Merge git://git.kernel.org/pub/scm/linux/kernel/git/davem/net
[sfrench/cifs-2.6.git] / drivers / scsi / scsi_lib.c
1 /*
2  * Copyright (C) 1999 Eric Youngdale
3  * Copyright (C) 2014 Christoph Hellwig
4  *
5  *  SCSI queueing library.
6  *      Initial versions: Eric Youngdale (eric@andante.org).
7  *                        Based upon conversations with large numbers
8  *                        of people at Linux Expo.
9  */
10
11 #include <linux/bio.h>
12 #include <linux/bitops.h>
13 #include <linux/blkdev.h>
14 #include <linux/completion.h>
15 #include <linux/kernel.h>
16 #include <linux/export.h>
17 #include <linux/init.h>
18 #include <linux/pci.h>
19 #include <linux/delay.h>
20 #include <linux/hardirq.h>
21 #include <linux/scatterlist.h>
22 #include <linux/blk-mq.h>
23 #include <linux/ratelimit.h>
24 #include <asm/unaligned.h>
25
26 #include <scsi/scsi.h>
27 #include <scsi/scsi_cmnd.h>
28 #include <scsi/scsi_dbg.h>
29 #include <scsi/scsi_device.h>
30 #include <scsi/scsi_driver.h>
31 #include <scsi/scsi_eh.h>
32 #include <scsi/scsi_host.h>
33 #include <scsi/scsi_transport.h> /* __scsi_init_queue() */
34 #include <scsi/scsi_dh.h>
35
36 #include <trace/events/scsi.h>
37
38 #include "scsi_debugfs.h"
39 #include "scsi_priv.h"
40 #include "scsi_logging.h"
41
42 static struct kmem_cache *scsi_sdb_cache;
43 static struct kmem_cache *scsi_sense_cache;
44 static struct kmem_cache *scsi_sense_isadma_cache;
45 static DEFINE_MUTEX(scsi_sense_cache_mutex);
46
47 static void scsi_mq_uninit_cmd(struct scsi_cmnd *cmd);
48
49 static inline struct kmem_cache *
50 scsi_select_sense_cache(bool unchecked_isa_dma)
51 {
52         return unchecked_isa_dma ? scsi_sense_isadma_cache : scsi_sense_cache;
53 }
54
55 static void scsi_free_sense_buffer(bool unchecked_isa_dma,
56                                    unsigned char *sense_buffer)
57 {
58         kmem_cache_free(scsi_select_sense_cache(unchecked_isa_dma),
59                         sense_buffer);
60 }
61
62 static unsigned char *scsi_alloc_sense_buffer(bool unchecked_isa_dma,
63         gfp_t gfp_mask, int numa_node)
64 {
65         return kmem_cache_alloc_node(scsi_select_sense_cache(unchecked_isa_dma),
66                                      gfp_mask, numa_node);
67 }
68
69 int scsi_init_sense_cache(struct Scsi_Host *shost)
70 {
71         struct kmem_cache *cache;
72         int ret = 0;
73
74         cache = scsi_select_sense_cache(shost->unchecked_isa_dma);
75         if (cache)
76                 return 0;
77
78         mutex_lock(&scsi_sense_cache_mutex);
79         if (shost->unchecked_isa_dma) {
80                 scsi_sense_isadma_cache =
81                         kmem_cache_create("scsi_sense_cache(DMA)",
82                                 SCSI_SENSE_BUFFERSIZE, 0,
83                                 SLAB_HWCACHE_ALIGN | SLAB_CACHE_DMA, NULL);
84                 if (!scsi_sense_isadma_cache)
85                         ret = -ENOMEM;
86         } else {
87                 scsi_sense_cache =
88                         kmem_cache_create_usercopy("scsi_sense_cache",
89                                 SCSI_SENSE_BUFFERSIZE, 0, SLAB_HWCACHE_ALIGN,
90                                 0, SCSI_SENSE_BUFFERSIZE, NULL);
91                 if (!scsi_sense_cache)
92                         ret = -ENOMEM;
93         }
94
95         mutex_unlock(&scsi_sense_cache_mutex);
96         return ret;
97 }
98
99 /*
100  * When to reinvoke queueing after a resource shortage. It's 3 msecs to
101  * not change behaviour from the previous unplug mechanism, experimentation
102  * may prove this needs changing.
103  */
104 #define SCSI_QUEUE_DELAY        3
105
106 static void
107 scsi_set_blocked(struct scsi_cmnd *cmd, int reason)
108 {
109         struct Scsi_Host *host = cmd->device->host;
110         struct scsi_device *device = cmd->device;
111         struct scsi_target *starget = scsi_target(device);
112
113         /*
114          * Set the appropriate busy bit for the device/host.
115          *
116          * If the host/device isn't busy, assume that something actually
117          * completed, and that we should be able to queue a command now.
118          *
119          * Note that the prior mid-layer assumption that any host could
120          * always queue at least one command is now broken.  The mid-layer
121          * will implement a user specifiable stall (see
122          * scsi_host.max_host_blocked and scsi_device.max_device_blocked)
123          * if a command is requeued with no other commands outstanding
124          * either for the device or for the host.
125          */
126         switch (reason) {
127         case SCSI_MLQUEUE_HOST_BUSY:
128                 atomic_set(&host->host_blocked, host->max_host_blocked);
129                 break;
130         case SCSI_MLQUEUE_DEVICE_BUSY:
131         case SCSI_MLQUEUE_EH_RETRY:
132                 atomic_set(&device->device_blocked,
133                            device->max_device_blocked);
134                 break;
135         case SCSI_MLQUEUE_TARGET_BUSY:
136                 atomic_set(&starget->target_blocked,
137                            starget->max_target_blocked);
138                 break;
139         }
140 }
141
142 static void scsi_mq_requeue_cmd(struct scsi_cmnd *cmd)
143 {
144         struct scsi_device *sdev = cmd->device;
145
146         if (cmd->request->rq_flags & RQF_DONTPREP) {
147                 cmd->request->rq_flags &= ~RQF_DONTPREP;
148                 scsi_mq_uninit_cmd(cmd);
149         } else {
150                 WARN_ON_ONCE(true);
151         }
152         blk_mq_requeue_request(cmd->request, true);
153         put_device(&sdev->sdev_gendev);
154 }
155
156 /**
157  * __scsi_queue_insert - private queue insertion
158  * @cmd: The SCSI command being requeued
159  * @reason:  The reason for the requeue
160  * @unbusy: Whether the queue should be unbusied
161  *
162  * This is a private queue insertion.  The public interface
163  * scsi_queue_insert() always assumes the queue should be unbusied
164  * because it's always called before the completion.  This function is
165  * for a requeue after completion, which should only occur in this
166  * file.
167  */
168 static void __scsi_queue_insert(struct scsi_cmnd *cmd, int reason, bool unbusy)
169 {
170         struct scsi_device *device = cmd->device;
171
172         SCSI_LOG_MLQUEUE(1, scmd_printk(KERN_INFO, cmd,
173                 "Inserting command %p into mlqueue\n", cmd));
174
175         scsi_set_blocked(cmd, reason);
176
177         /*
178          * Decrement the counters, since these commands are no longer
179          * active on the host/device.
180          */
181         if (unbusy)
182                 scsi_device_unbusy(device);
183
184         /*
185          * Requeue this command.  It will go before all other commands
186          * that are already in the queue. Schedule requeue work under
187          * lock such that the kblockd_schedule_work() call happens
188          * before blk_cleanup_queue() finishes.
189          */
190         cmd->result = 0;
191
192         /*
193          * Before a SCSI command is dispatched,
194          * get_device(&sdev->sdev_gendev) is called and the host,
195          * target and device busy counters are increased. Since
196          * requeuing a request causes these actions to be repeated and
197          * since scsi_device_unbusy() has already been called,
198          * put_device(&device->sdev_gendev) must still be called. Call
199          * put_device() after blk_mq_requeue_request() to avoid that
200          * removal of the SCSI device can start before requeueing has
201          * happened.
202          */
203         blk_mq_requeue_request(cmd->request, true);
204         put_device(&device->sdev_gendev);
205 }
206
207 /*
208  * Function:    scsi_queue_insert()
209  *
210  * Purpose:     Insert a command in the midlevel queue.
211  *
212  * Arguments:   cmd    - command that we are adding to queue.
213  *              reason - why we are inserting command to queue.
214  *
215  * Lock status: Assumed that lock is not held upon entry.
216  *
217  * Returns:     Nothing.
218  *
219  * Notes:       We do this for one of two cases.  Either the host is busy
220  *              and it cannot accept any more commands for the time being,
221  *              or the device returned QUEUE_FULL and can accept no more
222  *              commands.
223  * Notes:       This could be called either from an interrupt context or a
224  *              normal process context.
225  */
226 void scsi_queue_insert(struct scsi_cmnd *cmd, int reason)
227 {
228         __scsi_queue_insert(cmd, reason, true);
229 }
230
231
232 /**
233  * __scsi_execute - insert request and wait for the result
234  * @sdev:       scsi device
235  * @cmd:        scsi command
236  * @data_direction: data direction
237  * @buffer:     data buffer
238  * @bufflen:    len of buffer
239  * @sense:      optional sense buffer
240  * @sshdr:      optional decoded sense header
241  * @timeout:    request timeout in seconds
242  * @retries:    number of times to retry request
243  * @flags:      flags for ->cmd_flags
244  * @rq_flags:   flags for ->rq_flags
245  * @resid:      optional residual length
246  *
247  * Returns the scsi_cmnd result field if a command was executed, or a negative
248  * Linux error code if we didn't get that far.
249  */
250 int __scsi_execute(struct scsi_device *sdev, const unsigned char *cmd,
251                  int data_direction, void *buffer, unsigned bufflen,
252                  unsigned char *sense, struct scsi_sense_hdr *sshdr,
253                  int timeout, int retries, u64 flags, req_flags_t rq_flags,
254                  int *resid)
255 {
256         struct request *req;
257         struct scsi_request *rq;
258         int ret = DRIVER_ERROR << 24;
259
260         req = blk_get_request(sdev->request_queue,
261                         data_direction == DMA_TO_DEVICE ?
262                         REQ_OP_SCSI_OUT : REQ_OP_SCSI_IN, BLK_MQ_REQ_PREEMPT);
263         if (IS_ERR(req))
264                 return ret;
265         rq = scsi_req(req);
266
267         if (bufflen &&  blk_rq_map_kern(sdev->request_queue, req,
268                                         buffer, bufflen, GFP_NOIO))
269                 goto out;
270
271         rq->cmd_len = COMMAND_SIZE(cmd[0]);
272         memcpy(rq->cmd, cmd, rq->cmd_len);
273         rq->retries = retries;
274         req->timeout = timeout;
275         req->cmd_flags |= flags;
276         req->rq_flags |= rq_flags | RQF_QUIET;
277
278         /*
279          * head injection *required* here otherwise quiesce won't work
280          */
281         blk_execute_rq(req->q, NULL, req, 1);
282
283         /*
284          * Some devices (USB mass-storage in particular) may transfer
285          * garbage data together with a residue indicating that the data
286          * is invalid.  Prevent the garbage from being misinterpreted
287          * and prevent security leaks by zeroing out the excess data.
288          */
289         if (unlikely(rq->resid_len > 0 && rq->resid_len <= bufflen))
290                 memset(buffer + (bufflen - rq->resid_len), 0, rq->resid_len);
291
292         if (resid)
293                 *resid = rq->resid_len;
294         if (sense && rq->sense_len)
295                 memcpy(sense, rq->sense, SCSI_SENSE_BUFFERSIZE);
296         if (sshdr)
297                 scsi_normalize_sense(rq->sense, rq->sense_len, sshdr);
298         ret = rq->result;
299  out:
300         blk_put_request(req);
301
302         return ret;
303 }
304 EXPORT_SYMBOL(__scsi_execute);
305
306 /*
307  * Function:    scsi_init_cmd_errh()
308  *
309  * Purpose:     Initialize cmd fields related to error handling.
310  *
311  * Arguments:   cmd     - command that is ready to be queued.
312  *
313  * Notes:       This function has the job of initializing a number of
314  *              fields related to error handling.   Typically this will
315  *              be called once for each command, as required.
316  */
317 static void scsi_init_cmd_errh(struct scsi_cmnd *cmd)
318 {
319         cmd->serial_number = 0;
320         scsi_set_resid(cmd, 0);
321         memset(cmd->sense_buffer, 0, SCSI_SENSE_BUFFERSIZE);
322         if (cmd->cmd_len == 0)
323                 cmd->cmd_len = scsi_command_size(cmd->cmnd);
324 }
325
326 /*
327  * Decrement the host_busy counter and wake up the error handler if necessary.
328  * Avoid as follows that the error handler is not woken up if shost->host_busy
329  * == shost->host_failed: use call_rcu() in scsi_eh_scmd_add() in combination
330  * with an RCU read lock in this function to ensure that this function in its
331  * entirety either finishes before scsi_eh_scmd_add() increases the
332  * host_failed counter or that it notices the shost state change made by
333  * scsi_eh_scmd_add().
334  */
335 static void scsi_dec_host_busy(struct Scsi_Host *shost)
336 {
337         unsigned long flags;
338
339         rcu_read_lock();
340         atomic_dec(&shost->host_busy);
341         if (unlikely(scsi_host_in_recovery(shost))) {
342                 spin_lock_irqsave(shost->host_lock, flags);
343                 if (shost->host_failed || shost->host_eh_scheduled)
344                         scsi_eh_wakeup(shost);
345                 spin_unlock_irqrestore(shost->host_lock, flags);
346         }
347         rcu_read_unlock();
348 }
349
350 void scsi_device_unbusy(struct scsi_device *sdev)
351 {
352         struct Scsi_Host *shost = sdev->host;
353         struct scsi_target *starget = scsi_target(sdev);
354
355         scsi_dec_host_busy(shost);
356
357         if (starget->can_queue > 0)
358                 atomic_dec(&starget->target_busy);
359
360         atomic_dec(&sdev->device_busy);
361 }
362
363 static void scsi_kick_queue(struct request_queue *q)
364 {
365         blk_mq_run_hw_queues(q, false);
366 }
367
368 /*
369  * Called for single_lun devices on IO completion. Clear starget_sdev_user,
370  * and call blk_run_queue for all the scsi_devices on the target -
371  * including current_sdev first.
372  *
373  * Called with *no* scsi locks held.
374  */
375 static void scsi_single_lun_run(struct scsi_device *current_sdev)
376 {
377         struct Scsi_Host *shost = current_sdev->host;
378         struct scsi_device *sdev, *tmp;
379         struct scsi_target *starget = scsi_target(current_sdev);
380         unsigned long flags;
381
382         spin_lock_irqsave(shost->host_lock, flags);
383         starget->starget_sdev_user = NULL;
384         spin_unlock_irqrestore(shost->host_lock, flags);
385
386         /*
387          * Call blk_run_queue for all LUNs on the target, starting with
388          * current_sdev. We race with others (to set starget_sdev_user),
389          * but in most cases, we will be first. Ideally, each LU on the
390          * target would get some limited time or requests on the target.
