Merge tag 'iommu-fixes-v5.0-rc4' of git://git.kernel.org/pub/scm/linux/kernel/git...
[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                 return BLK_STS_NEXUS;
659         case DID_ALLOC_FAILURE:
660                 set_host_byte(cmd, DID_OK);
661                 return BLK_STS_NOSPC;
662         case DID_MEDIUM_ERROR:
663                 set_host_byte(cmd, DID_OK);
664                 return BLK_STS_MEDIUM;
665         default:
666                 return BLK_STS_IOERR;
667         }
668 }
669
670 /* Helper for scsi_io_completion() when "reprep" action required. */
671 static void scsi_io_completion_reprep(struct scsi_cmnd *cmd,
672                                       struct request_queue *q)
673 {
674         /* A new command will be prepared and issued. */
675         scsi_mq_requeue_cmd(cmd);
676 }
677
678 /* Helper for scsi_io_completion() when special action required. */
679 static void scsi_io_completion_action(struct scsi_cmnd *cmd, int result)
680 {
681         struct request_queue *q = cmd->device->request_queue;
682         struct request *req = cmd->request;
683         int level = 0;
684         enum {ACTION_FAIL, ACTION_REPREP, ACTION_RETRY,
685               ACTION_DELAYED_RETRY} action;
686         unsigned long wait_for = (cmd->allowed + 1) * req->timeout;
687         struct scsi_sense_hdr sshdr;
688         bool sense_valid;
689         bool sense_current = true;      /* false implies "deferred sense" */
690         blk_status_t blk_stat;
691
692         sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
693         if (sense_valid)
694                 sense_current = !scsi_sense_is_deferred(&sshdr);
695
696         blk_stat = scsi_result_to_blk_status(cmd, result);
697
698         if (host_byte(result) == DID_RESET) {
699                 /* Third party bus reset or reset for error recovery
700                  * reasons.  Just retry the command and see what
701                  * happens.
702                  */
703                 action = ACTION_RETRY;
704         } else if (sense_valid && sense_current) {
705                 switch (sshdr.sense_key) {
706                 case UNIT_ATTENTION:
707                         if (cmd->device->removable) {
708                                 /* Detected disc change.  Set a bit
709                                  * and quietly refuse further access.
710                                  */
711                                 cmd->device->changed = 1;
712                                 action = ACTION_FAIL;
713                         } else {
714                                 /* Must have been a power glitch, or a
715                                  * bus reset.  Could not have been a
716                                  * media change, so we just retry the
717                                  * command and see what happens.
718                                  */
719                                 action = ACTION_RETRY;
720                         }
721                         break;
722                 case ILLEGAL_REQUEST:
723                         /* If we had an ILLEGAL REQUEST returned, then
724                          * we may have performed an unsupported
725                          * command.  The only thing this should be
726                          * would be a ten byte read where only a six
727                          * byte read was supported.  Also, on a system
728                          * where READ CAPACITY failed, we may have
729                          * read past the end of the disk.
730                          */
731                         if ((cmd->device->use_10_for_rw &&
732                             sshdr.asc == 0x20 && sshdr.ascq == 0x00) &&
733                             (cmd->cmnd[0] == READ_10 ||
734                              cmd->cmnd[0] == WRITE_10)) {
735                                 /* This will issue a new 6-byte command. */
736                                 cmd->device->use_10_for_rw = 0;
737                                 action = ACTION_REPREP;
738                         } else if (sshdr.asc == 0x10) /* DIX */ {
739                                 action = ACTION_FAIL;
740                                 blk_stat = BLK_STS_PROTECTION;
741                         /* INVALID COMMAND OPCODE or INVALID FIELD IN CDB */
742                         } else if (sshdr.asc == 0x20 || sshdr.asc == 0x24) {
743                                 action = ACTION_FAIL;
744                                 blk_stat = BLK_STS_TARGET;
745                         } else
746                                 action = ACTION_FAIL;
747                         break;
748                 case ABORTED_COMMAND:
749                         action = ACTION_FAIL;
750                         if (sshdr.asc == 0x10) /* DIF */
751                                 blk_stat = BLK_STS_PROTECTION;
752                         break;
753                 case NOT_READY:
754                         /* If the device is in the process of becoming
755                          * ready, or has a temporary blockage, retry.
756                          */
757                         if (sshdr.asc == 0x04) {
758                                 switch (sshdr.ascq) {
759                                 case 0x01: /* becoming ready */
760                                 case 0x04: /* format in progress */
761                                 case 0x05: /* rebuild in progress */
762                                 case 0x06: /* recalculation in progress */
763                                 case 0x07: /* operation in progress */
764                                 case 0x08: /* Long write in progress */
765                                 case 0x09: /* self test in progress */
766                                 case 0x14: /* space allocation in progress */
767                                 case 0x1a: /* start stop unit in progress */
768                                 case 0x1b: /* sanitize in progress */
769                                 case 0x1d: /* configuration in progress */
770                                 case 0x24: /* depopulation in progress */
771                                         action = ACTION_DELAYED_RETRY;
772                                         break;
773                                 default:
774                                         action = ACTION_FAIL;
775                                         break;
776                                 }
777                         } else
778                                 action = ACTION_FAIL;
779                         break;
780                 case VOLUME_OVERFLOW:
781                         /* See SSC3rXX or current. */
782                         action = ACTION_FAIL;
783                         break;
784                 default:
785                         action = ACTION_FAIL;
786                         break;
787                 }
788         } else
789                 action = ACTION_FAIL;
790
791         if (action != ACTION_FAIL &&
792             time_before(cmd->jiffies_at_alloc + wait_for, jiffies))
793                 action = ACTION_FAIL;
794
795         switch (action) {
796         case ACTION_FAIL:
797                 /* Give up and fail the remainder of the request */
798                 if (!(req->rq_flags & RQF_QUIET)) {
799                         static DEFINE_RATELIMIT_STATE(_rs,
800                                         DEFAULT_RATELIMIT_INTERVAL,
801                                         DEFAULT_RATELIMIT_BURST);
802
803                         if (unlikely(scsi_logging_level))
804                                 level =
805                                      SCSI_LOG_LEVEL(SCSI_LOG_MLCOMPLETE_SHIFT,
806                                                     SCSI_LOG_MLCOMPLETE_BITS);
807
808                         /*
809                          * if logging is enabled the failure will be printed
810                          * in scsi_log_completion(), so avoid duplicate messages
811                          */
812                         if (!level && __ratelimit(&_rs)) {
813                                 scsi_print_result(cmd, NULL, FAILED);
814                                 if (driver_byte(result) == DRIVER_SENSE)
815                                         scsi_print_sense(cmd);
816                                 scsi_print_command(cmd);
817                         }
818                 }
819                 if (!scsi_end_request(req, blk_stat, blk_rq_err_bytes(req), 0))
820                         return;
821                 /*FALLTHRU*/
822         case ACTION_REPREP:
823                 scsi_io_completion_reprep(cmd, q);
824                 break;
825         case ACTION_RETRY:
826                 /* Retry the same command immediately */
827                 __scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY, false);
828                 break;
829         case ACTION_DELAYED_RETRY:
830                 /* Retry the same command after a delay */
831                 __scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY, false);
832                 break;
833         }
834 }
835
836 /*
837  * Helper for scsi_io_completion() when cmd->result is non-zero. Returns a
838  * new result that may suppress further error checking. Also modifies
839  * *blk_statp in some cases.
840  */
841 static int scsi_io_completion_nz_result(struct scsi_cmnd *cmd, int result,
842                                         blk_status_t *blk_statp)
843 {
844         bool sense_valid;
845         bool sense_current = true;      /* false implies "deferred sense" */
846         struct request *req = cmd->request;
847         struct scsi_sense_hdr sshdr;
848
849         sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
850         if (sense_valid)
851                 sense_current = !scsi_sense_is_deferred(&sshdr);
852
853         if (blk_rq_is_passthrough(req)) {
854                 if (sense_valid) {
855                         /*
856                          * SG_IO wants current and deferred errors
857                          */
858                         scsi_req(req)->sense_len =
859                                 min(8 + cmd->sense_buffer[7],
860                                     SCSI_SENSE_BUFFERSIZE);
861                 }
862                 if (sense_current)
863                         *blk_statp = scsi_result_to_blk_status(cmd, result);
864         } else if (blk_rq_bytes(req) == 0 && sense_current) {
865                 /*
866                  * Flush commands do not transfers any data, and thus cannot use
867                  * good_bytes != blk_rq_bytes(req) as the signal for an error.
868                  * This sets *blk_statp explicitly for the problem case.
869                  */
870                 *blk_statp = scsi_result_to_blk_status(cmd, result);
871         }
872         /*
873          * Recovered errors need reporting, but they're always treated as
874          * success, so fiddle the result code here.  For passthrough requests
875          * we already took a copy of the original into sreq->result which
876          * is what gets returned to the user
877          */
878         if (sense_valid && (sshdr.sense_key == RECOVERED_ERROR)) {
879                 bool do_print = true;
880                 /*
881                  * if ATA PASS-THROUGH INFORMATION AVAILABLE [0x0, 0x1d]
882                  * skip print since caller wants ATA registers. Only occurs
883                  * on SCSI ATA PASS_THROUGH commands when CK_COND=1
884                  */
885                 if ((sshdr.asc == 0x0) && (sshdr.ascq == 0x1d))
886                         do_print = false;
887                 else if (req->rq_flags & RQF_QUIET)
888                         do_print = false;
889                 if (do_print)
890                         scsi_print_sense(cmd);
891                 result = 0;
892                 /* for passthrough, *blk_statp may be set */
893                 *blk_statp = BLK_STS_OK;
894         }
895         /*
896          * Another corner case: the SCSI status byte is non-zero but 'good'.
897          * Example: PRE-FETCH command returns SAM_STAT_CONDITION_MET when
898          * it is able to fit nominated LBs in its cache (and SAM_STAT_GOOD
899          * if it can't fit). Treat SAM_STAT_CONDITION_MET and the related
900          * intermediate statuses (both obsolete in SAM-4) as good.
901          */
902         if (status_byte(result) && scsi_status_is_good(result)) {
903                 result = 0;
904                 *blk_statp = BLK_STS_OK;
905         }
906         return result;
907 }
908
909 /*
910  * Function:    scsi_io_completion()
911  *
912  * Purpose:     Completion processing for block device I/O requests.
913  *
914  * Arguments:   cmd   - command that is finished.
915  *
916  * Lock status: Assumed that no lock is held upon entry.
917  *
918  * Returns:     Nothing
919  *
920  * Notes:       We will finish off the specified number of sectors.  If we
921  *              are done, the command block will be released and the queue
922  *              function will be goosed.  If we are not done then we have to
923  *              figure out what to do next:
924  *
925  *              a) We can call scsi_requeue_command().  The request
926  *                 will be unprepared and put back on the queue.  Then
927  *                 a new command will be created for it.  This should
928  *                 be used if we made forward progress, or if we want
929  *                 to switch from READ(10) to READ(6) for example.
930  *
931  *              b) We can call __scsi_queue_insert().  The request will
932  *                 be put back on the queue and retried using the same
933  *                 command as before, possibly after a delay.
934  *
935  *              c) We can call scsi_end_request() with blk_stat other than
936  *                 BLK_STS_OK, to fail the remainder of the request.
