scsi: Suppress a kernel warning in case the prep function returns BLKPREP_DEFER
[sfrench/cifs-2.6.git] / drivers / scsi / scsi_lib.c
1 /*
2  * Copyright (C) 1999 Eric Youngdale
3  * Copyright (C) 2014 Christoph Hellwig
4  *
5  *  SCSI queueing library.
6  *      Initial versions: Eric Youngdale (eric@andante.org).
7  *                        Based upon conversations with large numbers
8  *                        of people at Linux Expo.
9  */
10
11 #include <linux/bio.h>
12 #include <linux/bitops.h>
13 #include <linux/blkdev.h>
14 #include <linux/completion.h>
15 #include <linux/kernel.h>
16 #include <linux/export.h>
17 #include <linux/init.h>
18 #include <linux/pci.h>
19 #include <linux/delay.h>
20 #include <linux/hardirq.h>
21 #include <linux/scatterlist.h>
22 #include <linux/blk-mq.h>
23 #include <linux/ratelimit.h>
24 #include <asm/unaligned.h>
25
26 #include <scsi/scsi.h>
27 #include <scsi/scsi_cmnd.h>
28 #include <scsi/scsi_dbg.h>
29 #include <scsi/scsi_device.h>
30 #include <scsi/scsi_driver.h>
31 #include <scsi/scsi_eh.h>
32 #include <scsi/scsi_host.h>
33 #include <scsi/scsi_transport.h> /* __scsi_init_queue() */
34 #include <scsi/scsi_dh.h>
35
36 #include <trace/events/scsi.h>
37
38 #include "scsi_debugfs.h"
39 #include "scsi_priv.h"
40 #include "scsi_logging.h"
41
42 static struct kmem_cache *scsi_sdb_cache;
43 static struct kmem_cache *scsi_sense_cache;
44 static struct kmem_cache *scsi_sense_isadma_cache;
45 static DEFINE_MUTEX(scsi_sense_cache_mutex);
46
47 static void scsi_mq_uninit_cmd(struct scsi_cmnd *cmd);
48
49 static inline struct kmem_cache *
50 scsi_select_sense_cache(bool unchecked_isa_dma)
51 {
52         return unchecked_isa_dma ? scsi_sense_isadma_cache : scsi_sense_cache;
53 }
54
55 static void scsi_free_sense_buffer(bool unchecked_isa_dma,
56                                    unsigned char *sense_buffer)
57 {
58         kmem_cache_free(scsi_select_sense_cache(unchecked_isa_dma),
59                         sense_buffer);
60 }
61
62 static unsigned char *scsi_alloc_sense_buffer(bool unchecked_isa_dma,
63         gfp_t gfp_mask, int numa_node)
64 {
65         return kmem_cache_alloc_node(scsi_select_sense_cache(unchecked_isa_dma),
66                                      gfp_mask, numa_node);
67 }
68
69 int scsi_init_sense_cache(struct Scsi_Host *shost)
70 {
71         struct kmem_cache *cache;
72         int ret = 0;
73
74         cache = scsi_select_sense_cache(shost->unchecked_isa_dma);
75         if (cache)
76                 return 0;
77
78         mutex_lock(&scsi_sense_cache_mutex);
79         if (shost->unchecked_isa_dma) {
80                 scsi_sense_isadma_cache =
81                         kmem_cache_create("scsi_sense_cache(DMA)",
82                         SCSI_SENSE_BUFFERSIZE, 0,
83                         SLAB_HWCACHE_ALIGN | SLAB_CACHE_DMA, NULL);
84                 if (!scsi_sense_isadma_cache)
85                         ret = -ENOMEM;
86         } else {
87                 scsi_sense_cache =
88                         kmem_cache_create("scsi_sense_cache",
89                         SCSI_SENSE_BUFFERSIZE, 0, SLAB_HWCACHE_ALIGN, NULL);
90                 if (!scsi_sense_cache)
91                         ret = -ENOMEM;
92         }
93
94         mutex_unlock(&scsi_sense_cache_mutex);
95         return ret;
96 }
97
98 /*
99  * When to reinvoke queueing after a resource shortage. It's 3 msecs to
100  * not change behaviour from the previous unplug mechanism, experimentation
101  * may prove this needs changing.
102  */
103 #define SCSI_QUEUE_DELAY        3
104
105 static void
106 scsi_set_blocked(struct scsi_cmnd *cmd, int reason)
107 {
108         struct Scsi_Host *host = cmd->device->host;
109         struct scsi_device *device = cmd->device;
110         struct scsi_target *starget = scsi_target(device);
111
112         /*
113          * Set the appropriate busy bit for the device/host.
114          *
115          * If the host/device isn't busy, assume that something actually
116          * completed, and that we should be able to queue a command now.
117          *
118          * Note that the prior mid-layer assumption that any host could
119          * always queue at least one command is now broken.  The mid-layer
120          * will implement a user specifiable stall (see
121          * scsi_host.max_host_blocked and scsi_device.max_device_blocked)
122          * if a command is requeued with no other commands outstanding
123          * either for the device or for the host.
124          */
125         switch (reason) {
126         case SCSI_MLQUEUE_HOST_BUSY:
127                 atomic_set(&host->host_blocked, host->max_host_blocked);
128                 break;
129         case SCSI_MLQUEUE_DEVICE_BUSY:
130         case SCSI_MLQUEUE_EH_RETRY:
131                 atomic_set(&device->device_blocked,
132                            device->max_device_blocked);
133                 break;
134         case SCSI_MLQUEUE_TARGET_BUSY:
135                 atomic_set(&starget->target_blocked,
136                            starget->max_target_blocked);
137                 break;
138         }
139 }
140
141 static void scsi_mq_requeue_cmd(struct scsi_cmnd *cmd)
142 {
143         struct scsi_device *sdev = cmd->device;
144
145         if (cmd->request->rq_flags & RQF_DONTPREP) {
146                 cmd->request->rq_flags &= ~RQF_DONTPREP;
147                 scsi_mq_uninit_cmd(cmd);
148         } else {
149                 WARN_ON_ONCE(true);
150         }
151         blk_mq_requeue_request(cmd->request, true);
152         put_device(&sdev->sdev_gendev);
153 }
154
155 /**
156  * __scsi_queue_insert - private queue insertion
157  * @cmd: The SCSI command being requeued
158  * @reason:  The reason for the requeue
159  * @unbusy: Whether the queue should be unbusied
160  *
161  * This is a private queue insertion.  The public interface
162  * scsi_queue_insert() always assumes the queue should be unbusied
163  * because it's always called before the completion.  This function is
164  * for a requeue after completion, which should only occur in this
165  * file.
166  */
167 static void __scsi_queue_insert(struct scsi_cmnd *cmd, int reason, int unbusy)
168 {
169         struct scsi_device *device = cmd->device;
170         struct request_queue *q = device->request_queue;
171         unsigned long flags;
172
173         SCSI_LOG_MLQUEUE(1, scmd_printk(KERN_INFO, cmd,
174                 "Inserting command %p into mlqueue\n", cmd));
175
176         scsi_set_blocked(cmd, reason);
177
178         /*
179          * Decrement the counters, since these commands are no longer
180          * active on the host/device.
181          */
182         if (unbusy)
183                 scsi_device_unbusy(device);
184
185         /*
186          * Requeue this command.  It will go before all other commands
187          * that are already in the queue. Schedule requeue work under
188          * lock such that the kblockd_schedule_work() call happens
189          * before blk_cleanup_queue() finishes.
190          */
191         cmd->result = 0;
192         if (q->mq_ops) {
193                 scsi_mq_requeue_cmd(cmd);
194                 return;
195         }
196         spin_lock_irqsave(q->queue_lock, flags);
197         blk_requeue_request(q, cmd->request);
198         kblockd_schedule_work(&device->requeue_work);
199         spin_unlock_irqrestore(q->queue_lock, flags);
200 }
201
202 /*
203  * Function:    scsi_queue_insert()
204  *
205  * Purpose:     Insert a command in the midlevel queue.
206  *
207  * Arguments:   cmd    - command that we are adding to queue.
208  *              reason - why we are inserting command to queue.
209  *
210  * Lock status: Assumed that lock is not held upon entry.
211  *
212  * Returns:     Nothing.
213  *
214  * Notes:       We do this for one of two cases.  Either the host is busy
215  *              and it cannot accept any more commands for the time being,
216  *              or the device returned QUEUE_FULL and can accept no more
217  *              commands.
218  * Notes:       This could be called either from an interrupt context or a
219  *              normal process context.
220  */
221 void scsi_queue_insert(struct scsi_cmnd *cmd, int reason)
222 {
223         __scsi_queue_insert(cmd, reason, 1);
224 }
225
226
227 /**
228  * scsi_execute - insert request and wait for the result
229  * @sdev:       scsi device
230  * @cmd:        scsi command
231  * @data_direction: data direction
232  * @buffer:     data buffer
233  * @bufflen:    len of buffer
234  * @sense:      optional sense buffer
235  * @sshdr:      optional decoded sense header
236  * @timeout:    request timeout in seconds
237  * @retries:    number of times to retry request
238  * @flags:      flags for ->cmd_flags
239  * @rq_flags:   flags for ->rq_flags
240  * @resid:      optional residual length
241  *
242  * Returns the scsi_cmnd result field if a command was executed, or a negative
243  * Linux error code if we didn't get that far.
244  */
245 int scsi_execute(struct scsi_device *sdev, const unsigned char *cmd,
246                  int data_direction, void *buffer, unsigned bufflen,
247                  unsigned char *sense, struct scsi_sense_hdr *sshdr,
248                  int timeout, int retries, u64 flags, req_flags_t rq_flags,
249                  int *resid)
250 {
251         struct request *req;
252         struct scsi_request *rq;
253         int ret = DRIVER_ERROR << 24;
254
255         req = blk_get_request(sdev->request_queue,
256                         data_direction == DMA_TO_DEVICE ?
257                         REQ_OP_SCSI_OUT : REQ_OP_SCSI_IN, __GFP_RECLAIM);
258         if (IS_ERR(req))
259                 return ret;
260         rq = scsi_req(req);
261
262         if (bufflen &&  blk_rq_map_kern(sdev->request_queue, req,
263                                         buffer, bufflen, __GFP_RECLAIM))
264                 goto out;
265
266         rq->cmd_len = COMMAND_SIZE(cmd[0]);
267         memcpy(rq->cmd, cmd, rq->cmd_len);
268         rq->retries = retries;
269         req->timeout = timeout;
270         req->cmd_flags |= flags;
271         req->rq_flags |= rq_flags | RQF_QUIET | RQF_PREEMPT;
272
273         /*
274          * head injection *required* here otherwise quiesce won't work
275          */
276         blk_execute_rq(req->q, NULL, req, 1);
277
278         /*
279          * Some devices (USB mass-storage in particular) may transfer
280          * garbage data together with a residue indicating that the data
281          * is invalid.  Prevent the garbage from being misinterpreted
282          * and prevent security leaks by zeroing out the excess data.
283          */
284         if (unlikely(rq->resid_len > 0 && rq->resid_len <= bufflen))
285                 memset(buffer + (bufflen - rq->resid_len), 0, rq->resid_len);
286
287         if (resid)
288                 *resid = rq->resid_len;
289         if (sense && rq->sense_len)
290                 memcpy(sense, rq->sense, SCSI_SENSE_BUFFERSIZE);
291         if (sshdr)
292                 scsi_normalize_sense(rq->sense, rq->sense_len, sshdr);
293         ret = rq->result;
294  out:
295         blk_put_request(req);
296
297         return ret;
298 }
299 EXPORT_SYMBOL(scsi_execute);
300
301 /*
302  * Function:    scsi_init_cmd_errh()
303  *
304  * Purpose:     Initialize cmd fields related to error handling.
305  *
306  * Arguments:   cmd     - command that is ready to be queued.
307  *
308  * Notes:       This function has the job of initializing a number of
309  *              fields related to error handling.   Typically this will
310  *              be called once for each command, as required.
311  */
312 static void scsi_init_cmd_errh(struct scsi_cmnd *cmd)
313 {
314         cmd->serial_number = 0;
315         scsi_set_resid(cmd, 0);
316         memset(cmd->sense_buffer, 0, SCSI_SENSE_BUFFERSIZE);
317         if (cmd->cmd_len == 0)
318                 cmd->cmd_len = scsi_command_size(cmd->cmnd);
319 }
320
321 void scsi_device_unbusy(struct scsi_device *sdev)
322 {
323         struct Scsi_Host *shost = sdev->host;
324         struct scsi_target *starget = scsi_target(sdev);
325         unsigned long flags;
326
327         atomic_dec(&shost->host_busy);
328         if (starget->can_queue > 0)
329                 atomic_dec(&starget->target_busy);
330
331         if (unlikely(scsi_host_in_recovery(shost) &&
332                      (shost->host_failed || shost->host_eh_scheduled))) {
333                 spin_lock_irqsave(shost->host_lock, flags);
334                 scsi_eh_wakeup(shost);
335                 spin_unlock_irqrestore(shost->host_lock, flags);
336         }
337
338         atomic_dec(&sdev->device_busy);
339 }
340
341 static void scsi_kick_queue(struct request_queue *q)
342 {
343         if (q->mq_ops)
344                 blk_mq_start_hw_queues(q);
345         else
346                 blk_run_queue(q);
347 }
348
349 /*
350  * Called for single_lun devices on IO completion. Clear starget_sdev_user,
351  * and call blk_run_queue for all the scsi_devices on the target -
352  * including current_sdev first.
353  *
354  * Called with *no* scsi locks held.
355  */
356 static void scsi_single_lun_run(struct scsi_device *current_sdev)
357 {
358         struct Scsi_Host *shost = current_sdev->host;
359         struct scsi_device *sdev, *tmp;
360         struct scsi_target *starget = scsi_target(current_sdev);
361         unsigned long flags;
362
363         spin_lock_irqsave(shost->host_lock, flags);
364         starget->starget_sdev_user = NULL;
365         spin_unlock_irqrestore(shost->host_lock, flags);
366
367         /*
368          * Call blk_run_queue for all LUNs on the target, starting with
369          * current_sdev. We race with others (to set starget_sdev_user),
370          * but in most cases, we will be first. Ideally, each LU on the
371          * target would get some limited time or requests on the target.
