Merge tag 'scsi-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/jejb/scsi
[muen/linux.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                 scsi_mq_requeue_cmd(cmd);
195                 return;
196         }
197         spin_lock_irqsave(q->queue_lock, flags);
198         blk_requeue_request(q, cmd->request);
199         kblockd_schedule_work(&device->requeue_work);
200         spin_unlock_irqrestore(q->queue_lock, flags);
201 }
202
203 /*
204  * Function:    scsi_queue_insert()
205  *
206  * Purpose:     Insert a command in the midlevel queue.
207  *
208  * Arguments:   cmd    - command that we are adding to queue.
209  *              reason - why we are inserting command to queue.
210  *
211  * Lock status: Assumed that lock is not held upon entry.
212  *
213  * Returns:     Nothing.
214  *
215  * Notes:       We do this for one of two cases.  Either the host is busy
216  *              and it cannot accept any more commands for the time being,
217  *              or the device returned QUEUE_FULL and can accept no more
218  *              commands.
219  * Notes:       This could be called either from an interrupt context or a
220  *              normal process context.
221  */
222 void scsi_queue_insert(struct scsi_cmnd *cmd, int reason)
223 {
224         __scsi_queue_insert(cmd, reason, true);
225 }
226
227
228 /**
229  * scsi_execute - insert request and wait for the result
230  * @sdev:       scsi device
231  * @cmd:        scsi command
232  * @data_direction: data direction
233  * @buffer:     data buffer
234  * @bufflen:    len of buffer
235  * @sense:      optional sense buffer
236  * @sshdr:      optional decoded sense header
237  * @timeout:    request timeout in seconds
238  * @retries:    number of times to retry request
239  * @flags:      flags for ->cmd_flags
240  * @rq_flags:   flags for ->rq_flags
241  * @resid:      optional residual length
242  *
243  * Returns the scsi_cmnd result field if a command was executed, or a negative
244  * Linux error code if we didn't get that far.
245  */
246 int scsi_execute(struct scsi_device *sdev, const unsigned char *cmd,
247                  int data_direction, void *buffer, unsigned bufflen,
248                  unsigned char *sense, struct scsi_sense_hdr *sshdr,
249                  int timeout, int retries, u64 flags, req_flags_t rq_flags,
250                  int *resid)
251 {
252         struct request *req;
253         struct scsi_request *rq;
254         int ret = DRIVER_ERROR << 24;
255
256         req = blk_get_request_flags(sdev->request_queue,
257                         data_direction == DMA_TO_DEVICE ?
258                         REQ_OP_SCSI_OUT : REQ_OP_SCSI_IN, BLK_MQ_REQ_PREEMPT);
259         if (IS_ERR(req))
260                 return ret;
261         rq = scsi_req(req);
262
263         if (bufflen &&  blk_rq_map_kern(sdev->request_queue, req,
264                                         buffer, bufflen, __GFP_RECLAIM))
265                 goto out;
266
267         rq->cmd_len = COMMAND_SIZE(cmd[0]);
268         memcpy(rq->cmd, cmd, rq->cmd_len);
269         rq->retries = retries;
270         req->timeout = timeout;
271         req->cmd_flags |= flags;
272         req->rq_flags |= rq_flags | RQF_QUIET;
273
274         /*
275          * head injection *required* here otherwise quiesce won't work
276          */
277         blk_execute_rq(req->q, NULL, req, 1);
278
279         /*
280          * Some devices (USB mass-storage in particular) may transfer
281          * garbage data together with a residue indicating that the data
282          * is invalid.  Prevent the garbage from being misinterpreted
283          * and prevent security leaks by zeroing out the excess data.
284          */
285         if (unlikely(rq->resid_len > 0 && rq->resid_len <= bufflen))
286                 memset(buffer + (bufflen - rq->resid_len), 0, rq->resid_len);
287
288         if (resid)
289                 *resid = rq->resid_len;
290         if (sense && rq->sense_len)
291                 memcpy(sense, rq->sense, SCSI_SENSE_BUFFERSIZE);
292         if (sshdr)
293                 scsi_normalize_sense(rq->sense, rq->sense_len, sshdr);
294         ret = rq->result;
295  out:
296         blk_put_request(req);
297
298         return ret;
299 }
300 EXPORT_SYMBOL(scsi_execute);
301
302 /*
303  * Function:    scsi_init_cmd_errh()
304  *
305  * Purpose:     Initialize cmd fields related to error handling.
306  *
307  * Arguments:   cmd     - command that is ready to be queued.
308  *
309  * Notes:       This function has the job of initializing a number of
310  *              fields related to error handling.   Typically this will
311  *              be called once for each command, as required.
312  */
313 static void scsi_init_cmd_errh(struct scsi_cmnd *cmd)
314 {
315         cmd->serial_number = 0;
316         scsi_set_resid(cmd, 0);
317         memset(cmd->sense_buffer, 0, SCSI_SENSE_BUFFERSIZE);
318         if (cmd->cmd_len == 0)
319                 cmd->cmd_len = scsi_command_size(cmd->cmnd);
320 }
321
322 /*
323  * Decrement the host_busy counter and wake up the error handler if necessary.
324  * Avoid as follows that the error handler is not woken up if shost->host_busy
325  * == shost->host_failed: use call_rcu() in scsi_eh_scmd_add() in combination
326  * with an RCU read lock in this function to ensure that this function in its
327  * entirety either finishes before scsi_eh_scmd_add() increases the
328  * host_failed counter or that it notices the shost state change made by
329  * scsi_eh_scmd_add().
330  */
331 static void scsi_dec_host_busy(struct Scsi_Host *shost)
332 {
333         unsigned long flags;
334
335         rcu_read_lock();
336         atomic_dec(&shost->host_busy);
337         if (unlikely(scsi_host_in_recovery(shost))) {
338                 spin_lock_irqsave(shost->host_lock, flags);
339                 if (shost->host_failed || shost->host_eh_scheduled)
340                         scsi_eh_wakeup(shost);
341                 spin_unlock_irqrestore(shost->host_lock, flags);
342         }
343         rcu_read_unlock();
344 }
345
346 void scsi_device_unbusy(struct scsi_device *sdev)
347 {
348         struct Scsi_Host *shost = sdev->host;
349         struct scsi_target *starget = scsi_target(sdev);
350
351         scsi_dec_host_busy(shost);
352
353         if (starget->can_queue > 0)
354                 atomic_dec(&starget->target_busy);
355
356         atomic_dec(&sdev->device_busy);
357 }
358
359 static void scsi_kick_queue(struct request_queue *q)
360 {
361         if (q->mq_ops)
362                 blk_mq_start_hw_queues(q);
363         else
364                 blk_run_queue(q);
365 }
366
367 /*
368  * Called for single_lun devices on IO completion. Clear starget_sdev_user,
369  * and call blk_run_queue for all the scsi_devices on the target -
370  * including current_sdev first.
371  *
372  * Called with *no* scsi locks held.
373  */
374 static void scsi_single_lun_run(struct scsi_device *current_sdev)
375 {
376         struct Scsi_Host *shost = current_sdev->host;
377         struct scsi_device *sdev, *tmp;
378         struct scsi_target *starget = scsi_target(current_sdev);
379         unsigned long flags;
380
381         spin_lock_irqsave(shost->host_lock, flags);
382         starget->starget_sdev_user = NULL;
383         spin_unlock_irqrestore(shost->host_lock, flags);
384
385         /*
386          * Call blk_run_queue for all LUNs on the target, starting with
387          * current_sdev. We race with others (to set starget_sdev_user),
388          * but in most cases, we will be first. Ideally, each LU on the
389          * target would get some limited time or requests on the target.
390          */
391         scsi_kick_queue(current_sdev->request_queue);
392
393         spin_lock_irqsave(shost->host_lock, flags);
394         if (starget->starget_sdev_user)
395                 goto out;
396         list_for_each_entry_safe(sdev, tmp, &starget->devices,
397                         same_target_siblings) {
398                 if (sdev == current_sdev)
399                         continue;
400                 if (scsi_device_get(sdev))
401                         continue;
402
403                 spin_unlock_irqrestore(shost->host_lock, flags);
404                 scsi_kick_queue(sdev->request_queue);
405                 spin_lock_irqsave(shost->host_lock, flags);
406         
407                 scsi_device_put(sdev);
408         }
409  out:
410         spin_unlock_irqrestore(shost->host_lock, flags);
411 }
412
413 static inline bool scsi_device_is_busy(struct scsi_device *sdev)
414 {
415         if (atomic_read(&sdev->device_busy) >= sdev->queue_depth)
416                 return true;
417         if (atomic_read(&sdev->device_blocked) > 0)
418                 return true;
419         return false;
420 }
421
422 static inline bool scsi_target_is_busy(struct scsi_target *starget)
423 {
424         if (starget->can_queue > 0) {
425                 if (atomic_read(&starget->target_busy) >= starget->can_queue)
426                         return true;
427                 if (atomic_read(&starget->target_blocked) > 0)
428                         return true;
429         }
430         return false;
431 }
432
433 static inline bool scsi_host_is_busy(struct Scsi_Host *shost)
434 {
435         if (shost->can_queue > 0 &&
436             atomic_read(&shost->host_busy) >= shost->can_queue)
437                 return true;
438         if (atomic_read(&shost->host_blocked) > 0)
439                 return true;
440         if (shost->host_self_blocked)
441                 return true;
442         return false;
443 }
444
445 static void scsi_starved_list_run(struct Scsi_Host *shost)
446 {
447         LIST_HEAD(starved_list);
448         struct scsi_device *sdev;
449         unsigned long flags;
450
451         spin_lock_irqsave(shost->host_lock, flags);
452         list_splice_init(&shost->starved_list, &starved_list);
453
454         while (!list_empty(&starved_list)) {
455                 struct request_queue *slq;
456
457                 /*
458                  * As long as shost is accepting commands and we have
459                  * starved queues, call blk_run_queue. scsi_request_fn
460                  * drops the queue_lock and can add us back to the
461                  * starved_list.
462                  *
463                  * host_lock protects the starved_list and starved_entry.
464                  * scsi_request_fn must get the host_lock before checking
465                  * or modifying starved_list or starved_entry.
466                  */
467                 if (scsi_host_is_busy(shost))
468                         break;
469
470                 sdev = list_entry(starved_list.next,
471                                   struct scsi_device, starved_entry);
472                 list_del_init(&sdev->starved_entry);
473                 if (scsi_target_is_busy(scsi_target(sdev))) {
474                         list_move_tail(&sdev->starved_entry,
475                                        &shost->starved_list);
476                         continue;
477                 }
478
479                 /*
480                  * Once we drop the host lock, a racing scsi_remove_device()
481                  * call may remove the sdev from the starved list and destroy
482                  * it and the queue.  Mitigate by taking a reference to the
483                  * queue and never touching the sdev again after we drop the
484                  * host lock.  Note: if __scsi_remove_device() invokes
485                  * blk_cleanup_queue() before the queue is run from this
486                  * function then blk_run_queue() will return immediately since
487                  * blk_cleanup_queue() marks the queue with QUEUE_FLAG_DYING.
488                  */
489                 slq = sdev->request_queue;
490                 if (!blk_get_queue(slq))
491                         continue;
492                 spin_unlock_irqrestore(shost->host_lock, flags);
493
494                 scsi_kick_queue(slq);
495                 blk_put_queue(slq);
496
497                 spin_lock_irqsave(shost->host_lock, flags);
498         }
499         /* put any unprocessed entries back */
500         list_splice(&starved_list, &shost->starved_list);
501         spin_unlock_irqrestore(shost->host_lock, flags);
502 }
503
504 /*
505  * Function:   scsi_run_queue()
506  *
507  * Purpose:    Select a proper request queue to serve next
508  *
509  * Arguments:  q       - last request's queue
510  *
511  * Returns:     Nothing
512  *
513  * Notes:      The previous command was completely finished, start
514  *             a new one if possible.
515  */
516 static void scsi_run_queue(struct request_queue *q)
517 {
518         struct scsi_device *sdev = q->queuedata;
519
520         if (scsi_target(sdev)->single_lun)
521                 scsi_single_lun_run(sdev);
522         if (!list_empty(&sdev->host->starved_list))
523                 scsi_starved_list_run(sdev->host);
524
525         if (q->mq_ops)
526                 blk_mq_run_hw_queues(q, false);
527         else
528                 blk_run_queue(q);
529 }
530
531 void scsi_requeue_run_queue(struct work_struct *work)
532 {
533         struct scsi_device *sdev;
534         struct request_queue *q;
535
536         sdev = container_of(work, struct scsi_device, requeue_work);
537         q = sdev->request_queue;
538         scsi_run_queue(q);
539 }
540
541 /*
542  * Function:    scsi_requeue_command()
543  *
544  * Purpose:     Handle post-processing of completed commands.
545  *
546  * Arguments:   q       - queue to operate on
547  *              cmd     - command that may need to be requeued.
548  *
549  * Returns:     Nothing
550  *
551  * Notes:       After command completion, there may be blocks left
552  *              over which weren't finished by the previous command
553  *              this can be for a number of reasons - the main one is
554  *              I/O errors in the middle of the request, in which case
555  *              we need to request the blocks that come after the bad
556  *              sector.
557  * Notes:       Upon return, cmd is a stale pointer.
558  */
559 static void scsi_requeue_command(struct request_queue *q, struct scsi_cmnd *cmd)
560 {
561         struct scsi_device *sdev = cmd->device;
562         struct request *req = cmd->request;
563         unsigned long flags;
564
565         spin_lock_irqsave(q->queue_lock, flags);
566         blk_unprep_request(req);
567         req->special = NULL;
568         scsi_put_command(cmd);
569         blk_requeue_request(q, req);
570         spin_unlock_irqrestore(q->queue_lock, flags);
571
572         scsi_run_queue(q);
573
574         put_device(&sdev->sdev_gendev);
575 }
576
577 void scsi_run_host_queues(struct Scsi_Host *shost)
578 {
579         struct scsi_device *sdev;
580
581         shost_for_each_device(sdev, shost)
582                 scsi_run_queue(sdev->request_queue);
583 }
584
585 static void scsi_uninit_cmd(struct scsi_cmnd *cmd)
586 {
587         if (!blk_rq_is_passthrough(cmd->request)) {
588                 struct scsi_driver *drv = scsi_cmd_to_driver(cmd);
589
590                 if (drv->uninit_command)
591                         drv->uninit_command(cmd);
592         }
593 }
594
595 static void scsi_mq_free_sgtables(struct scsi_cmnd *cmd)
596 {
597         struct scsi_data_buffer *sdb;
598
599         if (cmd->sdb.table.nents)
600                 sg_free_table_chained(&cmd->sdb.table, true);
601         if (cmd->request->next_rq) {
602                 sdb = cmd->request->next_rq->special;
603                 if (sdb)
604                         sg_free_table_chained(&sdb->table, true);
605         }
606         if (scsi_prot_sg_count(cmd))
607                 sg_free_table_chained(&cmd->prot_sdb->table, true);
608 }
609
610 static void scsi_mq_uninit_cmd(struct scsi_cmnd *cmd)
611 {
612         scsi_mq_free_sgtables(cmd);
613         scsi_uninit_cmd(cmd);
614         scsi_del_cmd_from_list(cmd);
615 }
616
617 /*
618  * Function:    scsi_release_buffers()
619  *
620  * Purpose:     Free resources allocate for a scsi_command.