937  */
938 void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes)
939 {
940         int result = cmd->result;
941         struct request_queue *q = cmd->device->request_queue;
942         struct request *req = cmd->request;
943         blk_status_t blk_stat = BLK_STS_OK;
944
945         if (unlikely(result))   /* a nz result may or may not be an error */
946                 result = scsi_io_completion_nz_result(cmd, result, &blk_stat);
947
948         if (unlikely(blk_rq_is_passthrough(req))) {
949                 /*
950                  * scsi_result_to_blk_status may have reset the host_byte
951                  */
952                 scsi_req(req)->result = cmd->result;
953                 scsi_req(req)->resid_len = scsi_get_resid(cmd);
954
955                 if (unlikely(scsi_bidi_cmnd(cmd))) {
956                         /*
957                          * Bidi commands Must be complete as a whole,
958                          * both sides at once.
959                          */
960                         scsi_req(req->next_rq)->resid_len = scsi_in(cmd)->resid;
961                         if (scsi_end_request(req, BLK_STS_OK, blk_rq_bytes(req),
962                                         blk_rq_bytes(req->next_rq)))
963                                 WARN_ONCE(true,
964                                           "Bidi command with remaining bytes");
965                         return;
966                 }
967         }
968
969         /* no bidi support yet, other than in pass-through */
970         if (unlikely(blk_bidi_rq(req))) {
971                 WARN_ONCE(true, "Only support bidi command in passthrough");
972                 scmd_printk(KERN_ERR, cmd, "Killing bidi command\n");
973                 if (scsi_end_request(req, BLK_STS_IOERR, blk_rq_bytes(req),
974                                      blk_rq_bytes(req->next_rq)))
975                         WARN_ONCE(true, "Bidi command with remaining bytes");
976                 return;
977         }
978
979         /*
980          * Next deal with any sectors which we were able to correctly
981          * handle.
982          */
983         SCSI_LOG_HLCOMPLETE(1, scmd_printk(KERN_INFO, cmd,
984                 "%u sectors total, %d bytes done.\n",
985                 blk_rq_sectors(req), good_bytes));
986
987         /*
988          * Next deal with any sectors which we were able to correctly
989          * handle. Failed, zero length commands always need to drop down
990          * to retry code. Fast path should return in this block.
991          */
992         if (likely(blk_rq_bytes(req) > 0 || blk_stat == BLK_STS_OK)) {
993                 if (likely(!scsi_end_request(req, blk_stat, good_bytes, 0)))
994                         return; /* no bytes remaining */
995         }
996
997         /* Kill remainder if no retries. */
998         if (unlikely(blk_stat && scsi_noretry_cmd(cmd))) {
999                 if (scsi_end_request(req, blk_stat, blk_rq_bytes(req), 0))
1000                         WARN_ONCE(true,
1001                             "Bytes remaining after failed, no-retry command");
1002                 return;
1003         }
1004
1005         /*
1006          * If there had been no error, but we have leftover bytes in the
1007          * requeues just queue the command up again.
1008          */
1009         if (likely(result == 0))
1010                 scsi_io_completion_reprep(cmd, q);
1011         else
1012                 scsi_io_completion_action(cmd, result);
1013 }
1014
1015 static blk_status_t scsi_init_sgtable(struct request *req,
1016                 struct scsi_data_buffer *sdb)
1017 {
1018         int count;
1019
1020         /*
1021          * If sg table allocation fails, requeue request later.
1022          */
1023         if (unlikely(sg_alloc_table_chained(&sdb->table,
1024                         blk_rq_nr_phys_segments(req), sdb->table.sgl)))
1025                 return BLK_STS_RESOURCE;
1026
1027         /* 
1028          * Next, walk the list, and fill in the addresses and sizes of
1029          * each segment.
1030          */
1031         count = blk_rq_map_sg(req->q, req, sdb->table.sgl);
1032         BUG_ON(count > sdb->table.nents);
1033         sdb->table.nents = count;
1034         sdb->length = blk_rq_payload_bytes(req);
1035         return BLK_STS_OK;
1036 }
1037
1038 /*
1039  * Function:    scsi_init_io()
1040  *
1041  * Purpose:     SCSI I/O initialize function.
1042  *
1043  * Arguments:   cmd   - Command descriptor we wish to initialize
1044  *
1045  * Returns:     BLK_STS_OK on success
1046  *              BLK_STS_RESOURCE if the failure is retryable
1047  *              BLK_STS_IOERR if the failure is fatal
1048  */
1049 blk_status_t scsi_init_io(struct scsi_cmnd *cmd)
1050 {
1051         struct request *rq = cmd->request;
1052         blk_status_t ret;
1053
1054         if (WARN_ON_ONCE(!blk_rq_nr_phys_segments(rq)))
1055                 return BLK_STS_IOERR;
1056
1057         ret = scsi_init_sgtable(rq, &cmd->sdb);
1058         if (ret)
1059                 return ret;
1060
1061         if (blk_bidi_rq(rq)) {
1062                 ret = scsi_init_sgtable(rq->next_rq, rq->next_rq->special);
1063                 if (ret)
1064                         goto out_free_sgtables;
1065         }
1066
1067         if (blk_integrity_rq(rq)) {
1068                 struct scsi_data_buffer *prot_sdb = cmd->prot_sdb;
1069                 int ivecs, count;
1070
1071                 if (WARN_ON_ONCE(!prot_sdb)) {
1072                         /*
1073                          * This can happen if someone (e.g. multipath)
1074                          * queues a command to a device on an adapter
1075                          * that does not support DIX.
1076                          */
1077                         ret = BLK_STS_IOERR;
1078                         goto out_free_sgtables;
1079                 }
1080
1081                 ivecs = blk_rq_count_integrity_sg(rq->q, rq->bio);
1082
1083                 if (sg_alloc_table_chained(&prot_sdb->table, ivecs,
1084                                 prot_sdb->table.sgl)) {
1085                         ret = BLK_STS_RESOURCE;
1086                         goto out_free_sgtables;
1087                 }
1088
1089                 count = blk_rq_map_integrity_sg(rq->q, rq->bio,
1090                                                 prot_sdb->table.sgl);
1091                 BUG_ON(count > ivecs);
1092                 BUG_ON(count > queue_max_integrity_segments(rq->q));
1093
1094                 cmd->prot_sdb = prot_sdb;
1095                 cmd->prot_sdb->table.nents = count;
1096         }
1097
1098         return BLK_STS_OK;
1099 out_free_sgtables:
1100         scsi_mq_free_sgtables(cmd);
1101         return ret;
1102 }
1103 EXPORT_SYMBOL(scsi_init_io);
1104
1105 /**
1106  * scsi_initialize_rq - initialize struct scsi_cmnd partially
1107  * @rq: Request associated with the SCSI command to be initialized.
1108  *
1109  * This function initializes the members of struct scsi_cmnd that must be
1110  * initialized before request processing starts and that won't be
1111  * reinitialized if a SCSI command is requeued.
1112  *
1113  * Called from inside blk_get_request() for pass-through requests and from
1114  * inside scsi_init_command() for filesystem requests.
1115  */
1116 static void scsi_initialize_rq(struct request *rq)
1117 {
1118         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1119
1120         scsi_req_init(&cmd->req);
1121         init_rcu_head(&cmd->rcu);
1122         cmd->jiffies_at_alloc = jiffies;
1123         cmd->retries = 0;
1124 }
1125
1126 /* Add a command to the list used by the aacraid and dpt_i2o drivers */
1127 void scsi_add_cmd_to_list(struct scsi_cmnd *cmd)
1128 {
1129         struct scsi_device *sdev = cmd->device;
1130         struct Scsi_Host *shost = sdev->host;
1131         unsigned long flags;
1132
1133         if (shost->use_cmd_list) {
1134                 spin_lock_irqsave(&sdev->list_lock, flags);
1135                 list_add_tail(&cmd->list, &sdev->cmd_list);
1136                 spin_unlock_irqrestore(&sdev->list_lock, flags);
1137         }
1138 }
1139
1140 /* Remove a command from the list used by the aacraid and dpt_i2o drivers */
1141 void scsi_del_cmd_from_list(struct scsi_cmnd *cmd)
1142 {
1143         struct scsi_device *sdev = cmd->device;
1144         struct Scsi_Host *shost = sdev->host;
1145         unsigned long flags;
1146
1147         if (shost->use_cmd_list) {
1148                 spin_lock_irqsave(&sdev->list_lock, flags);
1149                 BUG_ON(list_empty(&cmd->list));
1150                 list_del_init(&cmd->list);
1151                 spin_unlock_irqrestore(&sdev->list_lock, flags);
1152         }
1153 }
1154
1155 /* Called after a request has been started. */
1156 void scsi_init_command(struct scsi_device *dev, struct scsi_cmnd *cmd)
1157 {
1158         void *buf = cmd->sense_buffer;
1159         void *prot = cmd->prot_sdb;
1160         struct request *rq = blk_mq_rq_from_pdu(cmd);
1161         unsigned int flags = cmd->flags & SCMD_PRESERVED_FLAGS;
1162         unsigned long jiffies_at_alloc;
1163         int retries;
1164
1165         if (!blk_rq_is_scsi(rq) && !(flags & SCMD_INITIALIZED)) {
1166                 flags |= SCMD_INITIALIZED;
1167                 scsi_initialize_rq(rq);
1168         }
1169
1170         jiffies_at_alloc = cmd->jiffies_at_alloc;
1171         retries = cmd->retries;
1172         /* zero out the cmd, except for the embedded scsi_request */
1173         memset((char *)cmd + sizeof(cmd->req), 0,
1174                 sizeof(*cmd) - sizeof(cmd->req) + dev->host->hostt->cmd_size);
1175
1176         cmd->device = dev;
1177         cmd->sense_buffer = buf;
1178         cmd->prot_sdb = prot;
1179         cmd->flags = flags;
1180         INIT_DELAYED_WORK(&cmd->abort_work, scmd_eh_abort_handler);
1181         cmd->jiffies_at_alloc = jiffies_at_alloc;
1182         cmd->retries = retries;
1183
1184         scsi_add_cmd_to_list(cmd);
1185 }
1186
1187 static blk_status_t scsi_setup_scsi_cmnd(struct scsi_device *sdev,
1188                 struct request *req)
1189 {
1190         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1191
1192         /*
1193          * Passthrough requests may transfer data, in which case they must
1194          * a bio attached to them.  Or they might contain a SCSI command
1195          * that does not transfer data, in which case they may optionally
1196          * submit a request without an attached bio.
1197          */
1198         if (req->bio) {
1199                 blk_status_t ret = scsi_init_io(cmd);
1200                 if (unlikely(ret != BLK_STS_OK))
1201                         return ret;
1202         } else {
1203                 BUG_ON(blk_rq_bytes(req));
1204
1205                 memset(&cmd->sdb, 0, sizeof(cmd->sdb));
1206         }
1207
1208         cmd->cmd_len = scsi_req(req)->cmd_len;
1209         cmd->cmnd = scsi_req(req)->cmd;
1210         cmd->transfersize = blk_rq_bytes(req);
1211         cmd->allowed = scsi_req(req)->retries;
1212         return BLK_STS_OK;
1213 }
1214
1215 /*
1216  * Setup a normal block command.  These are simple request from filesystems
1217  * that still need to be translated to SCSI CDBs from the ULD.