372          */
373         scsi_kick_queue(current_sdev->request_queue);
374
375         spin_lock_irqsave(shost->host_lock, flags);
376         if (starget->starget_sdev_user)
377                 goto out;
378         list_for_each_entry_safe(sdev, tmp, &starget->devices,
379                         same_target_siblings) {
380                 if (sdev == current_sdev)
381                         continue;
382                 if (scsi_device_get(sdev))
383                         continue;
384
385                 spin_unlock_irqrestore(shost->host_lock, flags);
386                 scsi_kick_queue(sdev->request_queue);
387                 spin_lock_irqsave(shost->host_lock, flags);
388         
389                 scsi_device_put(sdev);
390         }
391  out:
392         spin_unlock_irqrestore(shost->host_lock, flags);
393 }
394
395 static inline bool scsi_device_is_busy(struct scsi_device *sdev)
396 {
397         if (atomic_read(&sdev->device_busy) >= sdev->queue_depth)
398                 return true;
399         if (atomic_read(&sdev->device_blocked) > 0)
400                 return true;
401         return false;
402 }
403
404 static inline bool scsi_target_is_busy(struct scsi_target *starget)
405 {
406         if (starget->can_queue > 0) {
407                 if (atomic_read(&starget->target_busy) >= starget->can_queue)
408                         return true;
409                 if (atomic_read(&starget->target_blocked) > 0)
410                         return true;
411         }
412         return false;
413 }
414
415 static inline bool scsi_host_is_busy(struct Scsi_Host *shost)
416 {
417         if (shost->can_queue > 0 &&
418             atomic_read(&shost->host_busy) >= shost->can_queue)
419                 return true;
420         if (atomic_read(&shost->host_blocked) > 0)
421                 return true;
422         if (shost->host_self_blocked)
423                 return true;
424         return false;
425 }
426
427 static void scsi_starved_list_run(struct Scsi_Host *shost)
428 {
429         LIST_HEAD(starved_list);
430         struct scsi_device *sdev;
431         unsigned long flags;
432
433         spin_lock_irqsave(shost->host_lock, flags);
434         list_splice_init(&shost->starved_list, &starved_list);
435
436         while (!list_empty(&starved_list)) {
437                 struct request_queue *slq;
438
439                 /*
440                  * As long as shost is accepting commands and we have
441                  * starved queues, call blk_run_queue. scsi_request_fn
442                  * drops the queue_lock and can add us back to the
443                  * starved_list.
444                  *
445                  * host_lock protects the starved_list and starved_entry.
446                  * scsi_request_fn must get the host_lock before checking
447                  * or modifying starved_list or starved_entry.
448                  */
449                 if (scsi_host_is_busy(shost))
450                         break;
451
452                 sdev = list_entry(starved_list.next,
453                                   struct scsi_device, starved_entry);
454                 list_del_init(&sdev->starved_entry);
455                 if (scsi_target_is_busy(scsi_target(sdev))) {
456                         list_move_tail(&sdev->starved_entry,
457                                        &shost->starved_list);
458                         continue;
459                 }
460
461                 /*
462                  * Once we drop the host lock, a racing scsi_remove_device()
463                  * call may remove the sdev from the starved list and destroy
464                  * it and the queue.  Mitigate by taking a reference to the
465                  * queue and never touching the sdev again after we drop the
466                  * host lock.  Note: if __scsi_remove_device() invokes
467                  * blk_cleanup_queue() before the queue is run from this
468                  * function then blk_run_queue() will return immediately since
469                  * blk_cleanup_queue() marks the queue with QUEUE_FLAG_DYING.
470                  */
471                 slq = sdev->request_queue;
472                 if (!blk_get_queue(slq))
473                         continue;
474                 spin_unlock_irqrestore(shost->host_lock, flags);
475
476                 scsi_kick_queue(slq);
477                 blk_put_queue(slq);
478
479                 spin_lock_irqsave(shost->host_lock, flags);
480         }
481         /* put any unprocessed entries back */
482         list_splice(&starved_list, &shost->starved_list);
483         spin_unlock_irqrestore(shost->host_lock, flags);
484 }
485
486 /*
487  * Function:   scsi_run_queue()
488  *
489  * Purpose:    Select a proper request queue to serve next
490  *
491  * Arguments:  q       - last request's queue
492  *
493  * Returns:     Nothing
494  *
495  * Notes:      The previous command was completely finished, start
496  *             a new one if possible.
497  */
498 static void scsi_run_queue(struct request_queue *q)
499 {
500         struct scsi_device *sdev = q->queuedata;
501
502         if (scsi_target(sdev)->single_lun)
503                 scsi_single_lun_run(sdev);
504         if (!list_empty(&sdev->host->starved_list))
505                 scsi_starved_list_run(sdev->host);
506
507         if (q->mq_ops)
508                 blk_mq_run_hw_queues(q, false);
509         else
510                 blk_run_queue(q);
511 }
512
513 void scsi_requeue_run_queue(struct work_struct *work)
514 {
515         struct scsi_device *sdev;
516         struct request_queue *q;
517
518         sdev = container_of(work, struct scsi_device, requeue_work);
519         q = sdev->request_queue;
520         scsi_run_queue(q);
521 }
522
523 /*
524  * Function:    scsi_requeue_command()
525  *
526  * Purpose:     Handle post-processing of completed commands.
527  *
528  * Arguments:   q       - queue to operate on
529  *              cmd     - command that may need to be requeued.
530  *
531  * Returns:     Nothing
532  *
533  * Notes:       After command completion, there may be blocks left
534  *              over which weren't finished by the previous command
535  *              this can be for a number of reasons - the main one is
536  *              I/O errors in the middle of the request, in which case
537  *              we need to request the blocks that come after the bad
538  *              sector.
539  * Notes:       Upon return, cmd is a stale pointer.
540  */
541 static void scsi_requeue_command(struct request_queue *q, struct scsi_cmnd *cmd)
542 {
543         struct scsi_device *sdev = cmd->device;
544         struct request *req = cmd->request;
545         unsigned long flags;
546
547         spin_lock_irqsave(q->queue_lock, flags);
548         blk_unprep_request(req);
549         req->special = NULL;
550         scsi_put_command(cmd);
551         blk_requeue_request(q, req);
552         spin_unlock_irqrestore(q->queue_lock, flags);
553
554         scsi_run_queue(q);
555
556         put_device(&sdev->sdev_gendev);
557 }
558
559 void scsi_run_host_queues(struct Scsi_Host *shost)
560 {
561         struct scsi_device *sdev;
562
563         shost_for_each_device(sdev, shost)
564                 scsi_run_queue(sdev->request_queue);
565 }
566
567 static void scsi_uninit_cmd(struct scsi_cmnd *cmd)
568 {
569         if (!blk_rq_is_passthrough(cmd->request)) {
570                 struct scsi_driver *drv = scsi_cmd_to_driver(cmd);
571
572                 if (drv->uninit_command)
573                         drv->uninit_command(cmd);
574         }
575 }
576
577 static void scsi_mq_free_sgtables(struct scsi_cmnd *cmd)
578 {
579         struct scsi_data_buffer *sdb;
580
581         if (cmd->sdb.table.nents)
582                 sg_free_table_chained(&cmd->sdb.table, true);
583         if (cmd->request->next_rq) {
584                 sdb = cmd->request->next_rq->special;
585                 if (sdb)
586                         sg_free_table_chained(&sdb->table, true);
587         }
588         if (scsi_prot_sg_count(cmd))
589                 sg_free_table_chained(&cmd->prot_sdb->table, true);
590 }
591
592 static void scsi_mq_uninit_cmd(struct scsi_cmnd *cmd)
593 {
594         scsi_mq_free_sgtables(cmd);
595         scsi_uninit_cmd(cmd);
596         scsi_del_cmd_from_list(cmd);
597 }
598
599 /*
600  * Function:    scsi_release_buffers()
601  *
602  * Purpose:     Free resources allocate for a scsi_command.
603  *
604  * Arguments:   cmd     - command that we are bailing.
605  *
606  * Lock status: Assumed that no lock is held upon entry.
607  *
608  * Returns:     Nothing
609  *
610  * Notes:       In the event that an upper level driver rejects a
611  *              command, we must release resources allocated during
612  *              the __init_io() function.  Primarily this would involve
613  *              the scatter-gather table.
614  */
615 static void scsi_release_buffers(struct scsi_cmnd *cmd)
616 {
617         if (cmd->sdb.table.nents)
618                 sg_free_table_chained(&cmd->sdb.table, false);
619
620         memset(&cmd->sdb, 0, sizeof(cmd->sdb));
621
622         if (scsi_prot_sg_count(cmd))
623                 sg_free_table_chained(&cmd->prot_sdb->table, false);
624 }
625
626 static void scsi_release_bidi_buffers(struct scsi_cmnd *cmd)
627 {
628         struct scsi_data_buffer *bidi_sdb = cmd->request->next_rq->special;
629
630         sg_free_table_chained(&bidi_sdb->table, false);
631         kmem_cache_free(scsi_sdb_cache, bidi_sdb);
632         cmd->request->next_rq->special = NULL;
633 }
634
635 static bool scsi_end_request(struct request *req, blk_status_t error,
636                 unsigned int bytes, unsigned int bidi_bytes)
637 {
638         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
639         struct scsi_device *sdev = cmd->device;
640         struct request_queue *q = sdev->request_queue;
641
642         if (blk_update_request(req, error, bytes))
643                 return true;
644
645         /* Bidi request must be completed as a whole */
646         if (unlikely(bidi_bytes) &&
647             blk_update_request(req->next_rq, error, bidi_bytes))
648                 return true;
649
650         if (blk_queue_add_random(q))
651                 add_disk_randomness(req->rq_disk);
652
653         if (!blk_rq_is_scsi(req)) {
654                 WARN_ON_ONCE(!(cmd->flags & SCMD_INITIALIZED));
655                 cmd->flags &= ~SCMD_INITIALIZED;
656         }
657
658         if (req->mq_ctx) {
659                 /*
660                  * In the MQ case the command gets freed by __blk_mq_end_request,
661                  * so we have to do all cleanup that depends on it earlier.
662                  *
663                  * We also can't kick the queues from irq context, so we
664                  * will have to defer it to a workqueue.
665                  */
666                 scsi_mq_uninit_cmd(cmd);
667
668                 __blk_mq_end_request(req, error);
669
670                 if (scsi_target(sdev)->single_lun ||
671                     !list_empty(&sdev->host->starved_list))
672                         kblockd_schedule_work(&sdev->requeue_work);
673                 else
674                         blk_mq_run_hw_queues(q, true);
675         } else {
676                 unsigned long flags;
677
678                 if (bidi_bytes)
679                         scsi_release_bidi_buffers(cmd);
680                 scsi_release_buffers(cmd);
681                 scsi_put_command(cmd);
682
683                 spin_lock_irqsave(q->queue_lock, flags);
684                 blk_finish_request(req, error);
685                 spin_unlock_irqrestore(q->queue_lock, flags);
686
687                 scsi_run_queue(q);
688         }
689
690         put_device(&sdev->sdev_gendev);
691         return false;
692 }
693
694 /**
695  * __scsi_error_from_host_byte - translate SCSI error code into errno
696  * @cmd:        SCSI command (unused)
697  * @result:     scsi error code
698  *
699  * Translate SCSI error code into block errors.
700  */
701 static blk_status_t __scsi_error_from_host_byte(struct scsi_cmnd *cmd,
702                 int result)
703 {
704         switch (host_byte(result)) {
705         case DID_TRANSPORT_FAILFAST:
706                 return BLK_STS_TRANSPORT;
707         case DID_TARGET_FAILURE:
708                 set_host_byte(cmd, DID_OK);
709                 return BLK_STS_TARGET;
710         case DID_NEXUS_FAILURE:
711                 return BLK_STS_NEXUS;
712         case DID_ALLOC_FAILURE:
713                 set_host_byte(cmd, DID_OK);
714                 return BLK_STS_NOSPC;
715         case DID_MEDIUM_ERROR:
716                 set_host_byte(cmd, DID_OK);
717                 return BLK_STS_MEDIUM;
718         default:
719                 return BLK_STS_IOERR;
720         }
721 }
722
723 /*
724  * Function:    scsi_io_completion()
725  *
726  * Purpose:     Completion processing for block device I/O requests.
727  *
728  * Arguments:   cmd   - command that is finished.
729  *
730  * Lock status: Assumed that no lock is held upon entry.
731  *
732  * Returns:     Nothing
733  *
734  * Notes:       We will finish off the specified number of sectors.  If we
735  *              are done, the command block will be released and the queue
736  *              function will be goosed.  If we are not done then we have to
737  *              figure out what to do next:
738  *
739  *              a) We can call scsi_requeue_command().  The request
740  *                 will be unprepared and put back on the queue.  Then
741  *                 a new command will be created for it.  This should
742  *                 be used if we made forward progress, or if we want
743  *                 to switch from READ(10) to READ(6) for example.
744  *
745  *              b) We can call __scsi_queue_insert().  The request will
746  *                 be put back on the queue and retried using the same
747  *                 command as before, possibly after a delay.
748  *
749  *              c) We can call scsi_end_request() with -EIO to fail
750  *                 the remainder of the request.
751  */
752 void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes)
753 {
754         int result = cmd->result;
755         struct request_queue *q = cmd->device->request_queue;
756         struct request *req = cmd->request;
757         blk_status_t error = BLK_STS_OK;
758         struct scsi_sense_hdr sshdr;
759         bool sense_valid = false;
760         int sense_deferred = 0, level = 0;
761         enum {ACTION_FAIL, ACTION_REPREP, ACTION_RETRY,
762               ACTION_DELAYED_RETRY} action;
763         unsigned long wait_for = (cmd->allowed + 1) * req->timeout;
764
765         if (result) {
766                 sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
767                 if (sense_valid)
768                         sense_deferred = scsi_sense_is_deferred(&sshdr);
769         }
770
771         if (blk_rq_is_passthrough(req)) {
772                 if (result) {
773                         if (sense_valid) {
774                                 /*
775                                  * SG_IO wants current and deferred errors
776                                  */
777                                 scsi_req(req)->sense_len =
778                                         min(8 + cmd->sense_buffer[7],
779                                             SCSI_SENSE_BUFFERSIZE);
780                         }
781                         if (!sense_deferred)
782                                 error = __scsi_error_from_host_byte(cmd, result);
783                 }
784                 /*
785                  * __scsi_error_from_host_byte may have reset the host_byte
786                  */
787                 scsi_req(req)->result = cmd->result;
788                 scsi_req(req)->resid_len = scsi_get_resid(cmd);
789
790                 if (scsi_bidi_cmnd(cmd)) {
791                         /*
792                          * Bidi commands Must be complete as a whole,
793                          * both sides at once.
794                          */
795                         scsi_req(req->next_rq)->resid_len = scsi_in(cmd)->resid;
796                         if (scsi_end_request(req, BLK_STS_OK, blk_rq_bytes(req),
797                                         blk_rq_bytes(req->next_rq)))
798                                 BUG();
799                         return;
800                 }
801         } else if (blk_rq_bytes(req) == 0 && result && !sense_deferred) {
802                 /*
803                  * Flush commands do not transfers any data, and thus cannot use
804                  * good_bytes != blk_rq_bytes(req) as the signal for an error.
805                  * This sets the error explicitly for the problem case.
806                  */
807                 error = __scsi_error_from_host_byte(cmd, result);
808         }
809
810         /* no bidi support for !blk_rq_is_passthrough yet */
811         BUG_ON(blk_bidi_rq(req));
812
813         /*
814          * Next deal with any sectors which we were able to correctly
815          * handle.