621  *
622  * Arguments:   cmd     - command that we are bailing.
623  *
624  * Lock status: Assumed that no lock is held upon entry.
625  *
626  * Returns:     Nothing
627  *
628  * Notes:       In the event that an upper level driver rejects a
629  *              command, we must release resources allocated during
630  *              the __init_io() function.  Primarily this would involve
631  *              the scatter-gather table.
632  */
633 static void scsi_release_buffers(struct scsi_cmnd *cmd)
634 {
635         if (cmd->sdb.table.nents)
636                 sg_free_table_chained(&cmd->sdb.table, false);
637
638         memset(&cmd->sdb, 0, sizeof(cmd->sdb));
639
640         if (scsi_prot_sg_count(cmd))
641                 sg_free_table_chained(&cmd->prot_sdb->table, false);
642 }
643
644 static void scsi_release_bidi_buffers(struct scsi_cmnd *cmd)
645 {
646         struct scsi_data_buffer *bidi_sdb = cmd->request->next_rq->special;
647
648         sg_free_table_chained(&bidi_sdb->table, false);
649         kmem_cache_free(scsi_sdb_cache, bidi_sdb);
650         cmd->request->next_rq->special = NULL;
651 }
652
653 static bool scsi_end_request(struct request *req, blk_status_t error,
654                 unsigned int bytes, unsigned int bidi_bytes)
655 {
656         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
657         struct scsi_device *sdev = cmd->device;
658         struct request_queue *q = sdev->request_queue;
659
660         if (blk_update_request(req, error, bytes))
661                 return true;
662
663         /* Bidi request must be completed as a whole */
664         if (unlikely(bidi_bytes) &&
665             blk_update_request(req->next_rq, error, bidi_bytes))
666                 return true;
667
668         if (blk_queue_add_random(q))
669                 add_disk_randomness(req->rq_disk);
670
671         if (!blk_rq_is_scsi(req)) {
672                 WARN_ON_ONCE(!(cmd->flags & SCMD_INITIALIZED));
673                 cmd->flags &= ~SCMD_INITIALIZED;
674                 destroy_rcu_head(&cmd->rcu);
675         }
676
677         if (req->mq_ctx) {
678                 /*
679                  * In the MQ case the command gets freed by __blk_mq_end_request,
680                  * so we have to do all cleanup that depends on it earlier.
681                  *
682                  * We also can't kick the queues from irq context, so we
683                  * will have to defer it to a workqueue.
684                  */
685                 scsi_mq_uninit_cmd(cmd);
686
687                 __blk_mq_end_request(req, error);
688
689                 if (scsi_target(sdev)->single_lun ||
690                     !list_empty(&sdev->host->starved_list))
691                         kblockd_schedule_work(&sdev->requeue_work);
692                 else
693                         blk_mq_run_hw_queues(q, true);
694         } else {
695                 unsigned long flags;
696
697                 if (bidi_bytes)
698                         scsi_release_bidi_buffers(cmd);
699                 scsi_release_buffers(cmd);
700                 scsi_put_command(cmd);
701
702                 spin_lock_irqsave(q->queue_lock, flags);
703                 blk_finish_request(req, error);
704                 spin_unlock_irqrestore(q->queue_lock, flags);
705
706                 scsi_run_queue(q);
707         }
708
709         put_device(&sdev->sdev_gendev);
710         return false;
711 }
712
713 /**
714  * __scsi_error_from_host_byte - translate SCSI error code into errno
715  * @cmd:        SCSI command (unused)
716  * @result:     scsi error code
717  *
718  * Translate SCSI error code into block errors.
719  */
720 static blk_status_t __scsi_error_from_host_byte(struct scsi_cmnd *cmd,
721                 int result)
722 {
723         switch (host_byte(result)) {
724         case DID_OK:
725                 return BLK_STS_OK;
726         case DID_TRANSPORT_FAILFAST:
727                 return BLK_STS_TRANSPORT;
728         case DID_TARGET_FAILURE:
729                 set_host_byte(cmd, DID_OK);
730                 return BLK_STS_TARGET;
731         case DID_NEXUS_FAILURE:
732                 return BLK_STS_NEXUS;
733         case DID_ALLOC_FAILURE:
734                 set_host_byte(cmd, DID_OK);
735                 return BLK_STS_NOSPC;
736         case DID_MEDIUM_ERROR:
737                 set_host_byte(cmd, DID_OK);
738                 return BLK_STS_MEDIUM;
739         default:
740                 return BLK_STS_IOERR;
741         }
742 }
743
744 /*
745  * Function:    scsi_io_completion()
746  *
747  * Purpose:     Completion processing for block device I/O requests.
748  *
749  * Arguments:   cmd   - command that is finished.
750  *
751  * Lock status: Assumed that no lock is held upon entry.
752  *
753  * Returns:     Nothing
754  *
755  * Notes:       We will finish off the specified number of sectors.  If we
756  *              are done, the command block will be released and the queue
757  *              function will be goosed.  If we are not done then we have to
758  *              figure out what to do next:
759  *
760  *              a) We can call scsi_requeue_command().  The request
761  *                 will be unprepared and put back on the queue.  Then
762  *                 a new command will be created for it.  This should
763  *                 be used if we made forward progress, or if we want
764  *                 to switch from READ(10) to READ(6) for example.
765  *
766  *              b) We can call __scsi_queue_insert().  The request will
767  *                 be put back on the queue and retried using the same
768  *                 command as before, possibly after a delay.
769  *
770  *              c) We can call scsi_end_request() with -EIO to fail
771  *                 the remainder of the request.
772  */
773 void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes)
774 {
775         int result = cmd->result;
776         struct request_queue *q = cmd->device->request_queue;
777         struct request *req = cmd->request;
778         blk_status_t error = BLK_STS_OK;
779         struct scsi_sense_hdr sshdr;
780         bool sense_valid = false;
781         int sense_deferred = 0, level = 0;
782         enum {ACTION_FAIL, ACTION_REPREP, ACTION_RETRY,
783               ACTION_DELAYED_RETRY} action;
784         unsigned long wait_for = (cmd->allowed + 1) * req->timeout;
785
786         if (result) {
787                 sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
788                 if (sense_valid)
789                         sense_deferred = scsi_sense_is_deferred(&sshdr);
790         }
791
792         if (blk_rq_is_passthrough(req)) {
793                 if (result) {
794                         if (sense_valid) {
795                                 /*
796                                  * SG_IO wants current and deferred errors
797                                  */
798                                 scsi_req(req)->sense_len =
799                                         min(8 + cmd->sense_buffer[7],
800                                             SCSI_SENSE_BUFFERSIZE);
801                         }
802                         if (!sense_deferred)
803                                 error = __scsi_error_from_host_byte(cmd, result);
804                 }
805                 /*
806                  * __scsi_error_from_host_byte may have reset the host_byte
807                  */
808                 scsi_req(req)->result = cmd->result;
809                 scsi_req(req)->resid_len = scsi_get_resid(cmd);
810
811                 if (scsi_bidi_cmnd(cmd)) {
812                         /*
813                          * Bidi commands Must be complete as a whole,
814                          * both sides at once.
815                          */
816                         scsi_req(req->next_rq)->resid_len = scsi_in(cmd)->resid;
817                         if (scsi_end_request(req, BLK_STS_OK, blk_rq_bytes(req),
818                                         blk_rq_bytes(req->next_rq)))
819                                 BUG();
820                         return;
821                 }
822         } else if (blk_rq_bytes(req) == 0 && result && !sense_deferred) {
823                 /*
824                  * Flush commands do not transfers any data, and thus cannot use
825                  * good_bytes != blk_rq_bytes(req) as the signal for an error.
826                  * This sets the error explicitly for the problem case.
827                  */
828                 error = __scsi_error_from_host_byte(cmd, result);
829         }
830
831         /* no bidi support for !blk_rq_is_passthrough yet */
832         BUG_ON(blk_bidi_rq(req));
833
834         /*
835          * Next deal with any sectors which we were able to correctly
836          * handle.
837          */
838         SCSI_LOG_HLCOMPLETE(1, scmd_printk(KERN_INFO, cmd,
839                 "%u sectors total, %d bytes done.\n",
840                 blk_rq_sectors(req), good_bytes));
841
842         /*
843          * Recovered errors need reporting, but they're always treated as
844          * success, so fiddle the result code here.  For passthrough requests
845          * we already took a copy of the original into sreq->result which
846          * is what gets returned to the user
847          */
848         if (sense_valid && (sshdr.sense_key == RECOVERED_ERROR)) {
849                 /* if ATA PASS-THROUGH INFORMATION AVAILABLE skip
850                  * print since caller wants ATA registers. Only occurs on
851                  * SCSI ATA PASS_THROUGH commands when CK_COND=1
852                  */
853                 if ((sshdr.asc == 0x0) && (sshdr.ascq == 0x1d))
854                         ;
855                 else if (!(req->rq_flags & RQF_QUIET))
856                         scsi_print_sense(cmd);
857                 result = 0;
858                 /* for passthrough error may be set */
859                 error = BLK_STS_OK;
860         }
861
862         /*
863          * special case: failed zero length commands always need to
864          * drop down into the retry code. Otherwise, if we finished
865          * all bytes in the request we are done now.
866          */
867         if (!(blk_rq_bytes(req) == 0 && error) &&
868             !scsi_end_request(req, error, good_bytes, 0))
869                 return;
870
871         /*
872          * Kill remainder if no retrys.
873          */
874         if (error && scsi_noretry_cmd(cmd)) {
875                 if (scsi_end_request(req, error, blk_rq_bytes(req), 0))
876                         BUG();
877                 return;
878         }
879
880         /*
881          * If there had been no error, but we have leftover bytes in the
882          * requeues just queue the command up again.
883          */
884         if (result == 0)
885                 goto requeue;
886
887         error = __scsi_error_from_host_byte(cmd, result);
888
889         if (host_byte(result) == DID_RESET) {
890                 /* Third party bus reset or reset for error recovery
891                  * reasons.  Just retry the command and see what
892                  * happens.
893                  */
894                 action = ACTION_RETRY;
895         } else if (sense_valid && !sense_deferred) {
896                 switch (sshdr.sense_key) {
897                 case UNIT_ATTENTION:
898                         if (cmd->device->removable) {
899                                 /* Detected disc change.  Set a bit
900                                  * and quietly refuse further access.
901                                  */
902                                 cmd->device->changed = 1;
903                                 action = ACTION_FAIL;
904                         } else {
905                                 /* Must have been a power glitch, or a
906                                  * bus reset.  Could not have been a
907                                  * media change, so we just retry the
908                                  * command and see what happens.
909                                  */
910                                 action = ACTION_RETRY;
911                         }
912                         break;
913                 case ILLEGAL_REQUEST:
914                         /* If we had an ILLEGAL REQUEST returned, then
915                          * we may have performed an unsupported
916                          * command.  The only thing this should be
917                          * would be a ten byte read where only a six
918                          * byte read was supported.  Also, on a system
919                          * where READ CAPACITY failed, we may have
920                          * read past the end of the disk.
921                          */
922                         if ((cmd->device->use_10_for_rw &&
923                             sshdr.asc == 0x20 && sshdr.ascq == 0x00) &&
924                             (cmd->cmnd[0] == READ_10 ||
925                              cmd->cmnd[0] == WRITE_10)) {
926                                 /* This will issue a new 6-byte command. */
927                                 cmd->device->use_10_for_rw = 0;
928                                 action = ACTION_REPREP;
929                         } else if (sshdr.asc == 0x10) /* DIX */ {
930                                 action = ACTION_FAIL;
931                                 error = BLK_STS_PROTECTION;
932                         /* INVALID COMMAND OPCODE or INVALID FIELD IN CDB */
933                         } else if (sshdr.asc == 0x20 || sshdr.asc == 0x24) {
934                                 action = ACTION_FAIL;
935                                 error = BLK_STS_TARGET;
936                         } else
937                                 action = ACTION_FAIL;
938                         break;
939                 case ABORTED_COMMAND:
940                         action = ACTION_FAIL;
941                         if (sshdr.asc == 0x10) /* DIF */
942                                 error = BLK_STS_PROTECTION;
943                         break;
944                 case NOT_READY:
945                         /* If the device is in the process of becoming
946                          * ready, or has a temporary blockage, retry.
947                          */
948                         if (sshdr.asc == 0x04) {
949                                 switch (sshdr.ascq) {
950                                 case 0x01: /* becoming ready */
951                                 case 0x04: /* format in progress */
952                                 case 0x05: /* rebuild in progress */
953                                 case 0x06: /* recalculation in progress */
954                                 case 0x07: /* operation in progress */
955                                 case 0x08: /* Long write in progress */
956                                 case 0x09: /* self test in progress */
957                                 case 0x14: /* space allocation in progress */
958                                         action = ACTION_DELAYED_RETRY;
959                                         break;
960                                 default:
961                                         action = ACTION_FAIL;
962                                         break;
963                                 }
964                         } else
965                                 action = ACTION_FAIL;
966                         break;
967                 case VOLUME_OVERFLOW:
968                         /* See SSC3rXX or current. */
969                         action = ACTION_FAIL;
970                         break;
971                 default:
972                         action = ACTION_FAIL;
973                         break;
974                 }
975         } else
976                 action = ACTION_FAIL;
977
978         if (action != ACTION_FAIL &&
979             time_before(cmd->jiffies_at_alloc + wait_for, jiffies))
980                 action = ACTION_FAIL;
981
982         switch (action) {
983         case ACTION_FAIL:
984                 /* Give up and fail the remainder of the request */
985                 if (!(req->rq_flags & RQF_QUIET)) {
986                         static DEFINE_RATELIMIT_STATE(_rs,
987                                         DEFAULT_RATELIMIT_INTERVAL,
988                                         DEFAULT_RATELIMIT_BURST);
989
990                         if (unlikely(scsi_logging_level))
991                                 level = SCSI_LOG_LEVEL(SCSI_LOG_MLCOMPLETE_SHIFT,
992                                                        SCSI_LOG_MLCOMPLETE_BITS);
993
994                         /*
995                          * if logging is enabled the failure will be printed
996                          * in scsi_log_completion(), so avoid duplicate messages
997                          */
998                         if (!level && __ratelimit(&_rs)) {
999                                 scsi_print_result(cmd, NULL, FAILED);
1000                                 if (driver_byte(result) & DRIVER_SENSE)
1001                                         scsi_print_sense(cmd);
1002                                 scsi_print_command(cmd);
1003                         }
1004                 }
1005                 if (!scsi_end_request(req, error, blk_rq_err_bytes(req), 0))
1006                         return;
1007                 /*FALLTHRU*/
1008         case ACTION_REPREP:
1009         requeue:
1010                 /* Unprep the request and put it back at the head of the queue.
1011                  * A new command will be prepared and issued.
