c7fccbb8f5545e463537389bc34b607de3dc9cb3
[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                 __blk_mq_end_request(req, error);
701
702                 if (scsi_target(sdev)->single_lun ||
703                     !list_empty(&sdev->host->starved_list))
704                         kblockd_schedule_work(&sdev->requeue_work);
705                 else
706                         blk_mq_run_hw_queues(q, true);
707         } else {
708                 unsigned long flags;
709
710                 if (bidi_bytes)
711                         scsi_release_bidi_buffers(cmd);
712                 scsi_release_buffers(cmd);
713                 scsi_put_command(cmd);
714
715                 spin_lock_irqsave(q->queue_lock, flags);
716                 blk_finish_request(req, error);
717                 spin_unlock_irqrestore(q->queue_lock, flags);
718
719                 scsi_run_queue(q);
720         }
721
722         put_device(&sdev->sdev_gendev);
723         return false;
724 }
725
726 /**
727  * scsi_result_to_blk_status - translate a SCSI result code into blk_status_t
728  * @cmd:        SCSI command
729  * @result:     scsi error code
730  *
731  * Translate a SCSI result code into a blk_status_t value. May reset the host
732  * byte of @cmd->result.
733  */
734 static blk_status_t scsi_result_to_blk_status(struct scsi_cmnd *cmd, int result)
735 {
736         switch (host_byte(result)) {
737         case DID_OK:
738                 /*
739                  * Also check the other bytes than the status byte in result
740                  * to handle the case when a SCSI LLD sets result to
741                  * DRIVER_SENSE << 24 without setting SAM_STAT_CHECK_CONDITION.
742                  */
743                 if (scsi_status_is_good(result) && (result & ~0xff) == 0)
744                         return BLK_STS_OK;
745                 return BLK_STS_IOERR;
746         case DID_TRANSPORT_FAILFAST:
747                 return BLK_STS_TRANSPORT;
748         case DID_TARGET_FAILURE:
749                 set_host_byte(cmd, DID_OK);
750                 return BLK_STS_TARGET;
751         case DID_NEXUS_FAILURE:
752                 return BLK_STS_NEXUS;
753         case DID_ALLOC_FAILURE:
754                 set_host_byte(cmd, DID_OK);
755                 return BLK_STS_NOSPC;
756         case DID_MEDIUM_ERROR:
757                 set_host_byte(cmd, DID_OK);
758                 return BLK_STS_MEDIUM;
759         default:
760                 return BLK_STS_IOERR;
761         }
762 }
763
764 /* Helper for scsi_io_completion() when "reprep" action required. */
765 static void scsi_io_completion_reprep(struct scsi_cmnd *cmd,
766                                       struct request_queue *q)
767 {
768         /* A new command will be prepared and issued. */
769         if (q->mq_ops) {
770                 scsi_mq_requeue_cmd(cmd);
771         } else {
772                 /* Unprep request and put it back at head of the queue. */
773                 scsi_release_buffers(cmd);
774                 scsi_requeue_command(q, cmd);
775         }
776 }
777
778 /* Helper for scsi_io_completion() when special action required. */
779 static void scsi_io_completion_action(struct scsi_cmnd *cmd, int result)
780 {
781         struct request_queue *q = cmd->device->request_queue;
782         struct request *req = cmd->request;
783         int level = 0;
784         enum {ACTION_FAIL, ACTION_REPREP, ACTION_RETRY,
785               ACTION_DELAYED_RETRY} action;
786         unsigned long wait_for = (cmd->allowed + 1) * req->timeout;
787         struct scsi_sense_hdr sshdr;
788         bool sense_valid;
789         bool sense_current = true;      /* false implies "deferred sense" */
790         blk_status_t blk_stat;
791
792         sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
793         if (sense_valid)
794                 sense_current = !scsi_sense_is_deferred(&sshdr);
795
796         blk_stat = scsi_result_to_blk_status(cmd, result);
797
798         if (host_byte(result) == DID_RESET) {
799                 /* Third party bus reset or reset for error recovery
800                  * reasons.  Just retry the command and see what
801                  * happens.
802                  */
803                 action = ACTION_RETRY;
804         } else if (sense_valid && sense_current) {
805                 switch (sshdr.sense_key) {
806                 case UNIT_ATTENTION:
807                         if (cmd->device->removable) {
808                                 /* Detected disc change.  Set a bit
809                                  * and quietly refuse further access.
810                                  */
811                                 cmd->device->changed = 1;
812                                 action = ACTION_FAIL;
813                         } else {
814                                 /* Must have been a power glitch, or a
815                                  * bus reset.  Could not have been a
816                                  * media change, so we just retry the
817                                  * command and see what happens.
818                                  */
819                                 action = ACTION_RETRY;
820                         }
821                         break;
822                 case ILLEGAL_REQUEST:
823                         /* If we had an ILLEGAL REQUEST returned, then
824                          * we may have performed an unsupported
825                          * command.  The only thing this should be
826                          * would be a ten byte read where only a six
827                          * byte read was supported.  Also, on a system
828                          * where READ CAPACITY failed, we may have
829                          * read past the end of the disk.
830                          */
831                         if ((cmd->device->use_10_for_rw &&
832                             sshdr.asc == 0x20 && sshdr.ascq == 0x00) &&
833                             (cmd->cmnd[0] == READ_10 ||
834                              cmd->cmnd[0] == WRITE_10)) {
835                                 /* This will issue a new 6-byte command. */
836                                 cmd->device->use_10_for_rw = 0;
837                                 action = ACTION_REPREP;
838                         } else if (sshdr.asc == 0x10) /* DIX */ {
839                                 action = ACTION_FAIL;
840                                 blk_stat = BLK_STS_PROTECTION;
841                         /* INVALID COMMAND OPCODE or INVALID FIELD IN CDB */
842                         } else if (sshdr.asc == 0x20 || sshdr.asc == 0x24) {
843                                 action = ACTION_FAIL;
844                                 blk_stat = BLK_STS_TARGET;
845                         } else
846                                 action = ACTION_FAIL;
847                         break;
848                 case ABORTED_COMMAND:
849                         action = ACTION_FAIL;
850                         if (sshdr.asc == 0x10) /* DIF */
851                                 blk_stat = BLK_STS_PROTECTION;
852                         break;
853                 case NOT_READY:
854                         /* If the device is in the process of becoming
855                          * ready, or has a temporary blockage, retry.
856                          */
857                         if (sshdr.asc == 0x04) {
858                                 switch (sshdr.ascq) {
859                                 case 0x01: /* becoming ready */
860                                 case 0x04: /* format in progress */
861                                 case 0x05: /* rebuild in progress */
862                                 case 0x06: /* recalculation in progress */
863                                 case 0x07: /* operation in progress */
864                                 case 0x08: /* Long write in progress */
865                                 case 0x09: /* self test in progress */
866                                 case 0x14: /* space allocation in progress */
867                                 case 0x1a: /* start stop unit in progress */
868                                 case 0x1b: /* sanitize in progress */
869                                 case 0x1d: /* configuration in progress */
870                                 case 0x24: /* depopulation in progress */
871                                         action = ACTION_DELAYED_RETRY;
872                                         break;
873                                 default:
874                                         action = ACTION_FAIL;
875                                         break;
876                                 }
877                         } else
878                                 action = ACTION_FAIL;
879                         break;
880                 case VOLUME_OVERFLOW:
881                         /* See SSC3rXX or current. */
882                         action = ACTION_FAIL;
883                         break;
884                 default:
885                         action = ACTION_FAIL;
886                         break;
887                 }
888         } else
889                 action = ACTION_FAIL;
890
891         if (action != ACTION_FAIL &&
892             time_before(cmd->jiffies_at_alloc + wait_for, jiffies))
893                 action = ACTION_FAIL;
894
895         switch (action) {
896         case ACTION_FAIL:
897                 /* Give up and fail the remainder of the request */
898                 if (!(req->rq_flags & RQF_QUIET)) {
899                         static DEFINE_RATELIMIT_STATE(_rs,
900                                         DEFAULT_RATELIMIT_INTERVAL,
901                                         DEFAULT_RATELIMIT_BURST);
902
903                         if (unlikely(scsi_logging_level))
904                                 level =
905                                      SCSI_LOG_LEVEL(SCSI_LOG_MLCOMPLETE_SHIFT,
906                                                     SCSI_LOG_MLCOMPLETE_BITS);
907
908                         /*
909                          * if logging is enabled the failure will be printed
910                          * in scsi_log_completion(), so avoid duplicate messages
911                          */
912                         if (!level && __ratelimit(&_rs)) {
913                                 scsi_print_result(cmd, NULL, FAILED);
914                                 if (driver_byte(result) == DRIVER_SENSE)
915                                         scsi_print_sense(cmd);
916                                 scsi_print_command(cmd);
917                         }
918                 }
919                 if (!scsi_end_request(req, blk_stat, blk_rq_err_bytes(req), 0))
920                         return;
921                 /*FALLTHRU*/
922         case ACTION_REPREP:
923                 scsi_io_completion_reprep(cmd, q);
924                 break;
925         case ACTION_RETRY:
926                 /* Retry the same command immediately */
927                 __scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY, false);
928                 break;
929         case ACTION_DELAYED_RETRY:
930                 /* Retry the same command after a delay */
931                 __scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY, false);
932                 break;
933         }
934 }
935
936 /*
937  * Helper for scsi_io_completion() when cmd->result is non-zero. Returns a
938  * new result that may suppress further error checking. Also modifies
939  * *blk_statp in some cases.
940  */
941 static int scsi_io_completion_nz_result(struct scsi_cmnd *cmd, int result,
942                                         blk_status_t *blk_statp)
943 {
944         bool sense_valid;
945         bool sense_current = true;      /* false implies "deferred sense" */
946         struct request *req = cmd->request;
947         struct scsi_sense_hdr sshdr;
948
949         sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
950         if (sense_valid)
951                 sense_current = !scsi_sense_is_deferred(&sshdr);
952
953         if (blk_rq_is_passthrough(req)) {
954                 if (sense_valid) {
955                         /*
956                          * SG_IO wants current and deferred errors
957                          */
958                         scsi_req(req)->sense_len =
959                                 min(8 + cmd->sense_buffer[7],
960                                     SCSI_SENSE_BUFFERSIZE);
961                 }
962                 if (sense_current)
963                         *blk_statp = scsi_result_to_blk_status(cmd, result);
964         } else if (blk_rq_bytes(req) == 0 && sense_current) {
965                 /*
966                  * Flush commands do not transfers any data, and thus cannot use
967                  * good_bytes != blk_rq_bytes(req) as the signal for an error.
968                  * This sets *blk_statp explicitly for the problem case.
969                  */
970                 *blk_statp = scsi_result_to_blk_status(cmd, result);
971         }
972         /*
973          * Recovered errors need reporting, but they're always treated as
974          * success, so fiddle the result code here.  For passthrough requests
975          * we already took a copy of the original into sreq->result which
976          * is what gets returned to the user
977          */
978         if (sense_valid && (sshdr.sense_key == RECOVERED_ERROR)) {
979                 bool do_print = true;
980                 /*
981                  * if ATA PASS-THROUGH INFORMATION AVAILABLE [0x0, 0x1d]
982                  * skip print since caller wants ATA registers. Only occurs
983                  * on SCSI ATA PASS_THROUGH commands when CK_COND=1
984                  */
985                 if ((sshdr.asc == 0x0) && (sshdr.ascq == 0x1d))
986                         do_print = false;
987                 else if (req->rq_flags & RQF_QUIET)
988                         do_print = false;
989                 if (do_print)
990                         scsi_print_sense(cmd);
991                 result = 0;
992                 /* for passthrough, *blk_statp may be set */
993                 *blk_statp = BLK_STS_OK;
994         }
995         /*
996          * Another corner case: the SCSI status byte is non-zero but 'good'.
997          * Example: PRE-FETCH command returns SAM_STAT_CONDITION_MET when
998          * it is able to fit nominated LBs in its cache (and SAM_STAT_GOOD
999          * if it can't fit). Treat SAM_STAT_CONDITION_MET and the related
1000          * intermediate statuses (both obsolete in SAM-4) as good.
1001          */
1002         if (status_byte(result) && scsi_status_is_good(result)) {
1003                 result = 0;
1004                 *blk_statp = BLK_STS_OK;
1005         }
1006         return result;
1007 }
1008
1009 /*
1010  * Function:    scsi_io_completion()
1011  *
1012  * Purpose:     Completion processing for block device I/O requests.
1013  *
1014  * Arguments:   cmd   - command that is finished.
1015  *
1016  * Lock status: Assumed that no lock is held upon entry.
1017  *
1018  * Returns:     Nothing
1019  *
1020  * Notes:       We will finish off the specified number of sectors.  If we
1021  *              are done, the command block will be released and the queue
1022  *              function will be goosed.  If we are not done then we have to
1023  *              figure out what to do next:
1024  *
1025  *              a) We can call scsi_requeue_command().  The request
1026  *                 will be unprepared and put back on the queue.  Then
1027  *                 a new command will be created for it.  This should
1028  *                 be used if we made forward progress, or if we want
1029  *                 to switch from READ(10) to READ(6) for example.
1030  *
1031  *              b) We can call __scsi_queue_insert().  The request will
1032  *                 be put back on the queue and retried using the same
1033  *                 command as before, possibly after a delay.
1034  *
1035  *              c) We can call scsi_end_request() with blk_stat other than
1036  *                 BLK_STS_OK, to fail the remainder of the request.
1037  */
1038 void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes)
1039 {
1040         int result = cmd->result;
1041         struct request_queue *q = cmd->device->request_queue;
1042         struct request *req = cmd->request;
1043         blk_status_t blk_stat = BLK_STS_OK;
1044
1045         if (unlikely(result))   /* a nz result may or may not be an error */
1046                 result = scsi_io_completion_nz_result(cmd, result, &blk_stat);
1047
1048         if (unlikely(blk_rq_is_passthrough(req))) {
1049                 /*
1050                  * scsi_result_to_blk_status may have reset the host_byte
1051                  */
1052                 scsi_req(req)->result = cmd->result;
1053                 scsi_req(req)->resid_len = scsi_get_resid(cmd);
1054
1055                 if (unlikely(scsi_bidi_cmnd(cmd))) {
1056                         /*
1057                          * Bidi commands Must be complete as a whole,
1058                          * both sides at once.
