Merge git://git.kernel.org/pub/scm/linux/kernel/git/mingo/linux-2.6-sched
[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/kernel.h>
22 #include <linux/freezer.h>
23 #include <linux/kthread.h>
24 #include <linux/interrupt.h>
25 #include <linux/blkdev.h>
26 #include <linux/delay.h>
27 #include <linux/scatterlist.h>
28
29 #include <scsi/scsi.h>
30 #include <scsi/scsi_cmnd.h>
31 #include <scsi/scsi_dbg.h>
32 #include <scsi/scsi_device.h>
33 #include <scsi/scsi_eh.h>
34 #include <scsi/scsi_transport.h>
35 #include <scsi/scsi_host.h>
36 #include <scsi/scsi_ioctl.h>
37
38 #include "scsi_priv.h"
39 #include "scsi_logging.h"
40 #include "scsi_transport_api.h"
41
42 #define SENSE_TIMEOUT           (10*HZ)
43
44 /*
45  * These should *probably* be handled by the host itself.
46  * Since it is allowed to sleep, it probably should.
47  */
48 #define BUS_RESET_SETTLE_TIME   (10)
49 #define HOST_RESET_SETTLE_TIME  (10)
50
51 /* called with shost->host_lock held */
52 void scsi_eh_wakeup(struct Scsi_Host *shost)
53 {
54         if (shost->host_busy == shost->host_failed) {
55                 wake_up_process(shost->ehandler);
56                 SCSI_LOG_ERROR_RECOVERY(5,
57                                 printk("Waking error handler thread\n"));
58         }
59 }
60
61 /**
62  * scsi_schedule_eh - schedule EH for SCSI host
63  * @shost:      SCSI host to invoke error handling on.
64  *
65  * Schedule SCSI EH without scmd.
66  **/
67 void scsi_schedule_eh(struct Scsi_Host *shost)
68 {
69         unsigned long flags;
70
71         spin_lock_irqsave(shost->host_lock, flags);
72
73         if (scsi_host_set_state(shost, SHOST_RECOVERY) == 0 ||
74             scsi_host_set_state(shost, SHOST_CANCEL_RECOVERY) == 0) {
75                 shost->host_eh_scheduled++;
76                 scsi_eh_wakeup(shost);
77         }
78
79         spin_unlock_irqrestore(shost->host_lock, flags);
80 }
81 EXPORT_SYMBOL_GPL(scsi_schedule_eh);
82
83 /**
84  * scsi_eh_scmd_add - add scsi cmd to error handling.
85  * @scmd:       scmd to run eh on.
86  * @eh_flag:    optional SCSI_EH flag.
87  *
88  * Return value:
89  *      0 on failure.
90  **/
91 int scsi_eh_scmd_add(struct scsi_cmnd *scmd, int eh_flag)
92 {
93         struct Scsi_Host *shost = scmd->device->host;
94         unsigned long flags;
95         int ret = 0;
96
97         if (!shost->ehandler)
98                 return 0;
99
100         spin_lock_irqsave(shost->host_lock, flags);
101         if (scsi_host_set_state(shost, SHOST_RECOVERY))
102                 if (scsi_host_set_state(shost, SHOST_CANCEL_RECOVERY))
103                         goto out_unlock;
104
105         ret = 1;
106         scmd->eh_eflags |= eh_flag;
107         list_add_tail(&scmd->eh_entry, &shost->eh_cmd_q);
108         shost->host_failed++;
109         scsi_eh_wakeup(shost);
110  out_unlock:
111         spin_unlock_irqrestore(shost->host_lock, flags);
112         return ret;
113 }
114
115 /**
116  * scsi_add_timer - Start timeout timer for a single scsi command.
117  * @scmd:       scsi command that is about to start running.
118  * @timeout:    amount of time to allow this command to run.
119  * @complete:   timeout function to call if timer isn't canceled.
120  *
121  * Notes:
122  *    This should be turned into an inline function.  Each scsi command
123  *    has its own timer, and as it is added to the queue, we set up the
124  *    timer.  When the command completes, we cancel the timer.
125  **/
126 void scsi_add_timer(struct scsi_cmnd *scmd, int timeout,
127                     void (*complete)(struct scsi_cmnd *))
128 {
129
130         /*
131          * If the clock was already running for this command, then
132          * first delete the timer.  The timer handling code gets rather
133          * confused if we don't do this.
134          */
135         if (scmd->eh_timeout.function)
136                 del_timer(&scmd->eh_timeout);
137
138         scmd->eh_timeout.data = (unsigned long)scmd;
139         scmd->eh_timeout.expires = jiffies + timeout;
140         scmd->eh_timeout.function = (void (*)(unsigned long)) complete;
141
142         SCSI_LOG_ERROR_RECOVERY(5, printk("%s: scmd: %p, time:"
143                                           " %d, (%p)\n", __FUNCTION__,
144                                           scmd, timeout, complete));
145
146         add_timer(&scmd->eh_timeout);
147 }
148
149 /**
150  * scsi_delete_timer - Delete/cancel timer for a given function.
151  * @scmd:       Cmd that we are canceling timer for
152  *
153  * Notes:
154  *     This should be turned into an inline function.
155  *
156  * Return value:
157  *     1 if we were able to detach the timer.  0 if we blew it, and the
158  *     timer function has already started to run.
159  **/
160 int scsi_delete_timer(struct scsi_cmnd *scmd)
161 {
162         int rtn;
163
164         rtn = del_timer(&scmd->eh_timeout);
165
166         SCSI_LOG_ERROR_RECOVERY(5, printk("%s: scmd: %p,"
167                                          " rtn: %d\n", __FUNCTION__,
168                                          scmd, rtn));
169
170         scmd->eh_timeout.data = (unsigned long)NULL;
171         scmd->eh_timeout.function = NULL;
172
173         return rtn;
174 }
175
176 /**
177  * scsi_times_out - Timeout function for normal scsi commands.
178  * @scmd:       Cmd that is timing out.
179  *
180  * Notes:
181  *     We do not need to lock this.  There is the potential for a race
182  *     only in that the normal completion handling might run, but if the
183  *     normal completion function determines that the timer has already
184  *     fired, then it mustn't do anything.
185  **/
186 void scsi_times_out(struct scsi_cmnd *scmd)
187 {
188         enum scsi_eh_timer_return (* eh_timed_out)(struct scsi_cmnd *);
189
190         scsi_log_completion(scmd, TIMEOUT_ERROR);
191
192         if (scmd->device->host->transportt->eh_timed_out)
193                 eh_timed_out = scmd->device->host->transportt->eh_timed_out;
194         else if (scmd->device->host->hostt->eh_timed_out)
195                 eh_timed_out = scmd->device->host->hostt->eh_timed_out;
196         else
197                 eh_timed_out = NULL;
198
199         if (eh_timed_out)
200                 switch (eh_timed_out(scmd)) {
201                 case EH_HANDLED:
202                         __scsi_done(scmd);
203                         return;
204                 case EH_RESET_TIMER:
205                         scsi_add_timer(scmd, scmd->timeout_per_command,
206                                        scsi_times_out);
207                         return;
208                 case EH_NOT_HANDLED:
209                         break;
210                 }
211
212         if (unlikely(!scsi_eh_scmd_add(scmd, SCSI_EH_CANCEL_CMD))) {
213                 scmd->result |= DID_TIME_OUT << 16;
214                 __scsi_done(scmd);
215         }
216 }
217
218 /**
219  * scsi_block_when_processing_errors - Prevent cmds from being queued.
220  * @sdev:       Device on which we are performing recovery.
221  *
222  * Description:
223  *     We block until the host is out of error recovery, and then check to
224  *     see whether the host or the device is offline.
225  *
226  * Return value:
227  *     0 when dev was taken offline by error recovery. 1 OK to proceed.
228  **/
229 int scsi_block_when_processing_errors(struct scsi_device *sdev)
230 {
231         int online;
232
233         wait_event(sdev->host->host_wait, !scsi_host_in_recovery(sdev->host));
234
235         online = scsi_device_online(sdev);
236
237         SCSI_LOG_ERROR_RECOVERY(5, printk("%s: rtn: %d\n", __FUNCTION__,
238                                           online));
239
240         return online;
241 }
242 EXPORT_SYMBOL(scsi_block_when_processing_errors);
243
244 #ifdef CONFIG_SCSI_LOGGING
245 /**
246  * scsi_eh_prt_fail_stats - Log info on failures.
247  * @shost:      scsi host being recovered.
248  * @work_q:     Queue of scsi cmds to process.