1012                  */
1013                 if (q->mq_ops) {
1014                         scsi_mq_requeue_cmd(cmd);
1015                 } else {
1016                         scsi_release_buffers(cmd);
1017                         scsi_requeue_command(q, cmd);
1018                 }
1019                 break;
1020         case ACTION_RETRY:
1021                 /* Retry the same command immediately */
1022                 __scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY, false);
1023                 break;
1024         case ACTION_DELAYED_RETRY:
1025                 /* Retry the same command after a delay */
1026                 __scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY, false);
1027                 break;
1028         }
1029 }
1030
1031 static int scsi_init_sgtable(struct request *req, struct scsi_data_buffer *sdb)
1032 {
1033         int count;
1034
1035         /*
1036          * If sg table allocation fails, requeue request later.
1037          */
1038         if (unlikely(sg_alloc_table_chained(&sdb->table,
1039                         blk_rq_nr_phys_segments(req), sdb->table.sgl)))
1040                 return BLKPREP_DEFER;
1041
1042         /* 
1043          * Next, walk the list, and fill in the addresses and sizes of
1044          * each segment.
1045          */
1046         count = blk_rq_map_sg(req->q, req, sdb->table.sgl);
1047         BUG_ON(count > sdb->table.nents);
1048         sdb->table.nents = count;
1049         sdb->length = blk_rq_payload_bytes(req);
1050         return BLKPREP_OK;
1051 }
1052
1053 /*
1054  * Function:    scsi_init_io()
1055  *
1056  * Purpose:     SCSI I/O initialize function.
1057  *
1058  * Arguments:   cmd   - Command descriptor we wish to initialize
1059  *
1060  * Returns:     0 on success
1061  *              BLKPREP_DEFER if the failure is retryable
1062  *              BLKPREP_KILL if the failure is fatal
1063  */
1064 int scsi_init_io(struct scsi_cmnd *cmd)
1065 {
1066         struct scsi_device *sdev = cmd->device;
1067         struct request *rq = cmd->request;
1068         bool is_mq = (rq->mq_ctx != NULL);
1069         int error = BLKPREP_KILL;
1070
1071         if (WARN_ON_ONCE(!blk_rq_nr_phys_segments(rq)))
1072                 goto err_exit;
1073
1074         error = scsi_init_sgtable(rq, &cmd->sdb);
1075         if (error)
1076                 goto err_exit;
1077
1078         if (blk_bidi_rq(rq)) {
1079                 if (!rq->q->mq_ops) {
1080                         struct scsi_data_buffer *bidi_sdb =
1081                                 kmem_cache_zalloc(scsi_sdb_cache, GFP_ATOMIC);
1082                         if (!bidi_sdb) {
1083                                 error = BLKPREP_DEFER;
1084                                 goto err_exit;
1085                         }
1086
1087                         rq->next_rq->special = bidi_sdb;
1088                 }
1089
1090                 error = scsi_init_sgtable(rq->next_rq, rq->next_rq->special);
1091                 if (error)
1092                         goto err_exit;
1093         }
1094
1095         if (blk_integrity_rq(rq)) {
1096                 struct scsi_data_buffer *prot_sdb = cmd->prot_sdb;
1097                 int ivecs, count;
1098
1099                 if (prot_sdb == NULL) {
1100                         /*
1101                          * This can happen if someone (e.g. multipath)
1102                          * queues a command to a device on an adapter
1103                          * that does not support DIX.
1104                          */
1105                         WARN_ON_ONCE(1);
1106                         error = BLKPREP_KILL;
1107                         goto err_exit;
1108                 }
1109
1110                 ivecs = blk_rq_count_integrity_sg(rq->q, rq->bio);
1111
1112                 if (sg_alloc_table_chained(&prot_sdb->table, ivecs,
1113                                 prot_sdb->table.sgl)) {
1114                         error = BLKPREP_DEFER;
1115                         goto err_exit;
1116                 }
1117
1118                 count = blk_rq_map_integrity_sg(rq->q, rq->bio,
1119                                                 prot_sdb->table.sgl);
1120                 BUG_ON(unlikely(count > ivecs));
1121                 BUG_ON(unlikely(count > queue_max_integrity_segments(rq->q)));
1122
1123                 cmd->prot_sdb = prot_sdb;
1124                 cmd->prot_sdb->table.nents = count;
1125         }
1126
1127         return BLKPREP_OK;
1128 err_exit:
1129         if (is_mq) {
1130                 scsi_mq_free_sgtables(cmd);
1131         } else {
1132                 scsi_release_buffers(cmd);
1133                 cmd->request->special = NULL;
1134                 scsi_put_command(cmd);
1135                 put_device(&sdev->sdev_gendev);
1136         }
1137         return error;
1138 }
1139 EXPORT_SYMBOL(scsi_init_io);
1140
1141 /**
1142  * scsi_initialize_rq - initialize struct scsi_cmnd partially
1143  * @rq: Request associated with the SCSI command to be initialized.
1144  *
1145  * This function initializes the members of struct scsi_cmnd that must be
1146  * initialized before request processing starts and that won't be
1147  * reinitialized if a SCSI command is requeued.
1148  *
1149  * Called from inside blk_get_request() for pass-through requests and from
1150  * inside scsi_init_command() for filesystem requests.
1151  */
1152 static void scsi_initialize_rq(struct request *rq)
1153 {
1154         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1155
1156         scsi_req_init(&cmd->req);
1157         init_rcu_head(&cmd->rcu);
1158         cmd->jiffies_at_alloc = jiffies;
1159         cmd->retries = 0;
1160 }
1161
1162 /* Add a command to the list used by the aacraid and dpt_i2o drivers */
1163 void scsi_add_cmd_to_list(struct scsi_cmnd *cmd)
1164 {
1165         struct scsi_device *sdev = cmd->device;
1166         struct Scsi_Host *shost = sdev->host;
1167         unsigned long flags;
1168
1169         if (shost->use_cmd_list) {
1170                 spin_lock_irqsave(&sdev->list_lock, flags);
1171                 list_add_tail(&cmd->list, &sdev->cmd_list);
1172                 spin_unlock_irqrestore(&sdev->list_lock, flags);
1173         }
1174 }
1175
1176 /* Remove a command from the list used by the aacraid and dpt_i2o drivers */
1177 void scsi_del_cmd_from_list(struct scsi_cmnd *cmd)
1178 {
1179         struct scsi_device *sdev = cmd->device;
1180         struct Scsi_Host *shost = sdev->host;
1181         unsigned long flags;
1182
1183         if (shost->use_cmd_list) {
1184                 spin_lock_irqsave(&sdev->list_lock, flags);
1185                 BUG_ON(list_empty(&cmd->list));
1186                 list_del_init(&cmd->list);
1187                 spin_unlock_irqrestore(&sdev->list_lock, flags);
1188         }
1189 }
1190
1191 /* Called after a request has been started. */
1192 void scsi_init_command(struct scsi_device *dev, struct scsi_cmnd *cmd)
1193 {
1194         void *buf = cmd->sense_buffer;
1195         void *prot = cmd->prot_sdb;
1196         struct request *rq = blk_mq_rq_from_pdu(cmd);
1197         unsigned int flags = cmd->flags & SCMD_PRESERVED_FLAGS;
1198         unsigned long jiffies_at_alloc;
1199         int retries;
1200
1201         if (!blk_rq_is_scsi(rq) && !(flags & SCMD_INITIALIZED)) {
1202                 flags |= SCMD_INITIALIZED;
1203                 scsi_initialize_rq(rq);
1204         }
1205
1206         jiffies_at_alloc = cmd->jiffies_at_alloc;
1207         retries = cmd->retries;
1208         /* zero out the cmd, except for the embedded scsi_request */
1209         memset((char *)cmd + sizeof(cmd->req), 0,
1210                 sizeof(*cmd) - sizeof(cmd->req) + dev->host->hostt->cmd_size);
1211
1212         cmd->device = dev;
1213         cmd->sense_buffer = buf;
1214         cmd->prot_sdb = prot;
1215         cmd->flags = flags;
1216         INIT_DELAYED_WORK(&cmd->abort_work, scmd_eh_abort_handler);
1217         cmd->jiffies_at_alloc = jiffies_at_alloc;
1218         cmd->retries = retries;
1219
1220         scsi_add_cmd_to_list(cmd);
1221 }
1222
1223 static int scsi_setup_scsi_cmnd(struct scsi_device *sdev, struct request *req)
1224 {
1225         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1226
1227         /*
1228          * Passthrough requests may transfer data, in which case they must
1229          * a bio attached to them.  Or they might contain a SCSI command
1230          * that does not transfer data, in which case they may optionally
1231          * submit a request without an attached bio.
1232          */
1233         if (req->bio) {
1234                 int ret = scsi_init_io(cmd);
1235                 if (unlikely(ret))
1236                         return ret;
1237         } else {
1238                 BUG_ON(blk_rq_bytes(req));
1239
1240                 memset(&cmd->sdb, 0, sizeof(cmd->sdb));
1241         }
1242
1243         cmd->cmd_len = scsi_req(req)->cmd_len;
1244         cmd->cmnd = scsi_req(req)->cmd;
1245         cmd->transfersize = blk_rq_bytes(req);
1246         cmd->allowed = scsi_req(req)->retries;
1247         return BLKPREP_OK;
1248 }
1249
1250 /*
1251  * Setup a normal block command.  These are simple request from filesystems
1252  * that still need to be translated to SCSI CDBs from the ULD.
1253  */
1254 static int scsi_setup_fs_cmnd(struct scsi_device *sdev, struct request *req)
1255 {
1256         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1257
1258         if (unlikely(sdev->handler && sdev->handler->prep_fn)) {
1259                 int ret = sdev->handler->prep_fn(sdev, req);
1260                 if (ret != BLKPREP_OK)
1261                         return ret;
1262         }
1263
1264         cmd->cmnd = scsi_req(req)->cmd = scsi_req(req)->__cmd;
1265         memset(cmd->cmnd, 0, BLK_MAX_CDB);
1266         return scsi_cmd_to_driver(cmd)->init_command(cmd);
1267 }
1268
1269 static int scsi_setup_cmnd(struct scsi_device *sdev, struct request *req)
1270 {
1271         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1272
1273         if (!blk_rq_bytes(req))
1274                 cmd->sc_data_direction = DMA_NONE;
1275         else if (rq_data_dir(req) == WRITE)
1276                 cmd->sc_data_direction = DMA_TO_DEVICE;
1277         else
1278                 cmd->sc_data_direction = DMA_FROM_DEVICE;
1279
1280         if (blk_rq_is_scsi(req))
1281                 return scsi_setup_scsi_cmnd(sdev, req);
1282         else
1283                 return scsi_setup_fs_cmnd(sdev, req);
1284 }
1285
1286 static int
1287 scsi_prep_state_check(struct scsi_device *sdev, struct request *req)
1288 {
1289         int ret = BLKPREP_OK;
1290
1291         /*
1292          * If the device is not in running state we will reject some
1293          * or all commands.
1294          */
1295         if (unlikely(sdev->sdev_state != SDEV_RUNNING)) {
1296                 switch (sdev->sdev_state) {
1297                 case SDEV_OFFLINE:
1298                 case SDEV_TRANSPORT_OFFLINE:
1299                         /*
1300                          * If the device is offline we refuse to process any
1301                          * commands.  The device must be brought online
1302                          * before trying any recovery commands.
1303                          */
1304                         sdev_printk(KERN_ERR, sdev,
1305                                     "rejecting I/O to offline device\n");
1306                         ret = BLKPREP_KILL;
1307                         break;
1308                 case SDEV_DEL:
1309                         /*
1310                          * If the device is fully deleted, we refuse to
1311                          * process any commands as well.
1312                          */
1313                         sdev_printk(KERN_ERR, sdev,
1314                                     "rejecting I/O to dead device\n");
1315                         ret = BLKPREP_KILL;
1316                         break;
1317                 case SDEV_BLOCK:
1318                 case SDEV_CREATED_BLOCK:
1319                         ret = BLKPREP_DEFER;
1320                         break;
1321                 case SDEV_QUIESCE:
1322                         /*
1323                          * If the devices is blocked we defer normal commands.
1324                          */
1325                         if (req && !(req->rq_flags & RQF_PREEMPT))
1326                                 ret = BLKPREP_DEFER;
1327                         break;
1328                 default:
1329                         /*
1330                          * For any other not fully online state we only allow
1331                          * special commands.  In particular any user initiated
1332                          * command is not allowed.
1333                          */
1334                         if (req && !(req->rq_flags & RQF_PREEMPT))
1335                                 ret = BLKPREP_KILL;
1336                         break;
1337                 }
1338         }
1339         return ret;
1340 }
1341
1342 static int
1343 scsi_prep_return(struct request_queue *q, struct request *req, int ret)
1344 {
1345         struct scsi_device *sdev = q->queuedata;
1346
1347         switch (ret) {
1348         case BLKPREP_KILL:
1349         case BLKPREP_INVALID:
1350                 scsi_req(req)->result = DID_NO_CONNECT << 16;
1351                 /* release the command and kill it */
1352                 if (req->special) {
1353                         struct scsi_cmnd *cmd = req->special;
1354                         scsi_release_buffers(cmd);
1355                         scsi_put_command(cmd);
1356                         put_device(&sdev->sdev_gendev);
1357                         req->special = NULL;
1358                 }
1359                 break;
1360         case BLKPREP_DEFER:
1361                 /*
1362                  * If we defer, the blk_peek_request() returns NULL, but the
1363                  * queue must be restarted, so we schedule a callback to happen
1364                  * shortly.
1365                  */
1366                 if (atomic_read(&sdev->device_busy) == 0)
1367                         blk_delay_queue(q, SCSI_QUEUE_DELAY);
1368                 break;
1369         default:
1370                 req->rq_flags |= RQF_DONTPREP;
1371         }
1372
1373         return ret;
1374 }
1375
1376 static int scsi_prep_fn(struct request_queue *q, struct request *req)
1377 {
1378         struct scsi_device *sdev = q->queuedata;
1379         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1380         int ret;
1381
1382         ret = scsi_prep_state_check(sdev, req);
1383         if (ret != BLKPREP_OK)
1384                 goto out;
1385
1386         if (!req->special) {
1387                 /* Bail if we can't get a reference to the device */
1388                 if (unlikely(!get_device(&sdev->sdev_gendev))) {
1389                         ret = BLKPREP_DEFER;
1390                         goto out;
1391                 }
1392
1393                 scsi_init_command(sdev, cmd);
1394                 req->special = cmd;
1395         }
1396
1397         cmd->tag = req->tag;
1398         cmd->request = req;
1399         cmd->prot_op = SCSI_PROT_NORMAL;
1400
1401         ret = scsi_setup_cmnd(sdev, req);
1402 out:
1403         return scsi_prep_return(q, req, ret);
1404 }
1405
1406 static void scsi_unprep_fn(struct request_queue *q, struct request *req)
1407 {
1408         scsi_uninit_cmd(blk_mq_rq_to_pdu(req));
1409 }
1410
1411 /*
1412  * scsi_dev_queue_ready: if we can send requests to sdev, return 1 else
1413  * return 0.
1414  *
1415  * Called with the queue_lock held.
