52b348c36d56b7cc4be4a03e444e8ad182679620
[linux-2.6.git] / drivers / scsi / scsi_error.c
1 /*
2  *  scsi_error.c Copyright (C) 1997 Eric Youngdale
3  *
4  *  SCSI error/timeout handling
5  *      Initial versions: Eric Youngdale.  Based upon conversations with
6  *                        Leonard Zubkoff and David Miller at Linux Expo, 
7  *                        ideas originating from all over the place.
8  *
9  *      Restructured scsi_unjam_host and associated functions.
10  *      September 04, 2002 Mike Anderson (andmike@us.ibm.com)
11  *
12  *      Forward port of Russell King's (rmk@arm.linux.org.uk) changes and
13  *      minor  cleanups.
14  *      September 30, 2002 Mike Anderson (andmike@us.ibm.com)
15  */
16
17 #include <linux/module.h>
18 #include <linux/sched.h>
19 #include <linux/timer.h>
20 #include <linux/string.h>
21 #include <linux/slab.h>
22 #include <linux/kernel.h>
23 #include <linux/kthread.h>
24 #include <linux/interrupt.h>
25 #include <linux/blkdev.h>
26 #include <linux/delay.h>
27
28 #include <scsi/scsi.h>
29 #include <scsi/scsi_dbg.h>
30 #include <scsi/scsi_device.h>
31 #include <scsi/scsi_eh.h>
32 #include <scsi/scsi_host.h>
33 #include <scsi/scsi_ioctl.h>
34 #include <scsi/scsi_request.h>
35
36 #include "scsi_priv.h"
37 #include "scsi_logging.h"
38
39 #define SENSE_TIMEOUT           (10*HZ)
40 #define START_UNIT_TIMEOUT      (30*HZ)
41
42 /*
43  * These should *probably* be handled by the host itself.
44  * Since it is allowed to sleep, it probably should.
45  */
46 #define BUS_RESET_SETTLE_TIME   (10)
47 #define HOST_RESET_SETTLE_TIME  (10)
48
49 /* called with shost->host_lock held */
50 void scsi_eh_wakeup(struct Scsi_Host *shost)
51 {
52         if (shost->host_busy == shost->host_failed) {
53                 wake_up_process(shost->ehandler);
54                 SCSI_LOG_ERROR_RECOVERY(5,
55                                 printk("Waking error handler thread\n"));
56         }
57 }
58
59 /**
60  * scsi_eh_scmd_add - add scsi cmd to error handling.
61  * @scmd:       scmd to run eh on.
62  * @eh_flag:    optional SCSI_EH flag.
63  *
64  * Return value:
65  *      0 on failure.
66  **/
67 int scsi_eh_scmd_add(struct scsi_cmnd *scmd, int eh_flag)
68 {
69         struct Scsi_Host *shost = scmd->device->host;
70         unsigned long flags;
71         int ret = 0;
72
73         if (!shost->ehandler)
74                 return 0;
75
76         spin_lock_irqsave(shost->host_lock, flags);
77         if (scsi_host_set_state(shost, SHOST_RECOVERY))
78                 if (scsi_host_set_state(shost, SHOST_CANCEL_RECOVERY))
79                         goto out_unlock;
80
81         ret = 1;
82         scmd->eh_eflags |= eh_flag;
83         list_add_tail(&scmd->eh_entry, &shost->eh_cmd_q);
84         shost->host_failed++;
85         scsi_eh_wakeup(shost);
86  out_unlock:
87         spin_unlock_irqrestore(shost->host_lock, flags);
88         return ret;
89 }
90
91 /**
92  * scsi_add_timer - Start timeout timer for a single scsi command.
93  * @scmd:       scsi command that is about to start running.
94  * @timeout:    amount of time to allow this command to run.
95  * @complete:   timeout function to call if timer isn't canceled.
96  *
97  * Notes:
98  *    This should be turned into an inline function.  Each scsi command
99  *    has its own timer, and as it is added to the queue, we set up the
100  *    timer.  When the command completes, we cancel the timer.
101  **/
102 void scsi_add_timer(struct scsi_cmnd *scmd, int timeout,
103                     void (*complete)(struct scsi_cmnd *))
104 {
105
106         /*
107          * If the clock was already running for this command, then
108          * first delete the timer.  The timer handling code gets rather
109          * confused if we don't do this.
110          */
111         if (scmd->eh_timeout.function)
112                 del_timer(&scmd->eh_timeout);
113
114         scmd->eh_timeout.data = (unsigned long)scmd;
115         scmd->eh_timeout.expires = jiffies + timeout;
116         scmd->eh_timeout.function = (void (*)(unsigned long)) complete;
117
118         SCSI_LOG_ERROR_RECOVERY(5, printk("%s: scmd: %p, time:"
119                                           " %d, (%p)\n", __FUNCTION__,
120                                           scmd, timeout, complete));
121
122         add_timer(&scmd->eh_timeout);
123 }
124
125 /**
126  * scsi_delete_timer - Delete/cancel timer for a given function.
127  * @scmd:       Cmd that we are canceling timer for
128  *
129  * Notes:
130  *     This should be turned into an inline function.
131  *
132  * Return value:
133  *     1 if we were able to detach the timer.  0 if we blew it, and the
134  *     timer function has already started to run.
135  **/
136 int scsi_delete_timer(struct scsi_cmnd *scmd)
137 {
138         int rtn;
139
140         rtn = del_timer(&scmd->eh_timeout);
141
142         SCSI_LOG_ERROR_RECOVERY(5, printk("%s: scmd: %p,"
143                                          " rtn: %d\n", __FUNCTION__,
144                                          scmd, rtn));
145
146         scmd->eh_timeout.data = (unsigned long)NULL;
147         scmd->eh_timeout.function = NULL;
148
149         return rtn;
150 }
151
152 /**
153  * scsi_times_out - Timeout function for normal scsi commands.
154  * @scmd:       Cmd that is timing out.
155  *
156  * Notes:
157  *     We do not need to lock this.  There is the potential for a race
158  *     only in that the normal completion handling might run, but if the
159  *     normal completion function determines that the timer has already
160  *     fired, then it mustn't do anything.
161  **/
162 void scsi_times_out(struct scsi_cmnd *scmd)
163 {
164         scsi_log_completion(scmd, TIMEOUT_ERROR);
165
166         if (scmd->device->host->hostt->eh_timed_out)
167                 switch (scmd->device->host->hostt->eh_timed_out(scmd)) {
168                 case EH_HANDLED:
169                         __scsi_done(scmd);
170                         return;
171                 case EH_RESET_TIMER:
172                         /* This allows a single retry even of a command
173                          * with allowed == 0 */
174                         if (scmd->retries++ > scmd->allowed)
175                                 break;
176                         scsi_add_timer(scmd, scmd->timeout_per_command,
177                                        scsi_times_out);
178                         return;
179                 case EH_NOT_HANDLED:
180                         break;
181                 }
182
183         if (unlikely(!scsi_eh_scmd_add(scmd, SCSI_EH_CANCEL_CMD))) {
184                 scmd->result |= DID_TIME_OUT << 16;
185                 __scsi_done(scmd);
186         }
187 }
188
189 /**
190  * scsi_block_when_processing_errors - Prevent cmds from being queued.
191  * @sdev:       Device on which we are performing recovery.
192  *
193  * Description:
194  *     We block until the host is out of error recovery, and then check to
195  *     see whether the host or the device is offline.
196  *
197  * Return value:
198  *     0 when dev was taken offline by error recovery. 1 OK to proceed.
199  **/
200 int scsi_block_when_processing_errors(struct scsi_device *sdev)
201 {
202         int online;
203
204         wait_event(sdev->host->host_wait, !scsi_host_in_recovery(sdev->host));
205
206         online = scsi_device_online(sdev);
207
208         SCSI_LOG_ERROR_RECOVERY(5, printk("%s: rtn: %d\n", __FUNCTION__,
209                                           online));
210
211         return online;
212 }
213 EXPORT_SYMBOL(scsi_block_when_processing_errors);
214
215 #ifdef CONFIG_SCSI_LOGGING
216 /**
217  * scsi_eh_prt_fail_stats - Log info on failures.
218  * @shost:      scsi host being recovered.
219  * @work_q:     Queue of scsi cmds to process.
220  **/
221 static inline void scsi_eh_prt_fail_stats(struct Scsi_Host *shost,
222                                           struct list_head *work_q)
223 {
224         struct scsi_cmnd *scmd;
225         struct scsi_device *sdev;
226         int total_failures = 0;
227         int cmd_failed = 0;
228         int cmd_cancel = 0;
229         int devices_failed = 0;
230
231         shost_for_each_device(sdev, shost) {
232                 list_for_each_entry(scmd, work_q, eh_entry) {
233                         if (scmd->device == sdev) {
234                                 ++total_failures;
235                                 if (scmd->eh_eflags & SCSI_EH_CANCEL_CMD)
236                                         ++cmd_cancel;
237                                 else 
238                                         ++cmd_failed;
239                         }
240                 }
241
242                 if (cmd_cancel || cmd_failed) {
243                         SCSI_LOG_ERROR_RECOVERY(3,
244                                 printk("%s: %d:%d:%d:%d cmds failed: %d,"
245                                        " cancel: %d\n",
246                                        __FUNCTION__, shost->host_no,
247                                        sdev->channel, sdev->id, sdev->lun,
248                                        cmd_failed, cmd_cancel));
249                         cmd_cancel = 0;
250                         cmd_failed = 0;
251                         ++devices_failed;
252                 }
253         }
254
255         SCSI_LOG_ERROR_RECOVERY(2, printk("Total of %d commands on %d"
256                                           " devices require eh work\n",
257                                   total_failures, devices_failed));
258 }
259 #endif
260
261 /**
262  * scsi_check_sense - Examine scsi cmd sense
263  * @scmd:       Cmd to have sense checked.