249  **/
250 static inline void scsi_eh_prt_fail_stats(struct Scsi_Host *shost,
251                                           struct list_head *work_q)
252 {
253         struct scsi_cmnd *scmd;
254         struct scsi_device *sdev;
255         int total_failures = 0;
256         int cmd_failed = 0;
257         int cmd_cancel = 0;
258         int devices_failed = 0;
259
260         shost_for_each_device(sdev, shost) {
261                 list_for_each_entry(scmd, work_q, eh_entry) {
262                         if (scmd->device == sdev) {
263                                 ++total_failures;
264                                 if (scmd->eh_eflags & SCSI_EH_CANCEL_CMD)
265                                         ++cmd_cancel;
266                                 else 
267                                         ++cmd_failed;
268                         }
269                 }
270
271                 if (cmd_cancel || cmd_failed) {
272                         SCSI_LOG_ERROR_RECOVERY(3,
273                                 sdev_printk(KERN_INFO, sdev,
274                                             "%s: cmds failed: %d, cancel: %d\n",
275                                             __FUNCTION__, cmd_failed,
276                                             cmd_cancel));
277                         cmd_cancel = 0;
278                         cmd_failed = 0;
279                         ++devices_failed;
280                 }
281         }
282
283         SCSI_LOG_ERROR_RECOVERY(2, printk("Total of %d commands on %d"
284                                           " devices require eh work\n",
285                                   total_failures, devices_failed));
286 }
287 #endif
288
289 /**
290  * scsi_check_sense - Examine scsi cmd sense
291  * @scmd:       Cmd to have sense checked.
292  *
293  * Return value:
294  *      SUCCESS or FAILED or NEEDS_RETRY
295  *
296  * Notes:
297  *      When a deferred error is detected the current command has
298  *      not been executed and needs retrying.
299  **/
300 static int scsi_check_sense(struct scsi_cmnd *scmd)
301 {
302         struct scsi_sense_hdr sshdr;
303
304         if (! scsi_command_normalize_sense(scmd, &sshdr))
305                 return FAILED;  /* no valid sense data */
306
307         if (scsi_sense_is_deferred(&sshdr))
308                 return NEEDS_RETRY;
309
310         /*
311          * Previous logic looked for FILEMARK, EOM or ILI which are
312          * mainly associated with tapes and returned SUCCESS.
313          */
314         if (sshdr.response_code == 0x70) {
315                 /* fixed format */
316                 if (scmd->sense_buffer[2] & 0xe0)
317                         return SUCCESS;
318         } else {
319                 /*
320                  * descriptor format: look for "stream commands sense data
321                  * descriptor" (see SSC-3). Assume single sense data
322                  * descriptor. Ignore ILI from SBC-2 READ LONG and WRITE LONG.
323                  */
324                 if ((sshdr.additional_length > 3) &&
325                     (scmd->sense_buffer[8] == 0x4) &&
326                     (scmd->sense_buffer[11] & 0xe0))
327                         return SUCCESS;
328         }
329
330         switch (sshdr.sense_key) {
331         case NO_SENSE:
332                 return SUCCESS;
333         case RECOVERED_ERROR:
334                 return /* soft_error */ SUCCESS;
335
336         case ABORTED_COMMAND:
337                 return NEEDS_RETRY;
338         case NOT_READY:
339         case UNIT_ATTENTION:
340                 /*
341                  * if we are expecting a cc/ua because of a bus reset that we
342                  * performed, treat this just as a retry.  otherwise this is
343                  * information that we should pass up to the upper-level driver
344                  * so that we can deal with it there.
345                  */
346                 if (scmd->device->expecting_cc_ua) {
347                         scmd->device->expecting_cc_ua = 0;
348                         return NEEDS_RETRY;
349                 }
350                 /*
351                  * if the device is in the process of becoming ready, we 
352                  * should retry.
353                  */
354                 if ((sshdr.asc == 0x04) && (sshdr.ascq == 0x01))
355                         return NEEDS_RETRY;
356                 /*
357                  * if the device is not started, we need to wake
358                  * the error handler to start the motor
359                  */
360                 if (scmd->device->allow_restart &&
361                     (sshdr.asc == 0x04) && (sshdr.ascq == 0x02))
362                         return FAILED;
363                 return SUCCESS;
364
365                 /* these three are not supported */
366         case COPY_ABORTED:
367         case VOLUME_OVERFLOW:
368         case MISCOMPARE:
369                 return SUCCESS;
370
371         case MEDIUM_ERROR:
372                 if (sshdr.asc == 0x11 || /* UNRECOVERED READ ERR */
373                     sshdr.asc == 0x13 || /* AMNF DATA FIELD */
374                     sshdr.asc == 0x14) { /* RECORD NOT FOUND */
375                         return SUCCESS;
376                 }
377                 return NEEDS_RETRY;
378
379         case HARDWARE_ERROR:
380                 if (scmd->device->retry_hwerror)
381                         return NEEDS_RETRY;
382                 else
383                         return SUCCESS;
384
385         case ILLEGAL_REQUEST:
386         case BLANK_CHECK:
387         case DATA_PROTECT:
388         default:
389                 return SUCCESS;
390         }
391 }
392
393 /**
394  * scsi_eh_completed_normally - Disposition a eh cmd on return from LLD.
395  * @scmd:       SCSI cmd to examine.
396  *
397  * Notes:
398  *    This is *only* called when we are examining the status of commands
399  *    queued during error recovery.  the main difference here is that we
400  *    don't allow for the possibility of retries here, and we are a lot
401  *    more restrictive about what we consider acceptable.
402  **/
403 static int scsi_eh_completed_normally(struct scsi_cmnd *scmd)
404 {
405         /*
406          * first check the host byte, to see if there is anything in there
407          * that would indicate what we need to do.
408          */
409         if (host_byte(scmd->result) == DID_RESET) {
410                 /*
411                  * rats.  we are already in the error handler, so we now
412                  * get to try and figure out what to do next.  if the sense
413                  * is valid, we have a pretty good idea of what to do.
414                  * if not, we mark it as FAILED.
415                  */
416                 return scsi_check_sense(scmd);
417         }
418         if (host_byte(scmd->result) != DID_OK)
419                 return FAILED;
420
421         /*
422          * next, check the message byte.
423          */
424         if (msg_byte(scmd->result) != COMMAND_COMPLETE)
425                 return FAILED;
426
427         /*
428          * now, check the status byte to see if this indicates
429          * anything special.
430          */
431         switch (status_byte(scmd->result)) {
432         case GOOD:
433         case COMMAND_TERMINATED:
434                 return SUCCESS;
435         case CHECK_CONDITION:
436                 return scsi_check_sense(scmd);
437         case CONDITION_GOOD:
438         case INTERMEDIATE_GOOD:
439         case INTERMEDIATE_C_GOOD:
440                 /*
441                  * who knows?  FIXME(eric)
442                  */
443                 return SUCCESS;
444         case BUSY:
445         case QUEUE_FULL:
446         case RESERVATION_CONFLICT:
447         default:
448                 return FAILED;
449         }
450         return FAILED;
451 }
452
453 /**
454  * scsi_eh_done - Completion function for error handling.
455  * @scmd:       Cmd that is done.
456  **/
457 static void scsi_eh_done(struct scsi_cmnd *scmd)
458 {
459         struct completion     *eh_action;
460
461         SCSI_LOG_ERROR_RECOVERY(3,
462                 printk("%s scmd: %p result: %x\n",
463                         __FUNCTION__, scmd, scmd->result));
464
465         eh_action = scmd->device->host->eh_action;
466         if (eh_action)
467                 complete(eh_action);
468 }
469
470 /**
471  * scsi_try_host_reset - ask host adapter to reset itself
472  * @scmd:       SCSI cmd to send hsot reset.
473  **/
474 static int scsi_try_host_reset(struct scsi_cmnd *scmd)
475 {
476         unsigned long flags;
477         int rtn;
478
479         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Snd Host RST\n",
480                                           __FUNCTION__));
481
482         if (!scmd->device->host->hostt->eh_host_reset_handler)
483                 return FAILED;
484
485         rtn = scmd->device->host->hostt->eh_host_reset_handler(scmd);
486
487         if (rtn == SUCCESS) {
488                 if (!scmd->device->host->hostt->skip_settle_delay)
489                         ssleep(HOST_RESET_SETTLE_TIME);
490                 spin_lock_irqsave(scmd->device->host->host_lock, flags);
491                 scsi_report_bus_reset(scmd->device->host,
492                                       scmd_channel(scmd));
493                 spin_unlock_irqrestore(scmd->device->host->host_lock, flags);
494         }
495
496         return rtn;
497 }
498
499 /**
500  * scsi_try_bus_reset - ask host to perform a bus reset
501  * @scmd:       SCSI cmd to send bus reset.
502  **/
503 static int scsi_try_bus_reset(struct scsi_cmnd *scmd)
504 {
505         unsigned long flags;
506         int rtn;
507
508         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Snd Bus RST\n",
509                                           __FUNCTION__));
510
511         if (!scmd->device->host->hostt->eh_bus_reset_handler)
512                 return FAILED;
513
514         rtn = scmd->device->host->hostt->eh_bus_reset_handler(scmd);
515
516         if (rtn == SUCCESS) {
517                 if (!scmd->device->host->hostt->skip_settle_delay)
518                         ssleep(BUS_RESET_SETTLE_TIME);
519                 spin_lock_irqsave(scmd->device->host->host_lock, flags);
520                 scsi_report_bus_reset(scmd->device->host,
521                                       scmd_channel(scmd));
522                 spin_unlock_irqrestore(scmd->device->host->host_lock, flags);
523         }
524
525         return rtn;
526 }
527
528 /**
529  * scsi_try_bus_device_reset - Ask host to perform a BDR on a dev
530  * @scmd:       SCSI cmd used to send BDR
531  *
532  * Notes:
533  *    There is no timeout for this operation.  if this operation is
534  *    unreliable for a given host, then the host itself needs to put a
535  *    timer on it, and set the host back to a consistent state prior to
536  *    returning.
