GNU Linux-libre 4.14.266-gnu1
[releases.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/gfp.h>
20 #include <linux/timer.h>
21 #include <linux/string.h>
22 #include <linux/kernel.h>
23 #include <linux/freezer.h>
24 #include <linux/kthread.h>
25 #include <linux/interrupt.h>
26 #include <linux/blkdev.h>
27 #include <linux/delay.h>
28 #include <linux/jiffies.h>
29
30 #include <scsi/scsi.h>
31 #include <scsi/scsi_cmnd.h>
32 #include <scsi/scsi_dbg.h>
33 #include <scsi/scsi_device.h>
34 #include <scsi/scsi_driver.h>
35 #include <scsi/scsi_eh.h>
36 #include <scsi/scsi_common.h>
37 #include <scsi/scsi_transport.h>
38 #include <scsi/scsi_host.h>
39 #include <scsi/scsi_ioctl.h>
40 #include <scsi/scsi_dh.h>
41 #include <scsi/sg.h>
42
43 #include "scsi_priv.h"
44 #include "scsi_logging.h"
45 #include "scsi_transport_api.h"
46
47 #include <trace/events/scsi.h>
48
49 #include <asm/unaligned.h>
50
51 static void scsi_eh_done(struct scsi_cmnd *scmd);
52
53 /*
54  * These should *probably* be handled by the host itself.
55  * Since it is allowed to sleep, it probably should.
56  */
57 #define BUS_RESET_SETTLE_TIME   (10)
58 #define HOST_RESET_SETTLE_TIME  (10)
59
60 static int scsi_eh_try_stu(struct scsi_cmnd *scmd);
61 static int scsi_try_to_abort_cmd(struct scsi_host_template *,
62                                  struct scsi_cmnd *);
63
64 /* called with shost->host_lock held */
65 void scsi_eh_wakeup(struct Scsi_Host *shost)
66 {
67         if (atomic_read(&shost->host_busy) == shost->host_failed) {
68                 trace_scsi_eh_wakeup(shost);
69                 wake_up_process(shost->ehandler);
70                 SCSI_LOG_ERROR_RECOVERY(5, shost_printk(KERN_INFO, shost,
71                         "Waking error handler thread\n"));
72         }
73 }
74
75 /**
76  * scsi_schedule_eh - schedule EH for SCSI host
77  * @shost:      SCSI host to invoke error handling on.
78  *
79  * Schedule SCSI EH without scmd.
80  */
81 void scsi_schedule_eh(struct Scsi_Host *shost)
82 {
83         unsigned long flags;
84
85         spin_lock_irqsave(shost->host_lock, flags);
86
87         if (scsi_host_set_state(shost, SHOST_RECOVERY) == 0 ||
88             scsi_host_set_state(shost, SHOST_CANCEL_RECOVERY) == 0) {
89                 shost->host_eh_scheduled++;
90                 scsi_eh_wakeup(shost);
91         }
92
93         spin_unlock_irqrestore(shost->host_lock, flags);
94 }
95 EXPORT_SYMBOL_GPL(scsi_schedule_eh);
96
97 static int scsi_host_eh_past_deadline(struct Scsi_Host *shost)
98 {
99         if (!shost->last_reset || shost->eh_deadline == -1)
100                 return 0;
101
102         /*
103          * 32bit accesses are guaranteed to be atomic
104          * (on all supported architectures), so instead
105          * of using a spinlock we can as well double check
106          * if eh_deadline has been set to 'off' during the
107          * time_before call.
108          */
109         if (time_before(jiffies, shost->last_reset + shost->eh_deadline) &&
110             shost->eh_deadline > -1)
111                 return 0;
112
113         return 1;
114 }
115
116 /**
117  * scmd_eh_abort_handler - Handle command aborts
118  * @work:       command to be aborted.
119  */
120 void
121 scmd_eh_abort_handler(struct work_struct *work)
122 {
123         struct scsi_cmnd *scmd =
124                 container_of(work, struct scsi_cmnd, abort_work.work);
125         struct scsi_device *sdev = scmd->device;
126         int rtn;
127
128         if (scsi_host_eh_past_deadline(sdev->host)) {
129                 SCSI_LOG_ERROR_RECOVERY(3,
130                         scmd_printk(KERN_INFO, scmd,
131                                     "eh timeout, not aborting\n"));
132         } else {
133                 SCSI_LOG_ERROR_RECOVERY(3,
134                         scmd_printk(KERN_INFO, scmd,
135                                     "aborting command\n"));
136                 rtn = scsi_try_to_abort_cmd(sdev->host->hostt, scmd);
137                 if (rtn == SUCCESS) {
138                         set_host_byte(scmd, DID_TIME_OUT);
139                         if (scsi_host_eh_past_deadline(sdev->host)) {
140                                 SCSI_LOG_ERROR_RECOVERY(3,
141                                         scmd_printk(KERN_INFO, scmd,
142                                                     "eh timeout, not retrying "
143                                                     "aborted command\n"));
144                         } else if (!scsi_noretry_cmd(scmd) &&
145                             (++scmd->retries <= scmd->allowed)) {
146                                 SCSI_LOG_ERROR_RECOVERY(3,
147                                         scmd_printk(KERN_WARNING, scmd,
148                                                     "retry aborted command\n"));
149                                 scsi_queue_insert(scmd, SCSI_MLQUEUE_EH_RETRY);
150                                 return;
151                         } else {
152                                 SCSI_LOG_ERROR_RECOVERY(3,
153                                         scmd_printk(KERN_WARNING, scmd,
154                                                     "finish aborted command\n"));
155                                 scsi_finish_command(scmd);
156                                 return;
157                         }
158                 } else {
159                         SCSI_LOG_ERROR_RECOVERY(3,
160                                 scmd_printk(KERN_INFO, scmd,
161                                             "cmd abort %s\n",
162                                             (rtn == FAST_IO_FAIL) ?
163                                             "not send" : "failed"));
164                 }
165         }
166
167         scsi_eh_scmd_add(scmd);
168 }
169
170 /**
171  * scsi_abort_command - schedule a command abort
172  * @scmd:       scmd to abort.
173  *
174  * We only need to abort commands after a command timeout
175  */
176 static int
177 scsi_abort_command(struct scsi_cmnd *scmd)
178 {
179         struct scsi_device *sdev = scmd->device;
180         struct Scsi_Host *shost = sdev->host;
181         unsigned long flags;
182
183         if (scmd->eh_eflags & SCSI_EH_ABORT_SCHEDULED) {
184                 /*
185                  * Retry after abort failed, escalate to next level.
186                  */
187                 SCSI_LOG_ERROR_RECOVERY(3,
188                         scmd_printk(KERN_INFO, scmd,
189                                     "previous abort failed\n"));
190                 BUG_ON(delayed_work_pending(&scmd->abort_work));
191                 return FAILED;
192         }
193
194         spin_lock_irqsave(shost->host_lock, flags);
195         if (shost->eh_deadline != -1 && !shost->last_reset)
196                 shost->last_reset = jiffies;
197         spin_unlock_irqrestore(shost->host_lock, flags);
198
199         scmd->eh_eflags |= SCSI_EH_ABORT_SCHEDULED;
200         SCSI_LOG_ERROR_RECOVERY(3,
201                 scmd_printk(KERN_INFO, scmd, "abort scheduled\n"));
202         queue_delayed_work(shost->tmf_work_q, &scmd->abort_work, HZ / 100);
203         return SUCCESS;
204 }
205
206 /**
207  * scsi_eh_reset - call into ->eh_action to reset internal counters
208  * @scmd:       scmd to run eh on.
209  *
210  * The scsi driver might be carrying internal state about the
211  * devices, so we need to call into the driver to reset the
212  * internal state once the error handler is started.
213  */
214 static void scsi_eh_reset(struct scsi_cmnd *scmd)
215 {
216         if (!blk_rq_is_passthrough(scmd->request)) {
217                 struct scsi_driver *sdrv = scsi_cmd_to_driver(scmd);
218                 if (sdrv->eh_reset)
219                         sdrv->eh_reset(scmd);
220         }
221 }
222
223 static void scsi_eh_inc_host_failed(struct rcu_head *head)
224 {
225         struct scsi_cmnd *scmd = container_of(head, typeof(*scmd), rcu);
226         struct Scsi_Host *shost = scmd->device->host;
227         unsigned long flags;
228
229         spin_lock_irqsave(shost->host_lock, flags);
230         shost->host_failed++;
231         scsi_eh_wakeup(shost);
232         spin_unlock_irqrestore(shost->host_lock, flags);
233 }
234
235 /**
236  * scsi_eh_scmd_add - add scsi cmd to error handling.
237  * @scmd:       scmd to run eh on.
238  */
239 void scsi_eh_scmd_add(struct scsi_cmnd *scmd)
240 {
241         struct Scsi_Host *shost = scmd->device->host;
242         unsigned long flags;
243         int ret;
244
245         WARN_ON_ONCE(!shost->ehandler);
246
247         spin_lock_irqsave(shost->host_lock, flags);
248         if (scsi_host_set_state(shost, SHOST_RECOVERY)) {
249                 ret = scsi_host_set_state(shost, SHOST_CANCEL_RECOVERY);
250                 WARN_ON_ONCE(ret);
251         }
252         if (shost->eh_deadline != -1 && !shost->last_reset)
253                 shost->last_reset = jiffies;
254
255         scsi_eh_reset(scmd);
256         list_add_tail(&scmd->eh_entry, &shost->eh_cmd_q);
257         spin_unlock_irqrestore(shost->host_lock, flags);
258         /*
259          * Ensure that all tasks observe the host state change before the
260          * host_failed change.
261          */
262         call_rcu(&scmd->rcu, scsi_eh_inc_host_failed);
263 }
264
265 /**
266  * scsi_times_out - Timeout function for normal scsi commands.
267  * @req:        request that is timing out.
268  *
269  * Notes:
270  *     We do not need to lock this.  There is the potential for a race
271  *     only in that the normal completion handling might run, but if the
272  *     normal completion function determines that the timer has already
273  *     fired, then it mustn't do anything.
274  */
275 enum blk_eh_timer_return scsi_times_out(struct request *req)
276 {
277         struct scsi_cmnd *scmd = blk_mq_rq_to_pdu(req);
278         enum blk_eh_timer_return rtn = BLK_EH_NOT_HANDLED;
279         struct Scsi_Host *host = scmd->device->host;
280
281         trace_scsi_dispatch_cmd_timeout(scmd);
282         scsi_log_completion(scmd, TIMEOUT_ERROR);
283
284         if (host->eh_deadline != -1 && !host->last_reset)
285                 host->last_reset = jiffies;
286
287         if (host->hostt->eh_timed_out)
288                 rtn = host->hostt->eh_timed_out(scmd);
289
290         if (rtn == BLK_EH_NOT_HANDLED) {
291                 if (scsi_abort_command(scmd) != SUCCESS) {
292                         set_host_byte(scmd, DID_TIME_OUT);
293                         scsi_eh_scmd_add(scmd);
294                 }
295         }
296
297         return rtn;
298 }
299
300 /**
301  * scsi_block_when_processing_errors - Prevent cmds from being queued.
302  * @sdev:       Device on which we are performing recovery.
303  *
304  * Description:
305  *     We block until the host is out of error recovery, and then check to
306  *     see whether the host or the device is offline.
307  *
308  * Return value:
309  *     0 when dev was taken offline by error recovery. 1 OK to proceed.
310  */
311 int scsi_block_when_processing_errors(struct scsi_device *sdev)
312 {
313         int online;
314
315         wait_event(sdev->host->host_wait, !scsi_host_in_recovery(sdev->host));
316
317         online = scsi_device_online(sdev);
318
319         SCSI_LOG_ERROR_RECOVERY(5, sdev_printk(KERN_INFO, sdev,
320                 "%s: rtn: %d\n", __func__, online));
321
322         return online;
323 }
324 EXPORT_SYMBOL(scsi_block_when_processing_errors);
325
326 #ifdef CONFIG_SCSI_LOGGING
327 /**
328  * scsi_eh_prt_fail_stats - Log info on failures.