391          */
392         scsi_kick_queue(current_sdev->request_queue);
393
394         spin_lock_irqsave(shost->host_lock, flags);
395         if (starget->starget_sdev_user)
396                 goto out;
397         list_for_each_entry_safe(sdev, tmp, &starget->devices,
398                         same_target_siblings) {
399                 if (sdev == current_sdev)
400                         continue;
401                 if (scsi_device_get(sdev))
402                         continue;
403
404                 spin_unlock_irqrestore(shost->host_lock, flags);
405                 scsi_kick_queue(sdev->request_queue);
406                 spin_lock_irqsave(shost->host_lock, flags);
407         
408                 scsi_device_put(sdev);
409         }
410  out:
411         spin_unlock_irqrestore(shost->host_lock, flags);
412 }
413
414 static inline bool scsi_device_is_busy(struct scsi_device *sdev)
415 {
416         if (atomic_read(&sdev->device_busy) >= sdev->queue_depth)
417                 return true;
418         if (atomic_read(&sdev->device_blocked) > 0)
419                 return true;
420         return false;
421 }
422
423 static inline bool scsi_target_is_busy(struct scsi_target *starget)
424 {
425         if (starget->can_queue > 0) {
426                 if (atomic_read(&starget->target_busy) >= starget->can_queue)
427                         return true;
428                 if (atomic_read(&starget->target_blocked) > 0)
429                         return true;
430         }
431         return false;
432 }
433
434 static inline bool scsi_host_is_busy(struct Scsi_Host *shost)
435 {
436         if (shost->can_queue > 0 &&
437             atomic_read(&shost->host_busy) >= shost->can_queue)
438                 return true;
439         if (atomic_read(&shost->host_blocked) > 0)
440                 return true;
441         if (shost->host_self_blocked)
442                 return true;
443         return false;
444 }
445
446 static void scsi_starved_list_run(struct Scsi_Host *shost)
447 {
448         LIST_HEAD(starved_list);
449         struct scsi_device *sdev;
450         unsigned long flags;
451
452         spin_lock_irqsave(shost->host_lock, flags);
453         list_splice_init(&shost->starved_list, &starved_list);
454
455         while (!list_empty(&starved_list)) {
456                 struct request_queue *slq;
457
458                 /*
459                  * As long as shost is accepting commands and we have
460                  * starved queues, call blk_run_queue. scsi_request_fn
461                  * drops the queue_lock and can add us back to the
462                  * starved_list.
463                  *
464                  * host_lock protects the starved_list and starved_entry.
465                  * scsi_request_fn must get the host_lock before checking
466                  * or modifying starved_list or starved_entry.
467                  */
468                 if (scsi_host_is_busy(shost))
469                         break;
470
471                 sdev = list_entry(starved_list.next,
472                                   struct scsi_device, starved_entry);
473                 list_del_init(&sdev->starved_entry);
474                 if (scsi_target_is_busy(scsi_target(sdev))) {
475                         list_move_tail(&sdev->starved_entry,
476                                        &shost->starved_list);
477                         continue;
478                 }
479
480                 /*
481                  * Once we drop the host lock, a racing scsi_remove_device()
482                  * call may remove the sdev from the starved list and destroy
483                  * it and the queue.  Mitigate by taking a reference to the
484                  * queue and never touching the sdev again after we drop the
485                  * host lock.  Note: if __scsi_remove_device() invokes
486                  * blk_cleanup_queue() before the queue is run from this
487                  * function then blk_run_queue() will return immediately since
488                  * blk_cleanup_queue() marks the queue with QUEUE_FLAG_DYING.
489                  */
490                 slq = sdev->request_queue;
491                 if (!blk_get_queue(slq))
492                         continue;
493                 spin_unlock_irqrestore(shost->host_lock, flags);
494
495                 scsi_kick_queue(slq);
496                 blk_put_queue(slq);
497
498                 spin_lock_irqsave(shost->host_lock, flags);
499         }
500         /* put any unprocessed entries back */
501         list_splice(&starved_list, &shost->starved_list);
502         spin_unlock_irqrestore(shost->host_lock, flags);
503 }
504
505 /*
506  * Function:   scsi_run_queue()
507  *
508  * Purpose:    Select a proper request queue to serve next
509  *
510  * Arguments:  q       - last request's queue
511  *
512  * Returns:     Nothing
513  *
514  * Notes:      The previous command was completely finished, start
515  *             a new one if possible.
516  */
517 static void scsi_run_queue(struct request_queue *q)
518 {
519         struct scsi_device *sdev = q->queuedata;
520
521         if (scsi_target(sdev)->single_lun)
522                 scsi_single_lun_run(sdev);
523         if (!list_empty(&sdev->host->starved_list))
524                 scsi_starved_list_run(sdev->host);
525
526         blk_mq_run_hw_queues(q, false);
527 }
528
529 void scsi_requeue_run_queue(struct work_struct *work)
530 {
531         struct scsi_device *sdev;
532         struct request_queue *q;
533
534         sdev = container_of(work, struct scsi_device, requeue_work);
535         q = sdev->request_queue;
536         scsi_run_queue(q);
537 }
538
539 void scsi_run_host_queues(struct Scsi_Host *shost)
540 {
541         struct scsi_device *sdev;
542
543         shost_for_each_device(sdev, shost)
544                 scsi_run_queue(sdev->request_queue);
545 }
546
547 static void scsi_uninit_cmd(struct scsi_cmnd *cmd)
548 {
549         if (!blk_rq_is_passthrough(cmd->request)) {
550                 struct scsi_driver *drv = scsi_cmd_to_driver(cmd);
551
552                 if (drv->uninit_command)
553                         drv->uninit_command(cmd);
554         }
555 }
556
557 static void scsi_mq_free_sgtables(struct scsi_cmnd *cmd)
558 {
559         struct scsi_data_buffer *sdb;
560
561         if (cmd->sdb.table.nents)
562                 sg_free_table_chained(&cmd->sdb.table, true);
563         if (cmd->request->next_rq) {
564                 sdb = cmd->request->next_rq->special;
565                 if (sdb)
566                         sg_free_table_chained(&sdb->table, true);
567         }
568         if (scsi_prot_sg_count(cmd))
569                 sg_free_table_chained(&cmd->prot_sdb->table, true);
570 }
571
572 static void scsi_mq_uninit_cmd(struct scsi_cmnd *cmd)
573 {
574         scsi_mq_free_sgtables(cmd);
575         scsi_uninit_cmd(cmd);
576         scsi_del_cmd_from_list(cmd);
577 }
578
579 /* Returns false when no more bytes to process, true if there are more */
580 static bool scsi_end_request(struct request *req, blk_status_t error,
581                 unsigned int bytes, unsigned int bidi_bytes)
582 {
583         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
584         struct scsi_device *sdev = cmd->device;
585         struct request_queue *q = sdev->request_queue;
586
587         if (blk_update_request(req, error, bytes))
588                 return true;
589
590         /* Bidi request must be completed as a whole */
591         if (unlikely(bidi_bytes) &&
592             blk_update_request(req->next_rq, error, bidi_bytes))
593                 return true;
594
595         if (blk_queue_add_random(q))
596                 add_disk_randomness(req->rq_disk);
597
598         if (!blk_rq_is_scsi(req)) {
599                 WARN_ON_ONCE(!(cmd->flags & SCMD_INITIALIZED));
600                 cmd->flags &= ~SCMD_INITIALIZED;
601                 destroy_rcu_head(&cmd->rcu);
602         }
603
604         /*
605          * In the MQ case the command gets freed by __blk_mq_end_request,
606          * so we have to do all cleanup that depends on it earlier.
607          *
608          * We also can't kick the queues from irq context, so we
609          * will have to defer it to a workqueue.
610          */
611         scsi_mq_uninit_cmd(cmd);
612
613         /*
614          * queue is still alive, so grab the ref for preventing it
615          * from being cleaned up during running queue.
616          */
617         percpu_ref_get(&q->q_usage_counter);
618
619         __blk_mq_end_request(req, error);
620
621         if (scsi_target(sdev)->single_lun ||
622             !list_empty(&sdev->host->starved_list))
623                 kblockd_schedule_work(&sdev->requeue_work);
624         else
625                 blk_mq_run_hw_queues(q, true);
626
627         percpu_ref_put(&q->q_usage_counter);
628         put_device(&sdev->sdev_gendev);
629         return false;
630 }
631
632 /**
633  * scsi_result_to_blk_status - translate a SCSI result code into blk_status_t
634  * @cmd:        SCSI command
635  * @result:     scsi error code
636  *
637  * Translate a SCSI result code into a blk_status_t value. May reset the host
638  * byte of @cmd->result.
639  */
640 static blk_status_t scsi_result_to_blk_status(struct scsi_cmnd *cmd, int result)
641 {
642         switch (host_byte(result)) {
643         case DID_OK:
644                 /*
645                  * Also check the other bytes than the status byte in result
646                  * to handle the case when a SCSI LLD sets result to
647                  * DRIVER_SENSE << 24 without setting SAM_STAT_CHECK_CONDITION.
648                  */
649                 if (scsi_status_is_good(result) && (result & ~0xff) == 0)
650                         return BLK_STS_OK;
651                 return BLK_STS_IOERR;
652         case DID_TRANSPORT_FAILFAST:
653                 return BLK_STS_TRANSPORT;
654         case DID_TARGET_FAILURE:
655                 set_host_byte(cmd, DID_OK);
656                 return BLK_STS_TARGET;
657         case DID_NEXUS_FAILURE:
658                 set_host_byte(cmd, DID_OK);
659                 return BLK_STS_NEXUS;
660         case DID_ALLOC_FAILURE:
661                 set_host_byte(cmd, DID_OK);
662                 return BLK_STS_NOSPC;
663         case DID_MEDIUM_ERROR:
664                 set_host_byte(cmd, DID_OK);
665                 return BLK_STS_MEDIUM;
666         default:
667                 return BLK_STS_IOERR;
668         }
669 }
670
671 /* Helper for scsi_io_completion() when "reprep" action required. */
672 static void scsi_io_completion_reprep(struct scsi_cmnd *cmd,
673                                       struct request_queue *q)
674 {
675         /* A new command will be prepared and issued. */
676         scsi_mq_requeue_cmd(cmd);
677 }
678
679 /* Helper for scsi_io_completion() when special action required. */
680 static void scsi_io_completion_action(struct scsi_cmnd *cmd, int result)
681 {
682         struct request_queue *q = cmd->device->request_queue;
683         struct request *req = cmd->request;
684         int level = 0;
685         enum {ACTION_FAIL, ACTION_REPREP, ACTION_RETRY,
686               ACTION_DELAYED_RETRY} action;
687         unsigned long wait_for = (cmd->allowed + 1) * req->timeout;
688         struct scsi_sense_hdr sshdr;
689         bool sense_valid;
690         bool sense_current = true;      /* false implies "deferred sense" */
691         blk_status_t blk_stat;
692
693         sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
694         if (sense_valid)
695                 sense_current = !scsi_sense_is_deferred(&sshdr);
696
697         blk_stat = scsi_result_to_blk_status(cmd, result);
698
699         if (host_byte(result) == DID_RESET) {
700                 /* Third party bus reset or reset for error recovery
701                  * reasons.  Just retry the command and see what
702                  * happens.
703                  */
704                 action = ACTION_RETRY;
705         } else if (sense_valid && sense_current) {
706                 switch (sshdr.sense_key) {
707                 case UNIT_ATTENTION:
708                         if (cmd->device->removable) {
709                                 /* Detected disc change.  Set a bit
710                                  * and quietly refuse further access.
711                                  */
712                                 cmd->device->changed = 1;
713                                 action = ACTION_FAIL;
714                         } else {
715                                 /* Must have been a power glitch, or a
716                                  * bus reset.  Could not have been a
717                                  * media change, so we just retry the
718                                  * command and see what happens.
719                                  */
720                                 action = ACTION_RETRY;
721                         }
722                         break;
723                 case ILLEGAL_REQUEST:
724                         /* If we had an ILLEGAL REQUEST returned, then
725                          * we may have performed an unsupported
726                          * command.  The only thing this should be
727                          * would be a ten byte read where only a six
728                          * byte read was supported.  Also, on a system
729                          * where READ CAPACITY failed, we may have
730                          * read past the end of the disk.
731                          */
732                         if ((cmd->device->use_10_for_rw &&
733                             sshdr.asc == 0x20 && sshdr.ascq == 0x00) &&
734                             (cmd->cmnd[0] == READ_10 ||
735                              cmd->cmnd[0] == WRITE_10)) {
736                                 /* This will issue a new 6-byte command. */
737                                 cmd->device->use_10_for_rw = 0;
738                                 action = ACTION_REPREP;
739                         } else if (sshdr.asc == 0x10) /* DIX */ {
740                                 action = ACTION_FAIL;
741                                 blk_stat = BLK_STS_PROTECTION;
742                         /* INVALID COMMAND OPCODE or INVALID FIELD IN CDB */
743                         } else if (sshdr.asc == 0x20 || sshdr.asc == 0x24) {
744                                 action = ACTION_FAIL;
745                                 blk_stat = BLK_STS_TARGET;
746                         } else
747                                 action = ACTION_FAIL;
748                         break;
749                 case ABORTED_COMMAND:
750                         action = ACTION_FAIL;
751                         if (sshdr.asc == 0x10) /* DIF */
752                                 blk_stat = BLK_STS_PROTECTION;
753                         break;
754                 case NOT_READY:
755                         /* If the device is in the process of becoming
756                          * ready, or has a temporary blockage, retry.