537  **/
538 static int scsi_try_bus_device_reset(struct scsi_cmnd *scmd)
539 {
540         int rtn;
541
542         if (!scmd->device->host->hostt->eh_device_reset_handler)
543                 return FAILED;
544
545         rtn = scmd->device->host->hostt->eh_device_reset_handler(scmd);
546         if (rtn == SUCCESS) {
547                 scmd->device->was_reset = 1;
548                 scmd->device->expecting_cc_ua = 1;
549         }
550
551         return rtn;
552 }
553
554 static int __scsi_try_to_abort_cmd(struct scsi_cmnd *scmd)
555 {
556         if (!scmd->device->host->hostt->eh_abort_handler)
557                 return FAILED;
558
559         return scmd->device->host->hostt->eh_abort_handler(scmd);
560 }
561
562 /**
563  * scsi_try_to_abort_cmd - Ask host to abort a running command.
564  * @scmd:       SCSI cmd to abort from Lower Level.
565  *
566  * Notes:
567  *    This function will not return until the user's completion function
568  *    has been called.  there is no timeout on this operation.  if the
569  *    author of the low-level driver wishes this operation to be timed,
570  *    they can provide this facility themselves.  helper functions in
571  *    scsi_error.c can be supplied to make this easier to do.
572  **/
573 static int scsi_try_to_abort_cmd(struct scsi_cmnd *scmd)
574 {
575         /*
576          * scsi_done was called just after the command timed out and before
577          * we had a chance to process it. (db)
578          */
579         if (scmd->serial_number == 0)
580                 return SUCCESS;
581         return __scsi_try_to_abort_cmd(scmd);
582 }
583
584 static void scsi_abort_eh_cmnd(struct scsi_cmnd *scmd)
585 {
586         if (__scsi_try_to_abort_cmd(scmd) != SUCCESS)
587                 if (scsi_try_bus_device_reset(scmd) != SUCCESS)
588                         if (scsi_try_bus_reset(scmd) != SUCCESS)
589                                 scsi_try_host_reset(scmd);
590 }
591
592 /**
593  * scsi_eh_prep_cmnd  - Save a scsi command info as part of error recory
594  * @scmd:       SCSI command structure to hijack
595  * @ses:        structure to save restore information
596  * @cmnd:       CDB to send. Can be NULL if no new cmnd is needed
597  * @cmnd_size:  size in bytes of @cmnd
598  * @sense_bytes: size of sense data to copy. or 0 (if != 0 @cmnd is ignored)
599  *
600  * This function is used to save a scsi command information before re-execution
601  * as part of the error recovery process.  If @sense_bytes is 0 the command
602  * sent must be one that does not transfer any data.  If @sense_bytes != 0
603  * @cmnd is ignored and this functions sets up a REQUEST_SENSE command
604  * and cmnd buffers to read @sense_bytes into @scmd->sense_buffer.
605  **/
606 void scsi_eh_prep_cmnd(struct scsi_cmnd *scmd, struct scsi_eh_save *ses,
607                         unsigned char *cmnd, int cmnd_size, unsigned sense_bytes)
608 {
609         struct scsi_device *sdev = scmd->device;
610
611         /*
612          * We need saved copies of a number of fields - this is because
613          * error handling may need to overwrite these with different values
614          * to run different commands, and once error handling is complete,
615          * we will need to restore these values prior to running the actual
616          * command.
617          */
618         ses->cmd_len = scmd->cmd_len;
619         memcpy(ses->cmnd, scmd->cmnd, sizeof(scmd->cmnd));
620         ses->data_direction = scmd->sc_data_direction;
621         ses->bufflen = scmd->request_bufflen;
622         ses->buffer = scmd->request_buffer;
623         ses->use_sg = scmd->use_sg;
624         ses->resid = scmd->resid;
625         ses->result = scmd->result;
626
627         if (sense_bytes) {
628                 scmd->request_bufflen = min_t(unsigned,
629                                        sizeof(scmd->sense_buffer), sense_bytes);
630                 sg_init_one(&ses->sense_sgl, scmd->sense_buffer,
631                                                        scmd->request_bufflen);
632                 scmd->request_buffer = &ses->sense_sgl;
633                 scmd->sc_data_direction = DMA_FROM_DEVICE;
634                 scmd->use_sg = 1;
635                 memset(scmd->cmnd, 0, sizeof(scmd->cmnd));
636                 scmd->cmnd[0] = REQUEST_SENSE;
637                 scmd->cmnd[4] = scmd->request_bufflen;
638                 scmd->cmd_len = COMMAND_SIZE(scmd->cmnd[0]);
639         } else {
640                 scmd->request_buffer = NULL;
641                 scmd->request_bufflen = 0;
642                 scmd->sc_data_direction = DMA_NONE;
643                 scmd->use_sg = 0;
644                 if (cmnd) {
645                         memset(scmd->cmnd, 0, sizeof(scmd->cmnd));
646                         memcpy(scmd->cmnd, cmnd, cmnd_size);
647                         scmd->cmd_len = COMMAND_SIZE(scmd->cmnd[0]);
648                 }
649         }
650
651         scmd->underflow = 0;
652
653         if (sdev->scsi_level <= SCSI_2 && sdev->scsi_level != SCSI_UNKNOWN)
654                 scmd->cmnd[1] = (scmd->cmnd[1] & 0x1f) |
655                         (sdev->lun << 5 & 0xe0);
656
657         /*
658          * Zero the sense buffer.  The scsi spec mandates that any
659          * untransferred sense data should be interpreted as being zero.
660          */
661         memset(scmd->sense_buffer, 0, sizeof(scmd->sense_buffer));
662 }
663 EXPORT_SYMBOL(scsi_eh_prep_cmnd);
664
665 /**
666  * scsi_eh_restore_cmnd  - Restore a scsi command info as part of error recory
667  * @scmd:       SCSI command structure to restore
668  * @ses:        saved information from a coresponding call to scsi_prep_eh_cmnd
669  *
670  * Undo any damage done by above scsi_prep_eh_cmnd().
671  **/
672 void scsi_eh_restore_cmnd(struct scsi_cmnd* scmd, struct scsi_eh_save *ses)
673 {
674         /*
675          * Restore original data
676          */
677         scmd->cmd_len = ses->cmd_len;
678         memcpy(scmd->cmnd, ses->cmnd, sizeof(scmd->cmnd));
679         scmd->sc_data_direction = ses->data_direction;
680         scmd->request_bufflen = ses->bufflen;
681         scmd->request_buffer = ses->buffer;
682         scmd->use_sg = ses->use_sg;
683         scmd->resid = ses->resid;
684         scmd->result = ses->result;
685 }
686 EXPORT_SYMBOL(scsi_eh_restore_cmnd);
687
688 /**
689  * scsi_send_eh_cmnd  - submit a scsi command as part of error recory
690  * @scmd:       SCSI command structure to hijack
691  * @cmnd:       CDB to send
692  * @cmnd_size:  size in bytes of @cmnd
693  * @timeout:    timeout for this request
694  * @sense_bytes: size of sense data to copy or 0
695  *
696  * This function is used to send a scsi command down to a target device
697  * as part of the error recovery process. See also scsi_eh_prep_cmnd() above.
698  *
699  * Return value:
700  *    SUCCESS or FAILED or NEEDS_RETRY
701  **/
702 static int scsi_send_eh_cmnd(struct scsi_cmnd *scmd, unsigned char *cmnd,
703                              int cmnd_size, int timeout, unsigned sense_bytes)
704 {
705         struct scsi_device *sdev = scmd->device;
706         struct Scsi_Host *shost = sdev->host;
707         DECLARE_COMPLETION_ONSTACK(done);
708         unsigned long timeleft;
709         unsigned long flags;
710         struct scsi_eh_save ses;
711         int rtn;
712
713         scsi_eh_prep_cmnd(scmd, &ses, cmnd, cmnd_size, sense_bytes);
714         shost->eh_action = &done;
715
716         spin_lock_irqsave(shost->host_lock, flags);
717         scsi_log_send(scmd);
718         shost->hostt->queuecommand(scmd, scsi_eh_done);
719         spin_unlock_irqrestore(shost->host_lock, flags);
720
721         timeleft = wait_for_completion_timeout(&done, timeout);
722
723         shost->eh_action = NULL;
724
725         scsi_log_completion(scmd, SUCCESS);
726
727         SCSI_LOG_ERROR_RECOVERY(3,
728                 printk("%s: scmd: %p, timeleft: %ld\n",
729                         __FUNCTION__, scmd, timeleft));
730
731         /*
732          * If there is time left scsi_eh_done got called, and we will
733          * examine the actual status codes to see whether the command
734          * actually did complete normally, else tell the host to forget
735          * about this command.
