len                   = 0;
1792
1793         scmd->sc_data_direction         = DMA_BIDIRECTIONAL;
1794         scmd->sc_request                = NULL;
1795         scmd->sc_magic                  = SCSI_CMND_MAGIC;
1796
1797         init_timer(&scmd->eh_timeout);
1798
1799         /*
1800          * Sometimes the command can get back into the timer chain,
1801          * so use the pid as an identifier.
1802          */
1803         scmd->pid                       = 0;
1804
1805         switch (flag) {
1806         case SCSI_TRY_RESET_DEVICE:
1807                 rtn = scsi_try_bus_device_reset(scmd);
1808                 if (rtn == SUCCESS)
1809                         break;
1810                 /* FALLTHROUGH */
1811         case SCSI_TRY_RESET_BUS:
1812                 rtn = scsi_try_bus_reset(scmd);
1813                 if (rtn == SUCCESS)
1814                         break;
1815                 /* FALLTHROUGH */
1816         case SCSI_TRY_RESET_HOST:
1817                 rtn = scsi_try_host_reset(scmd);
1818                 break;
1819         default:
1820                 rtn = FAILED;
1821         }
1822
1823         scsi_next_command(scmd);
1824         return rtn;
1825 }
1826 EXPORT_SYMBOL(scsi_reset_provider);
1827
1828 /**
1829  * scsi_normalize_sense - normalize main elements from either fixed or
1830  *                      descriptor sense data format into a common format.
1831  *
1832  * @sense_buffer:       byte array containing sense data returned by device
1833  * @sb_len:             number of valid bytes in sense_buffer
1834  * @sshdr:              pointer to instance of structure that common
1835  *                      elements are written to.
1836  *
1837  * Notes:
1838  *      The "main elements" from sense data are: response_code, sense_key,
1839  *      asc, ascq and additional_length (only for descriptor format).
1840  *
1841  *      Typically this function can be called after a device has
1842  *      responded to a SCSI command with the CHECK_CONDITION status.
1843  *
1844  * Return value:
1845  *      1 if valid sense data information found, else 0;
1846  **/
1847 int scsi_normalize_sense(const u8 *sense_buffer, int sb_len,
1848                          struct scsi_sense_hdr *sshdr)
1849 {
1850         if (!sense_buffer || !sb_len || (sense_buffer[0] & 0x70) != 0x70)
1851                 return 0;
1852
1853         memset(sshdr, 0, sizeof(struct scsi_sense_hdr));
1854
1855         sshdr->response_code = (sense_buffer[0] & 0x7f);
1856         if (sshdr->response_code >= 0x72) {
1857                 /*
1858                  * descriptor format
1859                  */
1860                 if (sb_len > 1)
1861                         sshdr->sense_key = (sense_buffer[1] & 0xf);
1862                 if (sb_len > 2)
1863                         sshdr->asc = sense_buffer[2];
1864                 if (sb_len > 3)
1865                         sshdr->ascq = sense_buffer[3];
1866                 if (sb_len > 7)
1867                         sshdr->additional_length = sense_buffer[7];
1868         } else {
1869                 /* 
1870                  * fixed format
1871                  */
1872                 if (sb_len > 2)
1873                         sshdr->sense_key = (sense_buffer[2] & 0xf);
1874                 if (sb_len > 7) {
1875                         sb_len = (sb_len < (sense_buffer[7] + 8)) ?
1876                                          sb_len : (sense_buffer[7] + 8);
1877                         if (sb_len > 12)
1878                                 sshdr->asc = sense_buffer[12];
1879                         if (sb_len > 13)
1880                                 sshdr->ascq = sense_buffer[13];
1881                 }
1882         }
1883
1884         return 1;
1885 }
1886 EXPORT_SYMBOL(scsi_normalize_sense);
1887
1888 int scsi_request_normalize_sense(struct scsi_request *sreq,
1889                                  struct scsi_sense_hdr *sshdr)
1890 {
1891         return scsi_normalize_sense(sreq->sr_sense_buffer,
1892                         sizeof(sreq->sr_sense_buffer), sshdr);
1893 }
1894 EXPORT_SYMBOL(scsi_request_normalize_sense);
1895
1896 int scsi_command_normalize_sense(struct scsi_cmnd *cmd,
1897                                  struct scsi_sense_hdr *sshdr)
1898 {
1899         return scsi_normalize_sense(cmd->sense_buffer,
1900                         sizeof(cmd->sense_buffer), sshdr);
1901 }
1902 EXPORT_SYMBOL(scsi_command_normalize_sense);
1903
1904 /**
1905  * scsi_sense_desc_find - search for a given descriptor type in
1906  *                      descriptor sense data format.