757                          */
758                         if (sshdr.asc == 0x04) {
759                                 switch (sshdr.ascq) {
760                                 case 0x01: /* becoming ready */
761                                 case 0x04: /* format in progress */
762                                 case 0x05: /* rebuild in progress */
763                                 case 0x06: /* recalculation in progress */
764                                 case 0x07: /* operation in progress */
765                                 case 0x08: /* Long write in progress */
766                                 case 0x09: /* self test in progress */
767                                 case 0x14: /* space allocation in progress */
768                                 case 0x1a: /* start stop unit in progress */
769                                 case 0x1b: /* sanitize in progress */
770                                 case 0x1d: /* configuration in progress */
771                                 case 0x24: /* depopulation in progress */
772                                         action = ACTION_DELAYED_RETRY;
773                                         break;
774                                 default:
775                                         action = ACTION_FAIL;
776                                         break;
777                                 }
778                         } else
779                                 action = ACTION_FAIL;
780                         break;
781                 case VOLUME_OVERFLOW:
782                         /* See SSC3rXX or current. */
783                         action = ACTION_FAIL;
784                         break;
785                 default:
786                         action = ACTION_FAIL;
787                         break;
788                 }
789         } else
790                 action = ACTION_FAIL;
791
792         if (action != ACTION_FAIL &&
793             time_before(cmd->jiffies_at_alloc + wait_for, jiffies))
794                 action = ACTION_FAIL;
795
796         switch (action) {
797         case ACTION_FAIL:
798                 /* Give up and fail the remainder of the request */
799                 if (!(req->rq_flags & RQF_QUIET)) {
800                         static DEFINE_RATELIMIT_STATE(_rs,
801                                         DEFAULT_RATELIMIT_INTERVAL,
802                                         DEFAULT_RATELIMIT_BURST);
803
804                         if (unlikely(scsi_logging_level))
805                                 level =
806                                      SCSI_LOG_LEVEL(SCSI_LOG_MLCOMPLETE_SHIFT,
807                                                     SCSI_LOG_MLCOMPLETE_BITS);
808
809                         /*
810                          * if logging is enabled the failure will be printed
811                          * in scsi_log_completion(), so avoid duplicate messages
812                          */
813                         if (!level && __ratelimit(&_rs)) {
814                                 scsi_print_result(cmd, NULL, FAILED);
815                                 if (driver_byte(result) == DRIVER_SENSE)
816                                         scsi_print_sense(cmd);
817                                 scsi_print_command(cmd);
818                         }
819                 }
820                 if (!scsi_end_request(req, blk_stat, blk_rq_err_bytes(req), 0))
821                         return;
822                 /*FALLTHRU*/
823         case ACTION_REPREP:
824                 scsi_io_completion_reprep(cmd, q);
825                 break;
826         case ACTION_RETRY:
827                 /* Retry the same command immediately */
828                 __scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY, false);
829                 break;
830         case ACTION_DELAYED_RETRY:
831                 /* Retry the same command after a delay */
832                 __scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY, false);
833                 break;
834         }
835 }
836
837 /*
838  * Helper for scsi_io_completion() when cmd->result is non-zero. Returns a
839  * new result that may suppress further error checking. Also modifies
840  * *blk_statp in some cases.
841  */
842 static int scsi_io_completion_nz_result(struct scsi_cmnd *cmd, int result,
843                                         blk_status_t *blk_statp)
844 {
845         bool sense_valid;
846         bool sense_current = true;      /* false implies "deferred sense" */
847         struct request *req = cmd->request;
848         struct scsi_sense_hdr sshdr;
849
850         sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
851         if (sense_valid)
852                 sense_current = !scsi_sense_is_deferred(&sshdr);
853
854         if (blk_rq_is_passthrough(req)) {
855                 if (sense_valid) {
856                         /*
857                          * SG_IO wants current and deferred errors
858                          */
859                         scsi_req(req)->sense_len =
860                                 min(8 + cmd->sense_buffer[7],
861                                     SCSI_SENSE_BUFFERSIZE);
862                 }
863                 if (sense_current)
864                         *blk_statp = scsi_result_to_blk_status(cmd, result);
865         } else if (blk_rq_bytes(req) == 0 && sense_current) {
866                 /*
867                  * Flush commands do not transfers any data, and thus cannot use
868                  * good_bytes != blk_rq_bytes(req) as the signal for an error.
869                  * This sets *blk_statp explicitly for the problem case.
870                  */
871                 *blk_statp = scsi_result_to_blk_status(cmd, result);
872         }
873         /*
874          * Recovered errors need reporting, but they're always treated as
875          * success, so fiddle the result code here.  For passthrough requests
876          * we already took a copy of the original into sreq->result which
877          * is what gets returned to the user
878          */
879         if (sense_valid && (sshdr.sense_key == RECOVERED_ERROR)) {
880                 bool do_print = true;
881                 /*
882                  * if ATA PASS-THROUGH INFORMATION AVAILABLE [0x0, 0x1d]
883                  * skip print since caller wants ATA registers. Only occurs
884                  * on SCSI ATA PASS_THROUGH commands when CK_COND=1
885                  */
886                 if ((sshdr.asc == 0x0) && (sshdr.ascq == 0x1d))
887                         do_print = false;
888                 else if (req->rq_flags & RQF_QUIET)
889                         do_print = false;
890                 if (do_print)
891                         scsi_print_sense(cmd);
892                 result = 0;
893                 /* for passthrough, *blk_statp may be set */
894                 *blk_statp = BLK_STS_OK;
895         }
896         /*
897          * Another corner case: the SCSI status byte is non-zero but 'good'.
898          * Example: PRE-FETCH command returns SAM_STAT_CONDITION_MET when
899          * it is able to fit nominated LBs in its cache (and SAM_STAT_GOOD
900          * if it can't fit). Treat SAM_STAT_CONDITION_MET and the related
901          * intermediate statuses (both obsolete in SAM-4) as good.
902          */
903         if (status_byte(result) && scsi_status_is_good(result)) {
904                 result = 0;
905                 *blk_statp = BLK_STS_OK;
906         }
907         return result;
908 }
909
910 /*
911  * Function:    scsi_io_completion()
912  *
913  * Purpose:     Completion processing for block device I/O requests.
914  *
915  * Arguments:   cmd   - command that is finished.
916  *
917  * Lock status: Assumed that no lock is held upon entry.
918  *
919  * Returns:     Nothing
920  *
921  * Notes:       We will finish off the specified number of sectors.  If we
922  *              are done, the command block will be released and the queue
923  *              function will be goosed.  If we are not done then we have to
924  *              figure out what to do next:
925  *
926  *              a) We can call scsi_requeue_command().  The request
927  *                 will be unprepared and put back on the queue.  Then
928  *                 a new command will be created for it.  This should
929  *                 be used if we made forward progress, or if we want
930  *                 to switch from READ(10) to READ(6) for example.
931  *
932  *              b) We can call __scsi_queue_insert().  The request will
933  *                 be put back on the queue and retried using the same
934  *                 command as before, possibly after a delay.
935  *
936  *              c) We can call scsi_end_request() with blk_stat other than
937  *                 BLK_STS_OK, to fail the remainder of the request.
938  */
939 void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes)
940 {
941         int result = cmd->result;
942         struct request_queue *q = cmd->device->request_queue;
943         struct request *req = cmd->request;
944         blk_status_t blk_stat = BLK_STS_OK;
945
946         if (unlikely(result))   /* a nz result may or may not be an error */
947                 result = scsi_io_completion_nz_result(cmd, result, &blk_stat);
948
949         if (unlikely(blk_rq_is_passthrough(req))) {
950                 /*
951                  * scsi_result_to_blk_status may have reset the host_byte
952                  */
953                 scsi_req(req)->result = cmd->result;
954                 scsi_req(req)->resid_len = scsi_get_resid(cmd);
955
956                 if (unlikely(scsi_bidi_cmnd(cmd))) {
957                         /*
958                          * Bidi commands Must be complete as a whole,
959                          * both sides at once.
960                          */
961                         scsi_req(req->next_rq)->resid_len = scsi_in(cmd)->resid;
962                         if (scsi_end_request(req, BLK_STS_OK, blk_rq_bytes(req),
963                                         blk_rq_bytes(req->next_rq)))
964                                 WARN_ONCE(true,
965                                           "Bidi command with remaining bytes");
966                         return;
967                 }
968         }
969
970         /* no bidi support yet, other than in pass-through */
971         if (unlikely(blk_bidi_rq(req))) {
972                 WARN_ONCE(true, "Only support bidi command in passthrough");
973                 scmd_printk(KERN_ERR, cmd, "Killing bidi command\n");
974                 if (scsi_end_request(req, BLK_STS_IOERR, blk_rq_bytes(req),
975                                      blk_rq_bytes(req->next_rq)))
976                         WARN_ONCE(true, "Bidi command with remaining bytes");
977                 return;
978         }
979
980         /*
981          * Next deal with any sectors which we were able to correctly
982          * handle.
983          */
984         SCSI_LOG_HLCOMPLETE(1, scmd_printk(KERN_INFO, cmd,
985                 "%u sectors total, %d bytes done.\n",
986                 blk_rq_sectors(req), good_bytes));
987
988         /*
989          * Next deal with any sectors which we were able to correctly
990          * handle. Failed, zero length commands always need to drop down
991          * to retry code. Fast path should return in this block.
992          */
993         if (likely(blk_rq_bytes(req) > 0 || blk_stat == BLK_STS_OK)) {
994                 if (likely(!scsi_end_request(req, blk_stat, good_bytes, 0)))
995                         return; /* no bytes remaining */
996         }
997
998         /* Kill remainder if no retries. */
999         if (unlikely(blk_stat && scsi_noretry_cmd(cmd))) {
1000                 if (scsi_end_request(req, blk_stat, blk_rq_bytes(req), 0))
1001                         WARN_ONCE(true,
1002                             "Bytes remaining after failed, no-retry command");
1003                 return;
1004         }
1005
1006         /*
1007          * If there had been no error, but we have leftover bytes in the
1008          * requeues just queue the command up again.
1009          */
1010         if (likely(result == 0))
1011                 scsi_io_completion_reprep(cmd, q);
1012         else
1013                 scsi_io_completion_action(cmd, result);
1014 }
1015
1016 static blk_status_t scsi_init_sgtable(struct request *req,
1017                 struct scsi_data_buffer *sdb)
1018 {
1019         int count;
1020
1021         /*
1022          * If sg table allocation fails, requeue request later.
1023          */
1024         if (unlikely(sg_alloc_table_chained(&sdb->table,
1025                         blk_rq_nr_phys_segments(req), sdb->table.sgl)))
1026                 return BLK_STS_RESOURCE;
1027
1028         /* 
1029          * Next, walk the list, and fill in the addresses and sizes of
1030          * each segment.
1031          */
1032         count = blk_rq_map_sg(req->q, req, sdb->table.sgl);
1033         BUG_ON(count > sdb->table.nents);
1034         sdb->table.nents = count;
1035         sdb->length = blk_rq_payload_bytes(req);
1036         return BLK_STS_OK;
1037 }
1038
1039 /*
1040  * Function:    scsi_init_io()
1041  *
1042  * Purpose:     SCSI I/O initialize function.
1043  *
1044  * Arguments:   cmd   - Command descriptor we wish to initialize
1045  *
1046  * Returns:     BLK_STS_OK on success
1047  *              BLK_STS_RESOURCE if the failure is retryable
1048  *              BLK_STS_IOERR if the failure is fatal
1049  */
1050 blk_status_t scsi_init_io(struct scsi_cmnd *cmd)
1051 {
1052         struct request *rq = cmd->request;
1053         blk_status_t ret;
1054
1055         if (WARN_ON_ONCE(!blk_rq_nr_phys_segments(rq)))
1056                 return BLK_STS_IOERR;
1057
1058         ret = scsi_init_sgtable(rq, &cmd->sdb);
1059         if (ret)
1060                 return ret;
1061
1062         if (blk_bidi_rq(rq)) {
1063                 ret = scsi_init_sgtable(rq->next_rq, rq->next_rq->special);
1064                 if (ret)
1065                         goto out_free_sgtables;
1066         }
1067
1068         if (blk_integrity_rq(rq)) {
1069                 struct scsi_data_buffer *prot_sdb = cmd->prot_sdb;
1070                 int ivecs, count;
1071
1072                 if (WARN_ON_ONCE(!prot_sdb)) {
1073                         /*
1074                          * This can happen if someone (e.g. multipath)
1075                          * queues a command to a device on an adapter
1076                          * that does not support DIX.
1077                          */
1078                         ret = BLK_STS_IOERR;
1079                         goto out_free_sgtables;
1080                 }
1081
1082                 ivecs = blk_rq_count_integrity_sg(rq->q, rq->bio);
1083
1084                 if (sg_alloc_table_chained(&prot_sdb->table, ivecs,
1085                                 prot_sdb->table.sgl)) {
1086                         ret = BLK_STS_RESOURCE;
1087                         goto out_free_sgtables;
1088                 }
1089
1090                 count = blk_rq_map_integrity_sg(rq->q, rq->bio,
1091                                                 prot_sdb->table.sgl);
1092                 BUG_ON(count > ivecs);
1093                 BUG_ON(count > queue_max_integrity_segments(rq->q));
1094
1095                 cmd->prot_sdb = prot_sdb;
1096                 cmd->prot_sdb->table.nents = count;
1097         }
1098
1099         return BLK_STS_OK;
1100 out_free_sgtables:
1101         scsi_mq_free_sgtables(cmd);
1102         return ret;
1103 }
1104 EXPORT_SYMBOL(scsi_init_io);
1105
1106 /**
1107  * scsi_initialize_rq - initialize struct scsi_cmnd partially
1108  * @rq: Request associated with the SCSI command to be initialized.
1109  *
1110  * This function initializes the members of struct scsi_cmnd that must be
1111  * initialized before request processing starts and that won't be
1112  * reinitialized if a SCSI command is requeued.
1113  *
1114  * Called from inside blk_get_request() for pass-through requests and from
1115  * inside scsi_init_command() for filesystem requests.
1116  */
1117 static void scsi_initialize_rq(struct request *rq)
1118 {
1119         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1120
1121         scsi_req_init(&cmd->req);
1122         init_rcu_head(&cmd->rcu);
1123         cmd->jiffies_at_alloc = jiffies;
1124         cmd->retries = 0;
1125 }
1126
1127 /* Add a command to the list used by the aacraid and dpt_i2o drivers */
1128 void scsi_add_cmd_to_list(struct scsi_cmnd *cmd)
1129 {
1130         struct scsi_device *sdev = cmd->device;
1131         struct Scsi_Host *shost = sdev->host;
1132         unsigned long flags;
1133
1134         if (shost->use_cmd_list) {
1135                 spin_lock_irqsave(&sdev->list_lock, flags);
1136                 list_add_tail(&cmd->list, &sdev->cmd_list);
1137                 spin_unlock_irqrestore(&sdev->list_lock, flags);
1138         }
1139 }
1140
1141 /* Remove a command from the list used by the aacraid and dpt_i2o drivers */
1142 void scsi_del_cmd_from_list(struct scsi_cmnd *cmd)
1143 {
1144         struct scsi_device *sdev = cmd->device;
1145         struct Scsi_Host *shost = sdev->host;
1146         unsigned long flags;
1147
1148         if (shost->use_cmd_list) {
1149                 spin_lock_irqsave(&sdev->list_lock, flags);
1150                 BUG_ON(list_empty(&cmd->list));
1151                 list_del_init(&cmd->list);
1152                 spin_unlock_irqrestore(&sdev->list_lock, flags);
1153         }
1154 }
1155
1156 /* Called after a request has been started. */
1157 void scsi_init_command(struct scsi_device *dev, struct scsi_cmnd *cmd)
1158 {
1159         void *buf = cmd->sense_buffer;
1160         void *prot = cmd->prot_sdb;
1161         struct request *rq = blk_mq_rq_from_pdu(cmd);
1162         unsigned int flags = cmd->flags & SCMD_PRESERVED_FLAGS;
1163         unsigned long jiffies_at_alloc;
1164         int retries;
1165
1166         if (!blk_rq_is_scsi(rq) && !(flags & SCMD_INITIALIZED)) {
1167                 flags |= SCMD_INITIALIZED;
1168                 scsi_initialize_rq(rq);
1169         }
1170
1171         jiffies_at_alloc = cmd->jiffies_at_alloc;
1172         retries = cmd->retries;
1173         /* zero out the cmd, except for the embedded scsi_request */
1174         memset((char *)cmd + sizeof(cmd->req), 0,
1175                 sizeof(*cmd) - sizeof(cmd->req) + dev->host->hostt->cmd_size);
1176
1177         cmd->device = dev;
1178         cmd->sense_buffer = buf;
1179         cmd->prot_sdb = prot;
1180         cmd->flags = flags;
1181         INIT_DELAYED_WORK(&cmd->abort_work, scmd_eh_abort_handler);
1182         cmd->jiffies_at_alloc = jiffies_at_alloc;
1183         cmd->retries = retries;
1184
1185         scsi_add_cmd_to_list(cmd);
1186 }
1187
1188 static blk_status_t scsi_setup_scsi_cmnd(struct scsi_device *sdev,
1189                 struct request *req)
1190 {
1191         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1192
1193         /*
1194          * Passthrough requests may transfer data, in which case they must
1195          * a bio attached to them.  Or they might contain a SCSI command
1196          * that does not transfer data, in which case they may optionally
1197          * submit a request without an attached bio.