264  *
265  * Return value:
266  *      SUCCESS or FAILED or NEEDS_RETRY
267  *
268  * Notes:
269  *      When a deferred error is detected the current command has
270  *      not been executed and needs retrying.
271  **/
272 static int scsi_check_sense(struct scsi_cmnd *scmd)
273 {
274         struct scsi_sense_hdr sshdr;
275
276         if (! scsi_command_normalize_sense(scmd, &sshdr))
277                 return FAILED;  /* no valid sense data */
278
279         if (scsi_sense_is_deferred(&sshdr))
280                 return NEEDS_RETRY;
281
282         /*
283          * Previous logic looked for FILEMARK, EOM or ILI which are
284          * mainly associated with tapes and returned SUCCESS.
285          */
286         if (sshdr.response_code == 0x70) {
287                 /* fixed format */
288                 if (scmd->sense_buffer[2] & 0xe0)
289                         return SUCCESS;
290         } else {
291                 /*
292                  * descriptor format: look for "stream commands sense data
293                  * descriptor" (see SSC-3). Assume single sense data
294                  * descriptor. Ignore ILI from SBC-2 READ LONG and WRITE LONG.
295                  */
296                 if ((sshdr.additional_length > 3) &&
297                     (scmd->sense_buffer[8] == 0x4) &&
298                     (scmd->sense_buffer[11] & 0xe0))
299                         return SUCCESS;
300         }
301
302         switch (sshdr.sense_key) {
303         case NO_SENSE:
304                 return SUCCESS;
305         case RECOVERED_ERROR:
306                 return /* soft_error */ SUCCESS;
307
308         case ABORTED_COMMAND:
309                 return NEEDS_RETRY;
310         case NOT_READY:
311         case UNIT_ATTENTION:
312                 /*
313                  * if we are expecting a cc/ua because of a bus reset that we
314                  * performed, treat this just as a retry.  otherwise this is
315                  * information that we should pass up to the upper-level driver
316                  * so that we can deal with it there.
317                  */
318                 if (scmd->device->expecting_cc_ua) {
319                         scmd->device->expecting_cc_ua = 0;
320                         return NEEDS_RETRY;
321                 }
322                 /*
323                  * if the device is in the process of becoming ready, we 
324                  * should retry.
325                  */
326                 if ((sshdr.asc == 0x04) && (sshdr.ascq == 0x01))
327                         return NEEDS_RETRY;
328                 /*
329                  * if the device is not started, we need to wake
330                  * the error handler to start the motor
331                  */
332                 if (scmd->device->allow_restart &&
333                     (sshdr.asc == 0x04) && (sshdr.ascq == 0x02))
334                         return FAILED;
335                 return SUCCESS;
336
337                 /* these three are not supported */
338         case COPY_ABORTED:
339         case VOLUME_OVERFLOW:
340         case MISCOMPARE:
341                 return SUCCESS;
342
343         case MEDIUM_ERROR:
344                 return NEEDS_RETRY;
345
346         case HARDWARE_ERROR:
347                 if (scmd->device->retry_hwerror)
348                         return NEEDS_RETRY;
349                 else
350                         return SUCCESS;
351
352         case ILLEGAL_REQUEST:
353         case BLANK_CHECK:
354         case DATA_PROTECT:
355         default:
356                 return SUCCESS;
357         }
358 }
359
360 /**
361  * scsi_eh_completed_normally - Disposition a eh cmd on return from LLD.
362  * @scmd:       SCSI cmd to examine.
363  *
364  * Notes:
365  *    This is *only* called when we are examining the status of commands
366  *    queued during error recovery.  the main difference here is that we
367  *    don't allow for the possibility of retries here, and we are a lot
368  *    more restrictive about what we consider acceptable.
369  **/
370 static int scsi_eh_completed_normally(struct scsi_cmnd *scmd)
371 {
372         /*
373          * first check the host byte, to see if there is anything in there
374          * that would indicate what we need to do.
375          */
376         if (host_byte(scmd->result) == DID_RESET) {
377                 /*
378                  * rats.  we are already in the error handler, so we now
379                  * get to try and figure out what to do next.  if the sense
380                  * is valid, we have a pretty good idea of what to do.
381                  * if not, we mark it as FAILED.
382                  */
383                 return scsi_check_sense(scmd);
384         }
385         if (host_byte(scmd->result) != DID_OK)
386                 return FAILED;
387
388         /*
389          * next, check the message byte.
390          */
391         if (msg_byte(scmd->result) != COMMAND_COMPLETE)
392                 return FAILED;
393
394         /*
395          * now, check the status byte to see if this indicates
396          * anything special.
397          */
398         switch (status_byte(scmd->result)) {
399         case GOOD:
400         case COMMAND_TERMINATED:
401                 return SUCCESS;
402         case CHECK_CONDITION:
403                 return scsi_check_sense(scmd);
404         case CONDITION_GOOD:
405         case INTERMEDIATE_GOOD:
406         case INTERMEDIATE_C_GOOD:
407                 /*
408                  * who knows?  FIXME(eric)
409                  */
410                 return SUCCESS;
411         case BUSY:
412         case QUEUE_FULL:
413         case RESERVATION_CONFLICT:
414         default:
415                 return FAILED;
416         }
417         return FAILED;
418 }
419
420 /**
421  * scsi_eh_times_out - timeout function for error handling.
422  * @scmd:       Cmd that is timing out.
423  *
424  * Notes:
425  *    During error handling, the kernel thread will be sleeping waiting
426  *    for some action to complete on the device.  our only job is to
427  *    record that it timed out, and to wake up the thread.
428  **/
429 static void scsi_eh_times_out(struct scsi_cmnd *scmd)
430 {
431         scmd->eh_eflags |= SCSI_EH_REC_TIMEOUT;
432         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: scmd:%p\n", __FUNCTION__,
433                                           scmd));
434
435         up(scmd->device->host->eh_action);
436 }
437
438 /**
439  * scsi_eh_done - Completion function for error handling.
440  * @scmd:       Cmd that is done.
441  **/
442 static void scsi_eh_done(struct scsi_cmnd *scmd)
443 {
444         /*
445          * if the timeout handler is already running, then just set the
446          * flag which says we finished late, and return.  we have no
447          * way of stopping the timeout handler from running, so we must
448          * always defer to it.
449          */
450         if (del_timer(&scmd->eh_timeout)) {
451                 scmd->request->rq_status = RQ_SCSI_DONE;
452
453                 SCSI_LOG_ERROR_RECOVERY(3, printk("%s scmd: %p result: %x\n",
454                                            __FUNCTION__, scmd, scmd->result));
455
456                 up(scmd->device->host->eh_action);
457         }
458 }
459
460 /**
461  * scsi_send_eh_cmnd  - send a cmd to a device as part of error recovery.
462  * @scmd:       SCSI Cmd to send.
463  * @timeout:    Timeout for cmd.
464  *
465  * Notes:
466  *    The initialization of the structures is quite a bit different in
467  *    this case, and furthermore, there is a different completion handler
468  *    vs scsi_dispatch_cmd.
469  * Return value:
470  *    SUCCESS or FAILED or NEEDS_RETRY
471  **/
472 static int scsi_send_eh_cmnd(struct scsi_cmnd *scmd, int timeout)
473 {
474         struct scsi_device *sdev = scmd->device;
475         struct Scsi_Host *shost = sdev->host;
476         DECLARE_MUTEX_LOCKED(sem);
477         unsigned long flags;
478         int rtn = SUCCESS;
479
480         /*
481          * we will use a queued command if possible, otherwise we will
482          * emulate the queuing and calling of completion function ourselves.
483          */
484         if (sdev->scsi_level <= SCSI_2)
485                 scmd->cmnd[1] = (scmd->cmnd[1] & 0x1f) |
486                         (sdev->lun << 5 & 0xe0);
487
488         scsi_add_timer(scmd, timeout, scsi_eh_times_out);
489
490         /*
491          * set up the semaphore so we wait for the command to complete.
492          */
493         shost->eh_action = &sem;
494         scmd->request->rq_status = RQ_SCSI_BUSY;
495
496         spin_lock_irqsave(shost->host_lock, flags);
497         scsi_log_send(scmd);
498         shost->hostt->queuecommand(scmd, scsi_eh_done);
499         spin_unlock_irqrestore(shost->host_lock, flags);
500
501         down(&sem);
502         scsi_log_completion(scmd, SUCCESS);
503
504         shost->eh_action = NULL;
505
506         /*
507          * see if timeout.  if so, tell the host to forget about it.