1218  */
1219 static blk_status_t scsi_setup_fs_cmnd(struct scsi_device *sdev,
1220                 struct request *req)
1221 {
1222         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1223
1224         if (unlikely(sdev->handler && sdev->handler->prep_fn)) {
1225                 blk_status_t ret = sdev->handler->prep_fn(sdev, req);
1226                 if (ret != BLK_STS_OK)
1227                         return ret;
1228         }
1229
1230         cmd->cmnd = scsi_req(req)->cmd = scsi_req(req)->__cmd;
1231         memset(cmd->cmnd, 0, BLK_MAX_CDB);
1232         return scsi_cmd_to_driver(cmd)->init_command(cmd);
1233 }
1234
1235 static blk_status_t scsi_setup_cmnd(struct scsi_device *sdev,
1236                 struct request *req)
1237 {
1238         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1239
1240         if (!blk_rq_bytes(req))
1241                 cmd->sc_data_direction = DMA_NONE;
1242         else if (rq_data_dir(req) == WRITE)
1243                 cmd->sc_data_direction = DMA_TO_DEVICE;
1244         else
1245                 cmd->sc_data_direction = DMA_FROM_DEVICE;
1246
1247         if (blk_rq_is_scsi(req))
1248                 return scsi_setup_scsi_cmnd(sdev, req);
1249         else
1250                 return scsi_setup_fs_cmnd(sdev, req);
1251 }
1252
1253 static blk_status_t
1254 scsi_prep_state_check(struct scsi_device *sdev, struct request *req)
1255 {
1256         switch (sdev->sdev_state) {
1257         case SDEV_OFFLINE:
1258         case SDEV_TRANSPORT_OFFLINE:
1259                 /*
1260                  * If the device is offline we refuse to process any
1261                  * commands.  The device must be brought online
1262                  * before trying any recovery commands.
1263                  */
1264                 sdev_printk(KERN_ERR, sdev,
1265                             "rejecting I/O to offline device\n");
1266                 return BLK_STS_IOERR;
1267         case SDEV_DEL:
1268                 /*
1269                  * If the device is fully deleted, we refuse to
1270                  * process any commands as well.
1271                  */
1272                 sdev_printk(KERN_ERR, sdev,
1273                             "rejecting I/O to dead device\n");
1274                 return BLK_STS_IOERR;
1275         case SDEV_BLOCK:
1276         case SDEV_CREATED_BLOCK:
1277                 return BLK_STS_RESOURCE;
1278         case SDEV_QUIESCE:
1279                 /*
1280                  * If the devices is blocked we defer normal commands.
1281                  */
1282                 if (req && !(req->rq_flags & RQF_PREEMPT))
1283                         return BLK_STS_RESOURCE;
1284                 return BLK_STS_OK;
1285         default:
1286                 /*
1287                  * For any other not fully online state we only allow
1288                  * special commands.  In particular any user initiated
1289                  * command is not allowed.
1290                  */
1291                 if (req && !(req->rq_flags & RQF_PREEMPT))
1292                         return BLK_STS_IOERR;
1293                 return BLK_STS_OK;
1294         }
1295 }
1296
1297 /*
1298  * scsi_dev_queue_ready: if we can send requests to sdev, return 1 else
1299  * return 0.
1300  *
1301  * Called with the queue_lock held.
1302  */
1303 static inline int scsi_dev_queue_ready(struct request_queue *q,
1304                                   struct scsi_device *sdev)
1305 {
1306         unsigned int busy;
1307
1308         busy = atomic_inc_return(&sdev->device_busy) - 1;
1309         if (atomic_read(&sdev->device_blocked)) {
1310                 if (busy)
1311                         goto out_dec;
1312
1313                 /*
1314                  * unblock after device_blocked iterates to zero
1315                  */
1316                 if (atomic_dec_return(&sdev->device_blocked) > 0)
1317                         goto out_dec;
1318                 SCSI_LOG_MLQUEUE(3, sdev_printk(KERN_INFO, sdev,
1319                                    "unblocking device at zero depth\n"));
1320         }
1321
1322         if (busy >= sdev->queue_depth)
1323                 goto out_dec;
1324
1325         return 1;
1326 out_dec:
1327         atomic_dec(&sdev->device_busy);
1328         return 0;
1329 }
1330
1331 /*
1332  * scsi_target_queue_ready: checks if there we can send commands to target
1333  * @sdev: scsi device on starget to check.
1334  */
1335 static inline int scsi_target_queue_ready(struct Scsi_Host *shost,
1336                                            struct scsi_device *sdev)
1337 {
1338         struct scsi_target *starget = scsi_target(sdev);
1339         unsigned int busy;
1340
1341         if (starget->single_lun) {
1342                 spin_lock_irq(shost->host_lock);
1343                 if (starget->starget_sdev_user &&
1344                     starget->starget_sdev_user != sdev) {
1345                         spin_unlock_irq(shost->host_lock);
1346                         return 0;
1347                 }
1348                 starget->starget_sdev_user = sdev;
1349                 spin_unlock_irq(shost->host_lock);
1350         }
1351
1352         if (starget->can_queue <= 0)
1353                 return 1;
1354
1355         busy = atomic_inc_return(&starget->target_busy) - 1;
1356         if (atomic_read(&starget->target_blocked) > 0) {
1357                 if (busy)
1358                         goto starved;
1359
1360                 /*
1361                  * unblock after target_blocked iterates to zero
1362                  */
1363                 if (atomic_dec_return(&starget->target_blocked) > 0)
1364                         goto out_dec;
1365
1366                 SCSI_LOG_MLQUEUE(3, starget_printk(KERN_INFO, starget,
1367                                  "unblocking target at zero depth\n"));
1368         }
1369
1370         if (busy >= starget->can_queue)
1371                 goto starved;
1372
1373         return 1;
1374
1375 starved:
1376         spin_lock_irq(shost->host_lock);
1377         list_move_tail(&sdev->starved_entry, &shost->starved_list);
1378         spin_unlock_irq(shost->host_lock);
1379 out_dec:
1380         if (starget->can_queue > 0)
1381                 atomic_dec(&starget->target_busy);
1382         return 0;
1383 }
1384
1385 /*
1386  * scsi_host_queue_ready: if we can send requests to shost, return 1 else
1387  * return 0. We must end up running the queue again whenever 0 is
1388  * returned, else IO can hang.
1389  */
1390 static inline int scsi_host_queue_ready(struct request_queue *q,
1391                                    struct Scsi_Host *shost,
1392                                    struct scsi_device *sdev)
1393 {
1394         unsigned int busy;
1395
1396         if (scsi_host_in_recovery(shost))
1397                 return 0;
1398
1399         busy = atomic_inc_return(&shost->host_busy) - 1;
1400         if (atomic_read(&shost->host_blocked) > 0) {
1401                 if (busy)
1402                         goto starved;
1403
1404                 /*
1405                  * unblock after host_blocked iterates to zero
1406                  */
1407                 if (atomic_dec_return(&shost->host_blocked) > 0)
1408                         goto out_dec;
1409
1410                 SCSI_LOG_MLQUEUE(3,
1411                         shost_printk(KERN_INFO, shost,
1412                                      "unblocking host at zero depth\n"));
1413         }
1414
1415         if (shost->can_queue > 0 && busy >= shost->can_queue)
1416                 goto starved;
1417         if (shost->host_self_blocked)
1418                 goto starved;
1419
1420         /* We're OK to process the command, so we can't be starved */
1421         if (!list_empty(&sdev->starved_entry)) {
1422                 spin_lock_irq(shost->host_lock);
1423                 if (!list_empty(&sdev->starved_entry))
1424                         list_del_init(&sdev->starved_entry);
1425                 spin_unlock_irq(shost->host_lock);
1426         }
1427
1428         return 1;
1429
1430 starved:
1431         spin_lock_irq(shost->host_lock);
1432         if (list_empty(&sdev->starved_entry))
1433                 list_add_tail(&sdev->starved_entry, &shost->starved_list);
1434         spin_unlock_irq(shost->host_lock);
1435 out_dec:
1436         scsi_dec_host_busy(shost);
1437         return 0;
1438 }
1439
1440 /*
1441  * Busy state exporting function for request stacking drivers.
1442  *
1443  * For efficiency, no lock is taken to check the busy state of
1444  * shost/starget/sdev, since the returned value is not guaranteed and
1445  * may be changed after request stacking drivers call the function,
1446  * regardless of taking lock or not.
1447  *
1448  * When scsi can't dispatch I/Os anymore and needs to kill I/Os scsi
1449  * needs to return 'not busy'. Otherwise, request stacking drivers
1450  * may hold requests forever.
1451  */
1452 static bool scsi_mq_lld_busy(struct request_queue *q)
1453 {
1454         struct scsi_device *sdev = q->queuedata;
1455         struct Scsi_Host *shost;
1456
1457         if (blk_queue_dying(q))
1458                 return false;
1459
1460         shost = sdev->host;
1461
1462         /*
1463          * Ignore host/starget busy state.
1464          * Since block layer does not have a concept of fairness across
1465          * multiple queues, congestion of host/starget needs to be handled
1466          * in SCSI layer.
1467          */
1468         if (scsi_host_in_recovery(shost) || scsi_device_is_busy(sdev))
1469                 return true;
1470
1471         return false;
1472 }
1473
1474 static void scsi_softirq_done(struct request *rq)
1475 {
1476         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1477         unsigned long wait_for = (cmd->allowed + 1) * rq->timeout;
1478         int disposition;
1479
1480         INIT_LIST_HEAD(&cmd->eh_entry);
1481
1482         atomic_inc(&cmd->device->iodone_cnt);
1483         if (cmd->result)
1484                 atomic_inc(&cmd->device->ioerr_cnt);
1485
1486         disposition = scsi_decide_disposition(cmd);
1487         if (disposition != SUCCESS &&
1488             time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) {
1489                 sdev_printk(KERN_ERR, cmd->device,
1490                             "timing out command, waited %lus\n",
1491                             wait_for/HZ);
1492                 disposition = SUCCESS;
1493         }
1494
1495         scsi_log_completion(cmd, disposition);
1496
1497         switch (disposition) {
1498                 case SUCCESS:
1499                         scsi_finish_command(cmd);
1500                         break;
1501                 case NEEDS_RETRY:
1502                         scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY);
1503                         break;
1504                 case ADD_TO_MLQUEUE:
1505                         scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY);
1506                         break;
1507                 default:
1508                         scsi_eh_scmd_add(cmd);
1509                         break;
1510         }
1511 }
1512
1513 /**
1514  * scsi_dispatch_command - Dispatch a command to the low-level driver.
1515  * @cmd: command block we are dispatching.
1516  *
1517  * Return: nonzero return request was rejected and device's queue needs to be
1518  * plugged.
1519  */
1520 static int scsi_dispatch_cmd(struct scsi_cmnd *cmd)
1521 {
1522         struct Scsi_Host *host = cmd->device->host;
1523         int rtn = 0;
1524
1525         atomic_inc(&cmd->device->iorequest_cnt);
1526
1527         /* check if the device is still usable */
1528         if (unlikely(cmd->device->sdev_state == SDEV_DEL)) {
1529                 /* in SDEV_DEL we error all commands. DID_NO_CONNECT
1530                  * returns an immediate error upwards, and signals
1531                  * that the device is no longer present */
1532                 cmd->result = DID_NO_CONNECT << 16;
1533                 goto done;
1534         }
1535
1536         /* Check to see if the scsi lld made this device blocked. */
1537         if (unlikely(scsi_device_blocked(cmd->device))) {
1538                 /*
1539                  * in blocked state, the command is just put back on
1540                  * the device queue.  The suspend state has already
1541                  * blocked the queue so future requests should not
1542                  * occur until the device transitions out of the
1543                  * suspend state.
1544                  */
1545                 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1546                         "queuecommand : device blocked\n"));
1547                 return SCSI_MLQUEUE_DEVICE_BUSY;
1548         }
1549
1550         /* Store the LUN value in cmnd, if needed. */
1551         if (cmd->device->lun_in_cdb)
1552                 cmd->cmnd[1] = (cmd->cmnd[1] & 0x1f) |
1553                                (cmd->device->lun << 5 & 0xe0);
1554
1555         scsi_log_send(cmd);
1556
1557         /*
1558          * Before we queue this command, check if the command
1559          * length exceeds what the host adapter can handle.
