Merge git://git.skbuff.net/gitroot/yoshfuji/linux-2.6-git-rfc3542
[sfrench/cifs-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                 up(shost->eh_wait);
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
72         if (shost->eh_wait == NULL)
73                 return 0;
74
75         spin_lock_irqsave(shost->host_lock, flags);
76
77         scmd->eh_eflags |= eh_flag;
78         list_add_tail(&scmd->eh_entry, &shost->eh_cmd_q);
79         scsi_host_set_state(shost, SHOST_RECOVERY);
80         shost->host_failed++;
81         scsi_eh_wakeup(shost);
82         spin_unlock_irqrestore(shost->host_lock, flags);
83         return 1;
84 }
85
86 /**
87  * scsi_add_timer - Start timeout timer for a single scsi command.
88  * @scmd:       scsi command that is about to start running.
89  * @timeout:    amount of time to allow this command to run.
90  * @complete:   timeout function to call if timer isn't canceled.
91  *
92  * Notes:
93  *    This should be turned into an inline function.  Each scsi command
94  *    has its own timer, and as it is added to the queue, we set up the
95  *    timer.  When the command completes, we cancel the timer.
96  **/
97 void scsi_add_timer(struct scsi_cmnd *scmd, int timeout,
98                     void (*complete)(struct scsi_cmnd *))
99 {
100
101         /*
102          * If the clock was already running for this command, then
103          * first delete the timer.  The timer handling code gets rather
104          * confused if we don't do this.
105          */
106         if (scmd->eh_timeout.function)
107                 del_timer(&scmd->eh_timeout);
108
109         scmd->eh_timeout.data = (unsigned long)scmd;
110         scmd->eh_timeout.expires = jiffies + timeout;
111         scmd->eh_timeout.function = (void (*)(unsigned long)) complete;
112
113         SCSI_LOG_ERROR_RECOVERY(5, printk("%s: scmd: %p, time:"
114                                           " %d, (%p)\n", __FUNCTION__,
115                                           scmd, timeout, complete));
116
117         add_timer(&scmd->eh_timeout);
118 }
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
147 /**
148  * scsi_times_out - Timeout function for normal scsi commands.
149  * @scmd:       Cmd that is timing out.
150  *
151  * Notes:
152  *     We do not need to lock this.  There is the potential for a race
153  *     only in that the normal completion handling might run, but if the
154  *     normal completion function determines that the timer has already
155  *     fired, then it mustn't do anything.
156  **/
157 void scsi_times_out(struct scsi_cmnd *scmd)
158 {
159         scsi_log_completion(scmd, TIMEOUT_ERROR);
160
161         if (scmd->device->host->hostt->eh_timed_out)
162                 switch (scmd->device->host->hostt->eh_timed_out(scmd)) {
163                 case EH_HANDLED:
164                         __scsi_done(scmd);
165                         return;
166                 case EH_RESET_TIMER:
167                         /* This allows a single retry even of a command
168                          * with allowed == 0 */
169                         if (scmd->retries++ > scmd->allowed)
170                                 break;
171                         scsi_add_timer(scmd, scmd->timeout_per_command,
172                                        scsi_times_out);
173                         return;
174                 case EH_NOT_HANDLED:
175                         break;
176                 }
177
178         if (unlikely(!scsi_eh_scmd_add(scmd, SCSI_EH_CANCEL_CMD))) {
179                 panic("Error handler thread not present at %p %p %s %d",
180                       scmd, scmd->device->host, __FILE__, __LINE__);
181         }
182 }
183
184 /**
185  * scsi_block_when_processing_errors - Prevent cmds from being queued.
186  * @sdev:       Device on which we are performing recovery.
187  *
188  * Description:
189  *     We block until the host is out of error recovery, and then check to
190  *     see whether the host or the device is offline.
191  *
192  * Return value:
193  *     0 when dev was taken offline by error recovery. 1 OK to proceed.
194  **/
195 int scsi_block_when_processing_errors(struct scsi_device *sdev)
196 {
197         int online;
198
199         wait_event(sdev->host->host_wait, (sdev->host->shost_state !=
200                                            SHOST_RECOVERY));
201
202         online = scsi_device_online(sdev);
203
204         SCSI_LOG_ERROR_RECOVERY(5, printk("%s: rtn: %d\n", __FUNCTION__,
205                                           online));
206
207         return online;
208 }
209 EXPORT_SYMBOL(scsi_block_when_processing_errors);
210
211 #ifdef CONFIG_SCSI_LOGGING
212 /**
213  * scsi_eh_prt_fail_stats - Log info on failures.
214  * @shost:      scsi host being recovered.
215  * @work_q:     Queue of scsi cmds to process.
216  **/
217 static inline void scsi_eh_prt_fail_stats(struct Scsi_Host *shost,
218                                           struct list_head *work_q)
219 {
220         struct scsi_cmnd *scmd;
221         struct scsi_device *sdev;
222         int total_failures = 0;
223         int cmd_failed = 0;
224         int cmd_cancel = 0;
225         int devices_failed = 0;
226
227         shost_for_each_device(sdev, shost) {
228                 list_for_each_entry(scmd, work_q, eh_entry) {
229                         if (scmd->device == sdev) {
230                                 ++total_failures;
231                                 if (scmd->eh_eflags & SCSI_EH_CANCEL_CMD)
232                                         ++cmd_cancel;
233                                 else 
234                                         ++cmd_failed;
235                         }
236                 }
237
238                 if (cmd_cancel || cmd_failed) {
239                         SCSI_LOG_ERROR_RECOVERY(3,
240                                 printk("%s: %d:%d:%d:%d cmds failed: %d,"
241                                        " cancel: %d\n",
242                                        __FUNCTION__, shost->host_no,
243                                        sdev->channel, sdev->id, sdev->lun,
244                                        cmd_failed, cmd_cancel));
245                         cmd_cancel = 0;
246                         cmd_failed = 0;
247                         ++devices_failed;
248                 }
249         }
250
251         SCSI_LOG_ERROR_RECOVERY(2, printk("Total of %d commands on %d"
252                                           " devices require eh work\n",
253                                   total_failures, devices_failed));
254 }
255 #endif
256
257 /**
258  * scsi_check_sense - Examine scsi cmd sense
259  * @scmd:       Cmd to have sense checked.
260  *
261  * Return value:
262  *      SUCCESS or FAILED or NEEDS_RETRY
263  *
264  * Notes:
265  *      When a deferred error is detected the current command has
266  *      not been executed and needs retrying.
267  **/
268 static int scsi_check_sense(struct scsi_cmnd *scmd)
269 {
270         struct scsi_sense_hdr sshdr;
271
272         if (! scsi_command_normalize_sense(scmd, &sshdr))
273                 return FAILED;  /* no valid sense data */
274
275         if (scsi_sense_is_deferred(&sshdr))
276                 return NEEDS_RETRY;
277
278         /*
279          * Previous logic looked for FILEMARK, EOM or ILI which are
280          * mainly associated with tapes and returned SUCCESS.
281          */
282         if (sshdr.response_code == 0x70) {
283                 /* fixed format */
284                 if (scmd->sense_buffer[2] & 0xe0)
285                         return SUCCESS;
286         } else {
287                 /*
288                  * descriptor format: look for "stream commands sense data
289                  * descriptor" (see SSC-3). Assume single sense data
290                  * descriptor. Ignore ILI from SBC-2 READ LONG and WRITE LONG.
291                  */
292                 if ((sshdr.additional_length > 3) &&
293                     (scmd->sense_buffer[8] == 0x4) &&
294                     (scmd->sense_buffer[11] & 0xe0))
295                         return SUCCESS;
296         }
297
298         switch (sshdr.sense_key) {
299         case NO_SENSE:
300                 return SUCCESS;
301         case RECOVERED_ERROR:
302                 return /* soft_error */ SUCCESS;
303
304         case ABORTED_COMMAND:
305                 return NEEDS_RETRY;
306         case NOT_READY:
307         case UNIT_ATTENTION:
308                 /*
309                  * if we are expecting a cc/ua because of a bus reset that we
310                  * performed, treat this just as a retry.  otherwise this is
311                  * information that we should pass up to the upper-level driver
312                  * so that we can deal with it there.
313                  */
314                 if (scmd->device->expecting_cc_ua) {
315                         scmd->device->expecting_cc_ua = 0;
316                         return NEEDS_RETRY;
317                 }
318                 /*
319                  * if the device is in the process of becoming ready, we 
320                  * should retry.