816          */
817         SCSI_LOG_HLCOMPLETE(1, scmd_printk(KERN_INFO, cmd,
818                 "%u sectors total, %d bytes done.\n",
819                 blk_rq_sectors(req), good_bytes));
820
821         /*
822          * Recovered errors need reporting, but they're always treated as
823          * success, so fiddle the result code here.  For passthrough requests
824          * we already took a copy of the original into sreq->result which
825          * is what gets returned to the user
826          */
827         if (sense_valid && (sshdr.sense_key == RECOVERED_ERROR)) {
828                 /* if ATA PASS-THROUGH INFORMATION AVAILABLE skip
829                  * print since caller wants ATA registers. Only occurs on
830                  * SCSI ATA PASS_THROUGH commands when CK_COND=1
831                  */
832                 if ((sshdr.asc == 0x0) && (sshdr.ascq == 0x1d))
833                         ;
834                 else if (!(req->rq_flags & RQF_QUIET))
835                         scsi_print_sense(cmd);
836                 result = 0;
837                 /* for passthrough error may be set */
838                 error = BLK_STS_OK;
839         }
840
841         /*
842          * special case: failed zero length commands always need to
843          * drop down into the retry code. Otherwise, if we finished
844          * all bytes in the request we are done now.
845          */
846         if (!(blk_rq_bytes(req) == 0 && error) &&
847             !scsi_end_request(req, error, good_bytes, 0))
848                 return;
849
850         /*
851          * Kill remainder if no retrys.
852          */
853         if (error && scsi_noretry_cmd(cmd)) {
854                 if (scsi_end_request(req, error, blk_rq_bytes(req), 0))
855                         BUG();
856                 return;
857         }
858
859         /*
860          * If there had been no error, but we have leftover bytes in the
861          * requeues just queue the command up again.
862          */
863         if (result == 0)
864                 goto requeue;
865
866         error = __scsi_error_from_host_byte(cmd, result);
867
868         if (host_byte(result) == DID_RESET) {
869                 /* Third party bus reset or reset for error recovery
870                  * reasons.  Just retry the command and see what
871                  * happens.
872                  */
873                 action = ACTION_RETRY;
874         } else if (sense_valid && !sense_deferred) {
875                 switch (sshdr.sense_key) {
876                 case UNIT_ATTENTION:
877                         if (cmd->device->removable) {
878                                 /* Detected disc change.  Set a bit
879                                  * and quietly refuse further access.
880                                  */
881                                 cmd->device->changed = 1;
882                                 action = ACTION_FAIL;
883                         } else {
884                                 /* Must have been a power glitch, or a
885                                  * bus reset.  Could not have been a
886                                  * media change, so we just retry the
887                                  * command and see what happens.
888                                  */
889                                 action = ACTION_RETRY;
890                         }
891                         break;
892                 case ILLEGAL_REQUEST:
893                         /* If we had an ILLEGAL REQUEST returned, then
894                          * we may have performed an unsupported
895                          * command.  The only thing this should be
896                          * would be a ten byte read where only a six
897                          * byte read was supported.  Also, on a system
898                          * where READ CAPACITY failed, we may have
899                          * read past the end of the disk.
900                          */
901                         if ((cmd->device->use_10_for_rw &&
902                             sshdr.asc == 0x20 && sshdr.ascq == 0x00) &&
903                             (cmd->cmnd[0] == READ_10 ||
904                              cmd->cmnd[0] == WRITE_10)) {
905                                 /* This will issue a new 6-byte command. */
906                                 cmd->device->use_10_for_rw = 0;
907                                 action = ACTION_REPREP;
908                         } else if (sshdr.asc == 0x10) /* DIX */ {
909                                 action = ACTION_FAIL;
910                                 error = BLK_STS_PROTECTION;
911                         /* INVALID COMMAND OPCODE or INVALID FIELD IN CDB */
912                         } else if (sshdr.asc == 0x20 || sshdr.asc == 0x24) {
913                                 action = ACTION_FAIL;
914                                 error = BLK_STS_TARGET;
915                         } else
916                                 action = ACTION_FAIL;
917                         break;
918                 case ABORTED_COMMAND:
919                         action = ACTION_FAIL;
920                         if (sshdr.asc == 0x10) /* DIF */
921                                 error = BLK_STS_PROTECTION;
922                         break;
923                 case NOT_READY:
924                         /* If the device is in the process of becoming
925                          * ready, or has a temporary blockage, retry.
926                          */
927                         if (sshdr.asc == 0x04) {
928                                 switch (sshdr.ascq) {
929                                 case 0x01: /* becoming ready */
930                                 case 0x04: /* format in progress */
931                                 case 0x05: /* rebuild in progress */
932                                 case 0x06: /* recalculation in progress */
933                                 case 0x07: /* operation in progress */
934                                 case 0x08: /* Long write in progress */
935                                 case 0x09: /* self test in progress */
936                                 case 0x14: /* space allocation in progress */
937                                         action = ACTION_DELAYED_RETRY;
938                                         break;
939                                 default:
940                                         action = ACTION_FAIL;
941                                         break;
942                                 }
943                         } else
944                                 action = ACTION_FAIL;
945                         break;
946                 case VOLUME_OVERFLOW:
947                         /* See SSC3rXX or current. */
948                         action = ACTION_FAIL;
949                         break;
950                 default:
951                         action = ACTION_FAIL;
952                         break;
953                 }
954         } else
955                 action = ACTION_FAIL;
956
957         if (action != ACTION_FAIL &&
958             time_before(cmd->jiffies_at_alloc + wait_for, jiffies))
959                 action = ACTION_FAIL;
960
961         switch (action) {
962         case ACTION_FAIL:
963                 /* Give up and fail the remainder of the request */
964                 if (!(req->rq_flags & RQF_QUIET)) {
965                         static DEFINE_RATELIMIT_STATE(_rs,
966                                         DEFAULT_RATELIMIT_INTERVAL,
967                                         DEFAULT_RATELIMIT_BURST);
968
969                         if (unlikely(scsi_logging_level))
970                                 level = SCSI_LOG_LEVEL(SCSI_LOG_MLCOMPLETE_SHIFT,
971                                                        SCSI_LOG_MLCOMPLETE_BITS);
972
973                         /*
974                          * if logging is enabled the failure will be printed
975                          * in scsi_log_completion(), so avoid duplicate messages
976                          */
977                         if (!level && __ratelimit(&_rs)) {
978                                 scsi_print_result(cmd, NULL, FAILED);
979                                 if (driver_byte(result) & DRIVER_SENSE)
980                                         scsi_print_sense(cmd);
981                                 scsi_print_command(cmd);
982                         }
983                 }
984                 if (!scsi_end_request(req, error, blk_rq_err_bytes(req), 0))
985                         return;
986                 /*FALLTHRU*/
987         case ACTION_REPREP:
988         requeue:
989                 /* Unprep the request and put it back at the head of the queue.
990                  * A new command will be prepared and issued.
991                  */
992                 if (q->mq_ops) {
993                         scsi_mq_requeue_cmd(cmd);
994                 } else {
995                         scsi_release_buffers(cmd);
996                         scsi_requeue_command(q, cmd);
997                 }
998                 break;
999         case ACTION_RETRY:
1000                 /* Retry the same command immediately */
1001                 __scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY, 0);
1002                 break;
1003         case ACTION_DELAYED_RETRY:
1004                 /* Retry the same command after a delay */
1005                 __scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY, 0);
1006                 break;
1007         }
1008 }
1009
1010 static int scsi_init_sgtable(struct request *req, struct scsi_data_buffer *sdb)
1011 {
1012         int count;
1013
1014         /*
1015          * If sg table allocation fails, requeue request later.
1016          */
1017         if (unlikely(sg_alloc_table_chained(&sdb->table,
1018                         blk_rq_nr_phys_segments(req), sdb->table.sgl)))
1019                 return BLKPREP_DEFER;
1020
1021         /* 
1022          * Next, walk the list, and fill in the addresses and sizes of
1023          * each segment.
1024          */
1025         count = blk_rq_map_sg(req->q, req, sdb->table.sgl);
1026         BUG_ON(count > sdb->table.nents);
1027         sdb->table.nents = count;
1028         sdb->length = blk_rq_payload_bytes(req);
1029         return BLKPREP_OK;
1030 }
1031
1032 /*
1033  * Function:    scsi_init_io()
1034  *
1035  * Purpose:     SCSI I/O initialize function.
1036  *
1037  * Arguments:   cmd   - Command descriptor we wish to initialize
1038  *
1039  * Returns:     0 on success
1040  *              BLKPREP_DEFER if the failure is retryable
1041  *              BLKPREP_KILL if the failure is fatal
1042  */
1043 int scsi_init_io(struct scsi_cmnd *cmd)
1044 {
1045         struct scsi_device *sdev = cmd->device;
1046         struct request *rq = cmd->request;
1047         bool is_mq = (rq->mq_ctx != NULL);
1048         int error = BLKPREP_KILL;
1049
1050         if (WARN_ON_ONCE(!blk_rq_nr_phys_segments(rq)))
1051                 goto err_exit;
1052
1053         error = scsi_init_sgtable(rq, &cmd->sdb);
1054         if (error)
1055                 goto err_exit;
1056
1057         if (blk_bidi_rq(rq)) {
1058                 if (!rq->q->mq_ops) {
1059                         struct scsi_data_buffer *bidi_sdb =
1060                                 kmem_cache_zalloc(scsi_sdb_cache, GFP_ATOMIC);
1061                         if (!bidi_sdb) {
1062                                 error = BLKPREP_DEFER;
1063                                 goto err_exit;
1064                         }
1065
1066                         rq->next_rq->special = bidi_sdb;
1067                 }
1068
1069                 error = scsi_init_sgtable(rq->next_rq, rq->next_rq->special);
1070                 if (error)
1071                         goto err_exit;
1072         }
1073
1074         if (blk_integrity_rq(rq)) {
1075                 struct scsi_data_buffer *prot_sdb = cmd->prot_sdb;
1076                 int ivecs, count;
1077
1078                 if (prot_sdb == NULL) {
1079                         /*
1080                          * This can happen if someone (e.g. multipath)
1081                          * queues a command to a device on an adapter
1082                          * that does not support DIX.
1083                          */
1084                         WARN_ON_ONCE(1);
1085                         error = BLKPREP_KILL;
1086                         goto err_exit;
1087                 }
1088
1089                 ivecs = blk_rq_count_integrity_sg(rq->q, rq->bio);
1090
1091                 if (sg_alloc_table_chained(&prot_sdb->table, ivecs,
1092                                 prot_sdb->table.sgl)) {
1093                         error = BLKPREP_DEFER;
1094                         goto err_exit;
1095                 }
1096
1097                 count = blk_rq_map_integrity_sg(rq->q, rq->bio,
1098                                                 prot_sdb->table.sgl);
1099                 BUG_ON(unlikely(count > ivecs));
1100                 BUG_ON(unlikely(count > queue_max_integrity_segments(rq->q)));
1101
1102                 cmd->prot_sdb = prot_sdb;
1103                 cmd->prot_sdb->table.nents = count;
1104         }
1105
1106         return BLKPREP_OK;
1107 err_exit:
1108         if (is_mq) {
1109                 scsi_mq_free_sgtables(cmd);
1110         } else {
1111                 scsi_release_buffers(cmd);
1112                 cmd->request->special = NULL;
1113                 scsi_put_command(cmd);
1114                 put_device(&sdev->sdev_gendev);
1115         }
1116         return error;
1117 }
1118 EXPORT_SYMBOL(scsi_init_io);
1119
1120 /**
1121  * scsi_initialize_rq - initialize struct scsi_cmnd partially
1122  * @rq: Request associated with the SCSI command to be initialized.
1123  *
1124  * This function initializes the members of struct scsi_cmnd that must be
1125  * initialized before request processing starts and that won't be
1126  * reinitialized if a SCSI command is requeued.
1127  *
1128  * Called from inside blk_get_request() for pass-through requests and from
1129  * inside scsi_init_command() for filesystem requests.
1130  */
1131 void scsi_initialize_rq(struct request *rq)
1132 {
1133         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1134
1135         scsi_req_init(&cmd->req);
1136         cmd->jiffies_at_alloc = jiffies;
1137         cmd->retries = 0;
1138 }
1139 EXPORT_SYMBOL(scsi_initialize_rq);
1140
1141 /* Add a command to the list used by the aacraid and dpt_i2o drivers */
1142 void scsi_add_cmd_to_list(struct scsi_cmnd *cmd)
1143 {
1144         struct scsi_device *sdev = cmd->device;
1145         struct Scsi_Host *shost = sdev->host;
1146         unsigned long flags;
1147
1148         if (shost->use_cmd_list) {
1149                 spin_lock_irqsave(&sdev->list_lock, flags);
1150                 list_add_tail(&cmd->list, &sdev->cmd_list);
1151                 spin_unlock_irqrestore(&sdev->list_lock, flags);
1152         }
1153 }
1154
1155 /* Remove a command from the list used by the aacraid and dpt_i2o drivers */
1156 void scsi_del_cmd_from_list(struct scsi_cmnd *cmd)
1157 {
1158         struct scsi_device *sdev = cmd->device;
1159         struct Scsi_Host *shost = sdev->host;
1160         unsigned long flags;
1161
1162         if (shost->use_cmd_list) {
1163                 spin_lock_irqsave(&sdev->list_lock, flags);
1164                 BUG_ON(list_empty(&cmd->list));
1165                 list_del_init(&cmd->list);
1166                 spin_unlock_irqrestore(&sdev->list_lock, flags);
1167         }
1168 }
1169
1170 /* Called after a request has been started. */
1171 void scsi_init_command(struct scsi_device *dev, struct scsi_cmnd *cmd)
1172 {
1173         void *buf = cmd->sense_buffer;
1174         void *prot = cmd->prot_sdb;
1175         struct request *rq = blk_mq_rq_from_pdu(cmd);
1176         unsigned int flags = cmd->flags & SCMD_PRESERVED_FLAGS;
1177         unsigned long jiffies_at_alloc;
1178         int retries;
1179
1180         if (!blk_rq_is_scsi(rq) && !(flags & SCMD_INITIALIZED)) {
1181                 flags |= SCMD_INITIALIZED;
1182                 scsi_initialize_rq(rq);
1183         }
1184
1185         jiffies_at_alloc = cmd->jiffies_at_alloc;
1186         retries = cmd->retries;
1187         /* zero out the cmd, except for the embedded scsi_request */
1188         memset((char *)cmd + sizeof(cmd->req), 0,
1189                 sizeof(*cmd) - sizeof(cmd->req) + dev->host->hostt->cmd_size);
1190
1191         cmd->device = dev;
1192         cmd->sense_buffer = buf;
1193         cmd->prot_sdb = prot;
1194         cmd->flags = flags;
1195         INIT_DELAYED_WORK(&cmd->abort_work, scmd_eh_abort_handler);
1196         cmd->jiffies_at_alloc = jiffies_at_alloc;
1197         cmd->retries = retries;
1198
1199         scsi_add_cmd_to_list(cmd);
1200 }
1201
1202 static int scsi_setup_scsi_cmnd(struct scsi_device *sdev, struct request *req)
1203 {
1204         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1205
1206         /*
1207          * Passthrough requests may transfer data, in which case they must
1208          * a bio attached to them.  Or they might contain a SCSI command
1209          * that does not transfer data, in which case they may optionally
1210          * submit a request without an attached bio.
1211          */
1212         if (req->bio) {
1213                 int ret = scsi_init_io(cmd);
1214                 if (unlikely(ret))
1215                         return ret;
1216         } else {
1217                 BUG_ON(blk_rq_bytes(req));
1218
1219                 memset(&cmd->sdb, 0, sizeof(cmd->sdb));
1220         }
1221
1222         cmd->cmd_len = scsi_req(req)->cmd_len;
1223         cmd->cmnd = scsi_req(req)->cmd;
1224         cmd->transfersize = blk_rq_bytes(req);
1225         cmd->allowed = scsi_req(req)->retries;
1226         return BLKPREP_OK;
1227 }
1228
1229 /*
1230  * Setup a normal block command.  These are simple request from filesystems
1231  * that still need to be translated to SCSI CDBs from the ULD.