1416  */
1417 static inline int scsi_dev_queue_ready(struct request_queue *q,
1418                                   struct scsi_device *sdev)
1419 {
1420         unsigned int busy;
1421
1422         busy = atomic_inc_return(&sdev->device_busy) - 1;
1423         if (atomic_read(&sdev->device_blocked)) {
1424                 if (busy)
1425                         goto out_dec;
1426
1427                 /*
1428                  * unblock after device_blocked iterates to zero
1429                  */
1430                 if (atomic_dec_return(&sdev->device_blocked) > 0) {
1431                         /*
1432                          * For the MQ case we take care of this in the caller.
1433                          */
1434                         if (!q->mq_ops)
1435                                 blk_delay_queue(q, SCSI_QUEUE_DELAY);
1436                         goto out_dec;
1437                 }
1438                 SCSI_LOG_MLQUEUE(3, sdev_printk(KERN_INFO, sdev,
1439                                    "unblocking device at zero depth\n"));
1440         }
1441
1442         if (busy >= sdev->queue_depth)
1443                 goto out_dec;
1444
1445         return 1;
1446 out_dec:
1447         atomic_dec(&sdev->device_busy);
1448         return 0;
1449 }
1450
1451 /*
1452  * scsi_target_queue_ready: checks if there we can send commands to target
1453  * @sdev: scsi device on starget to check.
1454  */
1455 static inline int scsi_target_queue_ready(struct Scsi_Host *shost,
1456                                            struct scsi_device *sdev)
1457 {
1458         struct scsi_target *starget = scsi_target(sdev);
1459         unsigned int busy;
1460
1461         if (starget->single_lun) {
1462                 spin_lock_irq(shost->host_lock);
1463                 if (starget->starget_sdev_user &&
1464                     starget->starget_sdev_user != sdev) {
1465                         spin_unlock_irq(shost->host_lock);
1466                         return 0;
1467                 }
1468                 starget->starget_sdev_user = sdev;
1469                 spin_unlock_irq(shost->host_lock);
1470         }
1471
1472         if (starget->can_queue <= 0)
1473                 return 1;
1474
1475         busy = atomic_inc_return(&starget->target_busy) - 1;
1476         if (atomic_read(&starget->target_blocked) > 0) {
1477                 if (busy)
1478                         goto starved;
1479
1480                 /*
1481                  * unblock after target_blocked iterates to zero
1482                  */
1483                 if (atomic_dec_return(&starget->target_blocked) > 0)
1484                         goto out_dec;
1485
1486                 SCSI_LOG_MLQUEUE(3, starget_printk(KERN_INFO, starget,
1487                                  "unblocking target at zero depth\n"));
1488         }
1489
1490         if (busy >= starget->can_queue)
1491                 goto starved;
1492
1493         return 1;
1494
1495 starved:
1496         spin_lock_irq(shost->host_lock);
1497         list_move_tail(&sdev->starved_entry, &shost->starved_list);
1498         spin_unlock_irq(shost->host_lock);
1499 out_dec:
1500         if (starget->can_queue > 0)
1501                 atomic_dec(&starget->target_busy);
1502         return 0;
1503 }
1504
1505 /*
1506  * scsi_host_queue_ready: if we can send requests to shost, return 1 else
1507  * return 0. We must end up running the queue again whenever 0 is
1508  * returned, else IO can hang.
1509  */
1510 static inline int scsi_host_queue_ready(struct request_queue *q,
1511                                    struct Scsi_Host *shost,
1512                                    struct scsi_device *sdev)
1513 {
1514         unsigned int busy;
1515
1516         if (scsi_host_in_recovery(shost))
1517                 return 0;
1518
1519         busy = atomic_inc_return(&shost->host_busy) - 1;
1520         if (atomic_read(&shost->host_blocked) > 0) {
1521                 if (busy)
1522                         goto starved;
1523
1524                 /*
1525                  * unblock after host_blocked iterates to zero
1526                  */
1527                 if (atomic_dec_return(&shost->host_blocked) > 0)
1528                         goto out_dec;
1529
1530                 SCSI_LOG_MLQUEUE(3,
1531                         shost_printk(KERN_INFO, shost,
1532                                      "unblocking host at zero depth\n"));
1533         }
1534
1535         if (shost->can_queue > 0 && busy >= shost->can_queue)
1536                 goto starved;
1537         if (shost->host_self_blocked)
1538                 goto starved;
1539
1540         /* We're OK to process the command, so we can't be starved */
1541         if (!list_empty(&sdev->starved_entry)) {
1542                 spin_lock_irq(shost->host_lock);
1543                 if (!list_empty(&sdev->starved_entry))
1544                         list_del_init(&sdev->starved_entry);
1545                 spin_unlock_irq(shost->host_lock);
1546         }
1547
1548         return 1;
1549
1550 starved:
1551         spin_lock_irq(shost->host_lock);
1552         if (list_empty(&sdev->starved_entry))
1553                 list_add_tail(&sdev->starved_entry, &shost->starved_list);
1554         spin_unlock_irq(shost->host_lock);
1555 out_dec:
1556         scsi_dec_host_busy(shost);
1557         return 0;
1558 }
1559
1560 /*
1561  * Busy state exporting function for request stacking drivers.
1562  *
1563  * For efficiency, no lock is taken to check the busy state of
1564  * shost/starget/sdev, since the returned value is not guaranteed and
1565  * may be changed after request stacking drivers call the function,
1566  * regardless of taking lock or not.
1567  *
1568  * When scsi can't dispatch I/Os anymore and needs to kill I/Os scsi
1569  * needs to return 'not busy'. Otherwise, request stacking drivers
1570  * may hold requests forever.
1571  */
1572 static int scsi_lld_busy(struct request_queue *q)
1573 {
1574         struct scsi_device *sdev = q->queuedata;
1575         struct Scsi_Host *shost;
1576
1577         if (blk_queue_dying(q))
1578                 return 0;
1579
1580         shost = sdev->host;
1581
1582         /*
1583          * Ignore host/starget busy state.
1584          * Since block layer does not have a concept of fairness across
1585          * multiple queues, congestion of host/starget needs to be handled
1586          * in SCSI layer.
1587          */
1588         if (scsi_host_in_recovery(shost) || scsi_device_is_busy(sdev))
1589                 return 1;
1590
1591         return 0;
1592 }
1593
1594 /*
1595  * Kill a request for a dead device
1596  */
1597 static void scsi_kill_request(struct request *req, struct request_queue *q)
1598 {
1599         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1600         struct scsi_device *sdev;
1601         struct scsi_target *starget;
1602         struct Scsi_Host *shost;
1603
1604         blk_start_request(req);
1605
1606         scmd_printk(KERN_INFO, cmd, "killing request\n");
1607
1608         sdev = cmd->device;
1609         starget = scsi_target(sdev);
1610         shost = sdev->host;
1611         scsi_init_cmd_errh(cmd);
1612         cmd->result = DID_NO_CONNECT << 16;
1613         atomic_inc(&cmd->device->iorequest_cnt);
1614
1615         /*
1616          * SCSI request completion path will do scsi_device_unbusy(),
1617          * bump busy counts.  To bump the counters, we need to dance
1618          * with the locks as normal issue path does.
1619          */
1620         atomic_inc(&sdev->device_busy);
1621         atomic_inc(&shost->host_busy);
1622         if (starget->can_queue > 0)
1623                 atomic_inc(&starget->target_busy);
1624
1625         blk_complete_request(req);
1626 }
1627
1628 static void scsi_softirq_done(struct request *rq)
1629 {
1630         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1631         unsigned long wait_for = (cmd->allowed + 1) * rq->timeout;
1632         int disposition;
1633
1634         INIT_LIST_HEAD(&cmd->eh_entry);
1635
1636         atomic_inc(&cmd->device->iodone_cnt);
1637         if (cmd->result)
1638                 atomic_inc(&cmd->device->ioerr_cnt);
1639
1640         disposition = scsi_decide_disposition(cmd);
1641         if (disposition != SUCCESS &&
1642             time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) {
1643                 sdev_printk(KERN_ERR, cmd->device,
1644                             "timing out command, waited %lus\n",
1645                             wait_for/HZ);
1646                 disposition = SUCCESS;
1647         }
1648
1649         scsi_log_completion(cmd, disposition);
1650
1651         switch (disposition) {
1652                 case SUCCESS:
1653                         scsi_finish_command(cmd);
1654                         break;
1655                 case NEEDS_RETRY:
1656                         scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY);
1657                         break;
1658                 case ADD_TO_MLQUEUE:
1659                         scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY);
1660                         break;
1661                 default:
1662                         scsi_eh_scmd_add(cmd);
1663                         break;
1664         }
1665 }
1666
1667 /**
1668  * scsi_dispatch_command - Dispatch a command to the low-level driver.
1669  * @cmd: command block we are dispatching.
1670  *
1671  * Return: nonzero return request was rejected and device's queue needs to be
1672  * plugged.
1673  */
1674 static int scsi_dispatch_cmd(struct scsi_cmnd *cmd)
1675 {
1676         struct Scsi_Host *host = cmd->device->host;
1677         int rtn = 0;
1678
1679         atomic_inc(&cmd->device->iorequest_cnt);
1680
1681         /* check if the device is still usable */
1682         if (unlikely(cmd->device->sdev_state == SDEV_DEL)) {
1683                 /* in SDEV_DEL we error all commands. DID_NO_CONNECT
1684                  * returns an immediate error upwards, and signals
1685                  * that the device is no longer present */
1686                 cmd->result = DID_NO_CONNECT << 16;
1687                 goto done;
1688         }
1689
1690         /* Check to see if the scsi lld made this device blocked. */
1691         if (unlikely(scsi_device_blocked(cmd->device))) {
1692                 /*
1693                  * in blocked state, the command is just put back on
1694                  * the device queue.  The suspend state has already
1695                  * blocked the queue so future requests should not
1696                  * occur until the device transitions out of the
1697                  * suspend state.
1698                  */
1699                 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1700                         "queuecommand : device blocked\n"));
1701                 return SCSI_MLQUEUE_DEVICE_BUSY;
1702         }
1703
1704         /* Store the LUN value in cmnd, if needed. */
1705         if (cmd->device->lun_in_cdb)
1706                 cmd->cmnd[1] = (cmd->cmnd[1] & 0x1f) |
1707                                (cmd->device->lun << 5 & 0xe0);
1708
1709         scsi_log_send(cmd);
1710
1711         /*
1712          * Before we queue this command, check if the command
1713          * length exceeds what the host adapter can handle.
1714          */
1715         if (cmd->cmd_len > cmd->device->host->max_cmd_len) {
1716                 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1717                                "queuecommand : command too long. "
1718                                "cdb_size=%d host->max_cmd_len=%d\n",
1719                                cmd->cmd_len, cmd->device->host->max_cmd_len));
1720                 cmd->result = (DID_ABORT << 16);
1721                 goto done;
1722         }
1723
1724         if (unlikely(host->shost_state == SHOST_DEL)) {
1725                 cmd->result = (DID_NO_CONNECT << 16);
1726                 goto done;
1727
1728         }
1729
1730         trace_scsi_dispatch_cmd_start(cmd);
1731         rtn = host->hostt->queuecommand(host, cmd);
1732         if (rtn) {
1733                 trace_scsi_dispatch_cmd_error(cmd, rtn);
1734                 if (rtn != SCSI_MLQUEUE_DEVICE_BUSY &&
1735                     rtn != SCSI_MLQUEUE_TARGET_BUSY)
1736                         rtn = SCSI_MLQUEUE_HOST_BUSY;
1737
1738                 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1739                         "queuecommand : request rejected\n"));
1740         }
1741
1742         return rtn;
1743  done:
1744         cmd->scsi_done(cmd);
1745         return 0;
1746 }
1747
1748 /**
1749  * scsi_done - Invoke completion on finished SCSI command.
1750  * @cmd: The SCSI Command for which a low-level device driver (LLDD) gives
1751  * ownership back to SCSI Core -- i.e. the LLDD has finished with it.
1752  *
1753  * Description: This function is the mid-level's (SCSI Core) interrupt routine,
1754  * which regains ownership of the SCSI command (de facto) from a LLDD, and
1755  * calls blk_complete_request() for further processing.
1756  *
1757  * This function is interrupt context safe.
1758  */
1759 static void scsi_done(struct scsi_cmnd *cmd)
1760 {
1761         trace_scsi_dispatch_cmd_done(cmd);
1762         blk_complete_request(cmd->request);
1763 }
1764
1765 /*
1766  * Function:    scsi_request_fn()
1767  *
1768  * Purpose:     Main strategy routine for SCSI.
1769  *
1770  * Arguments:   q       - Pointer to actual queue.
1771  *
1772  * Returns:     Nothing
1773  *
1774  * Lock status: request queue lock assumed to be held when called.
1775  *
1776  * Note: See sd_zbc.c sd_zbc_write_lock_zone() for write order
1777  * protection for ZBC disks.
1778  */
1779 static void scsi_request_fn(struct request_queue *q)
1780         __releases(q->queue_lock)
1781         __acquires(q->queue_lock)
1782 {
1783         struct scsi_device *sdev = q->queuedata;
1784         struct Scsi_Host *shost;
1785         struct scsi_cmnd *cmd;
1786         struct request *req;
1787
1788         /*
1789          * To start with, we keep looping until the queue is empty, or until
1790          * the host is no longer able to accept any more requests.
1791          */
1792         shost = sdev->host;
1793         for (;;) {
1794                 int rtn;
1795                 /*
1796                  * get next queueable request.  We do this early to make sure
1797                  * that the request is fully prepared even if we cannot
1798                  * accept it.
1799                  */
1800                 req = blk_peek_request(q);
1801                 if (!req)
1802                         break;
1803
1804                 if (unlikely(!scsi_device_online(sdev))) {
1805                         sdev_printk(KERN_ERR, sdev,
1806                                     "rejecting I/O to offline device\n");
1807                         scsi_kill_request(req, q);
1808                         continue;
1809                 }
1810
1811                 if (!scsi_dev_queue_ready(q, sdev))
1812                         break;
1813
1814                 /*
1815                  * Remove the request from the request list.
1816                  */
1817                 if (!(blk_queue_tagged(q) && !blk_queue_start_tag(q, req)))
1818                         blk_start_request(req);
1819
1820                 spin_unlock_irq(q->queue_lock);
1821                 cmd = blk_mq_rq_to_pdu(req);
1822                 if (cmd != req->special) {
1823                         printk(KERN_CRIT "impossible request in %s.\n"
1824                                          "please mail a stack trace to "
1825                                          "linux-scsi@vger.kernel.org\n",
1826                                          __func__);
1827                         blk_dump_rq_flags(req, "foo");
1828                         BUG();
1829                 }
1830
1831                 /*
1832                  * We hit this when the driver is using a host wide
1833                  * tag map. For device level tag maps the queue_depth check
1834                  * in the device ready fn would prevent us from trying
1835                  * to allocate a tag. Since the map is a shared host resource
1836                  * we add the dev to the starved list so it eventually gets
1837                  * a run when a tag is freed.