1059                          */
1060                         scsi_req(req->next_rq)->resid_len = scsi_in(cmd)->resid;
1061                         if (scsi_end_request(req, BLK_STS_OK, blk_rq_bytes(req),
1062                                         blk_rq_bytes(req->next_rq)))
1063                                 WARN_ONCE(true,
1064                                           "Bidi command with remaining bytes");
1065                         return;
1066                 }
1067         }
1068
1069         /* no bidi support yet, other than in pass-through */
1070         if (unlikely(blk_bidi_rq(req))) {
1071                 WARN_ONCE(true, "Only support bidi command in passthrough");
1072                 scmd_printk(KERN_ERR, cmd, "Killing bidi command\n");
1073                 if (scsi_end_request(req, BLK_STS_IOERR, blk_rq_bytes(req),
1074                                      blk_rq_bytes(req->next_rq)))
1075                         WARN_ONCE(true, "Bidi command with remaining bytes");
1076                 return;
1077         }
1078
1079         /*
1080          * Next deal with any sectors which we were able to correctly
1081          * handle.
1082          */
1083         SCSI_LOG_HLCOMPLETE(1, scmd_printk(KERN_INFO, cmd,
1084                 "%u sectors total, %d bytes done.\n",
1085                 blk_rq_sectors(req), good_bytes));
1086
1087         /*
1088          * Next deal with any sectors which we were able to correctly
1089          * handle. Failed, zero length commands always need to drop down
1090          * to retry code. Fast path should return in this block.
1091          */
1092         if (likely(blk_rq_bytes(req) > 0 || blk_stat == BLK_STS_OK)) {
1093                 if (likely(!scsi_end_request(req, blk_stat, good_bytes, 0)))
1094                         return; /* no bytes remaining */
1095         }
1096
1097         /* Kill remainder if no retries. */
1098         if (unlikely(blk_stat && scsi_noretry_cmd(cmd))) {
1099                 if (scsi_end_request(req, blk_stat, blk_rq_bytes(req), 0))
1100                         WARN_ONCE(true,
1101                             "Bytes remaining after failed, no-retry command");
1102                 return;
1103         }
1104
1105         /*
1106          * If there had been no error, but we have leftover bytes in the
1107          * requeues just queue the command up again.
1108          */
1109         if (likely(result == 0))
1110                 scsi_io_completion_reprep(cmd, q);
1111         else
1112                 scsi_io_completion_action(cmd, result);
1113 }
1114
1115 static int scsi_init_sgtable(struct request *req, struct scsi_data_buffer *sdb)
1116 {
1117         int count;
1118
1119         /*
1120          * If sg table allocation fails, requeue request later.
1121          */
1122         if (unlikely(sg_alloc_table_chained(&sdb->table,
1123                         blk_rq_nr_phys_segments(req), sdb->table.sgl)))
1124                 return BLKPREP_DEFER;
1125
1126         /* 
1127          * Next, walk the list, and fill in the addresses and sizes of
1128          * each segment.
1129          */
1130         count = blk_rq_map_sg(req->q, req, sdb->table.sgl);
1131         BUG_ON(count > sdb->table.nents);
1132         sdb->table.nents = count;
1133         sdb->length = blk_rq_payload_bytes(req);
1134         return BLKPREP_OK;
1135 }
1136
1137 /*
1138  * Function:    scsi_init_io()
1139  *
1140  * Purpose:     SCSI I/O initialize function.
1141  *
1142  * Arguments:   cmd   - Command descriptor we wish to initialize
1143  *
1144  * Returns:     0 on success
1145  *              BLKPREP_DEFER if the failure is retryable
1146  *              BLKPREP_KILL if the failure is fatal
1147  */
1148 int scsi_init_io(struct scsi_cmnd *cmd)
1149 {
1150         struct scsi_device *sdev = cmd->device;
1151         struct request *rq = cmd->request;
1152         bool is_mq = (rq->mq_ctx != NULL);
1153         int error = BLKPREP_KILL;
1154
1155         if (WARN_ON_ONCE(!blk_rq_nr_phys_segments(rq)))
1156                 goto err_exit;
1157
1158         error = scsi_init_sgtable(rq, &cmd->sdb);
1159         if (error)
1160                 goto err_exit;
1161
1162         if (blk_bidi_rq(rq)) {
1163                 if (!rq->q->mq_ops) {
1164                         struct scsi_data_buffer *bidi_sdb =
1165                                 kmem_cache_zalloc(scsi_sdb_cache, GFP_ATOMIC);
1166                         if (!bidi_sdb) {
1167                                 error = BLKPREP_DEFER;
1168                                 goto err_exit;
1169                         }
1170
1171                         rq->next_rq->special = bidi_sdb;
1172                 }
1173
1174                 error = scsi_init_sgtable(rq->next_rq, rq->next_rq->special);
1175                 if (error)
1176                         goto err_exit;
1177         }
1178
1179         if (blk_integrity_rq(rq)) {
1180                 struct scsi_data_buffer *prot_sdb = cmd->prot_sdb;
1181                 int ivecs, count;
1182
1183                 if (prot_sdb == NULL) {
1184                         /*
1185                          * This can happen if someone (e.g. multipath)
1186                          * queues a command to a device on an adapter
1187                          * that does not support DIX.
1188                          */
1189                         WARN_ON_ONCE(1);
1190                         error = BLKPREP_KILL;
1191                         goto err_exit;
1192                 }
1193
1194                 ivecs = blk_rq_count_integrity_sg(rq->q, rq->bio);
1195
1196                 if (sg_alloc_table_chained(&prot_sdb->table, ivecs,
1197                                 prot_sdb->table.sgl)) {
1198                         error = BLKPREP_DEFER;
1199                         goto err_exit;
1200                 }
1201
1202                 count = blk_rq_map_integrity_sg(rq->q, rq->bio,
1203                                                 prot_sdb->table.sgl);
1204                 BUG_ON(count > ivecs);
1205                 BUG_ON(count > queue_max_integrity_segments(rq->q));
1206
1207                 cmd->prot_sdb = prot_sdb;
1208                 cmd->prot_sdb->table.nents = count;
1209         }
1210
1211         return BLKPREP_OK;
1212 err_exit:
1213         if (is_mq) {
1214                 scsi_mq_free_sgtables(cmd);
1215         } else {
1216                 scsi_release_buffers(cmd);
1217                 cmd->request->special = NULL;
1218                 scsi_put_command(cmd);
1219                 put_device(&sdev->sdev_gendev);
1220         }
1221         return error;
1222 }
1223 EXPORT_SYMBOL(scsi_init_io);
1224
1225 /**
1226  * scsi_initialize_rq - initialize struct scsi_cmnd partially
1227  * @rq: Request associated with the SCSI command to be initialized.
1228  *
1229  * This function initializes the members of struct scsi_cmnd that must be
1230  * initialized before request processing starts and that won't be
1231  * reinitialized if a SCSI command is requeued.
1232  *
1233  * Called from inside blk_get_request() for pass-through requests and from
1234  * inside scsi_init_command() for filesystem requests.
1235  */
1236 static void scsi_initialize_rq(struct request *rq)
1237 {
1238         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1239
1240         scsi_req_init(&cmd->req);
1241         init_rcu_head(&cmd->rcu);
1242         cmd->jiffies_at_alloc = jiffies;
1243         cmd->retries = 0;
1244 }
1245
1246 /* Add a command to the list used by the aacraid and dpt_i2o drivers */
1247 void scsi_add_cmd_to_list(struct scsi_cmnd *cmd)
1248 {
1249         struct scsi_device *sdev = cmd->device;
1250         struct Scsi_Host *shost = sdev->host;
1251         unsigned long flags;
1252
1253         if (shost->use_cmd_list) {
1254                 spin_lock_irqsave(&sdev->list_lock, flags);
1255                 list_add_tail(&cmd->list, &sdev->cmd_list);
1256                 spin_unlock_irqrestore(&sdev->list_lock, flags);
1257         }
1258 }
1259
1260 /* Remove a command from the list used by the aacraid and dpt_i2o drivers */
1261 void scsi_del_cmd_from_list(struct scsi_cmnd *cmd)
1262 {
1263         struct scsi_device *sdev = cmd->device;
1264         struct Scsi_Host *shost = sdev->host;
1265         unsigned long flags;
1266
1267         if (shost->use_cmd_list) {
1268                 spin_lock_irqsave(&sdev->list_lock, flags);
1269                 BUG_ON(list_empty(&cmd->list));
1270                 list_del_init(&cmd->list);
1271                 spin_unlock_irqrestore(&sdev->list_lock, flags);
1272         }
1273 }
1274
1275 /* Called after a request has been started. */
1276 void scsi_init_command(struct scsi_device *dev, struct scsi_cmnd *cmd)
1277 {
1278         void *buf = cmd->sense_buffer;
1279         void *prot = cmd->prot_sdb;
1280         struct request *rq = blk_mq_rq_from_pdu(cmd);
1281         unsigned int flags = cmd->flags & SCMD_PRESERVED_FLAGS;
1282         unsigned long jiffies_at_alloc;
1283         int retries;
1284
1285         if (!blk_rq_is_scsi(rq) && !(flags & SCMD_INITIALIZED)) {
1286                 flags |= SCMD_INITIALIZED;
1287                 scsi_initialize_rq(rq);
1288         }
1289
1290         jiffies_at_alloc = cmd->jiffies_at_alloc;
1291         retries = cmd->retries;
1292         /* zero out the cmd, except for the embedded scsi_request */
1293         memset((char *)cmd + sizeof(cmd->req), 0,
1294                 sizeof(*cmd) - sizeof(cmd->req) + dev->host->hostt->cmd_size);
1295
1296         cmd->device = dev;
1297         cmd->sense_buffer = buf;
1298         cmd->prot_sdb = prot;
1299         cmd->flags = flags;
1300         INIT_DELAYED_WORK(&cmd->abort_work, scmd_eh_abort_handler);
1301         cmd->jiffies_at_alloc = jiffies_at_alloc;
1302         cmd->retries = retries;
1303
1304         scsi_add_cmd_to_list(cmd);
1305 }
1306
1307 static int scsi_setup_scsi_cmnd(struct scsi_device *sdev, struct request *req)
1308 {
1309         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1310
1311         /*
1312          * Passthrough requests may transfer data, in which case they must
1313          * a bio attached to them.  Or they might contain a SCSI command
1314          * that does not transfer data, in which case they may optionally
1315          * submit a request without an attached bio.
1316          */
1317         if (req->bio) {
1318                 int ret = scsi_init_io(cmd);
1319                 if (unlikely(ret))
1320                         return ret;
1321         } else {
1322                 BUG_ON(blk_rq_bytes(req));
1323
1324                 memset(&cmd->sdb, 0, sizeof(cmd->sdb));
1325         }
1326
1327         cmd->cmd_len = scsi_req(req)->cmd_len;
1328         cmd->cmnd = scsi_req(req)->cmd;
1329         cmd->transfersize = blk_rq_bytes(req);
1330         cmd->allowed = scsi_req(req)->retries;
1331         return BLKPREP_OK;
1332 }
1333
1334 /*
1335  * Setup a normal block command.  These are simple request from filesystems
1336  * that still need to be translated to SCSI CDBs from the ULD.
1337  */
1338 static int scsi_setup_fs_cmnd(struct scsi_device *sdev, struct request *req)
1339 {
1340         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1341
1342         if (unlikely(sdev->handler && sdev->handler->prep_fn)) {
1343                 int ret = sdev->handler->prep_fn(sdev, req);
1344                 if (ret != BLKPREP_OK)
1345                         return ret;
1346         }
1347
1348         cmd->cmnd = scsi_req(req)->cmd = scsi_req(req)->__cmd;
1349         memset(cmd->cmnd, 0, BLK_MAX_CDB);
1350         return scsi_cmd_to_driver(cmd)->init_command(cmd);
1351 }
1352
1353 static int scsi_setup_cmnd(struct scsi_device *sdev, struct request *req)
1354 {
1355         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1356
1357         if (!blk_rq_bytes(req))
1358                 cmd->sc_data_direction = DMA_NONE;
1359         else if (rq_data_dir(req) == WRITE)
1360                 cmd->sc_data_direction = DMA_TO_DEVICE;
1361         else
1362                 cmd->sc_data_direction = DMA_FROM_DEVICE;
1363
1364         if (blk_rq_is_scsi(req))
1365                 return scsi_setup_scsi_cmnd(sdev, req);
1366         else
1367                 return scsi_setup_fs_cmnd(sdev, req);
1368 }
1369
1370 static int
1371 scsi_prep_state_check(struct scsi_device *sdev, struct request *req)
1372 {
1373         int ret = BLKPREP_OK;
1374
1375         /*
1376          * If the device is not in running state we will reject some
1377          * or all commands.
1378          */
1379         if (unlikely(sdev->sdev_state != SDEV_RUNNING)) {
1380                 switch (sdev->sdev_state) {
1381                 case SDEV_OFFLINE:
1382                 case SDEV_TRANSPORT_OFFLINE:
1383                         /*
1384                          * If the device is offline we refuse to process any
1385                          * commands.  The device must be brought online
1386                          * before trying any recovery commands.
1387                          */
1388                         sdev_printk(KERN_ERR, sdev,
1389                                     "rejecting I/O to offline device\n");
1390                         ret = BLKPREP_KILL;
1391                         break;
1392                 case SDEV_DEL:
1393                         /*
1394                          * If the device is fully deleted, we refuse to
1395                          * process any commands as well.
