SCSI: fix queue cleanup race before queue initialization is done
[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_usercopy("scsi_sense_cache",
89                                 SCSI_SENSE_BUFFERSIZE, 0, SLAB_HWCACHE_ALIGN,
90                                 0, SCSI_SENSE_BUFFERSIZE, NULL);
91                 if (!scsi_sense_cache)
92                         ret = -ENOMEM;
93         }
94
95         mutex_unlock(&scsi_sense_cache_mutex);
96         return ret;
97 }
98
99 /*
100  * When to reinvoke queueing after a resource shortage. It's 3 msecs to
101  * not change behaviour from the previous unplug mechanism, experimentation
102  * may prove this needs changing.
103  */
104 #define SCSI_QUEUE_DELAY        3
105
106 static void
107 scsi_set_blocked(struct scsi_cmnd *cmd, int reason)
108 {
109         struct Scsi_Host *host = cmd->device->host;
110         struct scsi_device *device = cmd->device;
111         struct scsi_target *starget = scsi_target(device);
112
113         /*
114          * Set the appropriate busy bit for the device/host.
115          *
116          * If the host/device isn't busy, assume that something actually
117          * completed, and that we should be able to queue a command now.
118          *
119          * Note that the prior mid-layer assumption that any host could
120          * always queue at least one command is now broken.  The mid-layer
121          * will implement a user specifiable stall (see
122          * scsi_host.max_host_blocked and scsi_device.max_device_blocked)
123          * if a command is requeued with no other commands outstanding
124          * either for the device or for the host.
125          */
126         switch (reason) {
127         case SCSI_MLQUEUE_HOST_BUSY:
128                 atomic_set(&host->host_blocked, host->max_host_blocked);
129                 break;
130         case SCSI_MLQUEUE_DEVICE_BUSY:
131         case SCSI_MLQUEUE_EH_RETRY:
132                 atomic_set(&device->device_blocked,
133                            device->max_device_blocked);
134                 break;
135         case SCSI_MLQUEUE_TARGET_BUSY:
136                 atomic_set(&starget->target_blocked,
137                            starget->max_target_blocked);
138                 break;
139         }
140 }
141
142 static void scsi_mq_requeue_cmd(struct scsi_cmnd *cmd)
143 {
144         struct scsi_device *sdev = cmd->device;
145
146         if (cmd->request->rq_flags & RQF_DONTPREP) {
147                 cmd->request->rq_flags &= ~RQF_DONTPREP;
148                 scsi_mq_uninit_cmd(cmd);
149         } else {
150                 WARN_ON_ONCE(true);
151         }
152         blk_mq_requeue_request(cmd->request, true);
153         put_device(&sdev->sdev_gendev);
154 }
155
156 /**
157  * __scsi_queue_insert - private queue insertion
158  * @cmd: The SCSI command being requeued
159  * @reason:  The reason for the requeue
160  * @unbusy: Whether the queue should be unbusied
161  *
162  * This is a private queue insertion.  The public interface
163  * scsi_queue_insert() always assumes the queue should be unbusied
164  * because it's always called before the completion.  This function is
165  * for a requeue after completion, which should only occur in this
166  * file.
167  */
168 static void __scsi_queue_insert(struct scsi_cmnd *cmd, int reason, bool unbusy)
169 {
170         struct scsi_device *device = cmd->device;
171         struct request_queue *q = device->request_queue;
172         unsigned long flags;
173
174         SCSI_LOG_MLQUEUE(1, scmd_printk(KERN_INFO, cmd,
175                 "Inserting command %p into mlqueue\n", cmd));
176
177         scsi_set_blocked(cmd, reason);
178
179         /*
180          * Decrement the counters, since these commands are no longer
181          * active on the host/device.
182          */
183         if (unbusy)
184                 scsi_device_unbusy(device);
185
186         /*
187          * Requeue this command.  It will go before all other commands
188          * that are already in the queue. Schedule requeue work under
189          * lock such that the kblockd_schedule_work() call happens
190          * before blk_cleanup_queue() finishes.
191          */
192         cmd->result = 0;
193         if (q->mq_ops) {
194                 /*
195                  * Before a SCSI command is dispatched,
196                  * get_device(&sdev->sdev_gendev) is called and the host,
197                  * target and device busy counters are increased. Since
198                  * requeuing a request causes these actions to be repeated and
199                  * since scsi_device_unbusy() has already been called,
200                  * put_device(&device->sdev_gendev) must still be called. Call
201                  * put_device() after blk_mq_requeue_request() to avoid that
202                  * removal of the SCSI device can start before requeueing has
203                  * happened.
204                  */
205                 blk_mq_requeue_request(cmd->request, true);
206                 put_device(&device->sdev_gendev);
207                 return;
208         }
209         spin_lock_irqsave(q->queue_lock, flags);
210         blk_requeue_request(q, cmd->request);
211         kblockd_schedule_work(&device->requeue_work);
212         spin_unlock_irqrestore(q->queue_lock, flags);
213 }
214
215 /*
216  * Function:    scsi_queue_insert()
217  *
218  * Purpose:     Insert a command in the midlevel queue.
219  *
220  * Arguments:   cmd    - command that we are adding to queue.
221  *              reason - why we are inserting command to queue.
222  *
223  * Lock status: Assumed that lock is not held upon entry.
224  *
225  * Returns:     Nothing.
226  *
227  * Notes:       We do this for one of two cases.  Either the host is busy
228  *              and it cannot accept any more commands for the time being,
229  *              or the device returned QUEUE_FULL and can accept no more
230  *              commands.
231  * Notes:       This could be called either from an interrupt context or a
232  *              normal process context.
233  */
234 void scsi_queue_insert(struct scsi_cmnd *cmd, int reason)
235 {
236         __scsi_queue_insert(cmd, reason, true);
237 }
238
239
240 /**
241  * __scsi_execute - insert request and wait for the result
242  * @sdev:       scsi device
243  * @cmd:        scsi command
244  * @data_direction: data direction
245  * @buffer:     data buffer
246  * @bufflen:    len of buffer
247  * @sense:      optional sense buffer
248  * @sshdr:      optional decoded sense header
249  * @timeout:    request timeout in seconds
250  * @retries:    number of times to retry request
251  * @flags:      flags for ->cmd_flags
252  * @rq_flags:   flags for ->rq_flags
253  * @resid:      optional residual length
254  *
255  * Returns the scsi_cmnd result field if a command was executed, or a negative
256  * Linux error code if we didn't get that far.
257  */
258 int __scsi_execute(struct scsi_device *sdev, const unsigned char *cmd,
259                  int data_direction, void *buffer, unsigned bufflen,
260                  unsigned char *sense, struct scsi_sense_hdr *sshdr,
261                  int timeout, int retries, u64 flags, req_flags_t rq_flags,
262                  int *resid)
263 {
264         struct request *req;
265         struct scsi_request *rq;
266         int ret = DRIVER_ERROR << 24;
267
268         req = blk_get_request(sdev->request_queue,
269                         data_direction == DMA_TO_DEVICE ?
270                         REQ_OP_SCSI_OUT : REQ_OP_SCSI_IN, BLK_MQ_REQ_PREEMPT);
271         if (IS_ERR(req))
272                 return ret;
273         rq = scsi_req(req);
274
275         if (bufflen &&  blk_rq_map_kern(sdev->request_queue, req,
276                                         buffer, bufflen, GFP_NOIO))
277                 goto out;
278
279         rq->cmd_len = COMMAND_SIZE(cmd[0]);
280         memcpy(rq->cmd, cmd, rq->cmd_len);
281         rq->retries = retries;
282         req->timeout = timeout;
283         req->cmd_flags |= flags;
284         req->rq_flags |= rq_flags | RQF_QUIET;
285
286         /*
287          * head injection *required* here otherwise quiesce won't work
288          */
289         blk_execute_rq(req->q, NULL, req, 1);
290
291         /*
292          * Some devices (USB mass-storage in particular) may transfer
293          * garbage data together with a residue indicating that the data
294          * is invalid.  Prevent the garbage from being misinterpreted
295          * and prevent security leaks by zeroing out the excess data.
296          */
297         if (unlikely(rq->resid_len > 0 && rq->resid_len <= bufflen))
298                 memset(buffer + (bufflen - rq->resid_len), 0, rq->resid_len);
299
300         if (resid)
301                 *resid = rq->resid_len;
302         if (sense && rq->sense_len)
303                 memcpy(sense, rq->sense, SCSI_SENSE_BUFFERSIZE);
304         if (sshdr)
305                 scsi_normalize_sense(rq->sense, rq->sense_len, sshdr);
306         ret = rq->result;
307  out:
308         blk_put_request(req);
309
310         return ret;
311 }
312 EXPORT_SYMBOL(__scsi_execute);
313
314 /*
315  * Function:    scsi_init_cmd_errh()
316  *
317  * Purpose:     Initialize cmd fields related to error handling.
318  *
319  * Arguments:   cmd     - command that is ready to be queued.
320  *
321  * Notes:       This function has the job of initializing a number of
322  *              fields related to error handling.   Typically this will
323  *              be called once for each command, as required.
324  */
325 static void scsi_init_cmd_errh(struct scsi_cmnd *cmd)
326 {
327         cmd->serial_number = 0;
328         scsi_set_resid(cmd, 0);
329         memset(cmd->sense_buffer, 0, SCSI_SENSE_BUFFERSIZE);
330         if (cmd->cmd_len == 0)
331                 cmd->cmd_len = scsi_command_size(cmd->cmnd);
332 }
333
334 /*
335  * Decrement the host_busy counter and wake up the error handler if necessary.
336  * Avoid as follows that the error handler is not woken up if shost->host_busy
337  * == shost->host_failed: use call_rcu() in scsi_eh_scmd_add() in combination
338  * with an RCU read lock in this function to ensure that this function in its
339  * entirety either finishes before scsi_eh_scmd_add() increases the
340  * host_failed counter or that it notices the shost state change made by
341  * scsi_eh_scmd_add().
342  */
343 static void scsi_dec_host_busy(struct Scsi_Host *shost)
344 {
345         unsigned long flags;
346
347         rcu_read_lock();
348         atomic_dec(&shost->host_busy);
349         if (unlikely(scsi_host_in_recovery(shost))) {
350                 spin_lock_irqsave(shost->host_lock, flags);
351                 if (shost->host_failed || shost->host_eh_scheduled)
352                         scsi_eh_wakeup(shost);
353                 spin_unlock_irqrestore(shost->host_lock, flags);
354         }
355         rcu_read_unlock();
356 }
357
358 void scsi_device_unbusy(struct scsi_device *sdev)
359 {
360         struct Scsi_Host *shost = sdev->host;
361         struct scsi_target *starget = scsi_target(sdev);
362
363         scsi_dec_host_busy(shost);
364
365         if (starget->can_queue > 0)
366                 atomic_dec(&starget->target_busy);
367
368         atomic_dec(&sdev->device_busy);
369 }
370
371 static void scsi_kick_queue(struct request_queue *q)
372 {
373         if (q->mq_ops)
374                 blk_mq_run_hw_queues(q, false);
375         else
376                 blk_run_queue(q);
377 }
378
379 /*
380  * Called for single_lun devices on IO completion. Clear starget_sdev_user,
381  * and call blk_run_queue for all the scsi_devices on the target -
382  * including current_sdev first.
383  *
384  * Called with *no* scsi locks held.
385  */
386 static void scsi_single_lun_run(struct scsi_device *current_sdev)
387 {
388         struct Scsi_Host *shost = current_sdev->host;
389         struct scsi_device *sdev, *tmp;
390         struct scsi_target *starget = scsi_target(current_sdev);
391         unsigned long flags;
392
393         spin_lock_irqsave(shost->host_lock, flags);
394         starget->starget_sdev_user = NULL;
395         spin_unlock_irqrestore(shost->host_lock, flags);
396
397         /*
398          * Call blk_run_queue for all LUNs on the target, starting with
399          * current_sdev. We race with others (to set starget_sdev_user),
400          * but in most cases, we will be first. Ideally, each LU on the
401          * target would get some limited time or requests on the target.
402          */
403         scsi_kick_queue(current_sdev->request_queue);
404
405         spin_lock_irqsave(shost->host_lock, flags);
406         if (starget->starget_sdev_user)
407                 goto out;
408         list_for_each_entry_safe(sdev, tmp, &starget->devices,
409                         same_target_siblings) {
410                 if (sdev == current_sdev)
411                         continue;
412                 if (scsi_device_get(sdev))
413                         continue;
414
415                 spin_unlock_irqrestore(shost->host_lock, flags);
416                 scsi_kick_queue(sdev->request_queue);
417                 spin_lock_irqsave(shost->host_lock, flags);
418         
419                 scsi_device_put(sdev);
420         }
421  out:
422         spin_unlock_irqrestore(shost->host_lock, flags);
423 }
424
425 static inline bool scsi_device_is_busy(struct scsi_device *sdev)
426 {
427         if (atomic_read(&sdev->device_busy) >= sdev->queue_depth)
428                 return true;
429         if (atomic_read(&sdev->device_blocked) > 0)
430                 return true;
431         return false;
432 }
433
434 static inline bool scsi_target_is_busy(struct scsi_target *starget)
435 {
436         if (starget->can_queue > 0) {
437                 if (atomic_read(&starget->target_busy) >= starget->can_queue)
438                         return true;
439                 if (atomic_read(&starget->target_blocked) > 0)
440                         return true;
441         }
442         return false;
443 }
444
445 static inline bool scsi_host_is_busy(struct Scsi_Host *shost)
446 {
447         if (shost->can_queue > 0 &&
448             atomic_read(&shost->host_busy) >= shost->can_queue)
449                 return true;
450         if (atomic_read(&shost->host_blocked) > 0)
451                 return true;
452         if (shost->host_self_blocked)
453                 return true;
454         return false;
455 }
456
457 static void scsi_starved_list_run(struct Scsi_Host *shost)
458 {
459         LIST_HEAD(starved_list);
460         struct scsi_device *sdev;
461         unsigned long flags;
462
463         spin_lock_irqsave(shost->host_lock, flags);
464         list_splice_init(&shost->starved_list, &starved_list);
465
466         while (!list_empty(&starved_list)) {
467                 struct request_queue *slq;
468
469                 /*
470                  * As long as shost is accepting commands and we have
471                  * starved queues, call blk_run_queue. scsi_request_fn
472                  * drops the queue_lock and can add us back to the
473                  * starved_list.
474                  *
475                  * host_lock protects the starved_list and starved_entry.
476                  * scsi_request_fn must get the host_lock before checking
477                  * or modifying starved_list or starved_entry.