329  * @shost:      scsi host being recovered.
330  * @work_q:     Queue of scsi cmds to process.
331  */
332 static inline void scsi_eh_prt_fail_stats(struct Scsi_Host *shost,
333                                           struct list_head *work_q)
334 {
335         struct scsi_cmnd *scmd;
336         struct scsi_device *sdev;
337         int total_failures = 0;
338         int cmd_failed = 0;
339         int cmd_cancel = 0;
340         int devices_failed = 0;
341
342         shost_for_each_device(sdev, shost) {
343                 list_for_each_entry(scmd, work_q, eh_entry) {
344                         if (scmd->device == sdev) {
345                                 ++total_failures;
346                                 if (scmd->eh_eflags & SCSI_EH_ABORT_SCHEDULED)
347                                         ++cmd_cancel;
348                                 else
349                                         ++cmd_failed;
350                         }
351                 }
352
353                 if (cmd_cancel || cmd_failed) {
354                         SCSI_LOG_ERROR_RECOVERY(3,
355                                 shost_printk(KERN_INFO, shost,
356                                             "%s: cmds failed: %d, cancel: %d\n",
357                                             __func__, cmd_failed,
358                                             cmd_cancel));
359                         cmd_cancel = 0;
360                         cmd_failed = 0;
361                         ++devices_failed;
362                 }
363         }
364
365         SCSI_LOG_ERROR_RECOVERY(2, shost_printk(KERN_INFO, shost,
366                                    "Total of %d commands on %d"
367                                    " devices require eh work\n",
368                                    total_failures, devices_failed));
369 }
370 #endif
371
372  /**
373  * scsi_report_lun_change - Set flag on all *other* devices on the same target
374  *                          to indicate that a UNIT ATTENTION is expected.
375  * @sdev:       Device reporting the UNIT ATTENTION
376  */
377 static void scsi_report_lun_change(struct scsi_device *sdev)
378 {
379         sdev->sdev_target->expecting_lun_change = 1;
380 }
381
382 /**
383  * scsi_report_sense - Examine scsi sense information and log messages for
384  *                     certain conditions, also issue uevents for some of them.
385  * @sdev:       Device reporting the sense code
386  * @sshdr:      sshdr to be examined
387  */
388 static void scsi_report_sense(struct scsi_device *sdev,
389                               struct scsi_sense_hdr *sshdr)
390 {
391         enum scsi_device_event evt_type = SDEV_EVT_MAXBITS;     /* i.e. none */
392
393         if (sshdr->sense_key == UNIT_ATTENTION) {
394                 if (sshdr->asc == 0x3f && sshdr->ascq == 0x03) {
395                         evt_type = SDEV_EVT_INQUIRY_CHANGE_REPORTED;
396                         sdev_printk(KERN_WARNING, sdev,
397                                     "Inquiry data has changed");
398                 } else if (sshdr->asc == 0x3f && sshdr->ascq == 0x0e) {
399                         evt_type = SDEV_EVT_LUN_CHANGE_REPORTED;
400                         scsi_report_lun_change(sdev);
401                         sdev_printk(KERN_WARNING, sdev,
402                                     "Warning! Received an indication that the "
403                                     "LUN assignments on this target have "
404                                     "changed. The Linux SCSI layer does not "
405                                     "automatically remap LUN assignments.\n");
406                 } else if (sshdr->asc == 0x3f)
407                         sdev_printk(KERN_WARNING, sdev,
408                                     "Warning! Received an indication that the "
409                                     "operating parameters on this target have "
410                                     "changed. The Linux SCSI layer does not "
411                                     "automatically adjust these parameters.\n");
412
413                 if (sshdr->asc == 0x38 && sshdr->ascq == 0x07) {
414                         evt_type = SDEV_EVT_SOFT_THRESHOLD_REACHED_REPORTED;
415                         sdev_printk(KERN_WARNING, sdev,
416                                     "Warning! Received an indication that the "
417                                     "LUN reached a thin provisioning soft "
418                                     "threshold.\n");
419                 }
420
421                 if (sshdr->asc == 0x2a && sshdr->ascq == 0x01) {
422                         evt_type = SDEV_EVT_MODE_PARAMETER_CHANGE_REPORTED;
423                         sdev_printk(KERN_WARNING, sdev,
424                                     "Mode parameters changed");
425                 } else if (sshdr->asc == 0x2a && sshdr->ascq == 0x06) {
426                         evt_type = SDEV_EVT_ALUA_STATE_CHANGE_REPORTED;
427                         sdev_printk(KERN_WARNING, sdev,
428                                     "Asymmetric access state changed");
429                 } else if (sshdr->asc == 0x2a && sshdr->ascq == 0x09) {
430                         evt_type = SDEV_EVT_CAPACITY_CHANGE_REPORTED;
431                         sdev_printk(KERN_WARNING, sdev,
432                                     "Capacity data has changed");
433                 } else if (sshdr->asc == 0x2a)
434                         sdev_printk(KERN_WARNING, sdev,
435                                     "Parameters changed");
436         }
437
438         if (evt_type != SDEV_EVT_MAXBITS) {
439                 set_bit(evt_type, sdev->pending_events);
440                 schedule_work(&sdev->event_work);
441         }
442 }
443
444 /**
445  * scsi_check_sense - Examine scsi cmd sense
446  * @scmd:       Cmd to have sense checked.
447  *
448  * Return value:
449  *      SUCCESS or FAILED or NEEDS_RETRY or ADD_TO_MLQUEUE
450  *
451  * Notes:
452  *      When a deferred error is detected the current command has
453  *      not been executed and needs retrying.
454  */
455 int scsi_check_sense(struct scsi_cmnd *scmd)
456 {
457         struct scsi_device *sdev = scmd->device;
458         struct scsi_sense_hdr sshdr;
459
460         if (! scsi_command_normalize_sense(scmd, &sshdr))
461                 return FAILED;  /* no valid sense data */
462
463         scsi_report_sense(sdev, &sshdr);
464
465         if (scsi_sense_is_deferred(&sshdr))
466                 return NEEDS_RETRY;
467
468         if (sdev->handler && sdev->handler->check_sense) {
469                 int rc;
470
471                 rc = sdev->handler->check_sense(sdev, &sshdr);
472                 if (rc != SCSI_RETURN_NOT_HANDLED)
473                         return rc;
474                 /* handler does not care. Drop down to default handling */
475         }
476
477         if (scmd->cmnd[0] == TEST_UNIT_READY && scmd->scsi_done != scsi_eh_done)
478                 /*
479                  * nasty: for mid-layer issued TURs, we need to return the
480                  * actual sense data without any recovery attempt.  For eh
481                  * issued ones, we need to try to recover and interpret
482                  */
483                 return SUCCESS;
484
485         /*
486          * Previous logic looked for FILEMARK, EOM or ILI which are
487          * mainly associated with tapes and returned SUCCESS.
488          */
489         if (sshdr.response_code == 0x70) {
490                 /* fixed format */
491                 if (scmd->sense_buffer[2] & 0xe0)
492                         return SUCCESS;
493         } else {
494                 /*
495                  * descriptor format: look for "stream commands sense data
496                  * descriptor" (see SSC-3). Assume single sense data
497                  * descriptor. Ignore ILI from SBC-2 READ LONG and WRITE LONG.
498                  */
499                 if ((sshdr.additional_length > 3) &&
500                     (scmd->sense_buffer[8] == 0x4) &&
501                     (scmd->sense_buffer[11] & 0xe0))
502                         return SUCCESS;
503         }
504
505         switch (sshdr.sense_key) {
506         case NO_SENSE:
507                 return SUCCESS;
508         case RECOVERED_ERROR:
509                 return /* soft_error */ SUCCESS;
510
511         case ABORTED_COMMAND:
512                 if (sshdr.asc == 0x10) /* DIF */
513                         return SUCCESS;
514
515                 return NEEDS_RETRY;
516         case NOT_READY:
517         case UNIT_ATTENTION:
518                 /*
519                  * if we are expecting a cc/ua because of a bus reset that we
520                  * performed, treat this just as a retry.  otherwise this is
521                  * information that we should pass up to the upper-level driver
522                  * so that we can deal with it there.
523                  */
524                 if (scmd->device->expecting_cc_ua) {
525                         /*
526                          * Because some device does not queue unit
527                          * attentions correctly, we carefully check
528                          * additional sense code and qualifier so as
529                          * not to squash media change unit attention.
530                          */
531                         if (sshdr.asc != 0x28 || sshdr.ascq != 0x00) {
532                                 scmd->device->expecting_cc_ua = 0;
533                                 return NEEDS_RETRY;
534                         }
535                 }
536                 /*
537                  * we might also expect a cc/ua if another LUN on the target
538                  * reported a UA with an ASC/ASCQ of 3F 0E -
539                  * REPORTED LUNS DATA HAS CHANGED.
540                  */
541                 if (scmd->device->sdev_target->expecting_lun_change &&
542                     sshdr.asc == 0x3f && sshdr.ascq == 0x0e)
543                         return NEEDS_RETRY;
544                 /*
545                  * if the device is in the process of becoming ready, we
546                  * should retry.
547                  */
548                 if ((sshdr.asc == 0x04) && (sshdr.ascq == 0x01))
549                         return NEEDS_RETRY;
550                 /*
551                  * if the device is not started, we need to wake
552                  * the error handler to start the motor
553                  */
554                 if (scmd->device->allow_restart &&
555                     (sshdr.asc == 0x04) && (sshdr.ascq == 0x02))
556                         return FAILED;
557                 /*
558                  * Pass the UA upwards for a determination in the completion
559                  * functions.
560                  */
561                 return SUCCESS;
562
563                 /* these are not supported */
564         case DATA_PROTECT:
565                 if (sshdr.asc == 0x27 && sshdr.ascq == 0x07) {
566                         /* Thin provisioning hard threshold reached */
567                         set_host_byte(scmd, DID_ALLOC_FAILURE);
568                         return SUCCESS;
569                 }
570                 /* FALLTHROUGH */
571         case COPY_ABORTED:
572         case VOLUME_OVERFLOW:
573         case MISCOMPARE:
574         case BLANK_CHECK:
575                 set_host_byte(scmd, DID_TARGET_FAILURE);
576                 return SUCCESS;
577
578         case MEDIUM_ERROR:
579                 if (sshdr.asc == 0x11 || /* UNRECOVERED READ ERR */
580                     sshdr.asc == 0x13 || /* AMNF DATA FIELD */
581                     sshdr.asc == 0x14) { /* RECORD NOT FOUND */
582                         set_host_byte(scmd, DID_MEDIUM_ERROR);
583                         return SUCCESS;
584                 }
585                 return NEEDS_RETRY;
586
587         case HARDWARE_ERROR:
588                 if (scmd->device->retry_hwerror)
589                         return ADD_TO_MLQUEUE;
590                 else
591                         set_host_byte(scmd, DID_TARGET_FAILURE);
592                 /* FALLTHROUGH */
593
594         case ILLEGAL_REQUEST:
595                 if (sshdr.asc == 0x20 || /* Invalid command operation code */
596                     sshdr.asc == 0x21 || /* Logical block address out of range */
597                     sshdr.asc == 0x24 || /* Invalid field in cdb */
598                     sshdr.asc == 0x26 || /* Parameter value invalid */
599                     sshdr.asc == 0x27) { /* Write protected */
600                         set_host_byte(scmd, DID_TARGET_FAILURE);
601                 }
602                 return SUCCESS;
603
604         default:
605                 return SUCCESS;
606         }
607 }
608 EXPORT_SYMBOL_GPL(scsi_check_sense);
609
610 static void scsi_handle_queue_ramp_up(struct scsi_device *sdev)
611 {
612         struct scsi_host_template *sht = sdev->host->hostt;
613         struct scsi_device *tmp_sdev;
614
615         if (!sht->track_queue_depth ||
616             sdev->queue_depth >= sdev->max_queue_depth)
617                 return;
618
619         if (time_before(jiffies,
620             sdev->last_queue_ramp_up + sdev->queue_ramp_up_period))
621                 return;
622
623         if (time_before(jiffies,
624             sdev->last_queue_full_time + sdev->queue_ramp_up_period))
625                 return;
626
627         /*
628          * Walk all devices of a target and do
629          * ramp up on them.