1232  */
1233 static int scsi_setup_fs_cmnd(struct scsi_device *sdev, struct request *req)
1234 {
1235         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1236
1237         if (unlikely(sdev->handler && sdev->handler->prep_fn)) {
1238                 int ret = sdev->handler->prep_fn(sdev, req);
1239                 if (ret != BLKPREP_OK)
1240                         return ret;
1241         }
1242
1243         cmd->cmnd = scsi_req(req)->cmd = scsi_req(req)->__cmd;
1244         memset(cmd->cmnd, 0, BLK_MAX_CDB);
1245         return scsi_cmd_to_driver(cmd)->init_command(cmd);
1246 }
1247
1248 static int scsi_setup_cmnd(struct scsi_device *sdev, struct request *req)
1249 {
1250         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1251
1252         if (!blk_rq_bytes(req))
1253                 cmd->sc_data_direction = DMA_NONE;
1254         else if (rq_data_dir(req) == WRITE)
1255                 cmd->sc_data_direction = DMA_TO_DEVICE;
1256         else
1257                 cmd->sc_data_direction = DMA_FROM_DEVICE;
1258
1259         if (blk_rq_is_scsi(req))
1260                 return scsi_setup_scsi_cmnd(sdev, req);
1261         else
1262                 return scsi_setup_fs_cmnd(sdev, req);
1263 }
1264
1265 static int
1266 scsi_prep_state_check(struct scsi_device *sdev, struct request *req)
1267 {
1268         int ret = BLKPREP_OK;
1269
1270         /*
1271          * If the device is not in running state we will reject some
1272          * or all commands.
1273          */
1274         if (unlikely(sdev->sdev_state != SDEV_RUNNING)) {
1275                 switch (sdev->sdev_state) {
1276                 case SDEV_OFFLINE:
1277                 case SDEV_TRANSPORT_OFFLINE:
1278                         /*
1279                          * If the device is offline we refuse to process any
1280                          * commands.  The device must be brought online
1281                          * before trying any recovery commands.
1282                          */
1283                         sdev_printk(KERN_ERR, sdev,
1284                                     "rejecting I/O to offline device\n");
1285                         ret = BLKPREP_KILL;
1286                         break;
1287                 case SDEV_DEL:
1288                         /*
1289                          * If the device is fully deleted, we refuse to
1290                          * process any commands as well.
1291                          */
1292                         sdev_printk(KERN_ERR, sdev,
1293                                     "rejecting I/O to dead device\n");
1294                         ret = BLKPREP_KILL;
1295                         break;
1296                 case SDEV_BLOCK:
1297                 case SDEV_CREATED_BLOCK:
1298                         ret = BLKPREP_DEFER;
1299                         break;
1300                 case SDEV_QUIESCE:
1301                         /*
1302                          * If the devices is blocked we defer normal commands.
1303                          */
1304                         if (!(req->rq_flags & RQF_PREEMPT))
1305                                 ret = BLKPREP_DEFER;
1306                         break;
1307                 default:
1308                         /*
1309                          * For any other not fully online state we only allow
1310                          * special commands.  In particular any user initiated
1311                          * command is not allowed.
1312                          */
1313                         if (!(req->rq_flags & RQF_PREEMPT))
1314                                 ret = BLKPREP_KILL;
1315                         break;
1316                 }
1317         }
1318         return ret;
1319 }
1320
1321 static int
1322 scsi_prep_return(struct request_queue *q, struct request *req, int ret)
1323 {
1324         struct scsi_device *sdev = q->queuedata;
1325
1326         switch (ret) {
1327         case BLKPREP_KILL:
1328         case BLKPREP_INVALID:
1329                 scsi_req(req)->result = DID_NO_CONNECT << 16;
1330                 /* release the command and kill it */
1331                 if (req->special) {
1332                         struct scsi_cmnd *cmd = req->special;
1333                         scsi_release_buffers(cmd);
1334                         scsi_put_command(cmd);
1335                         put_device(&sdev->sdev_gendev);
1336                         req->special = NULL;
1337                 }
1338                 break;
1339         case BLKPREP_DEFER:
1340                 /*
1341                  * If we defer, the blk_peek_request() returns NULL, but the
1342                  * queue must be restarted, so we schedule a callback to happen
1343                  * shortly.
1344                  */
1345                 if (atomic_read(&sdev->device_busy) == 0)
1346                         blk_delay_queue(q, SCSI_QUEUE_DELAY);
1347                 break;
1348         default:
1349                 req->rq_flags |= RQF_DONTPREP;
1350         }
1351
1352         return ret;
1353 }
1354
1355 static int scsi_prep_fn(struct request_queue *q, struct request *req)
1356 {
1357         struct scsi_device *sdev = q->queuedata;
1358         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1359         int ret;
1360
1361         ret = scsi_prep_state_check(sdev, req);
1362         if (ret != BLKPREP_OK)
1363                 goto out;
1364
1365         if (!req->special) {
1366                 /* Bail if we can't get a reference to the device */
1367                 if (unlikely(!get_device(&sdev->sdev_gendev))) {
1368                         ret = BLKPREP_DEFER;
1369                         goto out;
1370                 }
1371
1372                 scsi_init_command(sdev, cmd);
1373                 req->special = cmd;
1374         }
1375
1376         cmd->tag = req->tag;
1377         cmd->request = req;
1378         cmd->prot_op = SCSI_PROT_NORMAL;
1379
1380         ret = scsi_setup_cmnd(sdev, req);
1381 out:
1382         return scsi_prep_return(q, req, ret);
1383 }
1384
1385 static void scsi_unprep_fn(struct request_queue *q, struct request *req)
1386 {
1387         scsi_uninit_cmd(blk_mq_rq_to_pdu(req));
1388 }
1389
1390 /*
1391  * scsi_dev_queue_ready: if we can send requests to sdev, return 1 else
1392  * return 0.
1393  *
1394  * Called with the queue_lock held.
1395  */
1396 static inline int scsi_dev_queue_ready(struct request_queue *q,
1397                                   struct scsi_device *sdev)
1398 {
1399         unsigned int busy;
1400
1401         busy = atomic_inc_return(&sdev->device_busy) - 1;
1402         if (atomic_read(&sdev->device_blocked)) {
1403                 if (busy)
1404                         goto out_dec;
1405
1406                 /*
1407                  * unblock after device_blocked iterates to zero
1408                  */
1409                 if (atomic_dec_return(&sdev->device_blocked) > 0) {
1410                         /*
1411                          * For the MQ case we take care of this in the caller.
1412                          */
1413                         if (!q->mq_ops)
1414                                 blk_delay_queue(q, SCSI_QUEUE_DELAY);
1415                         goto out_dec;
1416                 }
1417                 SCSI_LOG_MLQUEUE(3, sdev_printk(KERN_INFO, sdev,
1418                                    "unblocking device at zero depth\n"));
1419         }
1420
1421         if (busy >= sdev->queue_depth)
1422                 goto out_dec;
1423
1424         return 1;
1425 out_dec:
1426         atomic_dec(&sdev->device_busy);
1427         return 0;
1428 }
1429
1430 /*
1431  * scsi_target_queue_ready: checks if there we can send commands to target
1432  * @sdev: scsi device on starget to check.
1433  */
1434 static inline int scsi_target_queue_ready(struct Scsi_Host *shost,
1435                                            struct scsi_device *sdev)
1436 {
1437         struct scsi_target *starget = scsi_target(sdev);
1438         unsigned int busy;
1439
1440         if (starget->single_lun) {
1441                 spin_lock_irq(shost->host_lock);
1442                 if (starget->starget_sdev_user &&
1443                     starget->starget_sdev_user != sdev) {
1444                         spin_unlock_irq(shost->host_lock);
1445                         return 0;
1446                 }
1447                 starget->starget_sdev_user = sdev;
1448                 spin_unlock_irq(shost->host_lock);
1449         }
1450
1451         if (starget->can_queue <= 0)
1452                 return 1;
1453
1454         busy = atomic_inc_return(&starget->target_busy) - 1;
1455         if (atomic_read(&starget->target_blocked) > 0) {
1456                 if (busy)
1457                         goto starved;
1458
1459                 /*
1460                  * unblock after target_blocked iterates to zero
1461                  */
1462                 if (atomic_dec_return(&starget->target_blocked) > 0)
1463                         goto out_dec;
1464
1465                 SCSI_LOG_MLQUEUE(3, starget_printk(KERN_INFO, starget,
1466                                  "unblocking target at zero depth\n"));
1467         }
1468
1469         if (busy >= starget->can_queue)
1470                 goto starved;
1471
1472         return 1;
1473
1474 starved:
1475         spin_lock_irq(shost->host_lock);
1476         list_move_tail(&sdev->starved_entry, &shost->starved_list);
1477         spin_unlock_irq(shost->host_lock);
1478 out_dec:
1479         if (starget->can_queue > 0)
1480                 atomic_dec(&starget->target_busy);
1481         return 0;
1482 }
1483
1484 /*
1485  * scsi_host_queue_ready: if we can send requests to shost, return 1 else
1486  * return 0. We must end up running the queue again whenever 0 is
1487  * returned, else IO can hang.
1488  */
1489 static inline int scsi_host_queue_ready(struct request_queue *q,
1490                                    struct Scsi_Host *shost,
1491                                    struct scsi_device *sdev)
1492 {
1493         unsigned int busy;
1494
1495         if (scsi_host_in_recovery(shost))
1496                 return 0;
1497
1498         busy = atomic_inc_return(&shost->host_busy) - 1;
1499         if (atomic_read(&shost->host_blocked) > 0) {
1500                 if (busy)
1501                         goto starved;
1502
1503                 /*
1504                  * unblock after host_blocked iterates to zero
1505                  */
1506                 if (atomic_dec_return(&shost->host_blocked) > 0)
1507                         goto out_dec;
1508
1509                 SCSI_LOG_MLQUEUE(3,
1510                         shost_printk(KERN_INFO, shost,
1511                                      "unblocking host at zero depth\n"));
1512         }
1513
1514         if (shost->can_queue > 0 && busy >= shost->can_queue)
1515                 goto starved;
1516         if (shost->host_self_blocked)
1517                 goto starved;
1518
1519         /* We're OK to process the command, so we can't be starved */
1520         if (!list_empty(&sdev->starved_entry)) {
1521                 spin_lock_irq(shost->host_lock);
1522                 if (!list_empty(&sdev->starved_entry))
1523                         list_del_init(&sdev->starved_entry);
1524                 spin_unlock_irq(shost->host_lock);
1525         }
1526
1527         return 1;
1528
1529 starved:
1530         spin_lock_irq(shost->host_lock);
1531         if (list_empty(&sdev->starved_entry))
1532                 list_add_tail(&sdev->starved_entry, &shost->starved_list);
1533         spin_unlock_irq(shost->host_lock);
1534 out_dec:
1535         atomic_dec(&shost->host_busy);
1536         return 0;
1537 }
1538
1539 /*
1540  * Busy state exporting function for request stacking drivers.
1541  *
1542  * For efficiency, no lock is taken to check the busy state of
1543  * shost/starget/sdev, since the returned value is not guaranteed and
1544  * may be changed after request stacking drivers call the function,
1545  * regardless of taking lock or not.
1546  *
1547  * When scsi can't dispatch I/Os anymore and needs to kill I/Os scsi
1548  * needs to return 'not busy'. Otherwise, request stacking drivers
1549  * may hold requests forever.
1550  */
1551 static int scsi_lld_busy(struct request_queue *q)
1552 {
1553         struct scsi_device *sdev = q->queuedata;
1554         struct Scsi_Host *shost;
1555
1556         if (blk_queue_dying(q))
1557                 return 0;
1558
1559         shost = sdev->host;
1560
1561         /*
1562          * Ignore host/starget busy state.
1563          * Since block layer does not have a concept of fairness across
1564          * multiple queues, congestion of host/starget needs to be handled
1565          * in SCSI layer.
1566          */
1567         if (scsi_host_in_recovery(shost) || scsi_device_is_busy(sdev))
1568                 return 1;
1569
1570         return 0;
1571 }
1572
1573 /*
1574  * Kill a request for a dead device
1575  */
1576 static void scsi_kill_request(struct request *req, struct request_queue *q)
1577 {
1578         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1579         struct scsi_device *sdev;
1580         struct scsi_target *starget;
1581         struct Scsi_Host *shost;
1582
1583         blk_start_request(req);
1584
1585         scmd_printk(KERN_INFO, cmd, "killing request\n");
1586
1587         sdev = cmd->device;
1588         starget = scsi_target(sdev);
1589         shost = sdev->host;
1590         scsi_init_cmd_errh(cmd);
1591         cmd->result = DID_NO_CONNECT << 16;
1592         atomic_inc(&cmd->device->iorequest_cnt);
1593
1594         /*
1595          * SCSI request completion path will do scsi_device_unbusy(),
1596          * bump busy counts.  To bump the counters, we need to dance
1597          * with the locks as normal issue path does.
1598          */
1599         atomic_inc(&sdev->device_busy);
1600         atomic_inc(&shost->host_busy);
1601         if (starget->can_queue > 0)
1602                 atomic_inc(&starget->target_busy);
1603
1604         blk_complete_request(req);
1605 }
1606
1607 static void scsi_softirq_done(struct request *rq)
1608 {
1609         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1610         unsigned long wait_for = (cmd->allowed + 1) * rq->timeout;
1611         int disposition;
1612
1613         INIT_LIST_HEAD(&cmd->eh_entry);
1614
1615         atomic_inc(&cmd->device->iodone_cnt);
1616         if (cmd->result)
1617                 atomic_inc(&cmd->device->ioerr_cnt);
1618
1619         disposition = scsi_decide_disposition(cmd);
1620         if (disposition != SUCCESS &&
1621             time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) {
1622                 sdev_printk(KERN_ERR, cmd->device,
1623                             "timing out command, waited %lus\n",
1624                             wait_for/HZ);
1625                 disposition = SUCCESS;
1626         }
1627
1628         scsi_log_completion(cmd, disposition);
1629
1630         switch (disposition) {
1631                 case SUCCESS:
1632                         scsi_finish_command(cmd);
1633                         break;
1634                 case NEEDS_RETRY:
1635                         scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY);
1636                         break;
1637                 case ADD_TO_MLQUEUE:
1638                         scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY);
1639                         break;
1640                 default:
1641                         scsi_eh_scmd_add(cmd);
1642                         break;
1643         }
1644 }
1645
1646 /**
1647  * scsi_dispatch_command - Dispatch a command to the low-level driver.
1648  * @cmd: command block we are dispatching.
1649  *
1650  * Return: nonzero return request was rejected and device's queue needs to be
1651  * plugged.