478                  */
479                 if (scsi_host_is_busy(shost))
480                         break;
481
482                 sdev = list_entry(starved_list.next,
483                                   struct scsi_device, starved_entry);
484                 list_del_init(&sdev->starved_entry);
485                 if (scsi_target_is_busy(scsi_target(sdev))) {
486                         list_move_tail(&sdev->starved_entry,
487                                        &shost->starved_list);
488                         continue;
489                 }
490
491                 /*
492                  * Once we drop the host lock, a racing scsi_remove_device()
493                  * call may remove the sdev from the starved list and destroy
494                  * it and the queue.  Mitigate by taking a reference to the
495                  * queue and never touching the sdev again after we drop the
496                  * host lock.  Note: if __scsi_remove_device() invokes
497                  * blk_cleanup_queue() before the queue is run from this
498                  * function then blk_run_queue() will return immediately since
499                  * blk_cleanup_queue() marks the queue with QUEUE_FLAG_DYING.
500                  */
501                 slq = sdev->request_queue;
502                 if (!blk_get_queue(slq))
503                         continue;
504                 spin_unlock_irqrestore(shost->host_lock, flags);
505
506                 scsi_kick_queue(slq);
507                 blk_put_queue(slq);
508
509                 spin_lock_irqsave(shost->host_lock, flags);
510         }
511         /* put any unprocessed entries back */
512         list_splice(&starved_list, &shost->starved_list);
513         spin_unlock_irqrestore(shost->host_lock, flags);
514 }
515
516 /*
517  * Function:   scsi_run_queue()
518  *
519  * Purpose:    Select a proper request queue to serve next
520  *
521  * Arguments:  q       - last request's queue
522  *
523  * Returns:     Nothing
524  *
525  * Notes:      The previous command was completely finished, start
526  *             a new one if possible.
527  */
528 static void scsi_run_queue(struct request_queue *q)
529 {
530         struct scsi_device *sdev = q->queuedata;
531
532         if (scsi_target(sdev)->single_lun)
533                 scsi_single_lun_run(sdev);
534         if (!list_empty(&sdev->host->starved_list))
535                 scsi_starved_list_run(sdev->host);
536
537         if (q->mq_ops)
538                 blk_mq_run_hw_queues(q, false);
539         else
540                 blk_run_queue(q);
541 }
542
543 void scsi_requeue_run_queue(struct work_struct *work)
544 {
545         struct scsi_device *sdev;
546         struct request_queue *q;
547
548         sdev = container_of(work, struct scsi_device, requeue_work);
549         q = sdev->request_queue;
550         scsi_run_queue(q);
551 }
552
553 /*
554  * Function:    scsi_requeue_command()
555  *
556  * Purpose:     Handle post-processing of completed commands.
557  *
558  * Arguments:   q       - queue to operate on
559  *              cmd     - command that may need to be requeued.
560  *
561  * Returns:     Nothing
562  *
563  * Notes:       After command completion, there may be blocks left
564  *              over which weren't finished by the previous command
565  *              this can be for a number of reasons - the main one is
566  *              I/O errors in the middle of the request, in which case
567  *              we need to request the blocks that come after the bad
568  *              sector.
569  * Notes:       Upon return, cmd is a stale pointer.
570  */
571 static void scsi_requeue_command(struct request_queue *q, struct scsi_cmnd *cmd)
572 {
573         struct scsi_device *sdev = cmd->device;
574         struct request *req = cmd->request;
575         unsigned long flags;
576
577         spin_lock_irqsave(q->queue_lock, flags);
578         blk_unprep_request(req);
579         req->special = NULL;
580         scsi_put_command(cmd);
581         blk_requeue_request(q, req);
582         spin_unlock_irqrestore(q->queue_lock, flags);
583
584         scsi_run_queue(q);
585
586         put_device(&sdev->sdev_gendev);
587 }
588
589 void scsi_run_host_queues(struct Scsi_Host *shost)
590 {
591         struct scsi_device *sdev;
592
593         shost_for_each_device(sdev, shost)
594                 scsi_run_queue(sdev->request_queue);
595 }
596
597 static void scsi_uninit_cmd(struct scsi_cmnd *cmd)
598 {
599         if (!blk_rq_is_passthrough(cmd->request)) {
600                 struct scsi_driver *drv = scsi_cmd_to_driver(cmd);
601
602                 if (drv->uninit_command)
603                         drv->uninit_command(cmd);
604         }
605 }
606
607 static void scsi_mq_free_sgtables(struct scsi_cmnd *cmd)
608 {
609         struct scsi_data_buffer *sdb;
610
611         if (cmd->sdb.table.nents)
612                 sg_free_table_chained(&cmd->sdb.table, true);
613         if (cmd->request->next_rq) {
614                 sdb = cmd->request->next_rq->special;
615                 if (sdb)
616                         sg_free_table_chained(&sdb->table, true);
617         }
618         if (scsi_prot_sg_count(cmd))
619                 sg_free_table_chained(&cmd->prot_sdb->table, true);
620 }
621
622 static void scsi_mq_uninit_cmd(struct scsi_cmnd *cmd)
623 {
624         scsi_mq_free_sgtables(cmd);
625         scsi_uninit_cmd(cmd);
626         scsi_del_cmd_from_list(cmd);
627 }
628
629 /*
630  * Function:    scsi_release_buffers()
631  *
632  * Purpose:     Free resources allocate for a scsi_command.
633  *
634  * Arguments:   cmd     - command that we are bailing.
635  *
636  * Lock status: Assumed that no lock is held upon entry.
637  *
638  * Returns:     Nothing
639  *
640  * Notes:       In the event that an upper level driver rejects a
641  *              command, we must release resources allocated during
642  *              the __init_io() function.  Primarily this would involve
643  *              the scatter-gather table.
644  */
645 static void scsi_release_buffers(struct scsi_cmnd *cmd)
646 {
647         if (cmd->sdb.table.nents)
648                 sg_free_table_chained(&cmd->sdb.table, false);
649
650         memset(&cmd->sdb, 0, sizeof(cmd->sdb));
651
652         if (scsi_prot_sg_count(cmd))
653                 sg_free_table_chained(&cmd->prot_sdb->table, false);
654 }
655
656 static void scsi_release_bidi_buffers(struct scsi_cmnd *cmd)
657 {
658         struct scsi_data_buffer *bidi_sdb = cmd->request->next_rq->special;
659
660         sg_free_table_chained(&bidi_sdb->table, false);
661         kmem_cache_free(scsi_sdb_cache, bidi_sdb);
662         cmd->request->next_rq->special = NULL;
663 }
664
665 /* Returns false when no more bytes to process, true if there are more */
666 static bool scsi_end_request(struct request *req, blk_status_t error,
667                 unsigned int bytes, unsigned int bidi_bytes)
668 {
669         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
670         struct scsi_device *sdev = cmd->device;
671         struct request_queue *q = sdev->request_queue;
672
673         if (blk_update_request(req, error, bytes))
674                 return true;
675
676         /* Bidi request must be completed as a whole */
677         if (unlikely(bidi_bytes) &&
678             blk_update_request(req->next_rq, error, bidi_bytes))
679                 return true;
680
681         if (blk_queue_add_random(q))
682                 add_disk_randomness(req->rq_disk);
683
684         if (!blk_rq_is_scsi(req)) {
685                 WARN_ON_ONCE(!(cmd->flags & SCMD_INITIALIZED));
686                 cmd->flags &= ~SCMD_INITIALIZED;
687                 destroy_rcu_head(&cmd->rcu);
688         }
689
690         if (req->mq_ctx) {
691                 /*
692                  * In the MQ case the command gets freed by __blk_mq_end_request,
693                  * so we have to do all cleanup that depends on it earlier.
694                  *
695                  * We also can't kick the queues from irq context, so we
696                  * will have to defer it to a workqueue.
697                  */
698                 scsi_mq_uninit_cmd(cmd);
699
700                 /*
701                  * queue is still alive, so grab the ref for preventing it
702                  * from being cleaned up during running queue.
703                  */
704                 percpu_ref_get(&q->q_usage_counter);
705
706                 __blk_mq_end_request(req, error);
707
708                 if (scsi_target(sdev)->single_lun ||
709                     !list_empty(&sdev->host->starved_list))
710                         kblockd_schedule_work(&sdev->requeue_work);
711                 else
712                         blk_mq_run_hw_queues(q, true);
713
714                 percpu_ref_put(&q->q_usage_counter);
715         } else {
716                 unsigned long flags;
717
718                 if (bidi_bytes)
719                         scsi_release_bidi_buffers(cmd);
720                 scsi_release_buffers(cmd);
721                 scsi_put_command(cmd);
722
723                 spin_lock_irqsave(q->queue_lock, flags);
724                 blk_finish_request(req, error);
725                 spin_unlock_irqrestore(q->queue_lock, flags);
726
727                 scsi_run_queue(q);
728         }
729
730         put_device(&sdev->sdev_gendev);
731         return false;
732 }
733
734 /**
735  * scsi_result_to_blk_status - translate a SCSI result code into blk_status_t
736  * @cmd:        SCSI command
737  * @result:     scsi error code
738  *
739  * Translate a SCSI result code into a blk_status_t value. May reset the host
740  * byte of @cmd->result.
741  */
742 static blk_status_t scsi_result_to_blk_status(struct scsi_cmnd *cmd, int result)
743 {
744         switch (host_byte(result)) {
745         case DID_OK:
746                 /*
747                  * Also check the other bytes than the status byte in result
748                  * to handle the case when a SCSI LLD sets result to
749                  * DRIVER_SENSE << 24 without setting SAM_STAT_CHECK_CONDITION.
750                  */
751                 if (scsi_status_is_good(result) && (result & ~0xff) == 0)
752                         return BLK_STS_OK;
753                 return BLK_STS_IOERR;
754         case DID_TRANSPORT_FAILFAST:
755                 return BLK_STS_TRANSPORT;
756         case DID_TARGET_FAILURE:
757                 set_host_byte(cmd, DID_OK);
758                 return BLK_STS_TARGET;
759         case DID_NEXUS_FAILURE:
760                 return BLK_STS_NEXUS;
761         case DID_ALLOC_FAILURE:
762                 set_host_byte(cmd, DID_OK);
763                 return BLK_STS_NOSPC;
764         case DID_MEDIUM_ERROR:
765                 set_host_byte(cmd, DID_OK);
766                 return BLK_STS_MEDIUM;
767         default:
768                 return BLK_STS_IOERR;
769         }
770 }
771
772 /* Helper for scsi_io_completion() when "reprep" action required. */
773 static void scsi_io_completion_reprep(struct scsi_cmnd *cmd,
774                                       struct request_queue *q)
775 {
776         /* A new command will be prepared and issued. */
777         if (q->mq_ops) {
778                 scsi_mq_requeue_cmd(cmd);
779         } else {
780                 /* Unprep request and put it back at head of the queue. */
781                 scsi_release_buffers(cmd);
782                 scsi_requeue_command(q, cmd);
783         }
784 }
785
786 /* Helper for scsi_io_completion() when special action required. */
787 static void scsi_io_completion_action(struct scsi_cmnd *cmd, int result)
788 {
789         struct request_queue *q = cmd->device->request_queue;
790         struct request *req = cmd->request;
791         int level = 0;
792         enum {ACTION_FAIL, ACTION_REPREP, ACTION_RETRY,
793               ACTION_DELAYED_RETRY} action;
794         unsigned long wait_for = (cmd->allowed + 1) * req->timeout;
795         struct scsi_sense_hdr sshdr;
796         bool sense_valid;
797         bool sense_current = true;      /* false implies "deferred sense" */
798         blk_status_t blk_stat;
799
800         sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
801         if (sense_valid)
802                 sense_current = !scsi_sense_is_deferred(&sshdr);
803
804         blk_stat = scsi_result_to_blk_status(cmd, result);
805
806         if (host_byte(result) == DID_RESET) {
807                 /* Third party bus reset or reset for error recovery
808                  * reasons.  Just retry the command and see what
809                  * happens.
810                  */
811                 action = ACTION_RETRY;
812         } else if (sense_valid && sense_current) {
813                 switch (sshdr.sense_key) {
814                 case UNIT_ATTENTION:
815                         if (cmd->device->removable) {
816                                 /* Detected disc change.  Set a bit
817                                  * and quietly refuse further access.
818                                  */
819                                 cmd->device->changed = 1;
820                                 action = ACTION_FAIL;
821                         } else {
822                                 /* Must have been a power glitch, or a
823                                  * bus reset.  Could not have been a
824                                  * media change, so we just retry the
825                                  * command and see what happens.
826                                  */
827                                 action = ACTION_RETRY;
828                         }
829                         break;
830                 case ILLEGAL_REQUEST:
831                         /* If we had an ILLEGAL REQUEST returned, then
832                          * we may have performed an unsupported
833                          * command.  The only thing this should be
834                          * would be a ten byte read where only a six
835                          * byte read was supported.  Also, on a system
836                          * where READ CAPACITY failed, we may have
837                          * read past the end of the disk.
838                          */
839                         if ((cmd->device->use_10_for_rw &&
840                             sshdr.asc == 0x20 && sshdr.ascq == 0x00) &&
841                             (cmd->cmnd[0] == READ_10 ||
842                              cmd->cmnd[0] == WRITE_10)) {
843                                 /* This will issue a new 6-byte command. */
844                                 cmd->device->use_10_for_rw = 0;
845                                 action = ACTION_REPREP;
846                         } else if (sshdr.asc == 0x10) /* DIX */ {
847                                 action = ACTION_FAIL;
848                                 blk_stat = BLK_STS_PROTECTION;
849                         /* INVALID COMMAND OPCODE or INVALID FIELD IN CDB */
850                         } else if (sshdr.asc == 0x20 || sshdr.asc == 0x24) {
851                                 action = ACTION_FAIL;
852                                 blk_stat = BLK_STS_TARGET;
853                         } else
854                                 action = ACTION_FAIL;
855                         break;
856                 case ABORTED_COMMAND:
857                         action = ACTION_FAIL;
858                         if (sshdr.asc == 0x10) /* DIF */
859                                 blk_stat = BLK_STS_PROTECTION;
860                         break;
861                 case NOT_READY:
862                         /* If the device is in the process of becoming
863                          * ready, or has a temporary blockage, retry.
