]> pilppa.org Git - linux-2.6-omap-h63xx.git/blob - drivers/scsi/scsi_error.c
[SCSI] expose eh_timed_out to the host template
[linux-2.6-omap-h63xx.git] / drivers / scsi / scsi_error.c
1 /*
2  *  scsi_error.c Copyright (C) 1997 Eric Youngdale
3  *
4  *  SCSI error/timeout handling
5  *      Initial versions: Eric Youngdale.  Based upon conversations with
6  *                        Leonard Zubkoff and David Miller at Linux Expo, 
7  *                        ideas originating from all over the place.
8  *
9  *      Restructured scsi_unjam_host and associated functions.
10  *      September 04, 2002 Mike Anderson (andmike@us.ibm.com)
11  *
12  *      Forward port of Russell King's (rmk@arm.linux.org.uk) changes and
13  *      minor  cleanups.
14  *      September 30, 2002 Mike Anderson (andmike@us.ibm.com)
15  */
16
17 #include <linux/module.h>
18 #include <linux/sched.h>
19 #include <linux/timer.h>
20 #include <linux/string.h>
21 #include <linux/slab.h>
22 #include <linux/kernel.h>
23 #include <linux/kthread.h>
24 #include <linux/interrupt.h>
25 #include <linux/blkdev.h>
26 #include <linux/delay.h>
27
28 #include <scsi/scsi.h>
29 #include <scsi/scsi_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         enum scsi_eh_timer_return (* eh_timed_out)(struct scsi_cmnd *);
188
189         scsi_log_completion(scmd, TIMEOUT_ERROR);
190
191         if (scmd->device->host->transportt->eh_timed_out)
192                 eh_timed_out = scmd->device->host->transportt->eh_timed_out;
193         else if (scmd->device->host->hostt->eh_timed_out)
194                 eh_timed_out = scmd->device->host->hostt->eh_timed_out;
195         else
196                 eh_timed_out = NULL;
197
198         if (eh_timed_out)
199                 switch (eh_timed_out(scmd)) {
200                 case EH_HANDLED:
201                         __scsi_done(scmd);
202                         return;
203                 case EH_RESET_TIMER:
204                         scsi_add_timer(scmd, scmd->timeout_per_command,
205                                        scsi_times_out);
206                         return;
207                 case EH_NOT_HANDLED:
208                         break;
209                 }
210
211         if (unlikely(!scsi_eh_scmd_add(scmd, SCSI_EH_CANCEL_CMD))) {
212                 scmd->result |= DID_TIME_OUT << 16;
213                 __scsi_done(scmd);
214         }
215 }
216
217 /**
218  * scsi_block_when_processing_errors - Prevent cmds from being queued.
219  * @sdev:       Device on which we are performing recovery.
220  *
221  * Description:
222  *     We block until the host is out of error recovery, and then check to
223  *     see whether the host or the device is offline.
224  *
225  * Return value:
226  *     0 when dev was taken offline by error recovery. 1 OK to proceed.
227  **/
228 int scsi_block_when_processing_errors(struct scsi_device *sdev)
229 {
230         int online;
231
232         wait_event(sdev->host->host_wait, !scsi_host_in_recovery(sdev->host));
233
234         online = scsi_device_online(sdev);
235
236         SCSI_LOG_ERROR_RECOVERY(5, printk("%s: rtn: %d\n", __FUNCTION__,
237                                           online));
238
239         return online;
240 }
241 EXPORT_SYMBOL(scsi_block_when_processing_errors);
242
243 #ifdef CONFIG_SCSI_LOGGING
244 /**
245  * scsi_eh_prt_fail_stats - Log info on failures.
246  * @shost:      scsi host being recovered.
247  * @work_q:     Queue of scsi cmds to process.
248  **/
249 static inline void scsi_eh_prt_fail_stats(struct Scsi_Host *shost,
250                                           struct list_head *work_q)
251 {
252         struct scsi_cmnd *scmd;
253         struct scsi_device *sdev;
254         int total_failures = 0;
255         int cmd_failed = 0;
256         int cmd_cancel = 0;
257         int devices_failed = 0;
258
259         shost_for_each_device(sdev, shost) {
260                 list_for_each_entry(scmd, work_q, eh_entry) {
261                         if (scmd->device == sdev) {
262                                 ++total_failures;
263                                 if (scmd->eh_eflags & SCSI_EH_CANCEL_CMD)
264                                         ++cmd_cancel;
265                                 else 
266                                         ++cmd_failed;
267                         }
268                 }
269
270                 if (cmd_cancel || cmd_failed) {
271                         SCSI_LOG_ERROR_RECOVERY(3,
272                                 sdev_printk(KERN_INFO, sdev,
273                                             "%s: cmds failed: %d, cancel: %d\n",
274                                             __FUNCTION__, cmd_failed,
275                                             cmd_cancel));
276                         cmd_cancel = 0;
277                         cmd_failed = 0;
278                         ++devices_failed;
279                 }
280         }
281
282         SCSI_LOG_ERROR_RECOVERY(2, printk("Total of %d commands on %d"
283                                           " devices require eh work\n",
284                                   total_failures, devices_failed));
285 }
286 #endif
287
288 /**
289  * scsi_check_sense - Examine scsi cmd sense
290  * @scmd:       Cmd to have sense checked.
291  *
292  * Return value:
293  *      SUCCESS or FAILED or NEEDS_RETRY
294  *
295  * Notes:
296  *      When a deferred error is detected the current command has
297  *      not been executed and needs retrying.
298  **/
299 static int scsi_check_sense(struct scsi_cmnd *scmd)
300 {
301         struct scsi_sense_hdr sshdr;
302
303         if (! scsi_command_normalize_sense(scmd, &sshdr))
304                 return FAILED;  /* no valid sense data */
305
306         if (scsi_sense_is_deferred(&sshdr))
307                 return NEEDS_RETRY;
308
309         /*
310          * Previous logic looked for FILEMARK, EOM or ILI which are
311          * mainly associated with tapes and returned SUCCESS.
312          */
313         if (sshdr.response_code == 0x70) {
314                 /* fixed format */
315                 if (scmd->sense_buffer[2] & 0xe0)
316                         return SUCCESS;
317         } else {
318                 /*
319                  * descriptor format: look for "stream commands sense data
320                  * descriptor" (see SSC-3). Assume single sense data
321                  * descriptor. Ignore ILI from SBC-2 READ LONG and WRITE LONG.
322                  */
323                 if ((sshdr.additional_length > 3) &&
324                     (scmd->sense_buffer[8] == 0x4) &&
325                     (scmd->sense_buffer[11] & 0xe0))
326                         return SUCCESS;
327         }
328
329         switch (sshdr.sense_key) {
330         case NO_SENSE:
331                 return SUCCESS;
332         case RECOVERED_ERROR:
333                 return /* soft_error */ SUCCESS;
334
335         case ABORTED_COMMAND:
336                 return NEEDS_RETRY;
337         case NOT_READY:
338         case UNIT_ATTENTION:
339                 /*
340                  * if we are expecting a cc/ua because of a bus reset that we
341                  * performed, treat this just as a retry.  otherwise this is
342                  * information that we should pass up to the upper-level driver
343                  * so that we can deal with it there.
344                  */
345                 if (scmd->device->expecting_cc_ua) {
346                         scmd->device->expecting_cc_ua = 0;
347                         return NEEDS_RETRY;
348                 }
349                 /*
350                  * if the device is in the process of becoming ready, we 
351                  * should retry.
352                  */
353                 if ((sshdr.asc == 0x04) && (sshdr.ascq == 0x01))
354                         return NEEDS_RETRY;
355                 /*
356                  * if the device is not started, we need to wake
357                  * the error handler to start the motor
358                  */
359                 if (scmd->device->allow_restart &&
360                     (sshdr.asc == 0x04) && (sshdr.ascq == 0x02))
361                         return FAILED;
362                 return SUCCESS;
363
364                 /* these three are not supported */
365         case COPY_ABORTED:
366         case VOLUME_OVERFLOW:
367         case MISCOMPARE:
368                 return SUCCESS;
369
370         case MEDIUM_ERROR:
371                 if (sshdr.asc == 0x11 || /* UNRECOVERED READ ERR */
372                     sshdr.asc == 0x13 || /* AMNF DATA FIELD */
373                     sshdr.asc == 0x14) { /* RECORD NOT FOUND */
374                         return SUCCESS;
375                 }
376                 return NEEDS_RETRY;
377
378         case HARDWARE_ERROR:
379                 if (scmd->device->retry_hwerror)
380                         return NEEDS_RETRY;
381                 else
382                         return SUCCESS;
383
384         case ILLEGAL_REQUEST:
385         case BLANK_CHECK:
386         case DATA_PROTECT:
387         default:
388                 return SUCCESS;
389         }
390 }
391
392 /**
393  * scsi_eh_completed_normally - Disposition a eh cmd on return from LLD.