1652  */
1653 static int scsi_dispatch_cmd(struct scsi_cmnd *cmd)
1654 {
1655         struct Scsi_Host *host = cmd->device->host;
1656         int rtn = 0;
1657
1658         atomic_inc(&cmd->device->iorequest_cnt);
1659
1660         /* check if the device is still usable */
1661         if (unlikely(cmd->device->sdev_state == SDEV_DEL)) {
1662                 /* in SDEV_DEL we error all commands. DID_NO_CONNECT
1663                  * returns an immediate error upwards, and signals
1664                  * that the device is no longer present */
1665                 cmd->result = DID_NO_CONNECT << 16;
1666                 goto done;
1667         }
1668
1669         /* Check to see if the scsi lld made this device blocked. */
1670         if (unlikely(scsi_device_blocked(cmd->device))) {
1671                 /*
1672                  * in blocked state, the command is just put back on
1673                  * the device queue.  The suspend state has already
1674                  * blocked the queue so future requests should not
1675                  * occur until the device transitions out of the
1676                  * suspend state.
1677                  */
1678                 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1679                         "queuecommand : device blocked\n"));
1680                 return SCSI_MLQUEUE_DEVICE_BUSY;
1681         }
1682
1683         /* Store the LUN value in cmnd, if needed. */
1684         if (cmd->device->lun_in_cdb)
1685                 cmd->cmnd[1] = (cmd->cmnd[1] & 0x1f) |
1686                                (cmd->device->lun << 5 & 0xe0);
1687
1688         scsi_log_send(cmd);
1689
1690         /*
1691          * Before we queue this command, check if the command
1692          * length exceeds what the host adapter can handle.
1693          */
1694         if (cmd->cmd_len > cmd->device->host->max_cmd_len) {
1695                 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1696                                "queuecommand : command too long. "
1697                                "cdb_size=%d host->max_cmd_len=%d\n",
1698                                cmd->cmd_len, cmd->device->host->max_cmd_len));
1699                 cmd->result = (DID_ABORT << 16);
1700                 goto done;
1701         }
1702
1703         if (unlikely(host->shost_state == SHOST_DEL)) {
1704                 cmd->result = (DID_NO_CONNECT << 16);
1705                 goto done;
1706
1707         }
1708
1709         trace_scsi_dispatch_cmd_start(cmd);
1710         rtn = host->hostt->queuecommand(host, cmd);
1711         if (rtn) {
1712                 trace_scsi_dispatch_cmd_error(cmd, rtn);
1713                 if (rtn != SCSI_MLQUEUE_DEVICE_BUSY &&
1714                     rtn != SCSI_MLQUEUE_TARGET_BUSY)
1715                         rtn = SCSI_MLQUEUE_HOST_BUSY;
1716
1717                 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1718                         "queuecommand : request rejected\n"));
1719         }
1720
1721         return rtn;
1722  done:
1723         cmd->scsi_done(cmd);
1724         return 0;
1725 }
1726
1727 /**
1728  * scsi_done - Invoke completion on finished SCSI command.
1729  * @cmd: The SCSI Command for which a low-level device driver (LLDD) gives
1730  * ownership back to SCSI Core -- i.e. the LLDD has finished with it.
1731  *
1732  * Description: This function is the mid-level's (SCSI Core) interrupt routine,
1733  * which regains ownership of the SCSI command (de facto) from a LLDD, and
1734  * calls blk_complete_request() for further processing.
1735  *
1736  * This function is interrupt context safe.
1737  */
1738 static void scsi_done(struct scsi_cmnd *cmd)
1739 {
1740         trace_scsi_dispatch_cmd_done(cmd);
1741         blk_complete_request(cmd->request);
1742 }
1743
1744 /*
1745  * Function:    scsi_request_fn()
1746  *
1747  * Purpose:     Main strategy routine for SCSI.
1748  *
1749  * Arguments:   q       - Pointer to actual queue.
1750  *
1751  * Returns:     Nothing
1752  *
1753  * Lock status: IO request lock assumed to be held when called.
1754  */
1755 static void scsi_request_fn(struct request_queue *q)
1756         __releases(q->queue_lock)
1757         __acquires(q->queue_lock)
1758 {
1759         struct scsi_device *sdev = q->queuedata;
1760         struct Scsi_Host *shost;
1761         struct scsi_cmnd *cmd;
1762         struct request *req;
1763
1764         /*
1765          * To start with, we keep looping until the queue is empty, or until
1766          * the host is no longer able to accept any more requests.
1767          */
1768         shost = sdev->host;
1769         for (;;) {
1770                 int rtn;
1771                 /*
1772                  * get next queueable request.  We do this early to make sure
1773                  * that the request is fully prepared even if we cannot
1774                  * accept it.
1775                  */
1776                 req = blk_peek_request(q);
1777                 if (!req)
1778                         break;
1779
1780                 if (unlikely(!scsi_device_online(sdev))) {
1781                         sdev_printk(KERN_ERR, sdev,
1782                                     "rejecting I/O to offline device\n");
1783                         scsi_kill_request(req, q);
1784                         continue;
1785                 }
1786
1787                 if (!scsi_dev_queue_ready(q, sdev))
1788                         break;
1789
1790                 /*
1791                  * Remove the request from the request list.
1792                  */
1793                 if (!(blk_queue_tagged(q) && !blk_queue_start_tag(q, req)))
1794                         blk_start_request(req);
1795
1796                 spin_unlock_irq(q->queue_lock);
1797                 cmd = blk_mq_rq_to_pdu(req);
1798                 if (cmd != req->special) {
1799                         printk(KERN_CRIT "impossible request in %s.\n"
1800                                          "please mail a stack trace to "
1801                                          "linux-scsi@vger.kernel.org\n",
1802                                          __func__);
1803                         blk_dump_rq_flags(req, "foo");
1804                         BUG();
1805                 }
1806
1807                 /*
1808                  * We hit this when the driver is using a host wide
1809                  * tag map. For device level tag maps the queue_depth check
1810                  * in the device ready fn would prevent us from trying
1811                  * to allocate a tag. Since the map is a shared host resource
1812                  * we add the dev to the starved list so it eventually gets
1813                  * a run when a tag is freed.
1814                  */
1815                 if (blk_queue_tagged(q) && !(req->rq_flags & RQF_QUEUED)) {
1816                         spin_lock_irq(shost->host_lock);
1817                         if (list_empty(&sdev->starved_entry))
1818                                 list_add_tail(&sdev->starved_entry,
1819                                               &shost->starved_list);
1820                         spin_unlock_irq(shost->host_lock);
1821                         goto not_ready;
1822                 }
1823
1824                 if (!scsi_target_queue_ready(shost, sdev))
1825                         goto not_ready;
1826
1827                 if (!scsi_host_queue_ready(q, shost, sdev))
1828                         goto host_not_ready;
1829         
1830                 if (sdev->simple_tags)
1831                         cmd->flags |= SCMD_TAGGED;
1832                 else
1833                         cmd->flags &= ~SCMD_TAGGED;
1834
1835                 /*
1836                  * Finally, initialize any error handling parameters, and set up
1837                  * the timers for timeouts.
1838                  */
1839                 scsi_init_cmd_errh(cmd);
1840
1841                 /*
1842                  * Dispatch the command to the low-level driver.
1843                  */
1844                 cmd->scsi_done = scsi_done;
1845                 rtn = scsi_dispatch_cmd(cmd);
1846                 if (rtn) {
1847                         scsi_queue_insert(cmd, rtn);
1848                         spin_lock_irq(q->queue_lock);
1849                         goto out_delay;
1850                 }
1851                 spin_lock_irq(q->queue_lock);
1852         }
1853
1854         return;
1855
1856  host_not_ready:
1857         if (scsi_target(sdev)->can_queue > 0)
1858                 atomic_dec(&scsi_target(sdev)->target_busy);
1859  not_ready:
1860         /*
1861          * lock q, handle tag, requeue req, and decrement device_busy. We
1862          * must return with queue_lock held.
1863          *
1864          * Decrementing device_busy without checking it is OK, as all such
1865          * cases (host limits or settings) should run the queue at some
1866          * later time.
1867          */
1868         spin_lock_irq(q->queue_lock);
1869         blk_requeue_request(q, req);
1870         atomic_dec(&sdev->device_busy);
1871 out_delay:
1872         if (!atomic_read(&sdev->device_busy) && !scsi_device_blocked(sdev))
1873                 blk_delay_queue(q, SCSI_QUEUE_DELAY);
1874 }
1875
1876 static inline blk_status_t prep_to_mq(int ret)
1877 {
1878         switch (ret) {
1879         case BLKPREP_OK:
1880                 return BLK_STS_OK;
1881         case BLKPREP_DEFER:
1882                 return BLK_STS_RESOURCE;
1883         default:
1884                 return BLK_STS_IOERR;
1885         }
1886 }
1887
1888 /* Size in bytes of the sg-list stored in the scsi-mq command-private data. */
1889 static unsigned int scsi_mq_sgl_size(struct Scsi_Host *shost)
1890 {
1891         return min_t(unsigned int, shost->sg_tablesize, SG_CHUNK_SIZE) *
1892                 sizeof(struct scatterlist);
1893 }
1894
1895 static int scsi_mq_prep_fn(struct request *req)
1896 {
1897         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1898         struct scsi_device *sdev = req->q->queuedata;
1899         struct Scsi_Host *shost = sdev->host;
1900         struct scatterlist *sg;
1901
1902         scsi_init_command(sdev, cmd);
1903
1904         req->special = cmd;
1905
1906         cmd->request = req;
1907
1908         cmd->tag = req->tag;
1909         cmd->prot_op = SCSI_PROT_NORMAL;
1910
1911         sg = (void *)cmd + sizeof(struct scsi_cmnd) + shost->hostt->cmd_size;
1912         cmd->sdb.table.sgl = sg;
1913
1914         if (scsi_host_get_prot(shost)) {
1915                 memset(cmd->prot_sdb, 0, sizeof(struct scsi_data_buffer));
1916
1917                 cmd->prot_sdb->table.sgl =
1918                         (struct scatterlist *)(cmd->prot_sdb + 1);
1919         }
1920
1921         if (blk_bidi_rq(req)) {
1922                 struct request *next_rq = req->next_rq;
1923                 struct scsi_data_buffer *bidi_sdb = blk_mq_rq_to_pdu(next_rq);
1924
1925                 memset(bidi_sdb, 0, sizeof(struct scsi_data_buffer));
1926                 bidi_sdb->table.sgl =
1927                         (struct scatterlist *)(bidi_sdb + 1);
1928
1929                 next_rq->special = bidi_sdb;
1930         }
1931
1932         blk_mq_start_request(req);
1933
1934         return scsi_setup_cmnd(sdev, req);
1935 }
1936
1937 static void scsi_mq_done(struct scsi_cmnd *cmd)
1938 {
1939         trace_scsi_dispatch_cmd_done(cmd);
1940         blk_mq_complete_request(cmd->request);
1941 }
1942
1943 static blk_status_t scsi_queue_rq(struct blk_mq_hw_ctx *hctx,
1944                          const struct blk_mq_queue_data *bd)
1945 {
1946         struct request *req = bd->rq;
1947         struct request_queue *q = req->q;
1948         struct scsi_device *sdev = q->queuedata;
1949         struct Scsi_Host *shost = sdev->host;
1950         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1951         blk_status_t ret;
1952         int reason;
1953
1954         ret = prep_to_mq(scsi_prep_state_check(sdev, req));
1955         if (ret != BLK_STS_OK)
1956                 goto out;
1957
1958         ret = BLK_STS_RESOURCE;
1959         if (!get_device(&sdev->sdev_gendev))
1960                 goto out;
1961
1962         if (!scsi_dev_queue_ready(q, sdev))
1963                 goto out_put_device;
1964         if (!scsi_target_queue_ready(shost, sdev))
1965                 goto out_dec_device_busy;
1966         if (!scsi_host_queue_ready(q, shost, sdev))
1967                 goto out_dec_target_busy;
1968
1969         if (!(req->rq_flags & RQF_DONTPREP)) {
1970                 ret = prep_to_mq(scsi_mq_prep_fn(req));
1971                 if (ret != BLK_STS_OK)
1972                         goto out_dec_host_busy;
1973                 req->rq_flags |= RQF_DONTPREP;
1974         } else {
1975                 blk_mq_start_request(req);
1976         }
1977
1978         if (sdev->simple_tags)
1979                 cmd->flags |= SCMD_TAGGED;
1980         else
1981                 cmd->flags &= ~SCMD_TAGGED;
1982
1983         scsi_init_cmd_errh(cmd);
1984         cmd->scsi_done = scsi_mq_done;
1985
1986         reason = scsi_dispatch_cmd(cmd);
1987         if (reason) {
1988                 scsi_set_blocked(cmd, reason);
1989                 ret = BLK_STS_RESOURCE;
1990                 goto out_dec_host_busy;
1991         }
1992
1993         return BLK_STS_OK;
1994
1995 out_dec_host_busy:
1996         atomic_dec(&shost->host_busy);
1997 out_dec_target_busy:
1998         if (scsi_target(sdev)->can_queue > 0)
1999                 atomic_dec(&scsi_target(sdev)->target_busy);
2000 out_dec_device_busy:
2001         atomic_dec(&sdev->device_busy);
2002 out_put_device:
2003         put_device(&sdev->sdev_gendev);
2004 out:
2005         switch (ret) {
2006         case BLK_STS_OK:
2007                 break;
2008         case BLK_STS_RESOURCE:
2009                 if (atomic_read(&sdev->device_busy) == 0 &&
2010                     !scsi_device_blocked(sdev))
2011                         blk_mq_delay_run_hw_queue(hctx, SCSI_QUEUE_DELAY);
2012                 break;
2013         default:
2014                 /*
2015                  * Make sure to release all allocated ressources when
2016                  * we hit an error, as we will never see this command
2017                  * again.
2018                  */
2019                 if (req->rq_flags & RQF_DONTPREP)
2020                         scsi_mq_uninit_cmd(cmd);
2021                 break;
2022         }
2023         return ret;
2024 }
2025
2026 static enum blk_eh_timer_return scsi_timeout(struct request *req,
2027                 bool reserved)
2028 {
2029         if (reserved)
2030                 return BLK_EH_RESET_TIMER;
2031         return scsi_times_out(req);
2032 }
2033
2034 static int scsi_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
2035                                 unsigned int hctx_idx, unsigned int numa_node)
2036 {
2037         struct Scsi_Host *shost = set->driver_data;
2038         const bool unchecked_isa_dma = shost->unchecked_isa_dma;
2039         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
2040         struct scatterlist *sg;
2041
2042         if (unchecked_isa_dma)
2043                 cmd->flags |= SCMD_UNCHECKED_ISA_DMA;
2044         cmd->sense_buffer = scsi_alloc_sense_buffer(unchecked_isa_dma,
2045                                                     GFP_KERNEL, numa_node);
2046         if (!cmd->sense_buffer)
2047                 return -ENOMEM;
2048         cmd->req.sense = cmd->sense_buffer;
2049
2050         if (scsi_host_get_prot(shost)) {
2051                 sg = (void *)cmd + sizeof(struct scsi_cmnd) +
2052                         shost->hostt->cmd_size;
2053                 cmd->prot_sdb = (void *)sg + scsi_mq_sgl_size(shost);
2054         }
2055
2056         return 0;
2057 }
2058
2059 static void scsi_mq_exit_request(struct blk_mq_tag_set *set, struct request *rq,
2060                                  unsigned int hctx_idx)
2061 {
2062         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
2063
2064         scsi_free_sense_buffer(cmd->flags & SCMD_UNCHECKED_ISA_DMA,
2065                                cmd->sense_buffer);
2066 }
2067
2068 static int scsi_map_queues(struct blk_mq_tag_set *set)
2069 {
2070         struct Scsi_Host *shost = container_of(set, struct Scsi_Host, tag_set);
2071
2072         if (shost->hostt->map_queues)
2073                 return shost->hostt->map_queues(shost);
2074         return blk_mq_map_queues(set);
2075 }
2076
2077 static u64 scsi_calculate_bounce_limit(struct Scsi_Host *shost)
2078 {
2079         struct device *host_dev;
2080         u64 bounce_limit = 0xffffffff;
2081
2082         if (shost->unchecked_isa_dma)
2083                 return BLK_BOUNCE_ISA;
2084         /*
2085          * Platforms with virtual-DMA translation
2086          * hardware have no practical limit.