864                          */
865                         if (sshdr.asc == 0x04) {
866                                 switch (sshdr.ascq) {
867                                 case 0x01: /* becoming ready */
868                                 case 0x04: /* format in progress */
869                                 case 0x05: /* rebuild in progress */
870                                 case 0x06: /* recalculation in progress */
871                                 case 0x07: /* operation in progress */
872                                 case 0x08: /* Long write in progress */
873                                 case 0x09: /* self test in progress */
874                                 case 0x14: /* space allocation in progress */
875                                 case 0x1a: /* start stop unit in progress */
876                                 case 0x1b: /* sanitize in progress */
877                                 case 0x1d: /* configuration in progress */
878                                 case 0x24: /* depopulation in progress */
879                                         action = ACTION_DELAYED_RETRY;
880                                         break;
881                                 default:
882                                         action = ACTION_FAIL;
883                                         break;
884                                 }
885                         } else
886                                 action = ACTION_FAIL;
887                         break;
888                 case VOLUME_OVERFLOW:
889                         /* See SSC3rXX or current. */
890                         action = ACTION_FAIL;
891                         break;
892                 default:
893                         action = ACTION_FAIL;
894                         break;
895                 }
896         } else
897                 action = ACTION_FAIL;
898
899         if (action != ACTION_FAIL &&
900             time_before(cmd->jiffies_at_alloc + wait_for, jiffies))
901                 action = ACTION_FAIL;
902
903         switch (action) {
904         case ACTION_FAIL:
905                 /* Give up and fail the remainder of the request */
906                 if (!(req->rq_flags & RQF_QUIET)) {
907                         static DEFINE_RATELIMIT_STATE(_rs,
908                                         DEFAULT_RATELIMIT_INTERVAL,
909                                         DEFAULT_RATELIMIT_BURST);
910
911                         if (unlikely(scsi_logging_level))
912                                 level =
913                                      SCSI_LOG_LEVEL(SCSI_LOG_MLCOMPLETE_SHIFT,
914                                                     SCSI_LOG_MLCOMPLETE_BITS);
915
916                         /*
917                          * if logging is enabled the failure will be printed
918                          * in scsi_log_completion(), so avoid duplicate messages
919                          */
920                         if (!level && __ratelimit(&_rs)) {
921                                 scsi_print_result(cmd, NULL, FAILED);
922                                 if (driver_byte(result) == DRIVER_SENSE)
923                                         scsi_print_sense(cmd);
924                                 scsi_print_command(cmd);
925                         }
926                 }
927                 if (!scsi_end_request(req, blk_stat, blk_rq_err_bytes(req), 0))
928                         return;
929                 /*FALLTHRU*/
930         case ACTION_REPREP:
931                 scsi_io_completion_reprep(cmd, q);
932                 break;
933         case ACTION_RETRY:
934                 /* Retry the same command immediately */
935                 __scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY, false);
936                 break;
937         case ACTION_DELAYED_RETRY:
938                 /* Retry the same command after a delay */
939                 __scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY, false);
940                 break;
941         }
942 }
943
944 /*
945  * Helper for scsi_io_completion() when cmd->result is non-zero. Returns a
946  * new result that may suppress further error checking. Also modifies
947  * *blk_statp in some cases.
948  */
949 static int scsi_io_completion_nz_result(struct scsi_cmnd *cmd, int result,
950                                         blk_status_t *blk_statp)
951 {
952         bool sense_valid;
953         bool sense_current = true;      /* false implies "deferred sense" */
954         struct request *req = cmd->request;
955         struct scsi_sense_hdr sshdr;
956
957         sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
958         if (sense_valid)
959                 sense_current = !scsi_sense_is_deferred(&sshdr);
960
961         if (blk_rq_is_passthrough(req)) {
962                 if (sense_valid) {
963                         /*
964                          * SG_IO wants current and deferred errors
965                          */
966                         scsi_req(req)->sense_len =
967                                 min(8 + cmd->sense_buffer[7],
968                                     SCSI_SENSE_BUFFERSIZE);
969                 }
970                 if (sense_current)
971                         *blk_statp = scsi_result_to_blk_status(cmd, result);
972         } else if (blk_rq_bytes(req) == 0 && sense_current) {
973                 /*
974                  * Flush commands do not transfers any data, and thus cannot use
975                  * good_bytes != blk_rq_bytes(req) as the signal for an error.
976                  * This sets *blk_statp explicitly for the problem case.
977                  */
978                 *blk_statp = scsi_result_to_blk_status(cmd, result);
979         }
980         /*
981          * Recovered errors need reporting, but they're always treated as
982          * success, so fiddle the result code here.  For passthrough requests
983          * we already took a copy of the original into sreq->result which
984          * is what gets returned to the user
985          */
986         if (sense_valid && (sshdr.sense_key == RECOVERED_ERROR)) {
987                 bool do_print = true;
988                 /*
989                  * if ATA PASS-THROUGH INFORMATION AVAILABLE [0x0, 0x1d]
990                  * skip print since caller wants ATA registers. Only occurs
991                  * on SCSI ATA PASS_THROUGH commands when CK_COND=1
992                  */
993                 if ((sshdr.asc == 0x0) && (sshdr.ascq == 0x1d))
994                         do_print = false;
995                 else if (req->rq_flags & RQF_QUIET)
996                         do_print = false;
997                 if (do_print)
998                         scsi_print_sense(cmd);
999                 result = 0;
1000                 /* for passthrough, *blk_statp may be set */
1001                 *blk_statp = BLK_STS_OK;
1002         }
1003         /*
1004          * Another corner case: the SCSI status byte is non-zero but 'good'.
1005          * Example: PRE-FETCH command returns SAM_STAT_CONDITION_MET when
1006          * it is able to fit nominated LBs in its cache (and SAM_STAT_GOOD
1007          * if it can't fit). Treat SAM_STAT_CONDITION_MET and the related
1008          * intermediate statuses (both obsolete in SAM-4) as good.
1009          */
1010         if (status_byte(result) && scsi_status_is_good(result)) {
1011                 result = 0;
1012                 *blk_statp = BLK_STS_OK;
1013         }
1014         return result;
1015 }
1016
1017 /*
1018  * Function:    scsi_io_completion()
1019  *
1020  * Purpose:     Completion processing for block device I/O requests.
1021  *
1022  * Arguments:   cmd   - command that is finished.
1023  *
1024  * Lock status: Assumed that no lock is held upon entry.
1025  *
1026  * Returns:     Nothing
1027  *
1028  * Notes:       We will finish off the specified number of sectors.  If we
1029  *              are done, the command block will be released and the queue
1030  *              function will be goosed.  If we are not done then we have to
1031  *              figure out what to do next:
1032  *
1033  *              a) We can call scsi_requeue_command().  The request
1034  *                 will be unprepared and put back on the queue.  Then
1035  *                 a new command will be created for it.  This should
1036  *                 be used if we made forward progress, or if we want
1037  *                 to switch from READ(10) to READ(6) for example.
1038  *
1039  *              b) We can call __scsi_queue_insert().  The request will
1040  *                 be put back on the queue and retried using the same
1041  *                 command as before, possibly after a delay.
1042  *
1043  *              c) We can call scsi_end_request() with blk_stat other than
1044  *                 BLK_STS_OK, to fail the remainder of the request.
1045  */
1046 void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes)
1047 {
1048         int result = cmd->result;
1049         struct request_queue *q = cmd->device->request_queue;
1050         struct request *req = cmd->request;
1051         blk_status_t blk_stat = BLK_STS_OK;
1052
1053         if (unlikely(result))   /* a nz result may or may not be an error */
1054                 result = scsi_io_completion_nz_result(cmd, result, &blk_stat);
1055
1056         if (unlikely(blk_rq_is_passthrough(req))) {
1057                 /*
1058                  * scsi_result_to_blk_status may have reset the host_byte
1059                  */
1060                 scsi_req(req)->result = cmd->result;
1061                 scsi_req(req)->resid_len = scsi_get_resid(cmd);
1062
1063                 if (unlikely(scsi_bidi_cmnd(cmd))) {
1064                         /*
1065                          * Bidi commands Must be complete as a whole,
1066                          * both sides at once.
1067                          */
1068                         scsi_req(req->next_rq)->resid_len = scsi_in(cmd)->resid;
1069                         if (scsi_end_request(req, BLK_STS_OK, blk_rq_bytes(req),
1070                                         blk_rq_bytes(req->next_rq)))
1071                                 WARN_ONCE(true,
1072                                           "Bidi command with remaining bytes");
1073                         return;
1074                 }
1075         }
1076
1077         /* no bidi support yet, other than in pass-through */
1078         if (unlikely(blk_bidi_rq(req))) {
1079                 WARN_ONCE(true, "Only support bidi command in passthrough");
1080                 scmd_printk(KERN_ERR, cmd, "Killing bidi command\n");
1081                 if (scsi_end_request(req, BLK_STS_IOERR, blk_rq_bytes(req),
1082                                      blk_rq_bytes(req->next_rq)))
1083                         WARN_ONCE(true, "Bidi command with remaining bytes");
1084                 return;
1085         }
1086
1087         /*
1088          * Next deal with any sectors which we were able to correctly
1089          * handle.
1090          */
1091         SCSI_LOG_HLCOMPLETE(1, scmd_printk(KERN_INFO, cmd,
1092                 "%u sectors total, %d bytes done.\n",
1093                 blk_rq_sectors(req), good_bytes));
1094
1095         /*
1096          * Next deal with any sectors which we were able to correctly
1097          * handle. Failed, zero length commands always need to drop down
1098          * to retry code. Fast path should return in this block.
1099          */
1100         if (likely(blk_rq_bytes(req) > 0 || blk_stat == BLK_STS_OK)) {
1101                 if (likely(!scsi_end_request(req, blk_stat, good_bytes, 0)))
1102                         return; /* no bytes remaining */
1103         }
1104
1105         /* Kill remainder if no retries. */
1106         if (unlikely(blk_stat && scsi_noretry_cmd(cmd))) {
1107                 if (scsi_end_request(req, blk_stat, blk_rq_bytes(req), 0))
1108                         WARN_ONCE(true,
1109                             "Bytes remaining after failed, no-retry command");
1110                 return;
1111         }
1112
1113         /*
1114          * If there had been no error, but we have leftover bytes in the
1115          * requeues just queue the command up again.
1116          */
1117         if (likely(result == 0))
1118                 scsi_io_completion_reprep(cmd, q);
1119         else
1120                 scsi_io_completion_action(cmd, result);
1121 }
1122
1123 static int scsi_init_sgtable(struct request *req, struct scsi_data_buffer *sdb)
1124 {
1125         int count;
1126
1127         /*
1128          * If sg table allocation fails, requeue request later.
1129          */
1130         if (unlikely(sg_alloc_table_chained(&sdb->table,
1131                         blk_rq_nr_phys_segments(req), sdb->table.sgl)))
1132                 return BLKPREP_DEFER;
1133
1134         /* 
1135          * Next, walk the list, and fill in the addresses and sizes of
1136          * each segment.
1137          */
1138         count = blk_rq_map_sg(req->q, req, sdb->table.sgl);
1139         BUG_ON(count > sdb->table.nents);
1140         sdb->table.nents = count;
1141         sdb->length = blk_rq_payload_bytes(req);
1142         return BLKPREP_OK;
1143 }
1144
1145 /*
1146  * Function:    scsi_init_io()
1147  *
1148  * Purpose:     SCSI I/O initialize function.
1149  *
1150  * Arguments:   cmd   - Command descriptor we wish to initialize
1151  *
1152  * Returns:     0 on success
1153  *              BLKPREP_DEFER if the failure is retryable
1154  *              BLKPREP_KILL if the failure is fatal
1155  */
1156 int scsi_init_io(struct scsi_cmnd *cmd)
1157 {
1158         struct scsi_device *sdev = cmd->device;
1159         struct request *rq = cmd->request;
1160         bool is_mq = (rq->mq_ctx != NULL);
1161         int error = BLKPREP_KILL;
1162
1163         if (WARN_ON_ONCE(!blk_rq_nr_phys_segments(rq)))
1164                 goto err_exit;
1165
1166         error = scsi_init_sgtable(rq, &cmd->sdb);
1167         if (error)
1168                 goto err_exit;
1169
1170         if (blk_bidi_rq(rq)) {
1171                 if (!rq->q->mq_ops) {
1172                         struct scsi_data_buffer *bidi_sdb =
1173                                 kmem_cache_zalloc(scsi_sdb_cache, GFP_ATOMIC);
1174                         if (!bidi_sdb) {
1175                                 error = BLKPREP_DEFER;
1176                                 goto err_exit;
1177                         }
1178
1179                         rq->next_rq->special = bidi_sdb;
1180                 }
1181
1182                 error = scsi_init_sgtable(rq->next_rq, rq->next_rq->special);
1183                 if (error)
1184                         goto err_exit;
1185         }
1186
1187         if (blk_integrity_rq(rq)) {
1188                 struct scsi_data_buffer *prot_sdb = cmd->prot_sdb;
1189                 int ivecs, count;
1190
1191                 if (prot_sdb == NULL) {
1192                         /*
1193                          * This can happen if someone (e.g. multipath)
1194                          * queues a command to a device on an adapter
1195                          * that does not support DIX.
1196                          */
1197                         WARN_ON_ONCE(1);
1198                         error = BLKPREP_KILL;
1199                         goto err_exit;
1200                 }
1201
1202                 ivecs = blk_rq_count_integrity_sg(rq->q, rq->bio);
1203
1204                 if (sg_alloc_table_chained(&prot_sdb->table, ivecs,
1205                                 prot_sdb->table.sgl)) {
1206                         error = BLKPREP_DEFER;
1207                         goto err_exit;
1208                 }
1209
1210                 count = blk_rq_map_integrity_sg(rq->q, rq->bio,
1211                                                 prot_sdb->table.sgl);
1212                 BUG_ON(count > ivecs);
1213                 BUG_ON(count > queue_max_integrity_segments(rq->q));
1214
1215                 cmd->prot_sdb = prot_sdb;
1216                 cmd->prot_sdb->table.nents = count;
1217         }
1218
1219         return BLKPREP_OK;
1220 err_exit:
1221         if (is_mq) {
1222                 scsi_mq_free_sgtables(cmd);
1223         } else {
1224                 scsi_release_buffers(cmd);
1225                 cmd->request->special = NULL;
1226                 scsi_put_command(cmd);
1227                 put_device(&sdev->sdev_gendev);
1228         }
1229         return error;
1230 }
1231 EXPORT_SYMBOL(scsi_init_io);
1232
1233 /**
1234  * scsi_initialize_rq - initialize struct scsi_cmnd partially
1235  * @rq: Request associated with the SCSI command to be initialized.
1236  *
1237  * This function initializes the members of struct scsi_cmnd that must be
1238  * initialized before request processing starts and that won't be
1239  * reinitialized if a SCSI command is requeued.
1240  *
1241  * Called from inside blk_get_request() for pass-through requests and from
1242  * inside scsi_init_command() for filesystem requests.