508          * in other words, we don't want a callback any more.
509          */
510         if (scmd->eh_eflags & SCSI_EH_REC_TIMEOUT) {
511                 scmd->eh_eflags &= ~SCSI_EH_REC_TIMEOUT;
512
513                 /*
514                  * as far as the low level driver is
515                  * concerned, this command is still active, so
516                  * we must give the low level driver a chance
517                  * to abort it. (db) 
518                  *
519                  * FIXME(eric) - we are not tracking whether we could
520                  * abort a timed out command or not.  not sure how
521                  * we should treat them differently anyways.
522                  */
523                 if (shost->hostt->eh_abort_handler)
524                         shost->hostt->eh_abort_handler(scmd);
525                         
526                 scmd->request->rq_status = RQ_SCSI_DONE;
527                 rtn = FAILED;
528         }
529
530         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: scmd: %p, rtn:%x\n",
531                                           __FUNCTION__, scmd, rtn));
532
533         /*
534          * now examine the actual status codes to see whether the command
535          * actually did complete normally.
536          */
537         if (rtn == SUCCESS) {
538                 rtn = scsi_eh_completed_normally(scmd);
539                 SCSI_LOG_ERROR_RECOVERY(3,
540                         printk("%s: scsi_eh_completed_normally %x\n",
541                                __FUNCTION__, rtn));
542                 switch (rtn) {
543                 case SUCCESS:
544                 case NEEDS_RETRY:
545                 case FAILED:
546                         break;
547                 default:
548                         rtn = FAILED;
549                         break;
550                 }
551         }
552
553         return rtn;
554 }
555
556 /**
557  * scsi_request_sense - Request sense data from a particular target.
558  * @scmd:       SCSI cmd for request sense.
559  *
560  * Notes:
561  *    Some hosts automatically obtain this information, others require
562  *    that we obtain it on our own. This function will *not* return until
563  *    the command either times out, or it completes.
564  **/
565 static int scsi_request_sense(struct scsi_cmnd *scmd)
566 {
567         static unsigned char generic_sense[6] =
568         {REQUEST_SENSE, 0, 0, 0, 252, 0};
569         unsigned char *scsi_result;
570         int saved_result;
571         int rtn;
572
573         memcpy(scmd->cmnd, generic_sense, sizeof(generic_sense));
574
575         scsi_result = kmalloc(252, GFP_ATOMIC | ((scmd->device->host->hostt->unchecked_isa_dma) ? __GFP_DMA : 0));
576
577
578         if (unlikely(!scsi_result)) {
579                 printk(KERN_ERR "%s: cannot allocate scsi_result.\n",
580                        __FUNCTION__);
581                 return FAILED;
582         }
583
584         /*
585          * zero the sense buffer.  some host adapters automatically always
586          * request sense, so it is not a good idea that
587          * scmd->request_buffer and scmd->sense_buffer point to the same
588          * address (db).  0 is not a valid sense code. 
589          */
590         memset(scmd->sense_buffer, 0, sizeof(scmd->sense_buffer));
591         memset(scsi_result, 0, 252);
592
593         saved_result = scmd->result;
594         scmd->request_buffer = scsi_result;
595         scmd->request_bufflen = 252;
596         scmd->use_sg = 0;
597         scmd->cmd_len = COMMAND_SIZE(scmd->cmnd[0]);
598         scmd->sc_data_direction = DMA_FROM_DEVICE;
599         scmd->underflow = 0;
600
601         rtn = scsi_send_eh_cmnd(scmd, SENSE_TIMEOUT);
602
603         /* last chance to have valid sense data */
604         if(!SCSI_SENSE_VALID(scmd)) {
605                 memcpy(scmd->sense_buffer, scmd->request_buffer,
606                        sizeof(scmd->sense_buffer));
607         }
608
609         kfree(scsi_result);
610
611         /*
612          * when we eventually call scsi_finish, we really wish to complete
613          * the original request, so let's restore the original data. (db)
614          */
615         scsi_setup_cmd_retry(scmd);
616         scmd->result = saved_result;
617         return rtn;
618 }
619
620 /**
621  * scsi_eh_finish_cmd - Handle a cmd that eh is finished with.
622  * @scmd:       Original SCSI cmd that eh has finished.
623  * @done_q:     Queue for processed commands.
624  *
625  * Notes:
626  *    We don't want to use the normal command completion while we are are
627  *    still handling errors - it may cause other commands to be queued,
628  *    and that would disturb what we are doing.  thus we really want to
629  *    keep a list of pending commands for final completion, and once we
630  *    are ready to leave error handling we handle completion for real.
631  **/
632 static void scsi_eh_finish_cmd(struct scsi_cmnd *scmd,
633                                struct list_head *done_q)
634 {
635         scmd->device->host->host_failed--;
636         scmd->eh_eflags = 0;
637
638         /*
639          * set this back so that the upper level can correctly free up
640          * things.
641          */
642         scsi_setup_cmd_retry(scmd);
643         list_move_tail(&scmd->eh_entry, done_q);
644 }
645
646 /**
647  * scsi_eh_get_sense - Get device sense data.
648  * @work_q:     Queue of commands to process.
649  * @done_q:     Queue of proccessed commands..
650  *
651  * Description:
652  *    See if we need to request sense information.  if so, then get it
653  *    now, so we have a better idea of what to do.  
654  *
655  * Notes:
656  *    This has the unfortunate side effect that if a shost adapter does
657  *    not automatically request sense information, that we end up shutting
658  *    it down before we request it.
659  *
660  *    All drivers should request sense information internally these days,
661  *    so for now all I have to say is tough noogies if you end up in here.
662  *
663  *    XXX: Long term this code should go away, but that needs an audit of
664  *         all LLDDs first.
665  **/
666 static int scsi_eh_get_sense(struct list_head *work_q,
667                              struct list_head *done_q)
668 {
669         struct scsi_cmnd *scmd, *next;
670         int rtn;
671
672         list_for_each_entry_safe(scmd, next, work_q, eh_entry) {
673                 if ((scmd->eh_eflags & SCSI_EH_CANCEL_CMD) ||
674                     SCSI_SENSE_VALID(scmd))
675                         continue;
676
677                 SCSI_LOG_ERROR_RECOVERY(2, printk("%s: requesting sense"
678                                                   " for id: %d\n",
679                                                   current->comm,
680                                                   scmd->device->id));
681                 rtn = scsi_request_sense(scmd);
682                 if (rtn != SUCCESS)
683                         continue;
684
685                 SCSI_LOG_ERROR_RECOVERY(3, printk("sense requested for %p"
686                                                   " result %x\n", scmd,
687                                                   scmd->result));
688                 SCSI_LOG_ERROR_RECOVERY(3, scsi_print_sense("bh", scmd));
689
690                 rtn = scsi_decide_disposition(scmd);
691
692                 /*
693                  * if the result was normal, then just pass it along to the
694                  * upper level.
695                  */
696                 if (rtn == SUCCESS)
697                         /* we don't want this command reissued, just
698                          * finished with the sense data, so set
699                          * retries to the max allowed to ensure it
700                          * won't get reissued */
701                         scmd->retries = scmd->allowed;
702                 else if (rtn != NEEDS_RETRY)
703                         continue;
704
705                 scsi_eh_finish_cmd(scmd, done_q);
706         }
707
708         return list_empty(work_q);
709 }
710
711 /**
712  * scsi_try_to_abort_cmd - Ask host to abort a running command.
713  * @scmd:       SCSI cmd to abort from Lower Level.
714  *
715  * Notes:
716  *    This function will not return until the user's completion function
717  *    has been called.  there is no timeout on this operation.  if the
718  *    author of the low-level driver wishes this operation to be timed,
719  *    they can provide this facility themselves.  helper functions in
720  *    scsi_error.c can be supplied to make this easier to do.
721  **/
722 static int scsi_try_to_abort_cmd(struct scsi_cmnd *scmd)
723 {
724         if (!scmd->device->host->hostt->eh_abort_handler)
725                 return FAILED;
726
727         /*
728          * scsi_done was called just after the command timed out and before
729          * we had a chance to process it. (db)
730          */
731         if (scmd->serial_number == 0)
732                 return SUCCESS;
733         return scmd->device->host->hostt->eh_abort_handler(scmd);
734 }
735
736 /**
737  * scsi_eh_tur - Send TUR to device.
738  * @scmd:       Scsi cmd to send TUR
739  *
740  * Return value:
741  *    0 - Device is ready. 1 - Device NOT ready.
742  **/
743 static int scsi_eh_tur(struct scsi_cmnd *scmd)
744 {
745         static unsigned char tur_command[6] = {TEST_UNIT_READY, 0, 0, 0, 0, 0};
746         int retry_cnt = 1, rtn;
747         int saved_result;
748
749 retry_tur:
750         memcpy(scmd->cmnd, tur_command, sizeof(tur_command));
751
752         /*
753          * zero the sense buffer.  the scsi spec mandates that any
754          * untransferred sense data should be interpreted as being zero.