736          */
737         if (timeleft) {
738                 rtn = scsi_eh_completed_normally(scmd);
739                 SCSI_LOG_ERROR_RECOVERY(3,
740                         printk("%s: scsi_eh_completed_normally %x\n",
741                                __FUNCTION__, rtn));
742
743                 switch (rtn) {
744                 case SUCCESS:
745                 case NEEDS_RETRY:
746                 case FAILED:
747                         break;
748                 default:
749                         rtn = FAILED;
750                         break;
751                 }
752         } else {
753                 scsi_abort_eh_cmnd(scmd);
754                 rtn = FAILED;
755         }
756
757         scsi_eh_restore_cmnd(scmd, &ses);
758         return rtn;
759 }
760
761 /**
762  * scsi_request_sense - Request sense data from a particular target.
763  * @scmd:       SCSI cmd for request sense.
764  *
765  * Notes:
766  *    Some hosts automatically obtain this information, others require
767  *    that we obtain it on our own. This function will *not* return until
768  *    the command either times out, or it completes.
769  **/
770 static int scsi_request_sense(struct scsi_cmnd *scmd)
771 {
772         return scsi_send_eh_cmnd(scmd, NULL, 0, SENSE_TIMEOUT, ~0);
773 }
774
775 /**
776  * scsi_eh_finish_cmd - Handle a cmd that eh is finished with.
777  * @scmd:       Original SCSI cmd that eh has finished.
778  * @done_q:     Queue for processed commands.
779  *
780  * Notes:
781  *    We don't want to use the normal command completion while we are are
782  *    still handling errors - it may cause other commands to be queued,
783  *    and that would disturb what we are doing.  thus we really want to
784  *    keep a list of pending commands for final completion, and once we
785  *    are ready to leave error handling we handle completion for real.
786  **/
787 void scsi_eh_finish_cmd(struct scsi_cmnd *scmd, struct list_head *done_q)
788 {
789         scmd->device->host->host_failed--;
790         scmd->eh_eflags = 0;
791         list_move_tail(&scmd->eh_entry, done_q);
792 }
793 EXPORT_SYMBOL(scsi_eh_finish_cmd);
794
795 /**
796  * scsi_eh_get_sense - Get device sense data.
797  * @work_q:     Queue of commands to process.
798  * @done_q:     Queue of proccessed commands..
799  *
800  * Description:
801  *    See if we need to request sense information.  if so, then get it
802  *    now, so we have a better idea of what to do.  
803  *
804  * Notes:
805  *    This has the unfortunate side effect that if a shost adapter does
806  *    not automatically request sense information, that we end up shutting
807  *    it down before we request it.
808  *
809  *    All drivers should request sense information internally these days,
810  *    so for now all I have to say is tough noogies if you end up in here.
811  *
812  *    XXX: Long term this code should go away, but that needs an audit of
813  *         all LLDDs first.
814  **/
815 int scsi_eh_get_sense(struct list_head *work_q,
816                       struct list_head *done_q)
817 {
818         struct scsi_cmnd *scmd, *next;
819         int rtn;
820
821         list_for_each_entry_safe(scmd, next, work_q, eh_entry) {
822                 if ((scmd->eh_eflags & SCSI_EH_CANCEL_CMD) ||
823                     SCSI_SENSE_VALID(scmd))
824                         continue;
825
826                 SCSI_LOG_ERROR_RECOVERY(2, scmd_printk(KERN_INFO, scmd,
827                                                   "%s: requesting sense\n",
828                                                   current->comm));
829                 rtn = scsi_request_sense(scmd);
830                 if (rtn != SUCCESS)
831                         continue;
832
833                 SCSI_LOG_ERROR_RECOVERY(3, printk("sense requested for %p"
834                                                   " result %x\n", scmd,
835                                                   scmd->result));
836                 SCSI_LOG_ERROR_RECOVERY(3, scsi_print_sense("bh", scmd));
837
838                 rtn = scsi_decide_disposition(scmd);
839
840                 /*
841                  * if the result was normal, then just pass it along to the
842                  * upper level.
843                  */
844                 if (rtn == SUCCESS)
845                         /* we don't want this command reissued, just
846                          * finished with the sense data, so set
847                          * retries to the max allowed to ensure it
848                          * won't get reissued */
849                         scmd->retries = scmd->allowed;
850                 else if (rtn != NEEDS_RETRY)
851                         continue;
852
853                 scsi_eh_finish_cmd(scmd, done_q);
854         }
855
856         return list_empty(work_q);
857 }
858 EXPORT_SYMBOL_GPL(scsi_eh_get_sense);
859
860 /**
861  * scsi_eh_tur - Send TUR to device.
862  * @scmd:       Scsi cmd to send TUR
863  *
864  * Return value:
865  *    0 - Device is ready. 1 - Device NOT ready.
866  **/
867 static int scsi_eh_tur(struct scsi_cmnd *scmd)
868 {
869         static unsigned char tur_command[6] = {TEST_UNIT_READY, 0, 0, 0, 0, 0};
870         int retry_cnt = 1, rtn;
871
872 retry_tur:
873         rtn = scsi_send_eh_cmnd(scmd, tur_command, 6, SENSE_TIMEOUT, 0);
874
875         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: scmd %p rtn %x\n",
876                 __FUNCTION__, scmd, rtn));
877
878         switch (rtn) {
879         case NEEDS_RETRY:
880                 if (retry_cnt--)
881                         goto retry_tur;
882                 /*FALLTHRU*/
883         case SUCCESS:
884                 return 0;
885         default:
886                 return 1;
887         }
888 }
889
890 /**
891  * scsi_eh_abort_cmds - abort canceled commands.
892  * @shost:      scsi host being recovered.
893  * @eh_done_q:  list_head for processed commands.
894  *
895  * Decription:
896  *    Try and see whether or not it makes sense to try and abort the
897  *    running command.  this only works out to be the case if we have one
898  *    command that has timed out.  if the command simply failed, it makes
899  *    no sense to try and abort the command, since as far as the shost
900  *    adapter is concerned, it isn't running.
901  **/
902 static int scsi_eh_abort_cmds(struct list_head *work_q,
903                               struct list_head *done_q)
904 {
905         struct scsi_cmnd *scmd, *next;
906         int rtn;
907
908         list_for_each_entry_safe(scmd, next, work_q, eh_entry) {
909                 if (!(scmd->eh_eflags & SCSI_EH_CANCEL_CMD))
910                         continue;
911                 SCSI_LOG_ERROR_RECOVERY(3, printk("%s: aborting cmd:"
912                                                   "0x%p\n", current->comm,
913                                                   scmd));
914                 rtn = scsi_try_to_abort_cmd(scmd);
915                 if (rtn == SUCCESS) {
916                         scmd->eh_eflags &= ~SCSI_EH_CANCEL_CMD;
917                         if (!scsi_device_online(scmd->device) ||
918                             !scsi_eh_tur(scmd)) {
919                                 scsi_eh_finish_cmd(scmd, done_q);
920                         }
921                                 
922                 } else
923                         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: aborting"
924                                                           " cmd failed:"
925                                                           "0x%p\n",
926                                                           current->comm,
927                                                           scmd));
928         }
929
930         return list_empty(work_q);
931 }
932
933 /**
934  * scsi_eh_try_stu - Send START_UNIT to device.
935  * @scmd:       Scsi cmd to send START_UNIT
936  *
937  * Return value:
938  *    0 - Device is ready. 1 - Device NOT ready.
939  **/
940 static int scsi_eh_try_stu(struct scsi_cmnd *scmd)
941 {
942         static unsigned char stu_command[6] = {START_STOP, 0, 0, 0, 1, 0};
943
944         if (scmd->device->allow_restart) {
945                 int i, rtn = NEEDS_RETRY;
946
947                 for (i = 0; rtn == NEEDS_RETRY && i < 2; i++)
948                         rtn = scsi_send_eh_cmnd(scmd, stu_command, 6,
949                                                 scmd->device->timeout, 0);
950
951                 if (rtn == SUCCESS)
952                         return 0;
953         }
954
955         return 1;
956 }
957
958  /**
959  * scsi_eh_stu - send START_UNIT if needed
960  * @shost:      scsi host being recovered.
961  * @eh_done_q:  list_head for processed commands.
962  *
963  * Notes:
964  *    If commands are failing due to not ready, initializing command required,
965  *      try revalidating the device, which will end up sending a start unit. 