1396                          */
1397                         sdev_printk(KERN_ERR, sdev,
1398                                     "rejecting I/O to dead device\n");
1399                         ret = BLKPREP_KILL;
1400                         break;
1401                 case SDEV_BLOCK:
1402                 case SDEV_CREATED_BLOCK:
1403                         ret = BLKPREP_DEFER;
1404                         break;
1405                 case SDEV_QUIESCE:
1406                         /*
1407                          * If the devices is blocked we defer normal commands.
1408                          */
1409                         if (req && !(req->rq_flags & RQF_PREEMPT))
1410                                 ret = BLKPREP_DEFER;
1411                         break;
1412                 default:
1413                         /*
1414                          * For any other not fully online state we only allow
1415                          * special commands.  In particular any user initiated
1416                          * command is not allowed.
1417                          */
1418                         if (req && !(req->rq_flags & RQF_PREEMPT))
1419                                 ret = BLKPREP_KILL;
1420                         break;
1421                 }
1422         }
1423         return ret;
1424 }
1425
1426 static int
1427 scsi_prep_return(struct request_queue *q, struct request *req, int ret)
1428 {
1429         struct scsi_device *sdev = q->queuedata;
1430
1431         switch (ret) {
1432         case BLKPREP_KILL:
1433         case BLKPREP_INVALID:
1434                 scsi_req(req)->result = DID_NO_CONNECT << 16;
1435                 /* release the command and kill it */
1436                 if (req->special) {
1437                         struct scsi_cmnd *cmd = req->special;
1438                         scsi_release_buffers(cmd);
1439                         scsi_put_command(cmd);
1440                         put_device(&sdev->sdev_gendev);
1441                         req->special = NULL;
1442                 }
1443                 break;
1444         case BLKPREP_DEFER:
1445                 /*
1446                  * If we defer, the blk_peek_request() returns NULL, but the
1447                  * queue must be restarted, so we schedule a callback to happen
1448                  * shortly.
1449                  */
1450                 if (atomic_read(&sdev->device_busy) == 0)
1451                         blk_delay_queue(q, SCSI_QUEUE_DELAY);
1452                 break;
1453         default:
1454                 req->rq_flags |= RQF_DONTPREP;
1455         }
1456
1457         return ret;
1458 }
1459
1460 static int scsi_prep_fn(struct request_queue *q, struct request *req)
1461 {
1462         struct scsi_device *sdev = q->queuedata;
1463         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1464         int ret;
1465
1466         ret = scsi_prep_state_check(sdev, req);
1467         if (ret != BLKPREP_OK)
1468                 goto out;
1469
1470         if (!req->special) {
1471                 /* Bail if we can't get a reference to the device */
1472                 if (unlikely(!get_device(&sdev->sdev_gendev))) {
1473                         ret = BLKPREP_DEFER;
1474                         goto out;
1475                 }
1476
1477                 scsi_init_command(sdev, cmd);
1478                 req->special = cmd;
1479         }
1480
1481         cmd->tag = req->tag;
1482         cmd->request = req;
1483         cmd->prot_op = SCSI_PROT_NORMAL;
1484
1485         ret = scsi_setup_cmnd(sdev, req);
1486 out:
1487         return scsi_prep_return(q, req, ret);
1488 }
1489
1490 static void scsi_unprep_fn(struct request_queue *q, struct request *req)
1491 {
1492         scsi_uninit_cmd(blk_mq_rq_to_pdu(req));
1493 }
1494
1495 /*
1496  * scsi_dev_queue_ready: if we can send requests to sdev, return 1 else
1497  * return 0.
1498  *
1499  * Called with the queue_lock held.
1500  */
1501 static inline int scsi_dev_queue_ready(struct request_queue *q,
1502                                   struct scsi_device *sdev)
1503 {
1504         unsigned int busy;
1505
1506         busy = atomic_inc_return(&sdev->device_busy) - 1;
1507         if (atomic_read(&sdev->device_blocked)) {
1508                 if (busy)
1509                         goto out_dec;
1510
1511                 /*
1512                  * unblock after device_blocked iterates to zero
1513                  */
1514                 if (atomic_dec_return(&sdev->device_blocked) > 0) {
1515                         /*
1516                          * For the MQ case we take care of this in the caller.
1517                          */
1518                         if (!q->mq_ops)
1519                                 blk_delay_queue(q, SCSI_QUEUE_DELAY);
1520                         goto out_dec;
1521                 }
1522                 SCSI_LOG_MLQUEUE(3, sdev_printk(KERN_INFO, sdev,
1523                                    "unblocking device at zero depth\n"));
1524         }
1525
1526         if (busy >= sdev->queue_depth)
1527                 goto out_dec;
1528
1529         return 1;
1530 out_dec:
1531         atomic_dec(&sdev->device_busy);
1532         return 0;
1533 }
1534
1535 /*
1536  * scsi_target_queue_ready: checks if there we can send commands to target
1537  * @sdev: scsi device on starget to check.
1538  */
1539 static inline int scsi_target_queue_ready(struct Scsi_Host *shost,
1540                                            struct scsi_device *sdev)
1541 {
1542         struct scsi_target *starget = scsi_target(sdev);
1543         unsigned int busy;
1544
1545         if (starget->single_lun) {
1546                 spin_lock_irq(shost->host_lock);
1547                 if (starget->starget_sdev_user &&
1548                     starget->starget_sdev_user != sdev) {
1549                         spin_unlock_irq(shost->host_lock);
1550                         return 0;
1551                 }
1552                 starget->starget_sdev_user = sdev;
1553                 spin_unlock_irq(shost->host_lock);
1554         }
1555
1556         if (starget->can_queue <= 0)
1557                 return 1;
1558
1559         busy = atomic_inc_return(&starget->target_busy) - 1;
1560         if (atomic_read(&starget->target_blocked) > 0) {
1561                 if (busy)
1562                         goto starved;
1563
1564                 /*
1565                  * unblock after target_blocked iterates to zero
1566                  */
1567                 if (atomic_dec_return(&starget->target_blocked) > 0)
1568                         goto out_dec;
1569
1570                 SCSI_LOG_MLQUEUE(3, starget_printk(KERN_INFO, starget,
1571                                  "unblocking target at zero depth\n"));
1572         }
1573
1574         if (busy >= starget->can_queue)
1575                 goto starved;
1576
1577         return 1;
1578
1579 starved:
1580         spin_lock_irq(shost->host_lock);
1581         list_move_tail(&sdev->starved_entry, &shost->starved_list);
1582         spin_unlock_irq(shost->host_lock);
1583 out_dec:
1584         if (starget->can_queue > 0)
1585                 atomic_dec(&starget->target_busy);
1586         return 0;
1587 }
1588
1589 /*
1590  * scsi_host_queue_ready: if we can send requests to shost, return 1 else
1591  * return 0. We must end up running the queue again whenever 0 is
1592  * returned, else IO can hang.
1593  */
1594 static inline int scsi_host_queue_ready(struct request_queue *q,
1595                                    struct Scsi_Host *shost,
1596                                    struct scsi_device *sdev)
1597 {
1598         unsigned int busy;
1599
1600         if (scsi_host_in_recovery(shost))
1601                 return 0;
1602
1603         busy = atomic_inc_return(&shost->host_busy) - 1;
1604         if (atomic_read(&shost->host_blocked) > 0) {
1605                 if (busy)
1606                         goto starved;
1607
1608                 /*
1609                  * unblock after host_blocked iterates to zero
1610                  */
1611                 if (atomic_dec_return(&shost->host_blocked) > 0)
1612                         goto out_dec;
1613
1614                 SCSI_LOG_MLQUEUE(3,
1615                         shost_printk(KERN_INFO, shost,
1616                                      "unblocking host at zero depth\n"));
1617         }
1618
1619         if (shost->can_queue > 0 && busy >= shost->can_queue)
1620                 goto starved;
1621         if (shost->host_self_blocked)
1622                 goto starved;
1623
1624         /* We're OK to process the command, so we can't be starved */
1625         if (!list_empty(&sdev->starved_entry)) {
1626                 spin_lock_irq(shost->host_lock);
1627                 if (!list_empty(&sdev->starved_entry))
1628                         list_del_init(&sdev->starved_entry);
1629                 spin_unlock_irq(shost->host_lock);
1630         }
1631
1632         return 1;
1633
1634 starved:
1635         spin_lock_irq(shost->host_lock);
1636         if (list_empty(&sdev->starved_entry))
1637                 list_add_tail(&sdev->starved_entry, &shost->starved_list);
1638         spin_unlock_irq(shost->host_lock);
1639 out_dec:
1640         scsi_dec_host_busy(shost);
1641         return 0;
1642 }
1643
1644 /*
1645  * Busy state exporting function for request stacking drivers.
1646  *
1647  * For efficiency, no lock is taken to check the busy state of
1648  * shost/starget/sdev, since the returned value is not guaranteed and
1649  * may be changed after request stacking drivers call the function,
1650  * regardless of taking lock or not.
1651  *
1652  * When scsi can't dispatch I/Os anymore and needs to kill I/Os scsi
1653  * needs to return 'not busy'. Otherwise, request stacking drivers
1654  * may hold requests forever.
1655  */
1656 static int scsi_lld_busy(struct request_queue *q)
1657 {
1658         struct scsi_device *sdev = q->queuedata;
1659         struct Scsi_Host *shost;
1660
1661         if (blk_queue_dying(q))
1662                 return 0;
1663
1664         shost = sdev->host;
1665
1666         /*
1667          * Ignore host/starget busy state.
1668          * Since block layer does not have a concept of fairness across
1669          * multiple queues, congestion of host/starget needs to be handled
1670          * in SCSI layer.
1671          */
1672         if (scsi_host_in_recovery(shost) || scsi_device_is_busy(sdev))
1673                 return 1;
1674
1675         return 0;
1676 }
1677
1678 /*
1679  * Kill a request for a dead device
1680  */
1681 static void scsi_kill_request(struct request *req, struct request_queue *q)
1682 {
1683         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1684         struct scsi_device *sdev;
1685         struct scsi_target *starget;
1686         struct Scsi_Host *shost;
1687
1688         blk_start_request(req);
1689
1690         scmd_printk(KERN_INFO, cmd, "killing request\n");
1691
1692         sdev = cmd->device;
1693         starget = scsi_target(sdev);
1694         shost = sdev->host;
1695         scsi_init_cmd_errh(cmd);
1696         cmd->result = DID_NO_CONNECT << 16;
1697         atomic_inc(&cmd->device->iorequest_cnt);
1698
1699         /*
1700          * SCSI request completion path will do scsi_device_unbusy(),
1701          * bump busy counts.  To bump the counters, we need to dance
1702          * with the locks as normal issue path does.
1703          */
1704         atomic_inc(&sdev->device_busy);
1705         atomic_inc(&shost->host_busy);
1706         if (starget->can_queue > 0)
1707                 atomic_inc(&starget->target_busy);
1708
1709         blk_complete_request(req);
1710 }
1711
1712 static void scsi_softirq_done(struct request *rq)
1713 {
1714         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1715         unsigned long wait_for = (cmd->allowed + 1) * rq->timeout;
1716         int disposition;
1717
1718         INIT_LIST_HEAD(&cmd->eh_entry);
1719
1720         atomic_inc(&cmd->device->iodone_cnt);
1721         if (cmd->result)
1722                 atomic_inc(&cmd->device->ioerr_cnt);
1723
1724         disposition = scsi_decide_disposition(cmd);
1725         if (disposition != SUCCESS &&
1726             time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) {
1727                 sdev_printk(KERN_ERR, cmd->device,
1728                             "timing out command, waited %lus\n",
1729                             wait_for/HZ);
1730                 disposition = SUCCESS;
1731         }
1732
1733         scsi_log_completion(cmd, disposition);
1734
1735         switch (disposition) {
1736                 case SUCCESS:
1737                         scsi_finish_command(cmd);
1738                         break;
1739                 case NEEDS_RETRY:
1740                         scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY);
1741                         break;
1742                 case ADD_TO_MLQUEUE:
1743                         scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY);
1744                         break;
1745                 default:
1746                         scsi_eh_scmd_add(cmd);
1747                         break;
1748         }
1749 }
1750
1751 /**
1752  * scsi_dispatch_command - Dispatch a command to the low-level driver.
1753  * @cmd: command block we are dispatching.
1754  *
1755  * Return: nonzero return request was rejected and device's queue needs to be
1756  * plugged.
1757  */
1758 static int scsi_dispatch_cmd(struct scsi_cmnd *cmd)
1759 {
1760         struct Scsi_Host *host = cmd->device->host;
1761         int rtn = 0;
1762
1763         atomic_inc(&cmd->device->iorequest_cnt);
1764
1765         /* check if the device is still usable */
1766         if (unlikely(cmd->device->sdev_state == SDEV_DEL)) {
1767                 /* in SDEV_DEL we error all commands. DID_NO_CONNECT
1768                  * returns an immediate error upwards, and signals
1769                  * that the device is no longer present */
1770                 cmd->result = DID_NO_CONNECT << 16;
1771                 goto done;
1772         }
1773
1774         /* Check to see if the scsi lld made this device blocked. */
1775         if (unlikely(scsi_device_blocked(cmd->device))) {
1776                 /*
1777                  * in blocked state, the command is just put back on
1778                  * the device queue.  The suspend state has already
1779                  * blocked the queue so future requests should not
1780                  * occur until the device transitions out of the
1781                  * suspend state.
1782                  */
1783                 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1784                         "queuecommand : device blocked\n"));
1785                 return SCSI_MLQUEUE_DEVICE_BUSY;
1786         }
1787
1788         /* Store the LUN value in cmnd, if needed. */
1789         if (cmd->device->lun_in_cdb)
1790                 cmd->cmnd[1] = (cmd->cmnd[1] & 0x1f) |
1791                                (cmd->device->lun << 5 & 0xe0);
1792
1793         scsi_log_send(cmd);
1794
1795         /*
1796          * Before we queue this command, check if the command
1797          * length exceeds what the host adapter can handle.