630          */
631         shost_for_each_device(tmp_sdev, sdev->host) {
632                 if (tmp_sdev->channel != sdev->channel ||
633                     tmp_sdev->id != sdev->id ||
634                     tmp_sdev->queue_depth == sdev->max_queue_depth)
635                         continue;
636
637                 scsi_change_queue_depth(tmp_sdev, tmp_sdev->queue_depth + 1);
638                 sdev->last_queue_ramp_up = jiffies;
639         }
640 }
641
642 static void scsi_handle_queue_full(struct scsi_device *sdev)
643 {
644         struct scsi_host_template *sht = sdev->host->hostt;
645         struct scsi_device *tmp_sdev;
646
647         if (!sht->track_queue_depth)
648                 return;
649
650         shost_for_each_device(tmp_sdev, sdev->host) {
651                 if (tmp_sdev->channel != sdev->channel ||
652                     tmp_sdev->id != sdev->id)
653                         continue;
654                 /*
655                  * We do not know the number of commands that were at
656                  * the device when we got the queue full so we start
657                  * from the highest possible value and work our way down.
658                  */
659                 scsi_track_queue_full(tmp_sdev, tmp_sdev->queue_depth - 1);
660         }
661 }
662
663 /**
664  * scsi_eh_completed_normally - Disposition a eh cmd on return from LLD.
665  * @scmd:       SCSI cmd to examine.
666  *
667  * Notes:
668  *    This is *only* called when we are examining the status of commands
669  *    queued during error recovery.  the main difference here is that we
670  *    don't allow for the possibility of retries here, and we are a lot
671  *    more restrictive about what we consider acceptable.
672  */
673 static int scsi_eh_completed_normally(struct scsi_cmnd *scmd)
674 {
675         /*
676          * first check the host byte, to see if there is anything in there
677          * that would indicate what we need to do.
678          */
679         if (host_byte(scmd->result) == DID_RESET) {
680                 /*
681                  * rats.  we are already in the error handler, so we now
682                  * get to try and figure out what to do next.  if the sense
683                  * is valid, we have a pretty good idea of what to do.
684                  * if not, we mark it as FAILED.
685                  */
686                 return scsi_check_sense(scmd);
687         }
688         if (host_byte(scmd->result) != DID_OK)
689                 return FAILED;
690
691         /*
692          * next, check the message byte.
693          */
694         if (msg_byte(scmd->result) != COMMAND_COMPLETE)
695                 return FAILED;
696
697         /*
698          * now, check the status byte to see if this indicates
699          * anything special.
700          */
701         switch (status_byte(scmd->result)) {
702         case GOOD:
703                 scsi_handle_queue_ramp_up(scmd->device);
704                 /* FALLTHROUGH */
705         case COMMAND_TERMINATED:
706                 return SUCCESS;
707         case CHECK_CONDITION:
708                 return scsi_check_sense(scmd);
709         case CONDITION_GOOD:
710         case INTERMEDIATE_GOOD:
711         case INTERMEDIATE_C_GOOD:
712                 /*
713                  * who knows?  FIXME(eric)
714                  */
715                 return SUCCESS;
716         case RESERVATION_CONFLICT:
717                 if (scmd->cmnd[0] == TEST_UNIT_READY)
718                         /* it is a success, we probed the device and
719                          * found it */
720                         return SUCCESS;
721                 /* otherwise, we failed to send the command */
722                 return FAILED;
723         case QUEUE_FULL:
724                 scsi_handle_queue_full(scmd->device);
725                 /* fall through */
726         case BUSY:
727                 return NEEDS_RETRY;
728         default:
729                 return FAILED;
730         }
731         return FAILED;
732 }
733
734 /**
735  * scsi_eh_done - Completion function for error handling.
736  * @scmd:       Cmd that is done.
737  */
738 static void scsi_eh_done(struct scsi_cmnd *scmd)
739 {
740         struct completion *eh_action;
741
742         SCSI_LOG_ERROR_RECOVERY(3, scmd_printk(KERN_INFO, scmd,
743                         "%s result: %x\n", __func__, scmd->result));
744
745         eh_action = scmd->device->host->eh_action;
746         if (eh_action)
747                 complete(eh_action);
748 }
749
750 /**
751  * scsi_try_host_reset - ask host adapter to reset itself
752  * @scmd:       SCSI cmd to send host reset.
753  */
754 static int scsi_try_host_reset(struct scsi_cmnd *scmd)
755 {
756         unsigned long flags;
757         int rtn;
758         struct Scsi_Host *host = scmd->device->host;
759         struct scsi_host_template *hostt = host->hostt;
760
761         SCSI_LOG_ERROR_RECOVERY(3,
762                 shost_printk(KERN_INFO, host, "Snd Host RST\n"));
763
764         if (!hostt->eh_host_reset_handler)
765                 return FAILED;
766
767         rtn = hostt->eh_host_reset_handler(scmd);
768
769         if (rtn == SUCCESS) {
770                 if (!hostt->skip_settle_delay)
771                         ssleep(HOST_RESET_SETTLE_TIME);
772                 spin_lock_irqsave(host->host_lock, flags);
773                 scsi_report_bus_reset(host, scmd_channel(scmd));
774                 spin_unlock_irqrestore(host->host_lock, flags);
775         }
776
777         return rtn;
778 }
779
780 /**
781  * scsi_try_bus_reset - ask host to perform a bus reset
782  * @scmd:       SCSI cmd to send bus reset.
783  */
784 static int scsi_try_bus_reset(struct scsi_cmnd *scmd)
785 {
786         unsigned long flags;
787         int rtn;
788         struct Scsi_Host *host = scmd->device->host;
789         struct scsi_host_template *hostt = host->hostt;
790
791         SCSI_LOG_ERROR_RECOVERY(3, scmd_printk(KERN_INFO, scmd,
792                 "%s: Snd Bus RST\n", __func__));
793
794         if (!hostt->eh_bus_reset_handler)
795                 return FAILED;
796
797         rtn = hostt->eh_bus_reset_handler(scmd);
798
799         if (rtn == SUCCESS) {
800                 if (!hostt->skip_settle_delay)
801                         ssleep(BUS_RESET_SETTLE_TIME);
802                 spin_lock_irqsave(host->host_lock, flags);
803                 scsi_report_bus_reset(host, scmd_channel(scmd));
804                 spin_unlock_irqrestore(host->host_lock, flags);
805         }
806
807         return rtn;
808 }
809
810 static void __scsi_report_device_reset(struct scsi_device *sdev, void *data)
811 {
812         sdev->was_reset = 1;
813         sdev->expecting_cc_ua = 1;
814 }
815
816 /**
817  * scsi_try_target_reset - Ask host to perform a target reset
818  * @scmd:       SCSI cmd used to send a target reset
819  *
820  * Notes:
821  *    There is no timeout for this operation.  if this operation is
822  *    unreliable for a given host, then the host itself needs to put a
823  *    timer on it, and set the host back to a consistent state prior to
824  *    returning.
825  */
826 static int scsi_try_target_reset(struct scsi_cmnd *scmd)
827 {
828         unsigned long flags;
829         int rtn;
830         struct Scsi_Host *host = scmd->device->host;
831         struct scsi_host_template *hostt = host->hostt;
832
833         if (!hostt->eh_target_reset_handler)
834                 return FAILED;
835
836         rtn = hostt->eh_target_reset_handler(scmd);
837         if (rtn == SUCCESS) {
838                 spin_lock_irqsave(host->host_lock, flags);
839                 __starget_for_each_device(scsi_target(scmd->device), NULL,
840                                           __scsi_report_device_reset);
841                 spin_unlock_irqrestore(host->host_lock, flags);
842         }
843
844         return rtn;
845 }
846
847 /**
848  * scsi_try_bus_device_reset - Ask host to perform a BDR on a dev
849  * @scmd:       SCSI cmd used to send BDR
850  *
851  * Notes:
852  *    There is no timeout for this operation.  if this operation is
853  *    unreliable for a given host, then the host itself needs to put a
854  *    timer on it, and set the host back to a consistent state prior to
855  *    returning.
856  */
857 static int scsi_try_bus_device_reset(struct scsi_cmnd *scmd)
858 {
859         int rtn;
860         struct scsi_host_template *hostt = scmd->device->host->hostt;
861
862         if (!hostt->eh_device_reset_handler)
863                 return FAILED;
864
865         rtn = hostt->eh_device_reset_handler(scmd);
866         if (rtn == SUCCESS)
867                 __scsi_report_device_reset(scmd->device, NULL);
868         return rtn;
869 }
870
871 /**
872  * scsi_try_to_abort_cmd - Ask host to abort a SCSI command
873  * @hostt:      SCSI driver host template
874  * @scmd:       SCSI cmd used to send a target reset
875  *
876  * Return value:
877  *      SUCCESS, FAILED, or FAST_IO_FAIL
878  *
879  * Notes:
880  *    SUCCESS does not necessarily indicate that the command
881  *    has been aborted; it only indicates that the LLDDs
882  *    has cleared all references to that command.
883  *    LLDDs should return FAILED only if an abort was required
884  *    but could not be executed. LLDDs should return FAST_IO_FAIL
885  *    if the device is temporarily unavailable (eg due to a
886  *    link down on FibreChannel)
887  */
888 static int scsi_try_to_abort_cmd(struct scsi_host_template *hostt,
889                                  struct scsi_cmnd *scmd)
890 {
891         if (!hostt->eh_abort_handler)
892                 return FAILED;
893
894         return hostt->eh_abort_handler(scmd);
895 }
896
897 static void scsi_abort_eh_cmnd(struct scsi_cmnd *scmd)
898 {
899         if (scsi_try_to_abort_cmd(scmd->device->host->hostt, scmd) != SUCCESS)
900                 if (scsi_try_bus_device_reset(scmd) != SUCCESS)
901                         if (scsi_try_target_reset(scmd) != SUCCESS)
902                                 if (scsi_try_bus_reset(scmd) != SUCCESS)
903                                         scsi_try_host_reset(scmd);
904 }
905
906 /**
907  * scsi_eh_prep_cmnd  - Save a scsi command info as part of error recovery
908  * @scmd:       SCSI command structure to hijack
909  * @ses:        structure to save restore information
910  * @cmnd:       CDB to send. Can be NULL if no new cmnd is needed
911  * @cmnd_size:  size in bytes of @cmnd (must be <= BLK_MAX_CDB)
912  * @sense_bytes: size of sense data to copy. or 0 (if != 0 @cmnd is ignored)
913  *
914  * This function is used to save a scsi command information before re-execution
915  * as part of the error recovery process.  If @sense_bytes is 0 the command
916  * sent must be one that does not transfer any data.  If @sense_bytes != 0
917  * @cmnd is ignored and this functions sets up a REQUEST_SENSE command
918  * and cmnd buffers to read @sense_bytes into @scmd->sense_buffer.
919  */
920 void scsi_eh_prep_cmnd(struct scsi_cmnd *scmd, struct scsi_eh_save *ses,
921                         unsigned char *cmnd, int cmnd_size, unsigned sense_bytes)
922 {
923         struct scsi_device *sdev = scmd->device;
924
925         /*
926          * We need saved copies of a number of fields - this is because
927          * error handling may need to overwrite these with different values
928          * to run different commands, and once error handling is complete,
929          * we will need to restore these values prior to running the actual
930          * command.