966  **/
967 static int scsi_eh_stu(struct Scsi_Host *shost,
968                               struct list_head *work_q,
969                               struct list_head *done_q)
970 {
971         struct scsi_cmnd *scmd, *stu_scmd, *next;
972         struct scsi_device *sdev;
973
974         shost_for_each_device(sdev, shost) {
975                 stu_scmd = NULL;
976                 list_for_each_entry(scmd, work_q, eh_entry)
977                         if (scmd->device == sdev && SCSI_SENSE_VALID(scmd) &&
978                             scsi_check_sense(scmd) == FAILED ) {
979                                 stu_scmd = scmd;
980                                 break;
981                         }
982
983                 if (!stu_scmd)
984                         continue;
985
986                 SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Sending START_UNIT to sdev:"
987                                                   " 0x%p\n", current->comm, sdev));
988
989                 if (!scsi_eh_try_stu(stu_scmd)) {
990                         if (!scsi_device_online(sdev) ||
991                             !scsi_eh_tur(stu_scmd)) {
992                                 list_for_each_entry_safe(scmd, next,
993                                                           work_q, eh_entry) {
994                                         if (scmd->device == sdev)
995                                                 scsi_eh_finish_cmd(scmd, done_q);
996                                 }
997                         }
998                 } else {
999                         SCSI_LOG_ERROR_RECOVERY(3,
1000                                                 printk("%s: START_UNIT failed to sdev:"
1001                                                        " 0x%p\n", current->comm, sdev));
1002                 }
1003         }
1004
1005         return list_empty(work_q);
1006 }
1007
1008
1009 /**
1010  * scsi_eh_bus_device_reset - send bdr if needed
1011  * @shost:      scsi host being recovered.
1012  * @eh_done_q:  list_head for processed commands.
1013  *
1014  * Notes:
1015  *    Try a bus device reset.  still, look to see whether we have multiple
1016  *    devices that are jammed or not - if we have multiple devices, it
1017  *    makes no sense to try bus_device_reset - we really would need to try
1018  *    a bus_reset instead. 
1019  **/
1020 static int scsi_eh_bus_device_reset(struct Scsi_Host *shost,
1021                                     struct list_head *work_q,
1022                                     struct list_head *done_q)
1023 {
1024         struct scsi_cmnd *scmd, *bdr_scmd, *next;
1025         struct scsi_device *sdev;
1026         int rtn;
1027
1028         shost_for_each_device(sdev, shost) {
1029                 bdr_scmd = NULL;
1030                 list_for_each_entry(scmd, work_q, eh_entry)
1031                         if (scmd->device == sdev) {
1032                                 bdr_scmd = scmd;
1033                                 break;
1034                         }
1035
1036                 if (!bdr_scmd)
1037                         continue;
1038
1039                 SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Sending BDR sdev:"
1040                                                   " 0x%p\n", current->comm,
1041                                                   sdev));
1042                 rtn = scsi_try_bus_device_reset(bdr_scmd);
1043                 if (rtn == SUCCESS) {
1044                         if (!scsi_device_online(sdev) ||
1045                             !scsi_eh_tur(bdr_scmd)) {
1046                                 list_for_each_entry_safe(scmd, next,
1047                                                          work_q, eh_entry) {
1048                                         if (scmd->device == sdev)
1049                                                 scsi_eh_finish_cmd(scmd,
1050                                                                    done_q);
1051                                 }
1052                         }
1053                 } else {
1054                         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: BDR"
1055                                                           " failed sdev:"
1056                                                           "0x%p\n",
1057                                                           current->comm,
1058                                                            sdev));
1059                 }
1060         }
1061
1062         return list_empty(work_q);
1063 }
1064
1065 /**
1066  * scsi_eh_bus_reset - send a bus reset 
1067  * @shost:      scsi host being recovered.
1068  * @eh_done_q:  list_head for processed commands.
1069  **/
1070 static int scsi_eh_bus_reset(struct Scsi_Host *shost,
1071                              struct list_head *work_q,
1072                              struct list_head *done_q)
1073 {
1074         struct scsi_cmnd *scmd, *chan_scmd, *next;
1075         unsigned int channel;
1076         int rtn;
1077
1078         /*
1079          * we really want to loop over the various channels, and do this on
1080          * a channel by channel basis.  we should also check to see if any
1081          * of the failed commands are on soft_reset devices, and if so, skip
1082          * the reset.  
1083          */
1084
1085         for (channel = 0; channel <= shost->max_channel; channel++) {
1086                 chan_scmd = NULL;
1087                 list_for_each_entry(scmd, work_q, eh_entry) {
1088                         if (channel == scmd_channel(scmd)) {
1089                                 chan_scmd = scmd;
1090                                 break;
1091                                 /*
1092                                  * FIXME add back in some support for
1093                                  * soft_reset devices.
1094                                  */
1095                         }
1096                 }
1097
1098                 if (!chan_scmd)
1099                         continue;
1100                 SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Sending BRST chan:"
1101                                                   " %d\n", current->comm,
1102                                                   channel));
1103                 rtn = scsi_try_bus_reset(chan_scmd);
1104                 if (rtn == SUCCESS) {
1105                         list_for_each_entry_safe(scmd, next, work_q, eh_entry) {
1106                                 if (channel == scmd_channel(scmd))
1107                                         if (!scsi_device_online(scmd->device) ||
1108                                             !scsi_eh_tur(scmd))
1109                                                 scsi_eh_finish_cmd(scmd,
1110                                                                    done_q);
1111                         }
1112                 } else {
1113                         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: BRST"
1114                                                           " failed chan: %d\n",
1115                                                           current->comm,
1116                                                           channel));
1117                 }
1118         }
1119         return list_empty(work_q);
1120 }
1121
1122 /**
1123  * scsi_eh_host_reset - send a host reset 
1124  * @work_q:     list_head for processed commands.
1125  * @done_q:     list_head for processed commands.
1126  **/
1127 static int scsi_eh_host_reset(struct list_head *work_q,
1128                               struct list_head *done_q)
1129 {
1130         struct scsi_cmnd *scmd, *next;
1131         int rtn;
1132
1133         if (!list_empty(work_q)) {
1134                 scmd = list_entry(work_q->next,
1135                                   struct scsi_cmnd, eh_entry);
1136
1137                 SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Sending HRST\n"
1138                                                   , current->comm));
1139
1140                 rtn = scsi_try_host_reset(scmd);
1141                 if (rtn == SUCCESS) {
1142                         list_for_each_entry_safe(scmd, next, work_q, eh_entry) {
1143                                 if (!scsi_device_online(scmd->device) ||
1144                                     (!scsi_eh_try_stu(scmd) && !scsi_eh_tur(scmd)) ||
1145                                     !scsi_eh_tur(scmd))
1146                                         scsi_eh_finish_cmd(scmd, done_q);
1147                         }
1148                 } else {
1149                         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: HRST"
1150                                                           " failed\n",
1151                                                           current->comm));
1152                 }
1153         }
1154         return list_empty(work_q);
1155 }
1156
1157 /**
1158  * scsi_eh_offline_sdevs - offline scsi devices that fail to recover
1159  * @work_q:     list_head for processed commands.
1160  * @done_q:     list_head for processed commands.
1161  *
1162  **/
1163 static void scsi_eh_offline_sdevs(struct list_head *work_q,
1164                                   struct list_head *done_q)
1165 {
1166         struct scsi_cmnd *scmd, *next;
1167
1168         list_for_each_entry_safe(scmd, next, work_q, eh_entry) {
1169                 sdev_printk(KERN_INFO, scmd->device, "Device offlined - "
1170                             "not ready after error recovery\n");
1171                 scsi_device_set_state(scmd->device, SDEV_OFFLINE);
1172                 if (scmd->eh_eflags & SCSI_EH_CANCEL_CMD) {
1173                         /*
1174                          * FIXME: Handle lost cmds.
1175                          */
1176                 }
1177                 scsi_eh_finish_cmd(scmd, done_q);
1178         }
1179         return;
1180 }
1181
1182 /**
1183  * scsi_decide_disposition - Disposition a cmd on return from LLD.
1184  * @scmd:       SCSI cmd to examine.
1185  *
1186  * Notes:
1187  *    This is *only* called when we are examining the status after sending
1188  *    out the actual data command.  any commands that are queued for error
1189  *    recovery (e.g. test_unit_ready) do *not* come through here.
1190  *
1191  *    When this routine returns failed, it means the error handler thread
1192  *    is woken.  In cases where the error code indicates an error that
1193  *    doesn't require the error handler read (i.e. we don't need to
1194  *    abort/reset), this function should return SUCCESS.
1195  **/
1196 int scsi_decide_disposition(struct scsi_cmnd *scmd)
1197 {
1198         int rtn;
1199
1200         /*
1201          * if the device is offline, then we clearly just pass the result back
1202          * up to the top level.
1203          */
1204         if (!scsi_device_online(scmd->device)) {
1205                 SCSI_LOG_ERROR_RECOVERY(5, printk("%s: device offline - report"
1206                                                   " as SUCCESS\n",
1207                                                   __FUNCTION__));
1208                 return SUCCESS;
1209         }
1210
1211         /*
1212          * first check the host byte, to see if there is anything in there
1213          * that would indicate what we need to do.
1214          */
1215         switch (host_byte(scmd->result)) {
1216         case DID_PASSTHROUGH:
1217                 /*
1218                  * no matter what, pass this through to the upper layer.