1798          */
1799         if (cmd->cmd_len > cmd->device->host->max_cmd_len) {
1800                 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1801                                "queuecommand : command too long. "
1802                                "cdb_size=%d host->max_cmd_len=%d\n",
1803                                cmd->cmd_len, cmd->device->host->max_cmd_len));
1804                 cmd->result = (DID_ABORT << 16);
1805                 goto done;
1806         }
1807
1808         if (unlikely(host->shost_state == SHOST_DEL)) {
1809                 cmd->result = (DID_NO_CONNECT << 16);
1810                 goto done;
1811
1812         }
1813
1814         trace_scsi_dispatch_cmd_start(cmd);
1815         rtn = host->hostt->queuecommand(host, cmd);
1816         if (rtn) {
1817                 trace_scsi_dispatch_cmd_error(cmd, rtn);
1818                 if (rtn != SCSI_MLQUEUE_DEVICE_BUSY &&
1819                     rtn != SCSI_MLQUEUE_TARGET_BUSY)
1820                         rtn = SCSI_MLQUEUE_HOST_BUSY;
1821
1822                 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1823                         "queuecommand : request rejected\n"));
1824         }
1825
1826         return rtn;
1827  done:
1828         cmd->scsi_done(cmd);
1829         return 0;
1830 }
1831
1832 /**
1833  * scsi_done - Invoke completion on finished SCSI command.
1834  * @cmd: The SCSI Command for which a low-level device driver (LLDD) gives
1835  * ownership back to SCSI Core -- i.e. the LLDD has finished with it.
1836  *
1837  * Description: This function is the mid-level's (SCSI Core) interrupt routine,
1838  * which regains ownership of the SCSI command (de facto) from a LLDD, and
1839  * calls blk_complete_request() for further processing.
1840  *
1841  * This function is interrupt context safe.
1842  */
1843 static void scsi_done(struct scsi_cmnd *cmd)
1844 {
1845         trace_scsi_dispatch_cmd_done(cmd);
1846         blk_complete_request(cmd->request);
1847 }
1848
1849 /*
1850  * Function:    scsi_request_fn()
1851  *
1852  * Purpose:     Main strategy routine for SCSI.
1853  *
1854  * Arguments:   q       - Pointer to actual queue.
1855  *
1856  * Returns:     Nothing
1857  *
1858  * Lock status: request queue lock assumed to be held when called.
1859  *
1860  * Note: See sd_zbc.c sd_zbc_write_lock_zone() for write order
1861  * protection for ZBC disks.
1862  */
1863 static void scsi_request_fn(struct request_queue *q)
1864         __releases(q->queue_lock)
1865         __acquires(q->queue_lock)
1866 {
1867         struct scsi_device *sdev = q->queuedata;
1868         struct Scsi_Host *shost;
1869         struct scsi_cmnd *cmd;
1870         struct request *req;
1871
1872         /*
1873          * To start with, we keep looping until the queue is empty, or until
1874          * the host is no longer able to accept any more requests.
1875          */
1876         shost = sdev->host;
1877         for (;;) {
1878                 int rtn;
1879                 /*
1880                  * get next queueable request.  We do this early to make sure
1881                  * that the request is fully prepared even if we cannot
1882                  * accept it.
1883                  */
1884                 req = blk_peek_request(q);
1885                 if (!req)
1886                         break;
1887
1888                 if (unlikely(!scsi_device_online(sdev))) {
1889                         sdev_printk(KERN_ERR, sdev,
1890                                     "rejecting I/O to offline device\n");
1891                         scsi_kill_request(req, q);
1892                         continue;
1893                 }
1894
1895                 if (!scsi_dev_queue_ready(q, sdev))
1896                         break;
1897
1898                 /*
1899                  * Remove the request from the request list.
1900                  */
1901                 if (!(blk_queue_tagged(q) && !blk_queue_start_tag(q, req)))
1902                         blk_start_request(req);
1903
1904                 spin_unlock_irq(q->queue_lock);
1905                 cmd = blk_mq_rq_to_pdu(req);
1906                 if (cmd != req->special) {
1907                         printk(KERN_CRIT "impossible request in %s.\n"
1908                                          "please mail a stack trace to "
1909                                          "linux-scsi@vger.kernel.org\n",
1910                                          __func__);
1911                         blk_dump_rq_flags(req, "foo");
1912                         BUG();
1913                 }
1914
1915                 /*
1916                  * We hit this when the driver is using a host wide
1917                  * tag map. For device level tag maps the queue_depth check
1918                  * in the device ready fn would prevent us from trying
1919                  * to allocate a tag. Since the map is a shared host resource
1920                  * we add the dev to the starved list so it eventually gets
1921                  * a run when a tag is freed.
1922                  */
1923                 if (blk_queue_tagged(q) && !(req->rq_flags & RQF_QUEUED)) {
1924                         spin_lock_irq(shost->host_lock);
1925                         if (list_empty(&sdev->starved_entry))
1926                                 list_add_tail(&sdev->starved_entry,
1927                                               &shost->starved_list);
1928                         spin_unlock_irq(shost->host_lock);
1929                         goto not_ready;
1930                 }
1931
1932                 if (!scsi_target_queue_ready(shost, sdev))
1933                         goto not_ready;
1934
1935                 if (!scsi_host_queue_ready(q, shost, sdev))
1936                         goto host_not_ready;
1937         
1938                 if (sdev->simple_tags)
1939                         cmd->flags |= SCMD_TAGGED;
1940                 else
1941                         cmd->flags &= ~SCMD_TAGGED;
1942
1943                 /*
1944                  * Finally, initialize any error handling parameters, and set up
1945                  * the timers for timeouts.
1946                  */
1947                 scsi_init_cmd_errh(cmd);
1948
1949                 /*
1950                  * Dispatch the command to the low-level driver.
1951                  */
1952                 cmd->scsi_done = scsi_done;
1953                 rtn = scsi_dispatch_cmd(cmd);
1954                 if (rtn) {
1955                         scsi_queue_insert(cmd, rtn);
1956                         spin_lock_irq(q->queue_lock);
1957                         goto out_delay;
1958                 }
1959                 spin_lock_irq(q->queue_lock);
1960         }
1961
1962         return;
1963
1964  host_not_ready:
1965         if (scsi_target(sdev)->can_queue > 0)
1966                 atomic_dec(&scsi_target(sdev)->target_busy);
1967  not_ready:
1968         /*
1969          * lock q, handle tag, requeue req, and decrement device_busy. We
1970          * must return with queue_lock held.
1971          *
1972          * Decrementing device_busy without checking it is OK, as all such
1973          * cases (host limits or settings) should run the queue at some
1974          * later time.
1975          */
1976         spin_lock_irq(q->queue_lock);
1977         blk_requeue_request(q, req);
1978         atomic_dec(&sdev->device_busy);
1979 out_delay:
1980         if (!atomic_read(&sdev->device_busy) && !scsi_device_blocked(sdev))
1981                 blk_delay_queue(q, SCSI_QUEUE_DELAY);
1982 }
1983
1984 static inline blk_status_t prep_to_mq(int ret)
1985 {
1986         switch (ret) {
1987         case BLKPREP_OK:
1988                 return BLK_STS_OK;
1989         case BLKPREP_DEFER:
1990                 return BLK_STS_RESOURCE;
1991         default:
1992                 return BLK_STS_IOERR;
1993         }
1994 }
1995
1996 /* Size in bytes of the sg-list stored in the scsi-mq command-private data. */
1997 static unsigned int scsi_mq_sgl_size(struct Scsi_Host *shost)
1998 {
1999         return min_t(unsigned int, shost->sg_tablesize, SG_CHUNK_SIZE) *
2000                 sizeof(struct scatterlist);
2001 }
2002
2003 static int scsi_mq_prep_fn(struct request *req)
2004 {
2005         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
2006         struct scsi_device *sdev = req->q->queuedata;
2007         struct Scsi_Host *shost = sdev->host;
2008         struct scatterlist *sg;
2009
2010         scsi_init_command(sdev, cmd);
2011
2012         req->special = cmd;
2013
2014         cmd->request = req;
2015
2016         cmd->tag = req->tag;
2017         cmd->prot_op = SCSI_PROT_NORMAL;
2018
2019         sg = (void *)cmd + sizeof(struct scsi_cmnd) + shost->hostt->cmd_size;
2020         cmd->sdb.table.sgl = sg;
2021
2022         if (scsi_host_get_prot(shost)) {
2023                 memset(cmd->prot_sdb, 0, sizeof(struct scsi_data_buffer));
2024
2025                 cmd->prot_sdb->table.sgl =
2026                         (struct scatterlist *)(cmd->prot_sdb + 1);
2027         }
2028
2029         if (blk_bidi_rq(req)) {
2030                 struct request *next_rq = req->next_rq;
2031                 struct scsi_data_buffer *bidi_sdb = blk_mq_rq_to_pdu(next_rq);
2032
2033                 memset(bidi_sdb, 0, sizeof(struct scsi_data_buffer));
2034                 bidi_sdb->table.sgl =
2035                         (struct scatterlist *)(bidi_sdb + 1);
2036
2037                 next_rq->special = bidi_sdb;
2038         }
2039
2040         blk_mq_start_request(req);
2041
2042         return scsi_setup_cmnd(sdev, req);
2043 }
2044
2045 static void scsi_mq_done(struct scsi_cmnd *cmd)
2046 {
2047         trace_scsi_dispatch_cmd_done(cmd);
2048         blk_mq_complete_request(cmd->request);
2049 }
2050
2051 static void scsi_mq_put_budget(struct blk_mq_hw_ctx *hctx)
2052 {
2053         struct request_queue *q = hctx->queue;
2054         struct scsi_device *sdev = q->queuedata;
2055
2056         atomic_dec(&sdev->device_busy);
2057         put_device(&sdev->sdev_gendev);
2058 }
2059
2060 static bool scsi_mq_get_budget(struct blk_mq_hw_ctx *hctx)
2061 {
2062         struct request_queue *q = hctx->queue;
2063         struct scsi_device *sdev = q->queuedata;
2064
2065         if (!get_device(&sdev->sdev_gendev))
2066                 goto out;
2067         if (!scsi_dev_queue_ready(q, sdev))
2068                 goto out_put_device;
2069
2070         return true;
2071
2072 out_put_device:
2073         put_device(&sdev->sdev_gendev);
2074 out:
2075         if (atomic_read(&sdev->device_busy) == 0 && !scsi_device_blocked(sdev))
2076                 blk_mq_delay_run_hw_queue(hctx, SCSI_QUEUE_DELAY);
2077         return false;
2078 }
2079
2080 static blk_status_t scsi_queue_rq(struct blk_mq_hw_ctx *hctx,
2081                          const struct blk_mq_queue_data *bd)
2082 {
2083         struct request *req = bd->rq;
2084         struct request_queue *q = req->q;
2085         struct scsi_device *sdev = q->queuedata;
2086         struct Scsi_Host *shost = sdev->host;
2087         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
2088         blk_status_t ret;
2089         int reason;
2090
2091         ret = prep_to_mq(scsi_prep_state_check(sdev, req));
2092         if (ret != BLK_STS_OK)
2093                 goto out_put_budget;
2094
2095         ret = BLK_STS_RESOURCE;
2096         if (!scsi_target_queue_ready(shost, sdev))
2097                 goto out_put_budget;
2098         if (!scsi_host_queue_ready(q, shost, sdev))
2099                 goto out_dec_target_busy;
2100
2101         if (!(req->rq_flags & RQF_DONTPREP)) {
2102                 ret = prep_to_mq(scsi_mq_prep_fn(req));
2103                 if (ret != BLK_STS_OK)
2104                         goto out_dec_host_busy;
2105                 req->rq_flags |= RQF_DONTPREP;
2106         } else {
2107                 blk_mq_start_request(req);
2108         }
2109
2110         if (sdev->simple_tags)
2111                 cmd->flags |= SCMD_TAGGED;
2112         else
2113                 cmd->flags &= ~SCMD_TAGGED;
2114
2115         scsi_init_cmd_errh(cmd);
2116         cmd->scsi_done = scsi_mq_done;
2117
2118         reason = scsi_dispatch_cmd(cmd);
2119         if (reason) {
2120                 scsi_set_blocked(cmd, reason);
2121                 ret = BLK_STS_RESOURCE;
2122                 goto out_dec_host_busy;
2123         }
2124
2125         return BLK_STS_OK;
2126
2127 out_dec_host_busy:
2128         scsi_dec_host_busy(shost);
2129 out_dec_target_busy:
2130         if (scsi_target(sdev)->can_queue > 0)
2131                 atomic_dec(&scsi_target(sdev)->target_busy);
2132 out_put_budget:
2133         scsi_mq_put_budget(hctx);
2134         switch (ret) {
2135         case BLK_STS_OK:
2136                 break;
2137         case BLK_STS_RESOURCE:
2138                 if (atomic_read(&sdev->device_busy) ||
2139                     scsi_device_blocked(sdev))
2140                         ret = BLK_STS_DEV_RESOURCE;
2141                 break;
2142         default:
2143                 /*
2144                  * Make sure to release all allocated ressources when
2145                  * we hit an error, as we will never see this command
2146                  * again.
