Merge master.kernel.org:/pub/scm/linux/kernel/git/davej/cpufreq
[pandora-kernel.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/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         set_bit(SHOST_RECOVERY, &shost->shost_state);
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, (!test_bit(SHOST_RECOVERY, &sdev->host->shost_state)));
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 1;
782 }
783
784 /**
785  * scsi_eh_abort_cmds - abort canceled commands.
786  * @shost:      scsi host being recovered.
787  * @eh_done_q:  list_head for processed commands.
788  *
789  * Decription:
790  *    Try and see whether or not it makes sense to try and abort the
791  *    running command.  this only works out to be the case if we have one
792  *    command that has timed out.  if the command simply failed, it makes
793  *    no sense to try and abort the command, since as far as the shost
794  *    adapter is concerned, it isn't running.
795  **/
796 static int scsi_eh_abort_cmds(struct list_head *work_q,
797                               struct list_head *done_q)
798 {
799         struct scsi_cmnd *scmd, *next;
800         int rtn;
801
802         list_for_each_entry_safe(scmd, next, work_q, eh_entry) {
803                 if (!(scmd->eh_eflags & SCSI_EH_CANCEL_CMD))
804                         continue;
805                 SCSI_LOG_ERROR_RECOVERY(3, printk("%s: aborting cmd:"
806                                                   "0x%p\n", current->comm,
807                                                   scmd));
808                 rtn = scsi_try_to_abort_cmd(scmd);
809                 if (rtn == SUCCESS) {
810                         scmd->eh_eflags &= ~SCSI_EH_CANCEL_CMD;
811                         if (!scsi_device_online(scmd->device) ||
812                             !scsi_eh_tur(scmd)) {
813                                 scsi_eh_finish_cmd(scmd, done_q);
814                         }
815                                 
816                 } else
817                         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: aborting"
818                                                           " cmd failed:"
819                                                           "0x%p\n",
820                                                           current->comm,
821                                                           scmd));
822         }
823
824         return list_empty(work_q);
825 }
826
827 /**
828  * scsi_try_bus_device_reset - Ask host to perform a BDR on a dev
829  * @scmd:       SCSI cmd used to send BDR       
830  *
831  * Notes:
832  *    There is no timeout for this operation.  if this operation is
833  *    unreliable for a given host, then the host itself needs to put a
834  *    timer on it, and set the host back to a consistent state prior to
835  *    returning.
836  **/
837 static int scsi_try_bus_device_reset(struct scsi_cmnd *scmd)
838 {
839         int rtn;
840
841         if (!scmd->device->host->hostt->eh_device_reset_handler)
842                 return FAILED;
843
844         rtn = scmd->device->host->hostt->eh_device_reset_handler(scmd);
845         if (rtn == SUCCESS) {
846                 scmd->device->was_reset = 1;
847                 scmd->device->expecting_cc_ua = 1;
848         }
849
850         return rtn;
851 }
852
853 /**
854  * scsi_eh_try_stu - Send START_UNIT to device.
855  * @scmd:       Scsi cmd to send START_UNIT
856  *
857  * Return value:
858  *    0 - Device is ready. 1 - Device NOT ready.
859  **/
860 static int scsi_eh_try_stu(struct scsi_cmnd *scmd)
861 {
862         static unsigned char stu_command[6] = {START_STOP, 0, 0, 0, 1, 0};
863         int rtn;
864         int saved_result;
865
866         if (!scmd->device->allow_restart)
867                 return 1;
868
869         memcpy(scmd->cmnd, stu_command, sizeof(stu_command));
870
871         /*
872          * zero the sense buffer.  the scsi spec mandates that any
873          * untransferred sense data should be interpreted as being zero.
874          */
875         memset(scmd->sense_buffer, 0, sizeof(scmd->sense_buffer));
876
877         saved_result = scmd->result;
878         scmd->request_buffer = NULL;
879         scmd->request_bufflen = 0;
880         scmd->use_sg = 0;
881         scmd->cmd_len = COMMAND_SIZE(scmd->cmnd[0]);
882         scmd->underflow = 0;
883         scmd->sc_data_direction = DMA_NONE;
884
885         rtn = scsi_send_eh_cmnd(scmd, START_UNIT_TIMEOUT);
886
887         /*
888          * when we eventually call scsi_finish, we really wish to complete
889          * the original request, so let's restore the original data. (db)
890          */
891         scsi_setup_cmd_retry(scmd);
892         scmd->result = saved_result;
893
894         /*
895          * hey, we are done.  let's look to see what happened.
896          */
897         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: scmd %p rtn %x\n",
898                 __FUNCTION__, scmd, rtn));
899         if (rtn == SUCCESS)
900                 return 0;
901         return 1;
902 }
903
904  /**
905  * scsi_eh_stu - send START_UNIT if needed
906  * @shost:      scsi host being recovered.
907  * @eh_done_q:  list_head for processed commands.
908  *
909  * Notes:
910  *    If commands are failing due to not ready, initializing command required,
911  *      try revalidating the device, which will end up sending a start unit. 