1243  */
1244 static void scsi_initialize_rq(struct request *rq)
1245 {
1246         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1247
1248         scsi_req_init(&cmd->req);
1249         init_rcu_head(&cmd->rcu);
1250         cmd->jiffies_at_alloc = jiffies;
1251         cmd->retries = 0;
1252 }
1253
1254 /* Add a command to the list used by the aacraid and dpt_i2o drivers */
1255 void scsi_add_cmd_to_list(struct scsi_cmnd *cmd)
1256 {
1257         struct scsi_device *sdev = cmd->device;
1258         struct Scsi_Host *shost = sdev->host;
1259         unsigned long flags;
1260
1261         if (shost->use_cmd_list) {
1262                 spin_lock_irqsave(&sdev->list_lock, flags);
1263                 list_add_tail(&cmd->list, &sdev->cmd_list);
1264                 spin_unlock_irqrestore(&sdev->list_lock, flags);
1265         }
1266 }
1267
1268 /* Remove a command from the list used by the aacraid and dpt_i2o drivers */
1269 void scsi_del_cmd_from_list(struct scsi_cmnd *cmd)
1270 {
1271         struct scsi_device *sdev = cmd->device;
1272         struct Scsi_Host *shost = sdev->host;
1273         unsigned long flags;
1274
1275         if (shost->use_cmd_list) {
1276                 spin_lock_irqsave(&sdev->list_lock, flags);
1277                 BUG_ON(list_empty(&cmd->list));
1278                 list_del_init(&cmd->list);
1279                 spin_unlock_irqrestore(&sdev->list_lock, flags);
1280         }
1281 }
1282
1283 /* Called after a request has been started. */
1284 void scsi_init_command(struct scsi_device *dev, struct scsi_cmnd *cmd)
1285 {
1286         void *buf = cmd->sense_buffer;
1287         void *prot = cmd->prot_sdb;
1288         struct request *rq = blk_mq_rq_from_pdu(cmd);
1289         unsigned int flags = cmd->flags & SCMD_PRESERVED_FLAGS;
1290         unsigned long jiffies_at_alloc;
1291         int retries;
1292
1293         if (!blk_rq_is_scsi(rq) && !(flags & SCMD_INITIALIZED)) {
1294                 flags |= SCMD_INITIALIZED;
1295                 scsi_initialize_rq(rq);
1296         }
1297
1298         jiffies_at_alloc = cmd->jiffies_at_alloc;
1299         retries = cmd->retries;
1300         /* zero out the cmd, except for the embedded scsi_request */
1301         memset((char *)cmd + sizeof(cmd->req), 0,
1302                 sizeof(*cmd) - sizeof(cmd->req) + dev->host->hostt->cmd_size);
1303
1304         cmd->device = dev;
1305         cmd->sense_buffer = buf;
1306         cmd->prot_sdb = prot;
1307         cmd->flags = flags;
1308         INIT_DELAYED_WORK(&cmd->abort_work, scmd_eh_abort_handler);
1309         cmd->jiffies_at_alloc = jiffies_at_alloc;
1310         cmd->retries = retries;
1311
1312         scsi_add_cmd_to_list(cmd);
1313 }
1314
1315 static int scsi_setup_scsi_cmnd(struct scsi_device *sdev, struct request *req)
1316 {
1317         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1318
1319         /*
1320          * Passthrough requests may transfer data, in which case they must
1321          * a bio attached to them.  Or they might contain a SCSI command
1322          * that does not transfer data, in which case they may optionally
1323          * submit a request without an attached bio.
1324          */
1325         if (req->bio) {
1326                 int ret = scsi_init_io(cmd);
1327                 if (unlikely(ret))
1328                         return ret;
1329         } else {
1330                 BUG_ON(blk_rq_bytes(req));
1331
1332                 memset(&cmd->sdb, 0, sizeof(cmd->sdb));
1333         }
1334
1335         cmd->cmd_len = scsi_req(req)->cmd_len;
1336         cmd->cmnd = scsi_req(req)->cmd;
1337         cmd->transfersize = blk_rq_bytes(req);
1338         cmd->allowed = scsi_req(req)->retries;
1339         return BLKPREP_OK;
1340 }
1341
1342 /*
1343  * Setup a normal block command.  These are simple request from filesystems
1344  * that still need to be translated to SCSI CDBs from the ULD.
1345  */
1346 static int scsi_setup_fs_cmnd(struct scsi_device *sdev, struct request *req)
1347 {
1348         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1349
1350         if (unlikely(sdev->handler && sdev->handler->prep_fn)) {
1351                 int ret = sdev->handler->prep_fn(sdev, req);
1352                 if (ret != BLKPREP_OK)
1353                         return ret;
1354         }
1355
1356         cmd->cmnd = scsi_req(req)->cmd = scsi_req(req)->__cmd;
1357         memset(cmd->cmnd, 0, BLK_MAX_CDB);
1358         return scsi_cmd_to_driver(cmd)->init_command(cmd);
1359 }
1360
1361 static int scsi_setup_cmnd(struct scsi_device *sdev, struct request *req)
1362 {
1363         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1364
1365         if (!blk_rq_bytes(req))
1366                 cmd->sc_data_direction = DMA_NONE;
1367         else if (rq_data_dir(req) == WRITE)
1368                 cmd->sc_data_direction = DMA_TO_DEVICE;
1369         else
1370                 cmd->sc_data_direction = DMA_FROM_DEVICE;
1371
1372         if (blk_rq_is_scsi(req))
1373                 return scsi_setup_scsi_cmnd(sdev, req);
1374         else
1375                 return scsi_setup_fs_cmnd(sdev, req);
1376 }
1377
1378 static int
1379 scsi_prep_state_check(struct scsi_device *sdev, struct request *req)
1380 {
1381         int ret = BLKPREP_OK;
1382
1383         /*
1384          * If the device is not in running state we will reject some
1385          * or all commands.
1386          */
1387         if (unlikely(sdev->sdev_state != SDEV_RUNNING)) {
1388                 switch (sdev->sdev_state) {
1389                 case SDEV_OFFLINE:
1390                 case SDEV_TRANSPORT_OFFLINE:
1391                         /*
1392                          * If the device is offline we refuse to process any
1393                          * commands.  The device must be brought online
1394                          * before trying any recovery commands.
1395                          */
1396                         sdev_printk(KERN_ERR, sdev,
1397                                     "rejecting I/O to offline device\n");
1398                         ret = BLKPREP_KILL;
1399                         break;
1400                 case SDEV_DEL:
1401                         /*
1402                          * If the device is fully deleted, we refuse to
1403                          * process any commands as well.
1404                          */
1405                         sdev_printk(KERN_ERR, sdev,
1406                                     "rejecting I/O to dead device\n");
1407                         ret = BLKPREP_KILL;
1408                         break;
1409                 case SDEV_BLOCK:
1410                 case SDEV_CREATED_BLOCK:
1411                         ret = BLKPREP_DEFER;
1412                         break;
1413                 case SDEV_QUIESCE:
1414                         /*
1415                          * If the devices is blocked we defer normal commands.
1416                          */
1417                         if (req && !(req->rq_flags & RQF_PREEMPT))
1418                                 ret = BLKPREP_DEFER;
1419                         break;
1420                 default:
1421                         /*
1422                          * For any other not fully online state we only allow
1423                          * special commands.  In particular any user initiated
1424                          * command is not allowed.
1425                          */
1426                         if (req && !(req->rq_flags & RQF_PREEMPT))
1427                                 ret = BLKPREP_KILL;
1428                         break;
1429                 }
1430         }
1431         return ret;
1432 }
1433
1434 static int
1435 scsi_prep_return(struct request_queue *q, struct request *req, int ret)
1436 {
1437         struct scsi_device *sdev = q->queuedata;
1438
1439         switch (ret) {
1440         case BLKPREP_KILL:
1441         case BLKPREP_INVALID:
1442                 scsi_req(req)->result = DID_NO_CONNECT << 16;
1443                 /* release the command and kill it */
1444                 if (req->special) {
1445                         struct scsi_cmnd *cmd = req->special;
1446                         scsi_release_buffers(cmd);
1447                         scsi_put_command(cmd);
1448                         put_device(&sdev->sdev_gendev);
1449                         req->special = NULL;
1450                 }
1451                 break;
1452         case BLKPREP_DEFER:
1453                 /*
1454                  * If we defer, the blk_peek_request() returns NULL, but the
1455                  * queue must be restarted, so we schedule a callback to happen
1456                  * shortly.
1457                  */
1458                 if (atomic_read(&sdev->device_busy) == 0)
1459                         blk_delay_queue(q, SCSI_QUEUE_DELAY);
1460                 break;
1461         default:
1462                 req->rq_flags |= RQF_DONTPREP;
1463         }
1464
1465         return ret;
1466 }
1467
1468 static int scsi_prep_fn(struct request_queue *q, struct request *req)
1469 {
1470         struct scsi_device *sdev = q->queuedata;
1471         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1472         int ret;
1473
1474         ret = scsi_prep_state_check(sdev, req);
1475         if (ret != BLKPREP_OK)
1476                 goto out;
1477
1478         if (!req->special) {
1479                 /* Bail if we can't get a reference to the device */
1480                 if (unlikely(!get_device(&sdev->sdev_gendev))) {
1481                         ret = BLKPREP_DEFER;
1482                         goto out;
1483                 }
1484
1485                 scsi_init_command(sdev, cmd);
1486                 req->special = cmd;
1487         }
1488
1489         cmd->tag = req->tag;
1490         cmd->request = req;
1491         cmd->prot_op = SCSI_PROT_NORMAL;
1492
1493         ret = scsi_setup_cmnd(sdev, req);
1494 out:
1495         return scsi_prep_return(q, req, ret);
1496 }
1497
1498 static void scsi_unprep_fn(struct request_queue *q, struct request *req)
1499 {
1500         scsi_uninit_cmd(blk_mq_rq_to_pdu(req));
1501 }
1502
1503 /*
1504  * scsi_dev_queue_ready: if we can send requests to sdev, return 1 else
1505  * return 0.
1506  *
1507  * Called with the queue_lock held.
1508  */
1509 static inline int scsi_dev_queue_ready(struct request_queue *q,
1510                                   struct scsi_device *sdev)
1511 {
1512         unsigned int busy;
1513
1514         busy = atomic_inc_return(&sdev->device_busy) - 1;
1515         if (atomic_read(&sdev->device_blocked)) {
1516                 if (busy)
1517                         goto out_dec;
1518
1519                 /*
1520                  * unblock after device_blocked iterates to zero
1521                  */
1522                 if (atomic_dec_return(&sdev->device_blocked) > 0) {
1523                         /*
1524                          * For the MQ case we take care of this in the caller.
1525                          */
1526                         if (!q->mq_ops)
1527                                 blk_delay_queue(q, SCSI_QUEUE_DELAY);
1528                         goto out_dec;
1529                 }
1530                 SCSI_LOG_MLQUEUE(3, sdev_printk(KERN_INFO, sdev,
1531                                    "unblocking device at zero depth\n"));
1532         }
1533
1534         if (busy >= sdev->queue_depth)
1535                 goto out_dec;
1536
1537         return 1;
1538 out_dec:
1539         atomic_dec(&sdev->device_busy);
1540         return 0;
1541 }
1542
1543 /*
1544  * scsi_target_queue_ready: checks if there we can send commands to target
1545  * @sdev: scsi device on starget to check.
1546  */
1547 static inline int scsi_target_queue_ready(struct Scsi_Host *shost,
1548                                            struct scsi_device *sdev)
1549 {
1550         struct scsi_target *starget = scsi_target(sdev);
1551         unsigned int busy;
1552
1553         if (starget->single_lun) {
1554                 spin_lock_irq(shost->host_lock);
1555                 if (starget->starget_sdev_user &&
1556                     starget->starget_sdev_user != sdev) {
1557                         spin_unlock_irq(shost->host_lock);
1558                         return 0;
1559                 }
1560                 starget->starget_sdev_user = sdev;
1561                 spin_unlock_irq(shost->host_lock);
1562         }
1563
1564         if (starget->can_queue <= 0)
1565                 return 1;
1566
1567         busy = atomic_inc_return(&starget->target_busy) - 1;
1568         if (atomic_read(&starget->target_blocked) > 0) {
1569                 if (busy)
1570                         goto starved;
1571
1572                 /*
1573                  * unblock after target_blocked iterates to zero
1574                  */
1575                 if (atomic_dec_return(&starget->target_blocked) > 0)
1576                         goto out_dec;
1577
1578                 SCSI_LOG_MLQUEUE(3, starget_printk(KERN_INFO, starget,
1579                                  "unblocking target at zero depth\n"));
1580         }
1581
1582         if (busy >= starget->can_queue)
1583                 goto starved;
1584
1585         return 1;
1586
1587 starved:
1588         spin_lock_irq(shost->host_lock);
1589         list_move_tail(&sdev->starved_entry, &shost->starved_list);
1590         spin_unlock_irq(shost->host_lock);
1591 out_dec:
1592         if (starget->can_queue > 0)
1593                 atomic_dec(&starget->target_busy);
1594         return 0;
1595 }
1596
1597 /*
1598  * scsi_host_queue_ready: if we can send requests to shost, return 1 else
1599  * return 0. We must end up running the queue again whenever 0 is
1600  * returned, else IO can hang.
1601  */
1602 static inline int scsi_host_queue_ready(struct request_queue *q,
1603                                    struct Scsi_Host *shost,
1604                                    struct scsi_device *sdev)
1605 {
1606         unsigned int busy;
1607
1608         if (scsi_host_in_recovery(shost))
1609                 return 0;
1610
1611         busy = atomic_inc_return(&shost->host_busy) - 1;
1612         if (atomic_read(&shost->host_blocked) > 0) {
1613                 if (busy)
1614                         goto starved;
1615
1616                 /*
1617                  * unblock after host_blocked iterates to zero
1618                  */
1619                 if (atomic_dec_return(&shost->host_blocked) > 0)
1620                         goto out_dec;
1621
1622                 SCSI_LOG_MLQUEUE(3,
1623                         shost_printk(KERN_INFO, shost,
1624                                      "unblocking host at zero depth\n"));
1625         }
1626
1627         if (shost->can_queue > 0 && busy >= shost->can_queue)
1628                 goto starved;
1629         if (shost->host_self_blocked)
1630                 goto starved;
1631
1632         /* We're OK to process the command, so we can't be starved */
1633         if (!list_empty(&sdev->starved_entry)) {
1634                 spin_lock_irq(shost->host_lock);
1635                 if (!list_empty(&sdev->starved_entry))
1636                         list_del_init(&sdev->starved_entry);
1637                 spin_unlock_irq(shost->host_lock);
1638         }
1639
1640         return 1;
1641
1642 starved:
1643         spin_lock_irq(shost->host_lock);
1644         if (list_empty(&sdev->starved_entry))
1645                 list_add_tail(&sdev->starved_entry, &shost->starved_list);
1646         spin_unlock_irq(shost->host_lock);
1647 out_dec:
1648         scsi_dec_host_busy(shost);
1649         return 0;
1650 }
1651
1652 /*
1653  * Busy state exporting function for request stacking drivers.
1654  *
1655  * For efficiency, no lock is taken to check the busy state of
1656  * shost/starget/sdev, since the returned value is not guaranteed and
1657  * may be changed after request stacking drivers call the function,
1658  * regardless of taking lock or not.
1659  *
1660  * When scsi can't dispatch I/Os anymore and needs to kill I/Os scsi
1661  * needs to return 'not busy'. Otherwise, request stacking drivers
1662  * may hold requests forever.
1663  */
1664 static int scsi_lld_busy(struct request_queue *q)
1665 {
1666         struct scsi_device *sdev = q->queuedata;
1667         struct Scsi_Host *shost;
1668
1669         if (blk_queue_dying(q))
1670                 return 0;
1671
1672         shost = sdev->host;
1673
1674         /*
1675          * Ignore host/starget busy state.