1838                  */
1839                 if (blk_queue_tagged(q) && !(req->rq_flags & RQF_QUEUED)) {
1840                         spin_lock_irq(shost->host_lock);
1841                         if (list_empty(&sdev->starved_entry))
1842                                 list_add_tail(&sdev->starved_entry,
1843                                               &shost->starved_list);
1844                         spin_unlock_irq(shost->host_lock);
1845                         goto not_ready;
1846                 }
1847
1848                 if (!scsi_target_queue_ready(shost, sdev))
1849                         goto not_ready;
1850
1851                 if (!scsi_host_queue_ready(q, shost, sdev))
1852                         goto host_not_ready;
1853         
1854                 if (sdev->simple_tags)
1855                         cmd->flags |= SCMD_TAGGED;
1856                 else
1857                         cmd->flags &= ~SCMD_TAGGED;
1858
1859                 /*
1860                  * Finally, initialize any error handling parameters, and set up
1861                  * the timers for timeouts.
1862                  */
1863                 scsi_init_cmd_errh(cmd);
1864
1865                 /*
1866                  * Dispatch the command to the low-level driver.
1867                  */
1868                 cmd->scsi_done = scsi_done;
1869                 rtn = scsi_dispatch_cmd(cmd);
1870                 if (rtn) {
1871                         scsi_queue_insert(cmd, rtn);
1872                         spin_lock_irq(q->queue_lock);
1873                         goto out_delay;
1874                 }
1875                 spin_lock_irq(q->queue_lock);
1876         }
1877
1878         return;
1879
1880  host_not_ready:
1881         if (scsi_target(sdev)->can_queue > 0)
1882                 atomic_dec(&scsi_target(sdev)->target_busy);
1883  not_ready:
1884         /*
1885          * lock q, handle tag, requeue req, and decrement device_busy. We
1886          * must return with queue_lock held.
1887          *
1888          * Decrementing device_busy without checking it is OK, as all such
1889          * cases (host limits or settings) should run the queue at some
1890          * later time.
1891          */
1892         spin_lock_irq(q->queue_lock);
1893         blk_requeue_request(q, req);
1894         atomic_dec(&sdev->device_busy);
1895 out_delay:
1896         if (!atomic_read(&sdev->device_busy) && !scsi_device_blocked(sdev))
1897                 blk_delay_queue(q, SCSI_QUEUE_DELAY);
1898 }
1899
1900 static inline blk_status_t prep_to_mq(int ret)
1901 {
1902         switch (ret) {
1903         case BLKPREP_OK:
1904                 return BLK_STS_OK;
1905         case BLKPREP_DEFER:
1906                 return BLK_STS_RESOURCE;
1907         default:
1908                 return BLK_STS_IOERR;
1909         }
1910 }
1911
1912 /* Size in bytes of the sg-list stored in the scsi-mq command-private data. */
1913 static unsigned int scsi_mq_sgl_size(struct Scsi_Host *shost)
1914 {
1915         return min_t(unsigned int, shost->sg_tablesize, SG_CHUNK_SIZE) *
1916                 sizeof(struct scatterlist);
1917 }
1918
1919 static int scsi_mq_prep_fn(struct request *req)
1920 {
1921         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1922         struct scsi_device *sdev = req->q->queuedata;
1923         struct Scsi_Host *shost = sdev->host;
1924         struct scatterlist *sg;
1925
1926         scsi_init_command(sdev, cmd);
1927
1928         req->special = cmd;
1929
1930         cmd->request = req;
1931
1932         cmd->tag = req->tag;
1933         cmd->prot_op = SCSI_PROT_NORMAL;
1934
1935         sg = (void *)cmd + sizeof(struct scsi_cmnd) + shost->hostt->cmd_size;
1936         cmd->sdb.table.sgl = sg;
1937
1938         if (scsi_host_get_prot(shost)) {
1939                 memset(cmd->prot_sdb, 0, sizeof(struct scsi_data_buffer));
1940
1941                 cmd->prot_sdb->table.sgl =
1942                         (struct scatterlist *)(cmd->prot_sdb + 1);
1943         }
1944
1945         if (blk_bidi_rq(req)) {
1946                 struct request *next_rq = req->next_rq;
1947                 struct scsi_data_buffer *bidi_sdb = blk_mq_rq_to_pdu(next_rq);
1948
1949                 memset(bidi_sdb, 0, sizeof(struct scsi_data_buffer));
1950                 bidi_sdb->table.sgl =
1951                         (struct scatterlist *)(bidi_sdb + 1);
1952
1953                 next_rq->special = bidi_sdb;
1954         }
1955
1956         blk_mq_start_request(req);
1957
1958         return scsi_setup_cmnd(sdev, req);
1959 }
1960
1961 static void scsi_mq_done(struct scsi_cmnd *cmd)
1962 {
1963         trace_scsi_dispatch_cmd_done(cmd);
1964         blk_mq_complete_request(cmd->request);
1965 }
1966
1967 static void scsi_mq_put_budget(struct blk_mq_hw_ctx *hctx)
1968 {
1969         struct request_queue *q = hctx->queue;
1970         struct scsi_device *sdev = q->queuedata;
1971
1972         atomic_dec(&sdev->device_busy);
1973         put_device(&sdev->sdev_gendev);
1974 }
1975
1976 static bool scsi_mq_get_budget(struct blk_mq_hw_ctx *hctx)
1977 {
1978         struct request_queue *q = hctx->queue;
1979         struct scsi_device *sdev = q->queuedata;
1980
1981         if (!get_device(&sdev->sdev_gendev))
1982                 goto out;
1983         if (!scsi_dev_queue_ready(q, sdev))
1984                 goto out_put_device;
1985
1986         return true;
1987
1988 out_put_device:
1989         put_device(&sdev->sdev_gendev);
1990 out:
1991         if (atomic_read(&sdev->device_busy) == 0 && !scsi_device_blocked(sdev))
1992                 blk_mq_delay_run_hw_queue(hctx, SCSI_QUEUE_DELAY);
1993         return false;
1994 }
1995
1996 static blk_status_t scsi_queue_rq(struct blk_mq_hw_ctx *hctx,
1997                          const struct blk_mq_queue_data *bd)
1998 {
1999         struct request *req = bd->rq;
2000         struct request_queue *q = req->q;
2001         struct scsi_device *sdev = q->queuedata;
2002         struct Scsi_Host *shost = sdev->host;
2003         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
2004         blk_status_t ret;
2005         int reason;
2006
2007         ret = prep_to_mq(scsi_prep_state_check(sdev, req));
2008         if (ret != BLK_STS_OK)
2009                 goto out_put_budget;
2010
2011         ret = BLK_STS_RESOURCE;
2012         if (!scsi_target_queue_ready(shost, sdev))
2013                 goto out_put_budget;
2014         if (!scsi_host_queue_ready(q, shost, sdev))
2015                 goto out_dec_target_busy;
2016
2017         if (!(req->rq_flags & RQF_DONTPREP)) {
2018                 ret = prep_to_mq(scsi_mq_prep_fn(req));
2019                 if (ret != BLK_STS_OK)
2020                         goto out_dec_host_busy;
2021                 req->rq_flags |= RQF_DONTPREP;
2022         } else {
2023                 blk_mq_start_request(req);
2024         }
2025
2026         if (sdev->simple_tags)
2027                 cmd->flags |= SCMD_TAGGED;
2028         else
2029                 cmd->flags &= ~SCMD_TAGGED;
2030
2031         scsi_init_cmd_errh(cmd);
2032         cmd->scsi_done = scsi_mq_done;
2033
2034         reason = scsi_dispatch_cmd(cmd);
2035         if (reason) {
2036                 scsi_set_blocked(cmd, reason);
2037                 ret = BLK_STS_RESOURCE;
2038                 goto out_dec_host_busy;
2039         }
2040
2041         return BLK_STS_OK;
2042
2043 out_dec_host_busy:
2044         scsi_dec_host_busy(shost);
2045 out_dec_target_busy:
2046         if (scsi_target(sdev)->can_queue > 0)
2047                 atomic_dec(&scsi_target(sdev)->target_busy);
2048 out_put_budget:
2049         scsi_mq_put_budget(hctx);
2050         switch (ret) {
2051         case BLK_STS_OK:
2052                 break;
2053         case BLK_STS_RESOURCE:
2054                 if (atomic_read(&sdev->device_busy) ||
2055                     scsi_device_blocked(sdev))
2056                         ret = BLK_STS_DEV_RESOURCE;
2057                 break;
2058         default:
2059                 /*
2060                  * Make sure to release all allocated ressources when
2061                  * we hit an error, as we will never see this command
2062                  * again.
2063                  */
2064                 if (req->rq_flags & RQF_DONTPREP)
2065                         scsi_mq_uninit_cmd(cmd);
2066                 break;
2067         }
2068         return ret;
2069 }
2070
2071 static enum blk_eh_timer_return scsi_timeout(struct request *req,
2072                 bool reserved)
2073 {
2074         if (reserved)
2075                 return BLK_EH_RESET_TIMER;
2076         return scsi_times_out(req);
2077 }
2078
2079 static int scsi_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
2080                                 unsigned int hctx_idx, unsigned int numa_node)
2081 {
2082         struct Scsi_Host *shost = set->driver_data;
2083         const bool unchecked_isa_dma = shost->unchecked_isa_dma;
2084         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
2085         struct scatterlist *sg;
2086
2087         if (unchecked_isa_dma)
2088                 cmd->flags |= SCMD_UNCHECKED_ISA_DMA;
2089         cmd->sense_buffer = scsi_alloc_sense_buffer(unchecked_isa_dma,
2090                                                     GFP_KERNEL, numa_node);
2091         if (!cmd->sense_buffer)
2092                 return -ENOMEM;
2093         cmd->req.sense = cmd->sense_buffer;
2094
2095         if (scsi_host_get_prot(shost)) {
2096                 sg = (void *)cmd + sizeof(struct scsi_cmnd) +
2097                         shost->hostt->cmd_size;
2098                 cmd->prot_sdb = (void *)sg + scsi_mq_sgl_size(shost);
2099         }
2100
2101         return 0;
2102 }
2103
2104 static void scsi_mq_exit_request(struct blk_mq_tag_set *set, struct request *rq,
2105                                  unsigned int hctx_idx)
2106 {
2107         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
2108
2109         scsi_free_sense_buffer(cmd->flags & SCMD_UNCHECKED_ISA_DMA,
2110                                cmd->sense_buffer);
2111 }
2112
2113 static int scsi_map_queues(struct blk_mq_tag_set *set)
2114 {
2115         struct Scsi_Host *shost = container_of(set, struct Scsi_Host, tag_set);
2116
2117         if (shost->hostt->map_queues)
2118                 return shost->hostt->map_queues(shost);
2119         return blk_mq_map_queues(set);
2120 }
2121
2122 static u64 scsi_calculate_bounce_limit(struct Scsi_Host *shost)
2123 {
2124         struct device *host_dev;
2125         u64 bounce_limit = 0xffffffff;
2126
2127         if (shost->unchecked_isa_dma)
2128                 return BLK_BOUNCE_ISA;
2129         /*
2130          * Platforms with virtual-DMA translation
2131          * hardware have no practical limit.
2132          */
2133         if (!PCI_DMA_BUS_IS_PHYS)
2134                 return BLK_BOUNCE_ANY;
2135
2136         host_dev = scsi_get_device(shost);
2137         if (host_dev && host_dev->dma_mask)
2138                 bounce_limit = (u64)dma_max_pfn(host_dev) << PAGE_SHIFT;
2139
2140         return bounce_limit;
2141 }
2142
2143 void __scsi_init_queue(struct Scsi_Host *shost, struct request_queue *q)
2144 {
2145         struct device *dev = shost->dma_dev;
2146
2147         queue_flag_set_unlocked(QUEUE_FLAG_SCSI_PASSTHROUGH, q);
2148
2149         /*
2150          * this limit is imposed by hardware restrictions
2151          */
2152         blk_queue_max_segments(q, min_t(unsigned short, shost->sg_tablesize,
2153                                         SG_MAX_SEGMENTS));
2154
2155         if (scsi_host_prot_dma(shost)) {
2156                 shost->sg_prot_tablesize =
2157                         min_not_zero(shost->sg_prot_tablesize,
2158                                      (unsigned short)SCSI_MAX_PROT_SG_SEGMENTS);
2159                 BUG_ON(shost->sg_prot_tablesize < shost->sg_tablesize);
2160                 blk_queue_max_integrity_segments(q, shost->sg_prot_tablesize);
2161         }
2162
2163         blk_queue_max_hw_sectors(q, shost->max_sectors);
2164         blk_queue_bounce_limit(q, scsi_calculate_bounce_limit(shost));
2165         blk_queue_segment_boundary(q, shost->dma_boundary);
2166         dma_set_seg_boundary(dev, shost->dma_boundary);
2167
2168         blk_queue_max_segment_size(q, dma_get_max_seg_size(dev));
2169
2170         if (!shost->use_clustering)
2171                 q->limits.cluster = 0;
2172
2173         /*
2174          * Set a reasonable default alignment:  The larger of 32-byte (dword),
2175          * which is a common minimum for HBAs, and the minimum DMA alignment,
2176          * which is set by the platform.
2177          *
2178          * Devices that require a bigger alignment can increase it later.