912  **/
913 static int scsi_eh_stu(struct Scsi_Host *shost,
914                               struct list_head *work_q,
915                               struct list_head *done_q)
916 {
917         struct scsi_cmnd *scmd, *stu_scmd, *next;
918         struct scsi_device *sdev;
919
920         shost_for_each_device(sdev, shost) {
921                 stu_scmd = NULL;
922                 list_for_each_entry(scmd, work_q, eh_entry)
923                         if (scmd->device == sdev && SCSI_SENSE_VALID(scmd) &&
924                             scsi_check_sense(scmd) == FAILED ) {
925                                 stu_scmd = scmd;
926                                 break;
927                         }
928
929                 if (!stu_scmd)
930                         continue;
931
932                 SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Sending START_UNIT to sdev:"
933                                                   " 0x%p\n", current->comm, sdev));
934
935                 if (!scsi_eh_try_stu(stu_scmd)) {
936                         if (!scsi_device_online(sdev) ||
937                             !scsi_eh_tur(stu_scmd)) {
938                                 list_for_each_entry_safe(scmd, next,
939                                                           work_q, eh_entry) {
940                                         if (scmd->device == sdev)
941                                                 scsi_eh_finish_cmd(scmd, done_q);
942                                 }
943                         }
944                 } else {
945                         SCSI_LOG_ERROR_RECOVERY(3,
946                                                 printk("%s: START_UNIT failed to sdev:"
947                                                        " 0x%p\n", current->comm, sdev));
948                 }
949         }
950
951         return list_empty(work_q);
952 }
953
954
955 /**
956  * scsi_eh_bus_device_reset - send bdr if needed
957  * @shost:      scsi host being recovered.
958  * @eh_done_q:  list_head for processed commands.
959  *
960  * Notes:
961  *    Try a bus device reset.  still, look to see whether we have multiple
962  *    devices that are jammed or not - if we have multiple devices, it
963  *    makes no sense to try bus_device_reset - we really would need to try
964  *    a bus_reset instead. 
965  **/
966 static int scsi_eh_bus_device_reset(struct Scsi_Host *shost,
967                                     struct list_head *work_q,
968                                     struct list_head *done_q)
969 {
970         struct scsi_cmnd *scmd, *bdr_scmd, *next;
971         struct scsi_device *sdev;
972         int rtn;
973
974         shost_for_each_device(sdev, shost) {
975                 bdr_scmd = NULL;
976                 list_for_each_entry(scmd, work_q, eh_entry)
977                         if (scmd->device == sdev) {
978                                 bdr_scmd = scmd;
979                                 break;
980                         }
981
982                 if (!bdr_scmd)
983                         continue;
984
985                 SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Sending BDR sdev:"
986                                                   " 0x%p\n", current->comm,
987                                                   sdev));
988                 rtn = scsi_try_bus_device_reset(bdr_scmd);
989                 if (rtn == SUCCESS) {
990                         if (!scsi_device_online(sdev) ||
991                             !scsi_eh_tur(bdr_scmd)) {
992                                 list_for_each_entry_safe(scmd, next,
993                                                          work_q, eh_entry) {
994                                         if (scmd->device == sdev)
995                                                 scsi_eh_finish_cmd(scmd,
996                                                                    done_q);
997                                 }
998                         }
999                 } else {
1000                         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: BDR"
1001                                                           " failed sdev:"
1002                                                           "0x%p\n",
1003                                                           current->comm,
1004                                                            sdev));
1005                 }
1006         }
1007
1008         return list_empty(work_q);
1009 }
1010
1011 /**
1012  * scsi_try_bus_reset - ask host to perform a bus reset
1013  * @scmd:       SCSI cmd to send bus reset.
1014  **/
1015 static int scsi_try_bus_reset(struct scsi_cmnd *scmd)
1016 {
1017         unsigned long flags;
1018         int rtn;
1019
1020         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Snd Bus RST\n",
1021                                           __FUNCTION__));
1022
1023         if (!scmd->device->host->hostt->eh_bus_reset_handler)
1024                 return FAILED;
1025
1026         rtn = scmd->device->host->hostt->eh_bus_reset_handler(scmd);
1027
1028         if (rtn == SUCCESS) {
1029                 if (!scmd->device->host->hostt->skip_settle_delay)
1030                         ssleep(BUS_RESET_SETTLE_TIME);
1031                 spin_lock_irqsave(scmd->device->host->host_lock, flags);
1032                 scsi_report_bus_reset(scmd->device->host, scmd->device->channel);
1033                 spin_unlock_irqrestore(scmd->device->host->host_lock, flags);
1034         }
1035
1036         return rtn;
1037 }
1038
1039 /**
1040  * scsi_try_host_reset - ask host adapter to reset itself
1041  * @scmd:       SCSI cmd to send hsot reset.
1042  **/
1043 static int scsi_try_host_reset(struct scsi_cmnd *scmd)
1044 {
1045         unsigned long flags;
1046         int rtn;
1047
1048         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Snd Host RST\n",
1049                                           __FUNCTION__));
1050
1051         if (!scmd->device->host->hostt->eh_host_reset_handler)
1052                 return FAILED;
1053
1054         rtn = scmd->device->host->hostt->eh_host_reset_handler(scmd);
1055
1056         if (rtn == SUCCESS) {
1057                 if (!scmd->device->host->hostt->skip_settle_delay)
1058                         ssleep(HOST_RESET_SETTLE_TIME);
1059                 spin_lock_irqsave(scmd->device->host->host_lock, flags);
1060                 scsi_report_bus_reset(scmd->device->host, scmd->device->channel);
1061                 spin_unlock_irqrestore(scmd->device->host->host_lock, flags);
1062         }
1063
1064         return rtn;
1065 }
1066
1067 /**
1068  * scsi_eh_bus_reset - send a bus reset 
1069  * @shost:      scsi host being recovered.
1070  * @eh_done_q:  list_head for processed commands.
1071  **/
1072 static int scsi_eh_bus_reset(struct Scsi_Host *shost,
1073                              struct list_head *work_q,
1074                              struct list_head *done_q)
1075 {
1076         struct scsi_cmnd *scmd, *chan_scmd, *next;
1077         unsigned int channel;
1078         int rtn;
1079
1080         /*
1081          * we really want to loop over the various channels, and do this on
1082          * a channel by channel basis.  we should also check to see if any
1083          * of the failed commands are on soft_reset devices, and if so, skip
1084          * the reset.  