394  * @scmd:       SCSI cmd to examine.
395  *
396  * Notes:
397  *    This is *only* called when we are examining the status of commands
398  *    queued during error recovery.  the main difference here is that we
399  *    don't allow for the possibility of retries here, and we are a lot
400  *    more restrictive about what we consider acceptable.
401  **/
402 static int scsi_eh_completed_normally(struct scsi_cmnd *scmd)
403 {
404         /*
405          * first check the host byte, to see if there is anything in there
406          * that would indicate what we need to do.
407          */
408         if (host_byte(scmd->result) == DID_RESET) {
409                 /*
410                  * rats.  we are already in the error handler, so we now
411                  * get to try and figure out what to do next.  if the sense
412                  * is valid, we have a pretty good idea of what to do.
413                  * if not, we mark it as FAILED.
414                  */
415                 return scsi_check_sense(scmd);
416         }
417         if (host_byte(scmd->result) != DID_OK)
418                 return FAILED;
419
420         /*
421          * next, check the message byte.
422          */
423         if (msg_byte(scmd->result) != COMMAND_COMPLETE)
424                 return FAILED;
425
426         /*
427          * now, check the status byte to see if this indicates
428          * anything special.
429          */
430         switch (status_byte(scmd->result)) {
431         case GOOD:
432         case COMMAND_TERMINATED:
433                 return SUCCESS;
434         case CHECK_CONDITION:
435                 return scsi_check_sense(scmd);
436         case CONDITION_GOOD:
437         case INTERMEDIATE_GOOD:
438         case INTERMEDIATE_C_GOOD:
439                 /*
440                  * who knows?  FIXME(eric)
441                  */
442                 return SUCCESS;
443         case BUSY:
444         case QUEUE_FULL:
445         case RESERVATION_CONFLICT:
446         default:
447                 return FAILED;
448         }
449         return FAILED;
450 }
451
452 /**
453  * scsi_eh_done - Completion function for error handling.
454  * @scmd:       Cmd that is done.
455  **/
456 static void scsi_eh_done(struct scsi_cmnd *scmd)
457 {
458         struct completion     *eh_action;
459
460         SCSI_LOG_ERROR_RECOVERY(3,
461                 printk("%s scmd: %p result: %x\n",
462                         __FUNCTION__, scmd, scmd->result));
463
464         eh_action = scmd->device->host->eh_action;
465         if (eh_action)
466                 complete(eh_action);
467 }
468
469 /**
470  * scsi_try_host_reset - ask host adapter to reset itself
471  * @scmd:       SCSI cmd to send hsot reset.
472  **/
473 static int scsi_try_host_reset(struct scsi_cmnd *scmd)
474 {
475         unsigned long flags;
476         int rtn;
477
478         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Snd Host RST\n",
479                                           __FUNCTION__));
480
481         if (!scmd->device->host->hostt->eh_host_reset_handler)
482                 return FAILED;
483
484         rtn = scmd->device->host->hostt->eh_host_reset_handler(scmd);
485
486         if (rtn == SUCCESS) {
487                 if (!scmd->device->host->hostt->skip_settle_delay)
488                         ssleep(HOST_RESET_SETTLE_TIME);
489                 spin_lock_irqsave(scmd->device->host->host_lock, flags);
490                 scsi_report_bus_reset(scmd->device->host,
491                                       scmd_channel(scmd));
492                 spin_unlock_irqrestore(scmd->device->host->host_lock, flags);
493         }
494
495         return rtn;
496 }
497
498 /**
499  * scsi_try_bus_reset - ask host to perform a bus reset
500  * @scmd:       SCSI cmd to send bus reset.
501  **/
502 static int scsi_try_bus_reset(struct scsi_cmnd *scmd)
503 {
504         unsigned long flags;
505         int rtn;
506
507         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Snd Bus RST\n",
508                                           __FUNCTION__));
509
510         if (!scmd->device->host->hostt->eh_bus_reset_handler)
511                 return FAILED;
512
513         rtn = scmd->device->host->hostt->eh_bus_reset_handler(scmd);
514
515         if (rtn == SUCCESS) {
516                 if (!scmd->device->host->hostt->skip_settle_delay)
517                         ssleep(BUS_RESET_SETTLE_TIME);
518                 spin_lock_irqsave(scmd->device->host->host_lock, flags);
519                 scsi_report_bus_reset(scmd->device->host,
520                                       scmd_channel(scmd));
521                 spin_unlock_irqrestore(scmd->device->host->host_lock, flags);
522         }
523
524         return rtn;
525 }
526
527 /**
528  * scsi_try_bus_device_reset - Ask host to perform a BDR on a dev
529  * @scmd:       SCSI cmd used to send BDR
530  *
531  * Notes:
532  *    There is no timeout for this operation.  if this operation is
533  *    unreliable for a given host, then the host itself needs to put a
534  *    timer on it, and set the host back to a consistent state prior to
535  *    returning.
536  **/
537 static int scsi_try_bus_device_reset(struct scsi_cmnd *scmd)
538 {
539         int rtn;
540
541         if (!scmd->device->host->hostt->eh_device_reset_handler)
542                 return FAILED;
543
544         rtn = scmd->device->host->hostt->eh_device_reset_handler(scmd);
545         if (rtn == SUCCESS) {
546                 scmd->device->was_reset = 1;
547                 scmd->device->expecting_cc_ua = 1;
548         }
549
550         return rtn;
551 }
552
553 static int __scsi_try_to_abort_cmd(struct scsi_cmnd *scmd)
554 {
555         if (!scmd->device->host->hostt->eh_abort_handler)
556                 return FAILED;
557
558         return scmd->device->host->hostt->eh_abort_handler(scmd);
559 }
560
561 /**
562  * scsi_try_to_abort_cmd - Ask host to abort a running command.
563  * @scmd:       SCSI cmd to abort from Lower Level.
564  *
565  * Notes:
566  *    This function will not return until the user's completion function
567  *    has been called.  there is no timeout on this operation.  if the
568  *    author of the low-level driver wishes this operation to be timed,
569  *    they can provide this facility themselves.  helper functions in
570  *    scsi_error.c can be supplied to make this easier to do.
571  **/
572 static int scsi_try_to_abort_cmd(struct scsi_cmnd *scmd)
573 {
574         /*
575          * scsi_done was called just after the command timed out and before
576          * we had a chance to process it. (db)
577          */
578         if (scmd->serial_number == 0)
579                 return SUCCESS;
580         return __scsi_try_to_abort_cmd(scmd);
581 }
582
583 static void scsi_abort_eh_cmnd(struct scsi_cmnd *scmd)
584 {
585         if (__scsi_try_to_abort_cmd(scmd) != SUCCESS)
586                 if (scsi_try_bus_device_reset(scmd) != SUCCESS)
587                         if (scsi_try_bus_reset(scmd) != SUCCESS)
588                                 scsi_try_host_reset(scmd);
589 }
590
591 /**
592  * scsi_send_eh_cmnd  - submit a scsi command as part of error recory
593  * @scmd:       SCSI command structure to hijack
594  * @cmnd:       CDB to send
595  * @cmnd_size:  size in bytes of @cmnd
596  * @timeout:    timeout for this request
597  * @copy_sense: request sense data if set to 1
598  *
599  * This function is used to send a scsi command down to a target device
600  * as part of the error recovery process.  If @copy_sense is 0 the command
601  * sent must be one that does not transfer any data.  If @copy_sense is 1
602  * the command must be REQUEST_SENSE and this functions copies out the
603  * sense buffer it got into @scmd->sense_buffer.