755          */
756         memset(scmd->sense_buffer, 0, sizeof(scmd->sense_buffer));
757
758         saved_result = scmd->result;
759         scmd->request_buffer = NULL;
760         scmd->request_bufflen = 0;
761         scmd->use_sg = 0;
762         scmd->cmd_len = COMMAND_SIZE(scmd->cmnd[0]);
763         scmd->underflow = 0;
764         scmd->sc_data_direction = DMA_NONE;
765
766         rtn = scsi_send_eh_cmnd(scmd, SENSE_TIMEOUT);
767
768         /*
769          * when we eventually call scsi_finish, we really wish to complete
770          * the original request, so let's restore the original data. (db)
771          */
772         scsi_setup_cmd_retry(scmd);
773         scmd->result = saved_result;
774
775         /*
776          * hey, we are done.  let's look to see what happened.
777          */
778         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: scmd %p rtn %x\n",
779                 __FUNCTION__, scmd, rtn));
780         if (rtn == SUCCESS)
781                 return 0;
782         else if (rtn == NEEDS_RETRY) {
783                 if (retry_cnt--)
784                         goto retry_tur;
785                 return 0;
786         }
787         return 1;
788 }
789
790 /**
791  * scsi_eh_abort_cmds - abort canceled commands.
792  * @shost:      scsi host being recovered.
793  * @eh_done_q:  list_head for processed commands.
794  *
795  * Decription:
796  *    Try and see whether or not it makes sense to try and abort the
797  *    running command.  this only works out to be the case if we have one
798  *    command that has timed out.  if the command simply failed, it makes
799  *    no sense to try and abort the command, since as far as the shost
800  *    adapter is concerned, it isn't running.
801  **/
802 static int scsi_eh_abort_cmds(struct list_head *work_q,
803                               struct list_head *done_q)
804 {
805         struct scsi_cmnd *scmd, *next;
806         int rtn;
807
808         list_for_each_entry_safe(scmd, next, work_q, eh_entry) {
809                 if (!(scmd->eh_eflags & SCSI_EH_CANCEL_CMD))
810                         continue;
811                 SCSI_LOG_ERROR_RECOVERY(3, printk("%s: aborting cmd:"
812                                                   "0x%p\n", current->comm,
813                                                   scmd));
814                 rtn = scsi_try_to_abort_cmd(scmd);
815                 if (rtn == SUCCESS) {
816                         scmd->eh_eflags &= ~SCSI_EH_CANCEL_CMD;
817                         if (!scsi_device_online(scmd->device) ||
818                             !scsi_eh_tur(scmd)) {
819                                 scsi_eh_finish_cmd(scmd, done_q);
820                         }
821                                 
822                 } else
823                         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: aborting"
824                                                           " cmd failed:"
825                                                           "0x%p\n",
826                                                           current->comm,
827                                                           scmd));
828         }
829
830         return list_empty(work_q);
831 }
832
833 /**
834  * scsi_try_bus_device_reset - Ask host to perform a BDR on a dev
835  * @scmd:       SCSI cmd used to send BDR       
836  *
837  * Notes:
838  *    There is no timeout for this operation.  if this operation is
839  *    unreliable for a given host, then the host itself needs to put a
840  *    timer on it, and set the host back to a consistent state prior to
841  *    returning.
842  **/
843 static int scsi_try_bus_device_reset(struct scsi_cmnd *scmd)
844 {
845         int rtn;
846
847         if (!scmd->device->host->hostt->eh_device_reset_handler)
848                 return FAILED;
849
850         rtn = scmd->device->host->hostt->eh_device_reset_handler(scmd);
851         if (rtn == SUCCESS) {
852                 scmd->device->was_reset = 1;
853                 scmd->device->expecting_cc_ua = 1;
854         }
855
856         return rtn;
857 }
858
859 /**
860  * scsi_eh_try_stu - Send START_UNIT to device.
861  * @scmd:       Scsi cmd to send START_UNIT
862  *
863  * Return value:
864  *    0 - Device is ready. 1 - Device NOT ready.
865  **/
866 static int scsi_eh_try_stu(struct scsi_cmnd *scmd)
867 {
868         static unsigned char stu_command[6] = {START_STOP, 0, 0, 0, 1, 0};
869         int rtn;
870         int saved_result;
871
872         if (!scmd->device->allow_restart)
873                 return 1;
874
875         memcpy(scmd->cmnd, stu_command, sizeof(stu_command));
876
877         /*
878          * zero the sense buffer.  the scsi spec mandates that any
879          * untransferred sense data should be interpreted as being zero.
880          */
881         memset(scmd->sense_buffer, 0, sizeof(scmd->sense_buffer));
882
883         saved_result = scmd->result;
884         scmd->request_buffer = NULL;
885         scmd->request_bufflen = 0;
886         scmd->use_sg = 0;
887         scmd->cmd_len = COMMAND_SIZE(scmd->cmnd[0]);
888         scmd->underflow = 0;
889         scmd->sc_data_direction = DMA_NONE;
890
891         rtn = scsi_send_eh_cmnd(scmd, START_UNIT_TIMEOUT);
892
893         /*
894          * when we eventually call scsi_finish, we really wish to complete
895          * the original request, so let's restore the original data. (db)
896          */
897         scsi_setup_cmd_retry(scmd);
898         scmd->result = saved_result;
899
900         /*
901          * hey, we are done.  let's look to see what happened.
902          */
903         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: scmd %p rtn %x\n",
904                 __FUNCTION__, scmd, rtn));
905         if (rtn == SUCCESS)
906                 return 0;
907         return 1;
908 }
909
910  /**
911  * scsi_eh_stu - send START_UNIT if needed
912  * @shost:      scsi host being recovered.
913  * @eh_done_q:  list_head for processed commands.
914  *
915  * Notes:
916  *    If commands are failing due to not ready, initializing command required,
917  *      try revalidating the device, which will end up sending a start unit. 
918  **/
919 static int scsi_eh_stu(struct Scsi_Host *shost,
920                               struct list_head *work_q,
921                               struct list_head *done_q)
922 {
923         struct scsi_cmnd *scmd, *stu_scmd, *next;
924         struct scsi_device *sdev;
925
926         shost_for_each_device(sdev, shost) {
927                 stu_scmd = NULL;
928                 list_for_each_entry(scmd, work_q, eh_entry)
929                         if (scmd->device == sdev && SCSI_SENSE_VALID(scmd) &&
930                             scsi_check_sense(scmd) == FAILED ) {
931                                 stu_scmd = scmd;
932                                 break;
933                         }
934
935                 if (!stu_scmd)
936                         continue;
937
938                 SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Sending START_UNIT to sdev:"
939                                                   " 0x%p\n", current->comm, sdev));
940
941                 if (!scsi_eh_try_stu(stu_scmd)) {
942                         if (!scsi_device_online(sdev) ||
943                             !scsi_eh_tur(stu_scmd)) {
944                                 list_for_each_entry_safe(scmd, next,
945                                                           work_q, eh_entry) {
946                                         if (scmd->device == sdev)
947                                                 scsi_eh_finish_cmd(scmd, done_q);
948                                 }
949                         }
950                 } else {
951                         SCSI_LOG_ERROR_RECOVERY(3,
952                                                 printk("%s: START_UNIT failed to sdev:"
953                                                        " 0x%p\n", current->comm, sdev));
954                 }
955         }
956
957         return list_empty(work_q);
958 }
959
960
961 /**
962  * scsi_eh_bus_device_reset - send bdr if needed
963  * @shost:      scsi host being recovered.
964  * @eh_done_q:  list_head for processed commands.
965  *
966  * Notes:
967  *    Try a bus device reset.  still, look to see whether we have multiple
968  *    devices that are jammed or not - if we have multiple devices, it
969  *    makes no sense to try bus_device_reset - we really would need to try
970  *    a bus_reset instead. 
971  **/
972 static int scsi_eh_bus_device_reset(struct Scsi_Host *shost,
973                                     struct list_head *work_q,
974                                     struct list_head *done_q)
975 {
976         struct scsi_cmnd *scmd, *bdr_scmd, *next;
977         struct scsi_device *sdev;
978         int rtn;
979
980         shost_for_each_device(sdev, shost) {
981                 bdr_scmd = NULL;
982                 list_for_each_entry(scmd, work_q, eh_entry)
983                         if (scmd->device == sdev) {
984                                 bdr_scmd = scmd;
985                                 break;
986                         }
987
988                 if (!bdr_scmd)
989                         continue;
990
991                 SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Sending BDR sdev:"
992                                                   " 0x%p\n", current->comm,
993                                                   sdev));
994                 rtn = scsi_try_bus_device_reset(bdr_scmd);
995                 if (rtn == SUCCESS) {
996                         if (!scsi_device_online(sdev) ||
997                             !scsi_eh_tur(bdr_scmd)) {
998                                 list_for_each_entry_safe(scmd, next,
999                                                          work_q, eh_entry) {
1000                                         if (scmd->device == sdev)
1001                                                 scsi_eh_finish_cmd(scmd,
1002                                                                    done_q);
1003                                 }
1004                         }
1005                 } else {
1006                         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: BDR"
1007                                                           " failed sdev:"
1008                                                           "0x%p\n",
1009                                                           current->comm,
1010                                                            sdev));
1011                 }
1012         }
1013
1014         return list_empty(work_q);
1015 }
1016
1017 /**
1018  * scsi_try_bus_reset - ask host to perform a bus reset
1019  * @scmd:       SCSI cmd to send bus reset.