1085          */
1086
1087         for (channel = 0; channel <= shost->max_channel; channel++) {
1088                 chan_scmd = NULL;
1089                 list_for_each_entry(scmd, work_q, eh_entry) {
1090                         if (channel == scmd->device->channel) {
1091                                 chan_scmd = scmd;
1092                                 break;
1093                                 /*
1094                                  * FIXME add back in some support for
1095                                  * soft_reset devices.
1096                                  */
1097                         }
1098                 }
1099
1100                 if (!chan_scmd)
1101                         continue;
1102                 SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Sending BRST chan:"
1103                                                   " %d\n", current->comm,
1104                                                   channel));
1105                 rtn = scsi_try_bus_reset(chan_scmd);
1106                 if (rtn == SUCCESS) {
1107                         list_for_each_entry_safe(scmd, next, work_q, eh_entry) {
1108                                 if (channel == scmd->device->channel)
1109                                         if (!scsi_device_online(scmd->device) ||
1110                                             !scsi_eh_tur(scmd))
1111                                                 scsi_eh_finish_cmd(scmd,
1112                                                                    done_q);
1113                         }
1114                 } else {
1115                         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: BRST"
1116                                                           " failed chan: %d\n",
1117                                                           current->comm,
1118                                                           channel));
1119                 }
1120         }
1121         return list_empty(work_q);
1122 }
1123
1124 /**
1125  * scsi_eh_host_reset - send a host reset 
1126  * @work_q:     list_head for processed commands.
1127  * @done_q:     list_head for processed commands.
1128  **/
1129 static int scsi_eh_host_reset(struct list_head *work_q,
1130                               struct list_head *done_q)
1131 {
1132         struct scsi_cmnd *scmd, *next;
1133         int rtn;
1134
1135         if (!list_empty(work_q)) {
1136                 scmd = list_entry(work_q->next,
1137                                   struct scsi_cmnd, eh_entry);
1138
1139                 SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Sending HRST\n"
1140                                                   , current->comm));
1141
1142                 rtn = scsi_try_host_reset(scmd);
1143                 if (rtn == SUCCESS) {
1144                         list_for_each_entry_safe(scmd, next, work_q, eh_entry) {
1145                                 if (!scsi_device_online(scmd->device) ||
1146                                     (!scsi_eh_try_stu(scmd) && !scsi_eh_tur(scmd)) ||
1147                                     !scsi_eh_tur(scmd))
1148                                         scsi_eh_finish_cmd(scmd, done_q);
1149                         }
1150                 } else {
1151                         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: HRST"
1152                                                           " failed\n",
1153                                                           current->comm));
1154                 }
1155         }
1156         return list_empty(work_q);
1157 }
1158
1159 /**
1160  * scsi_eh_offline_sdevs - offline scsi devices that fail to recover
1161  * @work_q:     list_head for processed commands.
1162  * @done_q:     list_head for processed commands.
1163  *
1164  **/
1165 static void scsi_eh_offline_sdevs(struct list_head *work_q,
1166                                   struct list_head *done_q)
1167 {
1168         struct scsi_cmnd *scmd, *next;
1169
1170         list_for_each_entry_safe(scmd, next, work_q, eh_entry) {
1171                 printk(KERN_INFO "scsi: Device offlined - not"
1172                                 " ready after error recovery: host"
1173                                 " %d channel %d id %d lun %d\n",
1174                                 scmd->device->host->host_no,
1175                                 scmd->device->channel,
1176                                 scmd->device->id,
1177                                 scmd->device->lun);
1178                 scsi_device_set_state(scmd->device, SDEV_OFFLINE);
1179                 if (scmd->eh_eflags & SCSI_EH_CANCEL_CMD) {
1180                         /*
1181                          * FIXME: Handle lost cmds.
1182                          */
1183                 }
1184                 scsi_eh_finish_cmd(scmd, done_q);
1185         }
1186         return;
1187 }
1188
1189 /**
1190  * scsi_decide_disposition - Disposition a cmd on return from LLD.
1191  * @scmd:       SCSI cmd to examine.
1192  *
1193  * Notes:
1194  *    This is *only* called when we are examining the status after sending
1195  *    out the actual data command.  any commands that are queued for error
1196  *    recovery (e.g. test_unit_ready) do *not* come through here.
1197  *
1198  *    When this routine returns failed, it means the error handler thread
1199  *    is woken.  In cases where the error code indicates an error that
1200  *    doesn't require the error handler read (i.e. we don't need to
1201  *    abort/reset), this function should return SUCCESS.
1202  **/
1203 int scsi_decide_disposition(struct scsi_cmnd *scmd)
1204 {
1205         int rtn;
1206
1207         /*
1208          * if the device is offline, then we clearly just pass the result back
1209          * up to the top level.
1210          */
1211         if (!scsi_device_online(scmd->device)) {
1212                 SCSI_LOG_ERROR_RECOVERY(5, printk("%s: device offline - report"
1213                                                   " as SUCCESS\n",
1214                                                   __FUNCTION__));
1215                 return SUCCESS;
1216         }
1217
1218         /*
1219          * first check the host byte, to see if there is anything in there
1220          * that would indicate what we need to do.
1221          */
1222         switch (host_byte(scmd->result)) {
1223         case DID_PASSTHROUGH:
1224                 /*
1225                  * no matter what, pass this through to the upper layer.
1226                  * nuke this special code so that it looks like we are saying
1227                  * did_ok.
1228                  */
1229                 scmd->result &= 0xff00ffff;
1230                 return SUCCESS;
1231         case DID_OK:
1232                 /*
1233                  * looks good.  drop through, and check the next byte.
1234                  */
1235                 break;
1236         case DID_NO_CONNECT:
1237         case DID_BAD_TARGET:
1238         case DID_ABORT:
1239                 /*
1240                  * note - this means that we just report the status back
1241                  * to the top level driver, not that we actually think
1242                  * that it indicates SUCCESS.
