Merge branch 'master' of git://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux-2.6
[linux-2.6] / drivers / scsi / scsi_error.c
1 /*
2  *  scsi_error.c Copyright (C) 1997 Eric Youngdale
3  *
4  *  SCSI error/timeout handling
5  *      Initial versions: Eric Youngdale.  Based upon conversations with
6  *                        Leonard Zubkoff and David Miller at Linux Expo, 
7  *                        ideas originating from all over the place.
8  *
9  *      Restructured scsi_unjam_host and associated functions.
10  *      September 04, 2002 Mike Anderson (andmike@us.ibm.com)
11  *
12  *      Forward port of Russell King's (rmk@arm.linux.org.uk) changes and
13  *      minor  cleanups.
14  *      September 30, 2002 Mike Anderson (andmike@us.ibm.com)
15  */
16
17 #include <linux/module.h>
18 #include <linux/sched.h>
19 #include <linux/timer.h>
20 #include <linux/string.h>
21 #include <linux/slab.h>
22 #include <linux/kernel.h>
23 #include <linux/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_cmnd.h>
30 #include <scsi/scsi_dbg.h>
31 #include <scsi/scsi_device.h>
32 #include <scsi/scsi_eh.h>
33 #include <scsi/scsi_transport.h>
34 #include <scsi/scsi_host.h>
35 #include <scsi/scsi_ioctl.h>
36
37 #include "scsi_priv.h"
38 #include "scsi_logging.h"
39
40 #define SENSE_TIMEOUT           (10*HZ)
41 #define START_UNIT_TIMEOUT      (30*HZ)
42
43 /*
44  * These should *probably* be handled by the host itself.
45  * Since it is allowed to sleep, it probably should.
46  */
47 #define BUS_RESET_SETTLE_TIME   (10)
48 #define HOST_RESET_SETTLE_TIME  (10)
49
50 /* called with shost->host_lock held */
51 void scsi_eh_wakeup(struct Scsi_Host *shost)
52 {
53         if (shost->host_busy == shost->host_failed) {
54                 wake_up_process(shost->ehandler);
55                 SCSI_LOG_ERROR_RECOVERY(5,
56                                 printk("Waking error handler thread\n"));
57         }
58 }
59
60 /**
61  * scsi_schedule_eh - schedule EH for SCSI host
62  * @shost:      SCSI host to invoke error handling on.
63  *
64  * Schedule SCSI EH without scmd.
65  **/
66 void scsi_schedule_eh(struct Scsi_Host *shost)
67 {
68         unsigned long flags;
69
70         spin_lock_irqsave(shost->host_lock, flags);
71
72         if (scsi_host_set_state(shost, SHOST_RECOVERY) == 0 ||
73             scsi_host_set_state(shost, SHOST_CANCEL_RECOVERY) == 0) {
74                 shost->host_eh_scheduled++;
75                 scsi_eh_wakeup(shost);
76         }
77
78         spin_unlock_irqrestore(shost->host_lock, flags);
79 }
80 EXPORT_SYMBOL_GPL(scsi_schedule_eh);
81
82 /**
83  * scsi_eh_scmd_add - add scsi cmd to error handling.
84  * @scmd:       scmd to run eh on.
85  * @eh_flag:    optional SCSI_EH flag.
86  *
87  * Return value:
88  *      0 on failure.
89  **/
90 int scsi_eh_scmd_add(struct scsi_cmnd *scmd, int eh_flag)
91 {
92         struct Scsi_Host *shost = scmd->device->host;
93         unsigned long flags;
94         int ret = 0;
95
96         if (!shost->ehandler)
97                 return 0;
98
99         spin_lock_irqsave(shost->host_lock, flags);
100         if (scsi_host_set_state(shost, SHOST_RECOVERY))
101                 if (scsi_host_set_state(shost, SHOST_CANCEL_RECOVERY))
102                         goto out_unlock;
103
104         ret = 1;
105         scmd->eh_eflags |= eh_flag;
106         list_add_tail(&scmd->eh_entry, &shost->eh_cmd_q);
107         shost->host_failed++;
108         scsi_eh_wakeup(shost);
109  out_unlock:
110         spin_unlock_irqrestore(shost->host_lock, flags);
111         return ret;
112 }
113
114 /**
115  * scsi_add_timer - Start timeout timer for a single scsi command.
116  * @scmd:       scsi command that is about to start running.
117  * @timeout:    amount of time to allow this command to run.
118  * @complete:   timeout function to call if timer isn't canceled.
119  *
120  * Notes:
121  *    This should be turned into an inline function.  Each scsi command
122  *    has its own timer, and as it is added to the queue, we set up the
123  *    timer.  When the command completes, we cancel the timer.
124  **/
125 void scsi_add_timer(struct scsi_cmnd *scmd, int timeout,
126                     void (*complete)(struct scsi_cmnd *))
127 {
128
129         /*
130          * If the clock was already running for this command, then
131          * first delete the timer.  The timer handling code gets rather
132          * confused if we don't do this.
133          */
134         if (scmd->eh_timeout.function)
135                 del_timer(&scmd->eh_timeout);
136
137         scmd->eh_timeout.data = (unsigned long)scmd;
138         scmd->eh_timeout.expires = jiffies + timeout;
139         scmd->eh_timeout.function = (void (*)(unsigned long)) complete;
140
141         SCSI_LOG_ERROR_RECOVERY(5, printk("%s: scmd: %p, time:"
142                                           " %d, (%p)\n", __FUNCTION__,
143                                           scmd, timeout, complete));
144
145         add_timer(&scmd->eh_timeout);
146 }
147
148 /**
149  * scsi_delete_timer - Delete/cancel timer for a given function.
150  * @scmd:       Cmd that we are canceling timer for
151  *
152  * Notes:
153  *     This should be turned into an inline function.
154  *
155  * Return value:
156  *     1 if we were able to detach the timer.  0 if we blew it, and the
157  *     timer function has already started to run.
158  **/
159 int scsi_delete_timer(struct scsi_cmnd *scmd)
160 {
161         int rtn;
162
163         rtn = del_timer(&scmd->eh_timeout);
164
165         SCSI_LOG_ERROR_RECOVERY(5, printk("%s: scmd: %p,"
166                                          " rtn: %d\n", __FUNCTION__,
167                                          scmd, rtn));
168
169         scmd->eh_timeout.data = (unsigned long)NULL;
170         scmd->eh_timeout.function = NULL;
171
172         return rtn;
173 }
174
175 /**
176  * scsi_times_out - Timeout function for normal scsi commands.
177  * @scmd:       Cmd that is timing out.
178  *
179  * Notes:
180  *     We do not need to lock this.  There is the potential for a race
181  *     only in that the normal completion handling might run, but if the
182  *     normal completion function determines that the timer has already
183  *     fired, then it mustn't do anything.
184  **/
185 void scsi_times_out(struct scsi_cmnd *scmd)
186 {
187         scsi_log_completion(scmd, TIMEOUT_ERROR);
188
189         if (scmd->device->host->transportt->eh_timed_out)
190                 switch (scmd->device->host->transportt->eh_timed_out(scmd)) {
191                 case EH_HANDLED:
192                         __scsi_done(scmd);
193                         return;
194                 case EH_RESET_TIMER:
195                         scsi_add_timer(scmd, scmd->timeout_per_command,
196                                        scsi_times_out);
197                         return;
198                 case EH_NOT_HANDLED:
199                         break;
200                 }
201
202         if (unlikely(!scsi_eh_scmd_add(scmd, SCSI_EH_CANCEL_CMD))) {
203                 scmd->result |= DID_TIME_OUT << 16;
204                 __scsi_done(scmd);
205         }
206 }
207
208 /**
209  * scsi_block_when_processing_errors - Prevent cmds from being queued.
210  * @sdev:       Device on which we are performing recovery.
211  *
212  * Description:
213  *     We block until the host is out of error recovery, and then check to
214  *     see whether the host or the device is offline.
215  *
216  * Return value:
217  *     0 when dev was taken offline by error recovery. 1 OK to proceed.
218  **/
219 int scsi_block_when_processing_errors(struct scsi_device *sdev)
220 {
221         int online;
222
223         wait_event(sdev->host->host_wait, !scsi_host_in_recovery(sdev->host));
224
225         online = scsi_device_online(sdev);
226
227         SCSI_LOG_ERROR_RECOVERY(5, printk("%s: rtn: %d\n", __FUNCTION__,
228                                           online));
229
230         return online;
231 }
232 EXPORT_SYMBOL(scsi_block_when_processing_errors);
233
234 #ifdef CONFIG_SCSI_LOGGING
235 /**
236  * scsi_eh_prt_fail_stats - Log info on failures.
237  * @shost:      scsi host being recovered.
238  * @work_q:     Queue of scsi cmds to process.