604  *
605  * Return value:
606  *    SUCCESS or FAILED or NEEDS_RETRY
607  **/
608 static int scsi_send_eh_cmnd(struct scsi_cmnd *scmd, unsigned char *cmnd,
609                              int cmnd_size, int timeout, int copy_sense)
610 {
611         struct scsi_device *sdev = scmd->device;
612         struct Scsi_Host *shost = sdev->host;
613         int old_result = scmd->result;
614         DECLARE_COMPLETION_ONSTACK(done);
615         unsigned long timeleft;
616         unsigned long flags;
617         struct scatterlist sgl;
618         unsigned char old_cmnd[MAX_COMMAND_SIZE];
619         enum dma_data_direction old_data_direction;
620         unsigned short old_use_sg;
621         unsigned char old_cmd_len;
622         unsigned old_bufflen;
623         void *old_buffer;
624         int rtn;
625
626         /*
627          * We need saved copies of a number of fields - this is because
628          * error handling may need to overwrite these with different values
629          * to run different commands, and once error handling is complete,
630          * we will need to restore these values prior to running the actual
631          * command.
632          */
633         old_buffer = scmd->request_buffer;
634         old_bufflen = scmd->request_bufflen;
635         memcpy(old_cmnd, scmd->cmnd, sizeof(scmd->cmnd));
636         old_data_direction = scmd->sc_data_direction;
637         old_cmd_len = scmd->cmd_len;
638         old_use_sg = scmd->use_sg;
639
640         memset(scmd->cmnd, 0, sizeof(scmd->cmnd));
641         memcpy(scmd->cmnd, cmnd, cmnd_size);
642
643         if (copy_sense) {
644                 gfp_t gfp_mask = GFP_ATOMIC;
645
646                 if (shost->hostt->unchecked_isa_dma)
647                         gfp_mask |= __GFP_DMA;
648
649                 sgl.page = alloc_page(gfp_mask);
650                 if (!sgl.page)
651                         return FAILED;
652                 sgl.offset = 0;
653                 sgl.length = 252;
654
655                 scmd->sc_data_direction = DMA_FROM_DEVICE;
656                 scmd->request_bufflen = sgl.length;
657                 scmd->request_buffer = &sgl;
658                 scmd->use_sg = 1;
659         } else {
660                 scmd->request_buffer = NULL;
661                 scmd->request_bufflen = 0;
662                 scmd->sc_data_direction = DMA_NONE;
663                 scmd->use_sg = 0;
664         }
665
666         scmd->underflow = 0;
667         scmd->cmd_len = COMMAND_SIZE(scmd->cmnd[0]);
668
669         if (sdev->scsi_level <= SCSI_2)
670                 scmd->cmnd[1] = (scmd->cmnd[1] & 0x1f) |
671                         (sdev->lun << 5 & 0xe0);
672
673         /*
674          * Zero the sense buffer.  The scsi spec mandates that any
675          * untransferred sense data should be interpreted as being zero.
676          */
677         memset(scmd->sense_buffer, 0, sizeof(scmd->sense_buffer));
678
679         shost->eh_action = &done;
680
681         spin_lock_irqsave(shost->host_lock, flags);
682         scsi_log_send(scmd);
683         shost->hostt->queuecommand(scmd, scsi_eh_done);
684         spin_unlock_irqrestore(shost->host_lock, flags);
685
686         timeleft = wait_for_completion_timeout(&done, timeout);
687
688         shost->eh_action = NULL;
689
690         scsi_log_completion(scmd, SUCCESS);
691
692         SCSI_LOG_ERROR_RECOVERY(3,
693                 printk("%s: scmd: %p, timeleft: %ld\n",
694                         __FUNCTION__, scmd, timeleft));
695
696         /*
697          * If there is time left scsi_eh_done got called, and we will
698          * examine the actual status codes to see whether the command
699          * actually did complete normally, else tell the host to forget
700          * about this command.
701          */
702         if (timeleft) {
703                 rtn = scsi_eh_completed_normally(scmd);
704                 SCSI_LOG_ERROR_RECOVERY(3,
705                         printk("%s: scsi_eh_completed_normally %x\n",
706                                __FUNCTION__, rtn));
707
708                 switch (rtn) {
709                 case SUCCESS:
710                 case NEEDS_RETRY:
711                 case FAILED:
712                         break;
713                 default:
714                         rtn = FAILED;
715                         break;
716                 }
717         } else {
718                 scsi_abort_eh_cmnd(scmd);
719                 rtn = FAILED;
720         }
721
722
723         /*
724          * Last chance to have valid sense data.
725          */
726         if (copy_sense) {
727                 if (!SCSI_SENSE_VALID(scmd)) {
728                         memcpy(scmd->sense_buffer, scmd->request_buffer,
729                                sizeof(scmd->sense_buffer));
730                 }
731                 __free_page(sgl.page);
732         }
733
734
735         /*
736          * Restore original data
737          */
738         scmd->request_buffer = old_buffer;
739         scmd->request_bufflen = old_bufflen;
740         memcpy(scmd->cmnd, old_cmnd, sizeof(scmd->cmnd));
741         scmd->sc_data_direction = old_data_direction;
742         scmd->cmd_len = old_cmd_len;
743         scmd->use_sg = old_use_sg;
744         scmd->result = old_result;
745         return rtn;
746 }
747
748 /**
749  * scsi_request_sense - Request sense data from a particular target.
750  * @scmd:       SCSI cmd for request sense.
751  *
752  * Notes:
753  *    Some hosts automatically obtain this information, others require
754  *    that we obtain it on our own. This function will *not* return until
755  *    the command either times out, or it completes.
756  **/
757 static int scsi_request_sense(struct scsi_cmnd *scmd)
758 {
759         static unsigned char generic_sense[6] =
760                 {REQUEST_SENSE, 0, 0, 0, 252, 0};
761
762         return scsi_send_eh_cmnd(scmd, generic_sense, 6, SENSE_TIMEOUT, 1);
763 }
764
765 /**
766  * scsi_eh_finish_cmd - Handle a cmd that eh is finished with.
767  * @scmd:       Original SCSI cmd that eh has finished.
768  * @done_q:     Queue for processed commands.
769  *
770  * Notes:
771  *    We don't want to use the normal command completion while we are are
772  *    still handling errors - it may cause other commands to be queued,
773  *    and that would disturb what we are doing.  thus we really want to
774  *    keep a list of pending commands for final completion, and once we
775  *    are ready to leave error handling we handle completion for real.
776  **/
777 void scsi_eh_finish_cmd(struct scsi_cmnd *scmd, struct list_head *done_q)
778 {
779         scmd->device->host->host_failed--;
780         scmd->eh_eflags = 0;
781         list_move_tail(&scmd->eh_entry, done_q);
782 }
783 EXPORT_SYMBOL(scsi_eh_finish_cmd);
784
785 /**
786  * scsi_eh_get_sense - Get device sense data.
787  * @work_q:     Queue of commands to process.
788  * @done_q:     Queue of proccessed commands..
789  *
790  * Description:
791  *    See if we need to request sense information.  if so, then get it
792  *    now, so we have a better idea of what to do.  
793  *
794  * Notes:
795  *    This has the unfortunate side effect that if a shost adapter does
796  *    not automatically request sense information, that we end up shutting
797  *    it down before we request it.
798  *
799  *    All drivers should request sense information internally these days,
800  *    so for now all I have to say is tough noogies if you end up in here.
801  *
802  *    XXX: Long term this code should go away, but that needs an audit of
803  *         all LLDDs first.
804  **/
805 int scsi_eh_get_sense(struct list_head *work_q,
806                       struct list_head *done_q)
807 {
808         struct scsi_cmnd *scmd, *next;
809         int rtn;
810
811         list_for_each_entry_safe(scmd, next, work_q, eh_entry) {
812                 if ((scmd->eh_eflags & SCSI_EH_CANCEL_CMD) ||
813                     SCSI_SENSE_VALID(scmd))
814                         continue;
815
816                 SCSI_LOG_ERROR_RECOVERY(2, scmd_printk(KERN_INFO, scmd,
817                                                   "%s: requesting sense\n",
818                                                   current->comm));
819                 rtn = scsi_request_sense(scmd);
820                 if (rtn != SUCCESS)
821                         continue;
822
823                 SCSI_LOG_ERROR_RECOVERY(3, printk("sense requested for %p"
824                                                   " result %x\n", scmd,
825                                                   scmd->result));
826                 SCSI_LOG_ERROR_RECOVERY(3, scsi_print_sense("bh", scmd));
827
828                 rtn = scsi_decide_disposition(scmd);
829
830                 /*
831                  * if the result was normal, then just pass it along to the
832                  * upper level.