1243                  */
1244                 return SUCCESS;
1245                 /*
1246                  * when the low level driver returns did_soft_error,
1247                  * it is responsible for keeping an internal retry counter 
1248                  * in order to avoid endless loops (db)
1249                  *
1250                  * actually this is a bug in this function here.  we should
1251                  * be mindful of the maximum number of retries specified
1252                  * and not get stuck in a loop.
1253                  */
1254         case DID_SOFT_ERROR:
1255                 goto maybe_retry;
1256         case DID_IMM_RETRY:
1257                 return NEEDS_RETRY;
1258
1259         case DID_REQUEUE:
1260                 return ADD_TO_MLQUEUE;
1261
1262         case DID_ERROR:
1263                 if (msg_byte(scmd->result) == COMMAND_COMPLETE &&
1264                     status_byte(scmd->result) == RESERVATION_CONFLICT)
1265                         /*
1266                          * execute reservation conflict processing code
1267                          * lower down
1268                          */
1269                         break;
1270                 /* fallthrough */
1271
1272         case DID_BUS_BUSY:
1273         case DID_PARITY:
1274                 goto maybe_retry;
1275         case DID_TIME_OUT:
1276                 /*
1277                  * when we scan the bus, we get timeout messages for
1278                  * these commands if there is no device available.
1279                  * other hosts report did_no_connect for the same thing.
1280                  */
1281                 if ((scmd->cmnd[0] == TEST_UNIT_READY ||
1282                      scmd->cmnd[0] == INQUIRY)) {
1283                         return SUCCESS;
1284                 } else {
1285                         return FAILED;
1286                 }
1287         case DID_RESET:
1288                 return SUCCESS;
1289         default:
1290                 return FAILED;
1291         }
1292
1293         /*
1294          * next, check the message byte.
1295          */
1296         if (msg_byte(scmd->result) != COMMAND_COMPLETE)
1297                 return FAILED;
1298
1299         /*
1300          * check the status byte to see if this indicates anything special.
1301          */
1302         switch (status_byte(scmd->result)) {
1303         case QUEUE_FULL:
1304                 /*
1305                  * the case of trying to send too many commands to a
1306                  * tagged queueing device.
1307                  */
1308         case BUSY:
1309                 /*
1310                  * device can't talk to us at the moment.  Should only
1311                  * occur (SAM-3) when the task queue is empty, so will cause
1312                  * the empty queue handling to trigger a stall in the
1313                  * device.
1314                  */
1315                 return ADD_TO_MLQUEUE;
1316         case GOOD:
1317         case COMMAND_TERMINATED:
1318         case TASK_ABORTED:
1319                 return SUCCESS;
1320         case CHECK_CONDITION:
1321                 rtn = scsi_check_sense(scmd);
1322                 if (rtn == NEEDS_RETRY)
1323                         goto maybe_retry;
1324                 /* if rtn == FAILED, we have no sense information;
1325                  * returning FAILED will wake the error handler thread
1326                  * to collect the sense and redo the decide
1327                  * disposition */
1328                 return rtn;
1329         case CONDITION_GOOD:
1330         case INTERMEDIATE_GOOD:
1331         case INTERMEDIATE_C_GOOD:
1332         case ACA_ACTIVE:
1333                 /*
1334                  * who knows?  FIXME(eric)
1335                  */
1336                 return SUCCESS;
1337
1338         case RESERVATION_CONFLICT:
1339                 printk(KERN_INFO "scsi: reservation conflict: host"
1340                                 " %d channel %d id %d lun %d\n",
1341                        scmd->device->host->host_no, scmd->device->channel,
1342                        scmd->device->id, scmd->device->lun);
1343                 return SUCCESS; /* causes immediate i/o error */
1344         default:
1345                 return FAILED;
1346         }
1347         return FAILED;
1348
1349       maybe_retry:
1350
1351         /* we requeue for retry because the error was retryable, and
1352          * the request was not marked fast fail.  Note that above,
1353          * even if the request is marked fast fail, we still requeue
1354          * for queue congestion conditions (QUEUE_FULL or BUSY) */
1355         if ((++scmd->retries) < scmd->allowed 
1356             && !blk_noretry_request(scmd->request)) {
1357                 return NEEDS_RETRY;
1358         } else {
1359                 /*
1360                  * no more retries - report this one back to upper level.
1361                  */
1362                 return SUCCESS;
1363         }
1364 }
1365
1366 /**
1367  * scsi_eh_lock_done - done function for eh door lock request
1368  * @scmd:       SCSI command block for the door lock request
1369  *
1370  * Notes:
1371  *      We completed the asynchronous door lock request, and it has either
1372  *      locked the door or failed.  We must free the command structures
1373  *      associated with this request.
1374  **/
1375 static void scsi_eh_lock_done(struct scsi_cmnd *scmd)
1376 {
1377         struct scsi_request *sreq = scmd->sc_request;
1378
1379         scsi_release_request(sreq);
1380 }
1381
1382
1383 /**
1384  * scsi_eh_lock_door - Prevent medium removal for the specified device
1385  * @sdev:       SCSI device to prevent medium removal
1386  *
1387  * Locking:
1388  *      We must be called from process context; scsi_allocate_request()
1389  *      may sleep.
1390  *
1391  * Notes:
1392  *      We queue up an asynchronous "ALLOW MEDIUM REMOVAL" request on the
1393  *      head of the devices request queue, and continue.
1394  *
1395  * Bugs:
1396  *      scsi_allocate_request() may sleep waiting for existing requests to
1397  *      be processed.  However, since we haven't kicked off any request
1398  *      processing for this host, this may deadlock.
1399  *
1400  *      If scsi_allocate_request() fails for what ever reason, we
1401  *      completely forget to lock the door.