321                  */
322                 if ((sshdr.asc == 0x04) && (sshdr.ascq == 0x01))
323                         return NEEDS_RETRY;
324                 /*
325                  * if the device is not started, we need to wake
326                  * the error handler to start the motor
327                  */
328                 if (scmd->device->allow_restart &&
329                     (sshdr.asc == 0x04) && (sshdr.ascq == 0x02))
330                         return FAILED;
331                 return SUCCESS;
332
333                 /* these three are not supported */
334         case COPY_ABORTED:
335         case VOLUME_OVERFLOW:
336         case MISCOMPARE:
337                 return SUCCESS;
338
339         case MEDIUM_ERROR:
340                 return NEEDS_RETRY;
341
342         case HARDWARE_ERROR:
343                 if (scmd->device->retry_hwerror)
344                         return NEEDS_RETRY;
345                 else
346                         return SUCCESS;
347
348         case ILLEGAL_REQUEST:
349         case BLANK_CHECK:
350         case DATA_PROTECT:
351         default:
352                 return SUCCESS;
353         }
354 }
355
356 /**
357  * scsi_eh_completed_normally - Disposition a eh cmd on return from LLD.
358  * @scmd:       SCSI cmd to examine.
359  *
360  * Notes:
361  *    This is *only* called when we are examining the status of commands
362  *    queued during error recovery.  the main difference here is that we
363  *    don't allow for the possibility of retries here, and we are a lot
364  *    more restrictive about what we consider acceptable.
365  **/
366 static int scsi_eh_completed_normally(struct scsi_cmnd *scmd)
367 {
368         /*
369          * first check the host byte, to see if there is anything in there
370          * that would indicate what we need to do.
371          */
372         if (host_byte(scmd->result) == DID_RESET) {
373                 /*
374                  * rats.  we are already in the error handler, so we now
375                  * get to try and figure out what to do next.  if the sense
376                  * is valid, we have a pretty good idea of what to do.
377                  * if not, we mark it as FAILED.
378                  */
379                 return scsi_check_sense(scmd);
380         }
381         if (host_byte(scmd->result) != DID_OK)
382                 return FAILED;
383
384         /*
385          * next, check the message byte.
386          */
387         if (msg_byte(scmd->result) != COMMAND_COMPLETE)
388                 return FAILED;
389
390         /*
391          * now, check the status byte to see if this indicates
392          * anything special.
393          */
394         switch (status_byte(scmd->result)) {
395         case GOOD:
396         case COMMAND_TERMINATED:
397                 return SUCCESS;
398         case CHECK_CONDITION:
399                 return scsi_check_sense(scmd);
400         case CONDITION_GOOD:
401         case INTERMEDIATE_GOOD:
402         case INTERMEDIATE_C_GOOD:
403                 /*
404                  * who knows?  FIXME(eric)
405                  */
406                 return SUCCESS;
407         case BUSY:
408         case QUEUE_FULL:
409         case RESERVATION_CONFLICT:
410         default:
411                 return FAILED;
412         }
413         return FAILED;
414 }
415
416 /**
417  * scsi_eh_times_out - timeout function for error handling.
418  * @scmd:       Cmd that is timing out.
419  *
420  * Notes:
421  *    During error handling, the kernel thread will be sleeping waiting
422  *    for some action to complete on the device.  our only job is to
423  *    record that it timed out, and to wake up the thread.
424  **/
425 static void scsi_eh_times_out(struct scsi_cmnd *scmd)
426 {
427         scmd->eh_eflags |= SCSI_EH_REC_TIMEOUT;
428         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: scmd:%p\n", __FUNCTION__,
429                                           scmd));
430
431         up(scmd->device->host->eh_action);
432 }
433
434 /**
435  * scsi_eh_done - Completion function for error handling.
436  * @scmd:       Cmd that is done.
437  **/
438 static void scsi_eh_done(struct scsi_cmnd *scmd)
439 {
440         /*
441          * if the timeout handler is already running, then just set the
442          * flag which says we finished late, and return.  we have no
443          * way of stopping the timeout handler from running, so we must
444          * always defer to it.
445          */
446         if (del_timer(&scmd->eh_timeout)) {
447                 scmd->request->rq_status = RQ_SCSI_DONE;
448
449                 SCSI_LOG_ERROR_RECOVERY(3, printk("%s scmd: %p result: %x\n",
450                                            __FUNCTION__, scmd, scmd->result));
451
452                 up(scmd->device->host->eh_action);
453         }
454 }
455
456 /**
457  * scsi_send_eh_cmnd  - send a cmd to a device as part of error recovery.
458  * @scmd:       SCSI Cmd to send.
459  * @timeout:    Timeout for cmd.
460  *
461  * Notes:
462  *    The initialization of the structures is quite a bit different in
463  *    this case, and furthermore, there is a different completion handler
464  *    vs scsi_dispatch_cmd.
465  * Return value:
466  *    SUCCESS or FAILED or NEEDS_RETRY
467  **/
468 static int scsi_send_eh_cmnd(struct scsi_cmnd *scmd, int timeout)
469 {
470         struct scsi_device *sdev = scmd->device;
471         struct Scsi_Host *shost = sdev->host;
472         DECLARE_MUTEX_LOCKED(sem);
473         unsigned long flags;
474         int rtn = SUCCESS;
475
476         /*
477          * we will use a queued command if possible, otherwise we will
478          * emulate the queuing and calling of completion function ourselves.
479          */
480         if (sdev->scsi_level <= SCSI_2)
481                 scmd->cmnd[1] = (scmd->cmnd[1] & 0x1f) |
482                         (sdev->lun << 5 & 0xe0);
483
484         scsi_add_timer(scmd, timeout, scsi_eh_times_out);
485
486         /*
487          * set up the semaphore so we wait for the command to complete.
488          */
489         shost->eh_action = &sem;
490         scmd->request->rq_status = RQ_SCSI_BUSY;
491
492         spin_lock_irqsave(shost->host_lock, flags);
493         scsi_log_send(scmd);
494         shost->hostt->queuecommand(scmd, scsi_eh_done);
495         spin_unlock_irqrestore(shost->host_lock, flags);
496
497         down(&sem);
498         scsi_log_completion(scmd, SUCCESS);
499
500         shost->eh_action = NULL;
501
502         /*
503          * see if timeout.  if so, tell the host to forget about it.
504          * in other words, we don't want a callback any more.
505          */
506         if (scmd->eh_eflags & SCSI_EH_REC_TIMEOUT) {
507                 scmd->eh_eflags &= ~SCSI_EH_REC_TIMEOUT;
508
509                 /*
510                  * as far as the low level driver is
511                  * concerned, this command is still active, so
512                  * we must give the low level driver a chance
513                  * to abort it. (db) 
514                  *
515                  * FIXME(eric) - we are not tracking whether we could
516                  * abort a timed out command or not.  not sure how
517                  * we should treat them differently anyways.
518                  */
519                 if (shost->hostt->eh_abort_handler)
520                         shost->hostt->eh_abort_handler(scmd);
521                         
522                 scmd->request->rq_status = RQ_SCSI_DONE;
523                 rtn = FAILED;
524         }
525
526         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: scmd: %p, rtn:%x\n",
527                                           __FUNCTION__, scmd, rtn));
528
529         /*
530          * now examine the actual status codes to see whether the command
531          * actually did complete normally.
532          */
533         if (rtn == SUCCESS) {
534                 rtn = scsi_eh_completed_normally(scmd);
535                 SCSI_LOG_ERROR_RECOVERY(3,
536                         printk("%s: scsi_eh_completed_normally %x\n",
537                                __FUNCTION__, rtn));
538                 switch (rtn) {
539                 case SUCCESS:
540                 case NEEDS_RETRY:
541                 case FAILED:
542                         break;
543                 default:
544                         rtn = FAILED;
545                         break;
546                 }
547         }
548
549         return rtn;
550 }
551
552 /**
553  * scsi_request_sense - Request sense data from a particular target.
554  * @scmd:       SCSI cmd for request sense.
555  *
556  * Notes:
557  *    Some hosts automatically obtain this information, others require
558  *    that we obtain it on our own. This function will *not* return until
559  *    the command either times out, or it completes.
560  **/
561 static int scsi_request_sense(struct scsi_cmnd *scmd)
562 {
563         static unsigned char generic_sense[6] =
564         {REQUEST_SENSE, 0, 0, 0, 252, 0};
565         unsigned char *scsi_result;
566         int saved_result;
567         int rtn;
568
569         memcpy(scmd->cmnd, generic_sense, sizeof(generic_sense));
570
571         scsi_result = kmalloc(252, GFP_ATOMIC | ((scmd->device->host->hostt->unchecked_isa_dma) ? __GFP_DMA : 0));
572
573
574         if (unlikely(!scsi_result)) {
575                 printk(KERN_ERR "%s: cannot allocate scsi_result.\n",
576                        __FUNCTION__);
577                 return FAILED;
578         }
579
580         /*
581          * zero the sense buffer.  some host adapters automatically always
582          * request sense, so it is not a good idea that
583          * scmd->request_buffer and scmd->sense_buffer point to the same
584          * address (db).  0 is not a valid sense code. 