2087          */
2088         if (!PCI_DMA_BUS_IS_PHYS)
2089                 return BLK_BOUNCE_ANY;
2090
2091         host_dev = scsi_get_device(shost);
2092         if (host_dev && host_dev->dma_mask)
2093                 bounce_limit = (u64)dma_max_pfn(host_dev) << PAGE_SHIFT;
2094
2095         return bounce_limit;
2096 }
2097
2098 void __scsi_init_queue(struct Scsi_Host *shost, struct request_queue *q)
2099 {
2100         struct device *dev = shost->dma_dev;
2101
2102         queue_flag_set_unlocked(QUEUE_FLAG_SCSI_PASSTHROUGH, q);
2103
2104         /*
2105          * this limit is imposed by hardware restrictions
2106          */
2107         blk_queue_max_segments(q, min_t(unsigned short, shost->sg_tablesize,
2108                                         SG_MAX_SEGMENTS));
2109
2110         if (scsi_host_prot_dma(shost)) {
2111                 shost->sg_prot_tablesize =
2112                         min_not_zero(shost->sg_prot_tablesize,
2113                                      (unsigned short)SCSI_MAX_PROT_SG_SEGMENTS);
2114                 BUG_ON(shost->sg_prot_tablesize < shost->sg_tablesize);
2115                 blk_queue_max_integrity_segments(q, shost->sg_prot_tablesize);
2116         }
2117
2118         blk_queue_max_hw_sectors(q, shost->max_sectors);
2119         blk_queue_bounce_limit(q, scsi_calculate_bounce_limit(shost));
2120         blk_queue_segment_boundary(q, shost->dma_boundary);
2121         dma_set_seg_boundary(dev, shost->dma_boundary);
2122
2123         blk_queue_max_segment_size(q, dma_get_max_seg_size(dev));
2124
2125         if (!shost->use_clustering)
2126                 q->limits.cluster = 0;
2127
2128         /*
2129          * set a reasonable default alignment on word boundaries: the
2130          * host and device may alter it using
2131          * blk_queue_update_dma_alignment() later.
2132          */
2133         blk_queue_dma_alignment(q, 0x03);
2134 }
2135 EXPORT_SYMBOL_GPL(__scsi_init_queue);
2136
2137 static int scsi_old_init_rq(struct request_queue *q, struct request *rq,
2138                             gfp_t gfp)
2139 {
2140         struct Scsi_Host *shost = q->rq_alloc_data;
2141         const bool unchecked_isa_dma = shost->unchecked_isa_dma;
2142         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
2143
2144         memset(cmd, 0, sizeof(*cmd));
2145
2146         if (unchecked_isa_dma)
2147                 cmd->flags |= SCMD_UNCHECKED_ISA_DMA;
2148         cmd->sense_buffer = scsi_alloc_sense_buffer(unchecked_isa_dma, gfp,
2149                                                     NUMA_NO_NODE);
2150         if (!cmd->sense_buffer)
2151                 goto fail;
2152         cmd->req.sense = cmd->sense_buffer;
2153
2154         if (scsi_host_get_prot(shost) >= SHOST_DIX_TYPE0_PROTECTION) {
2155                 cmd->prot_sdb = kmem_cache_zalloc(scsi_sdb_cache, gfp);
2156                 if (!cmd->prot_sdb)
2157                         goto fail_free_sense;
2158         }
2159
2160         return 0;
2161
2162 fail_free_sense:
2163         scsi_free_sense_buffer(unchecked_isa_dma, cmd->sense_buffer);
2164 fail:
2165         return -ENOMEM;
2166 }
2167
2168 static void scsi_old_exit_rq(struct request_queue *q, struct request *rq)
2169 {
2170         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
2171
2172         if (cmd->prot_sdb)
2173                 kmem_cache_free(scsi_sdb_cache, cmd->prot_sdb);
2174         scsi_free_sense_buffer(cmd->flags & SCMD_UNCHECKED_ISA_DMA,
2175                                cmd->sense_buffer);
2176 }
2177
2178 struct request_queue *scsi_old_alloc_queue(struct scsi_device *sdev)
2179 {
2180         struct Scsi_Host *shost = sdev->host;
2181         struct request_queue *q;
2182
2183         q = blk_alloc_queue_node(GFP_KERNEL, NUMA_NO_NODE);
2184         if (!q)
2185                 return NULL;
2186         q->cmd_size = sizeof(struct scsi_cmnd) + shost->hostt->cmd_size;
2187         q->rq_alloc_data = shost;
2188         q->request_fn = scsi_request_fn;
2189         q->init_rq_fn = scsi_old_init_rq;
2190         q->exit_rq_fn = scsi_old_exit_rq;
2191         q->initialize_rq_fn = scsi_initialize_rq;
2192
2193         if (blk_init_allocated_queue(q) < 0) {
2194                 blk_cleanup_queue(q);
2195                 return NULL;
2196         }
2197
2198         __scsi_init_queue(shost, q);
2199         blk_queue_prep_rq(q, scsi_prep_fn);
2200         blk_queue_unprep_rq(q, scsi_unprep_fn);
2201         blk_queue_softirq_done(q, scsi_softirq_done);
2202         blk_queue_rq_timed_out(q, scsi_times_out);
2203         blk_queue_lld_busy(q, scsi_lld_busy);
2204         return q;
2205 }
2206
2207 static const struct blk_mq_ops scsi_mq_ops = {
2208         .queue_rq       = scsi_queue_rq,
2209         .complete       = scsi_softirq_done,
2210         .timeout        = scsi_timeout,
2211 #ifdef CONFIG_BLK_DEBUG_FS
2212         .show_rq        = scsi_show_rq,
2213 #endif
2214         .init_request   = scsi_mq_init_request,
2215         .exit_request   = scsi_mq_exit_request,
2216         .initialize_rq_fn = scsi_initialize_rq,
2217         .map_queues     = scsi_map_queues,
2218 };
2219
2220 struct request_queue *scsi_mq_alloc_queue(struct scsi_device *sdev)
2221 {
2222         sdev->request_queue = blk_mq_init_queue(&sdev->host->tag_set);
2223         if (IS_ERR(sdev->request_queue))
2224                 return NULL;
2225
2226         sdev->request_queue->queuedata = sdev;
2227         __scsi_init_queue(sdev->host, sdev->request_queue);
2228         return sdev->request_queue;
2229 }
2230
2231 int scsi_mq_setup_tags(struct Scsi_Host *shost)
2232 {
2233         unsigned int cmd_size, sgl_size;
2234
2235         sgl_size = scsi_mq_sgl_size(shost);
2236         cmd_size = sizeof(struct scsi_cmnd) + shost->hostt->cmd_size + sgl_size;
2237         if (scsi_host_get_prot(shost))
2238                 cmd_size += sizeof(struct scsi_data_buffer) + sgl_size;
2239
2240         memset(&shost->tag_set, 0, sizeof(shost->tag_set));
2241         shost->tag_set.ops = &scsi_mq_ops;
2242         shost->tag_set.nr_hw_queues = shost->nr_hw_queues ? : 1;
2243         shost->tag_set.queue_depth = shost->can_queue;
2244         shost->tag_set.cmd_size = cmd_size;
2245         shost->tag_set.numa_node = NUMA_NO_NODE;
2246         shost->tag_set.flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_SG_MERGE;
2247         shost->tag_set.flags |=
2248                 BLK_ALLOC_POLICY_TO_MQ_FLAG(shost->hostt->tag_alloc_policy);
2249         shost->tag_set.driver_data = shost;
2250
2251         return blk_mq_alloc_tag_set(&shost->tag_set);
2252 }
2253
2254 void scsi_mq_destroy_tags(struct Scsi_Host *shost)
2255 {
2256         blk_mq_free_tag_set(&shost->tag_set);
2257 }
2258
2259 /**
2260  * scsi_device_from_queue - return sdev associated with a request_queue
2261  * @q: The request queue to return the sdev from
2262  *
2263  * Return the sdev associated with a request queue or NULL if the
2264  * request_queue does not reference a SCSI device.
2265  */
2266 struct scsi_device *scsi_device_from_queue(struct request_queue *q)
2267 {
2268         struct scsi_device *sdev = NULL;
2269
2270         if (q->mq_ops) {
2271                 if (q->mq_ops == &scsi_mq_ops)
2272                         sdev = q->queuedata;
2273         } else if (q->request_fn == scsi_request_fn)
2274                 sdev = q->queuedata;
2275         if (!sdev || !get_device(&sdev->sdev_gendev))
2276                 sdev = NULL;
2277
2278         return sdev;
2279 }
2280 EXPORT_SYMBOL_GPL(scsi_device_from_queue);
2281
2282 /*
2283  * Function:    scsi_block_requests()
2284  *
2285  * Purpose:     Utility function used by low-level drivers to prevent further
2286  *              commands from being queued to the device.
2287  *
2288  * Arguments:   shost       - Host in question
2289  *
2290  * Returns:     Nothing
2291  *
2292  * Lock status: No locks are assumed held.
2293  *
2294  * Notes:       There is no timer nor any other means by which the requests
2295  *              get unblocked other than the low-level driver calling
2296  *              scsi_unblock_requests().
2297  */
2298 void scsi_block_requests(struct Scsi_Host *shost)
2299 {
2300         shost->host_self_blocked = 1;
2301 }
2302 EXPORT_SYMBOL(scsi_block_requests);
2303
2304 /*
2305  * Function:    scsi_unblock_requests()
2306  *
2307  * Purpose:     Utility function used by low-level drivers to allow further
2308  *              commands from being queued to the device.
2309  *
2310  * Arguments:   shost       - Host in question
2311  *
2312  * Returns:     Nothing
2313  *
2314  * Lock status: No locks are assumed held.
2315  *
2316  * Notes:       There is no timer nor any other means by which the requests
2317  *              get unblocked other than the low-level driver calling
2318  *              scsi_unblock_requests().
2319  *
2320  *              This is done as an API function so that changes to the
2321  *              internals of the scsi mid-layer won't require wholesale
2322  *              changes to drivers that use this feature.
2323  */
2324 void scsi_unblock_requests(struct Scsi_Host *shost)
2325 {
2326         shost->host_self_blocked = 0;
2327         scsi_run_host_queues(shost);
2328 }
2329 EXPORT_SYMBOL(scsi_unblock_requests);
2330
2331 int __init scsi_init_queue(void)
2332 {
2333         scsi_sdb_cache = kmem_cache_create("scsi_data_buffer",
2334                                            sizeof(struct scsi_data_buffer),
2335                                            0, 0, NULL);
2336         if (!scsi_sdb_cache) {
2337                 printk(KERN_ERR "SCSI: can't init scsi sdb cache\n");
2338                 return -ENOMEM;
2339         }
2340
2341         return 0;
2342 }
2343
2344 void scsi_exit_queue(void)
2345 {
2346         kmem_cache_destroy(scsi_sense_cache);
2347         kmem_cache_destroy(scsi_sense_isadma_cache);
2348         kmem_cache_destroy(scsi_sdb_cache);
2349 }
2350
2351 /**
2352  *      scsi_mode_select - issue a mode select
2353  *      @sdev:  SCSI device to be queried
2354  *      @pf:    Page format bit (1 == standard, 0 == vendor specific)
2355  *      @sp:    Save page bit (0 == don't save, 1 == save)
2356  *      @modepage: mode page being requested
2357  *      @buffer: request buffer (may not be smaller than eight bytes)
2358  *      @len:   length of request buffer.
2359  *      @timeout: command timeout
2360  *      @retries: number of retries before failing
2361  *      @data: returns a structure abstracting the mode header data
2362  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
2363  *              must be SCSI_SENSE_BUFFERSIZE big.
2364  *
2365  *      Returns zero if successful; negative error number or scsi
2366  *      status on error
2367  *
2368  */
2369 int
2370 scsi_mode_select(struct scsi_device *sdev, int pf, int sp, int modepage,
2371                  unsigned char *buffer, int len, int timeout, int retries,
2372                  struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
2373 {
2374         unsigned char cmd[10];
2375         unsigned char *real_buffer;
2376         int ret;
2377
2378         memset(cmd, 0, sizeof(cmd));
2379         cmd[1] = (pf ? 0x10 : 0) | (sp ? 0x01 : 0);
2380
2381         if (sdev->use_10_for_ms) {
2382                 if (len > 65535)
2383                         return -EINVAL;
2384                 real_buffer = kmalloc(8 + len, GFP_KERNEL);
2385                 if (!real_buffer)
2386                         return -ENOMEM;
2387                 memcpy(real_buffer + 8, buffer, len);
2388                 len += 8;
2389                 real_buffer[0] = 0;
2390                 real_buffer[1] = 0;
2391                 real_buffer[2] = data->medium_type;
2392                 real_buffer[3] = data->device_specific;
2393                 real_buffer[4] = data->longlba ? 0x01 : 0;
2394                 real_buffer[5] = 0;
2395                 real_buffer[6] = data->block_descriptor_length >> 8;
2396                 real_buffer[7] = data->block_descriptor_length;
2397
2398                 cmd[0] = MODE_SELECT_10;
2399                 cmd[7] = len >> 8;
2400                 cmd[8] = len;
2401         } else {
2402                 if (len > 255 || data->block_descriptor_length > 255 ||
2403                     data->longlba)
2404                         return -EINVAL;
2405
2406                 real_buffer = kmalloc(4 + len, GFP_KERNEL);
2407                 if (!real_buffer)
2408                         return -ENOMEM;
2409                 memcpy(real_buffer + 4, buffer, len);
2410                 len += 4;
2411                 real_buffer[0] = 0;
2412                 real_buffer[1] = data->medium_type;
2413                 real_buffer[2] = data->device_specific;
2414                 real_buffer[3] = data->block_descriptor_length;
2415                 
2416
2417                 cmd[0] = MODE_SELECT;
2418                 cmd[4] = len;
2419         }
2420
2421         ret = scsi_execute_req(sdev, cmd, DMA_TO_DEVICE, real_buffer, len,
2422                                sshdr, timeout, retries, NULL);
2423         kfree(real_buffer);
2424         return ret;
2425 }
2426 EXPORT_SYMBOL_GPL(scsi_mode_select);
2427
2428 /**
2429  *      scsi_mode_sense - issue a mode sense, falling back from 10 to six bytes if necessary.