239  **/
240 static inline void scsi_eh_prt_fail_stats(struct Scsi_Host *shost,
241                                           struct list_head *work_q)
242 {
243         struct scsi_cmnd *scmd;
244         struct scsi_device *sdev;
245         int total_failures = 0;
246         int cmd_failed = 0;
247         int cmd_cancel = 0;
248         int devices_failed = 0;
249
250         shost_for_each_device(sdev, shost) {
251                 list_for_each_entry(scmd, work_q, eh_entry) {
252                         if (scmd->device == sdev) {
253                                 ++total_failures;
254                                 if (scmd->eh_eflags & SCSI_EH_CANCEL_CMD)
255                                         ++cmd_cancel;
256                                 else 
257                                         ++cmd_failed;
258                         }
259                 }
260
261                 if (cmd_cancel || cmd_failed) {
262                         SCSI_LOG_ERROR_RECOVERY(3,
263                                 sdev_printk(KERN_INFO, sdev,
264                                             "%s: cmds failed: %d, cancel: %d\n",
265                                             __FUNCTION__, cmd_failed,
266                                             cmd_cancel));
267                         cmd_cancel = 0;
268                         cmd_failed = 0;
269                         ++devices_failed;
270                 }
271         }
272
273         SCSI_LOG_ERROR_RECOVERY(2, printk("Total of %d commands on %d"
274                                           " devices require eh work\n",
275                                   total_failures, devices_failed));
276 }
277 #endif
278
279 /**
280  * scsi_check_sense - Examine scsi cmd sense
281  * @scmd:       Cmd to have sense checked.
282  *
283  * Return value:
284  *      SUCCESS or FAILED or NEEDS_RETRY
285  *
286  * Notes:
287  *      When a deferred error is detected the current command has
288  *      not been executed and needs retrying.
289  **/
290 static int scsi_check_sense(struct scsi_cmnd *scmd)
291 {
292         struct scsi_sense_hdr sshdr;
293
294         if (! scsi_command_normalize_sense(scmd, &sshdr))
295                 return FAILED;  /* no valid sense data */
296
297         if (scsi_sense_is_deferred(&sshdr))
298                 return NEEDS_RETRY;
299
300         /*
301          * Previous logic looked for FILEMARK, EOM or ILI which are
302          * mainly associated with tapes and returned SUCCESS.
303          */
304         if (sshdr.response_code == 0x70) {
305                 /* fixed format */
306                 if (scmd->sense_buffer[2] & 0xe0)
307                         return SUCCESS;
308         } else {
309                 /*
310                  * descriptor format: look for "stream commands sense data
311                  * descriptor" (see SSC-3). Assume single sense data
312                  * descriptor. Ignore ILI from SBC-2 READ LONG and WRITE LONG.
313                  */
314                 if ((sshdr.additional_length > 3) &&
315                     (scmd->sense_buffer[8] == 0x4) &&
316                     (scmd->sense_buffer[11] & 0xe0))
317                         return SUCCESS;
318         }
319
320         switch (sshdr.sense_key) {
321         case NO_SENSE:
322                 return SUCCESS;
323         case RECOVERED_ERROR:
324                 return /* soft_error */ SUCCESS;
325
326         case ABORTED_COMMAND:
327                 return NEEDS_RETRY;
328         case NOT_READY:
329         case UNIT_ATTENTION:
330                 /*
331                  * if we are expecting a cc/ua because of a bus reset that we
332                  * performed, treat this just as a retry.  otherwise this is
333                  * information that we should pass up to the upper-level driver
334                  * so that we can deal with it there.
335                  */
336                 if (scmd->device->expecting_cc_ua) {
337                         scmd->device->expecting_cc_ua = 0;
338                         return NEEDS_RETRY;
339                 }
340                 /*
341                  * if the device is in the process of becoming ready, we 
342                  * should retry.
343                  */
344                 if ((sshdr.asc == 0x04) && (sshdr.ascq == 0x01))
345                         return NEEDS_RETRY;
346                 /*
347                  * if the device is not started, we need to wake
348                  * the error handler to start the motor
349                  */
350                 if (scmd->device->allow_restart &&
351                     (sshdr.asc == 0x04) && (sshdr.ascq == 0x02))
352                         return FAILED;
353                 return SUCCESS;
354
355                 /* these three are not supported */
356         case COPY_ABORTED:
357         case VOLUME_OVERFLOW:
358         case MISCOMPARE:
359                 return SUCCESS;
360
361         case MEDIUM_ERROR:
362                 return NEEDS_RETRY;
363
364         case HARDWARE_ERROR:
365                 if (scmd->device->retry_hwerror)
366                         return NEEDS_RETRY;
367                 else
368                         return SUCCESS;
369
370         case ILLEGAL_REQUEST:
371         case BLANK_CHECK:
372         case DATA_PROTECT:
373         default:
374                 return SUCCESS;
375         }
376 }
377
378 /**
379  * scsi_eh_completed_normally - Disposition a eh cmd on return from LLD.
380  * @scmd:       SCSI cmd to examine.
381  *
382  * Notes:
383  *    This is *only* called when we are examining the status of commands
384  *    queued during error recovery.  the main difference here is that we
385  *    don't allow for the possibility of retries here, and we are a lot
386  *    more restrictive about what we consider acceptable.
387  **/
388 static int scsi_eh_completed_normally(struct scsi_cmnd *scmd)
389 {
390         /*
391          * first check the host byte, to see if there is anything in there
392          * that would indicate what we need to do.
393          */
394         if (host_byte(scmd->result) == DID_RESET) {
395                 /*
396                  * rats.  we are already in the error handler, so we now
397                  * get to try and figure out what to do next.  if the sense
398                  * is valid, we have a pretty good idea of what to do.
399                  * if not, we mark it as FAILED.
400                  */
401                 return scsi_check_sense(scmd);
402         }
403         if (host_byte(scmd->result) != DID_OK)
404                 return FAILED;
405
406         /*
407          * next, check the message byte.
408          */
409         if (msg_byte(scmd->result) != COMMAND_COMPLETE)
410                 return FAILED;
411
412         /*
413          * now, check the status byte to see if this indicates
414          * anything special.
415          */
416         switch (status_byte(scmd->result)) {
417         case GOOD:
418         case COMMAND_TERMINATED:
419                 return SUCCESS;
420         case CHECK_CONDITION:
421                 return scsi_check_sense(scmd);
422         case CONDITION_GOOD:
423         case INTERMEDIATE_GOOD:
424         case INTERMEDIATE_C_GOOD:
425                 /*
426                  * who knows?  FIXME(eric)
427                  */
428                 return SUCCESS;
429         case BUSY:
430         case QUEUE_FULL:
431         case RESERVATION_CONFLICT:
432         default:
433                 return FAILED;
434         }
435         return FAILED;
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         struct completion     *eh_action;
445
446         SCSI_LOG_ERROR_RECOVERY(3,
447                 printk("%s scmd: %p result: %x\n",
448                         __FUNCTION__, scmd, scmd->result));
449
450         eh_action = scmd->device->host->eh_action;
451         if (eh_action)
452                 complete(eh_action);
453 }
454
455 /**
456  * scsi_send_eh_cmnd  - submit a scsi command as part of error recory
457  * @scmd:       SCSI command structure to hijack
458  * @cmnd:       CDB to send
459  * @cmnd_size:  size in bytes of @cmnd
460  * @timeout:    timeout for this request
461  * @copy_sense: request sense data if set to 1
462  *
463  * This function is used to send a scsi command down to a target device
464  * as part of the error recovery process.  If @copy_sense is 0 the command
465  * sent must be one that does not transfer any data.  If @copy_sense is 1
466  * the command must be REQUEST_SENSE and this functions copies out the
467  * sense buffer it got into @scmd->sense_buffer.
468  *
469  * Return value:
470  *    SUCCESS or FAILED or NEEDS_RETRY
471  **/
472 static int scsi_send_eh_cmnd(struct scsi_cmnd *scmd, unsigned char *cmnd,
473                              int cmnd_size, int timeout, int copy_sense)
474 {
475         struct scsi_device *sdev = scmd->device;
476         struct Scsi_Host *shost = sdev->host;
477         int old_result = scmd->result;
478         DECLARE_COMPLETION_ONSTACK(done);
479         unsigned long timeleft;
480         unsigned long flags;
481         struct scatterlist sgl;
482         unsigned char old_cmnd[MAX_COMMAND_SIZE];
483         enum dma_data_direction old_data_direction;
484         unsigned short old_use_sg;
485         unsigned char old_cmd_len;
486         unsigned old_bufflen;
487         void *old_buffer;
488         int rtn;
489
490         /*
491          * We need saved copies of a number of fields - this is because
492          * error handling may need to overwrite these with different values
493          * to run different commands, and once error handling is complete,
494          * we will need to restore these values prior to running the actual
495          * command.