931          */
932         ses->cmd_len = scmd->cmd_len;
933         ses->cmnd = scmd->cmnd;
934         ses->data_direction = scmd->sc_data_direction;
935         ses->sdb = scmd->sdb;
936         ses->next_rq = scmd->request->next_rq;
937         ses->result = scmd->result;
938         ses->resid_len = scmd->req.resid_len;
939         ses->underflow = scmd->underflow;
940         ses->prot_op = scmd->prot_op;
941         ses->eh_eflags = scmd->eh_eflags;
942
943         scmd->prot_op = SCSI_PROT_NORMAL;
944         scmd->eh_eflags = 0;
945         scmd->cmnd = ses->eh_cmnd;
946         memset(scmd->cmnd, 0, BLK_MAX_CDB);
947         memset(&scmd->sdb, 0, sizeof(scmd->sdb));
948         scmd->request->next_rq = NULL;
949         scmd->result = 0;
950         scmd->req.resid_len = 0;
951
952         if (sense_bytes) {
953                 scmd->sdb.length = min_t(unsigned, SCSI_SENSE_BUFFERSIZE,
954                                          sense_bytes);
955                 sg_init_one(&ses->sense_sgl, scmd->sense_buffer,
956                             scmd->sdb.length);
957                 scmd->sdb.table.sgl = &ses->sense_sgl;
958                 scmd->sc_data_direction = DMA_FROM_DEVICE;
959                 scmd->sdb.table.nents = scmd->sdb.table.orig_nents = 1;
960                 scmd->cmnd[0] = REQUEST_SENSE;
961                 scmd->cmnd[4] = scmd->sdb.length;
962                 scmd->cmd_len = COMMAND_SIZE(scmd->cmnd[0]);
963         } else {
964                 scmd->sc_data_direction = DMA_NONE;
965                 if (cmnd) {
966                         BUG_ON(cmnd_size > BLK_MAX_CDB);
967                         memcpy(scmd->cmnd, cmnd, cmnd_size);
968                         scmd->cmd_len = COMMAND_SIZE(scmd->cmnd[0]);
969                 }
970         }
971
972         scmd->underflow = 0;
973
974         if (sdev->scsi_level <= SCSI_2 && sdev->scsi_level != SCSI_UNKNOWN)
975                 scmd->cmnd[1] = (scmd->cmnd[1] & 0x1f) |
976                         (sdev->lun << 5 & 0xe0);
977
978         /*
979          * Zero the sense buffer.  The scsi spec mandates that any
980          * untransferred sense data should be interpreted as being zero.
981          */
982         memset(scmd->sense_buffer, 0, SCSI_SENSE_BUFFERSIZE);
983 }
984 EXPORT_SYMBOL(scsi_eh_prep_cmnd);
985
986 /**
987  * scsi_eh_restore_cmnd  - Restore a scsi command info as part of error recovery
988  * @scmd:       SCSI command structure to restore
989  * @ses:        saved information from a coresponding call to scsi_eh_prep_cmnd
990  *
991  * Undo any damage done by above scsi_eh_prep_cmnd().
992  */
993 void scsi_eh_restore_cmnd(struct scsi_cmnd* scmd, struct scsi_eh_save *ses)
994 {
995         /*
996          * Restore original data
997          */
998         scmd->cmd_len = ses->cmd_len;
999         scmd->cmnd = ses->cmnd;
1000         scmd->sc_data_direction = ses->data_direction;
1001         scmd->sdb = ses->sdb;
1002         scmd->request->next_rq = ses->next_rq;
1003         scmd->result = ses->result;
1004         scmd->req.resid_len = ses->resid_len;
1005         scmd->underflow = ses->underflow;
1006         scmd->prot_op = ses->prot_op;
1007         scmd->eh_eflags = ses->eh_eflags;
1008 }
1009 EXPORT_SYMBOL(scsi_eh_restore_cmnd);
1010
1011 /**
1012  * scsi_send_eh_cmnd  - submit a scsi command as part of error recovery
1013  * @scmd:       SCSI command structure to hijack
1014  * @cmnd:       CDB to send
1015  * @cmnd_size:  size in bytes of @cmnd
1016  * @timeout:    timeout for this request
1017  * @sense_bytes: size of sense data to copy or 0
1018  *
1019  * This function is used to send a scsi command down to a target device
1020  * as part of the error recovery process. See also scsi_eh_prep_cmnd() above.
1021  *
1022  * Return value:
1023  *    SUCCESS or FAILED or NEEDS_RETRY
1024  */
1025 static int scsi_send_eh_cmnd(struct scsi_cmnd *scmd, unsigned char *cmnd,
1026                              int cmnd_size, int timeout, unsigned sense_bytes)
1027 {
1028         struct scsi_device *sdev = scmd->device;
1029         struct Scsi_Host *shost = sdev->host;
1030         DECLARE_COMPLETION_ONSTACK(done);
1031         unsigned long timeleft = timeout;
1032         struct scsi_eh_save ses;
1033         const unsigned long stall_for = msecs_to_jiffies(100);
1034         int rtn;
1035
1036 retry:
1037         scsi_eh_prep_cmnd(scmd, &ses, cmnd, cmnd_size, sense_bytes);
1038         shost->eh_action = &done;
1039
1040         scsi_log_send(scmd);
1041         scmd->scsi_done = scsi_eh_done;
1042         rtn = shost->hostt->queuecommand(shost, scmd);
1043         if (rtn) {
1044                 if (timeleft > stall_for) {
1045                         scsi_eh_restore_cmnd(scmd, &ses);
1046                         timeleft -= stall_for;
1047                         msleep(jiffies_to_msecs(stall_for));
1048                         goto retry;
1049                 }
1050                 /* signal not to enter either branch of the if () below */
1051                 timeleft = 0;
1052                 rtn = FAILED;
1053         } else {
1054                 timeleft = wait_for_completion_timeout(&done, timeout);
1055                 rtn = SUCCESS;
1056         }
1057
1058         shost->eh_action = NULL;
1059
1060         scsi_log_completion(scmd, rtn);
1061
1062         SCSI_LOG_ERROR_RECOVERY(3, scmd_printk(KERN_INFO, scmd,
1063                         "%s timeleft: %ld\n",
1064                         __func__, timeleft));
1065
1066         /*
1067          * If there is time left scsi_eh_done got called, and we will examine
1068          * the actual status codes to see whether the command actually did
1069          * complete normally, else if we have a zero return and no time left,
1070          * the command must still be pending, so abort it and return FAILED.
1071          * If we never actually managed to issue the command, because
1072          * ->queuecommand() kept returning non zero, use the rtn = FAILED
1073          * value above (so don't execute either branch of the if)
1074          */
1075         if (timeleft) {
1076                 rtn = scsi_eh_completed_normally(scmd);
1077                 SCSI_LOG_ERROR_RECOVERY(3, scmd_printk(KERN_INFO, scmd,
1078                         "%s: scsi_eh_completed_normally %x\n", __func__, rtn));
1079
1080                 switch (rtn) {
1081                 case SUCCESS:
1082                 case NEEDS_RETRY:
1083                 case FAILED:
1084                         break;
1085                 case ADD_TO_MLQUEUE:
1086                         rtn = NEEDS_RETRY;
1087                         break;
1088                 default:
1089                         rtn = FAILED;
1090                         break;
1091                 }
1092         } else if (rtn != FAILED) {
1093                 scsi_abort_eh_cmnd(scmd);
1094                 rtn = FAILED;
1095         }
1096
1097         scsi_eh_restore_cmnd(scmd, &ses);
1098
1099         return rtn;
1100 }
1101
1102 /**
1103  * scsi_request_sense - Request sense data from a particular target.
1104  * @scmd:       SCSI cmd for request sense.
1105  *
1106  * Notes:
1107  *    Some hosts automatically obtain this information, others require
1108  *    that we obtain it on our own. This function will *not* return until
1109  *    the command either times out, or it completes.
1110  */
1111 static int scsi_request_sense(struct scsi_cmnd *scmd)
1112 {
1113         return scsi_send_eh_cmnd(scmd, NULL, 0, scmd->device->eh_timeout, ~0);
1114 }
1115
1116 static int scsi_eh_action(struct scsi_cmnd *scmd, int rtn)
1117 {
1118         if (!blk_rq_is_passthrough(scmd->request)) {
1119                 struct scsi_driver *sdrv = scsi_cmd_to_driver(scmd);
1120                 if (sdrv->eh_action)
1121                         rtn = sdrv->eh_action(scmd, rtn);
1122         }
1123         return rtn;
1124 }
1125
1126 /**
1127  * scsi_eh_finish_cmd - Handle a cmd that eh is finished with.
1128  * @scmd:       Original SCSI cmd that eh has finished.
1129  * @done_q:     Queue for processed commands.
1130  *
1131  * Notes:
1132  *    We don't want to use the normal command completion while we are are
1133  *    still handling errors - it may cause other commands to be queued,
1134  *    and that would disturb what we are doing.  Thus we really want to
1135  *    keep a list of pending commands for final completion, and once we
1136  *    are ready to leave error handling we handle completion for real.
1137  */
1138 void scsi_eh_finish_cmd(struct scsi_cmnd *scmd, struct list_head *done_q)
1139 {
1140         list_move_tail(&scmd->eh_entry, done_q);
1141 }
1142 EXPORT_SYMBOL(scsi_eh_finish_cmd);
1143
1144 /**
1145  * scsi_eh_get_sense - Get device sense data.
1146  * @work_q:     Queue of commands to process.
1147  * @done_q:     Queue of processed commands.
1148  *
1149  * Description:
1150  *    See if we need to request sense information.  if so, then get it
1151  *    now, so we have a better idea of what to do.
1152  *
1153  * Notes:
1154  *    This has the unfortunate side effect that if a shost adapter does
1155  *    not automatically request sense information, we end up shutting
1156  *    it down before we request it.
1157  *
1158  *    All drivers should request sense information internally these days,
1159  *    so for now all I have to say is tough noogies if you end up in here.
1160  *
1161  *    XXX: Long term this code should go away, but that needs an audit of
1162  *         all LLDDs first.
1163  */
1164 int scsi_eh_get_sense(struct list_head *work_q,
1165                       struct list_head *done_q)
1166 {
1167         struct scsi_cmnd *scmd, *next;
1168         struct Scsi_Host *shost;
1169         int rtn;
1170
1171         /*
1172          * If SCSI_EH_ABORT_SCHEDULED has been set, it is timeout IO,
1173          * should not get sense.
1174          */
1175         list_for_each_entry_safe(scmd, next, work_q, eh_entry) {
1176                 if ((scmd->eh_eflags & SCSI_EH_ABORT_SCHEDULED) ||
1177                     SCSI_SENSE_VALID(scmd))
1178                         continue;
1179
1180                 shost = scmd->device->host;
1181                 if (scsi_host_eh_past_deadline(shost)) {
1182                         SCSI_LOG_ERROR_RECOVERY(3,
1183                                 scmd_printk(KERN_INFO, scmd,
1184                                             "%s: skip request sense, past eh deadline\n",
1185                                              current->comm));
1186                         break;
1187                 }
1188                 if (status_byte(scmd->result) != CHECK_CONDITION)
1189                         /*
1190                          * don't request sense if there's no check condition
1191                          * status because the error we're processing isn't one
1192                          * that has a sense code (and some devices get
1193                          * confused by sense requests out of the blue)
1194                          */
1195                         continue;
1196
1197                 SCSI_LOG_ERROR_RECOVERY(2, scmd_printk(KERN_INFO, scmd,
1198                                                   "%s: requesting sense\n",
1199                                                   current->comm));
1200                 rtn = scsi_request_sense(scmd);
1201                 if (rtn != SUCCESS)
1202                         continue;
1203
1204                 SCSI_LOG_ERROR_RECOVERY(3, scmd_printk(KERN_INFO, scmd,
1205                         "sense requested, result %x\n", scmd->result));
1206                 SCSI_LOG_ERROR_RECOVERY(3, scsi_print_sense(scmd));
1207
1208                 rtn = scsi_decide_disposition(scmd);
1209
1210                 /*
1211                  * if the result was normal, then just pass it along to the
1212                  * upper level.
1213                  */
1214                 if (rtn == SUCCESS)
1215                         /* we don't want this command reissued, just
1216                          * finished with the sense data, so set
1217                          * retries to the max allowed to ensure it
1218                          * won't get reissued */
1219                         scmd->retries = scmd->allowed;
1220                 else if (rtn != NEEDS_RETRY)
1221                         continue;
1222
1223                 scsi_eh_finish_cmd(scmd, done_q);
1224         }
1225
1226         return list_empty(work_q);
1227 }
1228 EXPORT_SYMBOL_GPL(scsi_eh_get_sense);
1229
1230 /**
1231  * scsi_eh_tur - Send TUR to device.