2179          */
2180         blk_queue_dma_alignment(q, max(4, dma_get_cache_alignment()) - 1);
2181 }
2182 EXPORT_SYMBOL_GPL(__scsi_init_queue);
2183
2184 static int scsi_old_init_rq(struct request_queue *q, struct request *rq,
2185                             gfp_t gfp)
2186 {
2187         struct Scsi_Host *shost = q->rq_alloc_data;
2188         const bool unchecked_isa_dma = shost->unchecked_isa_dma;
2189         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
2190
2191         memset(cmd, 0, sizeof(*cmd));
2192
2193         if (unchecked_isa_dma)
2194                 cmd->flags |= SCMD_UNCHECKED_ISA_DMA;
2195         cmd->sense_buffer = scsi_alloc_sense_buffer(unchecked_isa_dma, gfp,
2196                                                     NUMA_NO_NODE);
2197         if (!cmd->sense_buffer)
2198                 goto fail;
2199         cmd->req.sense = cmd->sense_buffer;
2200
2201         if (scsi_host_get_prot(shost) >= SHOST_DIX_TYPE0_PROTECTION) {
2202                 cmd->prot_sdb = kmem_cache_zalloc(scsi_sdb_cache, gfp);
2203                 if (!cmd->prot_sdb)
2204                         goto fail_free_sense;
2205         }
2206
2207         return 0;
2208
2209 fail_free_sense:
2210         scsi_free_sense_buffer(unchecked_isa_dma, cmd->sense_buffer);
2211 fail:
2212         return -ENOMEM;
2213 }
2214
2215 static void scsi_old_exit_rq(struct request_queue *q, struct request *rq)
2216 {
2217         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
2218
2219         if (cmd->prot_sdb)
2220                 kmem_cache_free(scsi_sdb_cache, cmd->prot_sdb);
2221         scsi_free_sense_buffer(cmd->flags & SCMD_UNCHECKED_ISA_DMA,
2222                                cmd->sense_buffer);
2223 }
2224
2225 struct request_queue *scsi_old_alloc_queue(struct scsi_device *sdev)
2226 {
2227         struct Scsi_Host *shost = sdev->host;
2228         struct request_queue *q;
2229
2230         q = blk_alloc_queue_node(GFP_KERNEL, NUMA_NO_NODE);
2231         if (!q)
2232                 return NULL;
2233         q->cmd_size = sizeof(struct scsi_cmnd) + shost->hostt->cmd_size;
2234         q->rq_alloc_data = shost;
2235         q->request_fn = scsi_request_fn;
2236         q->init_rq_fn = scsi_old_init_rq;
2237         q->exit_rq_fn = scsi_old_exit_rq;
2238         q->initialize_rq_fn = scsi_initialize_rq;
2239
2240         if (blk_init_allocated_queue(q) < 0) {
2241                 blk_cleanup_queue(q);
2242                 return NULL;
2243         }
2244
2245         __scsi_init_queue(shost, q);
2246         blk_queue_prep_rq(q, scsi_prep_fn);
2247         blk_queue_unprep_rq(q, scsi_unprep_fn);
2248         blk_queue_softirq_done(q, scsi_softirq_done);
2249         blk_queue_rq_timed_out(q, scsi_times_out);
2250         blk_queue_lld_busy(q, scsi_lld_busy);
2251         return q;
2252 }
2253
2254 static const struct blk_mq_ops scsi_mq_ops = {
2255         .get_budget     = scsi_mq_get_budget,
2256         .put_budget     = scsi_mq_put_budget,
2257         .queue_rq       = scsi_queue_rq,
2258         .complete       = scsi_softirq_done,
2259         .timeout        = scsi_timeout,
2260 #ifdef CONFIG_BLK_DEBUG_FS
2261         .show_rq        = scsi_show_rq,
2262 #endif
2263         .init_request   = scsi_mq_init_request,
2264         .exit_request   = scsi_mq_exit_request,
2265         .initialize_rq_fn = scsi_initialize_rq,
2266         .map_queues     = scsi_map_queues,
2267 };
2268
2269 struct request_queue *scsi_mq_alloc_queue(struct scsi_device *sdev)
2270 {
2271         sdev->request_queue = blk_mq_init_queue(&sdev->host->tag_set);
2272         if (IS_ERR(sdev->request_queue))
2273                 return NULL;
2274
2275         sdev->request_queue->queuedata = sdev;
2276         __scsi_init_queue(sdev->host, sdev->request_queue);
2277         return sdev->request_queue;
2278 }
2279
2280 int scsi_mq_setup_tags(struct Scsi_Host *shost)
2281 {
2282         unsigned int cmd_size, sgl_size;
2283
2284         sgl_size = scsi_mq_sgl_size(shost);
2285         cmd_size = sizeof(struct scsi_cmnd) + shost->hostt->cmd_size + sgl_size;
2286         if (scsi_host_get_prot(shost))
2287                 cmd_size += sizeof(struct scsi_data_buffer) + sgl_size;
2288
2289         memset(&shost->tag_set, 0, sizeof(shost->tag_set));
2290         shost->tag_set.ops = &scsi_mq_ops;
2291         shost->tag_set.nr_hw_queues = shost->nr_hw_queues ? : 1;
2292         shost->tag_set.queue_depth = shost->can_queue;
2293         shost->tag_set.cmd_size = cmd_size;
2294         shost->tag_set.numa_node = NUMA_NO_NODE;
2295         shost->tag_set.flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_SG_MERGE;
2296         shost->tag_set.flags |=
2297                 BLK_ALLOC_POLICY_TO_MQ_FLAG(shost->hostt->tag_alloc_policy);
2298         shost->tag_set.driver_data = shost;
2299
2300         return blk_mq_alloc_tag_set(&shost->tag_set);
2301 }
2302
2303 void scsi_mq_destroy_tags(struct Scsi_Host *shost)
2304 {
2305         blk_mq_free_tag_set(&shost->tag_set);
2306 }
2307
2308 /**
2309  * scsi_device_from_queue - return sdev associated with a request_queue
2310  * @q: The request queue to return the sdev from
2311  *
2312  * Return the sdev associated with a request queue or NULL if the
2313  * request_queue does not reference a SCSI device.
2314  */
2315 struct scsi_device *scsi_device_from_queue(struct request_queue *q)
2316 {
2317         struct scsi_device *sdev = NULL;
2318
2319         if (q->mq_ops) {
2320                 if (q->mq_ops == &scsi_mq_ops)
2321                         sdev = q->queuedata;
2322         } else if (q->request_fn == scsi_request_fn)
2323                 sdev = q->queuedata;
2324         if (!sdev || !get_device(&sdev->sdev_gendev))
2325                 sdev = NULL;
2326
2327         return sdev;
2328 }
2329 EXPORT_SYMBOL_GPL(scsi_device_from_queue);
2330
2331 /*
2332  * Function:    scsi_block_requests()
2333  *
2334  * Purpose:     Utility function used by low-level drivers to prevent further
2335  *              commands from being queued to the device.
2336  *
2337  * Arguments:   shost       - Host in question
2338  *
2339  * Returns:     Nothing
2340  *
2341  * Lock status: No locks are assumed held.
2342  *
2343  * Notes:       There is no timer nor any other means by which the requests
2344  *              get unblocked other than the low-level driver calling
2345  *              scsi_unblock_requests().
2346  */
2347 void scsi_block_requests(struct Scsi_Host *shost)
2348 {
2349         shost->host_self_blocked = 1;
2350 }
2351 EXPORT_SYMBOL(scsi_block_requests);
2352
2353 /*
2354  * Function:    scsi_unblock_requests()
2355  *
2356  * Purpose:     Utility function used by low-level drivers to allow further
2357  *              commands from being queued to the device.
2358  *
2359  * Arguments:   shost       - Host in question
2360  *
2361  * Returns:     Nothing
2362  *
2363  * Lock status: No locks are assumed held.
2364  *
2365  * Notes:       There is no timer nor any other means by which the requests
2366  *              get unblocked other than the low-level driver calling
2367  *              scsi_unblock_requests().
2368  *
2369  *              This is done as an API function so that changes to the
2370  *              internals of the scsi mid-layer won't require wholesale
2371  *              changes to drivers that use this feature.
2372  */
2373 void scsi_unblock_requests(struct Scsi_Host *shost)
2374 {
2375         shost->host_self_blocked = 0;
2376         scsi_run_host_queues(shost);
2377 }
2378 EXPORT_SYMBOL(scsi_unblock_requests);
2379
2380 int __init scsi_init_queue(void)
2381 {
2382         scsi_sdb_cache = kmem_cache_create("scsi_data_buffer",
2383                                            sizeof(struct scsi_data_buffer),
2384                                            0, 0, NULL);
2385         if (!scsi_sdb_cache) {
2386                 printk(KERN_ERR "SCSI: can't init scsi sdb cache\n");
2387                 return -ENOMEM;
2388         }
2389
2390         return 0;
2391 }
2392
2393 void scsi_exit_queue(void)
2394 {
2395         kmem_cache_destroy(scsi_sense_cache);
2396         kmem_cache_destroy(scsi_sense_isadma_cache);
2397         kmem_cache_destroy(scsi_sdb_cache);
2398 }
2399
2400 /**
2401  *      scsi_mode_select - issue a mode select
2402  *      @sdev:  SCSI device to be queried
2403  *      @pf:    Page format bit (1 == standard, 0 == vendor specific)
2404  *      @sp:    Save page bit (0 == don't save, 1 == save)
2405  *      @modepage: mode page being requested
2406  *      @buffer: request buffer (may not be smaller than eight bytes)
2407  *      @len:   length of request buffer.
2408  *      @timeout: command timeout
2409  *      @retries: number of retries before failing
2410  *      @data: returns a structure abstracting the mode header data
2411  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
2412  *              must be SCSI_SENSE_BUFFERSIZE big.
2413  *
2414  *      Returns zero if successful; negative error number or scsi
2415  *      status on error
2416  *
2417  */
2418 int
2419 scsi_mode_select(struct scsi_device *sdev, int pf, int sp, int modepage,
2420                  unsigned char *buffer, int len, int timeout, int retries,
2421                  struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
2422 {
2423         unsigned char cmd[10];
2424         unsigned char *real_buffer;
2425         int ret;
2426
2427         memset(cmd, 0, sizeof(cmd));
2428         cmd[1] = (pf ? 0x10 : 0) | (sp ? 0x01 : 0);
2429
2430         if (sdev->use_10_for_ms) {
2431                 if (len > 65535)
2432                         return -EINVAL;
2433                 real_buffer = kmalloc(8 + len, GFP_KERNEL);
2434                 if (!real_buffer)
2435                         return -ENOMEM;
2436                 memcpy(real_buffer + 8, buffer, len);
2437                 len += 8;
2438                 real_buffer[0] = 0;
2439                 real_buffer[1] = 0;
2440                 real_buffer[2] = data->medium_type;
2441                 real_buffer[3] = data->device_specific;
2442                 real_buffer[4] = data->longlba ? 0x01 : 0;
2443                 real_buffer[5] = 0;
2444                 real_buffer[6] = data->block_descriptor_length >> 8;
2445                 real_buffer[7] = data->block_descriptor_length;
2446
2447                 cmd[0] = MODE_SELECT_10;
2448                 cmd[7] = len >> 8;
2449                 cmd[8] = len;
2450         } else {
2451                 if (len > 255 || data->block_descriptor_length > 255 ||
2452                     data->longlba)
2453                         return -EINVAL;
2454
2455                 real_buffer = kmalloc(4 + len, GFP_KERNEL);
2456                 if (!real_buffer)
2457                         return -ENOMEM;
2458                 memcpy(real_buffer + 4, buffer, len);
2459                 len += 4;
2460                 real_buffer[0] = 0;
2461                 real_buffer[1] = data->medium_type;
2462                 real_buffer[2] = data->device_specific;
2463                 real_buffer[3] = data->block_descriptor_length;
2464                 
2465
2466                 cmd[0] = MODE_SELECT;
2467                 cmd[4] = len;
2468         }
2469
2470         ret = scsi_execute_req(sdev, cmd, DMA_TO_DEVICE, real_buffer, len,
2471                                sshdr, timeout, retries, NULL);
2472         kfree(real_buffer);
2473         return ret;
2474 }
2475 EXPORT_SYMBOL_GPL(scsi_mode_select);
2476
2477 /**
2478  *      scsi_mode_sense - issue a mode sense, falling back from 10 to six bytes if necessary.
2479  *      @sdev:  SCSI device to be queried
2480  *      @dbd:   set if mode sense will allow block descriptors to be returned
2481  *      @modepage: mode page being requested
2482  *      @buffer: request buffer (may not be smaller than eight bytes)
2483  *      @len:   length of request buffer.
2484  *      @timeout: command timeout
2485  *      @retries: number of retries before failing
2486  *      @data: returns a structure abstracting the mode header data
2487  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
2488  *              must be SCSI_SENSE_BUFFERSIZE big.
2489  *
2490  *      Returns zero if unsuccessful, or the header offset (either 4
2491  *      or 8 depending on whether a six or ten byte command was
2492  *      issued) if successful.
2493  */
2494 int
2495 scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
2496                   unsigned char *buffer, int len, int timeout, int retries,
2497                   struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
2498 {
2499         unsigned char cmd[12];
2500         int use_10_for_ms;
2501         int header_length;
2502         int result, retry_count = retries;
2503         struct scsi_sense_hdr my_sshdr;
2504
2505         memset(data, 0, sizeof(*data));
2506         memset(&cmd[0], 0, 12);
2507         cmd[1] = dbd & 0x18;    /* allows DBD and LLBA bits */
2508         cmd[2] = modepage;
2509
2510         /* caller might not be interested in sense, but we need it */
2511         if (!sshdr)
2512                 sshdr = &my_sshdr;
2513
2514  retry:
2515         use_10_for_ms = sdev->use_10_for_ms;
2516
2517         if (use_10_for_ms) {
2518                 if (len < 8)
2519                         len = 8;
2520
2521                 cmd[0] = MODE_SENSE_10;
2522                 cmd[8] = len;
2523                 header_length = 8;
2524         } else {
2525                 if (len < 4)
2526                         len = 4;
2527
2528                 cmd[0] = MODE_SENSE;
2529                 cmd[4] = len;
2530                 header_length = 4;
2531         }
2532
2533         memset(buffer, 0, len);
2534
2535         result = scsi_execute_req(sdev, cmd, DMA_FROM_DEVICE, buffer, len,
2536                                   sshdr, timeout, retries, NULL);
2537
2538         /* This code looks awful: what it's doing is making sure an
2539          * ILLEGAL REQUEST sense return identifies the actual command
2540          * byte as the problem.  MODE_SENSE commands can return
2541          * ILLEGAL REQUEST if the code page isn't supported */
2542
2543         if (use_10_for_ms && !scsi_status_is_good(result) &&
2544             (driver_byte(result) & DRIVER_SENSE)) {
2545                 if (scsi_sense_valid(sshdr)) {
2546                         if ((sshdr->sense_key == ILLEGAL_REQUEST) &&
2547                             (sshdr->asc == 0x20) && (sshdr->ascq == 0)) {
2548                                 /* 
2549                                  * Invalid command operation code
2550                                  */
2551                                 sdev->use_10_for_ms = 0;
2552                                 goto retry;
2553                         }
2554                 }
2555         }
2556
2557         if(scsi_status_is_good(result)) {
2558                 if (unlikely(buffer[0] == 0x86 && buffer[1] == 0x0b &&
2559                              (modepage == 6 || modepage == 8))) {
2560                         /* Initio breakage? */
2561                         header_length = 0;
2562                         data->length = 13;
2563                         data->medium_type = 0;
2564                         data->device_specific = 0;
2565                         data->longlba = 0;
2566                         data->block_descriptor_length = 0;
2567                 } else if(use_10_for_ms) {
2568                         data->length = buffer[0]*256 + buffer[1] + 2;
2569                         data->medium_type = buffer[2];
2570                         data->device_specific = buffer[3];
2571                         data->longlba = buffer[4] & 0x01;
2572                         data->block_descriptor_length = buffer[6]*256
2573                                 + buffer[7];
2574                 } else {
2575                         data->length = buffer[0] + 1;
2576                         data->medium_type = buffer[1];
2577                         data->device_specific = buffer[2];
2578                         data->block_descriptor_length = buffer[3];
2579                 }
2580                 data->header_length = header_length;
2581         } else if ((status_byte(result) == CHECK_CONDITION) &&
2582                    scsi_sense_valid(sshdr) &&
2583                    sshdr->sense_key == UNIT_ATTENTION && retry_count) {
2584                 retry_count--;
2585                 goto retry;
2586         }
2587
2588         return result;
2589 }
2590 EXPORT_SYMBOL(scsi_mode_sense);
2591
2592 /**
2593  *      scsi_test_unit_ready - test if unit is ready
2594  *      @sdev:  scsi device to change the state of.
2595  *      @timeout: command timeout
2596  *      @retries: number of retries before failing
2597  *      @sshdr: outpout pointer for decoded sense information.
2598  *
2599  *      Returns zero if unsuccessful or an error if TUR failed.  For
2600  *      removable media, UNIT_ATTENTION sets ->changed flag.