1402  **/
1403 static void scsi_eh_lock_door(struct scsi_device *sdev)
1404 {
1405         struct scsi_request *sreq = scsi_allocate_request(sdev, GFP_KERNEL);
1406
1407         if (unlikely(!sreq)) {
1408                 printk(KERN_ERR "%s: request allocate failed,"
1409                        "prevent media removal cmd not sent\n", __FUNCTION__);
1410                 return;
1411         }
1412
1413         sreq->sr_cmnd[0] = ALLOW_MEDIUM_REMOVAL;
1414         sreq->sr_cmnd[1] = 0;
1415         sreq->sr_cmnd[2] = 0;
1416         sreq->sr_cmnd[3] = 0;
1417         sreq->sr_cmnd[4] = SCSI_REMOVAL_PREVENT;
1418         sreq->sr_cmnd[5] = 0;
1419         sreq->sr_data_direction = DMA_NONE;
1420         sreq->sr_bufflen = 0;
1421         sreq->sr_buffer = NULL;
1422         sreq->sr_allowed = 5;
1423         sreq->sr_done = scsi_eh_lock_done;
1424         sreq->sr_timeout_per_command = 10 * HZ;
1425         sreq->sr_cmd_len = COMMAND_SIZE(sreq->sr_cmnd[0]);
1426
1427         scsi_insert_special_req(sreq, 1);
1428 }
1429
1430
1431 /**
1432  * scsi_restart_operations - restart io operations to the specified host.
1433  * @shost:      Host we are restarting.
1434  *
1435  * Notes:
1436  *    When we entered the error handler, we blocked all further i/o to
1437  *    this device.  we need to 'reverse' this process.
1438  **/
1439 static void scsi_restart_operations(struct Scsi_Host *shost)
1440 {
1441         struct scsi_device *sdev;
1442
1443         /*
1444          * If the door was locked, we need to insert a door lock request
1445          * onto the head of the SCSI request queue for the device.  There
1446          * is no point trying to lock the door of an off-line device.
1447          */
1448         shost_for_each_device(sdev, shost) {
1449                 if (scsi_device_online(sdev) && sdev->locked)
1450                         scsi_eh_lock_door(sdev);
1451         }
1452
1453         /*
1454          * next free up anything directly waiting upon the host.  this
1455          * will be requests for character device operations, and also for
1456          * ioctls to queued block devices.
1457          */
1458         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: waking up host to restart\n",
1459                                           __FUNCTION__));
1460
1461         clear_bit(SHOST_RECOVERY, &shost->shost_state);
1462
1463         wake_up(&shost->host_wait);
1464
1465         /*
1466          * finally we need to re-initiate requests that may be pending.  we will
1467          * have had everything blocked while error handling is taking place, and
1468          * now that error recovery is done, we will need to ensure that these
1469          * requests are started.
1470          */
1471         scsi_run_host_queues(shost);
1472 }
1473
1474 /**
1475  * scsi_eh_ready_devs - check device ready state and recover if not.
1476  * @shost:      host to be recovered.
1477  * @eh_done_q:  list_head for processed commands.
1478  *
1479  **/
1480 static void scsi_eh_ready_devs(struct Scsi_Host *shost,
1481                                struct list_head *work_q,
1482                                struct list_head *done_q)
1483 {
1484         if (!scsi_eh_stu(shost, work_q, done_q))
1485                 if (!scsi_eh_bus_device_reset(shost, work_q, done_q))
1486                         if (!scsi_eh_bus_reset(shost, work_q, done_q))
1487                                 if (!scsi_eh_host_reset(work_q, done_q))
1488                                         scsi_eh_offline_sdevs(work_q, done_q);
1489 }
1490
1491 /**
1492  * scsi_eh_flush_done_q - finish processed commands or retry them.
1493  * @done_q:     list_head of processed commands.
1494  *
1495  **/
1496 static void scsi_eh_flush_done_q(struct list_head *done_q)
1497 {
1498         struct scsi_cmnd *scmd, *next;
1499
1500         list_for_each_entry_safe(scmd, next, done_q, eh_entry) {
1501                 list_del_init(&scmd->eh_entry);
1502                 if (scsi_device_online(scmd->device) &&
1503                     !blk_noretry_request(scmd->request) &&
1504                     (++scmd->retries < scmd->allowed)) {
1505                         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: flush"
1506                                                           " retry cmd: %p\n",
1507                                                           current->comm,
1508                                                           scmd));
1509                                 scsi_queue_insert(scmd, SCSI_MLQUEUE_EH_RETRY);
1510                 } else {
1511                         /*
1512                          * If just we got sense for the device (called
1513                          * scsi_eh_get_sense), scmd->result is already
1514                          * set, do not set DRIVER_TIMEOUT.
1515                          */
1516                         if (!scmd->result)
1517                                 scmd->result |= (DRIVER_TIMEOUT << 24);
1518                         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: flush finish"
1519                                                         " cmd: %p\n",
1520                                                         current->comm, scmd));
1521                         scsi_finish_command(scmd);
1522                 }
1523         }
1524 }
1525
1526 /**
1527  * scsi_unjam_host - Attempt to fix a host which has a cmd that failed.
1528  * @shost:      Host to unjam.
1529  *
1530  * Notes:
1531  *    When we come in here, we *know* that all commands on the bus have
1532  *    either completed, failed or timed out.  we also know that no further
1533  *    commands are being sent to the host, so things are relatively quiet
1534  *    and we have freedom to fiddle with things as we wish.
1535  *
1536  *    This is only the *default* implementation.  it is possible for
1537  *    individual drivers to supply their own version of this function, and
1538  *    if the maintainer wishes to do this, it is strongly suggested that
1539  *    this function be taken as a template and modified.  this function
1540  *    was designed to correctly handle problems for about 95% of the
1541  *    different cases out there, and it should always provide at least a
1542  *    reasonable amount of error recovery.