833                  */
834                 if (rtn == SUCCESS)
835                         /* we don't want this command reissued, just
836                          * finished with the sense data, so set
837                          * retries to the max allowed to ensure it
838                          * won't get reissued */
839                         scmd->retries = scmd->allowed;
840                 else if (rtn != NEEDS_RETRY)
841                         continue;
842
843                 scsi_eh_finish_cmd(scmd, done_q);
844         }
845
846         return list_empty(work_q);
847 }
848 EXPORT_SYMBOL_GPL(scsi_eh_get_sense);
849
850 /**
851  * scsi_eh_tur - Send TUR to device.
852  * @scmd:       Scsi cmd to send TUR
853  *
854  * Return value:
855  *    0 - Device is ready. 1 - Device NOT ready.
856  **/
857 static int scsi_eh_tur(struct scsi_cmnd *scmd)
858 {
859         static unsigned char tur_command[6] = {TEST_UNIT_READY, 0, 0, 0, 0, 0};
860         int retry_cnt = 1, rtn;
861
862 retry_tur:
863         rtn = scsi_send_eh_cmnd(scmd, tur_command, 6, SENSE_TIMEOUT, 0);
864
865         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: scmd %p rtn %x\n",
866                 __FUNCTION__, scmd, rtn));
867
868         switch (rtn) {
869         case NEEDS_RETRY:
870                 if (retry_cnt--)
871                         goto retry_tur;
872                 /*FALLTHRU*/
873         case SUCCESS:
874                 return 0;
875         default:
876                 return 1;
877         }
878 }
879
880 /**
881  * scsi_eh_abort_cmds - abort canceled commands.
882  * @shost:      scsi host being recovered.
883  * @eh_done_q:  list_head for processed commands.
884  *
885  * Decription:
886  *    Try and see whether or not it makes sense to try and abort the
887  *    running command.  this only works out to be the case if we have one
888  *    command that has timed out.  if the command simply failed, it makes
889  *    no sense to try and abort the command, since as far as the shost
890  *    adapter is concerned, it isn't running.
891  **/
892 static int scsi_eh_abort_cmds(struct list_head *work_q,
893                               struct list_head *done_q)
894 {
895         struct scsi_cmnd *scmd, *next;
896         int rtn;
897
898         list_for_each_entry_safe(scmd, next, work_q, eh_entry) {
899                 if (!(scmd->eh_eflags & SCSI_EH_CANCEL_CMD))
900                         continue;
901                 SCSI_LOG_ERROR_RECOVERY(3, printk("%s: aborting cmd:"
902                                                   "0x%p\n", current->comm,
903                                                   scmd));
904                 rtn = scsi_try_to_abort_cmd(scmd);
905                 if (rtn == SUCCESS) {
906                         scmd->eh_eflags &= ~SCSI_EH_CANCEL_CMD;
907                         if (!scsi_device_online(scmd->device) ||
908                             !scsi_eh_tur(scmd)) {
909                                 scsi_eh_finish_cmd(scmd, done_q);
910                         }
911                                 
912                 } else
913                         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: aborting"
914                                                           " cmd failed:"
915                                                           "0x%p\n",
916                                                           current->comm,
917                                                           scmd));
918         }
919
920         return list_empty(work_q);
921 }
922
923 /**
924  * scsi_eh_try_stu - Send START_UNIT to device.
925  * @scmd:       Scsi cmd to send START_UNIT
926  *
927  * Return value:
928  *    0 - Device is ready. 1 - Device NOT ready.
929  **/
930 static int scsi_eh_try_stu(struct scsi_cmnd *scmd)
931 {
932         static unsigned char stu_command[6] = {START_STOP, 0, 0, 0, 1, 0};
933
934         if (scmd->device->allow_restart) {
935                 int rtn;
936
937                 rtn = scsi_send_eh_cmnd(scmd, stu_command, 6,
938                                         START_UNIT_TIMEOUT, 0);
939                 if (rtn == SUCCESS)
940                         return 0;
941         }
942
943         return 1;
944 }
945
946  /**
947  * scsi_eh_stu - send START_UNIT if needed
948  * @shost:      scsi host being recovered.
949  * @eh_done_q:  list_head for processed commands.
950  *
951  * Notes:
952  *    If commands are failing due to not ready, initializing command required,
953  *      try revalidating the device, which will end up sending a start unit. 
954  **/
955 static int scsi_eh_stu(struct Scsi_Host *shost,
956                               struct list_head *work_q,
957                               struct list_head *done_q)
958 {
959         struct scsi_cmnd *scmd, *stu_scmd, *next;
960         struct scsi_device *sdev;
961
962         shost_for_each_device(sdev, shost) {
963                 stu_scmd = NULL;
964                 list_for_each_entry(scmd, work_q, eh_entry)
965                         if (scmd->device == sdev && SCSI_SENSE_VALID(scmd) &&
966                             scsi_check_sense(scmd) == FAILED ) {
967                                 stu_scmd = scmd;
968                                 break;
969                         }
970
971                 if (!stu_scmd)
972                         continue;
973
974                 SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Sending START_UNIT to sdev:"
975                                                   " 0x%p\n", current->comm, sdev));
976
977                 if (!scsi_eh_try_stu(stu_scmd)) {
978                         if (!scsi_device_online(sdev) ||
979                             !scsi_eh_tur(stu_scmd)) {
980                                 list_for_each_entry_safe(scmd, next,
981                                                           work_q, eh_entry) {
982                                         if (scmd->device == sdev)
983                                                 scsi_eh_finish_cmd(scmd, done_q);
984                                 }
985                         }
986                 } else {
987                         SCSI_LOG_ERROR_RECOVERY(3,
988                                                 printk("%s: START_UNIT failed to sdev:"
989                                                        " 0x%p\n", current->comm, sdev));
990                 }
991         }
992
993         return list_empty(work_q);
994 }
995
996
997 /**
998  * scsi_eh_bus_device_reset - send bdr if needed
999  * @shost:      scsi host being recovered.
1000  * @eh_done_q:  list_head for processed commands.
1001  *
1002  * Notes:
1003  *    Try a bus device reset.  still, look to see whether we have multiple
1004  *    devices that are jammed or not - if we have multiple devices, it
1005  *    makes no sense to try bus_device_reset - we really would need to try
1006  *    a bus_reset instead. 
1007  **/
1008 static int scsi_eh_bus_device_reset(struct Scsi_Host *shost,
1009                                     struct list_head *work_q,
1010                                     struct list_head *done_q)
1011 {
1012         struct scsi_cmnd *scmd, *bdr_scmd, *next;
1013         struct scsi_device *sdev;
1014         int rtn;
1015
1016         shost_for_each_device(sdev, shost) {
1017                 bdr_scmd = NULL;
1018                 list_for_each_entry(scmd, work_q, eh_entry)
1019                         if (scmd->device == sdev) {
1020                                 bdr_scmd = scmd;
1021                                 break;
1022                         }
1023
1024                 if (!bdr_scmd)
1025                         continue;
1026
1027                 SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Sending BDR sdev:"
1028                                                   " 0x%p\n", current->comm,
1029                                                   sdev));
1030                 rtn = scsi_try_bus_device_reset(bdr_scmd);
1031                 if (rtn == SUCCESS) {
1032                         if (!scsi_device_online(sdev) ||
1033                             !scsi_eh_tur(bdr_scmd)) {
1034                                 list_for_each_entry_safe(scmd, next,
1035                                                          work_q, eh_entry) {
1036                                         if (scmd->device == sdev)
1037                                                 scsi_eh_finish_cmd(scmd,
1038                                                                    done_q);
1039                                 }
1040                         }
1041                 } else {
1042                         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: BDR"
1043                                                           " failed sdev:"
1044                                                           "0x%p\n",
1045                                                           current->comm,
1046                                                            sdev));
1047                 }
1048         }
1049
1050         return list_empty(work_q);
1051 }
1052
1053 /**
1054  * scsi_eh_bus_reset - send a bus reset 
1055  * @shost:      scsi host being recovered.