2147                  */
2148                 if (req->rq_flags & RQF_DONTPREP)
2149                         scsi_mq_uninit_cmd(cmd);
2150                 break;
2151         }
2152         return ret;
2153 }
2154
2155 static enum blk_eh_timer_return scsi_timeout(struct request *req,
2156                 bool reserved)
2157 {
2158         if (reserved)
2159                 return BLK_EH_RESET_TIMER;
2160         return scsi_times_out(req);
2161 }
2162
2163 static int scsi_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
2164                                 unsigned int hctx_idx, unsigned int numa_node)
2165 {
2166         struct Scsi_Host *shost = set->driver_data;
2167         const bool unchecked_isa_dma = shost->unchecked_isa_dma;
2168         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
2169         struct scatterlist *sg;
2170
2171         if (unchecked_isa_dma)
2172                 cmd->flags |= SCMD_UNCHECKED_ISA_DMA;
2173         cmd->sense_buffer = scsi_alloc_sense_buffer(unchecked_isa_dma,
2174                                                     GFP_KERNEL, numa_node);
2175         if (!cmd->sense_buffer)
2176                 return -ENOMEM;
2177         cmd->req.sense = cmd->sense_buffer;
2178
2179         if (scsi_host_get_prot(shost)) {
2180                 sg = (void *)cmd + sizeof(struct scsi_cmnd) +
2181                         shost->hostt->cmd_size;
2182                 cmd->prot_sdb = (void *)sg + scsi_mq_sgl_size(shost);
2183         }
2184
2185         return 0;
2186 }
2187
2188 static void scsi_mq_exit_request(struct blk_mq_tag_set *set, struct request *rq,
2189                                  unsigned int hctx_idx)
2190 {
2191         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
2192
2193         scsi_free_sense_buffer(cmd->flags & SCMD_UNCHECKED_ISA_DMA,
2194                                cmd->sense_buffer);
2195 }
2196
2197 static int scsi_map_queues(struct blk_mq_tag_set *set)
2198 {
2199         struct Scsi_Host *shost = container_of(set, struct Scsi_Host, tag_set);
2200
2201         if (shost->hostt->map_queues)
2202                 return shost->hostt->map_queues(shost);
2203         return blk_mq_map_queues(set);
2204 }
2205
2206 void __scsi_init_queue(struct Scsi_Host *shost, struct request_queue *q)
2207 {
2208         struct device *dev = shost->dma_dev;
2209
2210         /*
2211          * this limit is imposed by hardware restrictions
2212          */
2213         blk_queue_max_segments(q, min_t(unsigned short, shost->sg_tablesize,
2214                                         SG_MAX_SEGMENTS));
2215
2216         if (scsi_host_prot_dma(shost)) {
2217                 shost->sg_prot_tablesize =
2218                         min_not_zero(shost->sg_prot_tablesize,
2219                                      (unsigned short)SCSI_MAX_PROT_SG_SEGMENTS);
2220                 BUG_ON(shost->sg_prot_tablesize < shost->sg_tablesize);
2221                 blk_queue_max_integrity_segments(q, shost->sg_prot_tablesize);
2222         }
2223
2224         blk_queue_max_hw_sectors(q, shost->max_sectors);
2225         if (shost->unchecked_isa_dma)
2226                 blk_queue_bounce_limit(q, BLK_BOUNCE_ISA);
2227         blk_queue_segment_boundary(q, shost->dma_boundary);
2228         dma_set_seg_boundary(dev, shost->dma_boundary);
2229
2230         blk_queue_max_segment_size(q, dma_get_max_seg_size(dev));
2231
2232         if (!shost->use_clustering)
2233                 q->limits.cluster = 0;
2234
2235         /*
2236          * Set a reasonable default alignment:  The larger of 32-byte (dword),
2237          * which is a common minimum for HBAs, and the minimum DMA alignment,
2238          * which is set by the platform.
2239          *
2240          * Devices that require a bigger alignment can increase it later.
2241          */
2242         blk_queue_dma_alignment(q, max(4, dma_get_cache_alignment()) - 1);
2243 }
2244 EXPORT_SYMBOL_GPL(__scsi_init_queue);
2245
2246 static int scsi_old_init_rq(struct request_queue *q, struct request *rq,
2247                             gfp_t gfp)
2248 {
2249         struct Scsi_Host *shost = q->rq_alloc_data;
2250         const bool unchecked_isa_dma = shost->unchecked_isa_dma;
2251         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
2252
2253         memset(cmd, 0, sizeof(*cmd));
2254
2255         if (unchecked_isa_dma)
2256                 cmd->flags |= SCMD_UNCHECKED_ISA_DMA;
2257         cmd->sense_buffer = scsi_alloc_sense_buffer(unchecked_isa_dma, gfp,
2258                                                     NUMA_NO_NODE);
2259         if (!cmd->sense_buffer)
2260                 goto fail;
2261         cmd->req.sense = cmd->sense_buffer;
2262
2263         if (scsi_host_get_prot(shost) >= SHOST_DIX_TYPE0_PROTECTION) {
2264                 cmd->prot_sdb = kmem_cache_zalloc(scsi_sdb_cache, gfp);
2265                 if (!cmd->prot_sdb)
2266                         goto fail_free_sense;
2267         }
2268
2269         return 0;
2270
2271 fail_free_sense:
2272         scsi_free_sense_buffer(unchecked_isa_dma, cmd->sense_buffer);
2273 fail:
2274         return -ENOMEM;
2275 }
2276
2277 static void scsi_old_exit_rq(struct request_queue *q, struct request *rq)
2278 {
2279         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
2280
2281         if (cmd->prot_sdb)
2282                 kmem_cache_free(scsi_sdb_cache, cmd->prot_sdb);
2283         scsi_free_sense_buffer(cmd->flags & SCMD_UNCHECKED_ISA_DMA,
2284                                cmd->sense_buffer);
2285 }
2286
2287 struct request_queue *scsi_old_alloc_queue(struct scsi_device *sdev)
2288 {
2289         struct Scsi_Host *shost = sdev->host;
2290         struct request_queue *q;
2291
2292         q = blk_alloc_queue_node(GFP_KERNEL, NUMA_NO_NODE, NULL);
2293         if (!q)
2294                 return NULL;
2295         q->cmd_size = sizeof(struct scsi_cmnd) + shost->hostt->cmd_size;
2296         q->rq_alloc_data = shost;
2297         q->request_fn = scsi_request_fn;
2298         q->init_rq_fn = scsi_old_init_rq;
2299         q->exit_rq_fn = scsi_old_exit_rq;
2300         q->initialize_rq_fn = scsi_initialize_rq;
2301
2302         if (blk_init_allocated_queue(q) < 0) {
2303                 blk_cleanup_queue(q);
2304                 return NULL;
2305         }
2306
2307         __scsi_init_queue(shost, q);
2308         blk_queue_flag_set(QUEUE_FLAG_SCSI_PASSTHROUGH, q);
2309         blk_queue_prep_rq(q, scsi_prep_fn);
2310         blk_queue_unprep_rq(q, scsi_unprep_fn);
2311         blk_queue_softirq_done(q, scsi_softirq_done);
2312         blk_queue_rq_timed_out(q, scsi_times_out);
2313         blk_queue_lld_busy(q, scsi_lld_busy);
2314         return q;
2315 }
2316
2317 static const struct blk_mq_ops scsi_mq_ops = {
2318         .get_budget     = scsi_mq_get_budget,
2319         .put_budget     = scsi_mq_put_budget,
2320         .queue_rq       = scsi_queue_rq,
2321         .complete       = scsi_softirq_done,
2322         .timeout        = scsi_timeout,
2323 #ifdef CONFIG_BLK_DEBUG_FS
2324         .show_rq        = scsi_show_rq,
2325 #endif
2326         .init_request   = scsi_mq_init_request,
2327         .exit_request   = scsi_mq_exit_request,
2328         .initialize_rq_fn = scsi_initialize_rq,
2329         .map_queues     = scsi_map_queues,
2330 };
2331
2332 struct request_queue *scsi_mq_alloc_queue(struct scsi_device *sdev)
2333 {
2334         sdev->request_queue = blk_mq_init_queue(&sdev->host->tag_set);
2335         if (IS_ERR(sdev->request_queue))
2336                 return NULL;
2337
2338         sdev->request_queue->queuedata = sdev;
2339         __scsi_init_queue(sdev->host, sdev->request_queue);
2340         blk_queue_flag_set(QUEUE_FLAG_SCSI_PASSTHROUGH, sdev->request_queue);
2341         return sdev->request_queue;
2342 }
2343
2344 int scsi_mq_setup_tags(struct Scsi_Host *shost)
2345 {
2346         unsigned int cmd_size, sgl_size;
2347
2348         sgl_size = scsi_mq_sgl_size(shost);
2349         cmd_size = sizeof(struct scsi_cmnd) + shost->hostt->cmd_size + sgl_size;
2350         if (scsi_host_get_prot(shost))
2351                 cmd_size += sizeof(struct scsi_data_buffer) + sgl_size;
2352
2353         memset(&shost->tag_set, 0, sizeof(shost->tag_set));
2354         shost->tag_set.ops = &scsi_mq_ops;
2355         shost->tag_set.nr_hw_queues = shost->nr_hw_queues ? : 1;
2356         shost->tag_set.queue_depth = shost->can_queue;
2357         shost->tag_set.cmd_size = cmd_size;
2358         shost->tag_set.numa_node = NUMA_NO_NODE;
2359         shost->tag_set.flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_SG_MERGE;
2360         shost->tag_set.flags |=
2361                 BLK_ALLOC_POLICY_TO_MQ_FLAG(shost->hostt->tag_alloc_policy);
2362         shost->tag_set.driver_data = shost;
2363
2364         return blk_mq_alloc_tag_set(&shost->tag_set);
2365 }
2366
2367 void scsi_mq_destroy_tags(struct Scsi_Host *shost)
2368 {
2369         blk_mq_free_tag_set(&shost->tag_set);
2370 }
2371
2372 /**
2373  * scsi_device_from_queue - return sdev associated with a request_queue
2374  * @q: The request queue to return the sdev from
2375  *
2376  * Return the sdev associated with a request queue or NULL if the
2377  * request_queue does not reference a SCSI device.
2378  */
2379 struct scsi_device *scsi_device_from_queue(struct request_queue *q)
2380 {
2381         struct scsi_device *sdev = NULL;
2382
2383         if (q->mq_ops) {
2384                 if (q->mq_ops == &scsi_mq_ops)
2385                         sdev = q->queuedata;
2386         } else if (q->request_fn == scsi_request_fn)
2387                 sdev = q->queuedata;
2388         if (!sdev || !get_device(&sdev->sdev_gendev))
2389                 sdev = NULL;
2390
2391         return sdev;
2392 }
2393 EXPORT_SYMBOL_GPL(scsi_device_from_queue);
2394
2395 /*
2396  * Function:    scsi_block_requests()
2397  *
2398  * Purpose:     Utility function used by low-level drivers to prevent further
2399  *              commands from being queued to the device.
2400  *
2401  * Arguments:   shost       - Host in question
2402  *
2403  * Returns:     Nothing
2404  *
2405  * Lock status: No locks are assumed held.
2406  *
2407  * Notes:       There is no timer nor any other means by which the requests
2408  *              get unblocked other than the low-level driver calling
2409  *              scsi_unblock_requests().
2410  */
2411 void scsi_block_requests(struct Scsi_Host *shost)
2412 {
2413         shost->host_self_blocked = 1;
2414 }
2415 EXPORT_SYMBOL(scsi_block_requests);
2416
2417 /*
2418  * Function:    scsi_unblock_requests()
2419  *
2420  * Purpose:     Utility function used by low-level drivers to allow further
2421  *              commands from being queued to the device.
2422  *
2423  * Arguments:   shost       - Host in question
2424  *
2425  * Returns:     Nothing
2426  *
2427  * Lock status: No locks are assumed held.
2428  *
2429  * Notes:       There is no timer nor any other means by which the requests
2430  *              get unblocked other than the low-level driver calling
2431  *              scsi_unblock_requests().
2432  *
2433  *              This is done as an API function so that changes to the
2434  *              internals of the scsi mid-layer won't require wholesale
2435  *              changes to drivers that use this feature.
2436  */
2437 void scsi_unblock_requests(struct Scsi_Host *shost)
2438 {
2439         shost->host_self_blocked = 0;
2440         scsi_run_host_queues(shost);
2441 }
2442 EXPORT_SYMBOL(scsi_unblock_requests);
2443
2444 int __init scsi_init_queue(void)
2445 {
2446         scsi_sdb_cache = kmem_cache_create("scsi_data_buffer",
2447                                            sizeof(struct scsi_data_buffer),
2448                                            0, 0, NULL);
2449         if (!scsi_sdb_cache) {
2450                 printk(KERN_ERR "SCSI: can't init scsi sdb cache\n");
2451                 return -ENOMEM;
2452         }
2453
2454         return 0;
2455 }
2456
2457 void scsi_exit_queue(void)
2458 {
2459         kmem_cache_destroy(scsi_sense_cache);
2460         kmem_cache_destroy(scsi_sense_isadma_cache);
2461         kmem_cache_destroy(scsi_sdb_cache);
2462 }
2463
2464 /**
2465  *      scsi_mode_select - issue a mode select
2466  *      @sdev:  SCSI device to be queried
2467  *      @pf:    Page format bit (1 == standard, 0 == vendor specific)
2468  *      @sp:    Save page bit (0 == don't save, 1 == save)
2469  *      @modepage: mode page being requested
2470  *      @buffer: request buffer (may not be smaller than eight bytes)
2471  *      @len:   length of request buffer.
2472  *      @timeout: command timeout
2473  *      @retries: number of retries before failing
2474  *      @data: returns a structure abstracting the mode header data
2475  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
2476  *              must be SCSI_SENSE_BUFFERSIZE big.