1198          */
1199         if (req->bio) {
1200                 blk_status_t ret = scsi_init_io(cmd);
1201                 if (unlikely(ret != BLK_STS_OK))
1202                         return ret;
1203         } else {
1204                 BUG_ON(blk_rq_bytes(req));
1205
1206                 memset(&cmd->sdb, 0, sizeof(cmd->sdb));
1207         }
1208
1209         cmd->cmd_len = scsi_req(req)->cmd_len;
1210         cmd->cmnd = scsi_req(req)->cmd;
1211         cmd->transfersize = blk_rq_bytes(req);
1212         cmd->allowed = scsi_req(req)->retries;
1213         return BLK_STS_OK;
1214 }
1215
1216 /*
1217  * Setup a normal block command.  These are simple request from filesystems
1218  * that still need to be translated to SCSI CDBs from the ULD.
1219  */
1220 static blk_status_t scsi_setup_fs_cmnd(struct scsi_device *sdev,
1221                 struct request *req)
1222 {
1223         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1224
1225         if (unlikely(sdev->handler && sdev->handler->prep_fn)) {
1226                 blk_status_t ret = sdev->handler->prep_fn(sdev, req);
1227                 if (ret != BLK_STS_OK)
1228                         return ret;
1229         }
1230
1231         cmd->cmnd = scsi_req(req)->cmd = scsi_req(req)->__cmd;
1232         memset(cmd->cmnd, 0, BLK_MAX_CDB);
1233         return scsi_cmd_to_driver(cmd)->init_command(cmd);
1234 }
1235
1236 static blk_status_t scsi_setup_cmnd(struct scsi_device *sdev,
1237                 struct request *req)
1238 {
1239         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1240
1241         if (!blk_rq_bytes(req))
1242                 cmd->sc_data_direction = DMA_NONE;
1243         else if (rq_data_dir(req) == WRITE)
1244                 cmd->sc_data_direction = DMA_TO_DEVICE;
1245         else
1246                 cmd->sc_data_direction = DMA_FROM_DEVICE;
1247
1248         if (blk_rq_is_scsi(req))
1249                 return scsi_setup_scsi_cmnd(sdev, req);
1250         else
1251                 return scsi_setup_fs_cmnd(sdev, req);
1252 }
1253
1254 static blk_status_t
1255 scsi_prep_state_check(struct scsi_device *sdev, struct request *req)
1256 {
1257         switch (sdev->sdev_state) {
1258         case SDEV_OFFLINE:
1259         case SDEV_TRANSPORT_OFFLINE:
1260                 /*
1261                  * If the device is offline we refuse to process any
1262                  * commands.  The device must be brought online
1263                  * before trying any recovery commands.
1264                  */
1265                 sdev_printk(KERN_ERR, sdev,
1266                             "rejecting I/O to offline device\n");
1267                 return BLK_STS_IOERR;
1268         case SDEV_DEL:
1269                 /*
1270                  * If the device is fully deleted, we refuse to
1271                  * process any commands as well.
1272                  */
1273                 sdev_printk(KERN_ERR, sdev,
1274                             "rejecting I/O to dead device\n");
1275                 return BLK_STS_IOERR;
1276         case SDEV_BLOCK:
1277         case SDEV_CREATED_BLOCK:
1278                 return BLK_STS_RESOURCE;
1279         case SDEV_QUIESCE:
1280                 /*
1281                  * If the devices is blocked we defer normal commands.
1282                  */
1283                 if (req && !(req->rq_flags & RQF_PREEMPT))
1284                         return BLK_STS_RESOURCE;
1285                 return BLK_STS_OK;
1286         default:
1287                 /*
1288                  * For any other not fully online state we only allow
1289                  * special commands.  In particular any user initiated
1290                  * command is not allowed.
1291                  */
1292                 if (req && !(req->rq_flags & RQF_PREEMPT))
1293                         return BLK_STS_IOERR;
1294                 return BLK_STS_OK;
1295         }
1296 }
1297
1298 /*
1299  * scsi_dev_queue_ready: if we can send requests to sdev, return 1 else
1300  * return 0.
1301  *
1302  * Called with the queue_lock held.
1303  */
1304 static inline int scsi_dev_queue_ready(struct request_queue *q,
1305                                   struct scsi_device *sdev)
1306 {
1307         unsigned int busy;
1308
1309         busy = atomic_inc_return(&sdev->device_busy) - 1;
1310         if (atomic_read(&sdev->device_blocked)) {
1311                 if (busy)
1312                         goto out_dec;
1313
1314                 /*
1315                  * unblock after device_blocked iterates to zero
1316                  */
1317                 if (atomic_dec_return(&sdev->device_blocked) > 0)
1318                         goto out_dec;
1319                 SCSI_LOG_MLQUEUE(3, sdev_printk(KERN_INFO, sdev,
1320                                    "unblocking device at zero depth\n"));
1321         }
1322
1323         if (busy >= sdev->queue_depth)
1324                 goto out_dec;
1325
1326         return 1;
1327 out_dec:
1328         atomic_dec(&sdev->device_busy);
1329         return 0;
1330 }
1331
1332 /*
1333  * scsi_target_queue_ready: checks if there we can send commands to target
1334  * @sdev: scsi device on starget to check.
1335  */
1336 static inline int scsi_target_queue_ready(struct Scsi_Host *shost,
1337                                            struct scsi_device *sdev)
1338 {
1339         struct scsi_target *starget = scsi_target(sdev);
1340         unsigned int busy;
1341
1342         if (starget->single_lun) {
1343                 spin_lock_irq(shost->host_lock);
1344                 if (starget->starget_sdev_user &&
1345                     starget->starget_sdev_user != sdev) {
1346                         spin_unlock_irq(shost->host_lock);
1347                         return 0;
1348                 }
1349                 starget->starget_sdev_user = sdev;
1350                 spin_unlock_irq(shost->host_lock);
1351         }
1352
1353         if (starget->can_queue <= 0)
1354                 return 1;
1355
1356         busy = atomic_inc_return(&starget->target_busy) - 1;
1357         if (atomic_read(&starget->target_blocked) > 0) {
1358                 if (busy)
1359                         goto starved;
1360
1361                 /*
1362                  * unblock after target_blocked iterates to zero
1363                  */
1364                 if (atomic_dec_return(&starget->target_blocked) > 0)
1365                         goto out_dec;
1366
1367                 SCSI_LOG_MLQUEUE(3, starget_printk(KERN_INFO, starget,
1368                                  "unblocking target at zero depth\n"));
1369         }
1370
1371         if (busy >= starget->can_queue)
1372                 goto starved;
1373
1374         return 1;
1375
1376 starved:
1377         spin_lock_irq(shost->host_lock);
1378         list_move_tail(&sdev->starved_entry, &shost->starved_list);
1379         spin_unlock_irq(shost->host_lock);
1380 out_dec:
1381         if (starget->can_queue > 0)
1382                 atomic_dec(&starget->target_busy);
1383         return 0;
1384 }
1385
1386 /*
1387  * scsi_host_queue_ready: if we can send requests to shost, return 1 else
1388  * return 0. We must end up running the queue again whenever 0 is
1389  * returned, else IO can hang.
1390  */
1391 static inline int scsi_host_queue_ready(struct request_queue *q,
1392                                    struct Scsi_Host *shost,
1393                                    struct scsi_device *sdev)
1394 {
1395         unsigned int busy;
1396
1397         if (scsi_host_in_recovery(shost))
1398                 return 0;
1399
1400         busy = atomic_inc_return(&shost->host_busy) - 1;
1401         if (atomic_read(&shost->host_blocked) > 0) {
1402                 if (busy)
1403                         goto starved;
1404
1405                 /*
1406                  * unblock after host_blocked iterates to zero
1407                  */
1408                 if (atomic_dec_return(&shost->host_blocked) > 0)
1409                         goto out_dec;
1410
1411                 SCSI_LOG_MLQUEUE(3,
1412                         shost_printk(KERN_INFO, shost,
1413                                      "unblocking host at zero depth\n"));
1414         }
1415
1416         if (shost->can_queue > 0 && busy >= shost->can_queue)
1417                 goto starved;
1418         if (shost->host_self_blocked)
1419                 goto starved;
1420
1421         /* We're OK to process the command, so we can't be starved */
1422         if (!list_empty(&sdev->starved_entry)) {
1423                 spin_lock_irq(shost->host_lock);
1424                 if (!list_empty(&sdev->starved_entry))
1425                         list_del_init(&sdev->starved_entry);
1426                 spin_unlock_irq(shost->host_lock);
1427         }
1428
1429         return 1;
1430
1431 starved:
1432         spin_lock_irq(shost->host_lock);
1433         if (list_empty(&sdev->starved_entry))
1434                 list_add_tail(&sdev->starved_entry, &shost->starved_list);
1435         spin_unlock_irq(shost->host_lock);
1436 out_dec:
1437         scsi_dec_host_busy(shost);
1438         return 0;
1439 }
1440
1441 /*
1442  * Busy state exporting function for request stacking drivers.
1443  *
1444  * For efficiency, no lock is taken to check the busy state of
1445  * shost/starget/sdev, since the returned value is not guaranteed and
1446  * may be changed after request stacking drivers call the function,
1447  * regardless of taking lock or not.
1448  *
1449  * When scsi can't dispatch I/Os anymore and needs to kill I/Os scsi
1450  * needs to return 'not busy'. Otherwise, request stacking drivers
1451  * may hold requests forever.
1452  */
1453 static bool scsi_mq_lld_busy(struct request_queue *q)
1454 {
1455         struct scsi_device *sdev = q->queuedata;
1456         struct Scsi_Host *shost;
1457
1458         if (blk_queue_dying(q))
1459                 return false;
1460
1461         shost = sdev->host;
1462
1463         /*
1464          * Ignore host/starget busy state.
1465          * Since block layer does not have a concept of fairness across
1466          * multiple queues, congestion of host/starget needs to be handled
1467          * in SCSI layer.
1468          */
1469         if (scsi_host_in_recovery(shost) || scsi_device_is_busy(sdev))
1470                 return true;
1471
1472         return false;
1473 }
1474
1475 static void scsi_softirq_done(struct request *rq)
1476 {
1477         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1478         unsigned long wait_for = (cmd->allowed + 1) * rq->timeout;
1479         int disposition;
1480
1481         INIT_LIST_HEAD(&cmd->eh_entry);
1482
1483         atomic_inc(&cmd->device->iodone_cnt);
1484         if (cmd->result)
1485                 atomic_inc(&cmd->device->ioerr_cnt);
1486
1487         disposition = scsi_decide_disposition(cmd);
1488         if (disposition != SUCCESS &&
1489             time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) {
1490                 sdev_printk(KERN_ERR, cmd->device,
1491                             "timing out command, waited %lus\n",
1492                             wait_for/HZ);
1493                 disposition = SUCCESS;
1494         }
1495
1496         scsi_log_completion(cmd, disposition);
1497
1498         switch (disposition) {
1499                 case SUCCESS:
1500                         scsi_finish_command(cmd);
1501                         break;
1502                 case NEEDS_RETRY:
1503                         scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY);
1504                         break;
1505                 case ADD_TO_MLQUEUE:
1506                         scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY);
1507                         break;
1508                 default:
1509                         scsi_eh_scmd_add(cmd);
1510                         break;
1511         }
1512 }
1513
1514 /**
1515  * scsi_dispatch_command - Dispatch a command to the low-level driver.
1516  * @cmd: command block we are dispatching.
1517  *
1518  * Return: nonzero return request was rejected and device's queue needs to be
1519  * plugged.
1520  */
1521 static int scsi_dispatch_cmd(struct scsi_cmnd *cmd)
1522 {
1523         struct Scsi_Host *host = cmd->device->host;
1524         int rtn = 0;
1525
1526         atomic_inc(&cmd->device->iorequest_cnt);
1527
1528         /* check if the device is still usable */
1529         if (unlikely(cmd->device->sdev_state == SDEV_DEL)) {
1530                 /* in SDEV_DEL we error all commands. DID_NO_CONNECT
1531                  * returns an immediate error upwards, and signals
1532                  * that the device is no longer present */
1533                 cmd->result = DID_NO_CONNECT << 16;
1534                 goto done;
1535         }
1536
1537         /* Check to see if the scsi lld made this device blocked. */
1538         if (unlikely(scsi_device_blocked(cmd->device))) {
1539                 /*
1540                  * in blocked state, the command is just put back on
1541                  * the device queue.  The suspend state has already
1542                  * blocked the queue so future requests should not
1543                  * occur until the device transitions out of the
1544                  * suspend state.
1545                  */
1546                 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1547                         "queuecommand : device blocked\n"));
1548                 return SCSI_MLQUEUE_DEVICE_BUSY;
1549         }
1550
1551         /* Store the LUN value in cmnd, if needed. */
1552         if (cmd->device->lun_in_cdb)
1553                 cmd->cmnd[1] = (cmd->cmnd[1] & 0x1f) |
1554                                (cmd->device->lun << 5 & 0xe0);
1555
1556         scsi_log_send(cmd);
1557
1558         /*
1559          * Before we queue this command, check if the command
1560          * length exceeds what the host adapter can handle.