1676          * Since block layer does not have a concept of fairness across
1677          * multiple queues, congestion of host/starget needs to be handled
1678          * in SCSI layer.
1679          */
1680         if (scsi_host_in_recovery(shost) || scsi_device_is_busy(sdev))
1681                 return 1;
1682
1683         return 0;
1684 }
1685
1686 /*
1687  * Kill a request for a dead device
1688  */
1689 static void scsi_kill_request(struct request *req, struct request_queue *q)
1690 {
1691         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1692         struct scsi_device *sdev;
1693         struct scsi_target *starget;
1694         struct Scsi_Host *shost;
1695
1696         blk_start_request(req);
1697
1698         scmd_printk(KERN_INFO, cmd, "killing request\n");
1699
1700         sdev = cmd->device;
1701         starget = scsi_target(sdev);
1702         shost = sdev->host;
1703         scsi_init_cmd_errh(cmd);
1704         cmd->result = DID_NO_CONNECT << 16;
1705         atomic_inc(&cmd->device->iorequest_cnt);
1706
1707         /*
1708          * SCSI request completion path will do scsi_device_unbusy(),
1709          * bump busy counts.  To bump the counters, we need to dance
1710          * with the locks as normal issue path does.
1711          */
1712         atomic_inc(&sdev->device_busy);
1713         atomic_inc(&shost->host_busy);
1714         if (starget->can_queue > 0)
1715                 atomic_inc(&starget->target_busy);
1716
1717         blk_complete_request(req);
1718 }
1719
1720 static void scsi_softirq_done(struct request *rq)
1721 {
1722         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1723         unsigned long wait_for = (cmd->allowed + 1) * rq->timeout;
1724         int disposition;
1725
1726         INIT_LIST_HEAD(&cmd->eh_entry);
1727
1728         atomic_inc(&cmd->device->iodone_cnt);
1729         if (cmd->result)
1730                 atomic_inc(&cmd->device->ioerr_cnt);
1731
1732         disposition = scsi_decide_disposition(cmd);
1733         if (disposition != SUCCESS &&
1734             time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) {
1735                 sdev_printk(KERN_ERR, cmd->device,
1736                             "timing out command, waited %lus\n",
1737                             wait_for/HZ);
1738                 disposition = SUCCESS;
1739         }
1740
1741         scsi_log_completion(cmd, disposition);
1742
1743         switch (disposition) {
1744                 case SUCCESS:
1745                         scsi_finish_command(cmd);
1746                         break;
1747                 case NEEDS_RETRY:
1748                         scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY);
1749                         break;
1750                 case ADD_TO_MLQUEUE:
1751                         scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY);
1752                         break;
1753                 default:
1754                         scsi_eh_scmd_add(cmd);
1755                         break;
1756         }
1757 }
1758
1759 /**
1760  * scsi_dispatch_command - Dispatch a command to the low-level driver.
1761  * @cmd: command block we are dispatching.
1762  *
1763  * Return: nonzero return request was rejected and device's queue needs to be
1764  * plugged.
1765  */
1766 static int scsi_dispatch_cmd(struct scsi_cmnd *cmd)
1767 {
1768         struct Scsi_Host *host = cmd->device->host;
1769         int rtn = 0;
1770
1771         atomic_inc(&cmd->device->iorequest_cnt);
1772
1773         /* check if the device is still usable */
1774         if (unlikely(cmd->device->sdev_state == SDEV_DEL)) {
1775                 /* in SDEV_DEL we error all commands. DID_NO_CONNECT
1776                  * returns an immediate error upwards, and signals
1777                  * that the device is no longer present */
1778                 cmd->result = DID_NO_CONNECT << 16;
1779                 goto done;
1780         }
1781
1782         /* Check to see if the scsi lld made this device blocked. */
1783         if (unlikely(scsi_device_blocked(cmd->device))) {
1784                 /*
1785                  * in blocked state, the command is just put back on
1786                  * the device queue.  The suspend state has already
1787                  * blocked the queue so future requests should not
1788                  * occur until the device transitions out of the
1789                  * suspend state.
1790                  */
1791                 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1792                         "queuecommand : device blocked\n"));
1793                 return SCSI_MLQUEUE_DEVICE_BUSY;
1794         }
1795
1796         /* Store the LUN value in cmnd, if needed. */
1797         if (cmd->device->lun_in_cdb)
1798                 cmd->cmnd[1] = (cmd->cmnd[1] & 0x1f) |
1799                                (cmd->device->lun << 5 & 0xe0);
1800
1801         scsi_log_send(cmd);
1802
1803         /*
1804          * Before we queue this command, check if the command
1805          * length exceeds what the host adapter can handle.
1806          */
1807         if (cmd->cmd_len > cmd->device->host->max_cmd_len) {
1808                 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1809                                "queuecommand : command too long. "
1810                                "cdb_size=%d host->max_cmd_len=%d\n",
1811                                cmd->cmd_len, cmd->device->host->max_cmd_len));
1812                 cmd->result = (DID_ABORT << 16);
1813                 goto done;
1814         }
1815
1816         if (unlikely(host->shost_state == SHOST_DEL)) {
1817                 cmd->result = (DID_NO_CONNECT << 16);
1818                 goto done;
1819
1820         }
1821
1822         trace_scsi_dispatch_cmd_start(cmd);
1823         rtn = host->hostt->queuecommand(host, cmd);
1824         if (rtn) {
1825                 trace_scsi_dispatch_cmd_error(cmd, rtn);
1826                 if (rtn != SCSI_MLQUEUE_DEVICE_BUSY &&
1827                     rtn != SCSI_MLQUEUE_TARGET_BUSY)
1828                         rtn = SCSI_MLQUEUE_HOST_BUSY;
1829
1830                 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1831                         "queuecommand : request rejected\n"));
1832         }
1833
1834         return rtn;
1835  done:
1836         cmd->scsi_done(cmd);
1837         return 0;
1838 }
1839
1840 /**
1841  * scsi_done - Invoke completion on finished SCSI command.
1842  * @cmd: The SCSI Command for which a low-level device driver (LLDD) gives
1843  * ownership back to SCSI Core -- i.e. the LLDD has finished with it.
1844  *
1845  * Description: This function is the mid-level's (SCSI Core) interrupt routine,
1846  * which regains ownership of the SCSI command (de facto) from a LLDD, and
1847  * calls blk_complete_request() for further processing.
1848  *
1849  * This function is interrupt context safe.
1850  */
1851 static void scsi_done(struct scsi_cmnd *cmd)
1852 {
1853         trace_scsi_dispatch_cmd_done(cmd);
1854         blk_complete_request(cmd->request);
1855 }
1856
1857 /*
1858  * Function:    scsi_request_fn()
1859  *
1860  * Purpose:     Main strategy routine for SCSI.
1861  *
1862  * Arguments:   q       - Pointer to actual queue.
1863  *
1864  * Returns:     Nothing
1865  *
1866  * Lock status: request queue lock assumed to be held when called.
1867  *
1868  * Note: See sd_zbc.c sd_zbc_write_lock_zone() for write order
1869  * protection for ZBC disks.
1870  */
1871 static void scsi_request_fn(struct request_queue *q)
1872         __releases(q->queue_lock)
1873         __acquires(q->queue_lock)
1874 {
1875         struct scsi_device *sdev = q->queuedata;
1876         struct Scsi_Host *shost;
1877         struct scsi_cmnd *cmd;
1878         struct request *req;
1879
1880         /*
1881          * To start with, we keep looping until the queue is empty, or until
1882          * the host is no longer able to accept any more requests.
1883          */
1884         shost = sdev->host;
1885         for (;;) {
1886                 int rtn;
1887                 /*
1888                  * get next queueable request.  We do this early to make sure
1889                  * that the request is fully prepared even if we cannot
1890                  * accept it.
1891                  */
1892                 req = blk_peek_request(q);
1893                 if (!req)
1894                         break;
1895
1896                 if (unlikely(!scsi_device_online(sdev))) {
1897                         sdev_printk(KERN_ERR, sdev,
1898                                     "rejecting I/O to offline device\n");
1899                         scsi_kill_request(req, q);
1900                         continue;
1901                 }
1902
1903                 if (!scsi_dev_queue_ready(q, sdev))
1904                         break;
1905
1906                 /*
1907                  * Remove the request from the request list.
1908                  */
1909                 if (!(blk_queue_tagged(q) && !blk_queue_start_tag(q, req)))
1910                         blk_start_request(req);
1911
1912                 spin_unlock_irq(q->queue_lock);
1913                 cmd = blk_mq_rq_to_pdu(req);
1914                 if (cmd != req->special) {
1915                         printk(KERN_CRIT "impossible request in %s.\n"
1916                                          "please mail a stack trace to "
1917                                          "linux-scsi@vger.kernel.org\n",
1918                                          __func__);
1919                         blk_dump_rq_flags(req, "foo");
1920                         BUG();
1921                 }
1922
1923                 /*
1924                  * We hit this when the driver is using a host wide
1925                  * tag map. For device level tag maps the queue_depth check
1926                  * in the device ready fn would prevent us from trying
1927                  * to allocate a tag. Since the map is a shared host resource
1928                  * we add the dev to the starved list so it eventually gets
1929                  * a run when a tag is freed.
1930                  */
1931                 if (blk_queue_tagged(q) && !(req->rq_flags & RQF_QUEUED)) {
1932                         spin_lock_irq(shost->host_lock);
1933                         if (list_empty(&sdev->starved_entry))
1934                                 list_add_tail(&sdev->starved_entry,
1935                                               &shost->starved_list);
1936                         spin_unlock_irq(shost->host_lock);
1937                         goto not_ready;
1938                 }
1939
1940                 if (!scsi_target_queue_ready(shost, sdev))
1941                         goto not_ready;
1942
1943                 if (!scsi_host_queue_ready(q, shost, sdev))
1944                         goto host_not_ready;
1945         
1946                 if (sdev->simple_tags)
1947                         cmd->flags |= SCMD_TAGGED;
1948                 else
1949                         cmd->flags &= ~SCMD_TAGGED;
1950
1951                 /*
1952                  * Finally, initialize any error handling parameters, and set up
1953                  * the timers for timeouts.
1954                  */
1955                 scsi_init_cmd_errh(cmd);
1956
1957                 /*
1958                  * Dispatch the command to the low-level driver.
1959                  */
1960                 cmd->scsi_done = scsi_done;
1961                 rtn = scsi_dispatch_cmd(cmd);
1962                 if (rtn) {
1963                         scsi_queue_insert(cmd, rtn);
1964                         spin_lock_irq(q->queue_lock);
1965                         goto out_delay;
1966                 }
1967                 spin_lock_irq(q->queue_lock);
1968         }
1969
1970         return;
1971
1972  host_not_ready:
1973         if (scsi_target(sdev)->can_queue > 0)
1974                 atomic_dec(&scsi_target(sdev)->target_busy);
1975  not_ready:
1976         /*
1977          * lock q, handle tag, requeue req, and decrement device_busy. We
1978          * must return with queue_lock held.
1979          *
1980          * Decrementing device_busy without checking it is OK, as all such
1981          * cases (host limits or settings) should run the queue at some
1982          * later time.
1983          */
1984         spin_lock_irq(q->queue_lock);
1985         blk_requeue_request(q, req);
1986         atomic_dec(&sdev->device_busy);
1987 out_delay:
1988         if (!atomic_read(&sdev->device_busy) && !scsi_device_blocked(sdev))
1989                 blk_delay_queue(q, SCSI_QUEUE_DELAY);
1990 }
1991
1992 static inline blk_status_t prep_to_mq(int ret)
1993 {
1994         switch (ret) {
1995         case BLKPREP_OK:
1996                 return BLK_STS_OK;
1997         case BLKPREP_DEFER:
1998                 return BLK_STS_RESOURCE;
1999         default:
2000                 return BLK_STS_IOERR;
2001         }
2002 }
2003
2004 /* Size in bytes of the sg-list stored in the scsi-mq command-private data. */
2005 static unsigned int scsi_mq_sgl_size(struct Scsi_Host *shost)
2006 {
2007         return min_t(unsigned int, shost->sg_tablesize, SG_CHUNK_SIZE) *
2008                 sizeof(struct scatterlist);
2009 }
2010
2011 static int scsi_mq_prep_fn(struct request *req)
2012 {
2013         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
2014         struct scsi_device *sdev = req->q->queuedata;
2015         struct Scsi_Host *shost = sdev->host;
2016         struct scatterlist *sg;
2017
2018         scsi_init_command(sdev, cmd);
2019
2020         req->special = cmd;
2021
2022         cmd->request = req;
2023
2024         cmd->tag = req->tag;
2025         cmd->prot_op = SCSI_PROT_NORMAL;
2026
2027         sg = (void *)cmd + sizeof(struct scsi_cmnd) + shost->hostt->cmd_size;
2028         cmd->sdb.table.sgl = sg;
2029
2030         if (scsi_host_get_prot(shost)) {
2031                 memset(cmd->prot_sdb, 0, sizeof(struct scsi_data_buffer));
2032
2033                 cmd->prot_sdb->table.sgl =
2034                         (struct scatterlist *)(cmd->prot_sdb + 1);
2035         }
2036
2037         if (blk_bidi_rq(req)) {
2038                 struct request *next_rq = req->next_rq;
2039                 struct scsi_data_buffer *bidi_sdb = blk_mq_rq_to_pdu(next_rq);
2040
2041                 memset(bidi_sdb, 0, sizeof(struct scsi_data_buffer));
2042                 bidi_sdb->table.sgl =
2043                         (struct scatterlist *)(bidi_sdb + 1);
2044
2045                 next_rq->special = bidi_sdb;
2046         }
2047
2048         blk_mq_start_request(req);
2049
2050         return scsi_setup_cmnd(sdev, req);
2051 }
2052
2053 static void scsi_mq_done(struct scsi_cmnd *cmd)
2054 {
2055         trace_scsi_dispatch_cmd_done(cmd);
2056         blk_mq_complete_request(cmd->request);
2057 }
2058
2059 static void scsi_mq_put_budget(struct blk_mq_hw_ctx *hctx)
2060 {
2061         struct request_queue *q = hctx->queue;
2062         struct scsi_device *sdev = q->queuedata;
2063
2064         atomic_dec(&sdev->device_busy);
2065         put_device(&sdev->sdev_gendev);
2066 }
2067
2068 static bool scsi_mq_get_budget(struct blk_mq_hw_ctx *hctx)
2069 {
2070         struct request_queue *q = hctx->queue;
2071         struct scsi_device *sdev = q->queuedata;
2072
2073         if (!get_device(&sdev->sdev_gendev))
2074                 goto out;
2075         if (!scsi_dev_queue_ready(q, sdev))
2076                 goto out_put_device;
2077
2078         return true;
2079
2080 out_put_device:
2081         put_device(&sdev->sdev_gendev);
2082 out:
2083         if (atomic_read(&sdev->device_busy) == 0 && !scsi_device_blocked(sdev))
2084                 blk_mq_delay_run_hw_queue(hctx, SCSI_QUEUE_DELAY);
2085         return false;
2086 }
2087
2088 static blk_status_t scsi_queue_rq(struct blk_mq_hw_ctx *hctx,
2089                          const struct blk_mq_queue_data *bd)
2090 {
2091         struct request *req = bd->rq;
2092         struct request_queue *q = req->q;
2093         struct scsi_device *sdev = q->queuedata;
2094         struct Scsi_Host *shost = sdev->host;
2095         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
2096         blk_status_t ret;
2097         int reason;
2098
2099         ret = prep_to_mq(scsi_prep_state_check(sdev, req));
2100         if (ret != BLK_STS_OK)
2101                 goto out_put_budget;
2102
2103         ret = BLK_STS_RESOURCE;
2104         if (!scsi_target_queue_ready(shost, sdev))
2105                 goto out_put_budget;
2106         if (!scsi_host_queue_ready(q, shost, sdev))
2107                 goto out_dec_target_busy;
2108
2109         if (!(req->rq_flags & RQF_DONTPREP)) {
2110                 ret = prep_to_mq(scsi_mq_prep_fn(req));
2111                 if (ret != BLK_STS_OK)
2112                         goto out_dec_host_busy;
2113                 req->rq_flags |= RQF_DONTPREP;
2114         } else {
2115                 blk_mq_start_request(req);
2116         }
2117
2118         if (sdev->simple_tags)
2119                 cmd->flags |= SCMD_TAGGED;
2120         else
2121                 cmd->flags &= ~SCMD_TAGGED;
2122
2123         scsi_init_cmd_errh(cmd);
2124         cmd->scsi_done = scsi_mq_done;
2125
2126         reason = scsi_dispatch_cmd(cmd);
2127         if (reason) {
2128                 scsi_set_blocked(cmd, reason);
2129                 ret = BLK_STS_RESOURCE;
2130                 goto out_dec_host_busy;
2131         }
2132
2133         return BLK_STS_OK;
2134
2135 out_dec_host_busy:
2136         scsi_dec_host_busy(shost);
2137 out_dec_target_busy:
2138         if (scsi_target(sdev)->can_queue > 0)
2139                 atomic_dec(&scsi_target(sdev)->target_busy);
2140 out_put_budget:
2141         scsi_mq_put_budget(hctx);
2142         switch (ret) {
2143         case BLK_STS_OK:
2144                 break;
2145         case BLK_STS_RESOURCE:
2146                 if (atomic_read(&sdev->device_busy) ||
2147                     scsi_device_blocked(sdev))
2148                         ret = BLK_STS_DEV_RESOURCE;
2149                 break;
2150         default:
2151                 /*
2152                  * Make sure to release all allocated ressources when
2153                  * we hit an error, as we will never see this command
2154                  * again.