1219                  * nuke this special code so that it looks like we are saying
1220                  * did_ok.
1221                  */
1222                 scmd->result &= 0xff00ffff;
1223                 return SUCCESS;
1224         case DID_OK:
1225                 /*
1226                  * looks good.  drop through, and check the next byte.
1227                  */
1228                 break;
1229         case DID_NO_CONNECT:
1230         case DID_BAD_TARGET:
1231         case DID_ABORT:
1232                 /*
1233                  * note - this means that we just report the status back
1234                  * to the top level driver, not that we actually think
1235                  * that it indicates SUCCESS.
1236                  */
1237                 return SUCCESS;
1238                 /*
1239                  * when the low level driver returns did_soft_error,
1240                  * it is responsible for keeping an internal retry counter 
1241                  * in order to avoid endless loops (db)
1242                  *
1243                  * actually this is a bug in this function here.  we should
1244                  * be mindful of the maximum number of retries specified
1245                  * and not get stuck in a loop.
1246                  */
1247         case DID_SOFT_ERROR:
1248                 goto maybe_retry;
1249         case DID_IMM_RETRY:
1250                 return NEEDS_RETRY;
1251
1252         case DID_REQUEUE:
1253                 return ADD_TO_MLQUEUE;
1254
1255         case DID_ERROR:
1256                 if (msg_byte(scmd->result) == COMMAND_COMPLETE &&
1257                     status_byte(scmd->result) == RESERVATION_CONFLICT)
1258                         /*
1259                          * execute reservation conflict processing code
1260                          * lower down
1261                          */
1262                         break;
1263                 /* fallthrough */
1264
1265         case DID_BUS_BUSY:
1266         case DID_PARITY:
1267                 goto maybe_retry;
1268         case DID_TIME_OUT:
1269                 /*
1270                  * when we scan the bus, we get timeout messages for
1271                  * these commands if there is no device available.
1272                  * other hosts report did_no_connect for the same thing.
1273                  */
1274                 if ((scmd->cmnd[0] == TEST_UNIT_READY ||
1275                      scmd->cmnd[0] == INQUIRY)) {
1276                         return SUCCESS;
1277                 } else {
1278                         return FAILED;
1279                 }
1280         case DID_RESET:
1281                 return SUCCESS;
1282         default:
1283                 return FAILED;
1284         }
1285
1286         /*
1287          * next, check the message byte.
1288          */
1289         if (msg_byte(scmd->result) != COMMAND_COMPLETE)
1290                 return FAILED;
1291
1292         /*
1293          * check the status byte to see if this indicates anything special.
1294          */
1295         switch (status_byte(scmd->result)) {
1296         case QUEUE_FULL:
1297                 /*
1298                  * the case of trying to send too many commands to a
1299                  * tagged queueing device.
1300                  */
1301         case BUSY:
1302                 /*
1303                  * device can't talk to us at the moment.  Should only
1304                  * occur (SAM-3) when the task queue is empty, so will cause
1305                  * the empty queue handling to trigger a stall in the
1306                  * device.
1307                  */
1308                 return ADD_TO_MLQUEUE;
1309         case GOOD:
1310         case COMMAND_TERMINATED:
1311         case TASK_ABORTED:
1312                 return SUCCESS;
1313         case CHECK_CONDITION:
1314                 rtn = scsi_check_sense(scmd);
1315                 if (rtn == NEEDS_RETRY)
1316                         goto maybe_retry;
1317                 /* if rtn == FAILED, we have no sense information;
1318                  * returning FAILED will wake the error handler thread
1319                  * to collect the sense and redo the decide
1320                  * disposition */
1321                 return rtn;
1322         case CONDITION_GOOD:
1323         case INTERMEDIATE_GOOD:
1324         case INTERMEDIATE_C_GOOD:
1325         case ACA_ACTIVE:
1326                 /*
1327                  * who knows?  FIXME(eric)
1328                  */
1329                 return SUCCESS;
1330
1331         case RESERVATION_CONFLICT:
1332                 sdev_printk(KERN_INFO, scmd->device,
1333                             "reservation conflict\n");
1334                 return SUCCESS; /* causes immediate i/o error */
1335         default:
1336                 return FAILED;
1337         }
1338         return FAILED;
1339
1340       maybe_retry:
1341
1342         /* we requeue for retry because the error was retryable, and
1343          * the request was not marked fast fail.  Note that above,
1344          * even if the request is marked fast fail, we still requeue
1345          * for queue congestion conditions (QUEUE_FULL or BUSY) */
1346         if ((++scmd->retries) <= scmd->allowed
1347             && !blk_noretry_request(scmd->request)) {
1348                 return NEEDS_RETRY;
1349         } else {
1350                 /*
1351                  * no more retries - report this one back to upper level.
1352                  */
1353                 return SUCCESS;
1354         }
1355 }
1356
1357 /**
1358  * scsi_eh_lock_door - Prevent medium removal for the specified device
1359  * @sdev:       SCSI device to prevent medium removal
1360  *
1361  * Locking:
1362  *      We must be called from process context; scsi_allocate_request()
1363  *      may sleep.
1364  *
1365  * Notes:
1366  *      We queue up an asynchronous "ALLOW MEDIUM REMOVAL" request on the
1367  *      head of the devices request queue, and continue.
1368  *
1369  * Bugs:
1370  *      scsi_allocate_request() may sleep waiting for existing requests to
1371  *      be processed.  However, since we haven't kicked off any request
1372  *      processing for this host, this may deadlock.
1373  *
1374  *      If scsi_allocate_request() fails for what ever reason, we
1375  *      completely forget to lock the door.
1376  **/
1377 static void scsi_eh_lock_door(struct scsi_device *sdev)
1378 {
1379         unsigned char cmnd[MAX_COMMAND_SIZE];
1380
1381         cmnd[0] = ALLOW_MEDIUM_REMOVAL;
1382         cmnd[1] = 0;
1383         cmnd[2] = 0;
1384         cmnd[3] = 0;
1385         cmnd[4] = SCSI_REMOVAL_PREVENT;
1386         cmnd[5] = 0;
1387
1388         scsi_execute_async(sdev, cmnd, 6, DMA_NONE, NULL, 0, 0, 10 * HZ,
1389                            5, NULL, NULL, GFP_KERNEL);
1390 }
1391
1392
1393 /**
1394  * scsi_restart_operations - restart io operations to the specified host.
1395  * @shost:      Host we are restarting.
1396  *
1397  * Notes:
1398  *    When we entered the error handler, we blocked all further i/o to
1399  *    this device.  we need to 'reverse' this process.
1400  **/
1401 static void scsi_restart_operations(struct Scsi_Host *shost)
1402 {
1403         struct scsi_device *sdev;
1404         unsigned long flags;
1405
1406         /*
1407          * If the door was locked, we need to insert a door lock request
1408          * onto the head of the SCSI request queue for the device.  There
1409          * is no point trying to lock the door of an off-line device.
1410          */
1411         shost_for_each_device(sdev, shost) {
1412                 if (scsi_device_online(sdev) && sdev->locked)
1413                         scsi_eh_lock_door(sdev);
1414         }
1415
1416         /*
1417          * next free up anything directly waiting upon the host.  this
1418          * will be requests for character device operations, and also for
1419          * ioctls to queued block devices.
1420          */
1421         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: waking up host to restart\n",
1422                                           __FUNCTION__));
1423
1424         spin_lock_irqsave(shost->host_lock, flags);
1425         if (scsi_host_set_state(shost, SHOST_RUNNING))
1426                 if (scsi_host_set_state(shost, SHOST_CANCEL))
1427                         BUG_ON(scsi_host_set_state(shost, SHOST_DEL));
1428         spin_unlock_irqrestore(shost->host_lock, flags);
1429
1430         wake_up(&shost->host_wait);
1431
1432         /*
1433          * finally we need to re-initiate requests that may be pending.  we will
1434          * have had everything blocked while error handling is taking place, and
1435          * now that error recovery is done, we will need to ensure that these
1436          * requests are started.
1437          */
1438         scsi_run_host_queues(shost);
1439 }
1440
1441 /**
1442  * scsi_eh_ready_devs - check device ready state and recover if not.
1443  * @shost:      host to be recovered.
1444  * @eh_done_q:  list_head for processed commands.
1445  *
1446  **/
1447 void scsi_eh_ready_devs(struct Scsi_Host *shost,
1448                         struct list_head *work_q,
1449                         struct list_head *done_q)
1450 {
1451         if (!scsi_eh_stu(shost, work_q, done_q))
1452                 if (!scsi_eh_bus_device_reset(shost, work_q, done_q))
1453                         if (!scsi_eh_bus_reset(shost, work_q, done_q))
1454                                 if (!scsi_eh_host_reset(work_q, done_q))
1455                                         scsi_eh_offline_sdevs(work_q, done_q);
1456 }
1457 EXPORT_SYMBOL_GPL(scsi_eh_ready_devs);
1458
1459 /**
1460  * scsi_eh_flush_done_q - finish processed commands or retry them.