1056  * @eh_done_q:  list_head for processed commands.
1057  **/
1058 static int scsi_eh_bus_reset(struct Scsi_Host *shost,
1059                              struct list_head *work_q,
1060                              struct list_head *done_q)
1061 {
1062         struct scsi_cmnd *scmd, *chan_scmd, *next;
1063         unsigned int channel;
1064         int rtn;
1065
1066         /*
1067          * we really want to loop over the various channels, and do this on
1068          * a channel by channel basis.  we should also check to see if any
1069          * of the failed commands are on soft_reset devices, and if so, skip
1070          * the reset.  
1071          */
1072
1073         for (channel = 0; channel <= shost->max_channel; channel++) {
1074                 chan_scmd = NULL;
1075                 list_for_each_entry(scmd, work_q, eh_entry) {
1076                         if (channel == scmd_channel(scmd)) {
1077                                 chan_scmd = scmd;
1078                                 break;
1079                                 /*
1080                                  * FIXME add back in some support for
1081                                  * soft_reset devices.
1082                                  */
1083                         }
1084                 }
1085
1086                 if (!chan_scmd)
1087                         continue;
1088                 SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Sending BRST chan:"
1089                                                   " %d\n", current->comm,
1090                                                   channel));
1091                 rtn = scsi_try_bus_reset(chan_scmd);
1092                 if (rtn == SUCCESS) {
1093                         list_for_each_entry_safe(scmd, next, work_q, eh_entry) {
1094                                 if (channel == scmd_channel(scmd))
1095                                         if (!scsi_device_online(scmd->device) ||
1096                                             !scsi_eh_tur(scmd))
1097                                                 scsi_eh_finish_cmd(scmd,
1098                                                                    done_q);
1099                         }
1100                 } else {
1101                         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: BRST"
1102                                                           " failed chan: %d\n",
1103                                                           current->comm,
1104                                                           channel));
1105                 }
1106         }
1107         return list_empty(work_q);
1108 }
1109
1110 /**
1111  * scsi_eh_host_reset - send a host reset 
1112  * @work_q:     list_head for processed commands.
1113  * @done_q:     list_head for processed commands.
1114  **/
1115 static int scsi_eh_host_reset(struct list_head *work_q,
1116                               struct list_head *done_q)
1117 {
1118         struct scsi_cmnd *scmd, *next;
1119         int rtn;
1120
1121         if (!list_empty(work_q)) {
1122                 scmd = list_entry(work_q->next,
1123                                   struct scsi_cmnd, eh_entry);
1124
1125                 SCSI_LOG_ERROR_RECOVERY(3, printk("%s: Sending HRST\n"
1126                                                   , current->comm));
1127
1128                 rtn = scsi_try_host_reset(scmd);
1129                 if (rtn == SUCCESS) {
1130                         list_for_each_entry_safe(scmd, next, work_q, eh_entry) {
1131                                 if (!scsi_device_online(scmd->device) ||
1132                                     (!scsi_eh_try_stu(scmd) && !scsi_eh_tur(scmd)) ||
1133                                     !scsi_eh_tur(scmd))
1134                                         scsi_eh_finish_cmd(scmd, done_q);
1135                         }
1136                 } else {
1137                         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: HRST"
1138                                                           " failed\n",
1139                                                           current->comm));
1140                 }
1141         }
1142         return list_empty(work_q);
1143 }
1144
1145 /**
1146  * scsi_eh_offline_sdevs - offline scsi devices that fail to recover
1147  * @work_q:     list_head for processed commands.
1148  * @done_q:     list_head for processed commands.
1149  *
1150  **/
1151 static void scsi_eh_offline_sdevs(struct list_head *work_q,
1152                                   struct list_head *done_q)
1153 {
1154         struct scsi_cmnd *scmd, *next;
1155
1156         list_for_each_entry_safe(scmd, next, work_q, eh_entry) {
1157                 sdev_printk(KERN_INFO, scmd->device,
1158                             "scsi: Device offlined - not"
1159                             " ready after error recovery\n");
1160                 scsi_device_set_state(scmd->device, SDEV_OFFLINE);
1161                 if (scmd->eh_eflags & SCSI_EH_CANCEL_CMD) {
1162                         /*
1163                          * FIXME: Handle lost cmds.
1164                          */
1165                 }
1166                 scsi_eh_finish_cmd(scmd, done_q);
1167         }
1168         return;
1169 }
1170
1171 /**
1172  * scsi_decide_disposition - Disposition a cmd on return from LLD.
1173  * @scmd:       SCSI cmd to examine.
1174  *
1175  * Notes:
1176  *    This is *only* called when we are examining the status after sending
1177  *    out the actual data command.  any commands that are queued for error
1178  *    recovery (e.g. test_unit_ready) do *not* come through here.
1179  *
1180  *    When this routine returns failed, it means the error handler thread
1181  *    is woken.  In cases where the error code indicates an error that
1182  *    doesn't require the error handler read (i.e. we don't need to
1183  *    abort/reset), this function should return SUCCESS.
1184  **/
1185 int scsi_decide_disposition(struct scsi_cmnd *scmd)
1186 {
1187         int rtn;
1188
1189         /*
1190          * if the device is offline, then we clearly just pass the result back
1191          * up to the top level.
1192          */
1193         if (!scsi_device_online(scmd->device)) {
1194                 SCSI_LOG_ERROR_RECOVERY(5, printk("%s: device offline - report"
1195                                                   " as SUCCESS\n",
1196                                                   __FUNCTION__));
1197                 return SUCCESS;
1198         }
1199
1200         /*
1201          * first check the host byte, to see if there is anything in there
1202          * that would indicate what we need to do.
1203          */
1204         switch (host_byte(scmd->result)) {
1205         case DID_PASSTHROUGH:
1206                 /*
1207                  * no matter what, pass this through to the upper layer.
1208                  * nuke this special code so that it looks like we are saying
1209                  * did_ok.
1210                  */
1211                 scmd->result &= 0xff00ffff;
1212                 return SUCCESS;
1213         case DID_OK:
1214                 /*
1215                  * looks good.  drop through, and check the next byte.
1216                  */
1217                 break;
1218         case DID_NO_CONNECT:
1219         case DID_BAD_TARGET:
1220         case DID_ABORT:
1221                 /*
1222                  * note - this means that we just report the status back
1223                  * to the top level driver, not that we actually think
1224                  * that it indicates SUCCESS.
1225                  */
1226                 return SUCCESS;
1227                 /*
1228                  * when the low level driver returns did_soft_error,
1229                  * it is responsible for keeping an internal retry counter 
1230                  * in order to avoid endless loops (db)
1231                  *
1232                  * actually this is a bug in this function here.  we should
1233                  * be mindful of the maximum number of retries specified
1234                  * and not get stuck in a loop.
1235                  */
1236         case DID_SOFT_ERROR:
1237                 goto maybe_retry;
1238         case DID_IMM_RETRY:
1239                 return NEEDS_RETRY;
1240
1241         case DID_REQUEUE:
1242                 return ADD_TO_MLQUEUE;
1243
1244         case DID_ERROR:
1245                 if (msg_byte(scmd->result) == COMMAND_COMPLETE &&
1246                     status_byte(scmd->result) == RESERVATION_CONFLICT)
1247                         /*
1248                          * execute reservation conflict processing code
1249                          * lower down
1250                          */
1251                         break;
1252                 /* fallthrough */
1253
1254         case DID_BUS_BUSY:
1255         case DID_PARITY:
1256                 goto maybe_retry;
1257         case DID_TIME_OUT:
1258                 /*
1259                  * when we scan the bus, we get timeout messages for
1260                  * these commands if there is no device available.
1261                  * other hosts report did_no_connect for the same thing.
1262                  */
1263                 if ((scmd->cmnd[0] == TEST_UNIT_READY ||
1264                      scmd->cmnd[0] == INQUIRY)) {
1265                         return SUCCESS;
1266                 } else {
1267                         return FAILED;
1268                 }
1269         case DID_RESET:
1270                 return SUCCESS;
1271         default:
1272                 return FAILED;
1273         }
1274
1275         /*
1276          * next, check the message byte.
1277          */
1278         if (msg_byte(scmd->result) != COMMAND_COMPLETE)
1279                 return FAILED;
1280
1281         /*
1282          * check the status byte to see if this indicates anything special.