1560          */
1561         if (cmd->cmd_len > cmd->device->host->max_cmd_len) {
1562                 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1563                                "queuecommand : command too long. "
1564                                "cdb_size=%d host->max_cmd_len=%d\n",
1565                                cmd->cmd_len, cmd->device->host->max_cmd_len));
1566                 cmd->result = (DID_ABORT << 16);
1567                 goto done;
1568         }
1569
1570         if (unlikely(host->shost_state == SHOST_DEL)) {
1571                 cmd->result = (DID_NO_CONNECT << 16);
1572                 goto done;
1573
1574         }
1575
1576         trace_scsi_dispatch_cmd_start(cmd);
1577         rtn = host->hostt->queuecommand(host, cmd);
1578         if (rtn) {
1579                 trace_scsi_dispatch_cmd_error(cmd, rtn);
1580                 if (rtn != SCSI_MLQUEUE_DEVICE_BUSY &&
1581                     rtn != SCSI_MLQUEUE_TARGET_BUSY)
1582                         rtn = SCSI_MLQUEUE_HOST_BUSY;
1583
1584                 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1585                         "queuecommand : request rejected\n"));
1586         }
1587
1588         return rtn;
1589  done:
1590         cmd->scsi_done(cmd);
1591         return 0;
1592 }
1593
1594 /* Size in bytes of the sg-list stored in the scsi-mq command-private data. */
1595 static unsigned int scsi_mq_sgl_size(struct Scsi_Host *shost)
1596 {
1597         return min_t(unsigned int, shost->sg_tablesize, SG_CHUNK_SIZE) *
1598                 sizeof(struct scatterlist);
1599 }
1600
1601 static blk_status_t scsi_mq_prep_fn(struct request *req)
1602 {
1603         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1604         struct scsi_device *sdev = req->q->queuedata;
1605         struct Scsi_Host *shost = sdev->host;
1606         struct scatterlist *sg;
1607
1608         scsi_init_command(sdev, cmd);
1609
1610         req->special = cmd;
1611
1612         cmd->request = req;
1613
1614         cmd->tag = req->tag;
1615         cmd->prot_op = SCSI_PROT_NORMAL;
1616
1617         sg = (void *)cmd + sizeof(struct scsi_cmnd) + shost->hostt->cmd_size;
1618         cmd->sdb.table.sgl = sg;
1619
1620         if (scsi_host_get_prot(shost)) {
1621                 memset(cmd->prot_sdb, 0, sizeof(struct scsi_data_buffer));
1622
1623                 cmd->prot_sdb->table.sgl =
1624                         (struct scatterlist *)(cmd->prot_sdb + 1);
1625         }
1626
1627         if (blk_bidi_rq(req)) {
1628                 struct request *next_rq = req->next_rq;
1629                 struct scsi_data_buffer *bidi_sdb = blk_mq_rq_to_pdu(next_rq);
1630
1631                 memset(bidi_sdb, 0, sizeof(struct scsi_data_buffer));
1632                 bidi_sdb->table.sgl =
1633                         (struct scatterlist *)(bidi_sdb + 1);
1634
1635                 next_rq->special = bidi_sdb;
1636         }
1637
1638         blk_mq_start_request(req);
1639
1640         return scsi_setup_cmnd(sdev, req);
1641 }
1642
1643 static void scsi_mq_done(struct scsi_cmnd *cmd)
1644 {
1645         if (unlikely(test_and_set_bit(SCMD_STATE_COMPLETE, &cmd->state)))
1646                 return;
1647         trace_scsi_dispatch_cmd_done(cmd);
1648
1649         /*
1650          * If the block layer didn't complete the request due to a timeout
1651          * injection, scsi must clear its internal completed state so that the
1652          * timeout handler will see it needs to escalate its own error
1653          * recovery.
1654          */
1655         if (unlikely(!blk_mq_complete_request(cmd->request)))
1656                 clear_bit(SCMD_STATE_COMPLETE, &cmd->state);
1657 }
1658
1659 static void scsi_mq_put_budget(struct blk_mq_hw_ctx *hctx)
1660 {
1661         struct request_queue *q = hctx->queue;
1662         struct scsi_device *sdev = q->queuedata;
1663
1664         atomic_dec(&sdev->device_busy);
1665         put_device(&sdev->sdev_gendev);
1666 }
1667
1668 static bool scsi_mq_get_budget(struct blk_mq_hw_ctx *hctx)
1669 {
1670         struct request_queue *q = hctx->queue;
1671         struct scsi_device *sdev = q->queuedata;
1672
1673         if (!get_device(&sdev->sdev_gendev))
1674                 goto out;
1675         if (!scsi_dev_queue_ready(q, sdev))
1676                 goto out_put_device;
1677
1678         return true;
1679
1680 out_put_device:
1681         put_device(&sdev->sdev_gendev);
1682 out:
1683         if (atomic_read(&sdev->device_busy) == 0 && !scsi_device_blocked(sdev))
1684                 blk_mq_delay_run_hw_queue(hctx, SCSI_QUEUE_DELAY);
1685         return false;
1686 }
1687
1688 static blk_status_t scsi_queue_rq(struct blk_mq_hw_ctx *hctx,
1689                          const struct blk_mq_queue_data *bd)
1690 {
1691         struct request *req = bd->rq;
1692         struct request_queue *q = req->q;
1693         struct scsi_device *sdev = q->queuedata;
1694         struct Scsi_Host *shost = sdev->host;
1695         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1696         blk_status_t ret;
1697         int reason;
1698
1699         /*
1700          * If the device is not in running state we will reject some or all
1701          * commands.
1702          */
1703         if (unlikely(sdev->sdev_state != SDEV_RUNNING)) {
1704                 ret = scsi_prep_state_check(sdev, req);
1705                 if (ret != BLK_STS_OK)
1706                         goto out_put_budget;
1707         }
1708
1709         ret = BLK_STS_RESOURCE;
1710         if (!scsi_target_queue_ready(shost, sdev))
1711                 goto out_put_budget;
1712         if (!scsi_host_queue_ready(q, shost, sdev))
1713                 goto out_dec_target_busy;
1714
1715         clear_bit(SCMD_STATE_COMPLETE, &cmd->state);
1716         if (!(req->rq_flags & RQF_DONTPREP)) {
1717                 ret = scsi_mq_prep_fn(req);
1718                 if (ret != BLK_STS_OK)
1719                         goto out_dec_host_busy;
1720                 req->rq_flags |= RQF_DONTPREP;
1721         } else {
1722                 blk_mq_start_request(req);
1723         }
1724
1725         if (sdev->simple_tags)
1726                 cmd->flags |= SCMD_TAGGED;
1727         else
1728                 cmd->flags &= ~SCMD_TAGGED;
1729
1730         scsi_init_cmd_errh(cmd);
1731         cmd->scsi_done = scsi_mq_done;
1732
1733         reason = scsi_dispatch_cmd(cmd);
1734         if (reason) {
1735                 scsi_set_blocked(cmd, reason);
1736                 ret = BLK_STS_RESOURCE;
1737                 goto out_dec_host_busy;
1738         }
1739
1740         return BLK_STS_OK;
1741
1742 out_dec_host_busy:
1743         scsi_dec_host_busy(shost);
1744 out_dec_target_busy:
1745         if (scsi_target(sdev)->can_queue > 0)
1746                 atomic_dec(&scsi_target(sdev)->target_busy);
1747 out_put_budget:
1748         scsi_mq_put_budget(hctx);
1749         switch (ret) {
1750         case BLK_STS_OK:
1751                 break;
1752         case BLK_STS_RESOURCE:
1753                 if (atomic_read(&sdev->device_busy) ||
1754                     scsi_device_blocked(sdev))
1755                         ret = BLK_STS_DEV_RESOURCE;
1756                 break;
1757         default:
1758                 /*
1759                  * Make sure to release all allocated ressources when
1760                  * we hit an error, as we will never see this command
1761                  * again.
1762                  */
1763                 if (req->rq_flags & RQF_DONTPREP)
1764                         scsi_mq_uninit_cmd(cmd);
1765                 break;
1766         }
1767         return ret;
1768 }
1769
1770 static enum blk_eh_timer_return scsi_timeout(struct request *req,
1771                 bool reserved)
1772 {
1773         if (reserved)
1774                 return BLK_EH_RESET_TIMER;
1775         return scsi_times_out(req);
1776 }
1777
1778 static int scsi_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
1779                                 unsigned int hctx_idx, unsigned int numa_node)
1780 {
1781         struct Scsi_Host *shost = set->driver_data;
1782         const bool unchecked_isa_dma = shost->unchecked_isa_dma;
1783         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1784         struct scatterlist *sg;
1785
1786         if (unchecked_isa_dma)
1787                 cmd->flags |= SCMD_UNCHECKED_ISA_DMA;
1788         cmd->sense_buffer = scsi_alloc_sense_buffer(unchecked_isa_dma,
1789                                                     GFP_KERNEL, numa_node);
1790         if (!cmd->sense_buffer)
1791                 return -ENOMEM;
1792         cmd->req.sense = cmd->sense_buffer;
1793
1794         if (scsi_host_get_prot(shost)) {
1795                 sg = (void *)cmd + sizeof(struct scsi_cmnd) +
1796                         shost->hostt->cmd_size;
1797                 cmd->prot_sdb = (void *)sg + scsi_mq_sgl_size(shost);
1798         }
1799
1800         return 0;
1801 }
1802
1803 static void scsi_mq_exit_request(struct blk_mq_tag_set *set, struct request *rq,
1804                                  unsigned int hctx_idx)
1805 {
1806         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1807
1808         scsi_free_sense_buffer(cmd->flags & SCMD_UNCHECKED_ISA_DMA,
1809                                cmd->sense_buffer);
1810 }
1811
1812 static int scsi_map_queues(struct blk_mq_tag_set *set)
1813 {
1814         struct Scsi_Host *shost = container_of(set, struct Scsi_Host, tag_set);
1815
1816         if (shost->hostt->map_queues)
1817                 return shost->hostt->map_queues(shost);
1818         return blk_mq_map_queues(&set->map[0]);
1819 }
1820
1821 void __scsi_init_queue(struct Scsi_Host *shost, struct request_queue *q)
1822 {
1823         struct device *dev = shost->dma_dev;
1824
1825         /*
1826          * this limit is imposed by hardware restrictions
1827          */
1828         blk_queue_max_segments(q, min_t(unsigned short, shost->sg_tablesize,
1829                                         SG_MAX_SEGMENTS));
1830
1831         if (scsi_host_prot_dma(shost)) {
1832                 shost->sg_prot_tablesize =
1833                         min_not_zero(shost->sg_prot_tablesize,
1834                                      (unsigned short)SCSI_MAX_PROT_SG_SEGMENTS);
1835                 BUG_ON(shost->sg_prot_tablesize < shost->sg_tablesize);
1836                 blk_queue_max_integrity_segments(q, shost->sg_prot_tablesize);
1837         }
1838
1839         blk_queue_max_hw_sectors(q, shost->max_sectors);
1840         if (shost->unchecked_isa_dma)
1841                 blk_queue_bounce_limit(q, BLK_BOUNCE_ISA);
1842         blk_queue_segment_boundary(q, shost->dma_boundary);
1843         dma_set_seg_boundary(dev, shost->dma_boundary);
1844
1845         blk_queue_max_segment_size(q, shost->max_segment_size);
1846         dma_set_max_seg_size(dev, shost->max_segment_size);
1847
1848         /*
1849          * Set a reasonable default alignment:  The larger of 32-byte (dword),
1850          * which is a common minimum for HBAs, and the minimum DMA alignment,
1851          * which is set by the platform.
1852          *
1853          * Devices that require a bigger alignment can increase it later.