1232  * @scmd:       &scsi_cmnd to send TUR
1233  *
1234  * Return value:
1235  *    0 - Device is ready. 1 - Device NOT ready.
1236  */
1237 static int scsi_eh_tur(struct scsi_cmnd *scmd)
1238 {
1239         static unsigned char tur_command[6] = {TEST_UNIT_READY, 0, 0, 0, 0, 0};
1240         int retry_cnt = 1, rtn;
1241
1242 retry_tur:
1243         rtn = scsi_send_eh_cmnd(scmd, tur_command, 6,
1244                                 scmd->device->eh_timeout, 0);
1245
1246         SCSI_LOG_ERROR_RECOVERY(3, scmd_printk(KERN_INFO, scmd,
1247                 "%s return: %x\n", __func__, rtn));
1248
1249         switch (rtn) {
1250         case NEEDS_RETRY:
1251                 if (retry_cnt--)
1252                         goto retry_tur;
1253                 /*FALLTHRU*/
1254         case SUCCESS:
1255                 return 0;
1256         default:
1257                 return 1;
1258         }
1259 }
1260
1261 /**
1262  * scsi_eh_test_devices - check if devices are responding from error recovery.
1263  * @cmd_list:   scsi commands in error recovery.
1264  * @work_q:     queue for commands which still need more error recovery
1265  * @done_q:     queue for commands which are finished
1266  * @try_stu:    boolean on if a STU command should be tried in addition to TUR.
1267  *
1268  * Decription:
1269  *    Tests if devices are in a working state.  Commands to devices now in
1270  *    a working state are sent to the done_q while commands to devices which
1271  *    are still failing to respond are returned to the work_q for more
1272  *    processing.
1273  **/
1274 static int scsi_eh_test_devices(struct list_head *cmd_list,
1275                                 struct list_head *work_q,
1276                                 struct list_head *done_q, int try_stu)
1277 {
1278         struct scsi_cmnd *scmd, *next;
1279         struct scsi_device *sdev;
1280         int finish_cmds;
1281
1282         while (!list_empty(cmd_list)) {
1283                 scmd = list_entry(cmd_list->next, struct scsi_cmnd, eh_entry);
1284                 sdev = scmd->device;
1285
1286                 if (!try_stu) {
1287                         if (scsi_host_eh_past_deadline(sdev->host)) {
1288                                 /* Push items back onto work_q */
1289                                 list_splice_init(cmd_list, work_q);
1290                                 SCSI_LOG_ERROR_RECOVERY(3,
1291                                         sdev_printk(KERN_INFO, sdev,
1292                                                     "%s: skip test device, past eh deadline",
1293                                                     current->comm));
1294                                 break;
1295                         }
1296                 }
1297
1298                 finish_cmds = !scsi_device_online(scmd->device) ||
1299                         (try_stu && !scsi_eh_try_stu(scmd) &&
1300                          !scsi_eh_tur(scmd)) ||
1301                         !scsi_eh_tur(scmd);
1302
1303                 list_for_each_entry_safe(scmd, next, cmd_list, eh_entry)
1304                         if (scmd->device == sdev) {
1305                                 if (finish_cmds &&
1306                                     (try_stu ||
1307                                      scsi_eh_action(scmd, SUCCESS) == SUCCESS))
1308                                         scsi_eh_finish_cmd(scmd, done_q);
1309                                 else
1310                                         list_move_tail(&scmd->eh_entry, work_q);
1311                         }
1312         }
1313         return list_empty(work_q);
1314 }
1315
1316 /**
1317  * scsi_eh_try_stu - Send START_UNIT to device.
1318  * @scmd:       &scsi_cmnd to send START_UNIT
1319  *
1320  * Return value:
1321  *    0 - Device is ready. 1 - Device NOT ready.
1322  */
1323 static int scsi_eh_try_stu(struct scsi_cmnd *scmd)
1324 {
1325         static unsigned char stu_command[6] = {START_STOP, 0, 0, 0, 1, 0};
1326
1327         if (scmd->device->allow_restart) {
1328                 int i, rtn = NEEDS_RETRY;
1329
1330                 for (i = 0; rtn == NEEDS_RETRY && i < 2; i++)
1331                         rtn = scsi_send_eh_cmnd(scmd, stu_command, 6, scmd->device->request_queue->rq_timeout, 0);
1332
1333                 if (rtn == SUCCESS)
1334                         return 0;
1335         }
1336
1337         return 1;
1338 }
1339
1340  /**
1341  * scsi_eh_stu - send START_UNIT if needed
1342  * @shost:      &scsi host being recovered.
1343  * @work_q:     &list_head for pending commands.
1344  * @done_q:     &list_head for processed commands.
1345  *
1346  * Notes:
1347  *    If commands are failing due to not ready, initializing command required,
1348  *      try revalidating the device, which will end up sending a start unit.
1349  */
1350 static int scsi_eh_stu(struct Scsi_Host *shost,
1351                               struct list_head *work_q,
1352                               struct list_head *done_q)
1353 {
1354         struct scsi_cmnd *scmd, *stu_scmd, *next;
1355         struct scsi_device *sdev;
1356
1357         shost_for_each_device(sdev, shost) {
1358                 if (scsi_host_eh_past_deadline(shost)) {
1359                         SCSI_LOG_ERROR_RECOVERY(3,
1360                                 sdev_printk(KERN_INFO, sdev,
1361                                             "%s: skip START_UNIT, past eh deadline\n",
1362                                             current->comm));
1363                         break;
1364                 }
1365                 stu_scmd = NULL;
1366                 list_for_each_entry(scmd, work_q, eh_entry)
1367                         if (scmd->device == sdev && SCSI_SENSE_VALID(scmd) &&
1368                             scsi_check_sense(scmd) == FAILED ) {
1369                                 stu_scmd = scmd;
1370                                 break;
1371                         }
1372
1373                 if (!stu_scmd)
1374                         continue;
1375
1376                 SCSI_LOG_ERROR_RECOVERY(3,
1377                         sdev_printk(KERN_INFO, sdev,
1378                                      "%s: Sending START_UNIT\n",
1379                                     current->comm));
1380
1381                 if (!scsi_eh_try_stu(stu_scmd)) {
1382                         if (!scsi_device_online(sdev) ||
1383                             !scsi_eh_tur(stu_scmd)) {
1384                                 list_for_each_entry_safe(scmd, next,
1385                                                           work_q, eh_entry) {
1386                                         if (scmd->device == sdev &&
1387                                             scsi_eh_action(scmd, SUCCESS) == SUCCESS)
1388                                                 scsi_eh_finish_cmd(scmd, done_q);
1389                                 }
1390                         }
1391                 } else {
1392                         SCSI_LOG_ERROR_RECOVERY(3,
1393                                 sdev_printk(KERN_INFO, sdev,
1394                                             "%s: START_UNIT failed\n",
1395                                             current->comm));
1396                 }
1397         }
1398
1399         return list_empty(work_q);
1400 }
1401
1402
1403 /**
1404  * scsi_eh_bus_device_reset - send bdr if needed
1405  * @shost:      scsi host being recovered.
1406  * @work_q:     &list_head for pending commands.
1407  * @done_q:     &list_head for processed commands.
1408  *
1409  * Notes:
1410  *    Try a bus device reset.  Still, look to see whether we have multiple
1411  *    devices that are jammed or not - if we have multiple devices, it
1412  *    makes no sense to try bus_device_reset - we really would need to try
1413  *    a bus_reset instead.
1414  */
1415 static int scsi_eh_bus_device_reset(struct Scsi_Host *shost,
1416                                     struct list_head *work_q,
1417                                     struct list_head *done_q)
1418 {
1419         struct scsi_cmnd *scmd, *bdr_scmd, *next;
1420         struct scsi_device *sdev;
1421         int rtn;
1422
1423         shost_for_each_device(sdev, shost) {
1424                 if (scsi_host_eh_past_deadline(shost)) {
1425                         SCSI_LOG_ERROR_RECOVERY(3,
1426                                 sdev_printk(KERN_INFO, sdev,
1427                                             "%s: skip BDR, past eh deadline\n",
1428                                              current->comm));
1429                         break;
1430                 }
1431                 bdr_scmd = NULL;
1432                 list_for_each_entry(scmd, work_q, eh_entry)
1433                         if (scmd->device == sdev) {
1434                                 bdr_scmd = scmd;
1435                                 break;
1436                         }
1437
1438                 if (!bdr_scmd)
1439                         continue;
1440
1441                 SCSI_LOG_ERROR_RECOVERY(3,
1442                         sdev_printk(KERN_INFO, sdev,
1443                                      "%s: Sending BDR\n", current->comm));
1444                 rtn = scsi_try_bus_device_reset(bdr_scmd);
1445                 if (rtn == SUCCESS || rtn == FAST_IO_FAIL) {
1446                         if (!scsi_device_online(sdev) ||
1447                             rtn == FAST_IO_FAIL ||
1448                             !scsi_eh_tur(bdr_scmd)) {
1449                                 list_for_each_entry_safe(scmd, next,
1450                                                          work_q, eh_entry) {
1451                                         if (scmd->device == sdev &&
1452                                             scsi_eh_action(scmd, rtn) != FAILED)
1453                                                 scsi_eh_finish_cmd(scmd,
1454                                                                    done_q);
1455                                 }
1456                         }
1457                 } else {
1458                         SCSI_LOG_ERROR_RECOVERY(3,
1459                                 sdev_printk(KERN_INFO, sdev,
1460                                             "%s: BDR failed\n", current->comm));
1461                 }
1462         }
1463
1464         return list_empty(work_q);
1465 }
1466
1467 /**
1468  * scsi_eh_target_reset - send target reset if needed
1469  * @shost:      scsi host being recovered.
1470  * @work_q:     &list_head for pending commands.
1471  * @done_q:     &list_head for processed commands.
1472  *
1473  * Notes:
1474  *    Try a target reset.
1475  */
1476 static int scsi_eh_target_reset(struct Scsi_Host *shost,
1477                                 struct list_head *work_q,
1478                                 struct list_head *done_q)
1479 {
1480         LIST_HEAD(tmp_list);
1481         LIST_HEAD(check_list);
1482
1483         list_splice_init(work_q, &tmp_list);
1484
1485         while (!list_empty(&tmp_list)) {
1486                 struct scsi_cmnd *next, *scmd;
1487                 int rtn;
1488                 unsigned int id;
1489
1490                 if (scsi_host_eh_past_deadline(shost)) {
1491                         /* push back on work queue for further processing */
1492                         list_splice_init(&check_list, work_q);
1493                         list_splice_init(&tmp_list, work_q);
1494                         SCSI_LOG_ERROR_RECOVERY(3,
1495                                 shost_printk(KERN_INFO, shost,
1496                                             "%s: Skip target reset, past eh deadline\n",
1497                                              current->comm));
1498                         return list_empty(work_q);
1499                 }
1500
1501                 scmd = list_entry(tmp_list.next, struct scsi_cmnd, eh_entry);
1502                 id = scmd_id(scmd);
1503
1504                 SCSI_LOG_ERROR_RECOVERY(3,
1505                         shost_printk(KERN_INFO, shost,
1506                                      "%s: Sending target reset to target %d\n",
1507                                      current->comm, id));
1508                 rtn = scsi_try_target_reset(scmd);
1509                 if (rtn != SUCCESS && rtn != FAST_IO_FAIL)
1510                         SCSI_LOG_ERROR_RECOVERY(3,
1511                                 shost_printk(KERN_INFO, shost,
1512                                              "%s: Target reset failed"
1513                                              " target: %d\n",
1514                                              current->comm, id));
1515                 list_for_each_entry_safe(scmd, next, &tmp_list, eh_entry) {
1516                         if (scmd_id(scmd) != id)
1517                                 continue;
1518
1519                         if (rtn == SUCCESS)
1520                                 list_move_tail(&scmd->eh_entry, &check_list);
1521                         else if (rtn == FAST_IO_FAIL)
1522                                 scsi_eh_finish_cmd(scmd, done_q);
1523                         else
1524                                 /* push back on work queue for further processing */
1525                                 list_move(&scmd->eh_entry, work_q);
1526                 }
1527         }
1528
1529         return scsi_eh_test_devices(&check_list, work_q, done_q, 0);
1530 }
1531
1532 /**
1533  * scsi_eh_bus_reset - send a bus reset
1534  * @shost:      &scsi host being recovered.