1283          */
1284         switch (status_byte(scmd->result)) {
1285         case QUEUE_FULL:
1286                 /*
1287                  * the case of trying to send too many commands to a
1288                  * tagged queueing device.
1289                  */
1290         case BUSY:
1291                 /*
1292                  * device can't talk to us at the moment.  Should only
1293                  * occur (SAM-3) when the task queue is empty, so will cause
1294                  * the empty queue handling to trigger a stall in the
1295                  * device.
1296                  */
1297                 return ADD_TO_MLQUEUE;
1298         case GOOD:
1299         case COMMAND_TERMINATED:
1300         case TASK_ABORTED:
1301                 return SUCCESS;
1302         case CHECK_CONDITION:
1303                 rtn = scsi_check_sense(scmd);
1304                 if (rtn == NEEDS_RETRY)
1305                         goto maybe_retry;
1306                 /* if rtn == FAILED, we have no sense information;
1307                  * returning FAILED will wake the error handler thread
1308                  * to collect the sense and redo the decide
1309                  * disposition */
1310                 return rtn;
1311         case CONDITION_GOOD:
1312         case INTERMEDIATE_GOOD:
1313         case INTERMEDIATE_C_GOOD:
1314         case ACA_ACTIVE:
1315                 /*
1316                  * who knows?  FIXME(eric)
1317                  */
1318                 return SUCCESS;
1319
1320         case RESERVATION_CONFLICT:
1321                 sdev_printk(KERN_INFO, scmd->device,
1322                             "reservation conflict\n");
1323                 return SUCCESS; /* causes immediate i/o error */
1324         default:
1325                 return FAILED;
1326         }
1327         return FAILED;
1328
1329       maybe_retry:
1330
1331         /* we requeue for retry because the error was retryable, and
1332          * the request was not marked fast fail.  Note that above,
1333          * even if the request is marked fast fail, we still requeue
1334          * for queue congestion conditions (QUEUE_FULL or BUSY) */
1335         if ((++scmd->retries) <= scmd->allowed
1336             && !blk_noretry_request(scmd->request)) {
1337                 return NEEDS_RETRY;
1338         } else {
1339                 /*
1340                  * no more retries - report this one back to upper level.
1341                  */
1342                 return SUCCESS;
1343         }
1344 }
1345
1346 /**
1347  * scsi_eh_lock_door - Prevent medium removal for the specified device
1348  * @sdev:       SCSI device to prevent medium removal
1349  *
1350  * Locking:
1351  *      We must be called from process context; scsi_allocate_request()
1352  *      may sleep.
1353  *
1354  * Notes:
1355  *      We queue up an asynchronous "ALLOW MEDIUM REMOVAL" request on the
1356  *      head of the devices request queue, and continue.
1357  *
1358  * Bugs:
1359  *      scsi_allocate_request() may sleep waiting for existing requests to
1360  *      be processed.  However, since we haven't kicked off any request
1361  *      processing for this host, this may deadlock.
1362  *
1363  *      If scsi_allocate_request() fails for what ever reason, we
1364  *      completely forget to lock the door.
1365  **/
1366 static void scsi_eh_lock_door(struct scsi_device *sdev)
1367 {
1368         unsigned char cmnd[MAX_COMMAND_SIZE];
1369
1370         cmnd[0] = ALLOW_MEDIUM_REMOVAL;
1371         cmnd[1] = 0;
1372         cmnd[2] = 0;
1373         cmnd[3] = 0;
1374         cmnd[4] = SCSI_REMOVAL_PREVENT;
1375         cmnd[5] = 0;
1376
1377         scsi_execute_async(sdev, cmnd, 6, DMA_NONE, NULL, 0, 0, 10 * HZ,
1378                            5, NULL, NULL, GFP_KERNEL);
1379 }
1380
1381
1382 /**
1383  * scsi_restart_operations - restart io operations to the specified host.
1384  * @shost:      Host we are restarting.
1385  *
1386  * Notes:
1387  *    When we entered the error handler, we blocked all further i/o to
1388  *    this device.  we need to 'reverse' this process.
1389  **/
1390 static void scsi_restart_operations(struct Scsi_Host *shost)
1391 {
1392         struct scsi_device *sdev;
1393         unsigned long flags;
1394
1395         /*
1396          * If the door was locked, we need to insert a door lock request
1397          * onto the head of the SCSI request queue for the device.  There
1398          * is no point trying to lock the door of an off-line device.
1399          */
1400         shost_for_each_device(sdev, shost) {
1401                 if (scsi_device_online(sdev) && sdev->locked)
1402                         scsi_eh_lock_door(sdev);
1403         }
1404
1405         /*
1406          * next free up anything directly waiting upon the host.  this
1407          * will be requests for character device operations, and also for
1408          * ioctls to queued block devices.
1409          */
1410         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: waking up host to restart\n",
1411                                           __FUNCTION__));
1412
1413         spin_lock_irqsave(shost->host_lock, flags);
1414         if (scsi_host_set_state(shost, SHOST_RUNNING))
1415                 if (scsi_host_set_state(shost, SHOST_CANCEL))
1416                         BUG_ON(scsi_host_set_state(shost, SHOST_DEL));
1417         spin_unlock_irqrestore(shost->host_lock, flags);
1418
1419         wake_up(&shost->host_wait);
1420
1421         /*
1422          * finally we need to re-initiate requests that may be pending.  we will
1423          * have had everything blocked while error handling is taking place, and
1424          * now that error recovery is done, we will need to ensure that these
1425          * requests are started.
1426          */
1427         scsi_run_host_queues(shost);
1428 }
1429
1430 /**
1431  * scsi_eh_ready_devs - check device ready state and recover if not.
1432  * @shost:      host to be recovered.
1433  * @eh_done_q:  list_head for processed commands.
1434  *
1435  **/
1436 void scsi_eh_ready_devs(struct Scsi_Host *shost,
1437                         struct list_head *work_q,
1438                         struct list_head *done_q)
1439 {
1440         if (!scsi_eh_stu(shost, work_q, done_q))
1441                 if (!scsi_eh_bus_device_reset(shost, work_q, done_q))
1442                         if (!scsi_eh_bus_reset(shost, work_q, done_q))
1443                                 if (!scsi_eh_host_reset(work_q, done_q))
1444                                         scsi_eh_offline_sdevs(work_q, done_q);
1445 }
1446 EXPORT_SYMBOL_GPL(scsi_eh_ready_devs);
1447
1448 /**
1449  * scsi_eh_flush_done_q - finish processed commands or retry them.
1450  * @done_q:     list_head of processed commands.
1451  *
1452  **/
1453 void scsi_eh_flush_done_q(struct list_head *done_q)
1454 {
1455         struct scsi_cmnd *scmd, *next;
1456
1457         list_for_each_entry_safe(scmd, next, done_q, eh_entry) {
1458                 list_del_init(&scmd->eh_entry);
1459                 if (scsi_device_online(scmd->device) &&
1460                     !blk_noretry_request(scmd->request) &&
1461                     (++scmd->retries <= scmd->allowed)) {
1462                         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: flush"
1463                                                           " retry cmd: %p\n",
1464                                                           current->comm,
1465                                                           scmd));
1466                                 scsi_queue_insert(scmd, SCSI_MLQUEUE_EH_RETRY);
1467                 } else {
1468                         /*
1469                          * If just we got sense for the device (called
1470                          * scsi_eh_get_sense), scmd->result is already
1471                          * set, do not set DRIVER_TIMEOUT.
1472                          */
1473                         if (!scmd->result)
1474                                 scmd->result |= (DRIVER_TIMEOUT << 24);
1475                         SCSI_LOG_ERROR_RECOVERY(3, printk("%s: flush finish"
1476                                                         " cmd: %p\n",
1477                                                         current->comm, scmd));
1478                         scsi_finish_command(scmd);
1479                 }
1480         }
1481 }
1482 EXPORT_SYMBOL(scsi_eh_flush_done_q);
1483
1484 /**
1485  * scsi_unjam_host - Attempt to fix a host which has a cmd that failed.
1486  * @shost:      Host to unjam.
1487  *
1488  * Notes:
1489  *    When we come in here, we *know* that all commands on the bus have
1490  *    either completed, failed or timed out.  we also know that no further
1491  *    commands are being sent to the host, so things are relatively quiet
1492  *    and we have freedom to fiddle with things as we wish.