2601  **/
2602 int
2603 scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries,
2604                      struct scsi_sense_hdr *sshdr)
2605 {
2606         char cmd[] = {
2607                 TEST_UNIT_READY, 0, 0, 0, 0, 0,
2608         };
2609         int result;
2610
2611         /* try to eat the UNIT_ATTENTION if there are enough retries */
2612         do {
2613                 result = scsi_execute_req(sdev, cmd, DMA_NONE, NULL, 0, sshdr,
2614                                           timeout, retries, NULL);
2615                 if (sdev->removable && scsi_sense_valid(sshdr) &&
2616                     sshdr->sense_key == UNIT_ATTENTION)
2617                         sdev->changed = 1;
2618         } while (scsi_sense_valid(sshdr) &&
2619                  sshdr->sense_key == UNIT_ATTENTION && --retries);
2620
2621         return result;
2622 }
2623 EXPORT_SYMBOL(scsi_test_unit_ready);
2624
2625 /**
2626  *      scsi_device_set_state - Take the given device through the device state model.
2627  *      @sdev:  scsi device to change the state of.
2628  *      @state: state to change to.
2629  *
2630  *      Returns zero if successful or an error if the requested
2631  *      transition is illegal.
2632  */
2633 int
2634 scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
2635 {
2636         enum scsi_device_state oldstate = sdev->sdev_state;
2637
2638         if (state == oldstate)
2639                 return 0;
2640
2641         switch (state) {
2642         case SDEV_CREATED:
2643                 switch (oldstate) {
2644                 case SDEV_CREATED_BLOCK:
2645                         break;
2646                 default:
2647                         goto illegal;
2648                 }
2649                 break;
2650                         
2651         case SDEV_RUNNING:
2652                 switch (oldstate) {
2653                 case SDEV_CREATED:
2654                 case SDEV_OFFLINE:
2655                 case SDEV_TRANSPORT_OFFLINE:
2656                 case SDEV_QUIESCE:
2657                 case SDEV_BLOCK:
2658                         break;
2659                 default:
2660                         goto illegal;
2661                 }
2662                 break;
2663
2664         case SDEV_QUIESCE:
2665                 switch (oldstate) {
2666                 case SDEV_RUNNING:
2667                 case SDEV_OFFLINE:
2668                 case SDEV_TRANSPORT_OFFLINE:
2669                         break;
2670                 default:
2671                         goto illegal;
2672                 }
2673                 break;
2674
2675         case SDEV_OFFLINE:
2676         case SDEV_TRANSPORT_OFFLINE:
2677                 switch (oldstate) {
2678                 case SDEV_CREATED:
2679                 case SDEV_RUNNING:
2680                 case SDEV_QUIESCE:
2681                 case SDEV_BLOCK:
2682                         break;
2683                 default:
2684                         goto illegal;
2685                 }
2686                 break;
2687
2688         case SDEV_BLOCK:
2689                 switch (oldstate) {
2690                 case SDEV_RUNNING:
2691                 case SDEV_CREATED_BLOCK:
2692                         break;
2693                 default:
2694                         goto illegal;
2695                 }
2696                 break;
2697
2698         case SDEV_CREATED_BLOCK:
2699                 switch (oldstate) {
2700                 case SDEV_CREATED:
2701                         break;
2702                 default:
2703                         goto illegal;
2704                 }
2705                 break;
2706
2707         case SDEV_CANCEL:
2708                 switch (oldstate) {
2709                 case SDEV_CREATED:
2710                 case SDEV_RUNNING:
2711                 case SDEV_QUIESCE:
2712                 case SDEV_OFFLINE:
2713                 case SDEV_TRANSPORT_OFFLINE:
2714                         break;
2715                 default:
2716                         goto illegal;
2717                 }
2718                 break;
2719
2720         case SDEV_DEL:
2721                 switch (oldstate) {
2722                 case SDEV_CREATED:
2723                 case SDEV_RUNNING:
2724                 case SDEV_OFFLINE:
2725                 case SDEV_TRANSPORT_OFFLINE:
2726                 case SDEV_CANCEL:
2727                 case SDEV_BLOCK:
2728                 case SDEV_CREATED_BLOCK:
2729                         break;
2730                 default:
2731                         goto illegal;
2732                 }
2733                 break;
2734
2735         }
2736         sdev->sdev_state = state;
2737         return 0;
2738
2739  illegal:
2740         SCSI_LOG_ERROR_RECOVERY(1,
2741                                 sdev_printk(KERN_ERR, sdev,
2742                                             "Illegal state transition %s->%s",
2743                                             scsi_device_state_name(oldstate),
2744                                             scsi_device_state_name(state))
2745                                 );
2746         return -EINVAL;
2747 }
2748 EXPORT_SYMBOL(scsi_device_set_state);
2749
2750 /**
2751  *      sdev_evt_emit - emit a single SCSI device uevent
2752  *      @sdev: associated SCSI device
2753  *      @evt: event to emit
2754  *
2755  *      Send a single uevent (scsi_event) to the associated scsi_device.
2756  */
2757 static void scsi_evt_emit(struct scsi_device *sdev, struct scsi_event *evt)
2758 {
2759         int idx = 0;
2760         char *envp[3];
2761
2762         switch (evt->evt_type) {
2763         case SDEV_EVT_MEDIA_CHANGE:
2764                 envp[idx++] = "SDEV_MEDIA_CHANGE=1";
2765                 break;
2766         case SDEV_EVT_INQUIRY_CHANGE_REPORTED:
2767                 scsi_rescan_device(&sdev->sdev_gendev);
2768                 envp[idx++] = "SDEV_UA=INQUIRY_DATA_HAS_CHANGED";
2769                 break;
2770         case SDEV_EVT_CAPACITY_CHANGE_REPORTED:
2771                 envp[idx++] = "SDEV_UA=CAPACITY_DATA_HAS_CHANGED";
2772                 break;
2773         case SDEV_EVT_SOFT_THRESHOLD_REACHED_REPORTED:
2774                envp[idx++] = "SDEV_UA=THIN_PROVISIONING_SOFT_THRESHOLD_REACHED";
2775                 break;
2776         case SDEV_EVT_MODE_PARAMETER_CHANGE_REPORTED:
2777                 envp[idx++] = "SDEV_UA=MODE_PARAMETERS_CHANGED";
2778                 break;
2779         case SDEV_EVT_LUN_CHANGE_REPORTED:
2780                 envp[idx++] = "SDEV_UA=REPORTED_LUNS_DATA_HAS_CHANGED";
2781                 break;
2782         case SDEV_EVT_ALUA_STATE_CHANGE_REPORTED:
2783                 envp[idx++] = "SDEV_UA=ASYMMETRIC_ACCESS_STATE_CHANGED";
2784                 break;
2785         case SDEV_EVT_POWER_ON_RESET_OCCURRED:
2786                 envp[idx++] = "SDEV_UA=POWER_ON_RESET_OCCURRED";
2787                 break;
2788         default:
2789                 /* do nothing */
2790                 break;
2791         }
2792
2793         envp[idx++] = NULL;
2794
2795         kobject_uevent_env(&sdev->sdev_gendev.kobj, KOBJ_CHANGE, envp);
2796 }
2797
2798 /**
2799  *      sdev_evt_thread - send a uevent for each scsi event
2800  *      @work: work struct for scsi_device
2801  *
2802  *      Dispatch queued events to their associated scsi_device kobjects
2803  *      as uevents.
2804  */
2805 void scsi_evt_thread(struct work_struct *work)
2806 {
2807         struct scsi_device *sdev;
2808         enum scsi_device_event evt_type;
2809         LIST_HEAD(event_list);
2810
2811         sdev = container_of(work, struct scsi_device, event_work);
2812
2813         for (evt_type = SDEV_EVT_FIRST; evt_type <= SDEV_EVT_LAST; evt_type++)
2814                 if (test_and_clear_bit(evt_type, sdev->pending_events))
2815                         sdev_evt_send_simple(sdev, evt_type, GFP_KERNEL);
2816
2817         while (1) {
2818                 struct scsi_event *evt;
2819                 struct list_head *this, *tmp;
2820                 unsigned long flags;
2821
2822                 spin_lock_irqsave(&sdev->list_lock, flags);
2823                 list_splice_init(&sdev->event_list, &event_list);
2824                 spin_unlock_irqrestore(&sdev->list_lock, flags);
2825
2826                 if (list_empty(&event_list))
2827                         break;
2828
2829                 list_for_each_safe(this, tmp, &event_list) {
2830                         evt = list_entry(this, struct scsi_event, node);
2831                         list_del(&evt->node);
2832                         scsi_evt_emit(sdev, evt);
2833                         kfree(evt);
2834                 }
2835         }
2836 }
2837
2838 /**
2839  *      sdev_evt_send - send asserted event to uevent thread
2840  *      @sdev: scsi_device event occurred on
2841  *      @evt: event to send
2842  *
2843  *      Assert scsi device event asynchronously.
2844  */
2845 void sdev_evt_send(struct scsi_device *sdev, struct scsi_event *evt)
2846 {
2847         unsigned long flags;
2848
2849 #if 0
2850         /* FIXME: currently this check eliminates all media change events
2851          * for polled devices.  Need to update to discriminate between AN
2852          * and polled events */
2853         if (!test_bit(evt->evt_type, sdev->supported_events)) {
2854                 kfree(evt);
2855                 return;
2856         }
2857 #endif
2858
2859         spin_lock_irqsave(&sdev->list_lock, flags);
2860         list_add_tail(&evt->node, &sdev->event_list);
2861         schedule_work(&sdev->event_work);
2862         spin_unlock_irqrestore(&sdev->list_lock, flags);
2863 }
2864 EXPORT_SYMBOL_GPL(sdev_evt_send);
2865
2866 /**
2867  *      sdev_evt_alloc - allocate a new scsi event
2868  *      @evt_type: type of event to allocate
2869  *      @gfpflags: GFP flags for allocation
2870  *
2871  *      Allocates and returns a new scsi_event.
2872  */
2873 struct scsi_event *sdev_evt_alloc(enum scsi_device_event evt_type,
2874                                   gfp_t gfpflags)
2875 {
2876         struct scsi_event *evt = kzalloc(sizeof(struct scsi_event), gfpflags);
2877         if (!evt)
2878                 return NULL;
2879
2880         evt->evt_type = evt_type;
2881         INIT_LIST_HEAD(&evt->node);
2882
2883         /* evt_type-specific initialization, if any */
2884         switch (evt_type) {
2885         case SDEV_EVT_MEDIA_CHANGE:
2886         case SDEV_EVT_INQUIRY_CHANGE_REPORTED:
2887         case SDEV_EVT_CAPACITY_CHANGE_REPORTED:
2888         case SDEV_EVT_SOFT_THRESHOLD_REACHED_REPORTED:
2889         case SDEV_EVT_MODE_PARAMETER_CHANGE_REPORTED:
2890         case SDEV_EVT_LUN_CHANGE_REPORTED:
2891         case SDEV_EVT_ALUA_STATE_CHANGE_REPORTED:
2892         case SDEV_EVT_POWER_ON_RESET_OCCURRED:
2893         default:
2894                 /* do nothing */
2895                 break;
2896         }
2897
2898         return evt;
2899 }
2900 EXPORT_SYMBOL_GPL(sdev_evt_alloc);
2901
2902 /**
2903  *      sdev_evt_send_simple - send asserted event to uevent thread
2904  *      @sdev: scsi_device event occurred on
2905  *      @evt_type: type of event to send
2906  *      @gfpflags: GFP flags for allocation
2907  *
2908  *      Assert scsi device event asynchronously, given an event type.
2909  */
2910 void sdev_evt_send_simple(struct scsi_device *sdev,
2911                           enum scsi_device_event evt_type, gfp_t gfpflags)
2912 {
2913         struct scsi_event *evt = sdev_evt_alloc(evt_type, gfpflags);
2914         if (!evt) {
2915                 sdev_printk(KERN_ERR, sdev, "event %d eaten due to OOM\n",
2916                             evt_type);
2917                 return;
2918         }
2919
2920         sdev_evt_send(sdev, evt);
2921 }
2922 EXPORT_SYMBOL_GPL(sdev_evt_send_simple);
2923
2924 /**
2925  * scsi_request_fn_active() - number of kernel threads inside scsi_request_fn()
2926  * @sdev: SCSI device to count the number of scsi_request_fn() callers for.
2927  */
2928 static int scsi_request_fn_active(struct scsi_device *sdev)
2929 {
2930         struct request_queue *q = sdev->request_queue;
2931         int request_fn_active;
2932
2933         WARN_ON_ONCE(sdev->host->use_blk_mq);
2934
2935         spin_lock_irq(q->queue_lock);
2936         request_fn_active = q->request_fn_active;
2937         spin_unlock_irq(q->queue_lock);
2938
2939         return request_fn_active;
2940 }
2941
2942 /**
2943  * scsi_wait_for_queuecommand() - wait for ongoing queuecommand() calls
2944  * @sdev: SCSI device pointer.
2945  *
2946  * Wait until the ongoing shost->hostt->queuecommand() calls that are
2947  * invoked from scsi_request_fn() have finished.
2948  */
2949 static void scsi_wait_for_queuecommand(struct scsi_device *sdev)
2950 {
2951         WARN_ON_ONCE(sdev->host->use_blk_mq);
2952
2953         while (scsi_request_fn_active(sdev))
2954                 msleep(20);
2955 }
2956
2957 /**
2958  *      scsi_device_quiesce - Block user issued commands.
2959  *      @sdev:  scsi device to quiesce.
2960  *
2961  *      This works by trying to transition to the SDEV_QUIESCE state
2962  *      (which must be a legal transition).  When the device is in this
2963  *      state, only special requests will be accepted, all others will
2964  *      be deferred.  Since special requests may also be requeued requests,
2965  *      a successful return doesn't guarantee the device will be 
2966  *      totally quiescent.
2967  *
2968  *      Must be called with user context, may sleep.
2969  *
2970  *      Returns zero if unsuccessful or an error if not.
2971  */
2972 int
2973 scsi_device_quiesce(struct scsi_device *sdev)
2974 {
2975         struct request_queue *q = sdev->request_queue;
2976         int err;
2977
2978         /*
2979          * It is allowed to call scsi_device_quiesce() multiple times from
2980          * the same context but concurrent scsi_device_quiesce() calls are
2981          * not allowed.
2982          */
2983         WARN_ON_ONCE(sdev->quiesced_by && sdev->quiesced_by != current);
2984
2985         blk_set_preempt_only(q);
2986
2987         blk_mq_freeze_queue(q);
2988         /*
2989          * Ensure that the effect of blk_set_preempt_only() will be visible
2990          * for percpu_ref_tryget() callers that occur after the queue
2991          * unfreeze even if the queue was already frozen before this function
2992          * was called. See also https://lwn.net/Articles/573497/.
2993          */
2994         synchronize_rcu();
2995         blk_mq_unfreeze_queue(q);
2996
2997         mutex_lock(&sdev->state_mutex);
2998         err = scsi_device_set_state(sdev, SDEV_QUIESCE);
2999         if (err == 0)
3000                 sdev->quiesced_by = current;
3001         else
3002                 blk_clear_preempt_only(q);
3003         mutex_unlock(&sdev->state_mutex);
3004
3005         return err;
3006 }
3007 EXPORT_SYMBOL(scsi_device_quiesce);
3008
3009 /**
3010  *      scsi_device_resume - Restart user issued commands to a quiesced device.
3011  *      @sdev:  scsi device to resume.
3012  *
3013  *      Moves the device from quiesced back to running and restarts the
3014  *      queues.