1020  **/
1021 static int scsi_try_bus_reset(struct scsi_cmnd *scmd)
1022 {
1023         unsigned long flags;
1024         int rtn;
1025
1026         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Snd Bus RST\n",
1027                                           __FUNCTION__));
1028
1029         if (!scmd->device->host->hostt->eh_bus_reset_handler)
1030                 return FAILED;
1031
1032         rtn = scmd->device->host->hostt->eh_bus_reset_handler(scmd);
1033
1034         if (rtn == SUCCESS) {
1035                 if (!scmd->device->host->hostt->skip_settle_delay)
1036                         ssleep(BUS_RESET_SETTLE_TIME);
1037                 spin_lock_irqsave(scmd->device->host->host_lock, flags);
1038                 scsi_report_bus_reset(scmd->device->host, scmd->device->channel);
1039                 spin_unlock_irqrestore(scmd->device->host->host_lock, flags);
1040         }
1041
1042         return rtn;
1043 }
1044
1045 /**
1046  * scsi_try_host_reset - ask host adapter to reset itself
1047  * @scmd:       SCSI cmd to send hsot reset.
1048  **/
1049 static int scsi_try_host_reset(struct scsi_cmnd *scmd)
1050 {
1051         unsigned long flags;
1052         int rtn;
1053
1054         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Snd Host RST\n",
1055                                           __FUNCTION__));
1056
1057         if (!scmd->device->host->hostt->eh_host_reset_handler)
1058                 return FAILED;
1059
1060         rtn = scmd->device->host->hostt->eh_host_reset_handler(scmd);
1061
1062         if (rtn == SUCCESS) {
1063                 if (!scmd->device->host->hostt->skip_settle_delay)
1064                         ssleep(HOST_RESET_SETTLE_TIME);
1065                 spin_lock_irqsave(scmd->device->host->host_lock, flags);
1066                 scsi_report_bus_reset(scmd->device->host, scmd->device->channel);
1067                 spin_unlock_irqrestore(scmd->device->host->host_lock, flags);
1068         }
1069
1070         return rtn;
1071 }
1072
1073 /**
1074  * scsi_eh_bus_reset - send a bus reset 
1075  * @shost:      scsi host being recovered.
1076  * @eh_done_q:  list_head for processed commands.
1077  **/
1078 static int scsi_eh_bus_reset(struct Scsi_Host *shost,
1079                              struct list_head *work_q,
1080                              struct list_head *done_q)
1081 {
1082         struct scsi_cmnd *scmd, *chan_scmd, *next;
1083         unsigned int channel;
1084         int rtn;
1085
1086         /*
1087          * we really want to loop over the various channels, and do this on
1088          * a channel by channel basis.  we should also check to see if any
1089          * of the failed commands are on soft_reset devices, and if so, skip
1090          * the reset.  
1091          */
1092
1093         for (channel = 0; channel <= shost->max_channel; channel++) {
1094                 chan_scmd = NULL;
1095                 list_for_each_entry(scmd, work_q, eh_entry) {
1096                         if (channel == scmd->device->channel) {
1097                                 chan_scmd = scmd;
1098                                 break;
1099                                 /*
1100                                  * FIXME add back in some support for
1101                                  * soft_reset devices.
1102                                  */
1103                         }
1104                 }
1105
1106                 if (!chan_scmd)
1107                         continue;
1108                 SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Sending BRST chan:"
1109                                                   " %d\n", current->comm,
1110                                                   channel));
1111                 rtn = scsi_try_bus_reset(chan_scmd);
1112                 if (rtn == SUCCESS) {
1113                         list_for_each_entry_safe(scmd, next, work_q, eh_entry) {
1114                                 if (channel == scmd->device->channel)
1115                                         if (!scsi_device_online(scmd->device) ||
1116                                             !scsi_eh_tur(scmd))
1117                                                 scsi_eh_finish_cmd(scmd,
1118                                                                    done_q);
1119                         }
1120                 } else {
1121                         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: BRST"
1122                                                           " failed chan: %d\n",
1123                                                           current->comm,
1124                                                           channel));
1125                 }
1126         }
1127         return list_empty(work_q);
1128 }
1129
1130 /**
1131  * scsi_eh_host_reset - send a host reset 
1132  * @work_q:     list_head for processed commands.
1133  * @done_q:     list_head for processed commands.
1134  **/
1135 static int scsi_eh_host_reset(struct list_head *work_q,
1136                               struct list_head *done_q)
1137 {
1138         struct scsi_cmnd *scmd, *next;
1139         int rtn;
1140
1141         if (!list_empty(work_q)) {
1142                 scmd = list_entry(work_q->next,
1143                                   struct scsi_cmnd, eh_entry);
1144
1145                 SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Sending HRST\n"
1146                                                   , current->comm));
1147
1148                 rtn = scsi_try_host_reset(scmd);
1149                 if (rtn == SUCCESS) {
1150                         list_for_each_entry_safe(scmd, next, work_q, eh_entry) {
1151                                 if (!scsi_device_online(scmd->device) ||
1152                                     (!scsi_eh_try_stu(scmd) && !scsi_eh_tur(scmd)) ||
1153                                     !scsi_eh_tur(scmd))
1154                                         scsi_eh_finish_cmd(scmd, done_q);
1155                         }
1156                 } else {
1157                         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: HRST"
1158                                                           " failed\n",
1159                                                           current->comm));
1160                 }
1161         }
1162         return list_empty(work_q);
1163 }
1164
1165 /**
1166  * scsi_eh_offline_sdevs - offline scsi devices that fail to recover
1167  * @work_q:     list_head for processed commands.
1168  * @done_q:     list_head for processed commands.
1169  *
1170  **/
1171 static void scsi_eh_offline_sdevs(struct list_head *work_q,
1172                                   struct list_head *done_q)
1173 {
1174         struct scsi_cmnd *scmd, *next;
1175
1176         list_for_each_entry_safe(scmd, next, work_q, eh_entry) {
1177                 printk(KERN_INFO "scsi: Device offlined - not"
1178                                 " ready after error recovery: host"
1179                                 " %d channel %d id %d lun %d\n",
1180                                 scmd->device->host->host_no,
1181                                 scmd->device->channel,
1182                                 scmd->device->id,
1183                                 scmd->device->lun);
1184                 scsi_device_set_state(scmd->device, SDEV_OFFLINE);
1185                 if (scmd->eh_eflags & SCSI_EH_CANCEL_CMD) {
1186                         /*
1187                          * FIXME: Handle lost cmds.
1188                          */
1189                 }
1190                 scsi_eh_finish_cmd(scmd, done_q);
1191         }
1192         return;
1193 }
1194
1195 /**
1196  * scsi_decide_disposition - Disposition a cmd on return from LLD.
1197  * @scmd:       SCSI cmd to examine.
1198  *
1199  * Notes:
1200  *    This is *only* called when we are examining the status after sending
1201  *    out the actual data command.  any commands that are queued for error
1202  *    recovery (e.g. test_unit_ready) do *not* come through here.
1203  *
1204  *    When this routine returns failed, it means the error handler thread
1205  *    is woken.  In cases where the error code indicates an error that
1206  *    doesn't require the error handler read (i.e. we don't need to
1207  *    abort/reset), this function should return SUCCESS.
1208  **/
1209 int scsi_decide_disposition(struct scsi_cmnd *scmd)
1210 {
1211         int rtn;
1212
1213         /*
1214          * if the device is offline, then we clearly just pass the result back
1215          * up to the top level.
1216          */
1217         if (!scsi_device_online(scmd->device)) {
1218                 SCSI_LOG_ERROR_RECOVERY(5, printk("%s: device offline - report"
1219                                                   " as SUCCESS\n",
1220                                                   __FUNCTION__));
1221                 return SUCCESS;
1222         }
1223
1224         /*
1225          * first check the host byte, to see if there is anything in there
1226          * that would indicate what we need to do.
1227          */
1228         switch (host_byte(scmd->result)) {
1229         case DID_PASSTHROUGH:
1230                 /*
1231                  * no matter what, pass this through to the upper layer.
1232                  * nuke this special code so that it looks like we are saying
1233                  * did_ok.
1234                  */
1235                 scmd->result &= 0xff00ffff;
1236                 return SUCCESS;
1237         case DID_OK:
1238                 /*
1239                  * looks good.  drop through, and check the next byte.
1240                  */
1241                 break;
1242         case DID_NO_CONNECT:
1243         case DID_BAD_TARGET:
1244         case DID_ABORT:
1245                 /*
1246                  * note - this means that we just report the status back
1247                  * to the top level driver, not that we actually think
1248                  * that it indicates SUCCESS.