1461  * @done_q:     list_head of processed commands.
1462  *
1463  **/
1464 void scsi_eh_flush_done_q(struct list_head *done_q)
1465 {
1466         struct scsi_cmnd *scmd, *next;
1467
1468         list_for_each_entry_safe(scmd, next, done_q, eh_entry) {
1469                 list_del_init(&scmd->eh_entry);
1470                 if (scsi_device_online(scmd->device) &&
1471                     !blk_noretry_request(scmd->request) &&
1472                     (++scmd->retries <= scmd->allowed)) {
1473                         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: flush"
1474                                                           " retry cmd: %p\n",
1475                                                           current->comm,
1476                                                           scmd));
1477                                 scsi_queue_insert(scmd, SCSI_MLQUEUE_EH_RETRY);
1478                 } else {
1479                         /*
1480                          * If just we got sense for the device (called
1481                          * scsi_eh_get_sense), scmd->result is already
1482                          * set, do not set DRIVER_TIMEOUT.
1483                          */
1484                         if (!scmd->result)
1485                                 scmd->result |= (DRIVER_TIMEOUT << 24);
1486                         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: flush finish"
1487                                                         " cmd: %p\n",
1488                                                         current->comm, scmd));
1489                         scsi_finish_command(scmd);
1490                 }
1491         }
1492 }
1493 EXPORT_SYMBOL(scsi_eh_flush_done_q);
1494
1495 /**
1496  * scsi_unjam_host - Attempt to fix a host which has a cmd that failed.
1497  * @shost:      Host to unjam.
1498  *
1499  * Notes:
1500  *    When we come in here, we *know* that all commands on the bus have
1501  *    either completed, failed or timed out.  we also know that no further
1502  *    commands are being sent to the host, so things are relatively quiet
1503  *    and we have freedom to fiddle with things as we wish.
1504  *
1505  *    This is only the *default* implementation.  it is possible for
1506  *    individual drivers to supply their own version of this function, and
1507  *    if the maintainer wishes to do this, it is strongly suggested that
1508  *    this function be taken as a template and modified.  this function
1509  *    was designed to correctly handle problems for about 95% of the
1510  *    different cases out there, and it should always provide at least a
1511  *    reasonable amount of error recovery.
1512  *
1513  *    Any command marked 'failed' or 'timeout' must eventually have
1514  *    scsi_finish_cmd() called for it.  we do all of the retry stuff
1515  *    here, so when we restart the host after we return it should have an
1516  *    empty queue.
1517  **/
1518 static void scsi_unjam_host(struct Scsi_Host *shost)
1519 {
1520         unsigned long flags;
1521         LIST_HEAD(eh_work_q);
1522         LIST_HEAD(eh_done_q);
1523
1524         spin_lock_irqsave(shost->host_lock, flags);
1525         list_splice_init(&shost->eh_cmd_q, &eh_work_q);
1526         spin_unlock_irqrestore(shost->host_lock, flags);
1527
1528         SCSI_LOG_ERROR_RECOVERY(1, scsi_eh_prt_fail_stats(shost, &eh_work_q));
1529
1530         if (!scsi_eh_get_sense(&eh_work_q, &eh_done_q))
1531                 if (!scsi_eh_abort_cmds(&eh_work_q, &eh_done_q))
1532                         scsi_eh_ready_devs(shost, &eh_work_q, &eh_done_q);
1533
1534         scsi_eh_flush_done_q(&eh_done_q);
1535 }
1536
1537 /**
1538  * scsi_error_handler - SCSI error handler thread
1539  * @data:       Host for which we are running.
1540  *
1541  * Notes:
1542  *    This is the main error handling loop.  This is run as a kernel thread
1543  *    for every SCSI host and handles all error handling activity.
1544  **/
1545 int scsi_error_handler(void *data)
1546 {
1547         struct Scsi_Host *shost = data;
1548
1549         /*
1550          * We use TASK_INTERRUPTIBLE so that the thread is not
1551          * counted against the load average as a running process.
1552          * We never actually get interrupted because kthread_run
1553          * disables singal delivery for the created thread.
1554          */
1555         set_current_state(TASK_INTERRUPTIBLE);
1556         while (!kthread_should_stop()) {
1557                 if ((shost->host_failed == 0 && shost->host_eh_scheduled == 0) ||
1558                     shost->host_failed != shost->host_busy) {
1559                         SCSI_LOG_ERROR_RECOVERY(1,
1560                                 printk("Error handler scsi_eh_%d sleeping\n",
1561                                         shost->host_no));
1562                         schedule();
1563                         set_current_state(TASK_INTERRUPTIBLE);
1564                         continue;
1565                 }
1566
1567                 __set_current_state(TASK_RUNNING);
1568                 SCSI_LOG_ERROR_RECOVERY(1,
1569                         printk("Error handler scsi_eh_%d waking up\n",
1570                                 shost->host_no));
1571
1572                 /*
1573                  * We have a host that is failing for some reason.  Figure out
1574                  * what we need to do to get it up and online again (if we can).
1575                  * If we fail, we end up taking the thing offline.
1576                  */
1577                 if (shost->transportt->eh_strategy_handler)
1578                         shost->transportt->eh_strategy_handler(shost);
1579                 else
1580                         scsi_unjam_host(shost);
1581
1582                 /*
1583                  * Note - if the above fails completely, the action is to take
1584                  * individual devices offline and flush the queue of any
1585                  * outstanding requests that may have been pending.  When we
1586                  * restart, we restart any I/O to any other devices on the bus
1587                  * which are still online.
1588                  */
1589                 scsi_restart_operations(shost);
1590                 set_current_state(TASK_INTERRUPTIBLE);
1591         }
1592         __set_current_state(TASK_RUNNING);
1593
1594         SCSI_LOG_ERROR_RECOVERY(1,
1595                 printk("Error handler scsi_eh_%d exiting\n", shost->host_no));
1596         shost->ehandler = NULL;
1597         return 0;
1598 }
1599
1600 /*
1601  * Function:    scsi_report_bus_reset()
1602  *
1603  * Purpose:     Utility function used by low-level drivers to report that
1604  *              they have observed a bus reset on the bus being handled.
1605  *
1606  * Arguments:   shost       - Host in question
1607  *              channel     - channel on which reset was observed.
1608  *
1609  * Returns:     Nothing
1610  *
1611  * Lock status: Host lock must be held.
1612  *
1613  * Notes:       This only needs to be called if the reset is one which
1614  *              originates from an unknown location.  Resets originated
1615  *              by the mid-level itself don't need to call this, but there
1616  *              should be no harm.
1617  *
1618  *              The main purpose of this is to make sure that a CHECK_CONDITION
1619  *              is properly treated.
1620  */
1621 void scsi_report_bus_reset(struct Scsi_Host *shost, int channel)
1622 {
1623         struct scsi_device *sdev;
1624
1625         __shost_for_each_device(sdev, shost) {
1626                 if (channel == sdev_channel(sdev)) {
1627                         sdev->was_reset = 1;
1628                         sdev->expecting_cc_ua = 1;
1629                 }
1630         }
1631 }
1632 EXPORT_SYMBOL(scsi_report_bus_reset);
1633
1634 /*
1635  * Function:    scsi_report_device_reset()
1636  *
1637  * Purpose:     Utility function used by low-level drivers to report that
1638  *              they have observed a device reset on the device being handled.
1639  *
1640  * Arguments:   shost       - Host in question
1641  *              channel     - channel on which reset was observed
1642  *              target      - target on which reset was observed
1643  *
1644  * Returns:     Nothing
1645  *
1646  * Lock status: Host lock must be held
1647  *
1648  * Notes:       This only needs to be called if the reset is one which
1649  *              originates from an unknown location.  Resets originated
1650  *              by the mid-level itself don't need to call this, but there
1651  *              should be no harm.
1652  *
1653  *              The main purpose of this is to make sure that a CHECK_CONDITION
1654  *              is properly treated.
1655  */
1656 void scsi_report_device_reset(struct Scsi_Host *shost, int channel, int target)
1657 {
1658         struct scsi_device *sdev;
1659
1660         __shost_for_each_device(sdev, shost) {
1661                 if (channel == sdev_channel(sdev) &&
1662                     target == sdev_id(sdev)) {
1663                         sdev->was_reset = 1;
1664                         sdev->expecting_cc_ua = 1;
1665                 }
1666         }
1667 }
1668 EXPORT_SYMBOL(scsi_report_device_reset);
1669
1670 static void
1671 scsi_reset_provider_done_command(struct scsi_cmnd *scmd)
1672 {
1673 }
1674
1675 /*
1676  * Function:    scsi_reset_provider
1677  *
1678  * Purpose:     Send requested reset to a bus or device at any phase.
1679  *
1680  * Arguments:   device  - device to send reset to
1681  *              flag - reset type (see scsi.h)
1682  *
1683  * Returns:     SUCCESS/FAILURE.
1684  *
1685  * Notes:       This is used by the SCSI Generic driver to provide
1686  *              Bus/Device reset capability.