1535  * @work_q:     &list_head for pending commands.
1536  * @done_q:     &list_head for processed commands.
1537  */
1538 static int scsi_eh_bus_reset(struct Scsi_Host *shost,
1539                              struct list_head *work_q,
1540                              struct list_head *done_q)
1541 {
1542         struct scsi_cmnd *scmd, *chan_scmd, *next;
1543         LIST_HEAD(check_list);
1544         unsigned int channel;
1545         int rtn;
1546
1547         /*
1548          * we really want to loop over the various channels, and do this on
1549          * a channel by channel basis.  we should also check to see if any
1550          * of the failed commands are on soft_reset devices, and if so, skip
1551          * the reset.
1552          */
1553
1554         for (channel = 0; channel <= shost->max_channel; channel++) {
1555                 if (scsi_host_eh_past_deadline(shost)) {
1556                         list_splice_init(&check_list, work_q);
1557                         SCSI_LOG_ERROR_RECOVERY(3,
1558                                 shost_printk(KERN_INFO, shost,
1559                                             "%s: skip BRST, past eh deadline\n",
1560                                              current->comm));
1561                         return list_empty(work_q);
1562                 }
1563
1564                 chan_scmd = NULL;
1565                 list_for_each_entry(scmd, work_q, eh_entry) {
1566                         if (channel == scmd_channel(scmd)) {
1567                                 chan_scmd = scmd;
1568                                 break;
1569                                 /*
1570                                  * FIXME add back in some support for
1571                                  * soft_reset devices.
1572                                  */
1573                         }
1574                 }
1575
1576                 if (!chan_scmd)
1577                         continue;
1578                 SCSI_LOG_ERROR_RECOVERY(3,
1579                         shost_printk(KERN_INFO, shost,
1580                                      "%s: Sending BRST chan: %d\n",
1581                                      current->comm, channel));
1582                 rtn = scsi_try_bus_reset(chan_scmd);
1583                 if (rtn == SUCCESS || rtn == FAST_IO_FAIL) {
1584                         list_for_each_entry_safe(scmd, next, work_q, eh_entry) {
1585                                 if (channel == scmd_channel(scmd)) {
1586                                         if (rtn == FAST_IO_FAIL)
1587                                                 scsi_eh_finish_cmd(scmd,
1588                                                                    done_q);
1589                                         else
1590                                                 list_move_tail(&scmd->eh_entry,
1591                                                                &check_list);
1592                                 }
1593                         }
1594                 } else {
1595                         SCSI_LOG_ERROR_RECOVERY(3,
1596                                 shost_printk(KERN_INFO, shost,
1597                                              "%s: BRST failed chan: %d\n",
1598                                              current->comm, channel));
1599                 }
1600         }
1601         return scsi_eh_test_devices(&check_list, work_q, done_q, 0);
1602 }
1603
1604 /**
1605  * scsi_eh_host_reset - send a host reset
1606  * @shost:      host to be reset.
1607  * @work_q:     &list_head for pending commands.
1608  * @done_q:     &list_head for processed commands.
1609  */
1610 static int scsi_eh_host_reset(struct Scsi_Host *shost,
1611                               struct list_head *work_q,
1612                               struct list_head *done_q)
1613 {
1614         struct scsi_cmnd *scmd, *next;
1615         LIST_HEAD(check_list);
1616         int rtn;
1617
1618         if (!list_empty(work_q)) {
1619                 scmd = list_entry(work_q->next,
1620                                   struct scsi_cmnd, eh_entry);
1621
1622                 SCSI_LOG_ERROR_RECOVERY(3,
1623                         shost_printk(KERN_INFO, shost,
1624                                      "%s: Sending HRST\n",
1625                                      current->comm));
1626
1627                 rtn = scsi_try_host_reset(scmd);
1628                 if (rtn == SUCCESS) {
1629                         list_splice_init(work_q, &check_list);
1630                 } else if (rtn == FAST_IO_FAIL) {
1631                         list_for_each_entry_safe(scmd, next, work_q, eh_entry) {
1632                                         scsi_eh_finish_cmd(scmd, done_q);
1633                         }
1634                 } else {
1635                         SCSI_LOG_ERROR_RECOVERY(3,
1636                                 shost_printk(KERN_INFO, shost,
1637                                              "%s: HRST failed\n",
1638                                              current->comm));
1639                 }
1640         }
1641         return scsi_eh_test_devices(&check_list, work_q, done_q, 1);
1642 }
1643
1644 /**
1645  * scsi_eh_offline_sdevs - offline scsi devices that fail to recover
1646  * @work_q:     &list_head for pending commands.
1647  * @done_q:     &list_head for processed commands.
1648  */
1649 static void scsi_eh_offline_sdevs(struct list_head *work_q,
1650                                   struct list_head *done_q)
1651 {
1652         struct scsi_cmnd *scmd, *next;
1653         struct scsi_device *sdev;
1654
1655         list_for_each_entry_safe(scmd, next, work_q, eh_entry) {
1656                 sdev_printk(KERN_INFO, scmd->device, "Device offlined - "
1657                             "not ready after error recovery\n");
1658                 sdev = scmd->device;
1659
1660                 mutex_lock(&sdev->state_mutex);
1661                 scsi_device_set_state(sdev, SDEV_OFFLINE);
1662                 mutex_unlock(&sdev->state_mutex);
1663
1664                 scsi_eh_finish_cmd(scmd, done_q);
1665         }
1666         return;
1667 }
1668
1669 /**
1670  * scsi_noretry_cmd - determine if command should be failed fast
1671  * @scmd:       SCSI cmd to examine.
1672  */
1673 int scsi_noretry_cmd(struct scsi_cmnd *scmd)
1674 {
1675         switch (host_byte(scmd->result)) {
1676         case DID_OK:
1677                 break;
1678         case DID_TIME_OUT:
1679                 goto check_type;
1680         case DID_BUS_BUSY:
1681                 return (scmd->request->cmd_flags & REQ_FAILFAST_TRANSPORT);
1682         case DID_PARITY:
1683                 return (scmd->request->cmd_flags & REQ_FAILFAST_DEV);
1684         case DID_ERROR:
1685                 if (msg_byte(scmd->result) == COMMAND_COMPLETE &&
1686                     status_byte(scmd->result) == RESERVATION_CONFLICT)
1687                         return 0;
1688                 /* fall through */
1689         case DID_SOFT_ERROR:
1690                 return (scmd->request->cmd_flags & REQ_FAILFAST_DRIVER);
1691         }
1692
1693         if (status_byte(scmd->result) != CHECK_CONDITION)
1694                 return 0;
1695
1696 check_type:
1697         /*
1698          * assume caller has checked sense and determined
1699          * the check condition was retryable.
1700          */
1701         if (scmd->request->cmd_flags & REQ_FAILFAST_DEV ||
1702             blk_rq_is_passthrough(scmd->request))
1703                 return 1;
1704         else
1705                 return 0;
1706 }
1707
1708 /**
1709  * scsi_decide_disposition - Disposition a cmd on return from LLD.
1710  * @scmd:       SCSI cmd to examine.
1711  *
1712  * Notes:
1713  *    This is *only* called when we are examining the status after sending
1714  *    out the actual data command.  any commands that are queued for error
1715  *    recovery (e.g. test_unit_ready) do *not* come through here.
1716  *
1717  *    When this routine returns failed, it means the error handler thread
1718  *    is woken.  In cases where the error code indicates an error that
1719  *    doesn't require the error handler read (i.e. we don't need to
1720  *    abort/reset), this function should return SUCCESS.
1721  */
1722 int scsi_decide_disposition(struct scsi_cmnd *scmd)
1723 {
1724         int rtn;
1725
1726         /*
1727          * if the device is offline, then we clearly just pass the result back
1728          * up to the top level.
1729          */
1730         if (!scsi_device_online(scmd->device)) {
1731                 SCSI_LOG_ERROR_RECOVERY(5, scmd_printk(KERN_INFO, scmd,
1732                         "%s: device offline - report as SUCCESS\n", __func__));
1733                 return SUCCESS;
1734         }
1735
1736         /*
1737          * first check the host byte, to see if there is anything in there
1738          * that would indicate what we need to do.
1739          */
1740         switch (host_byte(scmd->result)) {
1741         case DID_PASSTHROUGH:
1742                 /*
1743                  * no matter what, pass this through to the upper layer.
1744                  * nuke this special code so that it looks like we are saying
1745                  * did_ok.
1746                  */
1747                 scmd->result &= 0xff00ffff;
1748                 return SUCCESS;
1749         case DID_OK:
1750                 /*
1751                  * looks good.  drop through, and check the next byte.
1752                  */
1753                 break;
1754         case DID_ABORT:
1755                 if (scmd->eh_eflags & SCSI_EH_ABORT_SCHEDULED) {
1756                         set_host_byte(scmd, DID_TIME_OUT);
1757                         return SUCCESS;
1758                 }
1759                 /* FALLTHROUGH */
1760         case DID_NO_CONNECT:
1761         case DID_BAD_TARGET:
1762                 /*
1763                  * note - this means that we just report the status back
1764                  * to the top level driver, not that we actually think
1765                  * that it indicates SUCCESS.
1766                  */
1767                 return SUCCESS;
1768                 /*
1769                  * when the low level driver returns did_soft_error,
1770                  * it is responsible for keeping an internal retry counter
1771                  * in order to avoid endless loops (db)
1772                  *
1773                  * actually this is a bug in this function here.  we should
1774                  * be mindful of the maximum number of retries specified
1775                  * and not get stuck in a loop.
1776                  */
1777         case DID_SOFT_ERROR:
1778                 goto maybe_retry;
1779         case DID_IMM_RETRY:
1780                 return NEEDS_RETRY;
1781
1782         case DID_REQUEUE:
1783                 return ADD_TO_MLQUEUE;
1784         case DID_TRANSPORT_DISRUPTED:
1785                 /*
1786                  * LLD/transport was disrupted during processing of the IO.
1787                  * The transport class is now blocked/blocking,
1788                  * and the transport will decide what to do with the IO
1789                  * based on its timers and recovery capablilities if
1790                  * there are enough retries.
1791                  */
1792                 goto maybe_retry;
1793         case DID_TRANSPORT_FAILFAST:
1794                 /*
1795                  * The transport decided to failfast the IO (most likely
1796                  * the fast io fail tmo fired), so send IO directly upwards.
1797                  */
1798                 return SUCCESS;
1799         case DID_ERROR:
1800                 if (msg_byte(scmd->result) == COMMAND_COMPLETE &&
1801                     status_byte(scmd->result) == RESERVATION_CONFLICT)
1802                         /*
1803                          * execute reservation conflict processing code
1804                          * lower down
1805                          */
1806                         break;
1807                 /* fallthrough */
1808         case DID_BUS_BUSY:
1809         case DID_PARITY:
1810                 goto maybe_retry;
1811         case DID_TIME_OUT:
1812                 /*
1813                  * when we scan the bus, we get timeout messages for
1814                  * these commands if there is no device available.
1815                  * other hosts report did_no_connect for the same thing.
1816                  */
1817                 if ((scmd->cmnd[0] == TEST_UNIT_READY ||
1818                      scmd->cmnd[0] == INQUIRY)) {
1819                         return SUCCESS;
1820                 } else {
1821                         return FAILED;
1822                 }
1823         case DID_RESET:
1824                 return SUCCESS;
1825         default:
1826                 return FAILED;
1827         }
1828
1829         /*
1830          * next, check the message byte.
