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