496          */
497         old_buffer = scmd->request_buffer;
498         old_bufflen = scmd->request_bufflen;
499         memcpy(old_cmnd, scmd->cmnd, sizeof(scmd->cmnd));
500         old_data_direction = scmd->sc_data_direction;
501         old_cmd_len = scmd->cmd_len;
502         old_use_sg = scmd->use_sg;
503
504         memset(scmd->cmnd, 0, sizeof(scmd->cmnd));
505         memcpy(scmd->cmnd, cmnd, cmnd_size);
506
507         if (copy_sense) {
508                 gfp_t gfp_mask = GFP_ATOMIC;
509
510                 if (shost->hostt->unchecked_isa_dma)
511                         gfp_mask |= __GFP_DMA;
512
513                 sgl.page = alloc_page(gfp_mask);
514                 if (!sgl.page)
515                         return FAILED;
516                 sgl.offset = 0;
517                 sgl.length = 252;
518
519                 scmd->sc_data_direction = DMA_FROM_DEVICE;
520                 scmd->request_bufflen = sgl.length;
521                 scmd->request_buffer = &sgl;
522                 scmd->use_sg = 1;
523         } else {
524                 scmd->request_buffer = NULL;
525                 scmd->request_bufflen = 0;
526                 scmd->sc_data_direction = DMA_NONE;
527                 scmd->use_sg = 0;
528         }
529
530         scmd->underflow = 0;
531         scmd->cmd_len = COMMAND_SIZE(scmd->cmnd[0]);
532
533         if (sdev->scsi_level <= SCSI_2)
534                 scmd->cmnd[1] = (scmd->cmnd[1] & 0x1f) |
535                         (sdev->lun << 5 & 0xe0);
536
537         /*
538          * Zero the sense buffer.  The scsi spec mandates that any
539          * untransferred sense data should be interpreted as being zero.
540          */
541         memset(scmd->sense_buffer, 0, sizeof(scmd->sense_buffer));
542
543         shost->eh_action = &done;
544
545         spin_lock_irqsave(shost->host_lock, flags);
546         scsi_log_send(scmd);
547         shost->hostt->queuecommand(scmd, scsi_eh_done);
548         spin_unlock_irqrestore(shost->host_lock, flags);
549
550         timeleft = wait_for_completion_timeout(&done, timeout);
551
552         shost->eh_action = NULL;
553
554         scsi_log_completion(scmd, SUCCESS);
555
556         SCSI_LOG_ERROR_RECOVERY(3,
557                 printk("%s: scmd: %p, timeleft: %ld\n",
558                         __FUNCTION__, scmd, timeleft));
559
560         /*
561          * If there is time left scsi_eh_done got called, and we will
562          * examine the actual status codes to see whether the command
563          * actually did complete normally, else tell the host to forget
564          * about this command.
565          */
566         if (timeleft) {
567                 rtn = scsi_eh_completed_normally(scmd);
568                 SCSI_LOG_ERROR_RECOVERY(3,
569                         printk("%s: scsi_eh_completed_normally %x\n",
570                                __FUNCTION__, rtn));
571
572                 switch (rtn) {
573                 case SUCCESS:
574                 case NEEDS_RETRY:
575                 case FAILED:
576                         break;
577                 default:
578                         rtn = FAILED;
579                         break;
580                 }
581         } else {
582                 /*
583                  * FIXME(eric) - we are not tracking whether we could
584                  * abort a timed out command or not.  not sure how
585                  * we should treat them differently anyways.
586                  */
587                 if (shost->hostt->eh_abort_handler)
588                         shost->hostt->eh_abort_handler(scmd);
589                 rtn = FAILED;
590         }
591
592
593         /*
594          * Last chance to have valid sense data.
595          */
596         if (copy_sense) {
597                 if (!SCSI_SENSE_VALID(scmd)) {
598                         memcpy(scmd->sense_buffer, scmd->request_buffer,
599                                sizeof(scmd->sense_buffer));
600                 }
601                 __free_page(sgl.page);
602         }
603
604
605         /*
606          * Restore original data
607          */
608         scmd->request_buffer = old_buffer;
609         scmd->request_bufflen = old_bufflen;
610         memcpy(scmd->cmnd, old_cmnd, sizeof(scmd->cmnd));
611         scmd->sc_data_direction = old_data_direction;
612         scmd->cmd_len = old_cmd_len;
613         scmd->use_sg = old_use_sg;
614         scmd->result = old_result;
615         return rtn;
616 }
617
618 /**
619  * scsi_request_sense - Request sense data from a particular target.
620  * @scmd:       SCSI cmd for request sense.
621  *
622  * Notes:
623  *    Some hosts automatically obtain this information, others require
624  *    that we obtain it on our own. This function will *not* return until
625  *    the command either times out, or it completes.
626  **/
627 static int scsi_request_sense(struct scsi_cmnd *scmd)
628 {
629         static unsigned char generic_sense[6] =
630                 {REQUEST_SENSE, 0, 0, 0, 252, 0};
631
632         return scsi_send_eh_cmnd(scmd, generic_sense, 6, SENSE_TIMEOUT, 1);
633 }
634
635 /**
636  * scsi_eh_finish_cmd - Handle a cmd that eh is finished with.
637  * @scmd:       Original SCSI cmd that eh has finished.
638  * @done_q:     Queue for processed commands.
639  *
640  * Notes:
641  *    We don't want to use the normal command completion while we are are
642  *    still handling errors - it may cause other commands to be queued,
643  *    and that would disturb what we are doing.  thus we really want to
644  *    keep a list of pending commands for final completion, and once we
645  *    are ready to leave error handling we handle completion for real.
646  **/
647 void scsi_eh_finish_cmd(struct scsi_cmnd *scmd, struct list_head *done_q)
648 {
649         scmd->device->host->host_failed--;
650         scmd->eh_eflags = 0;
651         list_move_tail(&scmd->eh_entry, done_q);
652 }
653 EXPORT_SYMBOL(scsi_eh_finish_cmd);
654
655 /**
656  * scsi_eh_get_sense - Get device sense data.
657  * @work_q:     Queue of commands to process.
658  * @done_q:     Queue of proccessed commands..
659  *
660  * Description:
661  *    See if we need to request sense information.  if so, then get it
662  *    now, so we have a better idea of what to do.  
663  *
664  * Notes:
665  *    This has the unfortunate side effect that if a shost adapter does
666  *    not automatically request sense information, that we end up shutting
667  *    it down before we request it.
668  *
669  *    All drivers should request sense information internally these days,
670  *    so for now all I have to say is tough noogies if you end up in here.
671  *
672  *    XXX: Long term this code should go away, but that needs an audit of
673  *         all LLDDs first.
674  **/
675 static int scsi_eh_get_sense(struct list_head *work_q,
676                              struct list_head *done_q)
677 {
678         struct scsi_cmnd *scmd, *next;
679         int rtn;
680
681         list_for_each_entry_safe(scmd, next, work_q, eh_entry) {
682                 if ((scmd->eh_eflags & SCSI_EH_CANCEL_CMD) ||
683                     SCSI_SENSE_VALID(scmd))
684                         continue;
685
686                 SCSI_LOG_ERROR_RECOVERY(2, scmd_printk(KERN_INFO, scmd,
687                                                   "%s: requesting sense\n",
688                                                   current->comm));
689                 rtn = scsi_request_sense(scmd);
690                 if (rtn != SUCCESS)
691                         continue;
692
693                 SCSI_LOG_ERROR_RECOVERY(3, printk("sense requested for %p"
694                                                   " result %x\n", scmd,
695                                                   scmd->result));
696                 SCSI_LOG_ERROR_RECOVERY(3, scsi_print_sense("bh", scmd));
697
698                 rtn = scsi_decide_disposition(scmd);
699
700                 /*
701                  * if the result was normal, then just pass it along to the
702                  * upper level.
703                  */
704                 if (rtn == SUCCESS)
705                         /* we don't want this command reissued, just
706                          * finished with the sense data, so set
707                          * retries to the max allowed to ensure it
708                          * won't get reissued */
709                         scmd->retries = scmd->allowed;
710                 else if (rtn != NEEDS_RETRY)
711                         continue;
712
713                 scsi_eh_finish_cmd(scmd, done_q);
714         }
715
716         return list_empty(work_q);
717 }
718
719 /**
720  * scsi_try_to_abort_cmd - Ask host to abort a running command.
721  * @scmd:       SCSI cmd to abort from Lower Level.
722  *
723  * Notes:
724  *    This function will not return until the user's completion function
725  *    has been called.  there is no timeout on this operation.  if the
726  *    author of the low-level driver wishes this operation to be timed,
727  *    they can provide this facility themselves.  helper functions in
728  *    scsi_error.c can be supplied to make this easier to do.