585          */
586         memset(scmd->sense_buffer, 0, sizeof(scmd->sense_buffer));
587         memset(scsi_result, 0, 252);
588
589         saved_result = scmd->result;
590         scmd->request_buffer = scsi_result;
591         scmd->request_bufflen = 252;
592         scmd->use_sg = 0;
593         scmd->cmd_len = COMMAND_SIZE(scmd->cmnd[0]);
594         scmd->sc_data_direction = DMA_FROM_DEVICE;
595         scmd->underflow = 0;
596
597         rtn = scsi_send_eh_cmnd(scmd, SENSE_TIMEOUT);
598
599         /* last chance to have valid sense data */
600         if(!SCSI_SENSE_VALID(scmd)) {
601                 memcpy(scmd->sense_buffer, scmd->request_buffer,
602                        sizeof(scmd->sense_buffer));
603         }
604
605         kfree(scsi_result);
606
607         /*
608          * when we eventually call scsi_finish, we really wish to complete
609          * the original request, so let's restore the original data. (db)
610          */
611         scsi_setup_cmd_retry(scmd);
612         scmd->result = saved_result;
613         return rtn;
614 }
615
616 /**
617  * scsi_eh_finish_cmd - Handle a cmd that eh is finished with.
618  * @scmd:       Original SCSI cmd that eh has finished.
619  * @done_q:     Queue for processed commands.
620  *
621  * Notes:
622  *    We don't want to use the normal command completion while we are are
623  *    still handling errors - it may cause other commands to be queued,
624  *    and that would disturb what we are doing.  thus we really want to
625  *    keep a list of pending commands for final completion, and once we
626  *    are ready to leave error handling we handle completion for real.
627  **/
628 static void scsi_eh_finish_cmd(struct scsi_cmnd *scmd,
629                                struct list_head *done_q)
630 {
631         scmd->device->host->host_failed--;
632         scmd->eh_eflags = 0;
633
634         /*
635          * set this back so that the upper level can correctly free up
636          * things.
637          */
638         scsi_setup_cmd_retry(scmd);
639         list_move_tail(&scmd->eh_entry, done_q);
640 }
641
642 /**
643  * scsi_eh_get_sense - Get device sense data.
644  * @work_q:     Queue of commands to process.
645  * @done_q:     Queue of proccessed commands..
646  *
647  * Description:
648  *    See if we need to request sense information.  if so, then get it
649  *    now, so we have a better idea of what to do.  
650  *
651  * Notes:
652  *    This has the unfortunate side effect that if a shost adapter does
653  *    not automatically request sense information, that we end up shutting
654  *    it down before we request it.
655  *
656  *    All drivers should request sense information internally these days,
657  *    so for now all I have to say is tough noogies if you end up in here.
658  *
659  *    XXX: Long term this code should go away, but that needs an audit of
660  *         all LLDDs first.
661  **/
662 static int scsi_eh_get_sense(struct list_head *work_q,
663                              struct list_head *done_q)
664 {
665         struct scsi_cmnd *scmd, *next;
666         int rtn;
667
668         list_for_each_entry_safe(scmd, next, work_q, eh_entry) {
669                 if ((scmd->eh_eflags & SCSI_EH_CANCEL_CMD) ||
670                     SCSI_SENSE_VALID(scmd))
671                         continue;
672
673                 SCSI_LOG_ERROR_RECOVERY(2, printk("%s: requesting sense"
674                                                   " for id: %d\n",
675                                                   current->comm,
676                                                   scmd->device->id));
677                 rtn = scsi_request_sense(scmd);
678                 if (rtn != SUCCESS)
679                         continue;
680
681                 SCSI_LOG_ERROR_RECOVERY(3, printk("sense requested for %p"
682                                                   " result %x\n", scmd,
683                                                   scmd->result));
684                 SCSI_LOG_ERROR_RECOVERY(3, scsi_print_sense("bh", scmd));
685
686                 rtn = scsi_decide_disposition(scmd);
687
688                 /*
689                  * if the result was normal, then just pass it along to the
690                  * upper level.
691                  */
692                 if (rtn == SUCCESS)
693                         /* we don't want this command reissued, just
694                          * finished with the sense data, so set
695                          * retries to the max allowed to ensure it
696                          * won't get reissued */
697                         scmd->retries = scmd->allowed;
698                 else if (rtn != NEEDS_RETRY)
699                         continue;
700
701                 scsi_eh_finish_cmd(scmd, done_q);
702         }
703
704         return list_empty(work_q);
705 }
706
707 /**
708  * scsi_try_to_abort_cmd - Ask host to abort a running command.
709  * @scmd:       SCSI cmd to abort from Lower Level.
710  *
711  * Notes:
712  *    This function will not return until the user's completion function
713  *    has been called.  there is no timeout on this operation.  if the
714  *    author of the low-level driver wishes this operation to be timed,
715  *    they can provide this facility themselves.  helper functions in
716  *    scsi_error.c can be supplied to make this easier to do.
717  **/
718 static int scsi_try_to_abort_cmd(struct scsi_cmnd *scmd)
719 {
720         if (!scmd->device->host->hostt->eh_abort_handler)
721                 return FAILED;
722
723         /*
724          * scsi_done was called just after the command timed out and before
725          * we had a chance to process it. (db)
726          */
727         if (scmd->serial_number == 0)
728                 return SUCCESS;
729         return scmd->device->host->hostt->eh_abort_handler(scmd);
730 }
731
732 /**
733  * scsi_eh_tur - Send TUR to device.
734  * @scmd:       Scsi cmd to send TUR
735  *
736  * Return value:
737  *    0 - Device is ready. 1 - Device NOT ready.
738  **/
739 static int scsi_eh_tur(struct scsi_cmnd *scmd)
740 {
741         static unsigned char tur_command[6] = {TEST_UNIT_READY, 0, 0, 0, 0, 0};
742         int retry_cnt = 1, rtn;
743         int saved_result;
744
745 retry_tur:
746         memcpy(scmd->cmnd, tur_command, sizeof(tur_command));
747
748         /*
749          * zero the sense buffer.  the scsi spec mandates that any
750          * untransferred sense data should be interpreted as being zero.
751          */
752         memset(scmd->sense_buffer, 0, sizeof(scmd->sense_buffer));
753
754         saved_result = scmd->result;
755         scmd->request_buffer = NULL;
756         scmd->request_bufflen = 0;
757         scmd->use_sg = 0;
758         scmd->cmd_len = COMMAND_SIZE(scmd->cmnd[0]);
759         scmd->underflow = 0;
760         scmd->sc_data_direction = DMA_NONE;
761
762         rtn = scsi_send_eh_cmnd(scmd, SENSE_TIMEOUT);
763
764         /*
765          * when we eventually call scsi_finish, we really wish to complete
766          * the original request, so let's restore the original data. (db)
767          */
768         scsi_setup_cmd_retry(scmd);
769         scmd->result = saved_result;
770
771         /*
772          * hey, we are done.  let's look to see what happened.
773          */
774         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: scmd %p rtn %x\n",
775                 __FUNCTION__, scmd, rtn));
776         if (rtn == SUCCESS)
777                 return 0;
778         else if (rtn == NEEDS_RETRY) {
779                 if (retry_cnt--)
780                         goto retry_tur;
781                 return 0;
782         }
783         return 1;
784 }
785
786 /**
787  * scsi_eh_abort_cmds - abort canceled commands.
788  * @shost:      scsi host being recovered.
789  * @eh_done_q:  list_head for processed commands.
790  *
791  * Decription:
792  *    Try and see whether or not it makes sense to try and abort the
793  *    running command.  this only works out to be the case if we have one
794  *    command that has timed out.  if the command simply failed, it makes
795  *    no sense to try and abort the command, since as far as the shost
796  *    adapter is concerned, it isn't running.
797  **/
798 static int scsi_eh_abort_cmds(struct list_head *work_q,
799                               struct list_head *done_q)
800 {
801         struct scsi_cmnd *scmd, *next;
802         int rtn;
803
804         list_for_each_entry_safe(scmd, next, work_q, eh_entry) {
805                 if (!(scmd->eh_eflags & SCSI_EH_CANCEL_CMD))
806                         continue;
807                 SCSI_LOG_ERROR_RECOVERY(3, printk("%s: aborting cmd:"
808                                                   "0x%p\n", current->comm,
809                                                   scmd));
810                 rtn = scsi_try_to_abort_cmd(scmd);
811                 if (rtn == SUCCESS) {
812                         scmd->eh_eflags &= ~SCSI_EH_CANCEL_CMD;
813                         if (!scsi_device_online(scmd->device) ||
814                             !scsi_eh_tur(scmd)) {
815                                 scsi_eh_finish_cmd(scmd, done_q);
816                         }
817                                 
818                 } else
819                         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: aborting"
820                                                           " cmd failed:"
821                                                           "0x%p\n",
822                                                           current->comm,
823                                                           scmd));
824         }
825
826         return list_empty(work_q);
827 }
828
829 /**
830  * scsi_try_bus_device_reset - Ask host to perform a BDR on a dev
831  * @scmd:       SCSI cmd used to send BDR       
832  *
833  * Notes:
834  *    There is no timeout for this operation.  if this operation is
835  *    unreliable for a given host, then the host itself needs to put a
836  *    timer on it, and set the host back to a consistent state prior to
837  *    returning.