2477  *
2478  *      Returns zero if successful; negative error number or scsi
2479  *      status on error
2480  *
2481  */
2482 int
2483 scsi_mode_select(struct scsi_device *sdev, int pf, int sp, int modepage,
2484                  unsigned char *buffer, int len, int timeout, int retries,
2485                  struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
2486 {
2487         unsigned char cmd[10];
2488         unsigned char *real_buffer;
2489         int ret;
2490
2491         memset(cmd, 0, sizeof(cmd));
2492         cmd[1] = (pf ? 0x10 : 0) | (sp ? 0x01 : 0);
2493
2494         if (sdev->use_10_for_ms) {
2495                 if (len > 65535)
2496                         return -EINVAL;
2497                 real_buffer = kmalloc(8 + len, GFP_KERNEL);
2498                 if (!real_buffer)
2499                         return -ENOMEM;
2500                 memcpy(real_buffer + 8, buffer, len);
2501                 len += 8;
2502                 real_buffer[0] = 0;
2503                 real_buffer[1] = 0;
2504                 real_buffer[2] = data->medium_type;
2505                 real_buffer[3] = data->device_specific;
2506                 real_buffer[4] = data->longlba ? 0x01 : 0;
2507                 real_buffer[5] = 0;
2508                 real_buffer[6] = data->block_descriptor_length >> 8;
2509                 real_buffer[7] = data->block_descriptor_length;
2510
2511                 cmd[0] = MODE_SELECT_10;
2512                 cmd[7] = len >> 8;
2513                 cmd[8] = len;
2514         } else {
2515                 if (len > 255 || data->block_descriptor_length > 255 ||
2516                     data->longlba)
2517                         return -EINVAL;
2518
2519                 real_buffer = kmalloc(4 + len, GFP_KERNEL);
2520                 if (!real_buffer)
2521                         return -ENOMEM;
2522                 memcpy(real_buffer + 4, buffer, len);
2523                 len += 4;
2524                 real_buffer[0] = 0;
2525                 real_buffer[1] = data->medium_type;
2526                 real_buffer[2] = data->device_specific;
2527                 real_buffer[3] = data->block_descriptor_length;
2528                 
2529
2530                 cmd[0] = MODE_SELECT;
2531                 cmd[4] = len;
2532         }
2533
2534         ret = scsi_execute_req(sdev, cmd, DMA_TO_DEVICE, real_buffer, len,
2535                                sshdr, timeout, retries, NULL);
2536         kfree(real_buffer);
2537         return ret;
2538 }
2539 EXPORT_SYMBOL_GPL(scsi_mode_select);
2540
2541 /**
2542  *      scsi_mode_sense - issue a mode sense, falling back from 10 to six bytes if necessary.
2543  *      @sdev:  SCSI device to be queried
2544  *      @dbd:   set if mode sense will allow block descriptors to be returned
2545  *      @modepage: mode page being requested
2546  *      @buffer: request buffer (may not be smaller than eight bytes)
2547  *      @len:   length of request buffer.
2548  *      @timeout: command timeout
2549  *      @retries: number of retries before failing
2550  *      @data: returns a structure abstracting the mode header data
2551  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
2552  *              must be SCSI_SENSE_BUFFERSIZE big.
2553  *
2554  *      Returns zero if unsuccessful, or the header offset (either 4
2555  *      or 8 depending on whether a six or ten byte command was
2556  *      issued) if successful.
2557  */
2558 int
2559 scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
2560                   unsigned char *buffer, int len, int timeout, int retries,
2561                   struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
2562 {
2563         unsigned char cmd[12];
2564         int use_10_for_ms;
2565         int header_length;
2566         int result, retry_count = retries;
2567         struct scsi_sense_hdr my_sshdr;
2568
2569         memset(data, 0, sizeof(*data));
2570         memset(&cmd[0], 0, 12);
2571         cmd[1] = dbd & 0x18;    /* allows DBD and LLBA bits */
2572         cmd[2] = modepage;
2573
2574         /* caller might not be interested in sense, but we need it */
2575         if (!sshdr)
2576                 sshdr = &my_sshdr;
2577
2578  retry:
2579         use_10_for_ms = sdev->use_10_for_ms;
2580
2581         if (use_10_for_ms) {
2582                 if (len < 8)
2583                         len = 8;
2584
2585                 cmd[0] = MODE_SENSE_10;
2586                 cmd[8] = len;
2587                 header_length = 8;
2588         } else {
2589                 if (len < 4)
2590                         len = 4;
2591
2592                 cmd[0] = MODE_SENSE;
2593                 cmd[4] = len;
2594                 header_length = 4;
2595         }
2596
2597         memset(buffer, 0, len);
2598
2599         result = scsi_execute_req(sdev, cmd, DMA_FROM_DEVICE, buffer, len,
2600                                   sshdr, timeout, retries, NULL);
2601
2602         /* This code looks awful: what it's doing is making sure an
2603          * ILLEGAL REQUEST sense return identifies the actual command
2604          * byte as the problem.  MODE_SENSE commands can return
2605          * ILLEGAL REQUEST if the code page isn't supported */
2606
2607         if (use_10_for_ms && !scsi_status_is_good(result) &&
2608             driver_byte(result) == DRIVER_SENSE) {
2609                 if (scsi_sense_valid(sshdr)) {
2610                         if ((sshdr->sense_key == ILLEGAL_REQUEST) &&
2611                             (sshdr->asc == 0x20) && (sshdr->ascq == 0)) {
2612                                 /* 
2613                                  * Invalid command operation code
2614                                  */
2615                                 sdev->use_10_for_ms = 0;
2616                                 goto retry;
2617                         }
2618                 }
2619         }
2620
2621         if(scsi_status_is_good(result)) {
2622                 if (unlikely(buffer[0] == 0x86 && buffer[1] == 0x0b &&
2623                              (modepage == 6 || modepage == 8))) {
2624                         /* Initio breakage? */
2625                         header_length = 0;
2626                         data->length = 13;
2627                         data->medium_type = 0;
2628                         data->device_specific = 0;
2629                         data->longlba = 0;
2630                         data->block_descriptor_length = 0;
2631                 } else if(use_10_for_ms) {
2632                         data->length = buffer[0]*256 + buffer[1] + 2;
2633                         data->medium_type = buffer[2];
2634                         data->device_specific = buffer[3];
2635                         data->longlba = buffer[4] & 0x01;
2636                         data->block_descriptor_length = buffer[6]*256
2637                                 + buffer[7];
2638                 } else {
2639                         data->length = buffer[0] + 1;
2640                         data->medium_type = buffer[1];
2641                         data->device_specific = buffer[2];
2642                         data->block_descriptor_length = buffer[3];
2643                 }
2644                 data->header_length = header_length;
2645         } else if ((status_byte(result) == CHECK_CONDITION) &&
2646                    scsi_sense_valid(sshdr) &&
2647                    sshdr->sense_key == UNIT_ATTENTION && retry_count) {
2648                 retry_count--;
2649                 goto retry;
2650         }
2651
2652         return result;
2653 }
2654 EXPORT_SYMBOL(scsi_mode_sense);
2655
2656 /**
2657  *      scsi_test_unit_ready - test if unit is ready
2658  *      @sdev:  scsi device to change the state of.
2659  *      @timeout: command timeout
2660  *      @retries: number of retries before failing
2661  *      @sshdr: outpout pointer for decoded sense information.
2662  *
2663  *      Returns zero if unsuccessful or an error if TUR failed.  For
2664  *      removable media, UNIT_ATTENTION sets ->changed flag.
2665  **/
2666 int
2667 scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries,
2668                      struct scsi_sense_hdr *sshdr)
2669 {
2670         char cmd[] = {
2671                 TEST_UNIT_READY, 0, 0, 0, 0, 0,
2672         };
2673         int result;
2674
2675         /* try to eat the UNIT_ATTENTION if there are enough retries */
2676         do {
2677                 result = scsi_execute_req(sdev, cmd, DMA_NONE, NULL, 0, sshdr,
2678                                           timeout, 1, NULL);
2679                 if (sdev->removable && scsi_sense_valid(sshdr) &&
2680                     sshdr->sense_key == UNIT_ATTENTION)
2681                         sdev->changed = 1;
2682         } while (scsi_sense_valid(sshdr) &&
2683                  sshdr->sense_key == UNIT_ATTENTION && --retries);
2684
2685         return result;
2686 }
2687 EXPORT_SYMBOL(scsi_test_unit_ready);
2688
2689 /**
2690  *      scsi_device_set_state - Take the given device through the device state model.
2691  *      @sdev:  scsi device to change the state of.
2692  *      @state: state to change to.
2693  *
2694  *      Returns zero if successful or an error if the requested
2695  *      transition is illegal.
2696  */
2697 int
2698 scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
2699 {
2700         enum scsi_device_state oldstate = sdev->sdev_state;
2701
2702         if (state == oldstate)
2703                 return 0;
2704
2705         switch (state) {
2706         case SDEV_CREATED:
2707                 switch (oldstate) {
2708                 case SDEV_CREATED_BLOCK:
2709                         break;
2710                 default:
2711                         goto illegal;
2712                 }
2713                 break;
2714                         
2715         case SDEV_RUNNING:
2716                 switch (oldstate) {
2717                 case SDEV_CREATED:
2718                 case SDEV_OFFLINE:
2719                 case SDEV_TRANSPORT_OFFLINE:
2720                 case SDEV_QUIESCE:
2721                 case SDEV_BLOCK:
2722                         break;
2723                 default:
2724                         goto illegal;
2725                 }
2726                 break;
2727
2728         case SDEV_QUIESCE:
2729                 switch (oldstate) {
2730                 case SDEV_RUNNING:
2731                 case SDEV_OFFLINE:
2732                 case SDEV_TRANSPORT_OFFLINE:
2733                         break;
2734                 default:
2735                         goto illegal;
2736                 }
2737                 break;
2738
2739         case SDEV_OFFLINE:
2740         case SDEV_TRANSPORT_OFFLINE:
2741                 switch (oldstate) {
2742                 case SDEV_CREATED:
2743                 case SDEV_RUNNING:
2744                 case SDEV_QUIESCE:
2745                 case SDEV_BLOCK:
2746                         break;
2747                 default:
2748                         goto illegal;
2749                 }
2750                 break;
2751
2752         case SDEV_BLOCK:
2753                 switch (oldstate) {
2754                 case SDEV_RUNNING:
2755                 case SDEV_CREATED_BLOCK:
2756                 case SDEV_OFFLINE:
2757                         break;
2758                 default:
2759                         goto illegal;
2760                 }
2761                 break;
2762
2763         case SDEV_CREATED_BLOCK:
2764                 switch (oldstate) {
2765                 case SDEV_CREATED:
2766                         break;
2767                 default:
2768                         goto illegal;
2769                 }
2770                 break;
2771
2772         case SDEV_CANCEL:
2773                 switch (oldstate) {
2774                 case SDEV_CREATED:
2775                 case SDEV_RUNNING:
2776                 case SDEV_QUIESCE:
2777                 case SDEV_OFFLINE:
2778                 case SDEV_TRANSPORT_OFFLINE:
2779                         break;
2780                 default:
2781                         goto illegal;
2782                 }
2783                 break;
2784
2785         case SDEV_DEL:
2786                 switch (oldstate) {
2787                 case SDEV_CREATED:
2788                 case SDEV_RUNNING:
2789                 case SDEV_OFFLINE:
2790                 case SDEV_TRANSPORT_OFFLINE:
2791                 case SDEV_CANCEL:
2792                 case SDEV_BLOCK:
2793                 case SDEV_CREATED_BLOCK:
2794                         break;
2795                 default:
2796                         goto illegal;
2797                 }
2798                 break;
2799
2800         }
2801         sdev->sdev_state = state;
2802         return 0;
2803
2804  illegal:
2805         SCSI_LOG_ERROR_RECOVERY(1,
2806                                 sdev_printk(KERN_ERR, sdev,
2807                                             "Illegal state transition %s->%s",
2808                                             scsi_device_state_name(oldstate),
2809                                             scsi_device_state_name(state))
2810                                 );
2811         return -EINVAL;
2812 }
2813 EXPORT_SYMBOL(scsi_device_set_state);
2814
2815 /**
2816  *      sdev_evt_emit - emit a single SCSI device uevent
2817  *      @sdev: associated SCSI device
2818  *      @evt: event to emit
2819  *
2820  *      Send a single uevent (scsi_event) to the associated scsi_device.
2821  */
2822 static void scsi_evt_emit(struct scsi_device *sdev, struct scsi_event *evt)
2823 {
2824         int idx = 0;
2825         char *envp[3];
2826
2827         switch (evt->evt_type) {
2828         case SDEV_EVT_MEDIA_CHANGE:
2829                 envp[idx++] = "SDEV_MEDIA_CHANGE=1";
2830                 break;
2831         case SDEV_EVT_INQUIRY_CHANGE_REPORTED:
2832                 scsi_rescan_device(&sdev->sdev_gendev);
2833                 envp[idx++] = "SDEV_UA=INQUIRY_DATA_HAS_CHANGED";
2834                 break;
2835         case SDEV_EVT_CAPACITY_CHANGE_REPORTED:
2836                 envp[idx++] = "SDEV_UA=CAPACITY_DATA_HAS_CHANGED";
2837                 break;
2838         case SDEV_EVT_SOFT_THRESHOLD_REACHED_REPORTED:
2839                envp[idx++] = "SDEV_UA=THIN_PROVISIONING_SOFT_THRESHOLD_REACHED";
2840                 break;
2841         case SDEV_EVT_MODE_PARAMETER_CHANGE_REPORTED:
2842                 envp[idx++] = "SDEV_UA=MODE_PARAMETERS_CHANGED";
2843                 break;
2844         case SDEV_EVT_LUN_CHANGE_REPORTED:
2845                 envp[idx++] = "SDEV_UA=REPORTED_LUNS_DATA_HAS_CHANGED";
2846                 break;
2847         case SDEV_EVT_ALUA_STATE_CHANGE_REPORTED:
2848                 envp[idx++] = "SDEV_UA=ASYMMETRIC_ACCESS_STATE_CHANGED";
2849                 break;
2850         case SDEV_EVT_POWER_ON_RESET_OCCURRED:
2851                 envp[idx++] = "SDEV_UA=POWER_ON_RESET_OCCURRED";
2852                 break;
2853         default:
2854                 /* do nothing */
2855                 break;
2856         }
2857
2858         envp[idx++] = NULL;
2859
2860         kobject_uevent_env(&sdev->sdev_gendev.kobj, KOBJ_CHANGE, envp);
2861 }
2862
2863 /**
2864  *      sdev_evt_thread - send a uevent for each scsi event
2865  *      @work: work struct for scsi_device
2866  *
2867  *      Dispatch queued events to their associated scsi_device kobjects
2868  *      as uevents.