1854          */
1855         blk_queue_dma_alignment(q, max(4, dma_get_cache_alignment()) - 1);
1856 }
1857 EXPORT_SYMBOL_GPL(__scsi_init_queue);
1858
1859 static const struct blk_mq_ops scsi_mq_ops = {
1860         .get_budget     = scsi_mq_get_budget,
1861         .put_budget     = scsi_mq_put_budget,
1862         .queue_rq       = scsi_queue_rq,
1863         .complete       = scsi_softirq_done,
1864         .timeout        = scsi_timeout,
1865 #ifdef CONFIG_BLK_DEBUG_FS
1866         .show_rq        = scsi_show_rq,
1867 #endif
1868         .init_request   = scsi_mq_init_request,
1869         .exit_request   = scsi_mq_exit_request,
1870         .initialize_rq_fn = scsi_initialize_rq,
1871         .busy           = scsi_mq_lld_busy,
1872         .map_queues     = scsi_map_queues,
1873 };
1874
1875 struct request_queue *scsi_mq_alloc_queue(struct scsi_device *sdev)
1876 {
1877         sdev->request_queue = blk_mq_init_queue(&sdev->host->tag_set);
1878         if (IS_ERR(sdev->request_queue))
1879                 return NULL;
1880
1881         sdev->request_queue->queuedata = sdev;
1882         __scsi_init_queue(sdev->host, sdev->request_queue);
1883         blk_queue_flag_set(QUEUE_FLAG_SCSI_PASSTHROUGH, sdev->request_queue);
1884         return sdev->request_queue;
1885 }
1886
1887 int scsi_mq_setup_tags(struct Scsi_Host *shost)
1888 {
1889         unsigned int cmd_size, sgl_size;
1890
1891         sgl_size = scsi_mq_sgl_size(shost);
1892         cmd_size = sizeof(struct scsi_cmnd) + shost->hostt->cmd_size + sgl_size;
1893         if (scsi_host_get_prot(shost))
1894                 cmd_size += sizeof(struct scsi_data_buffer) + sgl_size;
1895
1896         memset(&shost->tag_set, 0, sizeof(shost->tag_set));
1897         shost->tag_set.ops = &scsi_mq_ops;
1898         shost->tag_set.nr_hw_queues = shost->nr_hw_queues ? : 1;
1899         shost->tag_set.queue_depth = shost->can_queue;
1900         shost->tag_set.cmd_size = cmd_size;
1901         shost->tag_set.numa_node = NUMA_NO_NODE;
1902         shost->tag_set.flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_SG_MERGE;
1903         shost->tag_set.flags |=
1904                 BLK_ALLOC_POLICY_TO_MQ_FLAG(shost->hostt->tag_alloc_policy);
1905         shost->tag_set.driver_data = shost;
1906
1907         return blk_mq_alloc_tag_set(&shost->tag_set);
1908 }
1909
1910 void scsi_mq_destroy_tags(struct Scsi_Host *shost)
1911 {
1912         blk_mq_free_tag_set(&shost->tag_set);
1913 }
1914
1915 /**
1916  * scsi_device_from_queue - return sdev associated with a request_queue
1917  * @q: The request queue to return the sdev from
1918  *
1919  * Return the sdev associated with a request queue or NULL if the
1920  * request_queue does not reference a SCSI device.
1921  */
1922 struct scsi_device *scsi_device_from_queue(struct request_queue *q)
1923 {
1924         struct scsi_device *sdev = NULL;
1925
1926         if (q->mq_ops == &scsi_mq_ops)
1927                 sdev = q->queuedata;
1928         if (!sdev || !get_device(&sdev->sdev_gendev))
1929                 sdev = NULL;
1930
1931         return sdev;
1932 }
1933 EXPORT_SYMBOL_GPL(scsi_device_from_queue);
1934
1935 /*
1936  * Function:    scsi_block_requests()
1937  *
1938  * Purpose:     Utility function used by low-level drivers to prevent further
1939  *              commands from being queued to the device.
1940  *
1941  * Arguments:   shost       - Host in question
1942  *
1943  * Returns:     Nothing
1944  *
1945  * Lock status: No locks are assumed held.
1946  *
1947  * Notes:       There is no timer nor any other means by which the requests
1948  *              get unblocked other than the low-level driver calling
1949  *              scsi_unblock_requests().
1950  */
1951 void scsi_block_requests(struct Scsi_Host *shost)
1952 {
1953         shost->host_self_blocked = 1;
1954 }
1955 EXPORT_SYMBOL(scsi_block_requests);
1956
1957 /*
1958  * Function:    scsi_unblock_requests()
1959  *
1960  * Purpose:     Utility function used by low-level drivers to allow further
1961  *              commands from being queued to the device.
1962  *
1963  * Arguments:   shost       - Host in question
1964  *
1965  * Returns:     Nothing
1966  *
1967  * Lock status: No locks are assumed held.
1968  *
1969  * Notes:       There is no timer nor any other means by which the requests
1970  *              get unblocked other than the low-level driver calling
1971  *              scsi_unblock_requests().
1972  *
1973  *              This is done as an API function so that changes to the
1974  *              internals of the scsi mid-layer won't require wholesale
1975  *              changes to drivers that use this feature.
1976  */
1977 void scsi_unblock_requests(struct Scsi_Host *shost)
1978 {
1979         shost->host_self_blocked = 0;
1980         scsi_run_host_queues(shost);
1981 }
1982 EXPORT_SYMBOL(scsi_unblock_requests);
1983
1984 int __init scsi_init_queue(void)
1985 {
1986         scsi_sdb_cache = kmem_cache_create("scsi_data_buffer",
1987                                            sizeof(struct scsi_data_buffer),
1988                                            0, 0, NULL);
1989         if (!scsi_sdb_cache) {
1990                 printk(KERN_ERR "SCSI: can't init scsi sdb cache\n");
1991                 return -ENOMEM;
1992         }
1993
1994         return 0;
1995 }
1996
1997 void scsi_exit_queue(void)
1998 {
1999         kmem_cache_destroy(scsi_sense_cache);
2000         kmem_cache_destroy(scsi_sense_isadma_cache);
2001         kmem_cache_destroy(scsi_sdb_cache);
2002 }
2003
2004 /**
2005  *      scsi_mode_select - issue a mode select
2006  *      @sdev:  SCSI device to be queried
2007  *      @pf:    Page format bit (1 == standard, 0 == vendor specific)
2008  *      @sp:    Save page bit (0 == don't save, 1 == save)
2009  *      @modepage: mode page being requested
2010  *      @buffer: request buffer (may not be smaller than eight bytes)
2011  *      @len:   length of request buffer.
2012  *      @timeout: command timeout
2013  *      @retries: number of retries before failing
2014  *      @data: returns a structure abstracting the mode header data
2015  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
2016  *              must be SCSI_SENSE_BUFFERSIZE big.
2017  *
2018  *      Returns zero if successful; negative error number or scsi
2019  *      status on error
2020  *
2021  */
2022 int
2023 scsi_mode_select(struct scsi_device *sdev, int pf, int sp, int modepage,
2024                  unsigned char *buffer, int len, int timeout, int retries,
2025                  struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
2026 {
2027         unsigned char cmd[10];
2028         unsigned char *real_buffer;
2029         int ret;
2030
2031         memset(cmd, 0, sizeof(cmd));
2032         cmd[1] = (pf ? 0x10 : 0) | (sp ? 0x01 : 0);
2033
2034         if (sdev->use_10_for_ms) {
2035                 if (len > 65535)
2036                         return -EINVAL;
2037                 real_buffer = kmalloc(8 + len, GFP_KERNEL);
2038                 if (!real_buffer)
2039                         return -ENOMEM;
2040                 memcpy(real_buffer + 8, buffer, len);
2041                 len += 8;
2042                 real_buffer[0] = 0;
2043                 real_buffer[1] = 0;
2044                 real_buffer[2] = data->medium_type;
2045                 real_buffer[3] = data->device_specific;
2046                 real_buffer[4] = data->longlba ? 0x01 : 0;
2047                 real_buffer[5] = 0;
2048                 real_buffer[6] = data->block_descriptor_length >> 8;
2049                 real_buffer[7] = data->block_descriptor_length;
2050
2051                 cmd[0] = MODE_SELECT_10;
2052                 cmd[7] = len >> 8;
2053                 cmd[8] = len;
2054         } else {
2055                 if (len > 255 || data->block_descriptor_length > 255 ||
2056                     data->longlba)
2057                         return -EINVAL;
2058
2059                 real_buffer = kmalloc(4 + len, GFP_KERNEL);
2060                 if (!real_buffer)
2061                         return -ENOMEM;
2062                 memcpy(real_buffer + 4, buffer, len);
2063                 len += 4;
2064                 real_buffer[0] = 0;
2065                 real_buffer[1] = data->medium_type;
2066                 real_buffer[2] = data->device_specific;
2067                 real_buffer[3] = data->block_descriptor_length;
2068                 
2069
2070                 cmd[0] = MODE_SELECT;
2071                 cmd[4] = len;
2072         }
2073
2074         ret = scsi_execute_req(sdev, cmd, DMA_TO_DEVICE, real_buffer, len,
2075                                sshdr, timeout, retries, NULL);
2076         kfree(real_buffer);
2077         return ret;
2078 }
2079 EXPORT_SYMBOL_GPL(scsi_mode_select);
2080
2081 /**
2082  *      scsi_mode_sense - issue a mode sense, falling back from 10 to six bytes if necessary.
2083  *      @sdev:  SCSI device to be queried
2084  *      @dbd:   set if mode sense will allow block descriptors to be returned
2085  *      @modepage: mode page being requested
2086  *      @buffer: request buffer (may not be smaller than eight bytes)
2087  *      @len:   length of request buffer.
2088  *      @timeout: command timeout
2089  *      @retries: number of retries before failing
2090  *      @data: returns a structure abstracting the mode header data
2091  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
2092  *              must be SCSI_SENSE_BUFFERSIZE big.
2093  *
2094  *      Returns zero if unsuccessful, or the header offset (either 4
2095  *      or 8 depending on whether a six or ten byte command was
2096  *      issued) if successful.