729  **/
730 static int scsi_try_to_abort_cmd(struct scsi_cmnd *scmd)
731 {
732         if (!scmd->device->host->hostt->eh_abort_handler)
733                 return FAILED;
734
735         /*
736          * scsi_done was called just after the command timed out and before
737          * we had a chance to process it. (db)
738          */
739         if (scmd->serial_number == 0)
740                 return SUCCESS;
741         return scmd->device->host->hostt->eh_abort_handler(scmd);
742 }
743
744 /**
745  * scsi_eh_tur - Send TUR to device.
746  * @scmd:       Scsi cmd to send TUR
747  *
748  * Return value:
749  *    0 - Device is ready. 1 - Device NOT ready.
750  **/
751 static int scsi_eh_tur(struct scsi_cmnd *scmd)
752 {
753         static unsigned char tur_command[6] = {TEST_UNIT_READY, 0, 0, 0, 0, 0};
754         int retry_cnt = 1, rtn;
755
756 retry_tur:
757         rtn = scsi_send_eh_cmnd(scmd, tur_command, 6, SENSE_TIMEOUT, 0);
758
759         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: scmd %p rtn %x\n",
760                 __FUNCTION__, scmd, rtn));
761
762         switch (rtn) {
763         case NEEDS_RETRY:
764                 if (retry_cnt--)
765                         goto retry_tur;
766                 /*FALLTHRU*/
767         case SUCCESS:
768                 return 0;
769         default:
770                 return 1;
771         }
772 }
773
774 /**
775  * scsi_eh_abort_cmds - abort canceled commands.
776  * @shost:      scsi host being recovered.
777  * @eh_done_q:  list_head for processed commands.
778  *
779  * Decription:
780  *    Try and see whether or not it makes sense to try and abort the
781  *    running command.  this only works out to be the case if we have one
782  *    command that has timed out.  if the command simply failed, it makes
783  *    no sense to try and abort the command, since as far as the shost
784  *    adapter is concerned, it isn't running.
785  **/
786 static int scsi_eh_abort_cmds(struct list_head *work_q,
787                               struct list_head *done_q)
788 {
789         struct scsi_cmnd *scmd, *next;
790         int rtn;
791
792         list_for_each_entry_safe(scmd, next, work_q, eh_entry) {
793                 if (!(scmd->eh_eflags & SCSI_EH_CANCEL_CMD))
794                         continue;
795                 SCSI_LOG_ERROR_RECOVERY(3, printk("%s: aborting cmd:"
796                                                   "0x%p\n", current->comm,
797                                                   scmd));
798                 rtn = scsi_try_to_abort_cmd(scmd);
799                 if (rtn == SUCCESS) {
800                         scmd->eh_eflags &= ~SCSI_EH_CANCEL_CMD;
801                         if (!scsi_device_online(scmd->device) ||
802                             !scsi_eh_tur(scmd)) {
803                                 scsi_eh_finish_cmd(scmd, done_q);
804                         }
805                                 
806                 } else
807                         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: aborting"
808                                                           " cmd failed:"
809                                                           "0x%p\n",
810                                                           current->comm,
811                                                           scmd));
812         }
813
814         return list_empty(work_q);
815 }
816
817 /**
818  * scsi_try_bus_device_reset - Ask host to perform a BDR on a dev
819  * @scmd:       SCSI cmd used to send BDR       
820  *
821  * Notes:
822  *    There is no timeout for this operation.  if this operation is
823  *    unreliable for a given host, then the host itself needs to put a
824  *    timer on it, and set the host back to a consistent state prior to
825  *    returning.
826  **/
827 static int scsi_try_bus_device_reset(struct scsi_cmnd *scmd)
828 {
829         int rtn;
830
831         if (!scmd->device->host->hostt->eh_device_reset_handler)
832                 return FAILED;
833
834         rtn = scmd->device->host->hostt->eh_device_reset_handler(scmd);
835         if (rtn == SUCCESS) {
836                 scmd->device->was_reset = 1;
837                 scmd->device->expecting_cc_ua = 1;
838         }
839
840         return rtn;
841 }
842
843 /**
844  * scsi_eh_try_stu - Send START_UNIT to device.
845  * @scmd:       Scsi cmd to send START_UNIT
846  *
847  * Return value:
848  *    0 - Device is ready. 1 - Device NOT ready.
849  **/
850 static int scsi_eh_try_stu(struct scsi_cmnd *scmd)
851 {
852         static unsigned char stu_command[6] = {START_STOP, 0, 0, 0, 1, 0};
853
854         if (scmd->device->allow_restart) {
855                 int rtn;
856
857                 rtn = scsi_send_eh_cmnd(scmd, stu_command, 6,
858                                         START_UNIT_TIMEOUT, 0);
859                 if (rtn == SUCCESS)
860                         return 0;
861         }
862
863         return 1;
864 }
865
866  /**
867  * scsi_eh_stu - send START_UNIT if needed
868  * @shost:      scsi host being recovered.
869  * @eh_done_q:  list_head for processed commands.
870  *
871  * Notes:
872  *    If commands are failing due to not ready, initializing command required,
873  *      try revalidating the device, which will end up sending a start unit. 
874  **/
875 static int scsi_eh_stu(struct Scsi_Host *shost,
876                               struct list_head *work_q,
877                               struct list_head *done_q)
878 {
879         struct scsi_cmnd *scmd, *stu_scmd, *next;
880         struct scsi_device *sdev;
881
882         shost_for_each_device(sdev, shost) {
883                 stu_scmd = NULL;
884                 list_for_each_entry(scmd, work_q, eh_entry)
885                         if (scmd->device == sdev && SCSI_SENSE_VALID(scmd) &&
886                             scsi_check_sense(scmd) == FAILED ) {
887                                 stu_scmd = scmd;
888                                 break;
889                         }
890
891                 if (!stu_scmd)
892                         continue;
893
894                 SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Sending START_UNIT to sdev:"
895                                                   " 0x%p\n", current->comm, sdev));
896
897                 if (!scsi_eh_try_stu(stu_scmd)) {
898                         if (!scsi_device_online(sdev) ||
899                             !scsi_eh_tur(stu_scmd)) {
900                                 list_for_each_entry_safe(scmd, next,
901                                                           work_q, eh_entry) {
902                                         if (scmd->device == sdev)
903                                                 scsi_eh_finish_cmd(scmd, done_q);
904                                 }
905                         }
906                 } else {
907                         SCSI_LOG_ERROR_RECOVERY(3,
908                                                 printk("%s: START_UNIT failed to sdev:"
909                                                        " 0x%p\n", current->comm, sdev));
910                 }
911         }
912
913         return list_empty(work_q);
914 }
915
916
917 /**
918  * scsi_eh_bus_device_reset - send bdr if needed
919  * @shost:      scsi host being recovered.
920  * @eh_done_q:  list_head for processed commands.
921  *
922  * Notes:
923  *    Try a bus device reset.  still, look to see whether we have multiple
924  *    devices that are jammed or not - if we have multiple devices, it
925  *    makes no sense to try bus_device_reset - we really would need to try
926  *    a bus_reset instead. 
927  **/
928 static int scsi_eh_bus_device_reset(struct Scsi_Host *shost,
929                                     struct list_head *work_q,
930                                     struct list_head *done_q)
931 {
932         struct scsi_cmnd *scmd, *bdr_scmd, *next;
933         struct scsi_device *sdev;
934         int rtn;
935
936         shost_for_each_device(sdev, shost) {
937                 bdr_scmd = NULL;
938                 list_for_each_entry(scmd, work_q, eh_entry)
939                         if (scmd->device == sdev) {
940                                 bdr_scmd = scmd;
941                                 break;
942                         }
943
944                 if (!bdr_scmd)
945                         continue;
946
947                 SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Sending BDR sdev:"
948                                                   " 0x%p\n", current->comm,
949                                                   sdev));
950                 rtn = scsi_try_bus_device_reset(bdr_scmd);
951                 if (rtn == SUCCESS) {
952                         if (!scsi_device_online(sdev) ||
953                             !scsi_eh_tur(bdr_scmd)) {
954                                 list_for_each_entry_safe(scmd, next,
955                                                          work_q, eh_entry) {
956                                         if (scmd->device == sdev)
957                                                 scsi_eh_finish_cmd(scmd,
958                                                                    done_q);
959                                 }
960                         }
961                 } else {
962                         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: BDR"
963                                                           " failed sdev:"
964                                                           "0x%p\n",
965                                                           current->comm,
966                                                            sdev));
967                 }
968         }
969
970         return list_empty(work_q);
971 }
972
973 /**
974  * scsi_try_bus_reset - ask host to perform a bus reset
975  * @scmd:       SCSI cmd to send bus reset.