838  **/
839 static int scsi_try_bus_device_reset(struct scsi_cmnd *scmd)
840 {
841         int rtn;
842
843         if (!scmd->device->host->hostt->eh_device_reset_handler)
844                 return FAILED;
845
846         rtn = scmd->device->host->hostt->eh_device_reset_handler(scmd);
847         if (rtn == SUCCESS) {
848                 scmd->device->was_reset = 1;
849                 scmd->device->expecting_cc_ua = 1;
850         }
851
852         return rtn;
853 }
854
855 /**
856  * scsi_eh_try_stu - Send START_UNIT to device.
857  * @scmd:       Scsi cmd to send START_UNIT
858  *
859  * Return value:
860  *    0 - Device is ready. 1 - Device NOT ready.
861  **/
862 static int scsi_eh_try_stu(struct scsi_cmnd *scmd)
863 {
864         static unsigned char stu_command[6] = {START_STOP, 0, 0, 0, 1, 0};
865         int rtn;
866         int saved_result;
867
868         if (!scmd->device->allow_restart)
869                 return 1;
870
871         memcpy(scmd->cmnd, stu_command, sizeof(stu_command));
872
873         /*
874          * zero the sense buffer.  the scsi spec mandates that any
875          * untransferred sense data should be interpreted as being zero.
876          */
877         memset(scmd->sense_buffer, 0, sizeof(scmd->sense_buffer));
878
879         saved_result = scmd->result;
880         scmd->request_buffer = NULL;
881         scmd->request_bufflen = 0;
882         scmd->use_sg = 0;
883         scmd->cmd_len = COMMAND_SIZE(scmd->cmnd[0]);
884         scmd->underflow = 0;
885         scmd->sc_data_direction = DMA_NONE;
886
887         rtn = scsi_send_eh_cmnd(scmd, START_UNIT_TIMEOUT);
888
889         /*
890          * when we eventually call scsi_finish, we really wish to complete
891          * the original request, so let's restore the original data. (db)
892          */
893         scsi_setup_cmd_retry(scmd);
894         scmd->result = saved_result;
895
896         /*
897          * hey, we are done.  let's look to see what happened.
898          */
899         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: scmd %p rtn %x\n",
900                 __FUNCTION__, scmd, rtn));
901         if (rtn == SUCCESS)
902                 return 0;
903         return 1;
904 }
905
906  /**
907  * scsi_eh_stu - send START_UNIT if needed
908  * @shost:      scsi host being recovered.
909  * @eh_done_q:  list_head for processed commands.
910  *
911  * Notes:
912  *    If commands are failing due to not ready, initializing command required,
913  *      try revalidating the device, which will end up sending a start unit. 
914  **/
915 static int scsi_eh_stu(struct Scsi_Host *shost,
916                               struct list_head *work_q,
917                               struct list_head *done_q)
918 {
919         struct scsi_cmnd *scmd, *stu_scmd, *next;
920         struct scsi_device *sdev;
921
922         shost_for_each_device(sdev, shost) {
923                 stu_scmd = NULL;
924                 list_for_each_entry(scmd, work_q, eh_entry)
925                         if (scmd->device == sdev && SCSI_SENSE_VALID(scmd) &&
926                             scsi_check_sense(scmd) == FAILED ) {
927                                 stu_scmd = scmd;
928                                 break;
929                         }
930
931                 if (!stu_scmd)
932                         continue;
933
934                 SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Sending START_UNIT to sdev:"
935                                                   " 0x%p\n", current->comm, sdev));
936
937                 if (!scsi_eh_try_stu(stu_scmd)) {
938                         if (!scsi_device_online(sdev) ||
939                             !scsi_eh_tur(stu_scmd)) {
940                                 list_for_each_entry_safe(scmd, next,
941                                                           work_q, eh_entry) {
942                                         if (scmd->device == sdev)
943                                                 scsi_eh_finish_cmd(scmd, done_q);
944                                 }
945                         }
946                 } else {
947                         SCSI_LOG_ERROR_RECOVERY(3,
948                                                 printk("%s: START_UNIT failed to sdev:"
949                                                        " 0x%p\n", current->comm, sdev));
950                 }
951         }
952
953         return list_empty(work_q);
954 }
955
956
957 /**
958  * scsi_eh_bus_device_reset - send bdr if needed
959  * @shost:      scsi host being recovered.
960  * @eh_done_q:  list_head for processed commands.
961  *
962  * Notes:
963  *    Try a bus device reset.  still, look to see whether we have multiple
964  *    devices that are jammed or not - if we have multiple devices, it
965  *    makes no sense to try bus_device_reset - we really would need to try
966  *    a bus_reset instead. 
967  **/
968 static int scsi_eh_bus_device_reset(struct Scsi_Host *shost,
969                                     struct list_head *work_q,
970                                     struct list_head *done_q)
971 {
972         struct scsi_cmnd *scmd, *bdr_scmd, *next;
973         struct scsi_device *sdev;
974         int rtn;
975
976         shost_for_each_device(sdev, shost) {
977                 bdr_scmd = NULL;
978                 list_for_each_entry(scmd, work_q, eh_entry)
979                         if (scmd->device == sdev) {
980                                 bdr_scmd = scmd;
981                                 break;
982                         }
983
984                 if (!bdr_scmd)
985                         continue;
986
987                 SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Sending BDR sdev:"
988                                                   " 0x%p\n", current->comm,
989                                                   sdev));
990                 rtn = scsi_try_bus_device_reset(bdr_scmd);
991                 if (rtn == SUCCESS) {
992                         if (!scsi_device_online(sdev) ||
993                             !scsi_eh_tur(bdr_scmd)) {
994                                 list_for_each_entry_safe(scmd, next,
995                                                          work_q, eh_entry) {
996                                         if (scmd->device == sdev)
997                                                 scsi_eh_finish_cmd(scmd,
998                                                                    done_q);
999                                 }
1000                         }
1001                 } else {
1002                         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: BDR"
1003                                                           " failed sdev:"
1004                                                           "0x%p\n",
1005                                                           current->comm,
1006                                                            sdev));
1007                 }
1008         }
1009
1010         return list_empty(work_q);
1011 }
1012
1013 /**
1014  * scsi_try_bus_reset - ask host to perform a bus reset
1015  * @scmd:       SCSI cmd to send bus reset.
1016  **/
1017 static int scsi_try_bus_reset(struct scsi_cmnd *scmd)
1018 {
1019         unsigned long flags;
1020         int rtn;
1021
1022         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Snd Bus RST\n",
1023                                           __FUNCTION__));
1024
1025         if (!scmd->device->host->hostt->eh_bus_reset_handler)
1026                 return FAILED;
1027
1028         rtn = scmd->device->host->hostt->eh_bus_reset_handler(scmd);
1029
1030         if (rtn == SUCCESS) {
1031                 if (!scmd->device->host->hostt->skip_settle_delay)
1032                         ssleep(BUS_RESET_SETTLE_TIME);
1033                 spin_lock_irqsave(scmd->device->host->host_lock, flags);
1034                 scsi_report_bus_reset(scmd->device->host, scmd->device->channel);
1035                 spin_unlock_irqrestore(scmd->device->host->host_lock, flags);
1036         }
1037
1038         return rtn;
1039 }
1040
1041 /**
1042  * scsi_try_host_reset - ask host adapter to reset itself
1043  * @scmd:       SCSI cmd to send hsot reset.
1044  **/
1045 static int scsi_try_host_reset(struct scsi_cmnd *scmd)
1046 {
1047         unsigned long flags;
1048         int rtn;
1049
1050         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Snd Host RST\n",
1051                                           __FUNCTION__));
1052
1053         if (!scmd->device->host->hostt->eh_host_reset_handler)
1054                 return FAILED;
1055
1056         rtn = scmd->device->host->hostt->eh_host_reset_handler(scmd);
1057
1058         if (rtn == SUCCESS) {
1059                 if (!scmd->device->host->hostt->skip_settle_delay)
1060                         ssleep(HOST_RESET_SETTLE_TIME);
1061                 spin_lock_irqsave(scmd->device->host->host_lock, flags);
1062                 scsi_report_bus_reset(scmd->device->host, scmd->device->channel);
1063                 spin_unlock_irqrestore(scmd->device->host->host_lock, flags);
1064         }
1065
1066         return rtn;
1067 }
1068
1069 /**
1070  * scsi_eh_bus_reset - send a bus reset 
1071  * @shost:      scsi host being recovered.