2097  */
2098 int
2099 scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
2100                   unsigned char *buffer, int len, int timeout, int retries,
2101                   struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
2102 {
2103         unsigned char cmd[12];
2104         int use_10_for_ms;
2105         int header_length;
2106         int result, retry_count = retries;
2107         struct scsi_sense_hdr my_sshdr;
2108
2109         memset(data, 0, sizeof(*data));
2110         memset(&cmd[0], 0, 12);
2111         cmd[1] = dbd & 0x18;    /* allows DBD and LLBA bits */
2112         cmd[2] = modepage;
2113
2114         /* caller might not be interested in sense, but we need it */
2115         if (!sshdr)
2116                 sshdr = &my_sshdr;
2117
2118  retry:
2119         use_10_for_ms = sdev->use_10_for_ms;
2120
2121         if (use_10_for_ms) {
2122                 if (len < 8)
2123                         len = 8;
2124
2125                 cmd[0] = MODE_SENSE_10;
2126                 cmd[8] = len;
2127                 header_length = 8;
2128         } else {
2129                 if (len < 4)
2130                         len = 4;
2131
2132                 cmd[0] = MODE_SENSE;
2133                 cmd[4] = len;
2134                 header_length = 4;
2135         }
2136
2137         memset(buffer, 0, len);
2138
2139         result = scsi_execute_req(sdev, cmd, DMA_FROM_DEVICE, buffer, len,
2140                                   sshdr, timeout, retries, NULL);
2141
2142         /* This code looks awful: what it's doing is making sure an
2143          * ILLEGAL REQUEST sense return identifies the actual command
2144          * byte as the problem.  MODE_SENSE commands can return
2145          * ILLEGAL REQUEST if the code page isn't supported */
2146
2147         if (use_10_for_ms && !scsi_status_is_good(result) &&
2148             driver_byte(result) == DRIVER_SENSE) {
2149                 if (scsi_sense_valid(sshdr)) {
2150                         if ((sshdr->sense_key == ILLEGAL_REQUEST) &&
2151                             (sshdr->asc == 0x20) && (sshdr->ascq == 0)) {
2152                                 /* 
2153                                  * Invalid command operation code
2154                                  */
2155                                 sdev->use_10_for_ms = 0;
2156                                 goto retry;
2157                         }
2158                 }
2159         }
2160
2161         if(scsi_status_is_good(result)) {
2162                 if (unlikely(buffer[0] == 0x86 && buffer[1] == 0x0b &&
2163                              (modepage == 6 || modepage == 8))) {
2164                         /* Initio breakage? */
2165                         header_length = 0;
2166                         data->length = 13;
2167                         data->medium_type = 0;
2168                         data->device_specific = 0;
2169                         data->longlba = 0;
2170                         data->block_descriptor_length = 0;
2171                 } else if(use_10_for_ms) {
2172                         data->length = buffer[0]*256 + buffer[1] + 2;
2173                         data->medium_type = buffer[2];
2174                         data->device_specific = buffer[3];
2175                         data->longlba = buffer[4] & 0x01;
2176                         data->block_descriptor_length = buffer[6]*256
2177                                 + buffer[7];
2178                 } else {
2179                         data->length = buffer[0] + 1;
2180                         data->medium_type = buffer[1];
2181                         data->device_specific = buffer[2];
2182                         data->block_descriptor_length = buffer[3];
2183                 }
2184                 data->header_length = header_length;
2185         } else if ((status_byte(result) == CHECK_CONDITION) &&
2186                    scsi_sense_valid(sshdr) &&
2187                    sshdr->sense_key == UNIT_ATTENTION && retry_count) {
2188                 retry_count--;
2189                 goto retry;
2190         }
2191
2192         return result;
2193 }
2194 EXPORT_SYMBOL(scsi_mode_sense);
2195
2196 /**
2197  *      scsi_test_unit_ready - test if unit is ready
2198  *      @sdev:  scsi device to change the state of.
2199  *      @timeout: command timeout
2200  *      @retries: number of retries before failing
2201  *      @sshdr: outpout pointer for decoded sense information.
2202  *
2203  *      Returns zero if unsuccessful or an error if TUR failed.  For
2204  *      removable media, UNIT_ATTENTION sets ->changed flag.
2205  **/
2206 int
2207 scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries,
2208                      struct scsi_sense_hdr *sshdr)
2209 {
2210         char cmd[] = {
2211                 TEST_UNIT_READY, 0, 0, 0, 0, 0,
2212         };
2213         int result;
2214
2215         /* try to eat the UNIT_ATTENTION if there are enough retries */
2216         do {
2217                 result = scsi_execute_req(sdev, cmd, DMA_NONE, NULL, 0, sshdr,
2218                                           timeout, 1, NULL);
2219                 if (sdev->removable && scsi_sense_valid(sshdr) &&
2220                     sshdr->sense_key == UNIT_ATTENTION)
2221                         sdev->changed = 1;
2222         } while (scsi_sense_valid(sshdr) &&
2223                  sshdr->sense_key == UNIT_ATTENTION && --retries);
2224
2225         return result;
2226 }
2227 EXPORT_SYMBOL(scsi_test_unit_ready);
2228
2229 /**
2230  *      scsi_device_set_state - Take the given device through the device state model.
2231  *      @sdev:  scsi device to change the state of.
2232  *      @state: state to change to.
2233  *
2234  *      Returns zero if successful or an error if the requested
2235  *      transition is illegal.
2236  */
2237 int
2238 scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
2239 {
2240         enum scsi_device_state oldstate = sdev->sdev_state;
2241
2242         if (state == oldstate)
2243                 return 0;
2244
2245         switch (state) {
2246         case SDEV_CREATED:
2247                 switch (oldstate) {
2248                 case SDEV_CREATED_BLOCK:
2249                         break;
2250                 default:
2251                         goto illegal;
2252                 }
2253                 break;
2254                         
2255         case SDEV_RUNNING:
2256                 switch (oldstate) {
2257                 case SDEV_CREATED:
2258                 case SDEV_OFFLINE:
2259                 case SDEV_TRANSPORT_OFFLINE:
2260                 case SDEV_QUIESCE:
2261                 case SDEV_BLOCK:
2262                         break;
2263                 default:
2264                         goto illegal;
2265                 }
2266                 break;
2267
2268         case SDEV_QUIESCE:
2269                 switch (oldstate) {
2270                 case SDEV_RUNNING:
2271                 case SDEV_OFFLINE:
2272                 case SDEV_TRANSPORT_OFFLINE:
2273                         break;
2274                 default:
2275                         goto illegal;
2276                 }
2277                 break;
2278
2279         case SDEV_OFFLINE:
2280         case SDEV_TRANSPORT_OFFLINE:
2281                 switch (oldstate) {
2282                 case SDEV_CREATED:
2283                 case SDEV_RUNNING:
2284                 case SDEV_QUIESCE:
2285                 case SDEV_BLOCK:
2286                         break;
2287                 default:
2288                         goto illegal;
2289                 }
2290                 break;
2291
2292         case SDEV_BLOCK:
2293                 switch (oldstate) {
2294                 case SDEV_RUNNING:
2295                 case SDEV_CREATED_BLOCK:
2296                 case SDEV_OFFLINE:
2297                         break;
2298                 default:
2299                         goto illegal;
2300                 }
2301                 break;
2302
2303         case SDEV_CREATED_BLOCK:
2304                 switch (oldstate) {
2305                 case SDEV_CREATED:
2306                         break;
2307                 default:
2308                         goto illegal;
2309                 }
2310                 break;
2311
2312         case SDEV_CANCEL:
2313                 switch (oldstate) {
2314                 case SDEV_CREATED:
2315                 case SDEV_RUNNING:
2316                 case SDEV_QUIESCE:
2317                 case SDEV_OFFLINE:
2318                 case SDEV_TRANSPORT_OFFLINE:
2319                         break;
2320                 default:
2321                         goto illegal;
2322                 }
2323                 break;
2324
2325         case SDEV_DEL:
2326                 switch (oldstate) {
2327                 case SDEV_CREATED:
2328                 case SDEV_RUNNING:
2329                 case SDEV_OFFLINE:
2330                 case SDEV_TRANSPORT_OFFLINE:
2331                 case SDEV_CANCEL:
2332                 case SDEV_BLOCK:
2333                 case SDEV_CREATED_BLOCK:
2334                         break;
2335                 default:
2336                         goto illegal;
2337                 }
2338                 break;
2339
2340         }
2341         sdev->sdev_state = state;
2342         return 0;
2343
2344  illegal:
2345         SCSI_LOG_ERROR_RECOVERY(1,
2346                                 sdev_printk(KERN_ERR, sdev,
2347                                             "Illegal state transition %s->%s",
2348                                             scsi_device_state_name(oldstate),
2349                                             scsi_device_state_name(state))
2350                                 );
2351         return -EINVAL;
2352 }
2353 EXPORT_SYMBOL(scsi_device_set_state);
2354
2355 /**
2356  *      sdev_evt_emit - emit a single SCSI device uevent
2357  *      @sdev: associated SCSI device
2358  *      @evt: event to emit
2359  *
2360  *      Send a single uevent (scsi_event) to the associated scsi_device.
2361  */
2362 static void scsi_evt_emit(struct scsi_device *sdev, struct scsi_event *evt)
2363 {
2364         int idx = 0;
2365         char *envp[3];
2366
2367         switch (evt->evt_type) {
2368         case SDEV_EVT_MEDIA_CHANGE:
2369                 envp[idx++] = "SDEV_MEDIA_CHANGE=1";
2370                 break;
2371         case SDEV_EVT_INQUIRY_CHANGE_REPORTED:
2372                 scsi_rescan_device(&sdev->sdev_gendev);
2373                 envp[idx++] = "SDEV_UA=INQUIRY_DATA_HAS_CHANGED";
2374                 break;
2375         case SDEV_EVT_CAPACITY_CHANGE_REPORTED:
2376                 envp[idx++] = "SDEV_UA=CAPACITY_DATA_HAS_CHANGED";
2377                 break;
2378         case SDEV_EVT_SOFT_THRESHOLD_REACHED_REPORTED:
2379                envp[idx++] = "SDEV_UA=THIN_PROVISIONING_SOFT_THRESHOLD_REACHED";
2380                 break;
2381         case SDEV_EVT_MODE_PARAMETER_CHANGE_REPORTED:
2382                 envp[idx++] = "SDEV_UA=MODE_PARAMETERS_CHANGED";
2383                 break;
2384         case SDEV_EVT_LUN_CHANGE_REPORTED:
2385                 envp[idx++] = "SDEV_UA=REPORTED_LUNS_DATA_HAS_CHANGED";
2386                 break;
2387         case SDEV_EVT_ALUA_STATE_CHANGE_REPORTED:
2388                 envp[idx++] = "SDEV_UA=ASYMMETRIC_ACCESS_STATE_CHANGED";
2389                 break;
2390         case SDEV_EVT_POWER_ON_RESET_OCCURRED:
2391                 envp[idx++] = "SDEV_UA=POWER_ON_RESET_OCCURRED";
2392                 break;
2393         default:
2394                 /* do nothing */
2395                 break;
2396         }
2397
2398         envp[idx++] = NULL;
2399
2400         kobject_uevent_env(&sdev->sdev_gendev.kobj, KOBJ_CHANGE, envp);
2401 }
2402
2403 /**
2404  *      sdev_evt_thread - send a uevent for each scsi event
2405  *      @work: work struct for scsi_device
2406  *
2407  *      Dispatch queued events to their associated scsi_device kobjects
2408  *      as uevents.
2409  */
2410 void scsi_evt_thread(struct work_struct *work)
2411 {
2412         struct scsi_device *sdev;
2413         enum scsi_device_event evt_type;
2414         LIST_HEAD(event_list);
2415
2416         sdev = container_of(work, struct scsi_device, event_work);
2417
2418         for (evt_type = SDEV_EVT_FIRST; evt_type <= SDEV_EVT_LAST; evt_type++)
2419                 if (test_and_clear_bit(evt_type, sdev->pending_events))
2420                         sdev_evt_send_simple(sdev, evt_type, GFP_KERNEL);
2421
2422         while (1) {
2423                 struct scsi_event *evt;
2424                 struct list_head *this, *tmp;
2425                 unsigned long flags;
2426
2427                 spin_lock_irqsave(&sdev->list_lock, flags);
2428                 list_splice_init(&sdev->event_list, &event_list);
2429                 spin_unlock_irqrestore(&sdev->list_lock, flags);
2430
2431                 if (list_empty(&event_list))
2432                         break;
2433
2434                 list_for_each_safe(this, tmp, &event_list) {
2435                         evt = list_entry(this, struct scsi_event, node);
2436                         list_del(&evt->node);
2437                         scsi_evt_emit(sdev, evt);
2438                         kfree(evt);
2439                 }
2440         }
2441 }
2442
2443 /**
2444  *      sdev_evt_send - send asserted event to uevent thread
2445  *      @sdev: scsi_device event occurred on
2446  *      @evt: event to send
2447  *
2448  *      Assert scsi device event asynchronously.