1543  *
1544  *    Any command marked 'failed' or 'timeout' must eventually have
1545  *    scsi_finish_cmd() called for it.  we do all of the retry stuff
1546  *    here, so when we restart the host after we return it should have an
1547  *    empty queue.
1548  **/
1549 static void scsi_unjam_host(struct Scsi_Host *shost)
1550 {
1551         unsigned long flags;
1552         LIST_HEAD(eh_work_q);
1553         LIST_HEAD(eh_done_q);
1554
1555         spin_lock_irqsave(shost->host_lock, flags);
1556         list_splice_init(&shost->eh_cmd_q, &eh_work_q);
1557         spin_unlock_irqrestore(shost->host_lock, flags);
1558
1559         SCSI_LOG_ERROR_RECOVERY(1, scsi_eh_prt_fail_stats(shost, &eh_work_q));
1560
1561         if (!scsi_eh_get_sense(&eh_work_q, &eh_done_q))
1562                 if (!scsi_eh_abort_cmds(&eh_work_q, &eh_done_q))
1563                         scsi_eh_ready_devs(shost, &eh_work_q, &eh_done_q);
1564
1565         scsi_eh_flush_done_q(&eh_done_q);
1566 }
1567
1568 /**
1569  * scsi_error_handler - Handle errors/timeouts of SCSI cmds.
1570  * @data:       Host for which we are running.
1571  *
1572  * Notes:
1573  *    This is always run in the context of a kernel thread.  The idea is
1574  *    that we start this thing up when the kernel starts up (one per host
1575  *    that we detect), and it immediately goes to sleep and waits for some
1576  *    event (i.e. failure).  When this takes place, we have the job of
1577  *    trying to unjam the bus and restarting things.
1578  **/
1579 int scsi_error_handler(void *data)
1580 {
1581         struct Scsi_Host *shost = (struct Scsi_Host *) data;
1582         int rtn;
1583         DECLARE_MUTEX_LOCKED(sem);
1584
1585         /*
1586          *    Flush resources
1587          */
1588
1589         daemonize("scsi_eh_%d", shost->host_no);
1590
1591         current->flags |= PF_NOFREEZE;
1592
1593         shost->eh_wait = &sem;
1594         shost->ehandler = current;
1595
1596         /*
1597          * Wake up the thread that created us.
1598          */
1599         SCSI_LOG_ERROR_RECOVERY(3, printk("Wake up parent of"
1600                                           " scsi_eh_%d\n",shost->host_no));
1601
1602         complete(shost->eh_notify);
1603
1604         while (1) {
1605                 /*
1606                  * If we get a signal, it means we are supposed to go
1607                  * away and die.  This typically happens if the user is
1608                  * trying to unload a module.
1609                  */
1610                 SCSI_LOG_ERROR_RECOVERY(1, printk("Error handler"
1611                                                   " scsi_eh_%d"
1612                                                   " sleeping\n",shost->host_no));
1613
1614                 /*
1615                  * Note - we always use down_interruptible with the semaphore
1616                  * even if the module was loaded as part of the kernel.  The
1617                  * reason is that down() will cause this thread to be counted
1618                  * in the load average as a running process, and down
1619                  * interruptible doesn't.  Given that we need to allow this
1620                  * thread to die if the driver was loaded as a module, using
1621                  * semaphores isn't unreasonable.
1622                  */
1623                 down_interruptible(&sem);
1624                 if (shost->eh_kill)
1625                         break;
1626
1627                 SCSI_LOG_ERROR_RECOVERY(1, printk("Error handler"
1628                                                   " scsi_eh_%d waking"
1629                                                   " up\n",shost->host_no));
1630
1631                 shost->eh_active = 1;
1632
1633                 /*
1634                  * We have a host that is failing for some reason.  Figure out
1635                  * what we need to do to get it up and online again (if we can).
1636                  * If we fail, we end up taking the thing offline.
1637                  */
1638                 if (shost->hostt->eh_strategy_handler) 
1639                         rtn = shost->hostt->eh_strategy_handler(shost);
1640                 else
1641                         scsi_unjam_host(shost);
1642
1643                 shost->eh_active = 0;
1644
1645                 /*
1646                  * Note - if the above fails completely, the action is to take
1647                  * individual devices offline and flush the queue of any
1648                  * outstanding requests that may have been pending.  When we
1649                  * restart, we restart any I/O to any other devices on the bus
1650                  * which are still online.
1651                  */
1652                 scsi_restart_operations(shost);
1653
1654         }
1655
1656         SCSI_LOG_ERROR_RECOVERY(1, printk("Error handler scsi_eh_%d"
1657                                           " exiting\n",shost->host_no));
1658
1659         /*
1660          * Make sure that nobody tries to wake us up again.
1661          */
1662         shost->eh_wait = NULL;
1663
1664         /*
1665          * Knock this down too.  From this point on, the host is flying
1666          * without a pilot.  If this is because the module is being unloaded,
1667          * that's fine.  If the user sent a signal to this thing, we are
1668          * potentially in real danger.
1669          */
1670         shost->eh_active = 0;
1671         shost->ehandler = NULL;
1672
1673         /*
1674          * If anyone is waiting for us to exit (i.e. someone trying to unload
1675          * a driver), then wake up that process to let them know we are on
1676          * the way out the door.
1677          */
1678         complete_and_exit(shost->eh_notify, 0);
1679         return 0;
1680 }
1681
1682 /*
1683  * Function:    scsi_report_bus_reset()
1684  *
1685  * Purpose:     Utility function used by low-level drivers to report that
1686  *              they have observed a bus reset on the bus being handled.
1687  *
1688  * Arguments:   shost       - Host in question
1689  *              channel     - channel on which reset was observed.
1690  *
1691  * Returns:     Nothing
1692  *
1693  * Lock status: Host lock must be held.