976  **/
977 static int scsi_try_bus_reset(struct scsi_cmnd *scmd)
978 {
979         unsigned long flags;
980         int rtn;
981
982         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Snd Bus RST\n",
983                                           __FUNCTION__));
984
985         if (!scmd->device->host->hostt->eh_bus_reset_handler)
986                 return FAILED;
987
988         rtn = scmd->device->host->hostt->eh_bus_reset_handler(scmd);
989
990         if (rtn == SUCCESS) {
991                 if (!scmd->device->host->hostt->skip_settle_delay)
992                         ssleep(BUS_RESET_SETTLE_TIME);
993                 spin_lock_irqsave(scmd->device->host->host_lock, flags);
994                 scsi_report_bus_reset(scmd->device->host,
995                                       scmd_channel(scmd));
996                 spin_unlock_irqrestore(scmd->device->host->host_lock, flags);
997         }
998
999         return rtn;
1000 }
1001
1002 /**
1003  * scsi_try_host_reset - ask host adapter to reset itself
1004  * @scmd:       SCSI cmd to send hsot reset.
1005  **/
1006 static int scsi_try_host_reset(struct scsi_cmnd *scmd)
1007 {
1008         unsigned long flags;
1009         int rtn;
1010
1011         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Snd Host RST\n",
1012                                           __FUNCTION__));
1013
1014         if (!scmd->device->host->hostt->eh_host_reset_handler)
1015                 return FAILED;
1016
1017         rtn = scmd->device->host->hostt->eh_host_reset_handler(scmd);
1018
1019         if (rtn == SUCCESS) {
1020                 if (!scmd->device->host->hostt->skip_settle_delay)
1021                         ssleep(HOST_RESET_SETTLE_TIME);
1022                 spin_lock_irqsave(scmd->device->host->host_lock, flags);
1023                 scsi_report_bus_reset(scmd->device->host,
1024                                       scmd_channel(scmd));
1025                 spin_unlock_irqrestore(scmd->device->host->host_lock, flags);
1026         }
1027
1028         return rtn;
1029 }
1030
1031 /**
1032  * scsi_eh_bus_reset - send a bus reset 
1033  * @shost:      scsi host being recovered.
1034  * @eh_done_q:  list_head for processed commands.
1035  **/
1036 static int scsi_eh_bus_reset(struct Scsi_Host *shost,
1037                              struct list_head *work_q,
1038                              struct list_head *done_q)
1039 {
1040         struct scsi_cmnd *scmd, *chan_scmd, *next;
1041         unsigned int channel;
1042         int rtn;
1043
1044         /*
1045          * we really want to loop over the various channels, and do this on
1046          * a channel by channel basis.  we should also check to see if any
1047          * of the failed commands are on soft_reset devices, and if so, skip
1048          * the reset.  
1049          */
1050
1051         for (channel = 0; channel <= shost->max_channel; channel++) {
1052                 chan_scmd = NULL;
1053                 list_for_each_entry(scmd, work_q, eh_entry) {
1054                         if (channel == scmd_channel(scmd)) {
1055                                 chan_scmd = scmd;
1056                                 break;
1057                                 /*
1058                                  * FIXME add back in some support for
1059                                  * soft_reset devices.
1060                                  */
1061                         }
1062                 }
1063
1064                 if (!chan_scmd)
1065                         continue;
1066                 SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Sending BRST chan:"
1067                                                   " %d\n", current->comm,
1068                                                   channel));
1069                 rtn = scsi_try_bus_reset(chan_scmd);
1070                 if (rtn == SUCCESS) {
1071                         list_for_each_entry_safe(scmd, next, work_q, eh_entry) {
1072                                 if (channel == scmd_channel(scmd))
1073                                         if (!scsi_device_online(scmd->device) ||
1074                                             !scsi_eh_tur(scmd))
1075                                                 scsi_eh_finish_cmd(scmd,
1076                                                                    done_q);
1077                         }
1078                 } else {
1079                         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: BRST"
1080                                                           " failed chan: %d\n",
1081                                                           current->comm,
1082                                                           channel));
1083                 }
1084         }
1085         return list_empty(work_q);
1086 }
1087
1088 /**
1089  * scsi_eh_host_reset - send a host reset 
1090  * @work_q:     list_head for processed commands.
1091  * @done_q:     list_head for processed commands.
1092  **/
1093 static int scsi_eh_host_reset(struct list_head *work_q,
1094                               struct list_head *done_q)
1095 {
1096         struct scsi_cmnd *scmd, *next;
1097         int rtn;
1098
1099         if (!list_empty(work_q)) {
1100                 scmd = list_entry(work_q->next,
1101                                   struct scsi_cmnd, eh_entry);
1102
1103                 SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Sending HRST\n"
1104                                                   , current->comm));
1105
1106                 rtn = scsi_try_host_reset(scmd);
1107                 if (rtn == SUCCESS) {
1108                         list_for_each_entry_safe(scmd, next, work_q, eh_entry) {
1109                                 if (!scsi_device_online(scmd->device) ||
1110                                     (!scsi_eh_try_stu(scmd) && !scsi_eh_tur(scmd)) ||
1111                                     !scsi_eh_tur(scmd))
1112                                         scsi_eh_finish_cmd(scmd, done_q);
1113                         }
1114                 } else {
1115                         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: HRST"
1116                                                           " failed\n",
1117                                                           current->comm));
1118                 }
1119         }
1120         return list_empty(work_q);
1121 }
1122
1123 /**
1124  * scsi_eh_offline_sdevs - offline scsi devices that fail to recover
1125  * @work_q:     list_head for processed commands.
1126  * @done_q:     list_head for processed commands.
1127  *
1128  **/
1129 static void scsi_eh_offline_sdevs(struct list_head *work_q,
1130                                   struct list_head *done_q)
1131 {
1132         struct scsi_cmnd *scmd, *next;
1133
1134         list_for_each_entry_safe(scmd, next, work_q, eh_entry) {
1135                 sdev_printk(KERN_INFO, scmd->device,
1136                             "scsi: Device offlined - not"
1137                             " ready after error recovery\n");
1138                 scsi_device_set_state(scmd->device, SDEV_OFFLINE);
1139                 if (scmd->eh_eflags & SCSI_EH_CANCEL_CMD) {
1140                         /*
1141                          * FIXME: Handle lost cmds.
1142                          */
1143                 }
1144                 scsi_eh_finish_cmd(scmd, done_q);
1145         }
1146         return;
1147 }
1148
1149 /**
1150  * scsi_decide_disposition - Disposition a cmd on return from LLD.
1151  * @scmd:       SCSI cmd to examine.
1152  *
1153  * Notes:
1154  *    This is *only* called when we are examining the status after sending
1155  *    out the actual data command.  any commands that are queued for error
1156  *    recovery (e.g. test_unit_ready) do *not* come through here.
1157  *
1158  *    When this routine returns failed, it means the error handler thread
1159  *    is woken.  In cases where the error code indicates an error that
1160  *    doesn't require the error handler read (i.e. we don't need to
1161  *    abort/reset), this function should return SUCCESS.
1162  **/
1163 int scsi_decide_disposition(struct scsi_cmnd *scmd)
1164 {
1165         int rtn;
1166
1167         /*
1168          * if the device is offline, then we clearly just pass the result back
1169          * up to the top level.
1170          */
1171         if (!scsi_device_online(scmd->device)) {
1172                 SCSI_LOG_ERROR_RECOVERY(5, printk("%s: device offline - report"
1173                                                   " as SUCCESS\n",
1174                                                   __FUNCTION__));
1175                 return SUCCESS;
1176         }
1177
1178         /*
1179          * first check the host byte, to see if there is anything in there
1180          * that would indicate what we need to do.
1181          */
1182         switch (host_byte(scmd->result)) {
1183         case DID_PASSTHROUGH:
1184                 /*
1185                  * no matter what, pass this through to the upper layer.
1186                  * nuke this special code so that it looks like we are saying
1187                  * did_ok.
1188                  */
1189                 scmd->result &= 0xff00ffff;
1190                 return SUCCESS;
1191         case DID_OK:
1192                 /*
1193                  * looks good.  drop through, and check the next byte.
1194                  */
1195                 break;
1196         case DID_NO_CONNECT:
1197         case DID_BAD_TARGET:
1198         case DID_ABORT:
1199                 /*
1200                  * note - this means that we just report the status back
1201                  * to the top level driver, not that we actually think
1202                  * that it indicates SUCCESS.
1203                  */
1204                 return SUCCESS;
1205                 /*
1206                  * when the low level driver returns did_soft_error,
1207                  * it is responsible for keeping an internal retry counter 
1208                  * in order to avoid endless loops (db)
1209                  *
1210                  * actually this is a bug in this function here.  we should
1211                  * be mindful of the maximum number of retries specified
1212                  * and not get stuck in a loop.