1687  */
1688 int
1689 scsi_reset_provider(struct scsi_device *dev, int flag)
1690 {
1691         struct scsi_cmnd *scmd = scsi_get_command(dev, GFP_KERNEL);
1692         struct Scsi_Host *shost = dev->host;
1693         struct request req;
1694         unsigned long flags;
1695         int rtn;
1696
1697         scmd->request = &req;
1698         memset(&scmd->eh_timeout, 0, sizeof(scmd->eh_timeout));
1699
1700         memset(&scmd->cmnd, '\0', sizeof(scmd->cmnd));
1701     
1702         scmd->scsi_done         = scsi_reset_provider_done_command;
1703         scmd->request_buffer            = NULL;
1704         scmd->request_bufflen           = 0;
1705
1706         scmd->cmd_len                   = 0;
1707
1708         scmd->sc_data_direction         = DMA_BIDIRECTIONAL;
1709
1710         init_timer(&scmd->eh_timeout);
1711
1712         spin_lock_irqsave(shost->host_lock, flags);
1713         shost->tmf_in_progress = 1;
1714         spin_unlock_irqrestore(shost->host_lock, flags);
1715
1716         switch (flag) {
1717         case SCSI_TRY_RESET_DEVICE:
1718                 rtn = scsi_try_bus_device_reset(scmd);
1719                 if (rtn == SUCCESS)
1720                         break;
1721                 /* FALLTHROUGH */
1722         case SCSI_TRY_RESET_BUS:
1723                 rtn = scsi_try_bus_reset(scmd);
1724                 if (rtn == SUCCESS)
1725                         break;
1726                 /* FALLTHROUGH */
1727         case SCSI_TRY_RESET_HOST:
1728                 rtn = scsi_try_host_reset(scmd);
1729                 break;
1730         default:
1731                 rtn = FAILED;
1732         }
1733
1734         spin_lock_irqsave(shost->host_lock, flags);
1735         shost->tmf_in_progress = 0;
1736         spin_unlock_irqrestore(shost->host_lock, flags);
1737
1738         /*
1739          * be sure to wake up anyone who was sleeping or had their queue
1740          * suspended while we performed the TMF.
1741          */
1742         SCSI_LOG_ERROR_RECOVERY(3,
1743                 printk("%s: waking up host to restart after TMF\n",
1744                 __FUNCTION__));
1745
1746         wake_up(&shost->host_wait);
1747
1748         scsi_run_host_queues(shost);
1749
1750         scsi_next_command(scmd);
1751         return rtn;
1752 }
1753 EXPORT_SYMBOL(scsi_reset_provider);
1754
1755 /**
1756  * scsi_normalize_sense - normalize main elements from either fixed or
1757  *                      descriptor sense data format into a common format.
1758  *
1759  * @sense_buffer:       byte array containing sense data returned by device
1760  * @sb_len:             number of valid bytes in sense_buffer
1761  * @sshdr:              pointer to instance of structure that common
1762  *                      elements are written to.
1763  *
1764  * Notes:
1765  *      The "main elements" from sense data are: response_code, sense_key,
1766  *      asc, ascq and additional_length (only for descriptor format).
1767  *
1768  *      Typically this function can be called after a device has
1769  *      responded to a SCSI command with the CHECK_CONDITION status.
1770  *
1771  * Return value:
1772  *      1 if valid sense data information found, else 0;
1773  **/
1774 int scsi_normalize_sense(const u8 *sense_buffer, int sb_len,
1775                          struct scsi_sense_hdr *sshdr)
1776 {
1777         if (!sense_buffer || !sb_len)
1778                 return 0;
1779
1780         memset(sshdr, 0, sizeof(struct scsi_sense_hdr));
1781
1782         sshdr->response_code = (sense_buffer[0] & 0x7f);
1783
1784         if (!scsi_sense_valid(sshdr))
1785                 return 0;
1786
1787         if (sshdr->response_code >= 0x72) {
1788                 /*
1789                  * descriptor format
1790                  */
1791                 if (sb_len > 1)
1792                         sshdr->sense_key = (sense_buffer[1] & 0xf);
1793                 if (sb_len > 2)
1794                         sshdr->asc = sense_buffer[2];
1795                 if (sb_len > 3)
1796                         sshdr->ascq = sense_buffer[3];
1797                 if (sb_len > 7)
1798                         sshdr->additional_length = sense_buffer[7];
1799         } else {
1800                 /* 
1801                  * fixed format
1802                  */
1803                 if (sb_len > 2)
1804                         sshdr->sense_key = (sense_buffer[2] & 0xf);
1805                 if (sb_len > 7) {
1806                         sb_len = (sb_len < (sense_buffer[7] + 8)) ?
1807                                          sb_len : (sense_buffer[7] + 8);
1808                         if (sb_len > 12)
1809                                 sshdr->asc = sense_buffer[12];
1810                         if (sb_len > 13)
1811                                 sshdr->ascq = sense_buffer[13];
1812                 }
1813         }
1814
1815         return 1;
1816 }
1817 EXPORT_SYMBOL(scsi_normalize_sense);
1818
1819 int scsi_command_normalize_sense(struct scsi_cmnd *cmd,
1820                                  struct scsi_sense_hdr *sshdr)
1821 {
1822         return scsi_normalize_sense(cmd->sense_buffer,
1823                         sizeof(cmd->sense_buffer), sshdr);
1824 }
1825 EXPORT_SYMBOL(scsi_command_normalize_sense);
1826
1827 /**
1828  * scsi_sense_desc_find - search for a given descriptor type in
1829  *                      descriptor sense data format.
1830  *
1831  * @sense_buffer:       byte array of descriptor format sense data
1832  * @sb_len:             number of valid bytes in sense_buffer
1833  * @desc_type:          value of descriptor type to find
1834  *                      (e.g. 0 -> information)
1835  *
1836  * Notes:
1837  *      only valid when sense data is in descriptor format
1838  *
1839  * Return value:
1840  *      pointer to start of (first) descriptor if found else NULL
1841  **/
1842 const u8 * scsi_sense_desc_find(const u8 * sense_buffer, int sb_len,
1843                                 int desc_type)
1844 {
1845         int add_sen_len, add_len, desc_len, k;
1846         const u8 * descp;
1847
1848         if ((sb_len < 8) || (0 == (add_sen_len = sense_buffer[7])))
1849                 return NULL;
1850         if ((sense_buffer[0] < 0x72) || (sense_buffer[0] > 0x73))
1851                 return NULL;
1852         add_sen_len = (add_sen_len < (sb_len - 8)) ?
1853                         add_sen_len : (sb_len - 8);
1854         descp = &sense_buffer[8];
1855         for (desc_len = 0, k = 0; k < add_sen_len; k += desc_len) {
1856                 descp += desc_len;
1857                 add_len = (k < (add_sen_len - 1)) ? descp[1]: -1;
1858                 desc_len = add_len + 2;
1859                 if (descp[0] == desc_type)
1860                         return descp;
1861                 if (add_len < 0) // short descriptor ??
1862                         break;
1863         }
1864         return NULL;
1865 }
1866 EXPORT_SYMBOL(scsi_sense_desc_find);
1867
1868 /**
1869  * scsi_get_sense_info_fld - attempts to get information field from
1870  *                      sense data (either fixed or descriptor format)
1871  *
1872  * @sense_buffer:       byte array of sense data
1873  * @sb_len:             number of valid bytes in sense_buffer
1874  * @info_out:           pointer to 64 integer where 8 or 4 byte information
1875  *                      field will be placed if found.
1876  *
1877  * Return value:
1878  *      1 if information field found, 0 if not found.
1879  **/
1880 int scsi_get_sense_info_fld(const u8 * sense_buffer, int sb_len,
1881                             u64 * info_out)
1882 {
1883         int j;
1884         const u8 * ucp;
1885         u64 ull;
1886
1887         if (sb_len < 7)
1888                 return 0;
1889         switch (sense_buffer[0] & 0x7f) {
1890         case 0x70:
1891         case 0x71:
1892                 if (sense_buffer[0] & 0x80) {
1893                         *info_out = (sense_buffer[3] << 24) +
1894                                     (sense_buffer[4] << 16) +
1895                                     (sense_buffer[5] << 8) + sense_buffer[6];
1896                         return 1;
1897                 } else
1898                         return 0;
1899         case 0x72:
1900         case 0x73:
1901                 ucp = scsi_sense_desc_find(sense_buffer, sb_len,
1902                                            0 /* info desc */);
1903                 if (ucp && (0xa == ucp[1])) {
1904                         ull = 0;
1905                         for (j = 0; j < 8; ++j) {
1906                                 if (j > 0)
1907                                         ull <<= 8;
1908                                 ull |= ucp[4 + j];
1909                         }
1910                         *info_out = ull;
1911                         return 1;
1912                 } else
1913                         return 0;
1914         default:
1915                 return 0;
1916         }
1917 }
1918 EXPORT_SYMBOL(scsi_get_sense_info_fld);