1072  * @eh_done_q:  list_head for processed commands.
1073  **/
1074 static int scsi_eh_bus_reset(struct Scsi_Host *shost,
1075                              struct list_head *work_q,
1076                              struct list_head *done_q)
1077 {
1078         struct scsi_cmnd *scmd, *chan_scmd, *next;
1079         unsigned int channel;
1080         int rtn;
1081
1082         /*
1083          * we really want to loop over the various channels, and do this on
1084          * a channel by channel basis.  we should also check to see if any
1085          * of the failed commands are on soft_reset devices, and if so, skip
1086          * the reset.  
1087          */
1088
1089         for (channel = 0; channel <= shost->max_channel; channel++) {
1090                 chan_scmd = NULL;
1091                 list_for_each_entry(scmd, work_q, eh_entry) {
1092                         if (channel == scmd->device->channel) {
1093                                 chan_scmd = scmd;
1094                                 break;
1095                                 /*
1096                                  * FIXME add back in some support for
1097                                  * soft_reset devices.
1098                                  */
1099                         }
1100                 }
1101
1102                 if (!chan_scmd)
1103                         continue;
1104                 SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Sending BRST chan:"
1105                                                   " %d\n", current->comm,
1106                                                   channel));
1107                 rtn = scsi_try_bus_reset(chan_scmd);
1108                 if (rtn == SUCCESS) {
1109                         list_for_each_entry_safe(scmd, next, work_q, eh_entry) {
1110                                 if (channel == scmd->device->channel)
1111                                         if (!scsi_device_online(scmd->device) ||
1112                                             !scsi_eh_tur(scmd))
1113                                                 scsi_eh_finish_cmd(scmd,
1114                                                                    done_q);
1115                         }
1116                 } else {
1117                         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: BRST"
1118                                                           " failed chan: %d\n",
1119                                                           current->comm,
1120                                                           channel));
1121                 }
1122         }
1123         return list_empty(work_q);
1124 }
1125
1126 /**
1127  * scsi_eh_host_reset - send a host reset 
1128  * @work_q:     list_head for processed commands.
1129  * @done_q:     list_head for processed commands.
1130  **/
1131 static int scsi_eh_host_reset(struct list_head *work_q,
1132                               struct list_head *done_q)
1133 {
1134         struct scsi_cmnd *scmd, *next;
1135         int rtn;
1136
1137         if (!list_empty(work_q)) {
1138                 scmd = list_entry(work_q->next,
1139                                   struct scsi_cmnd, eh_entry);
1140
1141                 SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Sending HRST\n"
1142                                                   , current->comm));
1143
1144                 rtn = scsi_try_host_reset(scmd);
1145                 if (rtn == SUCCESS) {
1146                         list_for_each_entry_safe(scmd, next, work_q, eh_entry) {
1147                                 if (!scsi_device_online(scmd->device) ||
1148                                     (!scsi_eh_try_stu(scmd) && !scsi_eh_tur(scmd)) ||
1149                                     !scsi_eh_tur(scmd))
1150                                         scsi_eh_finish_cmd(scmd, done_q);
1151                         }
1152                 } else {
1153                         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: HRST"
1154                                                           " failed\n",
1155                                                           current->comm));
1156                 }
1157         }
1158         return list_empty(work_q);
1159 }
1160
1161 /**
1162  * scsi_eh_offline_sdevs - offline scsi devices that fail to recover
1163  * @work_q:     list_head for processed commands.
1164  * @done_q:     list_head for processed commands.
1165  *
1166  **/
1167 static void scsi_eh_offline_sdevs(struct list_head *work_q,
1168                                   struct list_head *done_q)
1169 {
1170         struct scsi_cmnd *scmd, *next;
1171
1172         list_for_each_entry_safe(scmd, next, work_q, eh_entry) {
1173                 printk(KERN_INFO "scsi: Device offlined - not"
1174                                 " ready after error recovery: host"
1175                                 " %d channel %d id %d lun %d\n",
1176                                 scmd->device->host->host_no,
1177                                 scmd->device->channel,
1178                                 scmd->device->id,
1179                                 scmd->device->lun);
1180                 scsi_device_set_state(scmd->device, SDEV_OFFLINE);
1181                 if (scmd->eh_eflags & SCSI_EH_CANCEL_CMD) {
1182                         /*
1183                          * FIXME: Handle lost cmds.
1184                          */
1185                 }
1186                 scsi_eh_finish_cmd(scmd, done_q);
1187         }
1188         return;
1189 }
1190
1191 /**
1192  * scsi_decide_disposition - Disposition a cmd on return from LLD.
1193  * @scmd:       SCSI cmd to examine.
1194  *
1195  * Notes:
1196  *    This is *only* called when we are examining the status after sending
1197  *    out the actual data command.  any commands that are queued for error
1198  *    recovery (e.g. test_unit_ready) do *not* come through here.
1199  *
1200  *    When this routine returns failed, it means the error handler thread
1201  *    is woken.  In cases where the error code indicates an error that
1202  *    doesn't require the error handler read (i.e. we don't need to
1203  *    abort/reset), this function should return SUCCESS.
1204  **/
1205 int scsi_decide_disposition(struct scsi_cmnd *scmd)
1206 {
1207         int rtn;
1208
1209         /*
1210          * if the device is offline, then we clearly just pass the result back
1211          * up to the top level.
1212          */
1213         if (!scsi_device_online(scmd->device)) {
1214                 SCSI_LOG_ERROR_RECOVERY(5, printk("%s: device offline - report"
1215                                                   " as SUCCESS\n",
1216                                                   __FUNCTION__));
1217                 return SUCCESS;
1218         }
1219
1220         /*
1221          * first check the host byte, to see if there is anything in there
1222          * that would indicate what we need to do.
1223          */
1224         switch (host_byte(scmd->result)) {
1225         case DID_PASSTHROUGH:
1226                 /*
1227                  * no matter what, pass this through to the upper layer.
1228                  * nuke this special code so that it looks like we are saying
1229                  * did_ok.
1230                  */
1231                 scmd->result &= 0xff00ffff;
1232                 return SUCCESS;
1233         case DID_OK:
1234                 /*
1235                  * looks good.  drop through, and check the next byte.
1236                  */
1237                 break;
1238         case DID_NO_CONNECT:
1239         case DID_BAD_TARGET:
1240         case DID_ABORT:
1241                 /*
1242                  * note - this means that we just report the status back
1243                  * to the top level driver, not that we actually think
1244                  * that it indicates SUCCESS.
1245                  */
1246                 return SUCCESS;
1247                 /*
1248                  * when the low level driver returns did_soft_error,
1249                  * it is responsible for keeping an internal retry counter 
1250                  * in order to avoid endless loops (db)
1251                  *
1252                  * actually this is a bug in this function here.  we should
1253                  * be mindful of the maximum number of retries specified
1254                  * and not get stuck in a loop.
1255                  */
1256         case DID_SOFT_ERROR:
1257                 goto maybe_retry;
1258         case DID_IMM_RETRY:
1259                 return NEEDS_RETRY;
1260
1261         case DID_REQUEUE:
1262                 return ADD_TO_MLQUEUE;
1263
1264         case DID_ERROR:
1265                 if (msg_byte(scmd->result) == COMMAND_COMPLETE &&
1266                     status_byte(scmd->result) == RESERVATION_CONFLICT)
1267                         /*
1268                          * execute reservation conflict processing code
1269                          * lower down
1270                          */
1271                         break;
1272                 /* fallthrough */
1273
1274         case DID_BUS_BUSY:
1275         case DID_PARITY:
1276                 goto maybe_retry;
1277         case DID_TIME_OUT:
1278                 /*
1279                  * when we scan the bus, we get timeout messages for
1280                  * these commands if there is no device available.
1281                  * other hosts report did_no_connect for the same thing.
1282                  */
1283                 if ((scmd->cmnd[0] == TEST_UNIT_READY ||
1284                      scmd->cmnd[0] == INQUIRY)) {
1285                         return SUCCESS;
1286                 } else {
1287                         return FAILED;
1288                 }
1289         case DID_RESET:
1290                 return SUCCESS;
1291         default:
1292                 return FAILED;
1293         }
1294
1295         /*
1296          * next, check the message byte.
1297          */
1298         if (msg_byte(scmd->result) != COMMAND_COMPLETE)
1299                 return FAILED;
1300
1301         /*
1302          * check the status byte to see if this indicates anything special.