1213                  */
1214         case DID_SOFT_ERROR:
1215                 goto maybe_retry;
1216         case DID_IMM_RETRY:
1217                 return NEEDS_RETRY;
1218
1219         case DID_REQUEUE:
1220                 return ADD_TO_MLQUEUE;
1221
1222         case DID_ERROR:
1223                 if (msg_byte(scmd->result) == COMMAND_COMPLETE &&
1224                     status_byte(scmd->result) == RESERVATION_CONFLICT)
1225                         /*
1226                          * execute reservation conflict processing code
1227                          * lower down
1228                          */
1229                         break;
1230                 /* fallthrough */
1231
1232         case DID_BUS_BUSY:
1233         case DID_PARITY:
1234                 goto maybe_retry;
1235         case DID_TIME_OUT:
1236                 /*
1237                  * when we scan the bus, we get timeout messages for
1238                  * these commands if there is no device available.
1239                  * other hosts report did_no_connect for the same thing.
1240                  */
1241                 if ((scmd->cmnd[0] == TEST_UNIT_READY ||
1242                      scmd->cmnd[0] == INQUIRY)) {
1243                         return SUCCESS;
1244                 } else {
1245                         return FAILED;
1246                 }
1247         case DID_RESET:
1248                 return SUCCESS;
1249         default:
1250                 return FAILED;
1251         }
1252
1253         /*
1254          * next, check the message byte.
1255          */
1256         if (msg_byte(scmd->result) != COMMAND_COMPLETE)
1257                 return FAILED;
1258
1259         /*
1260          * check the status byte to see if this indicates anything special.
1261          */
1262         switch (status_byte(scmd->result)) {
1263         case QUEUE_FULL:
1264                 /*
1265                  * the case of trying to send too many commands to a
1266                  * tagged queueing device.
1267                  */
1268         case BUSY:
1269                 /*
1270                  * device can't talk to us at the moment.  Should only
1271                  * occur (SAM-3) when the task queue is empty, so will cause
1272                  * the empty queue handling to trigger a stall in the
1273                  * device.
1274                  */
1275                 return ADD_TO_MLQUEUE;
1276         case GOOD:
1277         case COMMAND_TERMINATED:
1278         case TASK_ABORTED:
1279                 return SUCCESS;
1280         case CHECK_CONDITION:
1281                 rtn = scsi_check_sense(scmd);
1282                 if (rtn == NEEDS_RETRY)
1283                         goto maybe_retry;
1284                 /* if rtn == FAILED, we have no sense information;
1285                  * returning FAILED will wake the error handler thread
1286                  * to collect the sense and redo the decide
1287                  * disposition */
1288                 return rtn;
1289         case CONDITION_GOOD:
1290         case INTERMEDIATE_GOOD:
1291         case INTERMEDIATE_C_GOOD:
1292         case ACA_ACTIVE:
1293                 /*
1294                  * who knows?  FIXME(eric)
1295                  */
1296                 return SUCCESS;
1297
1298         case RESERVATION_CONFLICT:
1299                 sdev_printk(KERN_INFO, scmd->device,
1300                             "reservation conflict\n");
1301                 return SUCCESS; /* causes immediate i/o error */
1302         default:
1303                 return FAILED;
1304         }
1305         return FAILED;
1306
1307       maybe_retry:
1308
1309         /* we requeue for retry because the error was retryable, and
1310          * the request was not marked fast fail.  Note that above,
1311          * even if the request is marked fast fail, we still requeue
1312          * for queue congestion conditions (QUEUE_FULL or BUSY) */
1313         if ((++scmd->retries) <= scmd->allowed
1314             && !blk_noretry_request(scmd->request)) {
1315                 return NEEDS_RETRY;
1316         } else {
1317                 /*
1318                  * no more retries - report this one back to upper level.
1319                  */
1320                 return SUCCESS;
1321         }
1322 }
1323
1324 /**
1325  * scsi_eh_lock_door - Prevent medium removal for the specified device
1326  * @sdev:       SCSI device to prevent medium removal
1327  *
1328  * Locking:
1329  *      We must be called from process context; scsi_allocate_request()
1330  *      may sleep.
1331  *
1332  * Notes:
1333  *      We queue up an asynchronous "ALLOW MEDIUM REMOVAL" request on the
1334  *      head of the devices request queue, and continue.
1335  *
1336  * Bugs:
1337  *      scsi_allocate_request() may sleep waiting for existing requests to
1338  *      be processed.  However, since we haven't kicked off any request
1339  *      processing for this host, this may deadlock.
1340  *
1341  *      If scsi_allocate_request() fails for what ever reason, we
1342  *      completely forget to lock the door.
1343  **/
1344 static void scsi_eh_lock_door(struct scsi_device *sdev)
1345 {
1346         unsigned char cmnd[MAX_COMMAND_SIZE];
1347
1348         cmnd[0] = ALLOW_MEDIUM_REMOVAL;
1349         cmnd[1] = 0;
1350         cmnd[2] = 0;
1351         cmnd[3] = 0;
1352         cmnd[4] = SCSI_REMOVAL_PREVENT;
1353         cmnd[5] = 0;
1354
1355         scsi_execute_async(sdev, cmnd, 6, DMA_NONE, NULL, 0, 0, 10 * HZ,
1356                            5, NULL, NULL, GFP_KERNEL);
1357 }
1358
1359
1360 /**
1361  * scsi_restart_operations - restart io operations to the specified host.
1362  * @shost:      Host we are restarting.
1363  *
1364  * Notes:
1365  *    When we entered the error handler, we blocked all further i/o to
1366  *    this device.  we need to 'reverse' this process.
1367  **/
1368 static void scsi_restart_operations(struct Scsi_Host *shost)
1369 {
1370         struct scsi_device *sdev;
1371         unsigned long flags;
1372
1373         /*
1374          * If the door was locked, we need to insert a door lock request
1375          * onto the head of the SCSI request queue for the device.  There
1376          * is no point trying to lock the door of an off-line device.
1377          */
1378         shost_for_each_device(sdev, shost) {
1379                 if (scsi_device_online(sdev) && sdev->locked)
1380                         scsi_eh_lock_door(sdev);
1381         }
1382
1383         /*
1384          * next free up anything directly waiting upon the host.  this
1385          * will be requests for character device operations, and also for
1386          * ioctls to queued block devices.
1387          */
1388         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: waking up host to restart\n",
1389                                           __FUNCTION__));
1390
1391         spin_lock_irqsave(shost->host_lock, flags);
1392         if (scsi_host_set_state(shost, SHOST_RUNNING))
1393                 if (scsi_host_set_state(shost, SHOST_CANCEL))
1394                         BUG_ON(scsi_host_set_state(shost, SHOST_DEL));
1395         spin_unlock_irqrestore(shost->host_lock, flags);
1396
1397         wake_up(&shost->host_wait);
1398
1399         /*
1400          * finally we need to re-initiate requests that may be pending.  we will
1401          * have had everything blocked while error handling is taking place, and
1402          * now that error recovery is done, we will need to ensure that these
1403          * requests are started.
1404          */
1405         scsi_run_host_queues(shost);
1406 }
1407
1408 /**
1409  * scsi_eh_ready_devs - check device ready state and recover if not.
1410  * @shost:      host to be recovered.
1411  * @eh_done_q:  list_head for processed commands.
1412  *
1413  **/
1414 static void scsi_eh_ready_devs(struct Scsi_Host *shost,
1415                                struct list_head *work_q,
1416                                struct list_head *done_q)
1417 {
1418         if (!scsi_eh_stu(shost, work_q, done_q))
1419                 if (!scsi_eh_bus_device_reset(shost, work_q, done_q))
1420                         if (!scsi_eh_bus_reset(shost, work_q, done_q))
1421                                 if (!scsi_eh_host_reset(work_q, done_q))
1422                                         scsi_eh_offline_sdevs(work_q, done_q);
1423 }
1424
1425 /**
1426  * scsi_eh_flush_done_q - finish processed commands or retry them.
1427  * @done_q:     list_head of processed commands.
1428  *
1429  **/
1430 void scsi_eh_flush_done_q(struct list_head *done_q)
1431 {
1432         struct scsi_cmnd *scmd, *next;
1433
1434         list_for_each_entry_safe(scmd, next, done_q, eh_entry) {
1435                 list_del_init(&scmd->eh_entry);
1436                 if (scsi_device_online(scmd->device) &&
1437                     !blk_noretry_request(scmd->request) &&
1438                     (++scmd->retries <= scmd->allowed)) {
1439                         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: flush"
1440                                                           " retry cmd: %p\n",
1441                                                           current->comm,
1442                                                           scmd));
1443                                 scsi_queue_insert(scmd, SCSI_MLQUEUE_EH_RETRY);
1444                 } else {
1445                         /*
1446                          * If just we got sense for the device (called
1447                          * scsi_eh_get_sense), scmd->result is already
1448                          * set, do not set DRIVER_TIMEOUT.