2869  */
2870 void scsi_evt_thread(struct work_struct *work)
2871 {
2872         struct scsi_device *sdev;
2873         enum scsi_device_event evt_type;
2874         LIST_HEAD(event_list);
2875
2876         sdev = container_of(work, struct scsi_device, event_work);
2877
2878         for (evt_type = SDEV_EVT_FIRST; evt_type <= SDEV_EVT_LAST; evt_type++)
2879                 if (test_and_clear_bit(evt_type, sdev->pending_events))
2880                         sdev_evt_send_simple(sdev, evt_type, GFP_KERNEL);
2881
2882         while (1) {
2883                 struct scsi_event *evt;
2884                 struct list_head *this, *tmp;
2885                 unsigned long flags;
2886
2887                 spin_lock_irqsave(&sdev->list_lock, flags);
2888                 list_splice_init(&sdev->event_list, &event_list);
2889                 spin_unlock_irqrestore(&sdev->list_lock, flags);
2890
2891                 if (list_empty(&event_list))
2892                         break;
2893
2894                 list_for_each_safe(this, tmp, &event_list) {
2895                         evt = list_entry(this, struct scsi_event, node);
2896                         list_del(&evt->node);
2897                         scsi_evt_emit(sdev, evt);
2898                         kfree(evt);
2899                 }
2900         }
2901 }
2902
2903 /**
2904  *      sdev_evt_send - send asserted event to uevent thread
2905  *      @sdev: scsi_device event occurred on
2906  *      @evt: event to send
2907  *
2908  *      Assert scsi device event asynchronously.
2909  */
2910 void sdev_evt_send(struct scsi_device *sdev, struct scsi_event *evt)
2911 {
2912         unsigned long flags;
2913
2914 #if 0
2915         /* FIXME: currently this check eliminates all media change events
2916          * for polled devices.  Need to update to discriminate between AN
2917          * and polled events */
2918         if (!test_bit(evt->evt_type, sdev->supported_events)) {
2919                 kfree(evt);
2920                 return;
2921         }
2922 #endif
2923
2924         spin_lock_irqsave(&sdev->list_lock, flags);
2925         list_add_tail(&evt->node, &sdev->event_list);
2926         schedule_work(&sdev->event_work);
2927         spin_unlock_irqrestore(&sdev->list_lock, flags);
2928 }
2929 EXPORT_SYMBOL_GPL(sdev_evt_send);
2930
2931 /**
2932  *      sdev_evt_alloc - allocate a new scsi event
2933  *      @evt_type: type of event to allocate
2934  *      @gfpflags: GFP flags for allocation
2935  *
2936  *      Allocates and returns a new scsi_event.
2937  */
2938 struct scsi_event *sdev_evt_alloc(enum scsi_device_event evt_type,
2939                                   gfp_t gfpflags)
2940 {
2941         struct scsi_event *evt = kzalloc(sizeof(struct scsi_event), gfpflags);
2942         if (!evt)
2943                 return NULL;
2944
2945         evt->evt_type = evt_type;
2946         INIT_LIST_HEAD(&evt->node);
2947
2948         /* evt_type-specific initialization, if any */
2949         switch (evt_type) {
2950         case SDEV_EVT_MEDIA_CHANGE:
2951         case SDEV_EVT_INQUIRY_CHANGE_REPORTED:
2952         case SDEV_EVT_CAPACITY_CHANGE_REPORTED:
2953         case SDEV_EVT_SOFT_THRESHOLD_REACHED_REPORTED:
2954         case SDEV_EVT_MODE_PARAMETER_CHANGE_REPORTED:
2955         case SDEV_EVT_LUN_CHANGE_REPORTED:
2956         case SDEV_EVT_ALUA_STATE_CHANGE_REPORTED:
2957         case SDEV_EVT_POWER_ON_RESET_OCCURRED:
2958         default:
2959                 /* do nothing */
2960                 break;
2961         }
2962
2963         return evt;
2964 }
2965 EXPORT_SYMBOL_GPL(sdev_evt_alloc);
2966
2967 /**
2968  *      sdev_evt_send_simple - send asserted event to uevent thread
2969  *      @sdev: scsi_device event occurred on
2970  *      @evt_type: type of event to send
2971  *      @gfpflags: GFP flags for allocation
2972  *
2973  *      Assert scsi device event asynchronously, given an event type.
2974  */
2975 void sdev_evt_send_simple(struct scsi_device *sdev,
2976                           enum scsi_device_event evt_type, gfp_t gfpflags)
2977 {
2978         struct scsi_event *evt = sdev_evt_alloc(evt_type, gfpflags);
2979         if (!evt) {
2980                 sdev_printk(KERN_ERR, sdev, "event %d eaten due to OOM\n",
2981                             evt_type);
2982                 return;
2983         }
2984
2985         sdev_evt_send(sdev, evt);
2986 }
2987 EXPORT_SYMBOL_GPL(sdev_evt_send_simple);
2988
2989 /**
2990  * scsi_request_fn_active() - number of kernel threads inside scsi_request_fn()
2991  * @sdev: SCSI device to count the number of scsi_request_fn() callers for.
2992  */
2993 static int scsi_request_fn_active(struct scsi_device *sdev)
2994 {
2995         struct request_queue *q = sdev->request_queue;
2996         int request_fn_active;
2997
2998         WARN_ON_ONCE(sdev->host->use_blk_mq);
2999
3000         spin_lock_irq(q->queue_lock);
3001         request_fn_active = q->request_fn_active;
3002         spin_unlock_irq(q->queue_lock);
3003
3004         return request_fn_active;
3005 }
3006
3007 /**
3008  * scsi_wait_for_queuecommand() - wait for ongoing queuecommand() calls
3009  * @sdev: SCSI device pointer.
3010  *
3011  * Wait until the ongoing shost->hostt->queuecommand() calls that are
3012  * invoked from scsi_request_fn() have finished.
3013  */
3014 static void scsi_wait_for_queuecommand(struct scsi_device *sdev)
3015 {
3016         WARN_ON_ONCE(sdev->host->use_blk_mq);
3017
3018         while (scsi_request_fn_active(sdev))
3019                 msleep(20);
3020 }
3021
3022 /**
3023  *      scsi_device_quiesce - Block user issued commands.
3024  *      @sdev:  scsi device to quiesce.
3025  *
3026  *      This works by trying to transition to the SDEV_QUIESCE state
3027  *      (which must be a legal transition).  When the device is in this
3028  *      state, only special requests will be accepted, all others will
3029  *      be deferred.  Since special requests may also be requeued requests,
3030  *      a successful return doesn't guarantee the device will be 
3031  *      totally quiescent.
3032  *
3033  *      Must be called with user context, may sleep.
3034  *
3035  *      Returns zero if unsuccessful or an error if not.
3036  */
3037 int
3038 scsi_device_quiesce(struct scsi_device *sdev)
3039 {
3040         struct request_queue *q = sdev->request_queue;
3041         int err;
3042
3043         /*
3044          * It is allowed to call scsi_device_quiesce() multiple times from
3045          * the same context but concurrent scsi_device_quiesce() calls are
3046          * not allowed.
3047          */
3048         WARN_ON_ONCE(sdev->quiesced_by && sdev->quiesced_by != current);
3049
3050         if (sdev->quiesced_by == current)
3051                 return 0;
3052
3053         blk_set_pm_only(q);
3054
3055         blk_mq_freeze_queue(q);
3056         /*
3057          * Ensure that the effect of blk_set_pm_only() will be visible
3058          * for percpu_ref_tryget() callers that occur after the queue
3059          * unfreeze even if the queue was already frozen before this function
3060          * was called. See also https://lwn.net/Articles/573497/.
3061          */
3062         synchronize_rcu();
3063         blk_mq_unfreeze_queue(q);
3064
3065         mutex_lock(&sdev->state_mutex);
3066         err = scsi_device_set_state(sdev, SDEV_QUIESCE);
3067         if (err == 0)
3068                 sdev->quiesced_by = current;
3069         else
3070                 blk_clear_pm_only(q);
3071         mutex_unlock(&sdev->state_mutex);
3072
3073         return err;
3074 }
3075 EXPORT_SYMBOL(scsi_device_quiesce);
3076
3077 /**
3078  *      scsi_device_resume - Restart user issued commands to a quiesced device.
3079  *      @sdev:  scsi device to resume.
3080  *
3081  *      Moves the device from quiesced back to running and restarts the
3082  *      queues.
3083  *
3084  *      Must be called with user context, may sleep.
3085  */
3086 void scsi_device_resume(struct scsi_device *sdev)
3087 {
3088         /* check if the device state was mutated prior to resume, and if
3089          * so assume the state is being managed elsewhere (for example
3090          * device deleted during suspend)
3091          */
3092         mutex_lock(&sdev->state_mutex);
3093         WARN_ON_ONCE(!sdev->quiesced_by);
3094         sdev->quiesced_by = NULL;
3095         blk_clear_pm_only(sdev->request_queue);
3096         if (sdev->sdev_state == SDEV_QUIESCE)
3097                 scsi_device_set_state(sdev, SDEV_RUNNING);
3098         mutex_unlock(&sdev->state_mutex);
3099 }
3100 EXPORT_SYMBOL(scsi_device_resume);
3101
3102 static void
3103 device_quiesce_fn(struct scsi_device *sdev, void *data)
3104 {
3105         scsi_device_quiesce(sdev);
3106 }
3107
3108 void
3109 scsi_target_quiesce(struct scsi_target *starget)
3110 {
3111         starget_for_each_device(starget, NULL, device_quiesce_fn);
3112 }
3113 EXPORT_SYMBOL(scsi_target_quiesce);
3114
3115 static void
3116 device_resume_fn(struct scsi_device *sdev, void *data)
3117 {
3118         scsi_device_resume(sdev);
3119 }
3120
3121 void
3122 scsi_target_resume(struct scsi_target *starget)
3123 {
3124         starget_for_each_device(starget, NULL, device_resume_fn);
3125 }
3126 EXPORT_SYMBOL(scsi_target_resume);
3127
3128 /**
3129  * scsi_internal_device_block_nowait - try to transition to the SDEV_BLOCK state
3130  * @sdev: device to block
3131  *
3132  * Pause SCSI command processing on the specified device. Does not sleep.
3133  *
3134  * Returns zero if successful or a negative error code upon failure.
3135  *
3136  * Notes:
3137  * This routine transitions the device to the SDEV_BLOCK state (which must be
3138  * a legal transition). When the device is in this state, command processing
3139  * is paused until the device leaves the SDEV_BLOCK state. See also
3140  * scsi_internal_device_unblock_nowait().
3141  */
3142 int scsi_internal_device_block_nowait(struct scsi_device *sdev)
3143 {
3144         struct request_queue *q = sdev->request_queue;
3145         unsigned long flags;
3146         int err = 0;
3147
3148         err = scsi_device_set_state(sdev, SDEV_BLOCK);
3149         if (err) {
3150                 err = scsi_device_set_state(sdev, SDEV_CREATED_BLOCK);
3151
3152                 if (err)
3153                         return err;
3154         }
3155
3156         /* 
3157          * The device has transitioned to SDEV_BLOCK.  Stop the
3158          * block layer from calling the midlayer with this device's
3159          * request queue. 
3160          */
3161         if (q->mq_ops) {
3162                 blk_mq_quiesce_queue_nowait(q);
3163         } else {
3164                 spin_lock_irqsave(q->queue_lock, flags);
3165                 blk_stop_queue(q);
3166                 spin_unlock_irqrestore(q->queue_lock, flags);
3167         }
3168
3169         return 0;
3170 }
3171 EXPORT_SYMBOL_GPL(scsi_internal_device_block_nowait);
3172
3173 /**
3174  * scsi_internal_device_block - try to transition to the SDEV_BLOCK state
3175  * @sdev: device to block
3176  *
3177  * Pause SCSI command processing on the specified device and wait until all
3178  * ongoing scsi_request_fn() / scsi_queue_rq() calls have finished. May sleep.
3179  *
3180  * Returns zero if successful or a negative error code upon failure.
3181  *
3182  * Note:
3183  * This routine transitions the device to the SDEV_BLOCK state (which must be
3184  * a legal transition). When the device is in this state, command processing
3185  * is paused until the device leaves the SDEV_BLOCK state. See also
3186  * scsi_internal_device_unblock().
3187  *
3188  * To do: avoid that scsi_send_eh_cmnd() calls queuecommand() after
3189  * scsi_internal_device_block() has blocked a SCSI device and also
3190  * remove the rport mutex lock and unlock calls from srp_queuecommand().
3191  */
3192 static int scsi_internal_device_block(struct scsi_device *sdev)
3193 {
3194         struct request_queue *q = sdev->request_queue;
3195         int err;
3196
3197         mutex_lock(&sdev->state_mutex);
3198         err = scsi_internal_device_block_nowait(sdev);
3199         if (err == 0) {
3200                 if (q->mq_ops)
3201                         blk_mq_quiesce_queue(q);
3202                 else
3203                         scsi_wait_for_queuecommand(sdev);
3204         }
3205         mutex_unlock(&sdev->state_mutex);
3206
3207         return err;
3208 }
3209  
3210 void scsi_start_queue(struct scsi_device *sdev)
3211 {
3212         struct request_queue *q = sdev->request_queue;
3213         unsigned long flags;
3214
3215         if (q->mq_ops) {
3216                 blk_mq_unquiesce_queue(q);
3217         } else {
3218                 spin_lock_irqsave(q->queue_lock, flags);
3219                 blk_start_queue(q);
3220                 spin_unlock_irqrestore(q->queue_lock, flags);
3221         }
3222 }
3223
3224 /**
3225  * scsi_internal_device_unblock_nowait - resume a device after a block request
3226  * @sdev:       device to resume
3227  * @new_state:  state to set the device to after unblocking
3228  *
3229  * Restart the device queue for a previously suspended SCSI device. Does not
3230  * sleep.
3231  *
3232  * Returns zero if successful or a negative error code upon failure.
3233  *
3234  * Notes:
3235  * This routine transitions the device to the SDEV_RUNNING state or to one of
3236  * the offline states (which must be a legal transition) allowing the midlayer
3237  * to goose the queue for this device.