1493  *
1494  *    This is only the *default* implementation.  it is possible for
1495  *    individual drivers to supply their own version of this function, and
1496  *    if the maintainer wishes to do this, it is strongly suggested that
1497  *    this function be taken as a template and modified.  this function
1498  *    was designed to correctly handle problems for about 95% of the
1499  *    different cases out there, and it should always provide at least a
1500  *    reasonable amount of error recovery.
1501  *
1502  *    Any command marked 'failed' or 'timeout' must eventually have
1503  *    scsi_finish_cmd() called for it.  we do all of the retry stuff
1504  *    here, so when we restart the host after we return it should have an
1505  *    empty queue.
1506  **/
1507 static void scsi_unjam_host(struct Scsi_Host *shost)
1508 {
1509         unsigned long flags;
1510         LIST_HEAD(eh_work_q);
1511         LIST_HEAD(eh_done_q);
1512
1513         spin_lock_irqsave(shost->host_lock, flags);
1514         list_splice_init(&shost->eh_cmd_q, &eh_work_q);
1515         spin_unlock_irqrestore(shost->host_lock, flags);
1516
1517         SCSI_LOG_ERROR_RECOVERY(1, scsi_eh_prt_fail_stats(shost, &eh_work_q));
1518
1519         if (!scsi_eh_get_sense(&eh_work_q, &eh_done_q))
1520                 if (!scsi_eh_abort_cmds(&eh_work_q, &eh_done_q))
1521                         scsi_eh_ready_devs(shost, &eh_work_q, &eh_done_q);
1522
1523         scsi_eh_flush_done_q(&eh_done_q);
1524 }
1525
1526 /**
1527  * scsi_error_handler - SCSI error handler thread
1528  * @data:       Host for which we are running.
1529  *
1530  * Notes:
1531  *    This is the main error handling loop.  This is run as a kernel thread
1532  *    for every SCSI host and handles all error handling activity.
1533  **/
1534 int scsi_error_handler(void *data)
1535 {
1536         struct Scsi_Host *shost = data;
1537
1538         current->flags |= PF_NOFREEZE;
1539
1540         /*
1541          * We use TASK_INTERRUPTIBLE so that the thread is not
1542          * counted against the load average as a running process.
1543          * We never actually get interrupted because kthread_run
1544          * disables singal delivery for the created thread.
1545          */
1546         set_current_state(TASK_INTERRUPTIBLE);
1547         while (!kthread_should_stop()) {
1548                 if ((shost->host_failed == 0 && shost->host_eh_scheduled == 0) ||
1549                     shost->host_failed != shost->host_busy) {
1550                         SCSI_LOG_ERROR_RECOVERY(1,
1551                                 printk("Error handler scsi_eh_%d sleeping\n",
1552                                         shost->host_no));
1553                         schedule();
1554                         set_current_state(TASK_INTERRUPTIBLE);
1555                         continue;
1556                 }
1557
1558                 __set_current_state(TASK_RUNNING);
1559                 SCSI_LOG_ERROR_RECOVERY(1,
1560                         printk("Error handler scsi_eh_%d waking up\n",
1561                                 shost->host_no));
1562
1563                 /*
1564                  * We have a host that is failing for some reason.  Figure out
1565                  * what we need to do to get it up and online again (if we can).
1566                  * If we fail, we end up taking the thing offline.
1567                  */
1568                 if (shost->transportt->eh_strategy_handler)
1569                         shost->transportt->eh_strategy_handler(shost);
1570                 else
1571                         scsi_unjam_host(shost);
1572
1573                 /*
1574                  * Note - if the above fails completely, the action is to take
1575                  * individual devices offline and flush the queue of any
1576                  * outstanding requests that may have been pending.  When we
1577                  * restart, we restart any I/O to any other devices on the bus
1578                  * which are still online.
1579                  */
1580                 scsi_restart_operations(shost);
1581                 set_current_state(TASK_INTERRUPTIBLE);
1582         }
1583         __set_current_state(TASK_RUNNING);
1584
1585         SCSI_LOG_ERROR_RECOVERY(1,
1586                 printk("Error handler scsi_eh_%d exiting\n", shost->host_no));
1587         shost->ehandler = NULL;
1588         return 0;
1589 }
1590
1591 /*
1592  * Function:    scsi_report_bus_reset()
1593  *
1594  * Purpose:     Utility function used by low-level drivers to report that
1595  *              they have observed a bus reset on the bus being handled.
1596  *
1597  * Arguments:   shost       - Host in question
1598  *              channel     - channel on which reset was observed.
1599  *
1600  * Returns:     Nothing
1601  *
1602  * Lock status: Host lock must be held.
1603  *
1604  * Notes:       This only needs to be called if the reset is one which
1605  *              originates from an unknown location.  Resets originated
1606  *              by the mid-level itself don't need to call this, but there
1607  *              should be no harm.
1608  *
1609  *              The main purpose of this is to make sure that a CHECK_CONDITION
1610  *              is properly treated.
1611  */
1612 void scsi_report_bus_reset(struct Scsi_Host *shost, int channel)
1613 {
1614         struct scsi_device *sdev;
1615
1616         __shost_for_each_device(sdev, shost) {
1617                 if (channel == sdev_channel(sdev)) {
1618                         sdev->was_reset = 1;
1619                         sdev->expecting_cc_ua = 1;
1620                 }
1621         }
1622 }
1623 EXPORT_SYMBOL(scsi_report_bus_reset);
1624
1625 /*
1626  * Function:    scsi_report_device_reset()
1627  *
1628  * Purpose:     Utility function used by low-level drivers to report that
1629  *              they have observed a device reset on the device being handled.
1630  *
1631  * Arguments:   shost       - Host in question
1632  *              channel     - channel on which reset was observed
1633  *              target      - target on which reset was observed
1634  *
1635  * Returns:     Nothing
1636  *
1637  * Lock status: Host lock must be held
1638  *
1639  * Notes:       This only needs to be called if the reset is one which
1640  *              originates from an unknown location.  Resets originated
1641  *              by the mid-level itself don't need to call this, but there
1642  *              should be no harm.
1643  *
1644  *              The main purpose of this is to make sure that a CHECK_CONDITION
1645  *              is properly treated.
1646  */
1647 void scsi_report_device_reset(struct Scsi_Host *shost, int channel, int target)
1648 {
1649         struct scsi_device *sdev;
1650
1651         __shost_for_each_device(sdev, shost) {
1652                 if (channel == sdev_channel(sdev) &&
1653                     target == sdev_id(sdev)) {
1654                         sdev->was_reset = 1;
1655                         sdev->expecting_cc_ua = 1;
1656                 }
1657         }
1658 }
1659 EXPORT_SYMBOL(scsi_report_device_reset);
1660
1661 static void
1662 scsi_reset_provider_done_command(struct scsi_cmnd *scmd)
1663 {
1664 }
1665
1666 /*
1667  * Function:    scsi_reset_provider
1668  *
1669  * Purpose:     Send requested reset to a bus or device at any phase.
1670  *
1671  * Arguments:   device  - device to send reset to
1672  *              flag - reset type (see scsi.h)
1673  *
1674  * Returns:     SUCCESS/FAILURE.
1675  *
1676  * Notes:       This is used by the SCSI Generic driver to provide
1677  *              Bus/Device reset capability.
1678  */
1679 int
1680 scsi_reset_provider(struct scsi_device *dev, int flag)
1681 {
1682         struct scsi_cmnd *scmd = scsi_get_command(dev, GFP_KERNEL);
1683         struct Scsi_Host *shost = dev->host;
1684         struct request req;
1685         unsigned long flags;
1686         int rtn;
1687
1688         scmd->request = &req;
1689         memset(&scmd->eh_timeout, 0, sizeof(scmd->eh_timeout));
1690
1691         memset(&scmd->cmnd, '\0', sizeof(scmd->cmnd));
1692     
1693         scmd->scsi_done         = scsi_reset_provider_done_command;
1694         scmd->done                      = NULL;
1695         scmd->request_buffer            = NULL;
1696         scmd->request_bufflen           = 0;
1697
1698         scmd->cmd_len                   = 0;
1699
1700         scmd->sc_data_direction         = DMA_BIDIRECTIONAL;
1701
1702         init_timer(&scmd->eh_timeout);
1703
1704         /*
1705          * Sometimes the command can get back into the timer chain,
1706          * so use the pid as an identifier.