1907  *
1908  * @sense_buffer:       byte array of descriptor format sense data
1909  * @sb_len:             number of valid bytes in sense_buffer
1910  * @desc_type:          value of descriptor type to find
1911  *                      (e.g. 0 -> information)
1912  *
1913  * Notes:
1914  *      only valid when sense data is in descriptor format
1915  *
1916  * Return value:
1917  *      pointer to start of (first) descriptor if found else NULL
1918  **/
1919 const u8 * scsi_sense_desc_find(const u8 * sense_buffer, int sb_len,
1920                                 int desc_type)
1921 {
1922         int add_sen_len, add_len, desc_len, k;
1923         const u8 * descp;
1924
1925         if ((sb_len < 8) || (0 == (add_sen_len = sense_buffer[7])))
1926                 return NULL;
1927         if ((sense_buffer[0] < 0x72) || (sense_buffer[0] > 0x73))
1928                 return NULL;
1929         add_sen_len = (add_sen_len < (sb_len - 8)) ?
1930                         add_sen_len : (sb_len - 8);
1931         descp = &sense_buffer[8];
1932         for (desc_len = 0, k = 0; k < add_sen_len; k += desc_len) {
1933                 descp += desc_len;
1934                 add_len = (k < (add_sen_len - 1)) ? descp[1]: -1;
1935                 desc_len = add_len + 2;
1936                 if (descp[0] == desc_type)
1937                         return descp;
1938                 if (add_len < 0) // short descriptor ??
1939                         break;
1940         }
1941         return NULL;
1942 }
1943 EXPORT_SYMBOL(scsi_sense_desc_find);
1944
1945 /**
1946  * scsi_get_sense_info_fld - attempts to get information field from
1947  *                      sense data (either fixed or descriptor format)
1948  *
1949  * @sense_buffer:       byte array of sense data
1950  * @sb_len:             number of valid bytes in sense_buffer
1951  * @info_out:           pointer to 64 integer where 8 or 4 byte information
1952  *                      field will be placed if found.
1953  *
1954  * Return value:
1955  *      1 if information field found, 0 if not found.
1956  **/
1957 int scsi_get_sense_info_fld(const u8 * sense_buffer, int sb_len,
1958                             u64 * info_out)
1959 {
1960         int j;
1961         const u8 * ucp;
1962         u64 ull;
1963
1964         if (sb_len < 7)
1965                 return 0;
1966         switch (sense_buffer[0] & 0x7f) {
1967         case 0x70:
1968         case 0x71:
1969                 if (sense_buffer[0] & 0x80) {
1970                         *info_out = (sense_buffer[3] << 24) +
1971                                     (sense_buffer[4] << 16) +
1972                                     (sense_buffer[5] << 8) + sense_buffer[6];
1973                         return 1;
1974                 } else
1975                         return 0;
1976         case 0x72:
1977         case 0x73:
1978                 ucp = scsi_sense_desc_find(sense_buffer, sb_len,
1979                                            0 /* info desc */);
1980                 if (ucp && (0xa == ucp[1])) {
1981                         ull = 0;
1982                         for (j = 0; j < 8; ++j) {
1983                                 if (j > 0)
1984                                         ull <<= 8;
1985                                 ull |= ucp[4 + j];
1986                         }
1987                         *info_out = ull;
1988                         return 1;
1989                 } else
1990                         return 0;
1991         default:
1992                 return 0;
1993         }
1994 }
1995 EXPORT_SYMBOL(scsi_get_sense_info_fld);