1694  *
1695  * Notes:       This only needs to be called if the reset is one which
1696  *              originates from an unknown location.  Resets originated
1697  *              by the mid-level itself don't need to call this, but there
1698  *              should be no harm.
1699  *
1700  *              The main purpose of this is to make sure that a CHECK_CONDITION
1701  *              is properly treated.
1702  */
1703 void scsi_report_bus_reset(struct Scsi_Host *shost, int channel)
1704 {
1705         struct scsi_device *sdev;
1706
1707         __shost_for_each_device(sdev, shost) {
1708                 if (channel == sdev->channel) {
1709                         sdev->was_reset = 1;
1710                         sdev->expecting_cc_ua = 1;
1711                 }
1712         }
1713 }
1714 EXPORT_SYMBOL(scsi_report_bus_reset);
1715
1716 /*
1717  * Function:    scsi_report_device_reset()
1718  *
1719  * Purpose:     Utility function used by low-level drivers to report that
1720  *              they have observed a device reset on the device being handled.
1721  *
1722  * Arguments:   shost       - Host in question
1723  *              channel     - channel on which reset was observed
1724  *              target      - target on which reset was observed
1725  *
1726  * Returns:     Nothing
1727  *
1728  * Lock status: Host lock must be held
1729  *
1730  * Notes:       This only needs to be called if the reset is one which
1731  *              originates from an unknown location.  Resets originated
1732  *              by the mid-level itself don't need to call this, but there
1733  *              should be no harm.
1734  *
1735  *              The main purpose of this is to make sure that a CHECK_CONDITION
1736  *              is properly treated.
1737  */
1738 void scsi_report_device_reset(struct Scsi_Host *shost, int channel, int target)
1739 {
1740         struct scsi_device *sdev;
1741
1742         __shost_for_each_device(sdev, shost) {
1743                 if (channel == sdev->channel &&
1744                     target == sdev->id) {
1745                         sdev->was_reset = 1;
1746                         sdev->expecting_cc_ua = 1;
1747                 }
1748         }
1749 }
1750 EXPORT_SYMBOL(scsi_report_device_reset);
1751
1752 static void
1753 scsi_reset_provider_done_command(struct scsi_cmnd *scmd)
1754 {
1755 }
1756
1757 /*
1758  * Function:    scsi_reset_provider
1759  *
1760  * Purpose:     Send requested reset to a bus or device at any phase.
1761  *
1762  * Arguments:   device  - device to send reset to
1763  *              flag - reset type (see scsi.h)
1764  *
1765  * Returns:     SUCCESS/FAILURE.
1766  *
1767  * Notes:       This is used by the SCSI Generic driver to provide
1768  *              Bus/Device reset capability.
1769  */
1770 int
1771 scsi_reset_provider(struct scsi_device *dev, int flag)
1772 {
1773         struct scsi_cmnd *scmd = scsi_get_command(dev, GFP_KERNEL);
1774         struct request req;
1775         int rtn;
1776
1777         scmd->request = &req;
1778         memset(&scmd->eh_timeout, 0, sizeof(scmd->eh_timeout));
1779         scmd->request->rq_status        = RQ_SCSI_BUSY;
1780
1781         memset(&scmd->cmnd, '\0', sizeof(scmd->cmnd));
1782     
1783         scmd->scsi_done         = scsi_reset_provider_done_command;
1784         scmd->done                      = NULL;
1785         scmd->buffer                    = NULL;
1786         scmd->bufflen                   = 0;
1787         scmd->request_buffer            = NULL;
1788         scmd->request_bufflen           = 0;
1789
1790         scmd->cmd_len                   = 0;
1791
1792         scmd->sc_data_direction         = DMA_BIDIRECTIONAL;
1793         scmd->sc_request                = NULL;
1794         scmd->sc_magic                  = SCSI_CMND_MAGIC;
1795
1796         init_timer(&scmd->eh_timeout);
1797
1798         /*
1799          * Sometimes the command can get back into the timer chain,
1800          * so use the pid as an identifier.
1801          */
1802         scmd->pid                       = 0;
1803
1804         switch (flag) {
1805         case SCSI_TRY_RESET_DEVICE:
1806                 rtn = scsi_try_bus_device_reset(scmd);
1807                 if (rtn == SUCCESS)
1808                         break;
1809                 /* FALLTHROUGH */
1810         case SCSI_TRY_RESET_BUS:
1811                 rtn = scsi_try_bus_reset(scmd);
1812                 if (rtn == SUCCESS)
1813                         break;
1814                 /* FALLTHROUGH */
1815         case SCSI_TRY_RESET_HOST:
1816                 rtn = scsi_try_host_reset(scmd);
1817                 break;
1818         default:
1819                 rtn = FAILED;
1820         }
1821
1822         scsi_next_command(scmd);
1823         return rtn;
1824 }
1825 EXPORT_SYMBOL(scsi_reset_provider);
1826
1827 /**
1828  * scsi_normalize_sense - normalize main elements from either fixed or
1829  *                      descriptor sense data format into a common format.
1830  *
1831  * @sense_buffer:       byte array containing sense data returned by device
1832  * @sb_len:             number of valid bytes in sense_buffer
1833  * @sshdr:              pointer to instance of structure that common
1834  *                      elements are written to.
1835  *
1836  * Notes:
1837  *      The "main elements" from sense data are: response_code, sense_key,
1838  *      asc, ascq and additional_length (only for descriptor format).
1839  *
1840  *      Typically this function can be called after a device has
1841  *      responded to a SCSI command with the CHECK_CONDITION status.