1449                          */
1450                         if (!scmd->result)
1451                                 scmd->result |= (DRIVER_TIMEOUT << 24);
1452                         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: flush finish"
1453                                                         " cmd: %p\n",
1454                                                         current->comm, scmd));
1455                         scsi_finish_command(scmd);
1456                 }
1457         }
1458 }
1459 EXPORT_SYMBOL(scsi_eh_flush_done_q);
1460
1461 /**
1462  * scsi_unjam_host - Attempt to fix a host which has a cmd that failed.
1463  * @shost:      Host to unjam.
1464  *
1465  * Notes:
1466  *    When we come in here, we *know* that all commands on the bus have
1467  *    either completed, failed or timed out.  we also know that no further
1468  *    commands are being sent to the host, so things are relatively quiet
1469  *    and we have freedom to fiddle with things as we wish.
1470  *
1471  *    This is only the *default* implementation.  it is possible for
1472  *    individual drivers to supply their own version of this function, and
1473  *    if the maintainer wishes to do this, it is strongly suggested that
1474  *    this function be taken as a template and modified.  this function
1475  *    was designed to correctly handle problems for about 95% of the
1476  *    different cases out there, and it should always provide at least a
1477  *    reasonable amount of error recovery.
1478  *
1479  *    Any command marked 'failed' or 'timeout' must eventually have
1480  *    scsi_finish_cmd() called for it.  we do all of the retry stuff
1481  *    here, so when we restart the host after we return it should have an
1482  *    empty queue.
1483  **/
1484 static void scsi_unjam_host(struct Scsi_Host *shost)
1485 {
1486         unsigned long flags;
1487         LIST_HEAD(eh_work_q);
1488         LIST_HEAD(eh_done_q);
1489
1490         spin_lock_irqsave(shost->host_lock, flags);
1491         list_splice_init(&shost->eh_cmd_q, &eh_work_q);
1492         spin_unlock_irqrestore(shost->host_lock, flags);
1493
1494         SCSI_LOG_ERROR_RECOVERY(1, scsi_eh_prt_fail_stats(shost, &eh_work_q));
1495
1496         if (!scsi_eh_get_sense(&eh_work_q, &eh_done_q))
1497                 if (!scsi_eh_abort_cmds(&eh_work_q, &eh_done_q))
1498                         scsi_eh_ready_devs(shost, &eh_work_q, &eh_done_q);
1499
1500         scsi_eh_flush_done_q(&eh_done_q);
1501 }
1502
1503 /**
1504  * scsi_error_handler - SCSI error handler thread
1505  * @data:       Host for which we are running.
1506  *
1507  * Notes:
1508  *    This is the main error handling loop.  This is run as a kernel thread
1509  *    for every SCSI host and handles all error handling activity.
1510  **/
1511 int scsi_error_handler(void *data)
1512 {
1513         struct Scsi_Host *shost = data;
1514
1515         current->flags |= PF_NOFREEZE;
1516
1517         /*
1518          * We use TASK_INTERRUPTIBLE so that the thread is not
1519          * counted against the load average as a running process.
1520          * We never actually get interrupted because kthread_run
1521          * disables singal delivery for the created thread.
1522          */
1523         set_current_state(TASK_INTERRUPTIBLE);
1524         while (!kthread_should_stop()) {
1525                 if ((shost->host_failed == 0 && shost->host_eh_scheduled == 0) ||
1526                     shost->host_failed != shost->host_busy) {
1527                         SCSI_LOG_ERROR_RECOVERY(1,
1528                                 printk("Error handler scsi_eh_%d sleeping\n",
1529                                         shost->host_no));
1530                         schedule();
1531                         set_current_state(TASK_INTERRUPTIBLE);
1532                         continue;
1533                 }
1534
1535                 __set_current_state(TASK_RUNNING);
1536                 SCSI_LOG_ERROR_RECOVERY(1,
1537                         printk("Error handler scsi_eh_%d waking up\n",
1538                                 shost->host_no));
1539
1540                 /*
1541                  * We have a host that is failing for some reason.  Figure out
1542                  * what we need to do to get it up and online again (if we can).
1543                  * If we fail, we end up taking the thing offline.
1544                  */
1545                 if (shost->transportt->eh_strategy_handler)
1546                         shost->transportt->eh_strategy_handler(shost);
1547                 else
1548                         scsi_unjam_host(shost);
1549
1550                 /*
1551                  * Note - if the above fails completely, the action is to take
1552                  * individual devices offline and flush the queue of any
1553                  * outstanding requests that may have been pending.  When we
1554                  * restart, we restart any I/O to any other devices on the bus
1555                  * which are still online.
1556                  */
1557                 scsi_restart_operations(shost);
1558                 set_current_state(TASK_INTERRUPTIBLE);
1559         }
1560         __set_current_state(TASK_RUNNING);
1561
1562         SCSI_LOG_ERROR_RECOVERY(1,
1563                 printk("Error handler scsi_eh_%d exiting\n", shost->host_no));
1564         shost->ehandler = NULL;
1565         return 0;
1566 }
1567
1568 /*
1569  * Function:    scsi_report_bus_reset()
1570  *
1571  * Purpose:     Utility function used by low-level drivers to report that
1572  *              they have observed a bus reset on the bus being handled.
1573  *
1574  * Arguments:   shost       - Host in question
1575  *              channel     - channel on which reset was observed.
1576  *
1577  * Returns:     Nothing
1578  *
1579  * Lock status: Host lock must be held.
1580  *
1581  * Notes:       This only needs to be called if the reset is one which
1582  *              originates from an unknown location.  Resets originated
1583  *              by the mid-level itself don't need to call this, but there
1584  *              should be no harm.
1585  *
1586  *              The main purpose of this is to make sure that a CHECK_CONDITION
1587  *              is properly treated.
1588  */
1589 void scsi_report_bus_reset(struct Scsi_Host *shost, int channel)
1590 {
1591         struct scsi_device *sdev;
1592
1593         __shost_for_each_device(sdev, shost) {
1594                 if (channel == sdev_channel(sdev)) {
1595                         sdev->was_reset = 1;
1596                         sdev->expecting_cc_ua = 1;
1597                 }
1598         }
1599 }
1600 EXPORT_SYMBOL(scsi_report_bus_reset);
1601
1602 /*
1603  * Function:    scsi_report_device_reset()
1604  *
1605  * Purpose:     Utility function used by low-level drivers to report that
1606  *              they have observed a device reset on the device being handled.
1607  *
1608  * Arguments:   shost       - Host in question
1609  *              channel     - channel on which reset was observed
1610  *              target      - target on which reset was observed
1611  *
1612  * Returns:     Nothing
1613  *
1614  * Lock status: Host lock must be held
1615  *
1616  * Notes:       This only needs to be called if the reset is one which
1617  *              originates from an unknown location.  Resets originated
1618  *              by the mid-level itself don't need to call this, but there
1619  *              should be no harm.
1620  *
1621  *              The main purpose of this is to make sure that a CHECK_CONDITION
1622  *              is properly treated.
1623  */
1624 void scsi_report_device_reset(struct Scsi_Host *shost, int channel, int target)
1625 {
1626         struct scsi_device *sdev;
1627
1628         __shost_for_each_device(sdev, shost) {
1629                 if (channel == sdev_channel(sdev) &&
1630                     target == sdev_id(sdev)) {
1631                         sdev->was_reset = 1;
1632                         sdev->expecting_cc_ua = 1;
1633                 }
1634         }
1635 }
1636 EXPORT_SYMBOL(scsi_report_device_reset);
1637
1638 static void
1639 scsi_reset_provider_done_command(struct scsi_cmnd *scmd)
1640 {
1641 }
1642
1643 /*
1644  * Function:    scsi_reset_provider
1645  *
1646  * Purpose:     Send requested reset to a bus or device at any phase.
1647  *
1648  * Arguments:   device  - device to send reset to
1649  *              flag - reset type (see scsi.h)
1650  *
1651  * Returns:     SUCCESS/FAILURE.
1652  *
1653  * Notes:       This is used by the SCSI Generic driver to provide
1654  *              Bus/Device reset capability.