1249                  */
1250                 return SUCCESS;
1251                 /*
1252                  * when the low level driver returns did_soft_error,
1253                  * it is responsible for keeping an internal retry counter 
1254                  * in order to avoid endless loops (db)
1255                  *
1256                  * actually this is a bug in this function here.  we should
1257                  * be mindful of the maximum number of retries specified
1258                  * and not get stuck in a loop.
1259                  */
1260         case DID_SOFT_ERROR:
1261                 goto maybe_retry;
1262         case DID_IMM_RETRY:
1263                 return NEEDS_RETRY;
1264
1265         case DID_REQUEUE:
1266                 return ADD_TO_MLQUEUE;
1267
1268         case DID_ERROR:
1269                 if (msg_byte(scmd->result) == COMMAND_COMPLETE &&
1270                     status_byte(scmd->result) == RESERVATION_CONFLICT)
1271                         /*
1272                          * execute reservation conflict processing code
1273                          * lower down
1274                          */
1275                         break;
1276                 /* fallthrough */
1277
1278         case DID_BUS_BUSY:
1279         case DID_PARITY:
1280                 goto maybe_retry;
1281         case DID_TIME_OUT:
1282                 /*
1283                  * when we scan the bus, we get timeout messages for
1284                  * these commands if there is no device available.
1285                  * other hosts report did_no_connect for the same thing.
1286                  */
1287                 if ((scmd->cmnd[0] == TEST_UNIT_READY ||
1288                      scmd->cmnd[0] == INQUIRY)) {
1289                         return SUCCESS;
1290                 } else {
1291                         return FAILED;
1292                 }
1293         case DID_RESET:
1294                 return SUCCESS;
1295         default:
1296                 return FAILED;
1297         }
1298
1299         /*
1300          * next, check the message byte.
1301          */
1302         if (msg_byte(scmd->result) != COMMAND_COMPLETE)
1303                 return FAILED;
1304
1305         /*
1306          * check the status byte to see if this indicates anything special.
1307          */
1308         switch (status_byte(scmd->result)) {
1309         case QUEUE_FULL:
1310                 /*
1311                  * the case of trying to send too many commands to a
1312                  * tagged queueing device.
1313                  */
1314         case BUSY:
1315                 /*
1316                  * device can't talk to us at the moment.  Should only
1317                  * occur (SAM-3) when the task queue is empty, so will cause
1318                  * the empty queue handling to trigger a stall in the
1319                  * device.
1320                  */
1321                 return ADD_TO_MLQUEUE;
1322         case GOOD:
1323         case COMMAND_TERMINATED:
1324         case TASK_ABORTED:
1325                 return SUCCESS;
1326         case CHECK_CONDITION:
1327                 rtn = scsi_check_sense(scmd);
1328                 if (rtn == NEEDS_RETRY)
1329                         goto maybe_retry;
1330                 /* if rtn == FAILED, we have no sense information;
1331                  * returning FAILED will wake the error handler thread
1332                  * to collect the sense and redo the decide
1333                  * disposition */
1334                 return rtn;
1335         case CONDITION_GOOD:
1336         case INTERMEDIATE_GOOD:
1337         case INTERMEDIATE_C_GOOD:
1338         case ACA_ACTIVE:
1339                 /*
1340                  * who knows?  FIXME(eric)
1341                  */
1342                 return SUCCESS;
1343
1344         case RESERVATION_CONFLICT:
1345                 printk(KERN_INFO "scsi: reservation conflict: host"
1346                                 " %d channel %d id %d lun %d\n",
1347                        scmd->device->host->host_no, scmd->device->channel,
1348                        scmd->device->id, scmd->device->lun);
1349                 return SUCCESS; /* causes immediate i/o error */
1350         default:
1351                 return FAILED;
1352         }
1353         return FAILED;
1354
1355       maybe_retry:
1356
1357         /* we requeue for retry because the error was retryable, and
1358          * the request was not marked fast fail.  Note that above,
1359          * even if the request is marked fast fail, we still requeue
1360          * for queue congestion conditions (QUEUE_FULL or BUSY) */
1361         if ((++scmd->retries) < scmd->allowed 
1362             && !blk_noretry_request(scmd->request)) {
1363                 return NEEDS_RETRY;
1364         } else {
1365                 /*
1366                  * no more retries - report this one back to upper level.
1367                  */
1368                 return SUCCESS;
1369         }
1370 }
1371
1372 /**
1373  * scsi_eh_lock_done - done function for eh door lock request
1374  * @scmd:       SCSI command block for the door lock request
1375  *
1376  * Notes:
1377  *      We completed the asynchronous door lock request, and it has either
1378  *      locked the door or failed.  We must free the command structures
1379  *      associated with this request.
1380  **/
1381 static void scsi_eh_lock_done(struct scsi_cmnd *scmd)
1382 {
1383         struct scsi_request *sreq = scmd->sc_request;
1384
1385         scsi_release_request(sreq);
1386 }
1387
1388
1389 /**
1390  * scsi_eh_lock_door - Prevent medium removal for the specified device
1391  * @sdev:       SCSI device to prevent medium removal
1392  *
1393  * Locking:
1394  *      We must be called from process context; scsi_allocate_request()
1395  *      may sleep.
1396  *
1397  * Notes:
1398  *      We queue up an asynchronous "ALLOW MEDIUM REMOVAL" request on the
1399  *      head of the devices request queue, and continue.
1400  *
1401  * Bugs:
1402  *      scsi_allocate_request() may sleep waiting for existing requests to
1403  *      be processed.  However, since we haven't kicked off any request
1404  *      processing for this host, this may deadlock.
1405  *
1406  *      If scsi_allocate_request() fails for what ever reason, we
1407  *      completely forget to lock the door.
1408  **/
1409 static void scsi_eh_lock_door(struct scsi_device *sdev)
1410 {
1411         struct scsi_request *sreq = scsi_allocate_request(sdev, GFP_KERNEL);
1412
1413         if (unlikely(!sreq)) {
1414                 printk(KERN_ERR "%s: request allocate failed,"
1415                        "prevent media removal cmd not sent\n", __FUNCTION__);
1416                 return;
1417         }
1418
1419         sreq->sr_cmnd[0] = ALLOW_MEDIUM_REMOVAL;
1420         sreq->sr_cmnd[1] = 0;
1421         sreq->sr_cmnd[2] = 0;
1422         sreq->sr_cmnd[3] = 0;
1423         sreq->sr_cmnd[4] = SCSI_REMOVAL_PREVENT;
1424         sreq->sr_cmnd[5] = 0;
1425         sreq->sr_data_direction = DMA_NONE;
1426         sreq->sr_bufflen = 0;
1427         sreq->sr_buffer = NULL;
1428         sreq->sr_allowed = 5;
1429         sreq->sr_done = scsi_eh_lock_done;
1430         sreq->sr_timeout_per_command = 10 * HZ;
1431         sreq->sr_cmd_len = COMMAND_SIZE(sreq->sr_cmnd[0]);
1432
1433         scsi_insert_special_req(sreq, 1);
1434 }
1435
1436
1437 /**
1438  * scsi_restart_operations - restart io operations to the specified host.
1439  * @shost:      Host we are restarting.
1440  *
1441  * Notes:
1442  *    When we entered the error handler, we blocked all further i/o to
1443  *    this device.  we need to 'reverse' this process.
1444  **/
1445 static void scsi_restart_operations(struct Scsi_Host *shost)
1446 {
1447         struct scsi_device *sdev;
1448         unsigned long flags;
1449
1450         /*
1451          * If the door was locked, we need to insert a door lock request
1452          * onto the head of the SCSI request queue for the device.  There
1453          * is no point trying to lock the door of an off-line device.
1454          */
1455         shost_for_each_device(sdev, shost) {
1456                 if (scsi_device_online(sdev) && sdev->locked)
1457                         scsi_eh_lock_door(sdev);
1458         }
1459
1460         /*
1461          * next free up anything directly waiting upon the host.  this
1462          * will be requests for character device operations, and also for
1463          * ioctls to queued block devices.
1464          */
1465         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: waking up host to restart\n",
1466                                           __FUNCTION__));
1467
1468         spin_lock_irqsave(shost->host_lock, flags);
1469         if (scsi_host_set_state(shost, SHOST_RUNNING))
1470                 if (scsi_host_set_state(shost, SHOST_CANCEL))
1471                         BUG_ON(scsi_host_set_state(shost, SHOST_DEL));
1472         spin_unlock_irqrestore(shost->host_lock, flags);
1473
1474         wake_up(&shost->host_wait);
1475
1476         /*
1477          * finally we need to re-initiate requests that may be pending.  we will
1478          * have had everything blocked while error handling is taking place, and
1479          * now that error recovery is done, we will need to ensure that these
1480          * requests are started.
1481          */
1482         scsi_run_host_queues(shost);
1483 }
1484
1485 /**
1486  * scsi_eh_ready_devs - check device ready state and recover if not.
1487  * @shost:      host to be recovered.
1488  * @eh_done_q:  list_head for processed commands.