2449  */
2450 void sdev_evt_send(struct scsi_device *sdev, struct scsi_event *evt)
2451 {
2452         unsigned long flags;
2453
2454 #if 0
2455         /* FIXME: currently this check eliminates all media change events
2456          * for polled devices.  Need to update to discriminate between AN
2457          * and polled events */
2458         if (!test_bit(evt->evt_type, sdev->supported_events)) {
2459                 kfree(evt);
2460                 return;
2461         }
2462 #endif
2463
2464         spin_lock_irqsave(&sdev->list_lock, flags);
2465         list_add_tail(&evt->node, &sdev->event_list);
2466         schedule_work(&sdev->event_work);
2467         spin_unlock_irqrestore(&sdev->list_lock, flags);
2468 }
2469 EXPORT_SYMBOL_GPL(sdev_evt_send);
2470
2471 /**
2472  *      sdev_evt_alloc - allocate a new scsi event
2473  *      @evt_type: type of event to allocate
2474  *      @gfpflags: GFP flags for allocation
2475  *
2476  *      Allocates and returns a new scsi_event.
2477  */
2478 struct scsi_event *sdev_evt_alloc(enum scsi_device_event evt_type,
2479                                   gfp_t gfpflags)
2480 {
2481         struct scsi_event *evt = kzalloc(sizeof(struct scsi_event), gfpflags);
2482         if (!evt)
2483                 return NULL;
2484
2485         evt->evt_type = evt_type;
2486         INIT_LIST_HEAD(&evt->node);
2487
2488         /* evt_type-specific initialization, if any */
2489         switch (evt_type) {
2490         case SDEV_EVT_MEDIA_CHANGE:
2491         case SDEV_EVT_INQUIRY_CHANGE_REPORTED:
2492         case SDEV_EVT_CAPACITY_CHANGE_REPORTED:
2493         case SDEV_EVT_SOFT_THRESHOLD_REACHED_REPORTED:
2494         case SDEV_EVT_MODE_PARAMETER_CHANGE_REPORTED:
2495         case SDEV_EVT_LUN_CHANGE_REPORTED:
2496         case SDEV_EVT_ALUA_STATE_CHANGE_REPORTED:
2497         case SDEV_EVT_POWER_ON_RESET_OCCURRED:
2498         default:
2499                 /* do nothing */
2500                 break;
2501         }
2502
2503         return evt;
2504 }
2505 EXPORT_SYMBOL_GPL(sdev_evt_alloc);
2506
2507 /**
2508  *      sdev_evt_send_simple - send asserted event to uevent thread
2509  *      @sdev: scsi_device event occurred on
2510  *      @evt_type: type of event to send
2511  *      @gfpflags: GFP flags for allocation
2512  *
2513  *      Assert scsi device event asynchronously, given an event type.
2514  */
2515 void sdev_evt_send_simple(struct scsi_device *sdev,
2516                           enum scsi_device_event evt_type, gfp_t gfpflags)
2517 {
2518         struct scsi_event *evt = sdev_evt_alloc(evt_type, gfpflags);
2519         if (!evt) {
2520                 sdev_printk(KERN_ERR, sdev, "event %d eaten due to OOM\n",
2521                             evt_type);
2522                 return;
2523         }
2524
2525         sdev_evt_send(sdev, evt);
2526 }
2527 EXPORT_SYMBOL_GPL(sdev_evt_send_simple);
2528
2529 /**
2530  *      scsi_device_quiesce - Block user issued commands.
2531  *      @sdev:  scsi device to quiesce.
2532  *
2533  *      This works by trying to transition to the SDEV_QUIESCE state
2534  *      (which must be a legal transition).  When the device is in this
2535  *      state, only special requests will be accepted, all others will
2536  *      be deferred.  Since special requests may also be requeued requests,
2537  *      a successful return doesn't guarantee the device will be 
2538  *      totally quiescent.
2539  *
2540  *      Must be called with user context, may sleep.
2541  *
2542  *      Returns zero if unsuccessful or an error if not.
2543  */
2544 int
2545 scsi_device_quiesce(struct scsi_device *sdev)
2546 {
2547         struct request_queue *q = sdev->request_queue;
2548         int err;
2549
2550         /*
2551          * It is allowed to call scsi_device_quiesce() multiple times from
2552          * the same context but concurrent scsi_device_quiesce() calls are
2553          * not allowed.
2554          */
2555         WARN_ON_ONCE(sdev->quiesced_by && sdev->quiesced_by != current);
2556
2557         if (sdev->quiesced_by == current)
2558                 return 0;
2559
2560         blk_set_pm_only(q);
2561
2562         blk_mq_freeze_queue(q);
2563         /*
2564          * Ensure that the effect of blk_set_pm_only() will be visible
2565          * for percpu_ref_tryget() callers that occur after the queue
2566          * unfreeze even if the queue was already frozen before this function
2567          * was called. See also https://lwn.net/Articles/573497/.
2568          */
2569         synchronize_rcu();
2570         blk_mq_unfreeze_queue(q);
2571
2572         mutex_lock(&sdev->state_mutex);
2573         err = scsi_device_set_state(sdev, SDEV_QUIESCE);
2574         if (err == 0)
2575                 sdev->quiesced_by = current;
2576         else
2577                 blk_clear_pm_only(q);
2578         mutex_unlock(&sdev->state_mutex);
2579
2580         return err;
2581 }
2582 EXPORT_SYMBOL(scsi_device_quiesce);
2583
2584 /**
2585  *      scsi_device_resume - Restart user issued commands to a quiesced device.
2586  *      @sdev:  scsi device to resume.
2587  *
2588  *      Moves the device from quiesced back to running and restarts the
2589  *      queues.
2590  *
2591  *      Must be called with user context, may sleep.
2592  */
2593 void scsi_device_resume(struct scsi_device *sdev)
2594 {
2595         /* check if the device state was mutated prior to resume, and if
2596          * so assume the state is being managed elsewhere (for example
2597          * device deleted during suspend)
2598          */
2599         mutex_lock(&sdev->state_mutex);
2600         WARN_ON_ONCE(!sdev->quiesced_by);
2601         sdev->quiesced_by = NULL;
2602         blk_clear_pm_only(sdev->request_queue);
2603         if (sdev->sdev_state == SDEV_QUIESCE)
2604                 scsi_device_set_state(sdev, SDEV_RUNNING);
2605         mutex_unlock(&sdev->state_mutex);
2606 }
2607 EXPORT_SYMBOL(scsi_device_resume);
2608
2609 static void
2610 device_quiesce_fn(struct scsi_device *sdev, void *data)
2611 {
2612         scsi_device_quiesce(sdev);
2613 }
2614
2615 void
2616 scsi_target_quiesce(struct scsi_target *starget)
2617 {
2618         starget_for_each_device(starget, NULL, device_quiesce_fn);
2619 }
2620 EXPORT_SYMBOL(scsi_target_quiesce);
2621
2622 static void
2623 device_resume_fn(struct scsi_device *sdev, void *data)
2624 {
2625         scsi_device_resume(sdev);
2626 }
2627
2628 void
2629 scsi_target_resume(struct scsi_target *starget)
2630 {
2631         starget_for_each_device(starget, NULL, device_resume_fn);
2632 }
2633 EXPORT_SYMBOL(scsi_target_resume);
2634
2635 /**
2636  * scsi_internal_device_block_nowait - try to transition to the SDEV_BLOCK state
2637  * @sdev: device to block
2638  *
2639  * Pause SCSI command processing on the specified device. Does not sleep.
2640  *
2641  * Returns zero if successful or a negative error code upon failure.
2642  *
2643  * Notes:
2644  * This routine transitions the device to the SDEV_BLOCK state (which must be
2645  * a legal transition). When the device is in this state, command processing
2646  * is paused until the device leaves the SDEV_BLOCK state. See also
2647  * scsi_internal_device_unblock_nowait().
2648  */
2649 int scsi_internal_device_block_nowait(struct scsi_device *sdev)
2650 {
2651         struct request_queue *q = sdev->request_queue;
2652         int err = 0;
2653
2654         err = scsi_device_set_state(sdev, SDEV_BLOCK);
2655         if (err) {
2656                 err = scsi_device_set_state(sdev, SDEV_CREATED_BLOCK);
2657
2658                 if (err)
2659                         return err;
2660         }
2661
2662         /* 
2663          * The device has transitioned to SDEV_BLOCK.  Stop the
2664          * block layer from calling the midlayer with this device's
2665          * request queue. 
2666          */
2667         blk_mq_quiesce_queue_nowait(q);
2668         return 0;
2669 }
2670 EXPORT_SYMBOL_GPL(scsi_internal_device_block_nowait);
2671
2672 /**
2673  * scsi_internal_device_block - try to transition to the SDEV_BLOCK state
2674  * @sdev: device to block
2675  *
2676  * Pause SCSI command processing on the specified device and wait until all
2677  * ongoing scsi_request_fn() / scsi_queue_rq() calls have finished. May sleep.
2678  *
2679  * Returns zero if successful or a negative error code upon failure.
2680  *
2681  * Note:
2682  * This routine transitions the device to the SDEV_BLOCK state (which must be
2683  * a legal transition). When the device is in this state, command processing
2684  * is paused until the device leaves the SDEV_BLOCK state. See also
2685  * scsi_internal_device_unblock().
2686  *
2687  * To do: avoid that scsi_send_eh_cmnd() calls queuecommand() after
2688  * scsi_internal_device_block() has blocked a SCSI device and also
2689  * remove the rport mutex lock and unlock calls from srp_queuecommand().
2690  */
2691 static int scsi_internal_device_block(struct scsi_device *sdev)
2692 {
2693         struct request_queue *q = sdev->request_queue;
2694         int err;
2695
2696         mutex_lock(&sdev->state_mutex);
2697         err = scsi_internal_device_block_nowait(sdev);
2698         if (err == 0)
2699                 blk_mq_quiesce_queue(q);
2700         mutex_unlock(&sdev->state_mutex);
2701
2702         return err;
2703 }
2704  
2705 void scsi_start_queue(struct scsi_device *sdev)
2706 {
2707         struct request_queue *q = sdev->request_queue;
2708
2709         blk_mq_unquiesce_queue(q);
2710 }
2711
2712 /**
2713  * scsi_internal_device_unblock_nowait - resume a device after a block request
2714  * @sdev:       device to resume
2715  * @new_state:  state to set the device to after unblocking
2716  *
2717  * Restart the device queue for a previously suspended SCSI device. Does not
2718  * sleep.
2719  *
2720  * Returns zero if successful or a negative error code upon failure.
2721  *
2722  * Notes:
2723  * This routine transitions the device to the SDEV_RUNNING state or to one of
2724  * the offline states (which must be a legal transition) allowing the midlayer
2725  * to goose the queue for this device.
2726  */
2727 int scsi_internal_device_unblock_nowait(struct scsi_device *sdev,
2728                                         enum scsi_device_state new_state)
2729 {
2730         /*
2731          * Try to transition the scsi device to SDEV_RUNNING or one of the
2732          * offlined states and goose the device queue if successful.
2733          */
2734         switch (sdev->sdev_state) {
2735         case SDEV_BLOCK:
2736         case SDEV_TRANSPORT_OFFLINE:
2737                 sdev->sdev_state = new_state;
2738                 break;
2739         case SDEV_CREATED_BLOCK:
2740                 if (new_state == SDEV_TRANSPORT_OFFLINE ||
2741                     new_state == SDEV_OFFLINE)
2742                         sdev->sdev_state = new_state;
2743                 else
2744                         sdev->sdev_state = SDEV_CREATED;
2745                 break;
2746         case SDEV_CANCEL:
2747         case SDEV_OFFLINE:
2748                 break;
2749         default:
2750                 return -EINVAL;
2751         }
2752         scsi_start_queue(sdev);
2753
2754         return 0;
2755 }
2756 EXPORT_SYMBOL_GPL(scsi_internal_device_unblock_nowait);
2757
2758 /**
2759  * scsi_internal_device_unblock - resume a device after a block request
2760  * @sdev:       device to resume
2761  * @new_state:  state to set the device to after unblocking
2762  *
2763  * Restart the device queue for a previously suspended SCSI device. May sleep.