1303          */
1304         switch (status_byte(scmd->result)) {
1305         case QUEUE_FULL:
1306                 /*
1307                  * the case of trying to send too many commands to a
1308                  * tagged queueing device.
1309                  */
1310         case BUSY:
1311                 /*
1312                  * device can't talk to us at the moment.  Should only
1313                  * occur (SAM-3) when the task queue is empty, so will cause
1314                  * the empty queue handling to trigger a stall in the
1315                  * device.
1316                  */
1317                 return ADD_TO_MLQUEUE;
1318         case GOOD:
1319         case COMMAND_TERMINATED:
1320         case TASK_ABORTED:
1321                 return SUCCESS;
1322         case CHECK_CONDITION:
1323                 rtn = scsi_check_sense(scmd);
1324                 if (rtn == NEEDS_RETRY)
1325                         goto maybe_retry;
1326                 /* if rtn == FAILED, we have no sense information;
1327                  * returning FAILED will wake the error handler thread
1328                  * to collect the sense and redo the decide
1329                  * disposition */
1330                 return rtn;
1331         case CONDITION_GOOD:
1332         case INTERMEDIATE_GOOD:
1333         case INTERMEDIATE_C_GOOD:
1334         case ACA_ACTIVE:
1335                 /*
1336                  * who knows?  FIXME(eric)
1337                  */
1338                 return SUCCESS;
1339
1340         case RESERVATION_CONFLICT:
1341                 printk(KERN_INFO "scsi: reservation conflict: host"
1342                                 " %d channel %d id %d lun %d\n",
1343                        scmd->device->host->host_no, scmd->device->channel,
1344                        scmd->device->id, scmd->device->lun);
1345                 return SUCCESS; /* causes immediate i/o error */
1346         default:
1347                 return FAILED;
1348         }
1349         return FAILED;
1350
1351       maybe_retry:
1352
1353         /* we requeue for retry because the error was retryable, and
1354          * the request was not marked fast fail.  Note that above,
1355          * even if the request is marked fast fail, we still requeue
1356          * for queue congestion conditions (QUEUE_FULL or BUSY) */
1357         if ((++scmd->retries) < scmd->allowed 
1358             && !blk_noretry_request(scmd->request)) {
1359                 return NEEDS_RETRY;
1360         } else {
1361                 /*
1362                  * no more retries - report this one back to upper level.
1363                  */
1364                 return SUCCESS;
1365         }
1366 }
1367
1368 /**
1369  * scsi_eh_lock_done - done function for eh door lock request
1370  * @scmd:       SCSI command block for the door lock request
1371  *
1372  * Notes:
1373  *      We completed the asynchronous door lock request, and it has either
1374  *      locked the door or failed.  We must free the command structures
1375  *      associated with this request.
1376  **/
1377 static void scsi_eh_lock_done(struct scsi_cmnd *scmd)
1378 {
1379         struct scsi_request *sreq = scmd->sc_request;
1380
1381         scsi_release_request(sreq);
1382 }
1383
1384
1385 /**
1386  * scsi_eh_lock_door - Prevent medium removal for the specified device
1387  * @sdev:       SCSI device to prevent medium removal
1388  *
1389  * Locking:
1390  *      We must be called from process context; scsi_allocate_request()
1391  *      may sleep.
1392  *
1393  * Notes:
1394  *      We queue up an asynchronous "ALLOW MEDIUM REMOVAL" request on the
1395  *      head of the devices request queue, and continue.
1396  *
1397  * Bugs:
1398  *      scsi_allocate_request() may sleep waiting for existing requests to
1399  *      be processed.  However, since we haven't kicked off any request
1400  *      processing for this host, this may deadlock.
1401  *
1402  *      If scsi_allocate_request() fails for what ever reason, we
1403  *      completely forget to lock the door.
1404  **/
1405 static void scsi_eh_lock_door(struct scsi_device *sdev)
1406 {
1407         struct scsi_request *sreq = scsi_allocate_request(sdev, GFP_KERNEL);
1408
1409         if (unlikely(!sreq)) {
1410                 printk(KERN_ERR "%s: request allocate failed,"
1411                        "prevent media removal cmd not sent\n", __FUNCTION__);
1412                 return;
1413         }
1414
1415         sreq->sr_cmnd[0] = ALLOW_MEDIUM_REMOVAL;
1416         sreq->sr_cmnd[1] = 0;
1417         sreq->sr_cmnd[2] = 0;
1418         sreq->sr_cmnd[3] = 0;
1419         sreq->sr_cmnd[4] = SCSI_REMOVAL_PREVENT;
1420         sreq->sr_cmnd[5] = 0;
1421         sreq->sr_data_direction = DMA_NONE;
1422         sreq->sr_bufflen = 0;
1423         sreq->sr_buffer = NULL;
1424         sreq->sr_allowed = 5;
1425         sreq->sr_done = scsi_eh_lock_done;
1426         sreq->sr_timeout_per_command = 10 * HZ;
1427         sreq->sr_cmd_len = COMMAND_SIZE(sreq->sr_cmnd[0]);
1428
1429         scsi_insert_special_req(sreq, 1);
1430 }
1431
1432
1433 /**
1434  * scsi_restart_operations - restart io operations to the specified host.
1435  * @shost:      Host we are restarting.
1436  *
1437  * Notes:
1438  *    When we entered the error handler, we blocked all further i/o to
1439  *    this device.  we need to 'reverse' this process.
1440  **/
1441 static void scsi_restart_operations(struct Scsi_Host *shost)
1442 {
1443         struct scsi_device *sdev;
1444
1445         /*
1446          * If the door was locked, we need to insert a door lock request
1447          * onto the head of the SCSI request queue for the device.  There
1448          * is no point trying to lock the door of an off-line device.
1449          */
1450         shost_for_each_device(sdev, shost) {
1451                 if (scsi_device_online(sdev) && sdev->locked)
1452                         scsi_eh_lock_door(sdev);
1453         }
1454
1455         /*
1456          * next free up anything directly waiting upon the host.  this
1457          * will be requests for character device operations, and also for
1458          * ioctls to queued block devices.
1459          */
1460         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: waking up host to restart\n",
1461                                           __FUNCTION__));
1462
1463         scsi_host_set_state(shost, SHOST_RUNNING);
1464
1465         wake_up(&shost->host_wait);
1466
1467         /*
1468          * finally we need to re-initiate requests that may be pending.  we will
1469          * have had everything blocked while error handling is taking place, and
1470          * now that error recovery is done, we will need to ensure that these
1471          * requests are started.
1472          */
1473         scsi_run_host_queues(shost);
1474 }
1475
1476 /**
1477  * scsi_eh_ready_devs - check device ready state and recover if not.
1478  * @shost:      host to be recovered.
1479  * @eh_done_q:  list_head for processed commands.
1480  *
1481  **/
1482 static void scsi_eh_ready_devs(struct Scsi_Host *shost,
1483                                struct list_head *work_q,
1484                                struct list_head *done_q)
1485 {
1486         if (!scsi_eh_stu(shost, work_q, done_q))
1487                 if (!scsi_eh_bus_device_reset(shost, work_q, done_q))
1488                         if (!scsi_eh_bus_reset(shost, work_q, done_q))
1489                                 if (!scsi_eh_host_reset(work_q, done_q))
1490                                         scsi_eh_offline_sdevs(work_q, done_q);
1491 }
1492
1493 /**
1494  * scsi_eh_flush_done_q - finish processed commands or retry them.
1495  * @done_q:     list_head of processed commands.
1496  *
1497  **/
1498 static void scsi_eh_flush_done_q(struct list_head *done_q)
1499 {
1500         struct scsi_cmnd *scmd, *next;
1501
1502         list_for_each_entry_safe(scmd, next, done_q, eh_entry) {
1503                 list_del_init(&scmd->eh_entry);
1504                 if (scsi_device_online(scmd->device) &&
1505                     !blk_noretry_request(scmd->request) &&
1506                     (++scmd->retries < scmd->allowed)) {
1507                         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: flush"
1508                                                           " retry cmd: %p\n",
1509                                                           current->comm,
1510                                                           scmd));
1511                                 scsi_queue_insert(scmd, SCSI_MLQUEUE_EH_RETRY);
1512                 } else {
1513                         /*
1514                          * If just we got sense for the device (called
1515                          * scsi_eh_get_sense), scmd->result is already
1516                          * set, do not set DRIVER_TIMEOUT.
1517                          */
1518                         if (!scmd->result)
1519                                 scmd->result |= (DRIVER_TIMEOUT << 24);
1520                         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: flush finish"
1521                                                         " cmd: %p\n",
1522                                                         current->comm, scmd));
1523                         scsi_finish_command(scmd);
1524                 }
1525         }
1526 }
1527
1528 /**
1529  * scsi_unjam_host - Attempt to fix a host which has a cmd that failed.