1489  *
1490  **/
1491 static void scsi_eh_ready_devs(struct Scsi_Host *shost,
1492                                struct list_head *work_q,
1493                                struct list_head *done_q)
1494 {
1495         if (!scsi_eh_stu(shost, work_q, done_q))
1496                 if (!scsi_eh_bus_device_reset(shost, work_q, done_q))
1497                         if (!scsi_eh_bus_reset(shost, work_q, done_q))
1498                                 if (!scsi_eh_host_reset(work_q, done_q))
1499                                         scsi_eh_offline_sdevs(work_q, done_q);
1500 }
1501
1502 /**
1503  * scsi_eh_flush_done_q - finish processed commands or retry them.
1504  * @done_q:     list_head of processed commands.
1505  *
1506  **/
1507 static void scsi_eh_flush_done_q(struct list_head *done_q)
1508 {
1509         struct scsi_cmnd *scmd, *next;
1510
1511         list_for_each_entry_safe(scmd, next, done_q, eh_entry) {
1512                 list_del_init(&scmd->eh_entry);
1513                 if (scsi_device_online(scmd->device) &&
1514                     !blk_noretry_request(scmd->request) &&
1515                     (++scmd->retries < scmd->allowed)) {
1516                         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: flush"
1517                                                           " retry cmd: %p\n",
1518                                                           current->comm,
1519                                                           scmd));
1520                                 scsi_queue_insert(scmd, SCSI_MLQUEUE_EH_RETRY);
1521                 } else {
1522                         /*
1523                          * If just we got sense for the device (called
1524                          * scsi_eh_get_sense), scmd->result is already
1525                          * set, do not set DRIVER_TIMEOUT.
1526                          */
1527                         if (!scmd->result)
1528                                 scmd->result |= (DRIVER_TIMEOUT << 24);
1529                         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: flush finish"
1530                                                         " cmd: %p\n",
1531                                                         current->comm, scmd));
1532                         scsi_finish_command(scmd);
1533                 }
1534         }
1535 }
1536
1537 /**
1538  * scsi_unjam_host - Attempt to fix a host which has a cmd that failed.
1539  * @shost:      Host to unjam.
1540  *
1541  * Notes:
1542  *    When we come in here, we *know* that all commands on the bus have
1543  *    either completed, failed or timed out.  we also know that no further
1544  *    commands are being sent to the host, so things are relatively quiet
1545  *    and we have freedom to fiddle with things as we wish.
1546  *
1547  *    This is only the *default* implementation.  it is possible for
1548  *    individual drivers to supply their own version of this function, and
1549  *    if the maintainer wishes to do this, it is strongly suggested that
1550  *    this function be taken as a template and modified.  this function
1551  *    was designed to correctly handle problems for about 95% of the
1552  *    different cases out there, and it should always provide at least a
1553  *    reasonable amount of error recovery.
1554  *
1555  *    Any command marked 'failed' or 'timeout' must eventually have
1556  *    scsi_finish_cmd() called for it.  we do all of the retry stuff
1557  *    here, so when we restart the host after we return it should have an
1558  *    empty queue.
1559  **/
1560 static void scsi_unjam_host(struct Scsi_Host *shost)
1561 {
1562         unsigned long flags;
1563         LIST_HEAD(eh_work_q);
1564         LIST_HEAD(eh_done_q);
1565
1566         spin_lock_irqsave(shost->host_lock, flags);
1567         list_splice_init(&shost->eh_cmd_q, &eh_work_q);
1568         spin_unlock_irqrestore(shost->host_lock, flags);
1569
1570         SCSI_LOG_ERROR_RECOVERY(1, scsi_eh_prt_fail_stats(shost, &eh_work_q));
1571
1572         if (!scsi_eh_get_sense(&eh_work_q, &eh_done_q))
1573                 if (!scsi_eh_abort_cmds(&eh_work_q, &eh_done_q))
1574                         scsi_eh_ready_devs(shost, &eh_work_q, &eh_done_q);
1575
1576         scsi_eh_flush_done_q(&eh_done_q);
1577 }
1578
1579 /**
1580  * scsi_error_handler - Handle errors/timeouts of SCSI cmds.
1581  * @data:       Host for which we are running.
1582  *
1583  * Notes:
1584  *    This is always run in the context of a kernel thread.  The idea is
1585  *    that we start this thing up when the kernel starts up (one per host
1586  *    that we detect), and it immediately goes to sleep and waits for some
1587  *    event (i.e. failure).  When this takes place, we have the job of
1588  *    trying to unjam the bus and restarting things.
1589  **/
1590 int scsi_error_handler(void *data)
1591 {
1592         struct Scsi_Host *shost = (struct Scsi_Host *) data;
1593         int rtn;
1594
1595         current->flags |= PF_NOFREEZE;
1596
1597         
1598         /*
1599          * Note - we always use TASK_INTERRUPTIBLE even if the module
1600          * was loaded as part of the kernel.  The reason is that
1601          * UNINTERRUPTIBLE would cause this thread to be counted in
1602          * the load average as a running process, and an interruptible
1603          * wait doesn't.
1604          */
1605         set_current_state(TASK_INTERRUPTIBLE);
1606         while (!kthread_should_stop()) {
1607                 if (shost->host_failed == 0 ||
1608                     shost->host_failed != shost->host_busy) {
1609                         SCSI_LOG_ERROR_RECOVERY(1, printk("Error handler"
1610                                                           " scsi_eh_%d"
1611                                                           " sleeping\n",
1612                                                           shost->host_no));
1613                         schedule();
1614                         set_current_state(TASK_INTERRUPTIBLE);
1615                         continue;
1616                 }
1617
1618                 __set_current_state(TASK_RUNNING);
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                 set_current_state(TASK_INTERRUPTIBLE);
1646         }
1647
1648         __set_current_state(TASK_RUNNING);
1649
1650         SCSI_LOG_ERROR_RECOVERY(1, printk("Error handler scsi_eh_%d"
1651                                           " exiting\n",shost->host_no));
1652
1653         /*
1654          * Make sure that nobody tries to wake us up again.
1655          */
1656         shost->ehandler = NULL;
1657         return 0;
1658 }
1659
1660 /*
1661  * Function:    scsi_report_bus_reset()
1662  *
1663  * Purpose:     Utility function used by low-level drivers to report that
1664  *              they have observed a bus reset on the bus being handled.
1665  *
1666  * Arguments:   shost       - Host in question
1667  *              channel     - channel on which reset was observed.
1668  *
1669  * Returns:     Nothing
1670  *
1671  * Lock status: Host lock must be held.
1672  *
1673  * Notes:       This only needs to be called if the reset is one which
1674  *              originates from an unknown location.  Resets originated
1675  *              by the mid-level itself don't need to call this, but there
1676  *              should be no harm.
1677  *
1678  *              The main purpose of this is to make sure that a CHECK_CONDITION
1679  *              is properly treated.
1680  */
1681 void scsi_report_bus_reset(struct Scsi_Host *shost, int channel)
1682 {
1683         struct scsi_device *sdev;
1684
1685         __shost_for_each_device(sdev, shost) {
1686                 if (channel == sdev->channel) {
1687                         sdev->was_reset = 1;
1688                         sdev->expecting_cc_ua = 1;
1689                 }
1690         }
1691 }
1692 EXPORT_SYMBOL(scsi_report_bus_reset);
1693
1694 /*
1695  * Function:    scsi_report_device_reset()
1696  *
1697  * Purpose:     Utility function used by low-level drivers to report that
1698  *              they have observed a device reset on the device being handled.
1699  *
1700  * Arguments:   shost       - Host in question
1701  *              channel     - channel on which reset was observed
1702  *              target      - target on which reset was observed
1703  *
1704  * Returns:     Nothing
1705  *
1706  * Lock status: Host lock must be held
1707  *
1708  * Notes:       This only needs to be called if the reset is one which
1709  *              originates from an unknown location.  Resets originated
1710  *              by the mid-level itself don't need to call this, but there
1711  *              should be no harm.
1712  *
1713  *              The main purpose of this is to make sure that a CHECK_CONDITION
1714  *              is properly treated.
1715  */
1716 void scsi_report_device_reset(struct Scsi_Host *shost, int channel, int target)
1717 {
1718         struct scsi_device *sdev;
1719
1720         __shost_for_each_device(sdev, shost) {
1721                 if (channel == sdev->channel &&
1722                     target == sdev->id) {
1723                         sdev->was_reset = 1;
1724                         sdev->expecting_cc_ua = 1;
1725                 }
1726         }
1727 }
1728 EXPORT_SYMBOL(scsi_report_device_reset);
1729
1730 static void
1731 scsi_reset_provider_done_command(struct scsi_cmnd *scmd)
1732 {
1733 }
1734
1735 /*
1736  * Function:    scsi_reset_provider
1737  *
1738  * Purpose:     Send requested reset to a bus or device at any phase.
1739  *
1740  * Arguments:   device  - device to send reset to
1741  *              flag - reset type (see scsi.h)
1742  *
1743  * Returns:     SUCCESS/FAILURE.