3015  *
3016  *      Must be called with user context, may sleep.
3017  */
3018 void scsi_device_resume(struct scsi_device *sdev)
3019 {
3020         /* check if the device state was mutated prior to resume, and if
3021          * so assume the state is being managed elsewhere (for example
3022          * device deleted during suspend)
3023          */
3024         mutex_lock(&sdev->state_mutex);
3025         WARN_ON_ONCE(!sdev->quiesced_by);
3026         sdev->quiesced_by = NULL;
3027         blk_clear_preempt_only(sdev->request_queue);
3028         if (sdev->sdev_state == SDEV_QUIESCE)
3029                 scsi_device_set_state(sdev, SDEV_RUNNING);
3030         mutex_unlock(&sdev->state_mutex);
3031 }
3032 EXPORT_SYMBOL(scsi_device_resume);
3033
3034 static void
3035 device_quiesce_fn(struct scsi_device *sdev, void *data)
3036 {
3037         scsi_device_quiesce(sdev);
3038 }
3039
3040 void
3041 scsi_target_quiesce(struct scsi_target *starget)
3042 {
3043         starget_for_each_device(starget, NULL, device_quiesce_fn);
3044 }
3045 EXPORT_SYMBOL(scsi_target_quiesce);
3046
3047 static void
3048 device_resume_fn(struct scsi_device *sdev, void *data)
3049 {
3050         scsi_device_resume(sdev);
3051 }
3052
3053 void
3054 scsi_target_resume(struct scsi_target *starget)
3055 {
3056         starget_for_each_device(starget, NULL, device_resume_fn);
3057 }
3058 EXPORT_SYMBOL(scsi_target_resume);
3059
3060 /**
3061  * scsi_internal_device_block_nowait - try to transition to the SDEV_BLOCK state
3062  * @sdev: device to block
3063  *
3064  * Pause SCSI command processing on the specified device. Does not sleep.
3065  *
3066  * Returns zero if successful or a negative error code upon failure.
3067  *
3068  * Notes:
3069  * This routine transitions the device to the SDEV_BLOCK state (which must be
3070  * a legal transition). When the device is in this state, command processing
3071  * is paused until the device leaves the SDEV_BLOCK state. See also
3072  * scsi_internal_device_unblock_nowait().
3073  */
3074 int scsi_internal_device_block_nowait(struct scsi_device *sdev)
3075 {
3076         struct request_queue *q = sdev->request_queue;
3077         unsigned long flags;
3078         int err = 0;
3079
3080         err = scsi_device_set_state(sdev, SDEV_BLOCK);
3081         if (err) {
3082                 err = scsi_device_set_state(sdev, SDEV_CREATED_BLOCK);
3083
3084                 if (err)
3085                         return err;
3086         }
3087
3088         /* 
3089          * The device has transitioned to SDEV_BLOCK.  Stop the
3090          * block layer from calling the midlayer with this device's
3091          * request queue. 
3092          */
3093         if (q->mq_ops) {
3094                 blk_mq_quiesce_queue_nowait(q);
3095         } else {
3096                 spin_lock_irqsave(q->queue_lock, flags);
3097                 blk_stop_queue(q);
3098                 spin_unlock_irqrestore(q->queue_lock, flags);
3099         }
3100
3101         return 0;
3102 }
3103 EXPORT_SYMBOL_GPL(scsi_internal_device_block_nowait);
3104
3105 /**
3106  * scsi_internal_device_block - try to transition to the SDEV_BLOCK state
3107  * @sdev: device to block
3108  *
3109  * Pause SCSI command processing on the specified device and wait until all
3110  * ongoing scsi_request_fn() / scsi_queue_rq() calls have finished. May sleep.
3111  *
3112  * Returns zero if successful or a negative error code upon failure.
3113  *
3114  * Note:
3115  * This routine transitions the device to the SDEV_BLOCK state (which must be
3116  * a legal transition). When the device is in this state, command processing
3117  * is paused until the device leaves the SDEV_BLOCK state. See also
3118  * scsi_internal_device_unblock().
3119  *
3120  * To do: avoid that scsi_send_eh_cmnd() calls queuecommand() after
3121  * scsi_internal_device_block() has blocked a SCSI device and also
3122  * remove the rport mutex lock and unlock calls from srp_queuecommand().
3123  */
3124 static int scsi_internal_device_block(struct scsi_device *sdev)
3125 {
3126         struct request_queue *q = sdev->request_queue;
3127         int err;
3128
3129         mutex_lock(&sdev->state_mutex);
3130         err = scsi_internal_device_block_nowait(sdev);
3131         if (err == 0) {
3132                 if (q->mq_ops)
3133                         blk_mq_quiesce_queue(q);
3134                 else
3135                         scsi_wait_for_queuecommand(sdev);
3136         }
3137         mutex_unlock(&sdev->state_mutex);
3138
3139         return err;
3140 }
3141  
3142 void scsi_start_queue(struct scsi_device *sdev)
3143 {
3144         struct request_queue *q = sdev->request_queue;
3145         unsigned long flags;
3146
3147         if (q->mq_ops) {
3148                 blk_mq_unquiesce_queue(q);
3149         } else {
3150                 spin_lock_irqsave(q->queue_lock, flags);
3151                 blk_start_queue(q);
3152                 spin_unlock_irqrestore(q->queue_lock, flags);
3153         }
3154 }
3155
3156 /**
3157  * scsi_internal_device_unblock_nowait - resume a device after a block request
3158  * @sdev:       device to resume
3159  * @new_state:  state to set the device to after unblocking
3160  *
3161  * Restart the device queue for a previously suspended SCSI device. Does not
3162  * sleep.
3163  *
3164  * Returns zero if successful or a negative error code upon failure.
3165  *
3166  * Notes:
3167  * This routine transitions the device to the SDEV_RUNNING state or to one of
3168  * the offline states (which must be a legal transition) allowing the midlayer
3169  * to goose the queue for this device.
3170  */
3171 int scsi_internal_device_unblock_nowait(struct scsi_device *sdev,
3172                                         enum scsi_device_state new_state)
3173 {
3174         /*
3175          * Try to transition the scsi device to SDEV_RUNNING or one of the
3176          * offlined states and goose the device queue if successful.
3177          */
3178         switch (sdev->sdev_state) {
3179         case SDEV_BLOCK:
3180         case SDEV_TRANSPORT_OFFLINE:
3181                 sdev->sdev_state = new_state;
3182                 break;
3183         case SDEV_CREATED_BLOCK:
3184                 if (new_state == SDEV_TRANSPORT_OFFLINE ||
3185                     new_state == SDEV_OFFLINE)
3186                         sdev->sdev_state = new_state;
3187                 else
3188                         sdev->sdev_state = SDEV_CREATED;
3189                 break;
3190         case SDEV_CANCEL:
3191         case SDEV_OFFLINE:
3192                 break;
3193         default:
3194                 return -EINVAL;
3195         }
3196         scsi_start_queue(sdev);
3197
3198         return 0;
3199 }
3200 EXPORT_SYMBOL_GPL(scsi_internal_device_unblock_nowait);
3201
3202 /**
3203  * scsi_internal_device_unblock - resume a device after a block request
3204  * @sdev:       device to resume
3205  * @new_state:  state to set the device to after unblocking
3206  *
3207  * Restart the device queue for a previously suspended SCSI device. May sleep.
3208  *
3209  * Returns zero if successful or a negative error code upon failure.
3210  *
3211  * Notes:
3212  * This routine transitions the device to the SDEV_RUNNING state or to one of
3213  * the offline states (which must be a legal transition) allowing the midlayer
3214  * to goose the queue for this device.
3215  */
3216 static int scsi_internal_device_unblock(struct scsi_device *sdev,
3217                                         enum scsi_device_state new_state)
3218 {
3219         int ret;
3220
3221         mutex_lock(&sdev->state_mutex);
3222         ret = scsi_internal_device_unblock_nowait(sdev, new_state);
3223         mutex_unlock(&sdev->state_mutex);
3224
3225         return ret;
3226 }
3227
3228 static void
3229 device_block(struct scsi_device *sdev, void *data)
3230 {
3231         scsi_internal_device_block(sdev);
3232 }
3233
3234 static int
3235 target_block(struct device *dev, void *data)
3236 {
3237         if (scsi_is_target_device(dev))
3238                 starget_for_each_device(to_scsi_target(dev), NULL,
3239                                         device_block);
3240         return 0;
3241 }
3242
3243 void
3244 scsi_target_block(struct device *dev)
3245 {
3246         if (scsi_is_target_device(dev))
3247                 starget_for_each_device(to_scsi_target(dev), NULL,
3248                                         device_block);
3249         else
3250                 device_for_each_child(dev, NULL, target_block);
3251 }
3252 EXPORT_SYMBOL_GPL(scsi_target_block);
3253
3254 static void
3255 device_unblock(struct scsi_device *sdev, void *data)
3256 {
3257         scsi_internal_device_unblock(sdev, *(enum scsi_device_state *)data);
3258 }
3259
3260 static int
3261 target_unblock(struct device *dev, void *data)
3262 {
3263         if (scsi_is_target_device(dev))
3264                 starget_for_each_device(to_scsi_target(dev), data,
3265                                         device_unblock);
3266         return 0;
3267 }
3268
3269 void
3270 scsi_target_unblock(struct device *dev, enum scsi_device_state new_state)
3271 {
3272         if (scsi_is_target_device(dev))
3273                 starget_for_each_device(to_scsi_target(dev), &new_state,
3274                                         device_unblock);
3275         else
3276                 device_for_each_child(dev, &new_state, target_unblock);
3277 }
3278 EXPORT_SYMBOL_GPL(scsi_target_unblock);
3279
3280 /**
3281  * scsi_kmap_atomic_sg - find and atomically map an sg-elemnt
3282  * @sgl:        scatter-gather list
3283  * @sg_count:   number of segments in sg
3284  * @offset:     offset in bytes into sg, on return offset into the mapped area
3285  * @len:        bytes to map, on return number of bytes mapped
3286  *
3287  * Returns virtual address of the start of the mapped page
3288  */
3289 void *scsi_kmap_atomic_sg(struct scatterlist *sgl, int sg_count,
3290                           size_t *offset, size_t *len)
3291 {
3292         int i;
3293         size_t sg_len = 0, len_complete = 0;
3294         struct scatterlist *sg;
3295         struct page *page;
3296
3297         WARN_ON(!irqs_disabled());
3298
3299         for_each_sg(sgl, sg, sg_count, i) {
3300                 len_complete = sg_len; /* Complete sg-entries */
3301                 sg_len += sg->length;
3302                 if (sg_len > *offset)
3303                         break;
3304         }
3305
3306         if (unlikely(i == sg_count)) {
3307                 printk(KERN_ERR "%s: Bytes in sg: %zu, requested offset %zu, "
3308                         "elements %d\n",
3309                        __func__, sg_len, *offset, sg_count);
3310                 WARN_ON(1);
3311                 return NULL;
3312         }
3313
3314         /* Offset starting from the beginning of first page in this sg-entry */
3315         *offset = *offset - len_complete + sg->offset;
3316
3317         /* Assumption: contiguous pages can be accessed as "page + i" */
3318         page = nth_page(sg_page(sg), (*offset >> PAGE_SHIFT));
3319         *offset &= ~PAGE_MASK;
3320
3321         /* Bytes in this sg-entry from *offset to the end of the page */
3322         sg_len = PAGE_SIZE - *offset;
3323         if (*len > sg_len)
3324                 *len = sg_len;
3325
3326         return kmap_atomic(page);
3327 }
3328 EXPORT_SYMBOL(scsi_kmap_atomic_sg);
3329
3330 /**
3331  * scsi_kunmap_atomic_sg - atomically unmap a virtual address, previously mapped with scsi_kmap_atomic_sg
3332  * @virt:       virtual address to be unmapped
3333  */
3334 void scsi_kunmap_atomic_sg(void *virt)
3335 {
3336         kunmap_atomic(virt);
3337 }
3338 EXPORT_SYMBOL(scsi_kunmap_atomic_sg);
3339
3340 void sdev_disable_disk_events(struct scsi_device *sdev)
3341 {
3342         atomic_inc(&sdev->disk_events_disable_depth);
3343 }
3344 EXPORT_SYMBOL(sdev_disable_disk_events);
3345
3346 void sdev_enable_disk_events(struct scsi_device *sdev)
3347 {
3348         if (WARN_ON_ONCE(atomic_read(&sdev->disk_events_disable_depth) <= 0))
3349                 return;
3350         atomic_dec(&sdev->disk_events_disable_depth);
3351 }
3352 EXPORT_SYMBOL(sdev_enable_disk_events);
3353
3354 /**
3355  * scsi_vpd_lun_id - return a unique device identification
3356  * @sdev: SCSI device
3357  * @id:   buffer for the identification
3358  * @id_len:  length of the buffer
3359  *
3360  * Copies a unique device identification into @id based
3361  * on the information in the VPD page 0x83 of the device.
3362  * The string will be formatted as a SCSI name string.
3363  *
3364  * Returns the length of the identification or error on failure.
3365  * If the identifier is longer than the supplied buffer the actual
3366  * identifier length is returned and the buffer is not zero-padded.
3367  */
3368 int scsi_vpd_lun_id(struct scsi_device *sdev, char *id, size_t id_len)
3369 {
3370         u8 cur_id_type = 0xff;
3371         u8 cur_id_size = 0;
3372         const unsigned char *d, *cur_id_str;
3373         const struct scsi_vpd *vpd_pg83;
3374         int id_size = -EINVAL;
3375
3376         rcu_read_lock();
3377         vpd_pg83 = rcu_dereference(sdev->vpd_pg83);
3378         if (!vpd_pg83) {
3379                 rcu_read_unlock();
3380                 return -ENXIO;
3381         }
3382
3383         /*
3384          * Look for the correct descriptor.
3385          * Order of preference for lun descriptor:
3386          * - SCSI name string
3387          * - NAA IEEE Registered Extended
3388          * - EUI-64 based 16-byte
3389          * - EUI-64 based 12-byte
3390          * - NAA IEEE Registered
3391          * - NAA IEEE Extended
3392          * - T10 Vendor ID
3393          * as longer descriptors reduce the likelyhood
3394          * of identification clashes.
3395          */
3396
3397         /* The id string must be at least 20 bytes + terminating NULL byte */
3398         if (id_len < 21) {
3399                 rcu_read_unlock();
3400                 return -EINVAL;
3401         }
3402
3403         memset(id, 0, id_len);
3404         d = vpd_pg83->data + 4;
3405         while (d < vpd_pg83->data + vpd_pg83->len) {
3406                 /* Skip designators not referring to the LUN */
3407                 if ((d[1] & 0x30) != 0x00)
3408                         goto next_desig;
3409
3410                 switch (d[1] & 0xf) {
3411                 case 0x1:
3412                         /* T10 Vendor ID */
3413                         if (cur_id_size > d[3])
3414                                 break;
3415                         /* Prefer anything */
3416                         if (cur_id_type > 0x01 && cur_id_type != 0xff)
3417                                 break;
3418                         cur_id_size = d[3];
3419                         if (cur_id_size + 4 > id_len)
3420                                 cur_id_size = id_len - 4;
3421          &