1561          */
1562         if (cmd->cmd_len > cmd->device->host->max_cmd_len) {
1563                 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1564                                "queuecommand : command too long. "
1565                                "cdb_size=%d host->max_cmd_len=%d\n",
1566                                cmd->cmd_len, cmd->device->host->max_cmd_len));
1567                 cmd->result = (DID_ABORT << 16);
1568                 goto done;
1569         }
1570
1571         if (unlikely(host->shost_state == SHOST_DEL)) {
1572                 cmd->result = (DID_NO_CONNECT << 16);
1573                 goto done;
1574
1575         }
1576
1577         trace_scsi_dispatch_cmd_start(cmd);
1578         rtn = host->hostt->queuecommand(host, cmd);
1579         if (rtn) {
1580                 trace_scsi_dispatch_cmd_error(cmd, rtn);
1581                 if (rtn != SCSI_MLQUEUE_DEVICE_BUSY &&
1582                     rtn != SCSI_MLQUEUE_TARGET_BUSY)
1583                         rtn = SCSI_MLQUEUE_HOST_BUSY;
1584
1585                 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1586                         "queuecommand : request rejected\n"));
1587         }
1588
1589         return rtn;
1590  done:
1591         cmd->scsi_done(cmd);
1592         return 0;
1593 }
1594
1595 /* Size in bytes of the sg-list stored in the scsi-mq command-private data. */
1596 static unsigned int scsi_mq_sgl_size(struct Scsi_Host *shost)
1597 {
1598         return min_t(unsigned int, shost->sg_tablesize, SG_CHUNK_SIZE) *
1599                 sizeof(struct scatterlist);
1600 }
1601
1602 static blk_status_t scsi_mq_prep_fn(struct request *req)
1603 {
1604         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1605         struct scsi_device *sdev = req->q->queuedata;
1606         struct Scsi_Host *shost = sdev->host;
1607         struct scatterlist *sg;
1608
1609         scsi_init_command(sdev, cmd);
1610
1611         req->special = cmd;
1612
1613         cmd->request = req;
1614
1615         cmd->tag = req->tag;
1616         cmd->prot_op = SCSI_PROT_NORMAL;
1617
1618         sg = (void *)cmd + sizeof(struct scsi_cmnd) + shost->hostt->cmd_size;
1619         cmd->sdb.table.sgl = sg;
1620
1621         if (scsi_host_get_prot(shost)) {
1622                 memset(cmd->prot_sdb, 0, sizeof(struct scsi_data_buffer));
1623
1624                 cmd->prot_sdb->table.sgl =
1625                         (struct scatterlist *)(cmd->prot_sdb + 1);
1626         }
1627
1628         if (blk_bidi_rq(req)) {
1629                 struct request *next_rq = req->next_rq;
1630                 struct scsi_data_buffer *bidi_sdb = blk_mq_rq_to_pdu(next_rq);
1631
1632                 memset(bidi_sdb, 0, sizeof(struct scsi_data_buffer));
1633                 bidi_sdb->table.sgl =
1634                         (struct scatterlist *)(bidi_sdb + 1);
1635
1636                 next_rq->special = bidi_sdb;
1637         }
1638
1639         blk_mq_start_request(req);
1640
1641         return scsi_setup_cmnd(sdev, req);
1642 }
1643
1644 static void scsi_mq_done(struct scsi_cmnd *cmd)
1645 {
1646         if (unlikely(test_and_set_bit(SCMD_STATE_COMPLETE, &cmd->state)))
1647                 return;
1648         trace_scsi_dispatch_cmd_done(cmd);
1649
1650         /*
1651          * If the block layer didn't complete the request due to a timeout
1652          * injection, scsi must clear its internal completed state so that the
1653          * timeout handler will see it needs to escalate its own error
1654          * recovery.
1655          */
1656         if (unlikely(!blk_mq_complete_request(cmd->request)))
1657                 clear_bit(SCMD_STATE_COMPLETE, &cmd->state);
1658 }
1659
1660 static void scsi_mq_put_budget(struct blk_mq_hw_ctx *hctx)
1661 {
1662         struct request_queue *q = hctx->queue;
1663         struct scsi_device *sdev = q->queuedata;
1664
1665         atomic_dec(&sdev->device_busy);
1666         put_device(&sdev->sdev_gendev);
1667 }
1668
1669 static bool scsi_mq_get_budget(struct blk_mq_hw_ctx *hctx)
1670 {
1671         struct request_queue *q = hctx->queue;
1672         struct scsi_device *sdev = q->queuedata;
1673
1674         if (!get_device(&sdev->sdev_gendev))
1675                 goto out;
1676         if (!scsi_dev_queue_ready(q, sdev))
1677                 goto out_put_device;
1678
1679         return true;
1680
1681 out_put_device:
1682         put_device(&sdev->sdev_gendev);
1683 out:
1684         if (atomic_read(&sdev->device_busy) == 0 && !scsi_device_blocked(sdev))
1685                 blk_mq_delay_run_hw_queue(hctx, SCSI_QUEUE_DELAY);
1686         return false;
1687 }
1688
1689 static blk_status_t scsi_queue_rq(struct blk_mq_hw_ctx *hctx,
1690                          const struct blk_mq_queue_data *bd)
1691 {
1692         struct request *req = bd->rq;
1693         struct request_queue *q = req->q;
1694         struct scsi_device *sdev = q->queuedata;
1695         struct Scsi_Host *shost = sdev->host;
1696         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1697         blk_status_t ret;
1698         int reason;
1699
1700         /*
1701          * If the device is not in running state we will reject some or all
1702          * commands.
1703          */
1704         if (unlikely(sdev->sdev_state != SDEV_RUNNING)) {
1705                 ret = scsi_prep_state_check(sdev, req);
1706                 if (ret != BLK_STS_OK)
1707                         goto out_put_budget;
1708         }
1709
1710         ret = BLK_STS_RESOURCE;
1711         if (!scsi_target_queue_ready(shost, sdev))
1712                 goto out_put_budget;
1713         if (!scsi_host_queue_ready(q, shost, sdev))
1714                 goto out_dec_target_busy;
1715
1716         clear_bit(SCMD_STATE_COMPLETE, &cmd->state);
1717         if (!(req->rq_flags & RQF_DONTPREP)) {
1718                 ret = scsi_mq_prep_fn(req);
1719                 if (ret != BLK_STS_OK)
1720                         goto out_dec_host_busy;
1721                 req->rq_flags |= RQF_DONTPREP;
1722         } else {
1723                 blk_mq_start_request(req);
1724         }
1725
1726         if (sdev->simple_tags)
1727                 cmd->flags |= SCMD_TAGGED;
1728         else
1729                 cmd->flags &= ~SCMD_TAGGED;
1730
1731         scsi_init_cmd_errh(cmd);
1732         cmd->scsi_done = scsi_mq_done;
1733
1734         reason = scsi_dispatch_cmd(cmd);
1735         if (reason) {
1736                 scsi_set_blocked(cmd, reason);
1737                 ret = BLK_STS_RESOURCE;
1738                 goto out_dec_host_busy;
1739         }
1740
1741         return BLK_STS_OK;
1742
1743 out_dec_host_busy:
1744         scsi_dec_host_busy(shost);
1745 out_dec_target_busy:
1746         if (scsi_target(sdev)->can_queue > 0)
1747                 atomic_dec(&scsi_target(sdev)->target_busy);
1748 out_put_budget:
1749         scsi_mq_put_budget(hctx);
1750         switch (ret) {
1751         case BLK_STS_OK:
1752                 break;
1753         case BLK_STS_RESOURCE:
1754                 if (atomic_read(&sdev->device_busy) ||
1755                     scsi_device_blocked(sdev))
1756                         ret = BLK_STS_DEV_RESOURCE;
1757                 break;
1758         default:
1759                 /*
1760                  * Make sure to release all allocated ressources when
1761                  * we hit an error, as we will never see this command
1762                  * again.
1763                  */
1764                 if (req->rq_flags & RQF_DONTPREP)
1765                         scsi_mq_uninit_cmd(cmd);
1766                 break;
1767         }
1768         return ret;
1769 }
1770
1771 static enum blk_eh_timer_return scsi_timeout(struct request *req,
1772                 bool reserved)
1773 {
1774         if (reserved)
1775                 return BLK_EH_RESET_TIMER;
1776         return scsi_times_out(req);
1777 }
1778
1779 static int scsi_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
1780                                 unsigned int hctx_idx, unsigned int numa_node)
1781 {
1782         struct Scsi_Host *shost = set->driver_data;
1783         const bool unchecked_isa_dma = shost->unchecked_isa_dma;
1784         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1785         struct scatterlist *sg;
1786
1787         if (unchecked_isa_dma)
1788                 cmd->flags |= SCMD_UNCHECKED_ISA_DMA;
1789         cmd->sense_buffer = scsi_alloc_sense_buffer(unchecked_isa_dma,
1790                                                     GFP_KERNEL, numa_node);
1791         if (!cmd->sense_buffer)
1792                 return -ENOMEM;
1793         cmd->req.sense = cmd->sense_buffer;
1794
1795         if (scsi_host_get_prot(shost)) {
1796                 sg = (void *)cmd + sizeof(struct scsi_cmnd) +
1797                         shost->hostt->cmd_size;
1798                 cmd->prot_sdb = (void *)sg + scsi_mq_sgl_size(shost);
1799         }
1800
1801         return 0;
1802 }
1803
1804 static void scsi_mq_exit_request(struct blk_mq_tag_set *set, struct request *rq,
1805                                  unsigned int hctx_idx)
1806 {
1807         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1808
1809         scsi_free_sense_buffer(cmd->flags & SCMD_UNCHECKED_ISA_DMA,
1810                                cmd->sense_buffer);
1811 }
1812
1813 static int scsi_map_queues(struct blk_mq_tag_set *set)
1814 {
1815         struct Scsi_Host *shost = container_of(set, struct Scsi_Host, tag_set);
1816
1817         if (shost->hostt->map_queues)
1818                 return shost->hostt->map_queues(shost);
1819         return blk_mq_map_queues(&set->map[0]);
1820 }
1821
1822 void __scsi_init_queue(struct Scsi_Host *shost, struct request_queue *q)
1823 {
1824         struct device *dev = shost->dma_dev;
1825
1826         /*
1827          * this limit is imposed by hardware restrictions
1828          */
1829         blk_queue_max_segments(q, min_t(unsigned short, shost->sg_tablesize,
1830                                         SG_MAX_SEGMENTS));
1831
1832         if (scsi_host_prot_dma(shost)) {
1833                 shost->sg_prot_tablesize =
1834                         min_not_zero(shost->sg_prot_tablesize,
1835                                      (unsigned short)SCSI_MAX_PROT_SG_SEGMENTS);
1836                 BUG_ON(shost->sg_prot_tablesize < shost->sg_tablesize);
1837                 blk_queue_max_integrity_segments(q, shost->sg_prot_tablesize);
1838         }
1839
1840         blk_queue_max_hw_sectors(q, shost->max_sectors);
1841         if (shost->unchecked_isa_dma)
1842                 blk_queue_bounce_limit(q, BLK_BOUNCE_ISA);
1843         blk_queue_segment_boundary(q, shost->dma_boundary);
1844         dma_set_seg_boundary(dev, shost->dma_boundary);
1845
1846         blk_queue_max_segment_size(q, shost->max_segment_size);
1847         dma_set_max_seg_size(dev, shost->max_segment_size);
1848
1849         /*
1850          * Set a reasonable default alignment:  The larger of 32-byte (dword),
1851          * which is a common minimum for HBAs, and the minimum DMA alignment,
1852          * which is set by the platform.
1853          *
1854          * Devices that require a bigger alignment can increase it later.
1855          */
1856         blk_queue_dma_alignment(q, max(4, dma_get_cache_alignment()) - 1);
1857 }
1858 EXPORT_SYMBOL_GPL(__scsi_init_queue);
1859
1860 static const struct blk_mq_ops scsi_mq_ops = {
1861         .get_budget     = scsi_mq_get_budget,
1862         .put_budget     = scsi_mq_put_budget,
1863         .queue_rq       = scsi_queue_rq,
1864         .complete       = scsi_softirq_done,
1865         .timeout        = scsi_timeout,
1866 #ifdef CONFIG_BLK_DEBUG_FS
1867         .show_rq        = scsi_show_rq,
1868 #endif
1869         .init_request   = scsi_mq_init_request,
1870         .exit_request   = scsi_mq_exit_request,
1871         .initialize_rq_fn = scsi_initialize_rq,
1872         .busy           = scsi_mq_lld_busy,
1873         .map_queues     = scsi_map_queues,
1874 };
1875
1876 struct request_queue *scsi_mq_alloc_queue(struct scsi_device *sdev)
1877 {
1878         sdev->request_queue = blk_mq_init_queue(&sdev->host->tag_set);
1879         if (IS_ERR(sdev->request_queue))
1880                 return NULL;
1881
1882         sdev->request_queue->queuedata = sdev;
1883         __scsi_init_queue(sdev->host, sdev->request_queue);
1884         blk_queue_flag_set(QUEUE_FLAG_SCSI_PASSTHROUGH, sdev->request_queue);
1885         return sdev->request_queue;
1886 }
1887
1888 int scsi_mq_setup_tags(struct Scsi_Host *shost)
1889 {
1890         unsigned int cmd_size, sgl_size;
1891
1892         sgl_size = scsi_mq_sgl_size(shost);
1893         cmd_size = sizeof(struct scsi_cmnd) + shost->hostt->cmd_size + sgl_size;
1894         if (scsi_host_get_prot(shost))
1895                 cmd_size += sizeof(struct scsi_data_buffer) + sgl_size;
1896
1897         memset(&shost->tag_set, 0, sizeof(shost->tag_set));
1898         shost->tag_set.ops = &scsi_mq_ops;
1899         shost->tag_set.nr_hw_queues = shost->nr_hw_queues ? : 1;
1900         shost->tag_set.queue_depth = shost->can_queue;
1901         shost->tag_set.cmd_size = cmd_size;
1902         shost->tag_set.numa_node = NUMA_NO_NODE;
1903         shost->tag_set.flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_SG_MERGE;
1904         shost->tag_set.flags |=
1905                 BLK_ALLOC_POLICY_TO_MQ_FLAG(shost->hostt->tag_alloc_policy);
1906         shost->tag_set.driver_data = shost;
1907
1908         return blk_mq_alloc_tag_set(&shost->tag_set);
1909 }
1910
1911 void scsi_mq_destroy_tags(struct Scsi_Host *shost)
1912 {
1913         blk_mq_free_tag_set(&shost->tag_set);
1914 }
1915
1916 /**
1917  * scsi_device_from_queue - return sdev associated with a request_queue
1918  * @q: The request queue to return the sdev from
1919  *
1920  * Return the sdev associated with a request queue or NULL if the
1921  * request_queue does not reference a SCSI device.