1744  *
1745  * Notes:       This is used by the SCSI Generic driver to provide
1746  *              Bus/Device reset capability.
1747  */
1748 int
1749 scsi_reset_provider(struct scsi_device *dev, int flag)
1750 {
1751         struct scsi_cmnd *scmd = scsi_get_command(dev, GFP_KERNEL);
1752         struct request req;
1753         int rtn;
1754
1755         scmd->request = &req;
1756         memset(&scmd->eh_timeout, 0, sizeof(scmd->eh_timeout));
1757         scmd->request->rq_status        = RQ_SCSI_BUSY;
1758
1759         memset(&scmd->cmnd, '\0', sizeof(scmd->cmnd));
1760     
1761         scmd->scsi_done         = scsi_reset_provider_done_command;
1762         scmd->done                      = NULL;
1763         scmd->buffer                    = NULL;
1764         scmd->bufflen                   = 0;
1765         scmd->request_buffer            = NULL;
1766         scmd->request_bufflen           = 0;
1767
1768         scmd->cmd_len                   = 0;
1769
1770         scmd->sc_data_direction         = DMA_BIDIRECTIONAL;
1771         scmd->sc_request                = NULL;
1772         scmd->sc_magic                  = SCSI_CMND_MAGIC;
1773
1774         init_timer(&scmd->eh_timeout);
1775
1776         /*
1777          * Sometimes the command can get back into the timer chain,
1778          * so use the pid as an identifier.
1779          */
1780         scmd->pid                       = 0;
1781
1782         switch (flag) {
1783         case SCSI_TRY_RESET_DEVICE:
1784                 rtn = scsi_try_bus_device_reset(scmd);
1785                 if (rtn == SUCCESS)
1786                         break;
1787                 /* FALLTHROUGH */
1788         case SCSI_TRY_RESET_BUS:
1789                 rtn = scsi_try_bus_reset(scmd);
1790                 if (rtn == SUCCESS)
1791                         break;
1792                 /* FALLTHROUGH */
1793         case SCSI_TRY_RESET_HOST:
1794                 rtn = scsi_try_host_reset(scmd);
1795                 break;
1796         default:
1797                 rtn = FAILED;
1798         }
1799
1800         scsi_next_command(scmd);
1801         return rtn;
1802 }
1803 EXPORT_SYMBOL(scsi_reset_provider);
1804
1805 /**
1806  * scsi_normalize_sense - normalize main elements from either fixed or
1807  *                      descriptor sense data format into a common format.
1808  *
1809  * @sense_buffer:       byte array containing sense data returned by device
1810  * @sb_len:             number of valid bytes in sense_buffer
1811  * @sshdr:              pointer to instance of structure that common
1812  *                      elements are written to.
1813  *
1814  * Notes:
1815  *      The "main elements" from sense data are: response_code, sense_key,
1816  *      asc, ascq and additional_length (only for descriptor format).
1817  *
1818  *      Typically this function can be called after a device has
1819  *      responded to a SCSI command with the CHECK_CONDITION status.
1820  *
1821  * Return value:
1822  *      1 if valid sense data information found, else 0;
1823  **/
1824 int scsi_normalize_sense(const u8 *sense_buffer, int sb_len,
1825                          struct scsi_sense_hdr *sshdr)
1826 {
1827         if (!sense_buffer || !sb_len)
1828                 return 0;
1829
1830         memset(sshdr, 0, sizeof(struct scsi_sense_hdr));
1831
1832         sshdr->response_code = (sense_buffer[0] & 0x7f);
1833
1834         if (!scsi_sense_valid(sshdr))
1835                 return 0;
1836
1837         if (sshdr->response_code >= 0x72) {
1838                 /*
1839                  * descriptor format
1840                  */
1841                 if (sb_len > 1)
1842                         sshdr->sense_key = (sense_buffer[1] & 0xf);
1843                 if (sb_len > 2)
1844                         sshdr->asc = sense_buffer[2];
1845                 if (sb_len > 3)
1846                         sshdr->ascq = sense_buffer[3];
1847                 if (sb_len > 7)
1848                         sshdr->additional_length = sense_buffer[7];
1849         } else {
1850                 /* 
1851                  * fixed format
1852                  */
1853                 if (sb_len > 2)
1854                         sshdr->sense_key = (sense_buffer[2] & 0xf);
1855                 if (sb_len > 7) {
1856                         sb_len = (sb_len < (sense_buffer[7] + 8)) ?
1857                                          sb_len : (sense_buffer[7] + 8);
1858                         if (sb_len > 12)
1859                                 sshdr->asc = sense_buffer[12];
1860                         if (sb_len > 13)
1861                                 sshdr->ascq = sense_buffer[13];
1862                 }
1863         }
1864
1865         return 1;
1866 }
1867 EXPORT_SYMBOL(scsi_normalize_sense);
1868
1869 int scsi_request_normalize_sense(struct scsi_request *sreq,
1870                                  struct scsi_sense_hdr *sshdr)
1871 {
1872         return scsi_normalize_sense(sreq->sr_sense_buffer,
1873                         sizeof(sreq->sr_sense_buffer), sshdr);
1874 }
1875 EXPORT_SYMBOL(scsi_request_normalize_sense);
1876
1877 int scsi_command_normalize_sense(struct scsi_cmnd *cmd,
1878                                  struct scsi_sense_hdr *sshdr)
1879 {
1880         return scsi_normalize_sense(cmd->sense_buffer,
1881                         sizeof(cmd->sense_buffer), sshdr);
1882 }
1883 EXPORT_SYMBOL(scsi_command_normalize_sense);
1884
1885 /**
1886  * scsi_sense_desc_find - search for a given descriptor type in
1887  *                      descriptor sense data format.
1888  *
1889  * @sense_buffer:       byte array of descriptor format sense data
1890  * @sb_len:             number of valid bytes in sense_buffer
1891  * @desc_type:          value of descriptor type to find
1892  *                      (e.g. 0 -> information)
1893  *
1894  * Notes:
1895  *      only valid when sense data is in descriptor format
1896  *
1897  * Return value:
1898  *      pointer to start of (first) descriptor if found else NULL
1899  **/
1900 const u8 * scsi_sense_desc_find(const u8 * sense_buffer, int sb_len,
1901                                 int desc_type)
1902 {
1903         int add_sen_len, add_len, desc_len, k;
1904         const u8 * descp;
1905
1906         if ((sb_len < 8) || (0 == (add_sen_len = sense_buffer[7])))
1907                 return NULL;
1908         if ((sense_buffer[0] < 0x72) || (sense_buffer[0] > 0x73))
1909                 return NULL;
1910         add_sen_len = (add_sen_len < (sb_len - 8)) ?
1911                         add_sen_len : (sb_len - 8);
1912         descp = &sense_buffer[8];
1913         for (desc_len = 0, k = 0; k < add_sen_len; k += desc_len) {
1914                 descp += desc_len;
1915                 add_len = (k < (add_sen_len - 1)) ? descp[1]: -1;
1916                 desc_len = add_len + 2;
1917                 if (descp[0] == desc_type)
1918                         return descp;
1919                 if (add_len < 0) // short descriptor ??
1920                         break;
1921         }
1922         return NULL;
1923 }
1924 EXPORT_SYMBOL(scsi_sense_desc_find);
1925
1926 /**
1927  * scsi_get_sense_info_fld - attempts to get information field from
1928  *                      sense data (either fixed or descriptor format)
1929  *
1930  * @sense_buffer:       byte array of sense data
1931  * @sb_len:             number of valid bytes in sense_buffer
1932  * @info_out:           pointer to 64 integer where 8 or 4 byte information
1933  *                      field will be placed if found.
1934  *
1935  * Return value:
1936  *      1 if information field found, 0 if not found.
1937  **/
1938 int scsi_get_sense_info_fld(const u8 * sense_buffer, int sb_len,
1939                             u64 * info_out)
1940 {
1941         int j;
1942         const u8 * ucp;
1943         u64 ull;
1944
1945         if (sb_len < 7)
1946                 return 0;
1947         switch (sense_buffer[0] & 0x7f) {
1948         case 0x70:
1949         case 0x71:
1950                 if (sense_buffer[0] & 0x80) {
1951                         *info_out = (sense_buffer[3] << 24) +
1952                                     (sense_buffer[4] << 16) +
1953                                     (sense_buffer[5] << 8) + sense_buffer[6];
1954                         return 1;
1955                 } else
1956                         return 0;
1957         case 0x72:
1958         case 0x73:
1959                 ucp = scsi_sense_desc_find(sense_buffer, sb_len,
1960                                            0 /* info desc */);
1961                 if (ucp && (0xa == ucp[1])) {
1962                         ull = 0;
1963                         for (j = 0; j < 8; ++j) {
1964                                 if (j > 0)
1965                                         ull <<= 8;
1966                                 ull |= ucp[4 + j];
1967                         }
1968                         *info_out = ull;
1969                         return 1;
1970                 } else
1971                         return 0;
1972         default:
1973                 return 0;
1974         }
1975 }
1976 EXPORT_SYMBOL(scsi_get_sense_info_fld);