1831          */
1832         if (msg_byte(scmd->result) != COMMAND_COMPLETE)
1833                 return FAILED;
1834
1835         /*
1836          * check the status byte to see if this indicates anything special.
1837          */
1838         switch (status_byte(scmd->result)) {
1839         case QUEUE_FULL:
1840                 scsi_handle_queue_full(scmd->device);
1841                 /*
1842                  * the case of trying to send too many commands to a
1843                  * tagged queueing device.
1844                  */
1845                 /* FALLTHROUGH */
1846         case BUSY:
1847                 /*
1848                  * device can't talk to us at the moment.  Should only
1849                  * occur (SAM-3) when the task queue is empty, so will cause
1850                  * the empty queue handling to trigger a stall in the
1851                  * device.
1852                  */
1853                 return ADD_TO_MLQUEUE;
1854         case GOOD:
1855                 if (scmd->cmnd[0] == REPORT_LUNS)
1856                         scmd->device->sdev_target->expecting_lun_change = 0;
1857                 scsi_handle_queue_ramp_up(scmd->device);
1858                 /* FALLTHROUGH */
1859         case COMMAND_TERMINATED:
1860                 return SUCCESS;
1861         case TASK_ABORTED:
1862                 goto maybe_retry;
1863         case CHECK_CONDITION:
1864                 rtn = scsi_check_sense(scmd);
1865                 if (rtn == NEEDS_RETRY)
1866                         goto maybe_retry;
1867                 /* if rtn == FAILED, we have no sense information;
1868                  * returning FAILED will wake the error handler thread
1869                  * to collect the sense and redo the decide
1870                  * disposition */
1871                 return rtn;
1872         case CONDITION_GOOD:
1873         case INTERMEDIATE_GOOD:
1874         case INTERMEDIATE_C_GOOD:
1875         case ACA_ACTIVE:
1876                 /*
1877                  * who knows?  FIXME(eric)
1878                  */
1879                 return SUCCESS;
1880
1881         case RESERVATION_CONFLICT:
1882                 sdev_printk(KERN_INFO, scmd->device,
1883                             "reservation conflict\n");
1884                 set_host_byte(scmd, DID_NEXUS_FAILURE);
1885                 return SUCCESS; /* causes immediate i/o error */
1886         default:
1887                 return FAILED;
1888         }
1889         return FAILED;
1890
1891       maybe_retry:
1892
1893         /* we requeue for retry because the error was retryable, and
1894          * the request was not marked fast fail.  Note that above,
1895          * even if the request is marked fast fail, we still requeue
1896          * for queue congestion conditions (QUEUE_FULL or BUSY) */
1897         if ((++scmd->retries) <= scmd->allowed
1898             && !scsi_noretry_cmd(scmd)) {
1899                 return NEEDS_RETRY;
1900         } else {
1901                 /*
1902                  * no more retries - report this one back to upper level.
1903                  */
1904                 return SUCCESS;
1905         }
1906 }
1907
1908 static void eh_lock_door_done(struct request *req, blk_status_t status)
1909 {
1910         __blk_put_request(req->q, req);
1911 }
1912
1913 /**
1914  * scsi_eh_lock_door - Prevent medium removal for the specified device
1915  * @sdev:       SCSI device to prevent medium removal
1916  *
1917  * Locking:
1918  *      We must be called from process context.
1919  *
1920  * Notes:
1921  *      We queue up an asynchronous "ALLOW MEDIUM REMOVAL" request on the
1922  *      head of the devices request queue, and continue.
1923  */
1924 static void scsi_eh_lock_door(struct scsi_device *sdev)
1925 {
1926         struct request *req;
1927         struct scsi_request *rq;
1928
1929         /*
1930          * blk_get_request with GFP_KERNEL (__GFP_RECLAIM) sleeps until a
1931          * request becomes available
1932          */
1933         req = blk_get_request(sdev->request_queue, REQ_OP_SCSI_IN, GFP_KERNEL);
1934         if (IS_ERR(req))
1935                 return;
1936         rq = scsi_req(req);
1937
1938         rq->cmd[0] = ALLOW_MEDIUM_REMOVAL;
1939         rq->cmd[1] = 0;
1940         rq->cmd[2] = 0;
1941         rq->cmd[3] = 0;
1942         rq->cmd[4] = SCSI_REMOVAL_PREVENT;
1943         rq->cmd[5] = 0;
1944         rq->cmd_len = COMMAND_SIZE(rq->cmd[0]);
1945
1946         req->rq_flags |= RQF_QUIET;
1947         req->timeout = 10 * HZ;
1948         rq->retries = 5;
1949
1950         blk_execute_rq_nowait(req->q, NULL, req, 1, eh_lock_door_done);
1951 }
1952
1953 /**
1954  * scsi_restart_operations - restart io operations to the specified host.
1955  * @shost:      Host we are restarting.
1956  *
1957  * Notes:
1958  *    When we entered the error handler, we blocked all further i/o to
1959  *    this device.  we need to 'reverse' this process.
1960  */
1961 static void scsi_restart_operations(struct Scsi_Host *shost)
1962 {
1963         struct scsi_device *sdev;
1964         unsigned long flags;
1965
1966         /*
1967          * If the door was locked, we need to insert a door lock request
1968          * onto the head of the SCSI request queue for the device.  There
1969          * is no point trying to lock the door of an off-line device.
1970          */
1971         shost_for_each_device(sdev, shost) {
1972                 if (scsi_device_online(sdev) && sdev->was_reset && sdev->locked) {
1973                         scsi_eh_lock_door(sdev);
1974                         sdev->was_reset = 0;
1975                 }
1976         }
1977
1978         /*
1979          * next free up anything directly waiting upon the host.  this
1980          * will be requests for character device operations, and also for
1981          * ioctls to queued block devices.
1982          */
1983         SCSI_LOG_ERROR_RECOVERY(3,
1984                 shost_printk(KERN_INFO, shost, "waking up host to restart\n"));
1985
1986         spin_lock_irqsave(shost->host_lock, flags);
1987         if (scsi_host_set_state(shost, SHOST_RUNNING))
1988                 if (scsi_host_set_state(shost, SHOST_CANCEL))
1989                         BUG_ON(scsi_host_set_state(shost, SHOST_DEL));
1990         spin_unlock_irqrestore(shost->host_lock, flags);
1991
1992         wake_up(&shost->host_wait);
1993
1994         /*
1995          * finally we need to re-initiate requests that may be pending.  we will
1996          * have had everything blocked while error handling is taking place, and
1997          * now that error recovery is done, we will need to ensure that these
1998          * requests are started.
1999          */
2000         scsi_run_host_queues(shost);
2001
2002         /*
2003          * if eh is active and host_eh_scheduled is pending we need to re-run
2004          * recovery.  we do this check after scsi_run_host_queues() to allow
2005          * everything pent up since the last eh run a chance to make forward
2006          * progress before we sync again.  Either we'll immediately re-run
2007          * recovery or scsi_device_unbusy() will wake us again when these
2008          * pending commands complete.
2009          */
2010         spin_lock_irqsave(shost->host_lock, flags);
2011         if (shost->host_eh_scheduled)
2012                 if (scsi_host_set_state(shost, SHOST_RECOVERY))
2013                         WARN_ON(scsi_host_set_state(shost, SHOST_CANCEL_RECOVERY));
2014         spin_unlock_irqrestore(shost->host_lock, flags);
2015 }
2016
2017 /**
2018  * scsi_eh_ready_devs - check device ready state and recover if not.
2019  * @shost:      host to be recovered.
2020  * @work_q:     &list_head for pending commands.
2021  * @done_q:     &list_head for processed commands.
2022  */
2023 void scsi_eh_ready_devs(struct Scsi_Host *shost,
2024                         struct list_head *work_q,
2025                         struct list_head *done_q)
2026 {
2027         if (!scsi_eh_stu(shost, work_q, done_q))
2028                 if (!scsi_eh_bus_device_reset(shost, work_q, done_q))
2029                         if (!scsi_eh_target_reset(shost, work_q, done_q))
2030                                 if (!scsi_eh_bus_reset(shost, work_q, done_q))
2031                                         if (!scsi_eh_host_reset(shost, work_q, done_q))
2032                                                 scsi_eh_offline_sdevs(work_q,
2033                                                                       done_q);
2034 }
2035 EXPORT_SYMBOL_GPL(scsi_eh_ready_devs);
2036
2037 /**
2038  * scsi_eh_flush_done_q - finish processed commands or retry them.
2039  * @done_q:     list_head of processed commands.
2040  */
2041 void scsi_eh_flush_done_q(struct list_head *done_q)
2042 {
2043         struct scsi_cmnd *scmd, *next;
2044
2045         list_for_each_entry_safe(scmd, next, done_q, eh_entry) {
2046                 list_del_init(&scmd->eh_entry);
2047                 if (scsi_device_online(scmd->device) &&
2048                     !scsi_noretry_cmd(scmd) &&
2049                     (++scmd->retries <= scmd->allowed)) {
2050                         SCSI_LOG_ERROR_RECOVERY(3,
2051                                 scmd_printk(KERN_INFO, scmd,
2052                                              "%s: flush retry cmd\n",
2053                                              current->comm));
2054                                 scsi_queue_insert(scmd, SCSI_MLQUEUE_EH_RETRY);
2055                 } else {
2056                         /*
2057                          * If just we got sense for the device (called
2058                          * scsi_eh_get_sense), scmd->result is already
2059                          * set, do not set DRIVER_TIMEOUT.
2060                          */
2061                         if (!scmd->result)
2062                                 scmd->result |= (DRIVER_TIMEOUT << 24);
2063                         SCSI_LOG_ERROR_RECOVERY(3,
2064                                 scmd_printk(KERN_INFO, scmd,
2065                                              "%s: flush finish cmd\n",
2066                                              current->comm));
2067                         scsi_finish_command(scmd);
2068                 }
2069         }
2070 }
2071 EXPORT_SYMBOL(scsi_eh_flush_done_q);
2072
2073 /**
2074  * scsi_unjam_host - Attempt to fix a host which has a cmd that failed.
2075  * @shost:      Host to unjam.
2076  *
2077  * Notes:
2078  *    When we come in here, we *know* that all commands on the bus have
2079  *    either completed, failed or timed out.  we also know that no further
2080  *    commands are being sent to the host, so things are relatively quiet
2081  *    and we have freedom to fiddle with things as we wish.
2082  *
2083  *    This is only the *default* implementation.  it is possible for
2084  *    individual drivers to supply their own version of this function, and
2085  *    if the maintainer wishes to do this, it is strongly suggested that
2086  *    this function be taken as a template and modified.  this function
2087  *    was designed to correctly handle problems for about 95% of the
2088  *    different cases out there, and it should always provide at least a
2089  *    reasonable amount of error recovery.
2090  *
2091  *    Any command marked 'failed' or 'timeout' must eventually have
2092  *    scsi_finish_cmd() called for it.  we do all of the retry stuff
2093  *    here, so when we restart the host after we return it should have an
2094  *    empty queue.
2095  */
2096 static void scsi_unjam_host(struct Scsi_Host *shost)
2097 {
2098         unsigned long flags;
2099         LIST_HEAD(eh_work_q);
2100         LIST_HEAD(eh_done_q);
2101
2102         spin_lock_irqsave(shost->host_lock, flags);
2103         list_splice_init(&shost->eh_cmd_q, &eh_work_q);
2104         spin_unlock_irqrestore(shost->host_lock, flags);
2105
2106         SCSI_LOG_ERROR_RECOVERY(1, scsi_eh_prt_fail_stats(shost, &eh_work_q));
2107
2108         if (!scsi_eh_get_sense(&eh_work_q, &eh_done_q))
2109                 scsi_eh_ready_devs(shost, &eh_work_q, &eh_done_q);
2110
2111         spin_lock_irqsave(shost->host_lock, flags);
2112         if (shost->eh_deadline != -1)
2113                 shost->last_reset = 0;
2114         spin_unlock_irqrestore(shost->host_lock, flags);
2115         scsi_eh_flush_done_q(&eh_done_q);
2116 }
2117
2118 /**
2119  * scsi_error_handler - SCSI error handler thread
2120  * @data:       Host for which we are running.