2430  *      @sdev:  SCSI device to be queried
2431  *      @dbd:   set if mode sense will allow block descriptors to be returned
2432  *      @modepage: mode page being requested
2433  *      @buffer: request buffer (may not be smaller than eight bytes)
2434  *      @len:   length of request buffer.
2435  *      @timeout: command timeout
2436  *      @retries: number of retries before failing
2437  *      @data: returns a structure abstracting the mode header data
2438  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
2439  *              must be SCSI_SENSE_BUFFERSIZE big.
2440  *
2441  *      Returns zero if unsuccessful, or the header offset (either 4
2442  *      or 8 depending on whether a six or ten byte command was
2443  *      issued) if successful.
2444  */
2445 int
2446 scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
2447                   unsigned char *buffer, int len, int timeout, int retries,
2448                   struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
2449 {
2450         unsigned char cmd[12];
2451         int use_10_for_ms;
2452         int header_length;
2453         int result, retry_count = retries;
2454         struct scsi_sense_hdr my_sshdr;
2455
2456         memset(data, 0, sizeof(*data));
2457         memset(&cmd[0], 0, 12);
2458         cmd[1] = dbd & 0x18;    /* allows DBD and LLBA bits */
2459         cmd[2] = modepage;
2460
2461         /* caller might not be interested in sense, but we need it */
2462         if (!sshdr)
2463                 sshdr = &my_sshdr;
2464
2465  retry:
2466         use_10_for_ms = sdev->use_10_for_ms;
2467
2468         if (use_10_for_ms) {
2469                 if (len < 8)
2470                         len = 8;
2471
2472                 cmd[0] = MODE_SENSE_10;
2473                 cmd[8] = len;
2474                 header_length = 8;
2475         } else {
2476                 if (len < 4)
2477                         len = 4;
2478
2479                 cmd[0] = MODE_SENSE;
2480                 cmd[4] = len;
2481                 header_length = 4;
2482         }
2483
2484         memset(buffer, 0, len);
2485
2486         result = scsi_execute_req(sdev, cmd, DMA_FROM_DEVICE, buffer, len,
2487                                   sshdr, timeout, retries, NULL);
2488
2489         /* This code looks awful: what it's doing is making sure an
2490          * ILLEGAL REQUEST sense return identifies the actual command
2491          * byte as the problem.  MODE_SENSE commands can return
2492          * ILLEGAL REQUEST if the code page isn't supported */
2493
2494         if (use_10_for_ms && !scsi_status_is_good(result) &&
2495             (driver_byte(result) & DRIVER_SENSE)) {
2496                 if (scsi_sense_valid(sshdr)) {
2497                         if ((sshdr->sense_key == ILLEGAL_REQUEST) &&
2498                             (sshdr->asc == 0x20) && (sshdr->ascq == 0)) {
2499                                 /* 
2500                                  * Invalid command operation code
2501                                  */
2502                                 sdev->use_10_for_ms = 0;
2503                                 goto retry;
2504                         }
2505                 }
2506         }
2507
2508         if(scsi_status_is_good(result)) {
2509                 if (unlikely(buffer[0] == 0x86 && buffer[1] == 0x0b &&
2510                              (modepage == 6 || modepage == 8))) {
2511                         /* Initio breakage? */
2512                         header_length = 0;
2513                         data->length = 13;
2514                         data->medium_type = 0;
2515                         data->device_specific = 0;
2516                         data->longlba = 0;
2517                         data->block_descriptor_length = 0;
2518                 } else if(use_10_for_ms) {
2519                         data->length = buffer[0]*256 + buffer[1] + 2;
2520                         data->medium_type = buffer[2];
2521                         data->device_specific = buffer[3];
2522                         data->longlba = buffer[4] & 0x01;
2523                         data->block_descriptor_length = buffer[6]*256
2524                                 + buffer[7];
2525                 } else {
2526                         data->length = buffer[0] + 1;
2527                         data->medium_type = buffer[1];
2528                         data->device_specific = buffer[2];
2529                         data->block_descriptor_length = buffer[3];
2530                 }
2531                 data->header_length = header_length;
2532         } else if ((status_byte(result) == CHECK_CONDITION) &&
2533                    scsi_sense_valid(sshdr) &&
2534                    sshdr->sense_key == UNIT_ATTENTION && retry_count) {
2535                 retry_count--;
2536                 goto retry;
2537         }
2538
2539         return result;
2540 }
2541 EXPORT_SYMBOL(scsi_mode_sense);
2542
2543 /**
2544  *      scsi_test_unit_ready - test if unit is ready
2545  *      @sdev:  scsi device to change the state of.
2546  *      @timeout: command timeout
2547  *      @retries: number of retries before failing
2548  *      @sshdr: outpout pointer for decoded sense information.
2549  *
2550  *      Returns zero if unsuccessful or an error if TUR failed.  For
2551  *      removable media, UNIT_ATTENTION sets ->changed flag.
2552  **/
2553 int
2554 scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries,
2555                      struct scsi_sense_hdr *sshdr)
2556 {
2557         char cmd[] = {
2558                 TEST_UNIT_READY, 0, 0, 0, 0, 0,
2559         };
2560         int result;
2561
2562         /* try to eat the UNIT_ATTENTION if there are enough retries */
2563         do {
2564                 result = scsi_execute_req(sdev, cmd, DMA_NONE, NULL, 0, sshdr,
2565                                           timeout, retries, NULL);
2566                 if (sdev->removable && scsi_sense_valid(sshdr) &&
2567                     sshdr->sense_key == UNIT_ATTENTION)
2568                         sdev->changed = 1;
2569         } while (scsi_sense_valid(sshdr) &&
2570                  sshdr->sense_key == UNIT_ATTENTION && --retries);
2571
2572         return result;
2573 }
2574 EXPORT_SYMBOL(scsi_test_unit_ready);
2575
2576 /**
2577  *      scsi_device_set_state - Take the given device through the device state model.
2578  *      @sdev:  scsi device to change the state of.
2579  *      @state: state to change to.
2580  *
2581  *      Returns zero if successful or an error if the requested
2582  *      transition is illegal.
2583  */
2584 int
2585 scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
2586 {
2587         enum scsi_device_state oldstate = sdev->sdev_state;
2588
2589         if (state == oldstate)
2590                 return 0;
2591
2592         switch (state) {
2593         case SDEV_CREATED:
2594                 switch (oldstate) {
2595                 case SDEV_CREATED_BLOCK:
2596                         break;
2597                 default:
2598                         goto illegal;
2599                 }
2600                 break;
2601                         
2602         case SDEV_RUNNING:
2603                 switch (oldstate) {
2604                 case SDEV_CREATED:
2605                 case SDEV_OFFLINE:
2606                 case SDEV_TRANSPORT_OFFLINE:
2607                 case SDEV_QUIESCE:
2608                 case SDEV_BLOCK:
2609                         break;
2610                 default:
2611                         goto illegal;
2612                 }
2613                 break;
2614
2615         case SDEV_QUIESCE:
2616                 switch (oldstate) {
2617                 case SDEV_RUNNING:
2618                 case SDEV_OFFLINE:
2619                 case SDEV_TRANSPORT_OFFLINE:
2620                         break;
2621                 default:
2622                         goto illegal;
2623                 }
2624                 break;
2625
2626         case SDEV_OFFLINE:
2627         case SDEV_TRANSPORT_OFFLINE:
2628                 switch (oldstate) {
2629                 case SDEV_CREATED:
2630                 case SDEV_RUNNING:
2631                 case SDEV_QUIESCE:
2632                 case SDEV_BLOCK:
2633                         break;
2634                 default:
2635                         goto illegal;
2636                 }
2637                 break;
2638
2639         case SDEV_BLOCK:
2640                 switch (oldstate) {
2641                 case SDEV_RUNNING:
2642                 case SDEV_CREATED_BLOCK:
2643                         break;
2644                 default:
2645                         goto illegal;
2646                 }
2647                 break;
2648
2649         case SDEV_CREATED_BLOCK:
2650                 switch (oldstate) {
2651                 case SDEV_CREATED:
2652                         break;
2653                 default:
2654                         goto illegal;
2655                 }
2656                 break;
2657
2658         case SDEV_CANCEL:
2659                 switch (oldstate) {
2660                 case SDEV_CREATED:
2661                 case SDEV_RUNNING:
2662                 case SDEV_QUIESCE:
2663                 case SDEV_OFFLINE:
2664                 case SDEV_TRANSPORT_OFFLINE:
2665                         break;
2666                 default:
2667                         goto illegal;
2668                 }
2669                 break;
2670
2671         case SDEV_DEL:
2672                 switch (oldstate) {
2673                 case SDEV_CREATED:
2674                 case SDEV_RUNNING:
2675                 case SDEV_OFFLINE:
2676                 case SDEV_TRANSPORT_OFFLINE:
2677                 case SDEV_CANCEL:
2678                 case SDEV_BLOCK:
2679                 case SDEV_CREATED_BLOCK:
2680                         break;
2681                 default:
2682                         goto illegal;
2683                 }
2684                 break;
2685
2686         }
2687         sdev->sdev_state = state;
2688         sysfs_notify(&sdev->sdev_gendev.kobj, NULL, "state");
2689         return 0;
2690
2691  illegal:
2692         SCSI_LOG_ERROR_RECOVERY(1,
2693                                 sdev_printk(KERN_ERR, sdev,
2694                                             "Illegal state transition %s->%s",
2695                                             scsi_device_state_name(oldstate),
2696                                             scsi_device_state_name(state))
2697                                 );
2698         return -EINVAL;
2699 }
2700 EXPORT_SYMBOL(scsi_device_set_state);
2701
2702 /**
2703  *      sdev_evt_emit - emit a single SCSI device uevent
2704  *      @sdev: associated SCSI device
2705  *      @evt: event to emit
2706  *
2707  *      Send a single uevent (scsi_event) to the associated scsi_device.
2708  */
2709 static void scsi_evt_emit(struct scsi_device *sdev, struct scsi_event *evt)
2710 {
2711         int idx = 0;
2712         char *envp[3];
2713
2714         switch (evt->evt_type) {
2715         case SDEV_EVT_MEDIA_CHANGE:
2716                 envp[idx++] = "SDEV_MEDIA_CHANGE=1";
2717                 break;
2718         case SDEV_EVT_INQUIRY_CHANGE_REPORTED:
2719                 scsi_rescan_device(&sdev->sdev_gendev);
2720                 envp[idx++] = "SDEV_UA=INQUIRY_DATA_HAS_CHANGED";
2721                 break;
2722         case SDEV_EVT_CAPACITY_CHANGE_REPORTED:
2723                 envp[idx++] = "SDEV_UA=CAPACITY_DATA_HAS_CHANGED";
2724                 break;
2725         case SDEV_EVT_SOFT_THRESHOLD_REACHED_REPORTED:
2726                envp[idx++] = "SDEV_UA=THIN_PROVISIONING_SOFT_THRESHOLD_REACHED";
2727                 break;
2728         case SDEV_EVT_MODE_PARAMETER_CHANGE_REPORTED:
2729                 envp[idx++] = "SDEV_UA=MODE_PARAMETERS_CHANGED";
2730                 break;
2731         case SDEV_EVT_LUN_CHANGE_REPORTED:
2732                 envp[idx++] = "SDEV_UA=REPORTED_LUNS_DATA_HAS_CHANGED";
2733                 break;
2734         case SDEV_EVT_ALUA_STATE_CHANGE_REPORTED:
2735                 envp[idx++] = "SDEV_UA=ASYMMETRIC_ACCESS_STATE_CHANGED";
2736                 break;
2737         default:
2738                 /* do nothing */
2739                 break;
2740         }
2741
2742         envp[idx++] = NULL;
2743
2744         kobject_uevent_env(&sdev->sdev_gendev.kobj, KOBJ_CHANGE, envp);
2745 }
2746
2747 /**
2748  *      sdev_evt_thread - send a uevent for each scsi event
2749  *      @work: work struct for scsi_device
2750  *
2751  *      Dispatch queued events to their associated scsi_device kobjects
2752  *      as uevents.
2753  */
2754 void scsi_evt_thread(struct work_struct *work)
2755 {
2756         struct scsi_device *sdev;
2757         enum scsi_device_event evt_type;
2758         LIST_HEAD(event_list);
2759
2760         sdev = container_of(work, struct scsi_device, event_work);
2761
2762         for (evt_type = SDEV_EVT_FIRST; evt_type <= SDEV_EVT_LAST; evt_type++)
2763                 if (test_and_clear_bit(evt_type, sdev->pending_events))
2764                         sdev_evt_send_simple(sdev, evt_type, GFP_KERNEL);
2765
2766         while (1) {
2767                 struct scsi_event *evt;
2768                 struct list_head *this, *tmp;
2769                 unsigned long flags;
2770
2771                 spin_lock_irqsave(&sdev->list_lock, flags);
2772                 list_splice_init(&sdev->event_list, &event_list);
2773                 spin_unlock_irqrestore(&sdev->list_lock, flags);
2774
2775                 if (list_empty(&event_list))
2776                         break;
2777
2778                 list_for_each_safe(this, tmp, &event_list) {
2779                         evt = list_entry(this, struct scsi_event, node);
2780                         list_del(&evt->node);
2781                         scsi_evt_emit(sdev, evt);
2782                         kfree(evt);
2783                 }
2784         }
2785 }
2786
2787 /**
2788  *      sdev_evt_send - send asserted event to uevent thread
2789  *      @sdev: scsi_device event occurred on
2790  *      @evt: event to send
2791  *
2792  *      Assert scsi device event asynchronously.
2793  */
2794 void sdev_evt_send(struct scsi_device *sdev, struct scsi_event *evt)
2795 {
2796         unsigned long flags;
2797
2798 #if 0
2799         /* FIXME: currently this check eliminates all media change events
2800          * for polled devices.  Need to update to discriminate between AN
2801          * and polled events */
2802         if (!test_bit(evt->evt_type, sdev->supported_events)) {
2803                 kfree(evt);
2804                 return;
2805         }
2806 #endif
2807
2808         spin_lock_irqsave(&sdev->list_lock, flags);
2809         list_add_tail(&evt->node, &sdev->event_list);
2810         schedule_work(&sdev->event_work);
2811         spin_unlock_irqrestore(&sdev->list_lock, flags);
2812 }
2813 EXPORT_SYMBOL_GPL(sdev_evt_send);
2814
2815 /**
2816  *      sdev_evt_alloc - allocate a new scsi event
2817  *      @evt_type: type of event to allocate
2818  *      @gfpflags: GFP flags for allocation
2819  *
2820  *      Allocates and returns a new scsi_event.