1922  */
1923 struct scsi_device *scsi_device_from_queue(struct request_queue *q)
1924 {
1925         struct scsi_device *sdev = NULL;
1926
1927         if (q->mq_ops == &scsi_mq_ops)
1928                 sdev = q->queuedata;
1929         if (!sdev || !get_device(&sdev->sdev_gendev))
1930                 sdev = NULL;
1931
1932         return sdev;
1933 }
1934 EXPORT_SYMBOL_GPL(scsi_device_from_queue);
1935
1936 /*
1937  * Function:    scsi_block_requests()
1938  *
1939  * Purpose:     Utility function used by low-level drivers to prevent further
1940  *              commands from being queued to the device.
1941  *
1942  * Arguments:   shost       - Host in question
1943  *
1944  * Returns:     Nothing
1945  *
1946  * Lock status: No locks are assumed held.
1947  *
1948  * Notes:       There is no timer nor any other means by which the requests
1949  *              get unblocked other than the low-level driver calling
1950  *              scsi_unblock_requests().
1951  */
1952 void scsi_block_requests(struct Scsi_Host *shost)
1953 {
1954         shost->host_self_blocked = 1;
1955 }
1956 EXPORT_SYMBOL(scsi_block_requests);
1957
1958 /*
1959  * Function:    scsi_unblock_requests()
1960  *
1961  * Purpose:     Utility function used by low-level drivers to allow further
1962  *              commands from being queued to the device.
1963  *
1964  * Arguments:   shost       - Host in question
1965  *
1966  * Returns:     Nothing
1967  *
1968  * Lock status: No locks are assumed held.
1969  *
1970  * Notes:       There is no timer nor any other means by which the requests
1971  *              get unblocked other than the low-level driver calling
1972  *              scsi_unblock_requests().
1973  *
1974  *              This is done as an API function so that changes to the
1975  *              internals of the scsi mid-layer won't require wholesale
1976  *              changes to drivers that use this feature.
1977  */
1978 void scsi_unblock_requests(struct Scsi_Host *shost)
1979 {
1980         shost->host_self_blocked = 0;
1981         scsi_run_host_queues(shost);
1982 }
1983 EXPORT_SYMBOL(scsi_unblock_requests);
1984
1985 int __init scsi_init_queue(void)
1986 {
1987         scsi_sdb_cache = kmem_cache_create("scsi_data_buffer",
1988                                            sizeof(struct scsi_data_buffer),
1989                                            0, 0, NULL);
1990         if (!scsi_sdb_cache) {
1991                 printk(KERN_ERR "SCSI: can't init scsi sdb cache\n");
1992                 return -ENOMEM;
1993         }
1994
1995         return 0;
1996 }
1997
1998 void scsi_exit_queue(void)
1999 {
2000         kmem_cache_destroy(scsi_sense_cache);
2001         kmem_cache_destroy(scsi_sense_isadma_cache);
2002         kmem_cache_destroy(scsi_sdb_cache);
2003 }
2004
2005 /**
2006  *      scsi_mode_select - issue a mode select
2007  *      @sdev:  SCSI device to be queried
2008  *      @pf:    Page format bit (1 == standard, 0 == vendor specific)
2009  *      @sp:    Save page bit (0 == don't save, 1 == save)
2010  *      @modepage: mode page being requested
2011  *      @buffer: request buffer (may not be smaller than eight bytes)
2012  *      @len:   length of request buffer.
2013  *      @timeout: command timeout
2014  *      @retries: number of retries before failing
2015  *      @data: returns a structure abstracting the mode header data
2016  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
2017  *              must be SCSI_SENSE_BUFFERSIZE big.
2018  *
2019  *      Returns zero if successful; negative error number or scsi
2020  *      status on error
2021  *
2022  */
2023 int
2024 scsi_mode_select(struct scsi_device *sdev, int pf, int sp, int modepage,
2025                  unsigned char *buffer, int len, int timeout, int retries,
2026                  struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
2027 {
2028         unsigned char cmd[10];
2029         unsigned char *real_buffer;
2030         int ret;
2031
2032         memset(cmd, 0, sizeof(cmd));
2033         cmd[1] = (pf ? 0x10 : 0) | (sp ? 0x01 : 0);
2034
2035         if (sdev->use_10_for_ms) {
2036                 if (len > 65535)
2037                         return -EINVAL;
2038                 real_buffer = kmalloc(8 + len, GFP_KERNEL);
2039                 if (!real_buffer)
2040                         return -ENOMEM;
2041                 memcpy(real_buffer + 8, buffer, len);
2042                 len += 8;
2043                 real_buffer[0] = 0;
2044                 real_buffer[1] = 0;
2045                 real_buffer[2] = data->medium_type;
2046                 real_buffer[3] = data->device_specific;
2047                 real_buffer[4] = data->longlba ? 0x01 : 0;
2048                 real_buffer[5] = 0;
2049                 real_buffer[6] = data->block_descriptor_length >> 8;
2050                 real_buffer[7] = data->block_descriptor_length;
2051
2052                 cmd[0] = MODE_SELECT_10;
2053                 cmd[7] = len >> 8;
2054                 cmd[8] = len;
2055         } else {
2056                 if (len > 255 || data->block_descriptor_length > 255 ||
2057                     data->longlba)
2058                         return -EINVAL;
2059
2060                 real_buffer = kmalloc(4 + len, GFP_KERNEL);
2061                 if (!real_buffer)
2062                         return -ENOMEM;
2063                 memcpy(real_buffer + 4, buffer, len);
2064                 len += 4;
2065                 real_buffer[0] = 0;
2066                 real_buffer[1] = data->medium_type;
2067                 real_buffer[2] = data->device_specific;
2068                 real_buffer[3] = data->block_descriptor_length;
2069                 
2070
2071                 cmd[0] = MODE_SELECT;
2072                 cmd[4] = len;
2073         }
2074
2075         ret = scsi_execute_req(sdev, cmd, DMA_TO_DEVICE, real_buffer, len,
2076                                sshdr, timeout, retries, NULL);
2077         kfree(real_buffer);
2078         return ret;
2079 }
2080 EXPORT_SYMBOL_GPL(scsi_mode_select);
2081
2082 /**
2083  *      scsi_mode_sense - issue a mode sense, falling back from 10 to six bytes if necessary.
2084  *      @sdev:  SCSI device to be queried
2085  *      @dbd:   set if mode sense will allow block descriptors to be returned
2086  *      @modepage: mode page being requested
2087  *      @buffer: request buffer (may not be smaller than eight bytes)
2088  *      @len:   length of request buffer.
2089  *      @timeout: command timeout
2090  *      @retries: number of retries before failing
2091  *      @data: returns a structure abstracting the mode header data
2092  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
2093  *              must be SCSI_SENSE_BUFFERSIZE big.
2094  *
2095  *      Returns zero if unsuccessful, or the header offset (either 4
2096  *      or 8 depending on whether a six or ten byte command was
2097  *      issued) if successful.
2098  */
2099 int
2100 scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
2101                   unsigned char *buffer, int len, int timeout, int retries,
2102                   struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
2103 {
2104         unsigned char cmd[12];
2105         int use_10_for_ms;
2106         int header_length;
2107         int result, retry_count = retries;
2108         struct scsi_sense_hdr my_sshdr;
2109
2110         memset(data, 0, sizeof(*data));
2111         memset(&cmd[0], 0, 12);
2112         cmd[1] = dbd & 0x18;    /* allows DBD and LLBA bits */
2113         cmd[2] = modepage;
2114
2115         /* caller might not be interested in sense, but we need it */
2116         if (!sshdr)
2117                 sshdr = &my_sshdr;
2118
2119  retry:
2120         use_10_for_ms = sdev->use_10_for_ms;
2121
2122         if (use_10_for_ms) {
2123                 if (len < 8)
2124                         len = 8;
2125
2126                 cmd[0] = MODE_SENSE_10;
2127                 cmd[8] = len;
2128                 header_length = 8;
2129         } else {
2130                 if (len < 4)
2131                         len = 4;
2132
2133                 cmd[0] = MODE_SENSE;
2134                 cmd[4] = len;
2135                 header_length = 4;
2136         }
2137
2138         memset(buffer, 0, len);
2139
2140         result = scsi_execute_req(sdev, cmd, DMA_FROM_DEVICE, buffer, len,
2141                                   sshdr, timeout, retries, NULL);
2142
2143         /* This code looks awful: what it's doing is making sure an
2144          * ILLEGAL REQUEST sense return identifies the actual command
2145          * byte as the problem.  MODE_SENSE commands can return
2146          * ILLEGAL REQUEST if the code page isn't supported */
2147
2148         if (use_10_for_ms && !scsi_status_is_good(result) &&
2149             driver_byte(result) == DRIVER_SENSE) {
2150                 if (scsi_sense_valid(sshdr)) {
2151                         if ((sshdr->sense_key == ILLEGAL_REQUEST) &&
2152                             (sshdr->asc == 0x20) && (sshdr->ascq == 0)) {
2153                                 /* 
2154                                  * Invalid command operation code
2155                                  */
2156                                 sdev->use_10_for_ms = 0;
2157                                 goto retry;
2158                         }
2159                 }
2160         }
2161
2162         if(scsi_status_is_good(result)) {
2163                 if (unlikely(buffer[0] == 0x86 && buffer[1] == 0x0b &&
2164                              (modepage == 6 || modepage == 8))) {
2165                         /* Initio breakage? */
2166                         header_length = 0;
2167                         data->length = 13;
2168                         data->medium_type = 0;
2169                         data->device_specific = 0;
2170                         data->longlba = 0;
2171                         data->block_descriptor_length = 0;
2172                 } else if(use_10_for_ms) {
2173                         data->length = buffer[0]*256 + buffer[1] + 2;
2174                         data->medium_type = buffer[2];
2175                         data->device_specific = buffer[3];
2176                         data->longlba = buffer[4] & 0x01;
2177                         data->block_descriptor_length = buffer[6]*256
2178                                 + buffer[7];
2179                 } else {
2180                         data->length = buffer[0] + 1;
2181                         data->medium_type = buffer[1];
2182                         data->device_specific = buffer[2];
2183                         data->block_descriptor_length = buffer[3];
2184                 }
2185                 data->header_length = header_length;
2186         } else if ((status_byte(result) == CHECK_CONDITION) &&
2187                    scsi_sense_valid(sshdr) &&
2188                    sshdr->sense_key == UNIT_ATTENTION && retry_count) {
2189                 retry_count--;
2190                 goto retry;
2191         }
2192
2193         return result;
2194 }
2195 EXPORT_SYMBOL(scsi_mode_sense);
2196
2197 /**
2198  *      scsi_test_unit_ready - test if unit is ready
2199  *      @sdev:  scsi device to change the state of.
2200  *      @timeout: command timeout
2201  *      @retries: number of retries before failing
2202  *      @sshdr: outpout pointer for decoded sense information.
2203  *
2204  *      Returns zero if unsuccessful or an error if TUR failed.  For
2205  *      removable media, UNIT_ATTENTION sets ->changed flag.
2206  **/
2207 int
2208 scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries,
2209                      struct scsi_sense_hdr *sshdr)
2210 {
2211         char cmd[] = {
2212                 TEST_UNIT_READY, 0, 0, 0, 0, 0,
2213         };
2214         int result;
2215
2216         /* try to eat the UNIT_ATTENTION if there are enough retries */
2217         do {
2218                 result = scsi_execute_req(sdev, cmd, DMA_NONE, NULL, 0, sshdr,
2219                                           timeout, 1, NULL);
2220                 if (sdev->removable && scsi_sense_valid(sshdr) &&
2221                     sshdr->sense_key == UNIT_ATTENTION)
2222                         sdev->changed = 1;
2223         } while (scsi_sense_valid(sshdr) &&
2224                  sshdr->sense_key == UNIT_ATTENTION && --retries);
2225
2226         return result;
2227 }
2228 EXPORT_SYMBOL(scsi_test_unit_ready);
2229
2230 /**
2231  *      scsi_device_set_state - Take the given device through the device state model.
2232  *      @sdev:  scsi device to change the state of.
2233  *      @state: state to change to.
2234  *
2235  *      Returns zero if successful or an error if the requested
2236  *      transition is illegal.
2237  */
2238 int
2239 scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
2240 {
2241         enum scsi_device_state oldstate = sdev->sdev_state;
2242
2243         if (state == oldstate)
2244                 return 0;
2245
2246         switch (state) {
2247         case SDEV_CREATED:
2248                 switch (oldstate) {
2249                 case SDEV_CREATED_BLOCK:
2250                         break;
2251                 default:
2252                         goto illegal;
2253                 }
2254                 break;
2255                         
2256         case SDEV_RUNNING:
2257                 switch (oldstate) {
2258                 case SDEV_CREATED:
2259                 case SDEV_OFFLINE:
2260                 case SDEV_TRANSPORT_OFFLINE:
2261                 case SDEV_QUIESCE:
2262                 case SDEV_BLOCK:
2263                         break;
2264                 default:
2265                         goto illegal;
2266                 }
2267                 break;
2268
2269         case SDEV_QUIESCE:
2270                 switch (oldstate) {
2271                 case SDEV_RUNNING:
2272                 case SDEV_OFFLINE:
2273                 case SDEV_TRANSPORT_OFFLINE:
2274                         break;
2275                 default:
2276                         goto illegal;
2277                 }
2278                 break;
2279
2280         case SDEV_OFFLINE:
2281         case SDEV_TRANSPORT_OFFLINE:
2282                 switch (oldstate) {
2283                 case SDEV_CREATED:
2284                 case SDEV_RUNNING:
2285                 case SDEV_QUIESCE:
2286                 case SDEV_BLOCK:
2287                         break;
2288                 default:
2289                         goto illegal;
2290                 }
2291                 break;
2292
2293         case SDEV_BLOCK:
2294                 switch (oldstate) {
2295                 case SDEV_RUNNING:
2296                 case SDEV_CREATED_BLOCK:
2297                 case SDEV_OFFLINE:
2298                         break;
2299                 default:
2300                         goto illegal;
2301                 }
2302                 break;
2303
2304         case SDEV_CREATED_BLOCK:
2305                 switch (oldstate) {
2306                 case SDEV_CREATED:
2307                         break;
2308                 default:
2309                         goto illegal;
2310                 }
2311                 break;
2312
2313         case SDEV_CANCEL:
2314                 switch (oldstate) {
2315                 case SDEV_CREATED:
2316                 case SDEV_RUNNING:
2317                 case SDEV_QUIESCE:
2318                 case SDEV_OFFLINE:
2319                 case SDEV_TRANSPORT_OFFLINE:
2320                         break;
2321                 default:
2322                         goto illegal;
2323                 }
2324                 break;
2325
2326         case SDEV_DEL:
2327                 switch (oldstate) {
2328                 case SDEV_CREATED:
2329                 case SDEV_RUNNING:
2330                 case SDEV_OFFLINE:
2331                 case SDEV_TRANSPORT_OFFLINE:
2332                 case SDEV_CANCEL:
2333                 case SDEV_BLOCK:
2334                 case SDEV_CREATED_BLOCK:
2335                         break;
2336                 default:
2337                         goto illegal;
2338                 }
2339                 break;
2340
2341         }
2342         sdev->sdev_state = state;
2343         return 0;
2344
2345  illegal:
2346         SCSI_LOG_ERROR_RECOVERY(1,
2347                                 sdev_printk(KERN_ERR, sdev,
2348                                             "Illegal state transition %s->%s",
2349                                             scsi_device_state_name(oldstate),
2350                                             scsi_device_state_name(state))
2351                                 );
2352         return -EINVAL;
2353 }
2354 EXPORT_SYMBOL(scsi_device_set_state);
2355
2356 /**
2357  *      sdev_evt_emit - emit a single SCSI device uevent
2358  *      @sdev: associated SCSI device
2359  *      @evt: event to emit
2360  *
2361  *      Send a single uevent (scsi_event) to the associated scsi_device.