1530  * @shost:      Host to unjam.
1531  *
1532  * Notes:
1533  *    When we come in here, we *know* that all commands on the bus have
1534  *    either completed, failed or timed out.  we also know that no further
1535  *    commands are being sent to the host, so things are relatively quiet
1536  *    and we have freedom to fiddle with things as we wish.
1537  *
1538  *    This is only the *default* implementation.  it is possible for
1539  *    individual drivers to supply their own version of this function, and
1540  *    if the maintainer wishes to do this, it is strongly suggested that
1541  *    this function be taken as a template and modified.  this function
1542  *    was designed to correctly handle problems for about 95% of the
1543  *    different cases out there, and it should always provide at least a
1544  *    reasonable amount of error recovery.
1545  *
1546  *    Any command marked 'failed' or 'timeout' must eventually have
1547  *    scsi_finish_cmd() called for it.  we do all of the retry stuff
1548  *    here, so when we restart the host after we return it should have an
1549  *    empty queue.
1550  **/
1551 static void scsi_unjam_host(struct Scsi_Host *shost)
1552 {
1553         unsigned long flags;
1554         LIST_HEAD(eh_work_q);
1555         LIST_HEAD(eh_done_q);
1556
1557         spin_lock_irqsave(shost->host_lock, flags);
1558         list_splice_init(&shost->eh_cmd_q, &eh_work_q);
1559         spin_unlock_irqrestore(shost->host_lock, flags);
1560
1561         SCSI_LOG_ERROR_RECOVERY(1, scsi_eh_prt_fail_stats(shost, &eh_work_q));
1562
1563         if (!scsi_eh_get_sense(&eh_work_q, &eh_done_q))
1564                 if (!scsi_eh_abort_cmds(&eh_work_q, &eh_done_q))
1565                         scsi_eh_ready_devs(shost, &eh_work_q, &eh_done_q);
1566
1567         scsi_eh_flush_done_q(&eh_done_q);
1568 }
1569
1570 /**
1571  * scsi_error_handler - Handle errors/timeouts of SCSI cmds.
1572  * @data:       Host for which we are running.
1573  *
1574  * Notes:
1575  *    This is always run in the context of a kernel thread.  The idea is
1576  *    that we start this thing up when the kernel starts up (one per host
1577  *    that we detect), and it immediately goes to sleep and waits for some
1578  *    event (i.e. failure).  When this takes place, we have the job of
1579  *    trying to unjam the bus and restarting things.
1580  **/
1581 int scsi_error_handler(void *data)
1582 {
1583         struct Scsi_Host *shost = (struct Scsi_Host *) data;
1584         int rtn;
1585         DECLARE_MUTEX_LOCKED(sem);
1586
1587         current->flags |= PF_NOFREEZE;
1588         shost->eh_wait = &sem;
1589
1590         /*
1591          * Wake up the thread that created us.
1592          */
1593         SCSI_LOG_ERROR_RECOVERY(3, printk("Wake up parent of"
1594                                           " scsi_eh_%d\n",shost->host_no));
1595
1596         while (1) {
1597                 /*
1598                  * If we get a signal, it means we are supposed to go
1599                  * away and die.  This typically happens if the user is
1600                  * trying to unload a module.
1601                  */
1602                 SCSI_LOG_ERROR_RECOVERY(1, printk("Error handler"
1603                                                   " scsi_eh_%d"
1604                                                   " sleeping\n",shost->host_no));
1605
1606                 /*
1607                  * Note - we always use down_interruptible with the semaphore
1608                  * even if the module was loaded as part of the kernel.  The
1609                  * reason is that down() will cause this thread to be counted
1610                  * in the load average as a running process, and down
1611                  * interruptible doesn't.  Given that we need to allow this
1612                  * thread to die if the driver was loaded as a module, using
1613                  * semaphores isn't unreasonable.
1614                  */
1615                 down_interruptible(&sem);
1616                 if (kthread_should_stop())
1617                         break;
1618
1619                 SCSI_LOG_ERROR_RECOVERY(1, printk("Error handler"
1620                                                   " scsi_eh_%d waking"
1621                                                   " up\n",shost->host_no));
1622
1623                 shost->eh_active = 1;
1624
1625                 /*
1626                  * We have a host that is failing for some reason.  Figure out
1627                  * what we need to do to get it up and online again (if we can).
1628                  * If we fail, we end up taking the thing offline.
1629                  */
1630                 if (shost->hostt->eh_strategy_handler) 
1631                         rtn = shost->hostt->eh_strategy_handler(shost);
1632                 else
1633                         scsi_unjam_host(shost);
1634
1635                 shost->eh_active = 0;
1636
1637                 /*
1638                  * Note - if the above fails completely, the action is to take
1639                  * individual devices offline and flush the queue of any
1640                  * outstanding requests that may have been pending.  When we
1641                  * restart, we restart any I/O to any other devices on the bus
1642                  * which are still online.
1643                  */
1644                 scsi_restart_operations(shost);
1645
1646         }
1647
1648         SCSI_LOG_ERROR_RECOVERY(1, printk("Error handler scsi_eh_%d"
1649                                           " exiting\n",shost->host_no));
1650
1651         /*
1652          * Make sure that nobody tries to wake us up again.
1653          */
1654         shost->eh_wait = NULL;
1655         return 0;
1656 }
1657
1658 /*
1659  * Function:    scsi_report_bus_reset()
1660  *
1661  * Purpose:     Utility function used by low-level drivers to report that
1662  *              they have observed a bus reset on the bus being handled.
1663  *
1664  * Arguments:   shost       - Host in question
1665  *              channel     - channel on which reset was observed.
1666  *
1667  * Returns:     Nothing
1668  *
1669  * Lock status: Host lock must be held.
1670  *
1671  * Notes:       This only needs to be called if the reset is one which
1672  *              originates from an unknown location.  Resets originated
1673  *              by the mid-level itself don't need to call this, but there
1674  *              should be no harm.
1675  *
1676  *              The main purpose of this is to make sure that a CHECK_CONDITION
1677  *              is properly treated.
1678  */
1679 void scsi_report_bus_reset(struct Scsi_Host *shost, int channel)
1680 {
1681         struct scsi_device *sdev;
1682
1683         __shost_for_each_device(sdev, shost) {
1684                 if (channel == sdev->channel) {
1685                         sdev->was_reset = 1;
1686                         sdev->expecting_cc_ua = 1;
1687                 }
1688         }
1689 }
1690 EXPORT_SYMBOL(scsi_report_bus_reset);
1691
1692 /*
1693  * Function:    scsi_report_device_reset()
1694  *
1695  * Purpose:     Utility function used by low-level drivers to report that
1696  *              they have observed a device reset on the device being handled.
1697  *
1698  * Arguments:   shost       - Host in question
1699  *              channel     - channel on which reset was observed
1700  *              target      - target on which reset was observed
1701  *
1702  * Returns:     Nothing
1703  *
1704  * Lock status: Host lock must be held
1705  *
1706  * Notes:       This only needs to be called if the reset is one which
1707  *              originates from an unknown location.  Resets originated
1708  *              by the mid-level itself don't need to call this, but there
1709  *              should be no harm.
1710  *
1711  *              The main purpose of this is to make sure that a CHECK_CONDITION
1712  *              is properly treated.
1713  */
1714 void scsi_report_device_reset(struct Scsi_Host *shost, int channel, int target)
1715 {
1716         struct scsi_device *sdev;
1717
1718         __shost_for_each_device(sdev, shost) {
1719                 if (channel == sdev->channel &&
1720                     target == sdev->id) {
1721                         sdev->was_reset = 1;
1722                         sdev->expecting_cc_ua = 1;
1723                 }
1724         }
1725 }
1726 EXPORT_SYMBOL(scsi_report_device_reset);
1727
1728 static void
1729 scsi_reset_provider_done_command(struct scsi_cmnd *scmd)
1730 {
1731 }
1732
1733 /*
1734  * Function:    scsi_reset_provider
1735  *
1736  * Purpose:     Send requested reset to a bus or device at any phase.
1737  *
1738  * Arguments:   device  - device to send reset to
1739  *              flag - reset type (see scsi.h)
1740  *
1741  * Returns:     SUCCESS/FAILURE.
1742  *
1743  * Notes:       This is used by the SCSI Generic driver to provide
1744  *              Bus/Device reset capability.