2155                  */
2156                 if (req->rq_flags & RQF_DONTPREP)
2157                         scsi_mq_uninit_cmd(cmd);
2158                 break;
2159         }
2160         return ret;
2161 }
2162
2163 static enum blk_eh_timer_return scsi_timeout(struct request *req,
2164                 bool reserved)
2165 {
2166         if (reserved)
2167                 return BLK_EH_RESET_TIMER;
2168         return scsi_times_out(req);
2169 }
2170
2171 static int scsi_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
2172                                 unsigned int hctx_idx, unsigned int numa_node)
2173 {
2174         struct Scsi_Host *shost = set->driver_data;
2175         const bool unchecked_isa_dma = shost->unchecked_isa_dma;
2176         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
2177         struct scatterlist *sg;
2178
2179         if (unchecked_isa_dma)
2180                 cmd->flags |= SCMD_UNCHECKED_ISA_DMA;
2181         cmd->sense_buffer = scsi_alloc_sense_buffer(unchecked_isa_dma,
2182                                                     GFP_KERNEL, numa_node);
2183         if (!cmd->sense_buffer)
2184                 return -ENOMEM;
2185         cmd->req.sense = cmd->sense_buffer;
2186
2187         if (scsi_host_get_prot(shost)) {
2188                 sg = (void *)cmd + sizeof(struct scsi_cmnd) +
2189                         shost->hostt->cmd_size;
2190                 cmd->prot_sdb = (void *)sg + scsi_mq_sgl_size(shost);
2191         }
2192
2193         return 0;
2194 }
2195
2196 static void scsi_mq_exit_request(struct blk_mq_tag_set *set, struct request *rq,
2197                                  unsigned int hctx_idx)
2198 {
2199         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
2200
2201         scsi_free_sense_buffer(cmd->flags & SCMD_UNCHECKED_ISA_DMA,
2202                                cmd->sense_buffer);
2203 }
2204
2205 static int scsi_map_queues(struct blk_mq_tag_set *set)
2206 {
2207         struct Scsi_Host *shost = container_of(set, struct Scsi_Host, tag_set);
2208
2209         if (shost->hostt->map_queues)
2210                 return shost->hostt->map_queues(shost);
2211         return blk_mq_map_queues(set);
2212 }
2213
2214 void __scsi_init_queue(struct Scsi_Host *shost, struct request_queue *q)
2215 {
2216         struct device *dev = shost->dma_dev;
2217
2218         /*
2219          * this limit is imposed by hardware restrictions
2220          */
2221         blk_queue_max_segments(q, min_t(unsigned short, shost->sg_tablesize,
2222                                         SG_MAX_SEGMENTS));
2223
2224         if (scsi_host_prot_dma(shost)) {
2225                 shost->sg_prot_tablesize =
2226                         min_not_zero(shost->sg_prot_tablesize,
2227                                      (unsigned short)SCSI_MAX_PROT_SG_SEGMENTS);
2228                 BUG_ON(shost->sg_prot_tablesize < shost->sg_tablesize);
2229                 blk_queue_max_integrity_segments(q, shost->sg_prot_tablesize);
2230         }
2231
2232         blk_queue_max_hw_sectors(q, shost->max_sectors);
2233         if (shost->unchecked_isa_dma)
2234                 blk_queue_bounce_limit(q, BLK_BOUNCE_ISA);
2235         blk_queue_segment_boundary(q, shost->dma_boundary);
2236         dma_set_seg_boundary(dev, shost->dma_boundary);
2237
2238         blk_queue_max_segment_size(q, dma_get_max_seg_size(dev));
2239
2240         if (!shost->use_clustering)
2241                 q->limits.cluster = 0;
2242
2243         /*
2244          * Set a reasonable default alignment:  The larger of 32-byte (dword),
2245          * which is a common minimum for HBAs, and the minimum DMA alignment,
2246          * which is set by the platform.
2247          *
2248          * Devices that require a bigger alignment can increase it later.
2249          */
2250         blk_queue_dma_alignment(q, max(4, dma_get_cache_alignment()) - 1);
2251 }
2252 EXPORT_SYMBOL_GPL(__scsi_init_queue);
2253
2254 static int scsi_old_init_rq(struct request_queue *q, struct request *rq,
2255                             gfp_t gfp)
2256 {
2257         struct Scsi_Host *shost = q->rq_alloc_data;
2258         const bool unchecked_isa_dma = shost->unchecked_isa_dma;
2259         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
2260
2261         memset(cmd, 0, sizeof(*cmd));
2262
2263         if (unchecked_isa_dma)
2264                 cmd->flags |= SCMD_UNCHECKED_ISA_DMA;
2265         cmd->sense_buffer = scsi_alloc_sense_buffer(unchecked_isa_dma, gfp,
2266                                                     NUMA_NO_NODE);
2267         if (!cmd->sense_buffer)
2268                 goto fail;
2269         cmd->req.sense = cmd->sense_buffer;
2270
2271         if (scsi_host_get_prot(shost) >= SHOST_DIX_TYPE0_PROTECTION) {
2272                 cmd->prot_sdb = kmem_cache_zalloc(scsi_sdb_cache, gfp);
2273                 if (!cmd->prot_sdb)
2274                         goto fail_free_sense;
2275         }
2276
2277         return 0;
2278
2279 fail_free_sense:
2280         scsi_free_sense_buffer(unchecked_isa_dma, cmd->sense_buffer);
2281 fail:
2282         return -ENOMEM;
2283 }
2284
2285 static void scsi_old_exit_rq(struct request_queue *q, struct request *rq)
2286 {
2287         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
2288
2289         if (cmd->prot_sdb)
2290                 kmem_cache_free(scsi_sdb_cache, cmd->prot_sdb);
2291         scsi_free_sense_buffer(cmd->flags & SCMD_UNCHECKED_ISA_DMA,
2292                                cmd->sense_buffer);
2293 }
2294
2295 struct request_queue *scsi_old_alloc_queue(struct scsi_device *sdev)
2296 {
2297         struct Scsi_Host *shost = sdev->host;
2298         struct request_queue *q;
2299
2300         q = blk_alloc_queue_node(GFP_KERNEL, NUMA_NO_NODE, NULL);
2301         if (!q)
2302                 return NULL;
2303         q->cmd_size = sizeof(struct scsi_cmnd) + shost->hostt->cmd_size;
2304         q->rq_alloc_data = shost;
2305         q->request_fn = scsi_request_fn;
2306         q->init_rq_fn = scsi_old_init_rq;
2307         q->exit_rq_fn = scsi_old_exit_rq;
2308         q->initialize_rq_fn = scsi_initialize_rq;
2309
2310         if (blk_init_allocated_queue(q) < 0) {
2311                 blk_cleanup_queue(q);
2312                 return NULL;
2313         }
2314
2315         __scsi_init_queue(shost, q);
2316         blk_queue_flag_set(QUEUE_FLAG_SCSI_PASSTHROUGH, q);
2317         blk_queue_prep_rq(q, scsi_prep_fn);
2318         blk_queue_unprep_rq(q, scsi_unprep_fn);
2319         blk_queue_softirq_done(q, scsi_softirq_done);
2320         blk_queue_rq_timed_out(q, scsi_times_out);
2321         blk_queue_lld_busy(q, scsi_lld_busy);
2322         return q;
2323 }
2324
2325 static const struct blk_mq_ops scsi_mq_ops = {
2326         .get_budget     = scsi_mq_get_budget,
2327         .put_budget     = scsi_mq_put_budget,
2328         .queue_rq       = scsi_queue_rq,
2329         .complete       = scsi_softirq_done,
2330         .timeout        = scsi_timeout,
2331 #ifdef CONFIG_BLK_DEBUG_FS
2332         .show_rq        = scsi_show_rq,
2333 #endif
2334         .init_request   = scsi_mq_init_request,
2335         .exit_request   = scsi_mq_exit_request,
2336         .initialize_rq_fn = scsi_initialize_rq,
2337         .map_queues     = scsi_map_queues,
2338 };
2339
2340 struct request_queue *scsi_mq_alloc_queue(struct scsi_device *sdev)
2341 {
2342         sdev->request_queue = blk_mq_init_queue(&sdev->host->tag_set);
2343         if (IS_ERR(sdev->request_queue))
2344                 return NULL;
2345
2346         sdev->request_queue->queuedata = sdev;
2347         __scsi_init_queue(sdev->host, sdev->request_queue);
2348         blk_queue_flag_set(QUEUE_FLAG_SCSI_PASSTHROUGH, sdev->request_queue);
2349         return sdev->request_queue;
2350 }
2351
2352 int scsi_mq_setup_tags(struct Scsi_Host *shost)
2353 {
2354         unsigned int cmd_size, sgl_size;
2355
2356         sgl_size = scsi_mq_sgl_size(shost);
2357         cmd_size = sizeof(struct scsi_cmnd) + shost->hostt->cmd_size + sgl_size;
2358         if (scsi_host_get_prot(shost))
2359                 cmd_size += sizeof(struct scsi_data_buffer) + sgl_size;
2360
2361         memset(&shost->tag_set, 0, sizeof(shost->tag_set));
2362         shost->tag_set.ops = &scsi_mq_ops;
2363         shost->tag_set.nr_hw_queues = shost->nr_hw_queues ? : 1;
2364         shost->tag_set.queue_depth = shost->can_queue;
2365         shost->tag_set.cmd_size = cmd_size;
2366         shost->tag_set.numa_node = NUMA_NO_NODE;
2367         shost->tag_set.flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_SG_MERGE;
2368         shost->tag_set.flags |=
2369                 BLK_ALLOC_POLICY_TO_MQ_FLAG(shost->hostt->tag_alloc_policy);
2370         shost->tag_set.driver_data = shost;
2371
2372         return blk_mq_alloc_tag_set(&shost->tag_set);
2373 }
2374
2375 void scsi_mq_destroy_tags(struct Scsi_Host *shost)
2376 {
2377         blk_mq_free_tag_set(&shost->tag_set);
2378 }
2379
2380 /**
2381  * scsi_device_from_queue - return sdev associated with a request_queue
2382  * @q: The request queue to return the sdev from
2383  *
2384  * Return the sdev associated with a request queue or NULL if the
2385  * request_queue does not reference a SCSI device.
2386  */
2387 struct scsi_device *scsi_device_from_queue(struct request_queue *q)
2388 {
2389         struct scsi_device *sdev = NULL;
2390
2391         if (q->mq_ops) {
2392                 if (q->mq_ops == &scsi_mq_ops)
2393                         sdev = q->queuedata;
2394         } else if (q->request_fn == scsi_request_fn)
2395                 sdev = q->queuedata;
2396         if (!sdev || !get_device(&sdev->sdev_gendev))
2397                 sdev = NULL;
2398
2399         return sdev;
2400 }
2401 EXPORT_SYMBOL_GPL(scsi_device_from_queue);
2402
2403 /*
2404  * Function:    scsi_block_requests()
2405  *
2406  * Purpose:     Utility function used by low-level drivers to prevent further
2407  *              commands from being queued to the device.
2408  *
2409  * Arguments:   shost       - Host in question
2410  *
2411  * Returns:     Nothing
2412  *
2413  * Lock status: No locks are assumed held.
2414  *
2415  * Notes:       There is no timer nor any other means by which the requests
2416  *              get unblocked other than the low-level driver calling
2417  *              scsi_unblock_requests().
2418  */
2419 void scsi_block_requests(struct Scsi_Host *shost)
2420 {
2421         shost->host_self_blocked = 1;
2422 }
2423 EXPORT_SYMBOL(scsi_block_requests);
2424
2425 /*
2426  * Function:    scsi_unblock_requests()
2427  *
2428  * Purpose:     Utility function used by low-level drivers to allow further
2429  *              commands from being queued to the device.
2430  *
2431  * Arguments:   shost       - Host in question
2432  *
2433  * Returns:     Nothing
2434  *
2435  * Lock status: No locks are assumed held.
2436  *
2437  * Notes:       There is no timer nor any other means by which the requests
2438  *              get unblocked other than the low-level driver calling
2439  *              scsi_unblock_requests().
2440  *
2441  *              This is done as an API function so that changes to the
2442  *              internals of the scsi mid-layer won't require wholesale
2443  *              changes to drivers that use this feature.