2362  */
2363 static void scsi_evt_emit(struct scsi_device *sdev, struct scsi_event *evt)
2364 {
2365         int idx = 0;
2366         char *envp[3];
2367
2368         switch (evt->evt_type) {
2369         case SDEV_EVT_MEDIA_CHANGE:
2370                 envp[idx++] = "SDEV_MEDIA_CHANGE=1";
2371                 break;
2372         case SDEV_EVT_INQUIRY_CHANGE_REPORTED:
2373                 scsi_rescan_device(&sdev->sdev_gendev);
2374                 envp[idx++] = "SDEV_UA=INQUIRY_DATA_HAS_CHANGED";
2375                 break;
2376         case SDEV_EVT_CAPACITY_CHANGE_REPORTED:
2377                 envp[idx++] = "SDEV_UA=CAPACITY_DATA_HAS_CHANGED";
2378                 break;
2379         case SDEV_EVT_SOFT_THRESHOLD_REACHED_REPORTED:
2380                envp[idx++] = "SDEV_UA=THIN_PROVISIONING_SOFT_THRESHOLD_REACHED";
2381                 break;
2382         case SDEV_EVT_MODE_PARAMETER_CHANGE_REPORTED:
2383                 envp[idx++] = "SDEV_UA=MODE_PARAMETERS_CHANGED";
2384                 break;
2385         case SDEV_EVT_LUN_CHANGE_REPORTED:
2386                 envp[idx++] = "SDEV_UA=REPORTED_LUNS_DATA_HAS_CHANGED";
2387                 break;
2388         case SDEV_EVT_ALUA_STATE_CHANGE_REPORTED:
2389                 envp[idx++] = "SDEV_UA=ASYMMETRIC_ACCESS_STATE_CHANGED";
2390                 break;
2391         case SDEV_EVT_POWER_ON_RESET_OCCURRED:
2392                 envp[idx++] = "SDEV_UA=POWER_ON_RESET_OCCURRED";
2393                 break;
2394         default:
2395                 /* do nothing */
2396                 break;
2397         }
2398
2399         envp[idx++] = NULL;
2400
2401         kobject_uevent_env(&sdev->sdev_gendev.kobj, KOBJ_CHANGE, envp);
2402 }
2403
2404 /**
2405  *      sdev_evt_thread - send a uevent for each scsi event
2406  *      @work: work struct for scsi_device
2407  *
2408  *      Dispatch queued events to their associated scsi_device kobjects
2409  *      as uevents.
2410  */
2411 void scsi_evt_thread(struct work_struct *work)
2412 {
2413         struct scsi_device *sdev;
2414         enum scsi_device_event evt_type;
2415         LIST_HEAD(event_list);
2416
2417         sdev = container_of(work, struct scsi_device, event_work);
2418
2419         for (evt_type = SDEV_EVT_FIRST; evt_type <= SDEV_EVT_LAST; evt_type++)
2420                 if (test_and_clear_bit(evt_type, sdev->pending_events))
2421                         sdev_evt_send_simple(sdev, evt_type, GFP_KERNEL);
2422
2423         while (1) {
2424                 struct scsi_event *evt;
2425                 struct list_head *this, *tmp;
2426                 unsigned long flags;
2427
2428                 spin_lock_irqsave(&sdev->list_lock, flags);
2429                 list_splice_init(&sdev->event_list, &event_list);
2430                 spin_unlock_irqrestore(&sdev->list_lock, flags);
2431
2432                 if (list_empty(&event_list))
2433                         break;
2434
2435                 list_for_each_safe(this, tmp, &event_list) {
2436                         evt = list_entry(this, struct scsi_event, node);
2437                         list_del(&evt->node);
2438                         scsi_evt_emit(sdev, evt);
2439                         kfree(evt);
2440                 }
2441         }
2442 }
2443
2444 /**
2445  *      sdev_evt_send - send asserted event to uevent thread
2446  *      @sdev: scsi_device event occurred on
2447  *      @evt: event to send
2448  *
2449  *      Assert scsi device event asynchronously.
2450  */
2451 void sdev_evt_send(struct scsi_device *sdev, struct scsi_event *evt)
2452 {
2453         unsigned long flags;
2454
2455 #if 0
2456         /* FIXME: currently this check eliminates all media change events
2457          * for polled devices.  Need to update to discriminate between AN
2458          * and polled events */
2459         if (!test_bit(evt->evt_type, sdev->supported_events)) {
2460                 kfree(evt);
2461                 return;
2462         }
2463 #endif
2464
2465         spin_lock_irqsave(&sdev->list_lock, flags);
2466         list_add_tail(&evt->node, &sdev->event_list);
2467         schedule_work(&sdev->event_work);
2468         spin_unlock_irqrestore(&sdev->list_lock, flags);
2469 }
2470 EXPORT_SYMBOL_GPL(sdev_evt_send);
2471
2472 /**
2473  *      sdev_evt_alloc - allocate a new scsi event
2474  *      @evt_type: type of event to allocate
2475  *      @gfpflags: GFP flags for allocation
2476  *
2477  *      Allocates and returns a new scsi_event.
2478  */
2479 struct scsi_event *sdev_evt_alloc(enum scsi_device_event evt_type,
2480                                   gfp_t gfpflags)
2481 {
2482         struct scsi_event *evt = kzalloc(sizeof(struct scsi_event), gfpflags);
2483         if (!evt)
2484                 return NULL;
2485
2486         evt->evt_type = evt_type;
2487         INIT_LIST_HEAD(&evt->node);
2488
2489         /* evt_type-specific initialization, if any */
2490         switch (evt_type) {
2491         case SDEV_EVT_MEDIA_CHANGE:
2492         case SDEV_EVT_INQUIRY_CHANGE_REPORTED:
2493         case SDEV_EVT_CAPACITY_CHANGE_REPORTED:
2494         case SDEV_EVT_SOFT_THRESHOLD_REACHED_REPORTED:
2495         case SDEV_EVT_MODE_PARAMETER_CHANGE_REPORTED:
2496         case SDEV_EVT_LUN_CHANGE_REPORTED:
2497         case SDEV_EVT_ALUA_STATE_CHANGE_REPORTED:
2498         case SDEV_EVT_POWER_ON_RESET_OCCURRED:
2499         default:
2500                 /* do nothing */
2501                 break;
2502         }
2503
2504         return evt;
2505 }
2506 EXPORT_SYMBOL_GPL(sdev_evt_alloc);
2507
2508 /**
2509  *      sdev_evt_send_simple - send asserted event to uevent thread
2510  *      @sdev: scsi_device event occurred on
2511  *      @evt_type: type of event to send
2512  *      @gfpflags: GFP flags for allocation
2513  *
2514  *      Assert scsi device event asynchronously, given an event type.
2515  */
2516 void sdev_evt_send_simple(struct scsi_device *sdev,
2517                           enum scsi_device_event evt_type, gfp_t gfpflags)
2518 {
2519         struct scsi_event *evt = sdev_evt_alloc(evt_type, gfpflags);
2520         if (!evt) {
2521                 sdev_printk(KERN_ERR, sdev, "event %d eaten due to OOM\n",
2522                             evt_type);
2523                 return;
2524         }
2525
2526         sdev_evt_send(sdev, evt);
2527 }
2528 EXPORT_SYMBOL_GPL(sdev_evt_send_simple);
2529
2530 /**
2531  *      scsi_device_quiesce - Block user issued commands.
2532  *      @sdev:  scsi device to quiesce.
2533  *
2534  *      This works by trying to transition to the SDEV_QUIESCE state
2535  *      (which must be a legal transition).  When the device is in this
2536  *      state, only special requests will be accepted, all others will
2537  *      be deferred.  Since special requests may also be requeued requests,
2538  *      a successful return doesn't guarantee the device will be 
2539  *      totally quiescent.
2540  *
2541  *      Must be called with user context, may sleep.
2542  *
2543  *      Returns zero if unsuccessful or an error if not.
2544  */
2545 int
2546 scsi_device_quiesce(struct scsi_device *sdev)
2547 {
2548         struct request_queue *q = sdev->request_queue;
2549         int err;
2550
2551         /*
2552          * It is allowed to call scsi_device_quiesce() multiple times from
2553          * the same context but concurrent scsi_device_quiesce() calls are
2554          * not allowed.
2555          */
2556         WARN_ON_ONCE(sdev->quiesced_by && sdev->quiesced_by != current);
2557
2558         if (sdev->quiesced_by == current)
2559                 return 0;
2560
2561         blk_set_pm_only(q);
2562
2563         blk_mq_freeze_queue(q);
2564         /*
2565          * Ensure that the effect of blk_set_pm_only() will be visible
2566          * for percpu_ref_tryget() callers that occur after the queue
2567          * unfreeze even if the queue was already frozen before this function
2568          * was called. See also https://lwn.net/Articles/573497/.
2569          */
2570         synchronize_rcu();
2571         blk_mq_unfreeze_queue(q);
2572
2573         mutex_lock(&sdev->state_mutex);
2574         err = scsi_device_set_state(sdev, SDEV_QUIESCE);
2575         if (err == 0)
2576                 sdev->quiesced_by = current;
2577         else
2578                 blk_clear_pm_only(q);
2579         mutex_unlock(&sdev->state_mutex);
2580
2581         return err;
2582 }
2583 EXPORT_SYMBOL(scsi_device_quiesce);
2584
2585 /**
2586  *      scsi_device_resume - Restart user issued commands to a quiesced device.
2587  *      @sdev:  scsi device to resume.
2588  *
2589  *      Moves the device from quiesced back to running and restarts the
2590  *      queues.
2591  *
2592  *      Must be called with user context, may sleep.
2593  */
2594 void scsi_device_resume(struct scsi_device *sdev)
2595 {
2596         /* check if the device state was mutated prior to resume, and if
2597          * so assume the state is being managed elsewhere (for example
2598          * device deleted during suspend)
2599          */
2600         mutex_lock(&sdev->state_mutex);
2601         WARN_ON_ONCE(!sdev->quiesced_by);
2602         sdev->quiesced_by = NULL;
2603         blk_clear_pm_only(sdev->request_queue);
2604         if (sdev->sdev_state == SDEV_QUIESCE)
2605                 scsi_device_set_state(sdev, SDEV_RUNNING);
2606         mutex_unlock(&sdev->state_mutex);
2607 }
2608 EXPORT_SYMBOL(scsi_device_resume);
2609
2610 static void
2611 device_quiesce_fn(struct scsi_device *sdev, void *data)
2612 {
2613         scsi_device_quiesce(sdev);
2614 }
2615
2616 void
2617 scsi_target_quiesce(struct scsi_target *starget)
2618 {
2619         starget_for_each_device(starget, NULL, device_quiesce_fn);
2620 }
2621 EXPORT_SYMBOL(scsi_target_quiesce);
2622
2623 static void
2624 device_resume_fn(struct scsi_device *sdev, void *data)
2625 {
2626         scsi_device_resume(sdev);
2627 }
2628
2629 void
2630 scsi_target_resume(struct scsi_target *starget)
2631 {
2632         starget_for_each_device(starget, NULL, device_resume_fn);
2633 }
2634 EXPORT_SYMBOL(scsi_target_resume);
2635
2636 /**
2637  * scsi_internal_device_block_nowait - try to transition to the SDEV_BLOCK state
2638  * @sdev: device to block
2639  *
2640  * Pause SCSI command processing on the specified device. Does not sleep.
2641  *
2642  * Returns zero if successful or a negative error code upon failure.
2643  *
2644  * Notes:
2645  * This routine transitions the device to the SDEV_BLOCK state (which must be
2646  * a legal transition). When the device is in this state, command processing
2647  * is paused until the device leaves the SDEV_BLOCK state. See also
2648  * scsi_internal_device_unblock_nowait().
2649  */
2650 int scsi_internal_device_block_nowait(struct scsi_device *sdev)
2651 {
2652         struct request_queue *q = sdev->request_queue;
2653         int err = 0;
2654
2655         err = scsi_device_set_state(sdev, SDEV_BLOCK);
2656         if (err) {
2657                 err = scsi_device_set_state(sdev, SDEV_CREATED_BLOCK);
2658
2659                 if (err)
2660                         return err;
2661         }
2662
2663         /* 
2664          * The device has transitioned to SDEV_BLOCK.  Stop the
2665          * block layer from calling the midlayer with this device's
2666          * request queue. 
2667          */
2668         blk_mq_quiesce_queue_nowait(q);
2669         return 0;
2670 }
2671 EXPORT_SYMBOL_GPL(scsi_internal_device_block_nowait);
2672
2673 /**
2674  * scsi_internal_device_block - try to transition to the SDEV_BLOCK state
2675  * @sdev: device to block
2676  *
2677  * Pause SCSI command processing on the specified device and wait until all
2678  * ongoing scsi_request_fn() / scsi_queue_rq() calls have finished. May sleep.
2679  *
2680  * Returns zero if successful or a negative error code upon failure.
2681  *
2682  * Note:
2683  * This routine transitions the device to the SDEV_BLOCK state (which must be
2684  * a legal transition). When the device is in this state, command processing
2685  * is paused until the device leaves the SDEV_BLOCK state. See also
2686  * scsi_internal_device_unblock().
2687  *
2688  * To do: avoid that scsi_send_eh_cmnd() calls queuecommand() after
2689  * scsi_internal_device_block() has blocked a SCSI device and also
2690  * remove the rport mutex lock and unlock calls from srp_queuecommand().