1745  */
1746 int
1747 scsi_reset_provider(struct scsi_device *dev, int flag)
1748 {
1749         struct scsi_cmnd *scmd = scsi_get_command(dev, GFP_KERNEL);
1750         struct request req;
1751         int rtn;
1752
1753         scmd->request = &req;
1754         memset(&scmd->eh_timeout, 0, sizeof(scmd->eh_timeout));
1755         scmd->request->rq_status        = RQ_SCSI_BUSY;
1756
1757         memset(&scmd->cmnd, '\0', sizeof(scmd->cmnd));
1758     
1759         scmd->scsi_done         = scsi_reset_provider_done_command;
1760         scmd->done                      = NULL;
1761         scmd->buffer                    = NULL;
1762         scmd->bufflen                   = 0;
1763         scmd->request_buffer            = NULL;
1764         scmd->request_bufflen           = 0;
1765
1766         scmd->cmd_len                   = 0;
1767
1768         scmd->sc_data_direction         = DMA_BIDIRECTIONAL;
1769         scmd->sc_request                = NULL;
1770         scmd->sc_magic                  = SCSI_CMND_MAGIC;
1771
1772         init_timer(&scmd->eh_timeout);
1773
1774         /*
1775          * Sometimes the command can get back into the timer chain,
1776          * so use the pid as an identifier.
1777          */
1778         scmd->pid                       = 0;
1779
1780         switch (flag) {
1781         case SCSI_TRY_RESET_DEVICE:
1782                 rtn = scsi_try_bus_device_reset(scmd);
1783                 if (rtn == SUCCESS)
1784                         break;
1785                 /* FALLTHROUGH */
1786         case SCSI_TRY_RESET_BUS:
1787                 rtn = scsi_try_bus_reset(scmd);
1788                 if (rtn == SUCCESS)
1789                         break;
1790                 /* FALLTHROUGH */
1791         case SCSI_TRY_RESET_HOST:
1792                 rtn = scsi_try_host_reset(scmd);
1793                 break;
1794         default:
1795                 rtn = FAILED;
1796         }
1797
1798         scsi_next_command(scmd);
1799         return rtn;
1800 }
1801 EXPORT_SYMBOL(scsi_reset_provider);
1802
1803 /**
1804  * scsi_normalize_sense - normalize main elements from either fixed or
1805  *                      descriptor sense data format into a common format.
1806  *
1807  * @sense_buffer:       byte array containing sense data returned by device
1808  * @sb_len:             number of valid bytes in sense_buffer
1809  * @sshdr:              pointer to instance of structure that common
1810  *                      elements are written to.
1811  *
1812  * Notes:
1813  *      The "main elements" from sense data are: response_code, sense_key,
1814  *      asc, ascq and additional_length (only for descriptor format).
1815  *
1816  *      Typically this function can be called after a device has
1817  *      responded to a SCSI command with the CHECK_CONDITION status.
1818  *
1819  * Return value:
1820  *      1 if valid sense data information found, else 0;
1821  **/
1822 int scsi_normalize_sense(const u8 *sense_buffer, int sb_len,
1823                          struct scsi_sense_hdr *sshdr)
1824 {
1825         if (!sense_buffer || !sb_len)
1826                 return 0;
1827
1828         memset(sshdr, 0, sizeof(struct scsi_sense_hdr));
1829
1830         sshdr->response_code = (sense_buffer[0] & 0x7f);
1831
1832         if (!scsi_sense_valid(sshdr))
1833                 return 0;
1834
1835         if (sshdr->response_code >= 0x72) {
1836                 /*
1837                  * descriptor format
1838                  */
1839                 if (sb_len > 1)
1840                         sshdr->sense_key = (sense_buffer[1] & 0xf);
1841                 if (sb_len > 2)
1842                         sshdr->asc = sense_buffer[2];
1843                 if (sb_len > 3)
1844                         sshdr->ascq = sense_buffer[3];
1845                 if (sb_len > 7)
1846                         sshdr->additional_length = sense_buffer[7];
1847         } else {
1848                 /* 
1849                  * fixed format
1850                  */
1851                 if (sb_len > 2)
1852                         sshdr->sense_key = (sense_buffer[2] & 0xf);
1853                 if (sb_len > 7) {
1854                         sb_len = (sb_len < (sense_buffer[7] + 8)) ?
1855                                          sb_len : (sense_buffer[7] + 8);
1856                         if (sb_len > 12)
1857                                 sshdr->asc = sense_buffer[12];
1858                         if (sb_len > 13)
1859                                 sshdr->ascq = sense_buffer[13];
1860                 }
1861         }
1862
1863         return 1;
1864 }
1865 EXPORT_SYMBOL(scsi_normalize_sense);
1866
1867 int scsi_request_normalize_sense(struct scsi_request *sreq,
1868                                  struct scsi_sense_hdr *sshdr)
1869 {
1870         return scsi_normalize_sense(sreq->sr_sense_buffer,
1871                         sizeof(sreq->sr_sense_buffer), sshdr);
1872 }
1873 EXPORT_SYMBOL(scsi_request_normalize_sense);
1874
1875 int scsi_command_normalize_sense(struct scsi_cmnd *cmd,
1876                                  struct scsi_sense_hdr *sshdr)
1877 {
1878         return scsi_normalize_sense(cmd->sense_buffer,
1879                         sizeof(cmd->sense_buffer), sshdr);
1880 }
1881 EXPORT_SYMBOL(scsi_command_normalize_sense);
1882
1883 /**
1884  * scsi_sense_desc_find - search for a given descriptor type in
1885  *                      descriptor sense data format.
1886  *
1887  * @sense_buffer:       byte array of descriptor format sense data
1888  * @sb_len:             number of valid bytes in sense_buffer
1889  * @desc_type:          value of descriptor type to find
1890  *                      (e.g. 0 -> information)
1891  *
1892  * Notes:
1893  *      only valid when sense data is in descriptor format
1894  *
1895  * Return value:
1896  *      pointer to start of (first) descriptor if found else NULL
1897  **/
1898 const u8 * scsi_sense_desc_find(const u8 * sense_buffer, int sb_len,
1899                                 int desc_type)
1900 {
1901         int add_sen_len, add_len, desc_len, k;
1902         const u8 * descp;
1903
1904         if ((sb_len < 8) || (0 == (add_sen_len = sense_buffer[7])))
1905                 return NULL;
1906         if ((sense_buffer[0] < 0x72) || (sense_buffer[0] > 0x73))
1907                 return NULL;
1908         add_sen_len = (add_sen_len < (sb_len - 8)) ?
1909                         add_sen_len : (sb_len - 8);
1910         descp = &sense_buffer[8];
1911         for (desc_len = 0, k = 0; k < add_sen_len; k += desc_len) {
1912                 descp += desc_len;
1913                 add_len = (k < (add_sen_len - 1)) ? descp[1]: -1;
1914                 desc_len = add_len + 2;
1915                 if (descp[0] == desc_type)
1916                         return descp;
1917                 if (add_len < 0) // short descriptor ??
1918                         break;
1919         }
1920         return NULL;
1921 }
1922 EXPORT_SYMBOL(scsi_sense_desc_find);
1923
1924 /**
1925  * scsi_get_sense_info_fld - attempts to get information field from
1926  *                      sense data (either fixed or descriptor format)
1927  *
1928  * @sense_buffer:       byte array of sense data
1929  * @sb_len:             number of valid bytes in sense_buffer
1930  * @info_out:           pointer to 64 integer where 8 or 4 byte information
1931  *                      field will be placed if found.
1932  *
1933  * Return value:
1934  *      1 if information field found, 0 if not found.
1935  **/
1936 int scsi_get_sense_info_fld(const u8 * sense_buffer, int sb_len,
1937                             u64 * info_out)
1938 {
1939         int j;
1940         const u8 * ucp;
1941         u64 ull;
1942
1943         if (sb_len < 7)
1944                 return 0;
1945         switch (sense_buffer[0] & 0x7f) {
1946         case 0x70:
1947         case 0x71:
1948                 if (sense_buffer[0] & 0x80) {
1949                         *info_out = (sense_buffer[3] << 24) +
1950                                     (sense_buffer[4] << 16) +
1951                                     (sense_buffer[5] << 8) + sense_buffer[6];
1952                         return 1;
1953                 } else
1954                         return 0;
1955         case 0x72:
1956         case 0x73:
1957                 ucp = scsi_sense_desc_find(sense_buffer, sb_len,
1958                                            0 /* info desc */);
1959                 if (ucp && (0xa == ucp[1])) {
1960                         ull = 0;
1961                         for (j = 0; j < 8; ++j) {
1962                                 if (j > 0)
1963                                         ull <<= 8;
1964                                 ull |= ucp[4 + j];
1965                         }
1966                         *info_out = ull;
1967                         return 1;
1968                 } else
1969                         return 0;
1970         default:
1971                 return 0;
1972         }
1973 }
1974 EXPORT_SYMBOL(scsi_get_sense_info_fld);