2444  */
2445 void scsi_unblock_requests(struct Scsi_Host *shost)
2446 {
2447         shost->host_self_blocked = 0;
2448         scsi_run_host_queues(shost);
2449 }
2450 EXPORT_SYMBOL(scsi_unblock_requests);
2451
2452 int __init scsi_init_queue(void)
2453 {
2454         scsi_sdb_cache = kmem_cache_create("scsi_data_buffer",
2455                                            sizeof(struct scsi_data_buffer),
2456                                            0, 0, NULL);
2457         if (!scsi_sdb_cache) {
2458                 printk(KERN_ERR "SCSI: can't init scsi sdb cache\n");
2459                 return -ENOMEM;
2460         }
2461
2462         return 0;
2463 }
2464
2465 void scsi_exit_queue(void)
2466 {
2467         kmem_cache_destroy(scsi_sense_cache);
2468         kmem_cache_destroy(scsi_sense_isadma_cache);
2469         kmem_cache_destroy(scsi_sdb_cache);
2470 }
2471
2472 /**
2473  *      scsi_mode_select - issue a mode select
2474  *      @sdev:  SCSI device to be queried
2475  *      @pf:    Page format bit (1 == standard, 0 == vendor specific)
2476  *      @sp:    Save page bit (0 == don't save, 1 == save)
2477  *      @modepage: mode page being requested
2478  *      @buffer: request buffer (may not be smaller than eight bytes)
2479  *      @len:   length of request buffer.
2480  *      @timeout: command timeout
2481  *      @retries: number of retries before failing
2482  *      @data: returns a structure abstracting the mode header data
2483  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
2484  *              must be SCSI_SENSE_BUFFERSIZE big.
2485  *
2486  *      Returns zero if successful; negative error number or scsi
2487  *      status on error
2488  *
2489  */
2490 int
2491 scsi_mode_select(struct scsi_device *sdev, int pf, int sp, int modepage,
2492                  unsigned char *buffer, int len, int timeout, int retries,
2493                  struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
2494 {
2495         unsigned char cmd[10];
2496         unsigned char *real_buffer;
2497         int ret;
2498
2499         memset(cmd, 0, sizeof(cmd));
2500         cmd[1] = (pf ? 0x10 : 0) | (sp ? 0x01 : 0);
2501
2502         if (sdev->use_10_for_ms) {
2503                 if (len > 65535)
2504                         return -EINVAL;
2505                 real_buffer = kmalloc(8 + len, GFP_KERNEL);
2506                 if (!real_buffer)
2507                         return -ENOMEM;
2508                 memcpy(real_buffer + 8, buffer, len);
2509                 len += 8;
2510                 real_buffer[0] = 0;
2511                 real_buffer[1] = 0;
2512                 real_buffer[2] = data->medium_type;
2513                 real_buffer[3] = data->device_specific;
2514                 real_buffer[4] = data->longlba ? 0x01 : 0;
2515                 real_buffer[5] = 0;
2516                 real_buffer[6] = data->block_descriptor_length >> 8;
2517                 real_buffer[7] = data->block_descriptor_length;
2518
2519                 cmd[0] = MODE_SELECT_10;
2520                 cmd[7] = len >> 8;
2521                 cmd[8] = len;
2522         } else {
2523                 if (len > 255 || data->block_descriptor_length > 255 ||
2524                     data->longlba)
2525                         return -EINVAL;
2526
2527                 real_buffer = kmalloc(4 + len, GFP_KERNEL);
2528                 if (!real_buffer)
2529                         return -ENOMEM;
2530                 memcpy(real_buffer + 4, buffer, len);
2531                 len += 4;
2532                 real_buffer[0] = 0;
2533                 real_buffer[1] = data->medium_type;
2534                 real_buffer[2] = data->device_specific;
2535                 real_buffer[3] = data->block_descriptor_length;
2536                 
2537
2538                 cmd[0] = MODE_SELECT;
2539                 cmd[4] = len;
2540         }
2541
2542         ret = scsi_execute_req(sdev, cmd, DMA_TO_DEVICE, real_buffer, len,
2543                                sshdr, timeout, retries, NULL);
2544         kfree(real_buffer);
2545         return ret;
2546 }
2547 EXPORT_SYMBOL_GPL(scsi_mode_select);
2548
2549 /**
2550  *      scsi_mode_sense - issue a mode sense, falling back from 10 to six bytes if necessary.
2551  *      @sdev:  SCSI device to be queried
2552  *      @dbd:   set if mode sense will allow block descriptors to be returned
2553  *      @modepage: mode page being requested
2554  *      @buffer: request buffer (may not be smaller than eight bytes)
2555  *      @len:   length of request buffer.
2556  *      @timeout: command timeout
2557  *      @retries: number of retries before failing
2558  *      @data: returns a structure abstracting the mode header data
2559  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
2560  *              must be SCSI_SENSE_BUFFERSIZE big.
2561  *
2562  *      Returns zero if unsuccessful, or the header offset (either 4
2563  *      or 8 depending on whether a six or ten byte command was
2564  *      issued) if successful.
2565  */
2566 int
2567 scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
2568                   unsigned char *buffer, int len, int timeout, int retries,
2569                   struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
2570 {
2571         unsigned char cmd[12];
2572         int use_10_for_ms;
2573         int header_length;
2574         int result, retry_count = retries;
2575         struct scsi_sense_hdr my_sshdr;
2576
2577         memset(data, 0, sizeof(*data));
2578         memset(&cmd[0], 0, 12);
2579         cmd[1] = dbd & 0x18;    /* allows DBD and LLBA bits */
2580         cmd[2] = modepage;
2581
2582         /* caller might not be interested in sense, but we need it */
2583         if (!sshdr)
2584                 sshdr = &my_sshdr;
2585
2586  retry:
2587         use_10_for_ms = sdev->use_10_for_ms;
2588
2589         if (use_10_for_ms) {
2590                 if (len < 8)
2591                         len = 8;
2592
2593                 cmd[0] = MODE_SENSE_10;
2594                 cmd[8] = len;
2595                 header_length = 8;
2596         } else {
2597                 if (len < 4)
2598                         len = 4;
2599
2600                 cmd[0] = MODE_SENSE;
2601                 cmd[4] = len;
2602                 header_length = 4;
2603         }
2604
2605         memset(buffer, 0, len);
2606
2607         result = scsi_execute_req(sdev, cmd, DMA_FROM_DEVICE, buffer, len,
2608                                   sshdr, timeout, retries, NULL);
2609
2610         /* This code looks awful: what it's doing is making sure an
2611          * ILLEGAL REQUEST sense return identifies the actual command
2612          * byte as the problem.  MODE_SENSE commands can return
2613          * ILLEGAL REQUEST if the code page isn't supported */
2614
2615         if (use_10_for_ms && !scsi_status_is_good(result) &&
2616             driver_byte(result) == DRIVER_SENSE) {
2617                 if (scsi_sense_valid(sshdr)) {
2618                         if ((sshdr->sense_key == ILLEGAL_REQUEST) &&
2619                             (sshdr->asc == 0x20) && (sshdr->ascq == 0)) {
2620                                 /* 
2621                                  * Invalid command operation code
2622                                  */
2623                                 sdev->use_10_for_ms = 0;
2624                                 goto retry;
2625                         }
2626                 }
2627         }
2628
2629         if(scsi_status_is_good(result)) {
2630                 if (unlikely(buffer[0] == 0x86 && buffer[1] == 0x0b &&
2631                              (modepage == 6 || modepage == 8))) {
2632                         /* Initio breakage? */
2633                         header_length = 0;
2634                         data->length = 13;
2635                         data->medium_type = 0;
2636                         data->device_specific = 0;
2637                         data->longlba = 0;
2638                         data->block_descriptor_length = 0;
2639                 } else if(use_10_for_ms) {
2640                         data->length = buffer[0]*256 + buffer[1] + 2;
2641                         data->medium_type = buffer[2];
2642                         data->device_specific = buffer[3];
2643                         data->longlba = buffer[4] & 0x01;
2644                         data->block_descriptor_length = buffer[6]*256
2645                                 + buffer[7];
2646                 } else {
2647                         data->length = buffer[0] + 1;
2648                         data->medium_type = buffer[1];
2649                         data->device_specific = buffer[2];
2650                         data->block_descriptor_length = buffer[3];
2651                 }
2652                 data->header_length = header_length;
2653         } else if ((status_byte(result) == CHECK_CONDITION) &&
2654                    scsi_sense_valid(sshdr) &&
2655                    sshdr->sense_key == UNIT_ATTENTION && retry_count) {
2656                 retry_count--;
2657                 goto retry;
2658         }
2659
2660         return result;
2661 }
2662 EXPORT_SYMBOL(scsi_mode_sense);
2663
2664 /**
2665  *      scsi_test_unit_ready - test if unit is ready
2666  *      @sdev:  scsi device to change the state of.
2667  *      @timeout: command timeout
2668  *      @retries: number of retries before failing
2669  *      @sshdr: outpout pointer for decoded sense information.
2670  *
2671  *      Returns zero if unsuccessful or an error if TUR failed.  For
2672  *      removable media, UNIT_ATTENTION sets ->changed flag.
2673  **/
2674 int
2675 scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries,
2676                      struct scsi_sense_hdr *sshdr)
2677 {
2678         char cmd[] = {
2679                 TEST_UNIT_READY, 0, 0, 0, 0, 0,
2680         };
2681         int result;
2682
2683         /* try to eat the UNIT_ATTENTION if there are enough retries */
2684         do {
2685                 result = scsi_execute_req(sdev, cmd, DMA_NONE, NULL, 0, sshdr,
2686                                           timeout, 1, NULL);
2687                 if (sdev->removable && scsi_sense_valid(sshdr) &&
2688                     sshdr->sense_key == UNIT_ATTENTION)
2689                         sdev->changed = 1;
2690         } while (scsi_sense_valid(sshdr) &&
2691                  sshdr->sense_key == UNIT_ATTENTION && --retries);
2692
2693         return result;
2694 }
2695 EXPORT_SYMBOL(scsi_test_unit_ready);
2696
2697 /**
2698  *      scsi_device_set_state - Take the given device through the device state model.
2699  *      @sdev:  scsi device to change the state of.
2700  *      @state: state to change to.
2701  *
2702  *      Returns zero if successful or an error if the requested
2703  *      transition is illegal.
2704  */
2705 int
2706 scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
2707 {
2708         enum scsi_device_state oldstate = sdev->sdev_state;
2709
2710         if (state == oldstate)
2711                 return 0;
2712
2713         switch (state) {
2714         case SDEV_CREATED:
2715                 switch (oldstate) {
2716                 case SDEV_CREATED_BLOCK:
2717                         break;
2718                 default:
2719                         goto illegal;
2720                 }
2721                 break;
2722                         
2723         case SDEV_RUNNING:
2724                 switch (oldstate) {
2725                 case SDEV_CREATED:
2726                 case SDEV_OFFLINE:
2727                 case SDEV_TRANSPORT_OFFLINE:
2728                 case SDEV_QUIESCE:
2729                 case SDEV_BLOCK:
2730                         break;
2731                 default:
2732                         goto illegal;
2733                 }
2734                 break;
2735
2736         case SDEV_QUIESCE:
2737                 switch (oldstate) {
2738                 case SDEV_RUNNING:
2739                 case SDEV_OFFLINE:
2740                 case SDEV_TRANSPORT_OFFLINE:
2741                         break;
2742                 default:
2743                         goto illegal;
2744                 }
2745                 break;
2746
2747         case SDEV_OFFLINE:
2748         case SDEV_TRANSPORT_OFFLINE:
2749                 switch (oldstate) {
2750                 case SDEV_CREATED:
2751                 case SDEV_RUNNING:
2752                 case SDEV_QUIESCE:
2753                 case SDEV_BLOCK:
2754                         break;
2755                 default:
2756                         goto illegal;
2757                 }
2758                 break;
2759
2760         case SDEV_BLOCK:
2761                 switch (oldstate) {
2762                 case SDEV_RUNNING:
2763                 case SDEV_CREATED_BLOCK:
2764                 case SDEV_OFFLINE:
2765                         break;
2766                 default:
2767                         goto illegal;
2768                 }
2769                 break;
2770
2771         case SDEV_CREATED_BLOCK:
2772                 switch (oldstate) {
2773                 case SDEV_CREATED:
2774                         break;
2775                 default:
2776                         goto illegal;
2777                 }
2778                 break;
2779
2780         case SDEV_CANCEL:
2781                 switch (oldstate) {
2782                 case SDEV_CREATED:
2783                 case SDEV_RUNNING:
2784                 case SDEV_QUIESCE:
2785                 case SDEV_OFFLINE:
2786                 case SDEV_TRANSPORT_OFFLINE:
2787                         break;
2788                 default:
2789                         goto illegal;
2790                 }
2791                 break;
2792
2793         case SDEV_DEL:
2794                 switch (oldstate) {
2795                 case SDEV_CREATED:
2796                 case SDEV_RUNNING:
2797                 case SDEV_OFFLINE:
2798                 case SDEV_TRANSPORT_OFFLINE:
2799                 case SDEV_CANCEL:
2800                 case SDEV_BLOCK:
2801                 case SDEV_CREATED_BLOCK:
2802                         break;
2803                 default:
2804                         goto illegal;
2805                 }
2806                 break;
2807
2808         }
2809         sdev->sdev_state = state;
2810         return 0;
2811
2812  illegal:
2813         SCSI_LOG_ERROR_RECOVERY(1,
2814                                 sdev_printk(KERN_ERR, sdev,
2815                                             "Illegal state transition %s->%s",
2816                                             scsi_device_state_name(oldstate),
2817                                             scsi_device_state_name(state))
2818                                 );
2819         return -EINVAL;
2820 }
2821 EXPORT_SYMBOL(scsi_device_set_state);
2822
2823 /**
2824  *      sdev_evt_emit - emit a single SCSI device uevent
2825  *      @sdev: associated SCSI device
2826  *      @evt: event to emit
2827  *
2828  *      Send a single uevent (scsi_event) to the associated scsi_device.