2821  */
2822 struct scsi_event *sdev_evt_alloc(enum scsi_device_event evt_type,
2823                                   gfp_t gfpflags)
2824 {
2825         struct scsi_event *evt = kzalloc(sizeof(struct scsi_event), gfpflags);
2826         if (!evt)
2827                 return NULL;
2828
2829         evt->evt_type = evt_type;
2830         INIT_LIST_HEAD(&evt->node);
2831
2832         /* evt_type-specific initialization, if any */
2833         switch (evt_type) {
2834         case SDEV_EVT_MEDIA_CHANGE:
2835         case SDEV_EVT_INQUIRY_CHANGE_REPORTED:
2836         case SDEV_EVT_CAPACITY_CHANGE_REPORTED:
2837         case SDEV_EVT_SOFT_THRESHOLD_REACHED_REPORTED:
2838         case SDEV_EVT_MODE_PARAMETER_CHANGE_REPORTED:
2839         case SDEV_EVT_LUN_CHANGE_REPORTED:
2840         case SDEV_EVT_ALUA_STATE_CHANGE_REPORTED:
2841         default:
2842                 /* do nothing */
2843                 break;
2844         }
2845
2846         return evt;
2847 }
2848 EXPORT_SYMBOL_GPL(sdev_evt_alloc);
2849
2850 /**
2851  *      sdev_evt_send_simple - send asserted event to uevent thread
2852  *      @sdev: scsi_device event occurred on
2853  *      @evt_type: type of event to send
2854  *      @gfpflags: GFP flags for allocation
2855  *
2856  *      Assert scsi device event asynchronously, given an event type.
2857  */
2858 void sdev_evt_send_simple(struct scsi_device *sdev,
2859                           enum scsi_device_event evt_type, gfp_t gfpflags)
2860 {
2861         struct scsi_event *evt = sdev_evt_alloc(evt_type, gfpflags);
2862         if (!evt) {
2863                 sdev_printk(KERN_ERR, sdev, "event %d eaten due to OOM\n",
2864                             evt_type);
2865                 return;
2866         }
2867
2868         sdev_evt_send(sdev, evt);
2869 }
2870 EXPORT_SYMBOL_GPL(sdev_evt_send_simple);
2871
2872 /**
2873  * scsi_request_fn_active() - number of kernel threads inside scsi_request_fn()
2874  * @sdev: SCSI device to count the number of scsi_request_fn() callers for.
2875  */
2876 static int scsi_request_fn_active(struct scsi_device *sdev)
2877 {
2878         struct request_queue *q = sdev->request_queue;
2879         int request_fn_active;
2880
2881         WARN_ON_ONCE(sdev->host->use_blk_mq);
2882
2883         spin_lock_irq(q->queue_lock);
2884         request_fn_active = q->request_fn_active;
2885         spin_unlock_irq(q->queue_lock);
2886
2887         return request_fn_active;
2888 }
2889
2890 /**
2891  * scsi_wait_for_queuecommand() - wait for ongoing queuecommand() calls
2892  * @sdev: SCSI device pointer.
2893  *
2894  * Wait until the ongoing shost->hostt->queuecommand() calls that are
2895  * invoked from scsi_request_fn() have finished.
2896  */
2897 static void scsi_wait_for_queuecommand(struct scsi_device *sdev)
2898 {
2899         WARN_ON_ONCE(sdev->host->use_blk_mq);
2900
2901         while (scsi_request_fn_active(sdev))
2902                 msleep(20);
2903 }
2904
2905 /**
2906  *      scsi_device_quiesce - Block user issued commands.
2907  *      @sdev:  scsi device to quiesce.
2908  *
2909  *      This works by trying to transition to the SDEV_QUIESCE state
2910  *      (which must be a legal transition).  When the device is in this
2911  *      state, only special requests will be accepted, all others will
2912  *      be deferred.  Since special requests may also be requeued requests,
2913  *      a successful return doesn't guarantee the device will be 
2914  *      totally quiescent.
2915  *
2916  *      Must be called with user context, may sleep.
2917  *
2918  *      Returns zero if unsuccessful or an error if not.
2919  */
2920 int
2921 scsi_device_quiesce(struct scsi_device *sdev)
2922 {
2923         int err;
2924
2925         mutex_lock(&sdev->state_mutex);
2926         err = scsi_device_set_state(sdev, SDEV_QUIESCE);
2927         mutex_unlock(&sdev->state_mutex);
2928
2929         if (err)
2930                 return err;
2931
2932         scsi_run_queue(sdev->request_queue);
2933         while (atomic_read(&sdev->device_busy)) {
2934                 msleep_interruptible(200);
2935                 scsi_run_queue(sdev->request_queue);
2936         }
2937         return 0;
2938 }
2939 EXPORT_SYMBOL(scsi_device_quiesce);
2940
2941 /**
2942  *      scsi_device_resume - Restart user issued commands to a quiesced device.
2943  *      @sdev:  scsi device to resume.
2944  *
2945  *      Moves the device from quiesced back to running and restarts the
2946  *      queues.
2947  *
2948  *      Must be called with user context, may sleep.
2949  */
2950 void scsi_device_resume(struct scsi_device *sdev)
2951 {
2952         /* check if the device state was mutated prior to resume, and if
2953          * so assume the state is being managed elsewhere (for example
2954          * device deleted during suspend)
2955          */
2956         mutex_lock(&sdev->state_mutex);
2957         if (sdev->sdev_state == SDEV_QUIESCE &&
2958             scsi_device_set_state(sdev, SDEV_RUNNING) == 0)
2959                 scsi_run_queue(sdev->request_queue);
2960         mutex_unlock(&sdev->state_mutex);
2961 }
2962 EXPORT_SYMBOL(scsi_device_resume);
2963
2964 static void
2965 device_quiesce_fn(struct scsi_device *sdev, void *data)
2966 {
2967         scsi_device_quiesce(sdev);
2968 }
2969
2970 void
2971 scsi_target_quiesce(struct scsi_target *starget)
2972 {
2973         starget_for_each_device(starget, NULL, device_quiesce_fn);
2974 }
2975 EXPORT_SYMBOL(scsi_target_quiesce);
2976
2977 static void
2978 device_resume_fn(struct scsi_device *sdev, void *data)
2979 {
2980         scsi_device_resume(sdev);
2981 }
2982
2983 void
2984 scsi_target_resume(struct scsi_target *starget)
2985 {
2986         starget_for_each_device(starget, NULL, device_resume_fn);
2987 }
2988 EXPORT_SYMBOL(scsi_target_resume);
2989
2990 /**
2991  * scsi_internal_device_block_nowait - try to transition to the SDEV_BLOCK state
2992  * @sdev: device to block
2993  *
2994  * Pause SCSI command processing on the specified device. Does not sleep.
2995  *
2996  * Returns zero if successful or a negative error code upon failure.
2997  *
2998  * Notes:
2999  * This routine transitions the device to the SDEV_BLOCK state (which must be
3000  * a legal transition). When the device is in this state, command processing
3001  * is paused until the device leaves the SDEV_BLOCK state. See also
3002  * scsi_internal_device_unblock_nowait().
3003  */
3004 int scsi_internal_device_block_nowait(struct scsi_device *sdev)
3005 {
3006         struct request_queue *q = sdev->request_queue;
3007         unsigned long flags;
3008         int err = 0;
3009
3010         err = scsi_device_set_state(sdev, SDEV_BLOCK);
3011         if (err) {
3012                 err = scsi_device_set_state(sdev, SDEV_CREATED_BLOCK);
3013
3014                 if (err)
3015                         return err;
3016         }
3017
3018         /* 
3019          * The device has transitioned to SDEV_BLOCK.  Stop the
3020          * block layer from calling the midlayer with this device's
3021          * request queue. 
3022          */
3023         if (q->mq_ops) {
3024                 blk_mq_quiesce_queue_nowait(q);
3025         } else {
3026                 spin_lock_irqsave(q->queue_lock, flags);
3027                 blk_stop_queue(q);
3028                 spin_unlock_irqrestore(q->queue_lock, flags);
3029         }
3030
3031         return 0;
3032 }
3033 EXPORT_SYMBOL_GPL(scsi_internal_device_block_nowait);
3034
3035 /**
3036  * scsi_internal_device_block - try to transition to the SDEV_BLOCK state
3037  * @sdev: device to block
3038  *
3039  * Pause SCSI command processing on the specified device and wait until all
3040  * ongoing scsi_request_fn() / scsi_queue_rq() calls have finished. May sleep.
3041  *
3042  * Returns zero if successful or a negative error code upon failure.
3043  *
3044  * Note:
3045  * This routine transitions the device to the SDEV_BLOCK state (which must be
3046  * a legal transition). When the device is in this state, command processing
3047  * is paused until the device leaves the SDEV_BLOCK state. See also
3048  * scsi_internal_device_unblock().
3049  *
3050  * To do: avoid that scsi_send_eh_cmnd() calls queuecommand() after
3051  * scsi_internal_device_block() has blocked a SCSI device and also
3052  * remove the rport mutex lock and unlock calls from srp_queuecommand().
3053  */
3054 static int scsi_internal_device_block(struct scsi_device *sdev)
3055 {
3056         struct request_queue *q = sdev->request_queue;
3057         int err;
3058
3059         mutex_lock(&sdev->state_mutex);
3060         err = scsi_internal_device_block_nowait(sdev);
3061         if (err == 0) {
3062                 if (q->mq_ops)
3063                         blk_mq_quiesce_queue(q);
3064                 else
3065                         scsi_wait_for_queuecommand(sdev);
3066         }
3067         mutex_unlock(&sdev->state_mutex);
3068
3069         return err;
3070 }
3071  
3072 void scsi_start_queue(struct scsi_device *sdev)
3073 {
3074         struct request_queue *q = sdev->request_queue;
3075         unsigned long flags;
3076
3077         if (q->mq_ops) {
3078                 blk_mq_unquiesce_queue(q);
3079         } else {
3080                 spin_lock_irqsave(q->queue_lock, flags);
3081                 blk_start_queue(q);
3082                 spin_unlock_irqrestore(q->queue_lock, flags);
3083         }
3084 }
3085
3086 /**
3087  * scsi_internal_device_unblock_nowait - resume a device after a block request
3088  * @sdev:       device to resume
3089  * @new_state:  state to set the device to after unblocking
3090  *
3091  * Restart the device queue for a previously suspended SCSI device. Does not
3092  * sleep.
3093  *
3094  * Returns zero if successful or a negative error code upon failure.
3095  *
3096  * Notes:
3097  * This routine transitions the device to the SDEV_RUNNING state or to one of
3098  * the offline states (which must be a legal transition) allowing the midlayer
3099  * to goose the queue for this device.
3100  */
3101 int scsi_internal_device_unblock_nowait(struct scsi_device *sdev,
3102                                         enum scsi_device_state new_state)
3103 {
3104         /*
3105          * Try to transition the scsi device to SDEV_RUNNING or one of the
3106          * offlined states and goose the device queue if successful.
3107          */
3108         switch (sdev->sdev_state) {
3109         case SDEV_BLOCK:
3110         case SDEV_TRANSPORT_OFFLINE:
3111                 sdev->sdev_state = new_state;
3112                 sysfs_notify(&sdev->sdev_gendev.kobj, NULL, "state");
3113                 break;
3114         case SDEV_CREATED_BLOCK:
3115                 if (new_state == SDEV_TRANSPORT_OFFLINE ||
3116                     new_state == SDEV_OFFLINE)
3117                         sdev->sdev_state = new_state;
3118                 else
3119                         sdev->sdev_state = SDEV_CREATED;
3120                 sysfs_notify(&sdev->sdev_gendev.kobj, NULL, "state");
3121                 break;
3122         case SDEV_CANCEL:
3123         case SDEV_OFFLINE:
3124                 break;
3125         default:
3126                 return -EINVAL;
3127         }
3128         scsi_start_queue(sdev);
3129
3130         return 0;
3131 }
3132 EXPORT_SYMBOL_GPL(scsi_internal_device_unblock_nowait);
3133
3134 /**
3135  * scsi_internal_device_unblock - resume a device after a block request
3136  * @sdev:       device to resume
3137  * @new_state:  state to set the device to after unblocking
3138  *
3139  * Restart the device queue for a previously suspended SCSI device. May sleep.
3140  *
3141  * Returns zero if successful or a negative error code upon failure.
3142  *
3143  * Notes:
3144  * This routine transitions the device to the SDEV_RUNNING state or to one of
3145  * the offline states (which must be a legal transition) allowing the midlayer
3146  * to goose the queue for this device.
3147  */
3148 static int scsi_internal_device_unblock(struct scsi_device *sdev,
3149                                         enum scsi_device_state new_state)
3150 {
3151         int ret;
3152
3153         mutex_lock(&sdev->state_mutex);
3154         ret = scsi_internal_device_unblock_nowait(sdev, new_state);
3155         mutex_unlock(&sdev->state_mutex);
3156
3157         return ret;
3158 }
3159
3160 static void
3161 device_block(struct scsi_device *sdev, void *data)
3162 {
3163         scsi_internal_device_block(sdev);
3164 }
3165
3166 static int
3167 target_block(struct device *dev, void *data)
3168 {
3169         if (scsi_is_target_device(dev))
3170                 starget_for_each_device(to_scsi_target(dev), NULL,
3171                                         device_block);
3172         return 0;
3173 }
3174
3175 void
3176 scsi_target_block(struct device *dev)
3177 {
3178         if (scsi_is_target_device(dev))
3179                 starget_for_each_device(to_scsi_target(dev), NULL,
3180                                         device_block);
3181         else
3182                 device_for_each_child(dev, NULL, target_block);
3183 }
3184 EXPORT_SYMBOL_GPL(scsi_target_block);
3185
3186 static void
3187 device_unblock(struct scsi_device *sdev, void *data)
3188 {
3189         scsi_internal_device_unblock(sdev, *(enum scsi_device_state *)data);
3190 }
3191
3192 static int
3193 target_unblock(struct device *dev, void *data)
3194 {
3195         if (scsi_is_target_device(dev))
3196                 starget_for_each_device(to_scsi_target(dev), data,
3197                                         device_unblock);
3198         return 0;
3199 }
3200
3201 void
3202 scsi_target_unblock(struct device *dev, enum scsi_device_state new_state)
3203 {
3204         if (scsi_is_target_device(dev))
3205                 starget_for_each_device(to_scsi_target(dev), &new_state,
3206                                         device_unblock);
3207         else
3208                 device_for_each_child(dev, &new_state, target_unblock);
3209 }
3210 EXPORT_SYMBOL_GPL(scsi_target_unblock);
3211
3212 /**
3213  * scsi_kmap_atomic_sg - find and atomically map an sg-elemnt
3214  * @sgl:        scatter-gather list
3215  * @sg_count:   number of segments in sg
3216  * @offset:     offset in bytes into sg, on return offset into the mapped area
3217  * @len:        bytes to map, on return number of bytes mapped
3218  *
3219  * Returns virtual address of the start of the mapped page
3220  */
3221 void *scsi_kmap_atomic_sg(struct scatterlist *sgl, int sg_count,
3222                           size_t *offset, size_t *len)
3223 {
3224         int i;
3225         size_t sg_len = 0, len_complete = 0;
3226         struct scatterlist *sg;
3227         struct page *page;
3228
3229         WARN_ON(!irqs_disabled());
3230
3231         for_each_sg(sgl, sg, sg_count, i) {
3232                 len_complete = sg_len; /* Complete sg-entries */
3233                 sg_len += sg->length;
3234                 if (sg_len > *offset)
3235                         break;
3236         }
3237
3238         if (unlikely(i == sg_count)) {
3239                 printk(KERN_ERR "%s: Bytes in sg: %zu, requested offset %zu, "
3240                         "elements %d\n",
3241                        __func__, sg_len, *offset, sg_count);
3242                 WARN_ON(1);
3243                 return NULL;
3244         }