1707          */
1708         scmd->pid                       = 0;
1709
1710         spin_lock_irqsave(shost->host_lock, flags);
1711         shost->tmf_in_progress = 1;
1712         spin_unlock_irqrestore(shost->host_lock, flags);
1713
1714         switch (flag) {
1715         case SCSI_TRY_RESET_DEVICE:
1716                 rtn = scsi_try_bus_device_reset(scmd);
1717                 if (rtn == SUCCESS)
1718                         break;
1719                 /* FALLTHROUGH */
1720         case SCSI_TRY_RESET_BUS:
1721                 rtn = scsi_try_bus_reset(scmd);
1722                 if (rtn == SUCCESS)
1723                         break;
1724                 /* FALLTHROUGH */
1725         case SCSI_TRY_RESET_HOST:
1726                 rtn = scsi_try_host_reset(scmd);
1727                 break;
1728         default:
1729                 rtn = FAILED;
1730         }
1731
1732         spin_lock_irqsave(shost->host_lock, flags);
1733         shost->tmf_in_progress = 0;
1734         spin_unlock_irqrestore(shost->host_lock, flags);
1735
1736         /*
1737          * be sure to wake up anyone who was sleeping or had their queue
1738          * suspended while we performed the TMF.
1739          */
1740         SCSI_LOG_ERROR_RECOVERY(3,
1741                 printk("%s: waking up host to restart after TMF\n",
1742                 __FUNCTION__));
1743
1744         wake_up(&shost->host_wait);
1745
1746         scsi_run_host_queues(shost);
1747
1748         scsi_next_command(scmd);
1749         return rtn;
1750 }
1751 EXPORT_SYMBOL(scsi_reset_provider);
1752
1753 /**
1754  * scsi_normalize_sense - normalize main elements from either fixed or
1755  *                      descriptor sense data format into a common format.
1756  *
1757  * @sense_buffer:       byte array containing sense data returned by device
1758  * @sb_len:             number of valid bytes in sense_buffer
1759  * @sshdr:              pointer to instance of structure that common
1760  *                      elements are written to.
1761  *
1762  * Notes:
1763  *      The "main elements" from sense data are: response_code, sense_key,
1764  *      asc, ascq and additional_length (only for descriptor format).
1765  *
1766  *      Typically this function can be called after a device has
1767  *      responded to a SCSI command with the CHECK_CONDITION status.
1768  *
1769  * Return value:
1770  *      1 if valid sense data information found, else 0;
1771  **/
1772 int scsi_normalize_sense(const u8 *sense_buffer, int sb_len,
1773                          struct scsi_sense_hdr *sshdr)
1774 {
1775         if (!sense_buffer || !sb_len)
1776                 return 0;
1777
1778         memset(sshdr, 0, sizeof(struct scsi_sense_hdr));
1779
1780         sshdr->response_code = (sense_buffer[0] & 0x7f);
1781
1782         if (!scsi_sense_valid(sshdr))
1783                 return 0;
1784
1785         if (sshdr->response_code >= 0x72) {
1786                 /*
1787                  * descriptor format
1788                  */
1789                 if (sb_len > 1)
1790                         sshdr->sense_key = (sense_buffer[1] & 0xf);
1791                 if (sb_len > 2)
1792                         sshdr->asc = sense_buffer[2];
1793                 if (sb_len > 3)
1794                         sshdr->ascq = sense_buffer[3];
1795                 if (sb_len > 7)
1796                         sshdr->additional_length = sense_buffer[7];
1797         } else {
1798                 /* 
1799                  * fixed format
1800                  */
1801                 if (sb_len > 2)
1802                         sshdr->sense_key = (sense_buffer[2] & 0xf);
1803                 if (sb_len > 7) {
1804                         sb_len = (sb_len < (sense_buffer[7] + 8)) ?
1805                                          sb_len : (sense_buffer[7] + 8);
1806                         if (sb_len > 12)
1807                                 sshdr->asc = sense_buffer[12];
1808                         if (sb_len > 13)
1809                                 sshdr->ascq = sense_buffer[13];
1810                 }
1811         }
1812
1813         return 1;
1814 }
1815 EXPORT_SYMBOL(scsi_normalize_sense);
1816
1817 int scsi_command_normalize_sense(struct scsi_cmnd *cmd,
1818                                  struct scsi_sense_hdr *sshdr)
1819 {
1820         return scsi_normalize_sense(cmd->sense_buffer,
1821                         sizeof(cmd->sense_buffer), sshdr);
1822 }
1823 EXPORT_SYMBOL(scsi_command_normalize_sense);
1824
1825 /**
1826  * scsi_sense_desc_find - search for a given descriptor type in
1827  *                      descriptor sense data format.
1828  *
1829  * @sense_buffer:       byte array of descriptor format sense data
1830  * @sb_len:             number of valid bytes in sense_buffer
1831  * @desc_type:          value of descriptor type to find
1832  *                      (e.g. 0 -> information)
1833  *
1834  * Notes:
1835  *      only valid when sense data is in descriptor format
1836  *
1837  * Return value:
1838  *      pointer to start of (first) descriptor if found else NULL
1839  **/
1840 const u8 * scsi_sense_desc_find(const u8 * sense_buffer, int sb_len,
1841                                 int desc_type)
1842 {
1843         int add_sen_len, add_len, desc_len, k;
1844         const u8 * descp;
1845
1846         if ((sb_len < 8) || (0 == (add_sen_len = sense_buffer[7])))
1847                 return NULL;
1848         if ((sense_buffer[0] < 0x72) || (sense_buffer[0] > 0x73))
1849                 return NULL;
1850         add_sen_len = (add_sen_len < (sb_len - 8)) ?
1851                         add_sen_len : (sb_len - 8);
1852         descp = &sense_buffer[8];
1853         for (desc_len = 0, k = 0; k < add_sen_len; k += desc_len) {
1854                 descp += desc_len;
1855                 add_len = (k < (add_sen_len - 1)) ? descp[1]: -1;
1856                 desc_len = add_len + 2;
1857                 if (descp[0] == desc_type)
1858                         return descp;
1859                 if (add_len < 0) // short descriptor ??
1860                         break;
1861         }
1862         return NULL;
1863 }
1864 EXPORT_SYMBOL(scsi_sense_desc_find);
1865
1866 /**
1867  * scsi_get_sense_info_fld - attempts to get information field from
1868  *                      sense data (either fixed or descriptor format)
1869  *
1870  * @sense_buffer:       byte array of sense data
1871  * @sb_len:             number of valid bytes in sense_buffer
1872  * @info_out:           pointer to 64 integer where 8 or 4 byte information
1873  *                      field will be placed if found.
1874  *
1875  * Return value:
1876  *      1 if information field found, 0 if not found.
1877  **/
1878 int scsi_get_sense_info_fld(const u8 * sense_buffer, int sb_len,
1879                             u64 * info_out)
1880 {
1881         int j;
1882         const u8 * ucp;
1883         u64 ull;
1884
1885         if (sb_len < 7)
1886                 return 0;
1887         switch (sense_buffer[0] & 0x7f) {
1888         case 0x70:
1889         case 0x71:
1890                 if (sense_buffer[0] & 0x80) {
1891                         *info_out = (sense_buffer[3] << 24) +
1892                                     (sense_buffer[4] << 16) +
1893                                     (sense_buffer[5] << 8) + sense_buffer[6];
1894                         return 1;
1895                 } else
1896                         return 0;
1897         case 0x72:
1898         case 0x73:
1899                 ucp = scsi_sense_desc_find(sense_buffer, sb_len,
1900                                            0 /* info desc */);
1901                 if (ucp && (0xa == ucp[1])) {
1902                         ull = 0;
1903                         for (j = 0; j < 8; ++j) {
1904                                 if (j > 0)
1905                                         ull <<= 8;
1906                                 ull |= ucp[4 + j];
1907                         }
1908                         *info_out = ull;
1909                         return 1;
1910                 } else
1911                         return 0;
1912         default:
1913                 return 0;
1914         }
1915 }
1916 EXPORT_SYMBOL(scsi_get_sense_info_fld);