2829  */
2830 static void scsi_evt_emit(struct scsi_device *sdev, struct scsi_event *evt)
2831 {
2832         int idx = 0;
2833         char *envp[3];
2834
2835         switch (evt->evt_type) {
2836         case SDEV_EVT_MEDIA_CHANGE:
2837                 envp[idx++] = "SDEV_MEDIA_CHANGE=1";
2838                 break;
2839         case SDEV_EVT_INQUIRY_CHANGE_REPORTED:
2840                 scsi_rescan_device(&sdev->sdev_gendev);
2841                 envp[idx++] = "SDEV_UA=INQUIRY_DATA_HAS_CHANGED";
2842                 break;
2843         case SDEV_EVT_CAPACITY_CHANGE_REPORTED:
2844                 envp[idx++] = "SDEV_UA=CAPACITY_DATA_HAS_CHANGED";
2845                 break;
2846         case SDEV_EVT_SOFT_THRESHOLD_REACHED_REPORTED:
2847                envp[idx++] = "SDEV_UA=THIN_PROVISIONING_SOFT_THRESHOLD_REACHED";
2848                 break;
2849         case SDEV_EVT_MODE_PARAMETER_CHANGE_REPORTED:
2850                 envp[idx++] = "SDEV_UA=MODE_PARAMETERS_CHANGED";
2851                 break;
2852         case SDEV_EVT_LUN_CHANGE_REPORTED:
2853                 envp[idx++] = "SDEV_UA=REPORTED_LUNS_DATA_HAS_CHANGED";
2854                 break;
2855         case SDEV_EVT_ALUA_STATE_CHANGE_REPORTED:
2856                 envp[idx++] = "SDEV_UA=ASYMMETRIC_ACCESS_STATE_CHANGED";
2857                 break;
2858         case SDEV_EVT_POWER_ON_RESET_OCCURRED:
2859                 envp[idx++] = "SDEV_UA=POWER_ON_RESET_OCCURRED";
2860                 break;
2861         default:
2862                 /* do nothing */
2863                 break;
2864         }
2865
2866         envp[idx++] = NULL;
2867
2868         kobject_uevent_env(&sdev->sdev_gendev.kobj, KOBJ_CHANGE, envp);
2869 }
2870
2871 /**
2872  *      sdev_evt_thread - send a uevent for each scsi event
2873  *      @work: work struct for scsi_device
2874  *
2875  *      Dispatch queued events to their associated scsi_device kobjects
2876  *      as uevents.
2877  */
2878 void scsi_evt_thread(struct work_struct *work)
2879 {
2880         struct scsi_device *sdev;
2881         enum scsi_device_event evt_type;
2882         LIST_HEAD(event_list);
2883
2884         sdev = container_of(work, struct scsi_device, event_work);
2885
2886         for (evt_type = SDEV_EVT_FIRST; evt_type <= SDEV_EVT_LAST; evt_type++)
2887                 if (test_and_clear_bit(evt_type, sdev->pending_events))
2888                         sdev_evt_send_simple(sdev, evt_type, GFP_KERNEL);
2889
2890         while (1) {
2891                 struct scsi_event *evt;
2892                 struct list_head *this, *tmp;
2893                 unsigned long flags;
2894
2895                 spin_lock_irqsave(&sdev->list_lock, flags);
2896                 list_splice_init(&sdev->event_list, &event_list);
2897                 spin_unlock_irqrestore(&sdev->list_lock, flags);
2898
2899                 if (list_empty(&event_list))
2900                         break;
2901
2902                 list_for_each_safe(this, tmp, &event_list) {
2903                         evt = list_entry(this, struct scsi_event, node);
2904                         list_del(&evt->node);
2905                         scsi_evt_emit(sdev, evt);
2906                         kfree(evt);
2907                 }
2908         }
2909 }
2910
2911 /**
2912  *      sdev_evt_send - send asserted event to uevent thread
2913  *      @sdev: scsi_device event occurred on
2914  *      @evt: event to send
2915  *
2916  *      Assert scsi device event asynchronously.
2917  */
2918 void sdev_evt_send(struct scsi_device *sdev, struct scsi_event *evt)
2919 {
2920         unsigned long flags;
2921
2922 #if 0
2923         /* FIXME: currently this check eliminates all media change events
2924          * for polled devices.  Need to update to discriminate between AN
2925          * and polled events */
2926         if (!test_bit(evt->evt_type, sdev->supported_events)) {
2927                 kfree(evt);
2928                 return;
2929         }
2930 #endif
2931
2932         spin_lock_irqsave(&sdev->list_lock, flags);
2933         list_add_tail(&evt->node, &sdev->event_list);
2934         schedule_work(&sdev->event_work);
2935         spin_unlock_irqrestore(&sdev->list_lock, flags);
2936 }
2937 EXPORT_SYMBOL_GPL(sdev_evt_send);
2938
2939 /**
2940  *      sdev_evt_alloc - allocate a new scsi event
2941  *      @evt_type: type of event to allocate
2942  *      @gfpflags: GFP flags for allocation
2943  *
2944  *      Allocates and returns a new scsi_event.
2945  */
2946 struct scsi_event *sdev_evt_alloc(enum scsi_device_event evt_type,
2947                                   gfp_t gfpflags)
2948 {
2949         struct scsi_event *evt = kzalloc(sizeof(struct scsi_event), gfpflags);
2950         if (!evt)
2951                 return NULL;
2952
2953         evt->evt_type = evt_type;
2954         INIT_LIST_HEAD(&evt->node);
2955
2956         /* evt_type-specific initialization, if any */
2957         switch (evt_type) {
2958         case SDEV_EVT_MEDIA_CHANGE:
2959         case SDEV_EVT_INQUIRY_CHANGE_REPORTED:
2960         case SDEV_EVT_CAPACITY_CHANGE_REPORTED:
2961         case SDEV_EVT_SOFT_THRESHOLD_REACHED_REPORTED:
2962         case SDEV_EVT_MODE_PARAMETER_CHANGE_REPORTED:
2963         case SDEV_EVT_LUN_CHANGE_REPORTED:
2964         case SDEV_EVT_ALUA_STATE_CHANGE_REPORTED:
2965         case SDEV_EVT_POWER_ON_RESET_OCCURRED:
2966         default:
2967                 /* do nothing */
2968                 break;
2969         }
2970
2971         return evt;
2972 }
2973 EXPORT_SYMBOL_GPL(sdev_evt_alloc);
2974
2975 /**
2976  *      sdev_evt_send_simple - send asserted event to uevent thread
2977  *      @sdev: scsi_device event occurred on
2978  *      @evt_type: type of event to send
2979  *      @gfpflags: GFP flags for allocation
2980  *
2981  *      Assert scsi device event asynchronously, given an event type.
2982  */
2983 void sdev_evt_send_simple(struct scsi_device *sdev,
2984                           enum scsi_device_event evt_type, gfp_t gfpflags)
2985 {
2986         struct scsi_event *evt = sdev_evt_alloc(evt_type, gfpflags);
2987         if (!evt) {
2988                 sdev_printk(KERN_ERR, sdev, "event %d eaten due to OOM\n",
2989                             evt_type);
2990                 return;
2991         }
2992
2993         sdev_evt_send(sdev, evt);
2994 }
2995 EXPORT_SYMBOL_GPL(sdev_evt_send_simple);
2996
2997 /**
2998  * scsi_request_fn_active() - number of kernel threads inside scsi_request_fn()
2999  * @sdev: SCSI device to count the number of scsi_request_fn() callers for.
3000  */
3001 static int scsi_request_fn_active(struct scsi_device *sdev)
3002 {
3003         struct request_queue *q = sdev->request_queue;
3004         int request_fn_active;
3005
3006         WARN_ON_ONCE(sdev->host->use_blk_mq);
3007
3008         spin_lock_irq(q->queue_lock);
3009         request_fn_active = q->request_fn_active;
3010         spin_unlock_irq(q->queue_lock);
3011
3012         return request_fn_active;
3013 }
3014
3015 /**
3016  * scsi_wait_for_queuecommand() - wait for ongoing queuecommand() calls
3017  * @sdev: SCSI device pointer.
3018  *
3019  * Wait until the ongoing shost->hostt->queuecommand() calls that are
3020  * invoked from scsi_request_fn() have finished.
3021  */
3022 static void scsi_wait_for_queuecommand(struct scsi_device *sdev)
3023 {
3024         WARN_ON_ONCE(sdev->host->use_blk_mq);
3025
3026         while (scsi_request_fn_active(sdev))
3027                 msleep(20);
3028 }
3029
3030 /**
3031  *      scsi_device_quiesce - Block user issued commands.
3032  *      @sdev:  scsi device to quiesce.
3033  *
3034  *      This works by trying to transition to the SDEV_QUIESCE state
3035  *      (which must be a legal transition).  When the device is in this
3036  *      state, only special requests will be accepted, all others will
3037  *      be deferred.  Since special requests may also be requeued requests,
3038  *      a successful return doesn't guarantee the device will be 
3039  *      totally quiescent.
3040  *
3041  *      Must be called with user context, may sleep.
3042  *
3043  *      Returns zero if unsuccessful or an error if not.
3044  */
3045 int
3046 scsi_device_quiesce(struct scsi_device *sdev)
3047 {
3048         struct request_queue *q = sdev->request_queue;
3049         int err;
3050
3051         /*
3052          * It is allowed to call scsi_device_quiesce() multiple times from
3053          * the same context but concurrent scsi_device_quiesce() calls are
3054          * not allowed.
3055          */
3056         WARN_ON_ONCE(sdev->quiesced_by && sdev->quiesced_by != current);
3057
3058         if (sdev->quiesced_by == current)
3059                 return 0;
3060
3061         blk_set_pm_only(q);
3062
3063         blk_mq_freeze_queue(q);
3064         /*
3065          * Ensure that the effect of blk_set_pm_only() will be visible
3066          * for percpu_ref_tryget() callers that occur after the queue
3067          * unfreeze even if the queue was already frozen before this function
3068          * was called. See also https://lwn.net/Articles/573497/.
3069          */
3070         synchronize_rcu();
3071         blk_mq_unfreeze_queue(q);
3072
3073         mutex_lock(&sdev->state_mutex);
3074         err = scsi_device_set_state(sdev, SDEV_QUIESCE);
3075         if (err == 0)
3076                 sdev->quiesced_by = current;
3077         else
3078                 blk_clear_pm_only(q);
3079         mutex_unlock(&sdev->state_mutex);
3080
3081         return err;
3082 }
3083 EXPORT_SYMBOL(scsi_device_quiesce);
3084
3085 /**
3086  *      scsi_device_resume - Restart user issued commands to a quiesced device.
3087  *      @sdev:  scsi device to resume.
3088  *
3089  *      Moves the device from quiesced back to running and restarts the
3090  *      queues.
3091  *
3092  *      Must be called with user context, may sleep.
3093  */
3094 void scsi_device_resume(struct scsi_device *sdev)
3095 {
3096         /* check if the device state was mutated prior to resume, and if
3097          * so assume the state is being managed elsewhere (for example
3098          * device deleted during suspend)
3099          */
3100         mutex_lock(&sdev->state_mutex);
3101         WARN_ON_ONCE(!sdev->quiesced_by);
3102         sdev->quiesced_by = NULL;
3103         blk_clear_pm_only(sdev->request_queue);
3104         if (sdev->sdev_state == SDEV_QUIESCE)
3105                 scsi_device_set_state(sdev, SDEV_RUNNING);
3106         mutex_unlock(&sdev->state_mutex);
3107 }
3108 EXPORT_SYMBOL(scsi_device_resume);
3109
3110 static void
3111 device_quiesce_fn(struct scsi_device *sdev, void *data)
3112 {
3113         scsi_device_quiesce(sdev);
3114 }
3115
3116 void
3117 scsi_target_quiesce(struct scsi_target *starget)
3118 {
3119         starget_for_each_device(starget, NULL, device_quiesce_fn);
3120 }
3121 EXPORT_SYMBOL(scsi_target_quiesce);
3122
3123 static void
3124 device_resume_fn(struct scsi_device *sdev, void *data)
3125 {
3126         scsi_device_resume(sdev);
3127 }
3128
3129 void
3130 scsi_target_resume(struct scsi_target *starget)
3131 {
3132         starget_for_each_device(starget, NULL, device_resume_fn);
3133 }
3134 EXPORT_SYMBOL(scsi_target_resume);
3135
3136 /**
3137  * scsi_internal_device_block_nowait - try to transition to the SDEV_BLOCK state
3138  * @sdev: device to block
3139  *
3140  * Pause SCSI command processing on the specified device. Does not sleep.
3141  *
3142  * Returns zero if successful or a negative error code upon failure.
3143  *
3144  * Notes:
3145  * This routine transitions the device to the SDEV_BLOCK state (which must be
3146  * a legal transition). When the device is in this state, command processing
3147  * is paused until the device leaves the SDEV_BLOCK state. See also
3148  * scsi_internal_device_unblock_nowait().
3149  */
3150 int scsi_internal_device_block_nowait(struct scsi_device *sdev)
3151 {
3152         struct request_queue *q = sdev->request_queue;
3153         unsigned long flags;
3154         int err = 0;
3155
3156         err = scsi_device_set_state(sdev, SDEV_BLOCK);
3157         if (err) {
3158                 err = scsi_device_set_state(sdev, SDEV_CREATED_BLOCK);
3159
3160                 if (err)
3161                         return err;
3162         }
3163
3164         /* 
3165          * The device has transitioned to SDEV_BLOCK.  Stop the
3166          * block layer from calling the midlayer with this device's
3167          * request queue. 
3168          */
3169         if (q->mq_ops) {
3170                 blk_mq_quiesce_queue_nowait(q);
3171         } else {
3172                 spin_lock_irqsave(q->queue_lock, flags);
3173                 blk_stop_queue(q);
3174                 spin_unlock_irqrestore(q->queue_lock, flags);
3175         }
3176
3177         return 0;
3178 }
3179 EXPORT_SYMBOL_GPL(scsi_internal_device_block_nowait);
3180
3181 /**
3182  * scsi_internal_device_block - try to transition to the SDEV_BLOCK state
3183  * @sdev: device to block
3184  *
3185  * Pause SCSI command processing on the specified device and wait until all
3186  * ongoing scsi_request_fn() / scsi_queue_rq() calls have finished. May sleep.
3187  *
3188  * Returns zero if successful or a negative error code upon failure.
3189  *
3190  * Note:
3191  * This routine transitions the device to the SDEV_BLOCK state (which must be
3192  * a legal transition). When the device is in this state, command processing
3193  * is paused until the device leaves the SDEV_BLOCK state. See also
3194  * scsi_internal_device_unblock().
3195  *
3196  * To do: avoid that scsi_send_eh_cmnd() calls queuecommand() after
3197  * scsi_internal_device_block() has blocked a SCSI device and also
3198  * remove the rport mutex lock and unlock calls from srp_queuecommand().
3199  */
3200 static int scsi_internal_device_block(struct scsi_device *sdev)
3201 {
3202         struct request_queue *q = sdev->request_queue;
3203         int err;
3204
3205         mutex_lock(&sdev->state_mutex);
3206         err = scsi_internal_device_block_nowait(sdev);
3207         if (err == 0) {
3208                 if (q->mq_ops)
3209                         blk_mq_quiesce_queue(q);
3210                 else
3211                         scsi_wait_for_queuecommand(sdev);
3212         }
3213         mutex_unlock(&sdev->state_mutex);
3214
3215         return err;
3216 }
3217  
3218 void scsi_start_queue(struct scsi_device *sdev)
3219 {
3220         struct request_queue *q = sdev->request_queue;
3221         unsigned long flags;
3222
3223         if (q->mq_ops) {
3224                 blk_mq_unquiesce_queue(q);
3225         } else {
3226                 spin_lock_irqsave(q->queue_lock, flags);
3227                 blk_start_queue(q);
3228                 spin_unlock