3238  */
3239 int scsi_internal_device_unblock_nowait(struct scsi_device *sdev,
3240                                         enum scsi_device_state new_state)
3241 {
3242         /*
3243          * Try to transition the scsi device to SDEV_RUNNING or one of the
3244          * offlined states and goose the device queue if successful.
3245          */
3246         switch (sdev->sdev_state) {
3247         case SDEV_BLOCK:
3248         case SDEV_TRANSPORT_OFFLINE:
3249                 sdev->sdev_state = new_state;
3250                 break;
3251         case SDEV_CREATED_BLOCK:
3252                 if (new_state == SDEV_TRANSPORT_OFFLINE ||
3253                     new_state == SDEV_OFFLINE)
3254                         sdev->sdev_state = new_state;
3255                 else
3256                         sdev->sdev_state = SDEV_CREATED;
3257                 break;
3258         case SDEV_CANCEL:
3259         case SDEV_OFFLINE:
3260                 break;
3261         default:
3262                 return -EINVAL;
3263         }
3264         scsi_start_queue(sdev);
3265
3266         return 0;
3267 }
3268 EXPORT_SYMBOL_GPL(scsi_internal_device_unblock_nowait);
3269
3270 /**
3271  * scsi_internal_device_unblock - resume a device after a block request
3272  * @sdev:       device to resume
3273  * @new_state:  state to set the device to after unblocking
3274  *
3275  * Restart the device queue for a previously suspended SCSI device. May sleep.
3276  *
3277  * Returns zero if successful or a negative error code upon failure.
3278  *
3279  * Notes:
3280  * This routine transitions the device to the SDEV_RUNNING state or to one of
3281  * the offline states (which must be a legal transition) allowing the midlayer
3282  * to goose the queue for this device.
3283  */
3284 static int scsi_internal_device_unblock(struct scsi_device *sdev,
3285                                         enum scsi_device_state new_state)
3286 {
3287         int ret;
3288
3289         mutex_lock(&sdev->state_mutex);
3290         ret = scsi_internal_device_unblock_nowait(sdev, new_state);
3291         mutex_unlock(&sdev->state_mutex);
3292
3293         return ret;
3294 }
3295
3296 static void
3297 device_block(struct scsi_device *sdev, void *data)
3298 {
3299         scsi_internal_device_block(sdev);
3300 }
3301
3302 static int
3303 target_block(struct device *dev, void *data)
3304 {
3305         if (scsi_is_target_device(dev))
3306                 starget_for_each_device(to_scsi_target(dev), NULL,
3307                                         device_block);
3308         return 0;
3309 }
3310
3311 void
3312 scsi_target_block(struct device *dev)
3313 {
3314         if (scsi_is_target_device(dev))
3315                 starget_for_each_device(to_scsi_target(dev), NULL,
3316                                         device_block);
3317         else
3318                 device_for_each_child(dev, NULL, target_block);
3319 }
3320 EXPORT_SYMBOL_GPL(scsi_target_block);
3321
3322 static void
3323 device_unblock(struct scsi_device *sdev, void *data)
3324 {
3325         scsi_internal_device_unblock(sdev, *(enum scsi_device_state *)data);
3326 }
3327
3328 static int
3329 target_unblock(struct device *dev, void *data)
3330 {
3331         if (scsi_is_target_device(dev))
3332                 starget_for_each_device(to_scsi_target(dev), data,
3333                                         device_unblock);
3334         return 0;
3335 }
3336
3337 void
3338 scsi_target_unblock(struct device *dev, enum scsi_device_state new_state)
3339 {
3340         if (scsi_is_target_device(dev))
3341                 starget_for_each_device(to_scsi_target(dev), &new_state,
3342                                         device_unblock);
3343         else
3344                 device_for_each_child(dev, &new_state, target_unblock);
3345 }
3346 EXPORT_SYMBOL_GPL(scsi_target_unblock);
3347
3348 /**
3349  * scsi_kmap_atomic_sg - find and atomically map an sg-elemnt
3350  * @sgl:        scatter-gather list
3351  * @sg_count:   number of segments in sg
3352  * @offset:     offset in bytes into sg, on return offset into the mapped area
3353  * @len:        bytes to map, on return number of bytes mapped
3354  *
3355  * Returns virtual address of the start of the mapped page
3356  */
3357 void *scsi_kmap_atomic_sg(struct scatterlist *sgl, int sg_count,
3358                           size_t *offset, size_t *len)
3359 {
3360         int i;
3361         size_t sg_len = 0, len_complete = 0;
3362         struct scatterlist *sg;
3363         struct page *page;
3364
3365         WARN_ON(!irqs_disabled());
3366
3367         for_each_sg(sgl, sg, sg_count, i) {
3368                 len_complete = sg_len; /* Complete sg-entries */
3369                 sg_len += sg->length;
3370                 if (sg_len > *offset)
3371                         break;
3372         }
3373
3374         if (unlikely(i == sg_count)) {
3375                 printk(KERN_ERR "%s: Bytes in sg: %zu, requested offset %zu, "
3376                         "elements %d\n",
3377                        __func__, sg_len, *offset, sg_count);
3378                 WARN_ON(1);
3379                 return NULL;
3380         }
3381
3382         /* Offset starting from the beginning of first page in this sg-entry */
3383         *offset = *offset - len_complete + sg->offset;
3384
3385         /* Assumption: contiguous pages can be accessed as "page + i" */
3386         page = nth_page(sg_page(sg), (*offset >> PAGE_SHIFT));
3387         *offset &= ~PAGE_MASK;
3388
3389         /* Bytes in this sg-entry from *offset to the end of the page */
3390         sg_len = PAGE_SIZE - *offset;
3391         if (*len > sg_len)
3392                 *len = sg_len;
3393
3394         return kmap_atomic(page);
3395 }
3396 EXPORT_SYMBOL(scsi_kmap_atomic_sg);
3397
3398 /**
3399  * scsi_kunmap_atomic_sg - atomically unmap a virtual address, previously mapped with scsi_kmap_atomic_sg
3400  * @virt:       virtual address to be unmapped
3401  */
3402 void scsi_kunmap_atomic_sg(void *virt)
3403 {
3404         kunmap_atomic(virt);
3405 }
3406 EXPORT_SYMBOL(scsi_kunmap_atomic_sg);
3407
3408 void sdev_disable_disk_events(struct scsi_device *sdev)
3409 {
3410         atomic_inc(&sdev->disk_events_disable_depth);
3411 }
3412 EXPORT_SYMBOL(sdev_disable_disk_events);
3413
3414 void sdev_enable_disk_events(struct scsi_device *sdev)
3415 {
3416         if (WARN_ON_ONCE(atomic_read(&sdev->disk_events_disable_depth) <= 0))
3417                 return;
3418         atomic_dec(&sdev->disk_events_disable_depth);
3419 }
3420 EXPORT_SYMBOL(sdev_enable_disk_events);
3421
3422 /**
3423  * scsi_vpd_lun_id - return a unique device identification
3424  * @sdev: SCSI device
3425  * @id:   buffer for the identification
3426  * @id_len:  length of the buffer
3427  *
3428  * Copies a unique device identification into @id based
3429  * on the information in the VPD page 0x83 of the device.
3430  * The string will be formatted as a SCSI name string.
3431  *
3432  * Returns the length of the identification or error on failure.
3433  * If the identifier is longer than the supplied buffer the actual
3434  * identifier length is returned and the buffer is not zero-padded.
3435  */
3436 int scsi_vpd_lun_id(struct scsi_device *sdev, char *id, size_t id_len)
3437 {
3438         u8 cur_id_type = 0xff;
3439         u8 cur_id_size = 0;
3440         const unsigned char *d, *cur_id_str;
3441         const struct scsi_vpd *vpd_pg83;
3442         int id_size = -EINVAL;
3443
3444         rcu_read_lock();
3445         vpd_pg83 = rcu_dereference(sdev->vpd_pg83);
3446         if (!vpd_pg83) {
3447                 rcu_read_unlock();
3448                 return -ENXIO;
3449         }
3450
3451         /*
3452          * Look for the correct descriptor.
3453          * Order of preference for lun descriptor:
3454          * - SCSI name string
3455          * - NAA IEEE Registered Extended
3456          * - EUI-64 based 16-byte
3457          * - EUI-64 based 12-byte
3458          * - NAA IEEE Registered
3459          * - NAA IEEE Extended
3460          * - T10 Vendor ID
3461          * as longer descriptors reduce the likelyhood
3462          * of identification clashes.
3463          */
3464
3465         /* The id string must be at least 20 bytes + terminating NULL byte */
3466         if (id_len < 21) {
3467                 rcu_read_unlock();
3468                 return -EINVAL;
3469         }
3470
3471         memset(id, 0, id_len);
3472         d = vpd_pg83->data + 4;
3473         while (d < vpd_pg83->data + vpd_pg83->len) {
3474                 /* Skip designators not referring to the LUN */
3475                 if ((d[1] & 0x30) != 0x00)
3476                         goto next_desig;
3477
3478                 switch (d[1] & 0xf) {
3479                 case 0x1:
3480                         /* T10 Vendor ID */
3481                         if (cur_id_size > d[3])
3482                                 break;
3483                         /* Prefer anything */
3484                         if (cur_id_type > 0x01 && cur_id_type != 0xff)
3485                                 break;
3486                         cur_id_size = d[3];
3487                         if (cur_id_size + 4 > id_len)
3488                                 cur_id_size = id_len - 4;
3489                         cur_id_str = d + 4;
3490                         cur_id_type = d[1] & 0xf;
3491                         id_size = snprintf(id, id_len, "t10.%*pE",
3492                                            cur_id_size, cur_id_str);
3493                         break;
3494                 case 0x2:
3495                         /* EUI-64 */
3496                         if (cur_id_size > d[3])
3497                                 break;
3498                         /* Prefer NAA IEEE Registered Extended */
3499                         if (cur_id_type == 0x3 &&
3500                             cur_id_size == d[3])
3501                                 break;
3502                         cur_id_size = d[3];
3503                         cur_id_str = d + 4;
3504                         cur_id_type = d[1] & 0xf;
3505                         switch (cur_id_size) {
3506                         case 8:
3507                                 id_size = snprintf(id, id_len,
3508                                                    "eui.%8phN",
3509                                                    cur_id_str);
3510                                 break;
3511                         case 12:
3512                                 id_size = snprintf(id, id_len,
3513                                                    "eui.%12phN",
3514                                                    cur_id_str);
3515                                 break;
3516                         case 16:
3517                                 id_size = snprintf(id, id_len,
3518                                                    "eui.%16phN",
3519                                                    cur_id_str);
3520                                 break;
3521                         default:
3522                                 cur_id_size = 0;
3523                                 break;
3524                         }
3525                         break;
3526                 case 0x3:
3527                         /* NAA */
3528                         if (cur_id_size > d[3])
3529                                 break;
3530                         cur_id_size = d[3];
3531                         cur_id_str = d + 4;
3532                         cur_id_type = d[1] & 0xf;
3533                         switch (cur_id_size) {
3534                         case 8:
3535                                 id_size = snprintf(id, id_len,
3536                                                    "naa.%8phN",
3537                                                    cur_id_str);
3538                                 break;
3539                         case 16:
3540                                 id_size = snprintf(id, id_len,
3541                                                    "naa.%16phN",
3542                                                    cur_id_str);
3543                                 break;
3544                         default:
3545                                 cur_id_size = 0;
3546                                 break;
3547                         }
3548                         break;
3549                 case 0x8:
3550                         /* SCSI name string */
3551                         if (cur_id_size + 4 > d[3])
3552                                 break;
3553                         /* Prefer others for truncated descriptor */
3554                         if (cur_id_size && d[3] > id_len)
3555                                 break;
3556                         cur_id_size = id_size = d[3];
3557                         cur_id_str = d + 4;
3558                         cur_id_type = d[1] & 0xf;
3559                         if (cur_id_size >= id_len)
3560                                 cur_id_size = id_len - 1;
3561                         memcpy(id, cur_id_str, cur_id_size);
3562                         /* Decrease priority for truncated descriptor */
3563                         if (cur_id_size != id_size)
3564                                 cur_id_size = 6;
3565                         break;
3566                 default:
3567                         break;
3568                 }
3569 next_desig:
3570                 d += d[3] + 4;
3571         }
3572         rcu_read_unlock();
3573
3574         return id_size;
3575 }
3576 EXPORT_SYMBOL(scsi_vpd_lun_id);
3577
3578 /*
3579  * scsi_vpd_tpg_id - return a target port group identifier
3580  * @sdev: SCSI device
3581  *
3582  * Returns the Target Port Group identifier from the information
3583  * froom VPD page 0x83 of the device.
3584  *
3585  * Returns the identifier or error on failure.
3586  */
3587 int scsi_vpd_tpg_id(struct scsi_device *sdev, int *rel_id)
3588 {
3589         const unsigned char *d;
3590         const struct scsi_vpd *vpd_pg83;
3591         int group_id = -EAGAIN, rel_port = -1;
3592
3593         rcu_read_lock();
3594         vpd_pg83 = rcu_dereference(sdev->vpd_pg83);
3595         if (!vpd_pg83) {
3596                 rcu_read_unlock();
3597                 return -ENXIO;
3598         }
3599
3600         d = vpd_pg83->data + 4;
3601         while (d < vpd_pg83->data + vpd_pg83->len) {
3602                 switch (d[1] & 0xf) {
3603                 case 0x4:
3604                         /* Relative target port */
3605                         rel_port = get_unaligned_be16(&d[6]);
3606                         break;
3607                 case 0x5:
3608                         /* Target port group */
3609                         group_id = get_unaligned_be16(&d[6]);
3610                         break;
3611                 default:
3612                         break;
3613                 }
3614                 d += d[3] + 4;
3615         }
3616         rcu_read_unlock();
3617
3618         if (group_id >= 0 && rel_id && rel_port != -1)
3619                 *rel_id = rel_port;
3620
3621         return group_id;
3622 }
3623 EXPORT_SYMBOL(scsi_vpd_tpg_id);