1842  *
1843  * Return value:
1844  *      1 if valid sense data information found, else 0;
1845  **/
1846 int scsi_normalize_sense(const u8 *sense_buffer, int sb_len,
1847                          struct scsi_sense_hdr *sshdr)
1848 {
1849         if (!sense_buffer || !sb_len || (sense_buffer[0] & 0x70) != 0x70)
1850                 return 0;
1851
1852         memset(sshdr, 0, sizeof(struct scsi_sense_hdr));
1853
1854         sshdr->response_code = (sense_buffer[0] & 0x7f);
1855         if (sshdr->response_code >= 0x72) {
1856                 /*
1857                  * descriptor format
1858                  */
1859                 if (sb_len > 1)
1860                         sshdr->sense_key = (sense_buffer[1] & 0xf);
1861                 if (sb_len > 2)
1862                         sshdr->asc = sense_buffer[2];
1863                 if (sb_len > 3)
1864                         sshdr->ascq = sense_buffer[3];
1865                 if (sb_len > 7)
1866                         sshdr->additional_length = sense_buffer[7];
1867         } else {
1868                 /* 
1869                  * fixed format
1870                  */
1871                 if (sb_len > 2)
1872                         sshdr->sense_key = (sense_buffer[2] & 0xf);
1873                 if (sb_len > 7) {
1874                         sb_len = (sb_len < (sense_buffer[7] + 8)) ?
1875                                          sb_len : (sense_buffer[7] + 8);
1876                         if (sb_len > 12)
1877                                 sshdr->asc = sense_buffer[12];
1878                         if (sb_len > 13)
1879                                 sshdr->ascq = sense_buffer[13];
1880                 }
1881         }
1882
1883         return 1;
1884 }
1885 EXPORT_SYMBOL(scsi_normalize_sense);
1886
1887 int scsi_request_normalize_sense(struct scsi_request *sreq,
1888                                  struct scsi_sense_hdr *sshdr)
1889 {
1890         return scsi_normalize_sense(sreq->sr_sense_buffer,
1891                         sizeof(sreq->sr_sense_buffer), sshdr);
1892 }
1893 EXPORT_SYMBOL(scsi_request_normalize_sense);
1894
1895 int scsi_command_normalize_sense(struct scsi_cmnd *cmd,
1896                                  struct scsi_sense_hdr *sshdr)
1897 {
1898         return scsi_normalize_sense(cmd->sense_buffer,
1899                         sizeof(cmd->sense_buffer), sshdr);
1900 }
1901 EXPORT_SYMBOL(scsi_command_normalize_sense);
1902
1903 /**
1904  * scsi_sense_desc_find - search for a given descriptor type in
1905  *                      descriptor sense data format.
1906  *
1907  * @sense_buffer:       byte array of descriptor format sense data
1908  * @sb_len:             number of valid bytes in sense_buffer
1909  * @desc_type:          value of descriptor type to find
1910  *                      (e.g. 0 -> information)
1911  *
1912  * Notes:
1913  *      only valid when sense data is in descriptor format
1914  *
1915  * Return value:
1916  *      pointer to start of (first) descriptor if found else NULL
1917  **/
1918 const u8 * scsi_sense_desc_find(const u8 * sense_buffer, int sb_len,
1919                                 int desc_type)
1920 {
1921         int add_sen_len, add_len, desc_len, k;
1922         const u8 * descp;
1923
1924         if ((sb_len < 8) || (0 == (add_sen_len = sense_buffer[7])))
1925                 return NULL;
1926         if ((sense_buffer[0] < 0x72) || (sense_buffer[0] > 0x73))
1927                 return NULL;
1928         add_sen_len = (add_sen_len < (sb_len - 8)) ?
1929                         add_sen_len : (sb_len - 8);
1930         descp = &sense_buffer[8];
1931         for (desc_len = 0, k = 0; k < add_sen_len; k += desc_len) {
1932                 descp += desc_len;
1933                 add_len = (k < (add_sen_len - 1)) ? descp[1]: -1;
1934                 desc_len = add_len + 2;
1935                 if (descp[0] == desc_type)
1936                         return descp;
1937                 if (add_len < 0) // short descriptor ??
1938                         break;
1939         }
1940         return NULL;
1941 }
1942 EXPORT_SYMBOL(scsi_sense_desc_find);
1943
1944 /**
1945  * scsi_get_sense_info_fld - attempts to get information field from
1946  *                      sense data (either fixed or descriptor format)
1947  *
1948  * @sense_buffer:       byte array of sense data
1949  * @sb_len:             number of valid bytes in sense_buffer
1950  * @info_out:           pointer to 64 integer where 8 or 4 byte information
1951  *                      field will be placed if found.
1952  *
1953  * Return value:
1954  *      1 if information field found, 0 if not found.
1955  **/
1956 int scsi_get_sense_info_fld(const u8 * sense_buffer, int sb_len,
1957                             u64 * info_out)
1958 {
1959         int j;
1960         const u8 * ucp;
1961         u64 ull;
1962
1963         if (sb_len < 7)
1964                 return 0;
1965         switch (sense_buffer[0] & 0x7f) {
1966         case 0x70:
1967         case 0x71:
1968                 if (sense_buffer[0] & 0x80) {
1969                         *info_out = (sense_buffer[3] << 24) +
1970                                     (sense_buffer[4] << 16) +
1971                                     (sense_buffer[5] << 8) + sense_buffer[6];
1972                         return 1;
1973                 } else
1974                         return 0;
1975         case 0x72:
1976         case 0x73:
1977                 ucp = scsi_sense_desc_find(sense_buffer, sb_len,
1978                                            0 /* info desc */);
1979                 if (ucp && (0xa == ucp[1])) {
1980                         ull = 0;
1981                         for (j = 0; j < 8; ++j) {
1982                                 if (j > 0)
1983                                         ull <<= 8;
1984                                 ull |= ucp[4 + j];
1985                         }
1986                         *info_out = ull;
1987                         return 1;
1988                 } else
1989                         return 0;
1990         default:
1991                 return 0;
1992         }
1993 }
1994 EXPORT_SYMBOL(scsi_get_sense_info_fld);