2691  */
2692 static int scsi_internal_device_block(struct scsi_device *sdev)
2693 {
2694         struct request_queue *q = sdev->request_queue;
2695         int err;
2696
2697         mutex_lock(&sdev->state_mutex);
2698         err = scsi_internal_device_block_nowait(sdev);
2699         if (err == 0)
2700                 blk_mq_quiesce_queue(q);
2701         mutex_unlock(&sdev->state_mutex);
2702
2703         return err;
2704 }
2705  
2706 void scsi_start_queue(struct scsi_device *sdev)
2707 {
2708         struct request_queue *q = sdev->request_queue;
2709
2710         blk_mq_unquiesce_queue(q);
2711 }
2712
2713 /**
2714  * scsi_internal_device_unblock_nowait - resume a device after a block request
2715  * @sdev:       device to resume
2716  * @new_state:  state to set the device to after unblocking
2717  *
2718  * Restart the device queue for a previously suspended SCSI device. Does not
2719  * sleep.
2720  *
2721  * Returns zero if successful or a negative error code upon failure.
2722  *
2723  * Notes:
2724  * This routine transitions the device to the SDEV_RUNNING state or to one of
2725  * the offline states (which must be a legal transition) allowing the midlayer
2726  * to goose the queue for this device.
2727  */
2728 int scsi_internal_device_unblock_nowait(struct scsi_device *sdev,
2729                                         enum scsi_device_state new_state)
2730 {
2731         /*
2732          * Try to transition the scsi device to SDEV_RUNNING or one of the
2733          * offlined states and goose the device queue if successful.
2734          */
2735         switch (sdev->sdev_state) {
2736         case SDEV_BLOCK:
2737         case SDEV_TRANSPORT_OFFLINE:
2738                 sdev->sdev_state = new_state;
2739                 break;
2740         case SDEV_CREATED_BLOCK:
2741                 if (new_state == SDEV_TRANSPORT_OFFLINE ||
2742                     new_state == SDEV_OFFLINE)
2743                         sdev->sdev_state = new_state;
2744                 else
2745                         sdev->sdev_state = SDEV_CREATED;
2746                 break;
2747         case SDEV_CANCEL:
2748         case SDEV_OFFLINE:
2749                 break;
2750         default:
2751                 return -EINVAL;
2752         }
2753         scsi_start_queue(sdev);
2754
2755         return 0;
2756 }
2757 EXPORT_SYMBOL_GPL(scsi_internal_device_unblock_nowait);
2758
2759 /**
2760  * scsi_internal_device_unblock - resume a device after a block request
2761  * @sdev:       device to resume
2762  * @new_state:  state to set the device to after unblocking
2763  *
2764  * Restart the device queue for a previously suspended SCSI device. May sleep.
2765  *
2766  * Returns zero if successful or a negative error code upon failure.
2767  *
2768  * Notes:
2769  * This routine transitions the device to the SDEV_RUNNING state or to one of
2770  * the offline states (which must be a legal transition) allowing the midlayer
2771  * to goose the queue for this device.
2772  */
2773 static int scsi_internal_device_unblock(struct scsi_device *sdev,
2774                                         enum scsi_device_state new_state)
2775 {
2776         int ret;
2777
2778         mutex_lock(&sdev->state_mutex);
2779         ret = scsi_internal_device_unblock_nowait(sdev, new_state);
2780         mutex_unlock(&sdev->state_mutex);
2781
2782         return ret;
2783 }
2784
2785 static void
2786 device_block(struct scsi_device *sdev, void *data)
2787 {
2788         scsi_internal_device_block(sdev);
2789 }
2790
2791 static int
2792 target_block(struct device *dev, void *data)
2793 {
2794         if (scsi_is_target_device(dev))
2795                 starget_for_each_device(to_scsi_target(dev), NULL,
2796                                         device_block);
2797         return 0;
2798 }
2799
2800 void
2801 scsi_target_block(struct device *dev)
2802 {
2803         if (scsi_is_target_device(dev))
2804                 starget_for_each_device(to_scsi_target(dev), NULL,
2805                                         device_block);
2806         else
2807                 device_for_each_child(dev, NULL, target_block);
2808 }
2809 EXPORT_SYMBOL_GPL(scsi_target_block);
2810
2811 static void
2812 device_unblock(struct scsi_device *sdev, void *data)
2813 {
2814         scsi_internal_device_unblock(sdev, *(enum scsi_device_state *)data);
2815 }
2816
2817 static int
2818 target_unblock(struct device *dev, void *data)
2819 {
2820         if (scsi_is_target_device(dev))
2821                 starget_for_each_device(to_scsi_target(dev), data,
2822                                         device_unblock);
2823         return 0;
2824 }
2825
2826 void
2827 scsi_target_unblock(struct device *dev, enum scsi_device_state new_state)
2828 {
2829         if (scsi_is_target_device(dev))
2830                 starget_for_each_device(to_scsi_target(dev), &new_state,
2831                                         device_unblock);
2832         else
2833                 device_for_each_child(dev, &new_state, target_unblock);
2834 }
2835 EXPORT_SYMBOL_GPL(scsi_target_unblock);
2836
2837 /**
2838  * scsi_kmap_atomic_sg - find and atomically map an sg-elemnt
2839  * @sgl:        scatter-gather list
2840  * @sg_count:   number of segments in sg
2841  * @offset:     offset in bytes into sg, on return offset into the mapped area
2842  * @len:        bytes to map, on return number of bytes mapped
2843  *
2844  * Returns virtual address of the start of the mapped page
2845  */
2846 void *scsi_kmap_atomic_sg(struct scatterlist *sgl, int sg_count,
2847                           size_t *offset, size_t *len)
2848 {
2849         int i;
2850         size_t sg_len = 0, len_complete = 0;
2851         struct scatterlist *sg;
2852         struct page *page;
2853
2854         WARN_ON(!irqs_disabled());
2855
2856         for_each_sg(sgl, sg, sg_count, i) {
2857                 len_complete = sg_len; /* Complete sg-entries */
2858                 sg_len += sg->length;
2859                 if (sg_len > *offset)
2860                         break;
2861         }
2862
2863         if (unlikely(i == sg_count)) {
2864                 printk(KERN_ERR "%s: Bytes in sg: %zu, requested offset %zu, "
2865                         "elements %d\n",
2866                        __func__, sg_len, *offset, sg_count);
2867                 WARN_ON(1);
2868                 return NULL;
2869         }
2870
2871         /* Offset starting from the beginning of first page in this sg-entry */
2872         *offset = *offset - len_complete + sg->offset;
2873
2874         /* Assumption: contiguous pages can be accessed as "page + i" */
2875         page = nth_page(sg_page(sg), (*offset >> PAGE_SHIFT));
2876         *offset &= ~PAGE_MASK;
2877
2878         /* Bytes in this sg-entry from *offset to the end of the page */
2879         sg_len = PAGE_SIZE - *offset;
2880         if (*len > sg_len)
2881                 *len = sg_len;
2882
2883         return kmap_atomic(page);
2884 }
2885 EXPORT_SYMBOL(scsi_kmap_atomic_sg);
2886
2887 /**
2888  * scsi_kunmap_atomic_sg - atomically unmap a virtual address, previously mapped with scsi_kmap_atomic_sg
2889  * @virt:       virtual address to be unmapped
2890  */
2891 void scsi_kunmap_atomic_sg(void *virt)
2892 {
2893         kunmap_atomic(virt);
2894 }
2895 EXPORT_SYMBOL(scsi_kunmap_atomic_sg);
2896
2897 void sdev_disable_disk_events(struct scsi_device *sdev)
2898 {
2899         atomic_inc(&sdev->disk_events_disable_depth);
2900 }
2901 EXPORT_SYMBOL(sdev_disable_disk_events);
2902
2903 void sdev_enable_disk_events(struct scsi_device *sdev)
2904 {
2905         if (WARN_ON_ONCE(atomic_read(&sdev->disk_events_disable_depth) <= 0))
2906                 return;
2907         atomic_dec(&sdev->disk_events_disable_depth);
2908 }
2909 EXPORT_SYMBOL(sdev_enable_disk_events);
2910
2911 /**
2912  * scsi_vpd_lun_id - return a unique device identification
2913  * @sdev: SCSI device
2914  * @id:   buffer for the identification
2915  * @id_len:  length of the buffer
2916  *
2917  * Copies a unique device identification into @id based
2918  * on the information in the VPD page 0x83 of the device.
2919  * The string will be formatted as a SCSI name string.
2920  *
2921  * Returns the length of the identification or error on failure.
2922  * If the identifier is longer than the supplied buffer the actual
2923  * identifier length is returned and the buffer is not zero-padded.
2924  */
2925 int scsi_vpd_lun_id(struct scsi_device *sdev, char *id, size_t id_len)
2926 {
2927         u8 cur_id_type = 0xff;
2928         u8 cur_id_size = 0;
2929         const unsigned char *d, *cur_id_str;
2930         const struct scsi_vpd *vpd_pg83;
2931         int id_size = -EINVAL;
2932
2933         rcu_read_lock();
2934         vpd_pg83 = rcu_dereference(sdev->vpd_pg83);
2935         if (!vpd_pg83) {
2936                 rcu_read_unlock();
2937                 return -ENXIO;
2938         }
2939
2940         /*
2941          * Look for the correct descriptor.
2942          * Order of preference for lun descriptor:
2943          * - SCSI name string
2944          * - NAA IEEE Registered Extended
2945          * - EUI-64 based 16-byte
2946          * - EUI-64 based 12-byte
2947          * - NAA IEEE Registered
2948          * - NAA IEEE Extended
2949          * - T10 Vendor ID
2950          * as longer descriptors reduce the likelyhood
2951          * of identification clashes.
2952          */
2953
2954         /* The id string must be at least 20 bytes + terminating NULL byte */
2955         if (id_len < 21) {
2956                 rcu_read_unlock();
2957                 return -EINVAL;
2958         }
2959
2960         memset(id, 0, id_len);
2961         d = vpd_pg83->data + 4;
2962         while (d < vpd_pg83->data + vpd_pg83->len) {
2963                 /* Skip designators not referring to the LUN */
2964                 if ((d[1] & 0x30) != 0x00)
2965                         goto next_desig;
2966
2967                 switch (d[1] & 0xf) {
2968                 case 0x1:
2969                         /* T10 Vendor ID */
2970                         if (cur_id_size > d[3])
2971                                 break;
2972                         /* Prefer anything */
2973                         if (cur_id_type > 0x01 && cur_id_type != 0xff)
2974                                 break;
2975                         cur_id_size = d[3];
2976                         if (cur_id_size + 4 > id_len)
2977                                 cur_id_size = id_len - 4;
2978                         cur_id_str = d + 4;
2979                         cur_id_type = d[1] & 0xf;
2980                         id_size = snprintf(id, id_len, "t10.%*pE",
2981                                            cur_id_size, cur_id_str);
2982                         break;
2983                 case 0x2:
2984                         /* EUI-64 */
2985                         if (cur_id_size > d[3])
2986                                 break;
2987                         /* Prefer NAA IEEE Registered Extended */
2988                         if (cur_id_type == 0x3 &&
2989                             cur_id_size == d[3])
2990                                 break;
2991                         cur_id_size = d[3];
2992                         cur_id_str = d + 4;
2993                         cur_id_type = d[1] & 0xf;
2994                         switch (cur_id_size) {
2995                         case 8:
2996                                 id_size = snprintf(id, id_len,
2997                                                    "eui.%8phN",
2998                                                    cur_id_str);
2999                                 break;
3000                         case 12:
3001                                 id_size = snprintf(id, id_len,
3002                                                    "eui.%12phN",
3003                                                    cur_id_str);
3004                                 break;
3005                         case 16:
3006                                 id_size = snprintf(id, id_len,
3007                                                    "eui.%16phN",
3008                                                    cur_id_str);
3009                                 break;
3010                         default:
3011                                 cur_id_size = 0;
3012                                 break;
3013                         }
3014                         break;
3015                 case 0x3:
3016                         /* NAA */
3017                         if (cur_id_size > d[3])
3018                                 break;
3019                         cur_id_size = d[3];
3020                         cur_id_str = d + 4;
3021                         cur_id_type = d[1] & 0xf;
3022                         switch (cur_id_size) {
3023                         case 8:
3024                                 id_size = snprintf(id, id_len,
3025                                                    "naa.%8phN",
3026                                                    cur_id_str);
3027                                 break;
3028                         case 16:
3029                                 id_size = snprintf(id, id_len,
3030                                                    "naa.%16phN",
3031                                                    cur_id_str);
3032                                 break;
3033                         default:
3034                                 cur_id_size = 0;
3035                                 break;
3036                         }
3037                         break;
3038                 case 0x8:
3039                         /* SCSI name string */
3040                         if (cur_id_size + 4 > d[3])
3041                                 break;
3042                         /* Prefer others for truncated descriptor */
3043                         if (cur_id_size && d[3] > id_len)
3044                                 break;
3045                         cur_id_size = id_size = d[3];
3046                         cur_id_str = d + 4;
3047                         cur_id_type = d[1] & 0xf;
3048                         if (cur_id_size >= id_len)
3049                                 cur_id_size = id_len - 1;
3050                         memcpy(id, cur_id_str, cur_id_size);
3051                         /* Decrease priority for truncated descriptor */
3052                         if (cur_id_size != id_size)
3053                                 cur_id_size = 6;
3054                         break;
3055                 default:
3056                         break;
3057                 }
3058 next_desig:
3059                 d += d[3] + 4;
3060         }
3061         rcu_read_unlock();
3062
3063         return id_size;
3064 }
3065 EXPORT_SYMBOL(scsi_vpd_lun_id);
3066
3067 /*
3068  * scsi_vpd_tpg_id - return a target port group identifier
3069  * @sdev: SCSI device
3070  *
3071  * Returns the Target Port Group identifier from the information
3072  * froom VPD page 0x83 of the device.
3073  *
3074  * Returns the identifier or error on failure.
3075  */
3076 int scsi_vpd_tpg_id(struct scsi_device *sdev, int *rel_id)
3077 {
3078         const unsigned char *d;
3079         const struct scsi_vpd *vpd_pg83;
3080         int group_id = -EAGAIN, rel_port = -1;
3081
3082         rcu_read_lock();
3083         vpd_pg83 = rcu_dereference(sdev->vpd_pg83);
3084         if (!vpd_pg83) {
3085                 rcu_read_unlock();
3086                 return -ENXIO;
3087         }
3088
3089         d = vpd_pg83->data + 4;
3090         while (d < vpd_pg83->data + vpd_pg83->len) {
3091                 switch (d[1] & 0xf) {
3092                 case 0x4:
3093                         /* Relative target port */
3094                         rel_port = get_unaligned_be16(&d[6]);
3095                         break;
3096                 case 0x5:
3097                         /* Target port group */
3098                         group_id = get_unaligned_be16(&d[6]);
3099                         break;
3100                 default:
3101                         break;
3102                 }
3103                 d += d[3] + 4;
3104         }
3105         rcu_read_unlock();
3106
3107         if (group_id >= 0 && rel_id && rel_port != -1)
3108                 *rel_id = rel_port;
3109
3110         return group_id;
3111 }
3112 EXPORT_SYMBOL(scsi_vpd_tpg_id);