1655  */
1656 int
1657 scsi_reset_provider(struct scsi_device *dev, int flag)
1658 {
1659         struct scsi_cmnd *scmd = scsi_get_command(dev, GFP_KERNEL);
1660         struct Scsi_Host *shost = dev->host;
1661         struct request req;
1662         unsigned long flags;
1663         int rtn;
1664
1665         scmd->request = &req;
1666         memset(&scmd->eh_timeout, 0, sizeof(scmd->eh_timeout));
1667
1668         memset(&scmd->cmnd, '\0', sizeof(scmd->cmnd));
1669     
1670         scmd->scsi_done         = scsi_reset_provider_done_command;
1671         scmd->done                      = NULL;
1672         scmd->request_buffer            = NULL;
1673         scmd->request_bufflen           = 0;
1674
1675         scmd->cmd_len                   = 0;
1676
1677         scmd->sc_data_direction         = DMA_BIDIRECTIONAL;
1678
1679         init_timer(&scmd->eh_timeout);
1680
1681         /*
1682          * Sometimes the command can get back into the timer chain,
1683          * so use the pid as an identifier.
1684          */
1685         scmd->pid                       = 0;
1686
1687         spin_lock_irqsave(shost->host_lock, flags);
1688         shost->tmf_in_progress = 1;
1689         spin_unlock_irqrestore(shost->host_lock, flags);
1690
1691         switch (flag) {
1692         case SCSI_TRY_RESET_DEVICE:
1693                 rtn = scsi_try_bus_device_reset(scmd);
1694                 if (rtn == SUCCESS)
1695                         break;
1696                 /* FALLTHROUGH */
1697         case SCSI_TRY_RESET_BUS:
1698                 rtn = scsi_try_bus_reset(scmd);
1699                 if (rtn == SUCCESS)
1700                         break;
1701                 /* FALLTHROUGH */
1702         case SCSI_TRY_RESET_HOST:
1703                 rtn = scsi_try_host_reset(scmd);
1704                 break;
1705         default:
1706                 rtn = FAILED;
1707         }
1708
1709         spin_lock_irqsave(shost->host_lock, flags);
1710         shost->tmf_in_progress = 0;
1711         spin_unlock_irqrestore(shost->host_lock, flags);
1712
1713         /*
1714          * be sure to wake up anyone who was sleeping or had their queue
1715          * suspended while we performed the TMF.
1716          */
1717         SCSI_LOG_ERROR_RECOVERY(3,
1718                 printk("%s: waking up host to restart after TMF\n",
1719                 __FUNCTION__));
1720
1721         wake_up(&shost->host_wait);
1722
1723         scsi_run_host_queues(shost);
1724
1725         scsi_next_command(scmd);
1726         return rtn;
1727 }
1728 EXPORT_SYMBOL(scsi_reset_provider);
1729
1730 /**
1731  * scsi_normalize_sense - normalize main elements from either fixed or
1732  *                      descriptor sense data format into a common format.
1733  *
1734  * @sense_buffer:       byte array containing sense data returned by device
1735  * @sb_len:             number of valid bytes in sense_buffer
1736  * @sshdr:              pointer to instance of structure that common
1737  *                      elements are written to.
1738  *
1739  * Notes:
1740  *      The "main elements" from sense data are: response_code, sense_key,
1741  *      asc, ascq and additional_length (only for descriptor format).
1742  *
1743  *      Typically this function can be called after a device has
1744  *      responded to a SCSI command with the CHECK_CONDITION status.
1745  *
1746  * Return value:
1747  *      1 if valid sense data information found, else 0;
1748  **/
1749 int scsi_normalize_sense(const u8 *sense_buffer, int sb_len,
1750                          struct scsi_sense_hdr *sshdr)
1751 {
1752         if (!sense_buffer || !sb_len)
1753                 return 0;
1754
1755         memset(sshdr, 0, sizeof(struct scsi_sense_hdr));
1756
1757         sshdr->response_code = (sense_buffer[0] & 0x7f);
1758
1759         if (!scsi_sense_valid(sshdr))
1760                 return 0;
1761
1762         if (sshdr->response_code >= 0x72) {
1763                 /*
1764                  * descriptor format
1765                  */
1766                 if (sb_len > 1)
1767                         sshdr->sense_key = (sense_buffer[1] & 0xf);
1768                 if (sb_len > 2)
1769                         sshdr->asc = sense_buffer[2];
1770                 if (sb_len > 3)
1771                         sshdr->ascq = sense_buffer[3];
1772                 if (sb_len > 7)
1773                         sshdr->additional_length = sense_buffer[7];
1774         } else {
1775                 /* 
1776                  * fixed format
1777                  */
1778                 if (sb_len > 2)
1779                         sshdr->sense_key = (sense_buffer[2] & 0xf);
1780                 if (sb_len > 7) {
1781                         sb_len = (sb_len < (sense_buffer[7] + 8)) ?
1782                                          sb_len : (sense_buffer[7] + 8);
1783                         if (sb_len > 12)
1784                                 sshdr->asc = sense_buffer[12];
1785                         if (sb_len > 13)
1786                                 sshdr->ascq = sense_buffer[13];
1787                 }
1788         }
1789
1790         return 1;
1791 }
1792 EXPORT_SYMBOL(scsi_normalize_sense);
1793
1794 int scsi_command_normalize_sense(struct scsi_cmnd *cmd,
1795                                  struct scsi_sense_hdr *sshdr)
1796 {
1797         return scsi_normalize_sense(cmd->sense_buffer,
1798                         sizeof(cmd->sense_buffer), sshdr);
1799 }
1800 EXPORT_SYMBOL(scsi_command_normalize_sense);
1801
1802 /**
1803  * scsi_sense_desc_find - search for a given descriptor type in
1804  *                      descriptor sense data format.
1805  *
1806  * @sense_buffer:       byte array of descriptor format sense data
1807  * @sb_len:             number of valid bytes in sense_buffer
1808  * @desc_type:          value of descriptor type to find
1809  *                      (e.g. 0 -> information)
1810  *
1811  * Notes:
1812  *      only valid when sense data is in descriptor format
1813  *
1814  * Return value:
1815  *      pointer to start of (first) descriptor if found else NULL
1816  **/
1817 const u8 * scsi_sense_desc_find(const u8 * sense_buffer, int sb_len,
1818                                 int desc_type)
1819 {
1820         int add_sen_len, add_len, desc_len, k;
1821         const u8 * descp;
1822
1823         if ((sb_len < 8) || (0 == (add_sen_len = sense_buffer[7])))
1824                 return NULL;
1825         if ((sense_buffer[0] < 0x72) || (sense_buffer[0] > 0x73))
1826                 return NULL;
1827         add_sen_len = (add_sen_len < (sb_len - 8)) ?
1828                         add_sen_len : (sb_len - 8);
1829         descp = &sense_buffer[8];
1830         for (desc_len = 0, k = 0; k < add_sen_len; k += desc_len) {
1831                 descp += desc_len;
1832                 add_len = (k < (add_sen_len - 1)) ? descp[1]: -1;
1833                 desc_len = add_len + 2;
1834                 if (descp[0] == desc_type)
1835                         return descp;
1836                 if (add_len < 0) // short descriptor ??
1837                         break;
1838         }
1839         return NULL;
1840 }
1841 EXPORT_SYMBOL(scsi_sense_desc_find);
1842
1843 /**
1844  * scsi_get_sense_info_fld - attempts to get information field from
1845  *                      sense data (either fixed or descriptor format)
1846  *
1847  * @sense_buffer:       byte array of sense data
1848  * @sb_len:             number of valid bytes in sense_buffer
1849  * @info_out:           pointer to 64 integer where 8 or 4 byte information
1850  *                      field will be placed if found.
1851  *
1852  * Return value:
1853  *      1 if information field found, 0 if not found.
1854  **/
1855 int scsi_get_sense_info_fld(const u8 * sense_buffer, int sb_len,
1856                             u64 * info_out)
1857 {
1858         int j;
1859         const u8 * ucp;
1860         u64 ull;
1861
1862         if (sb_len < 7)
1863                 return 0;
1864         switch (sense_buffer[0] & 0x7f) {
1865         case 0x70:
1866         case 0x71:
1867                 if (sense_buffer[0] & 0x80) {
1868                         *info_out = (sense_buffer[3] << 24) +
1869                                     (sense_buffer[4] << 16) +
1870                                     (sense_buffer[5] << 8) + sense_buffer[6];
1871                         return 1;
1872                 } else
1873                         return 0;
1874         case 0x72:
1875         case 0x73:
1876                 ucp = scsi_sense_desc_find(sense_buffer, sb_len,
1877                                            0 /* info desc */);
1878                 if (ucp && (0xa == ucp[1])) {
1879                         ull = 0;
1880                         for (j = 0; j < 8; ++j) {
1881                                 if (j > 0)
1882                                         ull <<= 8;
1883                                 ull |= ucp[4 + j];
1884                         }
1885                         *info_out = ull;
1886                         return 1;
1887                 } else
1888                         return 0;
1889         default:
1890                 return 0;
1891         }
1892 }
1893 EXPORT_SYMBOL(scsi_get_sense_info_fld);