2121  *
2122  * Notes:
2123  *    This is the main error handling loop.  This is run as a kernel thread
2124  *    for every SCSI host and handles all error handling activity.
2125  */
2126 int scsi_error_handler(void *data)
2127 {
2128         struct Scsi_Host *shost = data;
2129
2130         /*
2131          * We use TASK_INTERRUPTIBLE so that the thread is not
2132          * counted against the load average as a running process.
2133          * We never actually get interrupted because kthread_run
2134          * disables signal delivery for the created thread.
2135          */
2136         while (true) {
2137                 /*
2138                  * The sequence in kthread_stop() sets the stop flag first
2139                  * then wakes the process.  To avoid missed wakeups, the task
2140                  * should always be in a non running state before the stop
2141                  * flag is checked
2142                  */
2143                 set_current_state(TASK_INTERRUPTIBLE);
2144                 if (kthread_should_stop())
2145                         break;
2146
2147                 if ((shost->host_failed == 0 && shost->host_eh_scheduled == 0) ||
2148                     shost->host_failed != atomic_read(&shost->host_busy)) {
2149                         SCSI_LOG_ERROR_RECOVERY(1,
2150                                 shost_printk(KERN_INFO, shost,
2151                                              "scsi_eh_%d: sleeping\n",
2152                                              shost->host_no));
2153                         schedule();
2154                         continue;
2155                 }
2156
2157                 __set_current_state(TASK_RUNNING);
2158                 SCSI_LOG_ERROR_RECOVERY(1,
2159                         shost_printk(KERN_INFO, shost,
2160                                      "scsi_eh_%d: waking up %d/%d/%d\n",
2161                                      shost->host_no, shost->host_eh_scheduled,
2162                                      shost->host_failed,
2163                                      atomic_read(&shost->host_busy)));
2164
2165                 /*
2166                  * We have a host that is failing for some reason.  Figure out
2167                  * what we need to do to get it up and online again (if we can).
2168                  * If we fail, we end up taking the thing offline.
2169                  */
2170                 if (!shost->eh_noresume && scsi_autopm_get_host(shost) != 0) {
2171                         SCSI_LOG_ERROR_RECOVERY(1,
2172                                 shost_printk(KERN_ERR, shost,
2173                                              "scsi_eh_%d: unable to autoresume\n",
2174                                              shost->host_no));
2175                         continue;
2176                 }
2177
2178                 if (shost->transportt->eh_strategy_handler)
2179                         shost->transportt->eh_strategy_handler(shost);
2180                 else
2181                         scsi_unjam_host(shost);
2182
2183                 /* All scmds have been handled */
2184                 shost->host_failed = 0;
2185
2186                 /*
2187                  * Note - if the above fails completely, the action is to take
2188                  * individual devices offline and flush the queue of any
2189                  * outstanding requests that may have been pending.  When we
2190                  * restart, we restart any I/O to any other devices on the bus
2191                  * which are still online.
2192                  */
2193                 scsi_restart_operations(shost);
2194                 if (!shost->eh_noresume)
2195                         scsi_autopm_put_host(shost);
2196         }
2197         __set_current_state(TASK_RUNNING);
2198
2199         SCSI_LOG_ERROR_RECOVERY(1,
2200                 shost_printk(KERN_INFO, shost,
2201                              "Error handler scsi_eh_%d exiting\n",
2202                              shost->host_no));
2203         shost->ehandler = NULL;
2204         return 0;
2205 }
2206
2207 /*
2208  * Function:    scsi_report_bus_reset()
2209  *
2210  * Purpose:     Utility function used by low-level drivers to report that
2211  *              they have observed a bus reset on the bus being handled.
2212  *
2213  * Arguments:   shost       - Host in question
2214  *              channel     - channel on which reset was observed.
2215  *
2216  * Returns:     Nothing
2217  *
2218  * Lock status: Host lock must be held.
2219  *
2220  * Notes:       This only needs to be called if the reset is one which
2221  *              originates from an unknown location.  Resets originated
2222  *              by the mid-level itself don't need to call this, but there
2223  *              should be no harm.
2224  *
2225  *              The main purpose of this is to make sure that a CHECK_CONDITION
2226  *              is properly treated.
2227  */
2228 void scsi_report_bus_reset(struct Scsi_Host *shost, int channel)
2229 {
2230         struct scsi_device *sdev;
2231
2232         __shost_for_each_device(sdev, shost) {
2233                 if (channel == sdev_channel(sdev))
2234                         __scsi_report_device_reset(sdev, NULL);
2235         }
2236 }
2237 EXPORT_SYMBOL(scsi_report_bus_reset);
2238
2239 /*
2240  * Function:    scsi_report_device_reset()
2241  *
2242  * Purpose:     Utility function used by low-level drivers to report that
2243  *              they have observed a device reset on the device being handled.
2244  *
2245  * Arguments:   shost       - Host in question
2246  *              channel     - channel on which reset was observed
2247  *              target      - target on which reset was observed
2248  *
2249  * Returns:     Nothing
2250  *
2251  * Lock status: Host lock must be held
2252  *
2253  * Notes:       This only needs to be called if the reset is one which
2254  *              originates from an unknown location.  Resets originated
2255  *              by the mid-level itself don't need to call this, but there
2256  *              should be no harm.
2257  *
2258  *              The main purpose of this is to make sure that a CHECK_CONDITION
2259  *              is properly treated.
2260  */
2261 void scsi_report_device_reset(struct Scsi_Host *shost, int channel, int target)
2262 {
2263         struct scsi_device *sdev;
2264
2265         __shost_for_each_device(sdev, shost) {
2266                 if (channel == sdev_channel(sdev) &&
2267                     target == sdev_id(sdev))
2268                         __scsi_report_device_reset(sdev, NULL);
2269         }
2270 }
2271 EXPORT_SYMBOL(scsi_report_device_reset);
2272
2273 static void
2274 scsi_reset_provider_done_command(struct scsi_cmnd *scmd)
2275 {
2276 }
2277
2278 /**
2279  * scsi_ioctl_reset: explicitly reset a host/bus/target/device
2280  * @dev:        scsi_device to operate on
2281  * @arg:        reset type (see sg.h)
2282  */
2283 int
2284 scsi_ioctl_reset(struct scsi_device *dev, int __user *arg)
2285 {
2286         struct scsi_cmnd *scmd;
2287         struct Scsi_Host *shost = dev->host;
2288         struct request *rq;
2289         unsigned long flags;
2290         int error = 0, rtn, val;
2291
2292         if (!capable(CAP_SYS_ADMIN) || !capable(CAP_SYS_RAWIO))
2293                 return -EACCES;
2294
2295         error = get_user(val, arg);
2296         if (error)
2297                 return error;
2298
2299         if (scsi_autopm_get_host(shost) < 0)
2300                 return -EIO;
2301
2302         error = -EIO;
2303         rq = kzalloc(sizeof(struct request) + sizeof(struct scsi_cmnd) +
2304                         shost->hostt->cmd_size, GFP_KERNEL);
2305         if (!rq)
2306                 goto out_put_autopm_host;
2307         blk_rq_init(NULL, rq);
2308
2309         scmd = (struct scsi_cmnd *)(rq + 1);
2310         scsi_init_command(dev, scmd);
2311         scmd->request = rq;
2312         scmd->cmnd = scsi_req(rq)->cmd;
2313
2314         scmd->scsi_done         = scsi_reset_provider_done_command;
2315         memset(&scmd->sdb, 0, sizeof(scmd->sdb));
2316
2317         scmd->cmd_len                   = 0;
2318
2319         scmd->sc_data_direction         = DMA_BIDIRECTIONAL;
2320
2321         spin_lock_irqsave(shost->host_lock, flags);
2322         shost->tmf_in_progress = 1;
2323         spin_unlock_irqrestore(shost->host_lock, flags);
2324
2325         switch (val & ~SG_SCSI_RESET_NO_ESCALATE) {
2326         case SG_SCSI_RESET_NOTHING:
2327                 rtn = SUCCESS;
2328                 break;
2329         case SG_SCSI_RESET_DEVICE:
2330                 rtn = scsi_try_bus_device_reset(scmd);
2331                 if (rtn == SUCCESS || (val & SG_SCSI_RESET_NO_ESCALATE))
2332                         break;
2333                 /* FALLTHROUGH */
2334         case SG_SCSI_RESET_TARGET:
2335                 rtn = scsi_try_target_reset(scmd);
2336                 if (rtn == SUCCESS || (val & SG_SCSI_RESET_NO_ESCALATE))
2337                         break;
2338                 /* FALLTHROUGH */
2339         case SG_SCSI_RESET_BUS:
2340                 rtn = scsi_try_bus_reset(scmd);
2341                 if (rtn == SUCCESS || (val & SG_SCSI_RESET_NO_ESCALATE))
2342                         break;
2343                 /* FALLTHROUGH */
2344         case SG_SCSI_RESET_HOST:
2345                 rtn = scsi_try_host_reset(scmd);
2346                 if (rtn == SUCCESS)
2347                         break;
2348                 /* FALLTHROUGH */
2349         default:
2350                 rtn = FAILED;
2351                 break;
2352         }
2353
2354         error = (rtn == SUCCESS) ? 0 : -EIO;
2355
2356         spin_lock_irqsave(shost->host_lock, flags);
2357         shost->tmf_in_progress = 0;
2358         spin_unlock_irqrestore(shost->host_lock, flags);
2359
2360         /*
2361          * be sure to wake up anyone who was sleeping or had their queue
2362          * suspended while we performed the TMF.
2363          */
2364         SCSI_LOG_ERROR_RECOVERY(3,
2365                 shost_printk(KERN_INFO, shost,
2366                              "waking up host to restart after TMF\n"));
2367
2368         wake_up(&shost->host_wait);
2369         scsi_run_host_queues(shost);
2370
2371         scsi_put_command(scmd);
2372         kfree(rq);
2373
2374 out_put_autopm_host:
2375         scsi_autopm_put_host(shost);
2376         return error;
2377 }
2378 EXPORT_SYMBOL(scsi_ioctl_reset);
2379
2380 bool scsi_command_normalize_sense(const struct scsi_cmnd *cmd,
2381                                   struct scsi_sense_hdr *sshdr)
2382 {
2383         return scsi_normalize_sense(cmd->sense_buffer,
2384                         SCSI_SENSE_BUFFERSIZE, sshdr);
2385 }
2386 EXPORT_SYMBOL(scsi_command_normalize_sense);
2387
2388 /**
2389  * scsi_get_sense_info_fld - get information field from sense data (either fixed or descriptor format)
2390  * @sense_buffer:       byte array of sense data
2391  * @sb_len:             number of valid bytes in sense_buffer
2392  * @info_out:           pointer to 64 integer where 8 or 4 byte information
2393  *                      field will be placed if found.
2394  *
2395  * Return value:
2396  *      true if information field found, false if not found.
2397  */
2398 bool scsi_get_sense_info_fld(const u8 *sense_buffer, int sb_len,
2399                              u64 *info_out)
2400 {
2401         const u8 * ucp;
2402
2403         if (sb_len < 7)
2404                 return false;
2405         switch (sense_buffer[0] & 0x7f) {
2406         case 0x70:
2407         case 0x71:
2408                 if (sense_buffer[0] & 0x80) {
2409                         *info_out = get_unaligned_be32(&sense_buffer[3]);
2410                         return true;
2411                 }
2412                 return false;
2413         case 0x72:
2414         case 0x73:
2415                 ucp = scsi_sense_desc_find(sense_buffer, sb_len,
2416                                            0 /* info desc */);
2417                 if (ucp && (0xa == ucp[1])) {
2418                         *info_out = get_unaligned_be64(&ucp[4]);
2419                         return true;
2420                 }
2421                 return false;
2422         default:
2423                 return false;
2424         }
2425 }
2426 EXPORT_SYMBOL(scsi_get_sense_info_fld);