2764  *
2765  * Returns zero if successful or a negative error code upon failure.
2766  *
2767  * Notes:
2768  * This routine transitions the device to the SDEV_RUNNING state or to one of
2769  * the offline states (which must be a legal transition) allowing the midlayer
2770  * to goose the queue for this device.
2771  */
2772 static int scsi_internal_device_unblock(struct scsi_device *sdev,
2773                                         enum scsi_device_state new_state)
2774 {
2775         int ret;
2776
2777         mutex_lock(&sdev->state_mutex);
2778         ret = scsi_internal_device_unblock_nowait(sdev, new_state);
2779         mutex_unlock(&sdev->state_mutex);
2780
2781         return ret;
2782 }
2783
2784 static void
2785 device_block(struct scsi_device *sdev, void *data)
2786 {
2787         scsi_internal_device_block(sdev);
2788 }
2789
2790 static int
2791 target_block(struct device *dev, void *data)
2792 {
2793         if (scsi_is_target_device(dev))
2794                 starget_for_each_device(to_scsi_target(dev), NULL,
2795                                         device_block);
2796         return 0;
2797 }
2798
2799 void
2800 scsi_target_block(struct device *dev)
2801 {
2802         if (scsi_is_target_device(dev))
2803                 starget_for_each_device(to_scsi_target(dev), NULL,
2804                                         device_block);
2805         else
2806                 device_for_each_child(dev, NULL, target_block);
2807 }
2808 EXPORT_SYMBOL_GPL(scsi_target_block);
2809
2810 static void
2811 device_unblock(struct scsi_device *sdev, void *data)
2812 {
2813         scsi_internal_device_unblock(sdev, *(enum scsi_device_state *)data);
2814 }
2815
2816 static int
2817 target_unblock(struct device *dev, void *data)
2818 {
2819         if (scsi_is_target_device(dev))
2820                 starget_for_each_device(to_scsi_target(dev), data,
2821                                         device_unblock);
2822         return 0;
2823 }
2824
2825 void
2826 scsi_target_unblock(struct device *dev, enum scsi_device_state new_state)
2827 {
2828         if (scsi_is_target_device(dev))
2829                 starget_for_each_device(to_scsi_target(dev), &new_state,
2830                                         device_unblock);
2831         else
2832                 device_for_each_child(dev, &new_state, target_unblock);
2833 }
2834 EXPORT_SYMBOL_GPL(scsi_target_unblock);
2835
2836 /**
2837  * scsi_kmap_atomic_sg - find and atomically map an sg-elemnt
2838  * @sgl:        scatter-gather list
2839  * @sg_count:   number of segments in sg
2840  * @offset:     offset in bytes into sg, on return offset into the mapped area
2841  * @len:        bytes to map, on return number of bytes mapped
2842  *
2843  * Returns virtual address of the start of the mapped page
2844  */
2845 void *scsi_kmap_atomic_sg(struct scatterlist *sgl, int sg_count,
2846                           size_t *offset, size_t *len)
2847 {
2848         int i;
2849         size_t sg_len = 0, len_complete = 0;
2850         struct scatterlist *sg;
2851         struct page *page;
2852
2853         WARN_ON(!irqs_disabled());
2854
2855         for_each_sg(sgl, sg, sg_count, i) {
2856                 len_complete = sg_len; /* Complete sg-entries */
2857                 sg_len += sg->length;
2858                 if (sg_len > *offset)
2859                         break;
2860         }
2861
2862         if (unlikely(i == sg_count)) {
2863                 printk(KERN_ERR "%s: Bytes in sg: %zu, requested offset %zu, "
2864                         "elements %d\n",
2865                        __func__, sg_len, *offset, sg_count);
2866                 WARN_ON(1);
2867                 return NULL;
2868         }
2869
2870         /* Offset starting from the beginning of first page in this sg-entry */
2871         *offset = *offset - len_complete + sg->offset;
2872
2873         /* Assumption: contiguous pages can be accessed as "page + i" */
2874         page = nth_page(sg_page(sg), (*offset >> PAGE_SHIFT));
2875         *offset &= ~PAGE_MASK;
2876
2877         /* Bytes in this sg-entry from *offset to the end of the page */
2878         sg_len = PAGE_SIZE - *offset;
2879         if (*len > sg_len)
2880                 *len = sg_len;
2881
2882         return kmap_atomic(page);
2883 }
2884 EXPORT_SYMBOL(scsi_kmap_atomic_sg);
2885
2886 /**
2887  * scsi_kunmap_atomic_sg - atomically unmap a virtual address, previously mapped with scsi_kmap_atomic_sg
2888  * @virt:       virtual address to be unmapped
2889  */
2890 void scsi_kunmap_atomic_sg(void *virt)
2891 {
2892         kunmap_atomic(virt);
2893 }
2894 EXPORT_SYMBOL(scsi_kunmap_atomic_sg);
2895
2896 void sdev_disable_disk_events(struct scsi_device *sdev)
2897 {
2898         atomic_inc(&sdev->disk_events_disable_depth);
2899 }
2900 EXPORT_SYMBOL(sdev_disable_disk_events);
2901
2902 void sdev_enable_disk_events(struct scsi_device *sdev)
2903 {
2904         if (WARN_ON_ONCE(atomic_read(&sdev->disk_events_disable_depth) <= 0))
2905                 return;
2906         atomic_dec(&sdev->disk_events_disable_depth);
2907 }
2908 EXPORT_SYMBOL(sdev_enable_disk_events);
2909
2910 /**
2911  * scsi_vpd_lun_id - return a unique device identification
2912  * @sdev: SCSI device
2913  * @id:   buffer for the identification
2914  * @id_len:  length of the buffer
2915  *
2916  * Copies a unique device identification into @id based
2917  * on the information in the VPD page 0x83 of the device.
2918  * The string will be formatted as a SCSI name string.
2919  *
2920  * Returns the length of the identification or error on failure.
2921  * If the identifier is longer than the supplied buffer the actual
2922  * identifier length is returned and the buffer is not zero-padded.
2923  */
2924 int scsi_vpd_lun_id(struct scsi_device *sdev, char *id, size_t id_len)
2925 {
2926         u8 cur_id_type = 0xff;
2927         u8 cur_id_size = 0;
2928         const unsigned char *d, *cur_id_str;
2929         const struct scsi_vpd *vpd_pg83;
2930         int id_size = -EINVAL;
2931
2932         rcu_read_lock();
2933         vpd_pg83 = rcu_dereference(sdev->vpd_pg83);
2934         if (!vpd_pg83) {
2935                 rcu_read_unlock();
2936                 return -ENXIO;
2937         }
2938
2939         /*
2940          * Look for the correct descriptor.
2941          * Order of preference for lun descriptor:
2942          * - SCSI name string
2943          * - NAA IEEE Registered Extended
2944          * - EUI-64 based 16-byte
2945          * - EUI-64 based 12-byte
2946          * - NAA IEEE Registered
2947          * - NAA IEEE Extended
2948          * - T10 Vendor ID
2949          * as longer descriptors reduce the likelyhood
2950          * of identification clashes.
2951          */
2952
2953         /* The id string must be at least 20 bytes + terminating NULL byte */
2954         if (id_len < 21) {
2955                 rcu_read_unlock();
2956                 return -EINVAL;
2957         }
2958
2959         memset(id, 0, id_len);
2960         d = vpd_pg83->data + 4;
2961         while (d < vpd_pg83->data + vpd_pg83->len) {
2962                 /* Skip designators not referring to the LUN */
2963                 if ((d[1] & 0x30) != 0x00)
2964                         goto next_desig;
2965
2966                 switch (d[1] & 0xf) {
2967                 case 0x1:
2968                         /* T10 Vendor ID */
2969                         if (cur_id_size > d[3])
2970                                 break;
2971                         /* Prefer anything */
2972                         if (cur_id_type > 0x01 && cur_id_type != 0xff)
2973                                 break;
2974                         cur_id_size = d[3];
2975                         if (cur_id_size + 4 > id_len)
2976                                 cur_id_size = id_len - 4;
2977                         cur_id_str = d + 4;
2978                         cur_id_type = d[1] & 0xf;
2979                         id_size = snprintf(id, id_len, "t10.%*pE",
2980                                            cur_id_size, cur_id_str);
2981                         break;
2982                 case 0x2:
2983                         /* EUI-64 */
2984                         if (cur_id_size > d[3])
2985                                 break;
2986                         /* Prefer NAA IEEE Registered Extended */
2987                         if (cur_id_type == 0x3 &&
2988                             cur_id_size == d[3])
2989                                 break;
2990                         cur_id_size = d[3];
2991                         cur_id_str = d + 4;
2992                         cur_id_type = d[1] & 0xf;
2993                         switch (cur_id_size) {
2994                         case 8:
2995                                 id_size = snprintf(id, id_len,
2996                                                    "eui.%8phN",
2997                                                    cur_id_str);
2998                                 break;
2999                         case 12:
3000                                 id_size = snprintf(id, id_len,
3001                                                    "eui.%12phN",
3002                                                    cur_id_str);
3003                                 break;
3004                         case 16:
3005                                 id_size = snprintf(id, id_len,
3006                                                    "eui.%16phN",
3007                                                    cur_id_str);
3008                                 break;
3009                         default:
3010                                 cur_id_size = 0;
3011                                 break;
3012                         }
3013                         break;
3014                 case 0x3:
3015                         /* NAA */
3016                         if (cur_id_size > d[3])
3017                                 break;
3018                         cur_id_size = d[3];
3019                         cur_id_str = d + 4;
3020                         cur_id_type = d[1] & 0xf;
3021                         switch (cur_id_size) {
3022                         case 8:
3023                                 id_size = snprintf(id, id_len,
3024                                                    "naa.%8phN",
3025                                                    cur_id_str);
3026                                 break;
3027                         case 16:
3028                                 id_size = snprintf(id, id_len,
3029                                                    "naa.%16phN",
3030                                                    cur_id_str);
3031                                 break;
3032                         default:
3033                                 cur_id_size = 0;
3034                                 break;
3035                         }
3036                         break;
3037                 case 0x8:
3038                         /* SCSI name string */
3039                         if (cur_id_size + 4 > d[3])
3040                                 break;
3041                         /* Prefer others for truncated descriptor */
3042                         if (cur_id_size && d[3] > id_len)
3043                                 break;
3044                         cur_id_size = id_size = d[3];
3045                         cur_id_str = d + 4;
3046                         cur_id_type = d[1] & 0xf;
3047                         if (cur_id_size >= id_len)
3048                                 cur_id_size = id_len - 1;
3049                         memcpy(id, cur_id_str, cur_id_size);
3050                         /* Decrease priority for truncated descriptor */
3051                         if (cur_id_size != id_size)
3052                                 cur_id_size = 6;
3053                         break;
3054                 default:
3055                         break;
3056                 }
3057 next_desig:
3058                 d += d[3] + 4;
3059         }
3060         rcu_read_unlock();
3061
3062         return id_size;
3063 }
3064 EXPORT_SYMBOL(scsi_vpd_lun_id);
3065
3066 /*
3067  * scsi_vpd_tpg_id - return a target port group identifier
3068  * @sdev: SCSI device
3069  *
3070  * Returns the Target Port Group identifier from the information
3071  * froom VPD page 0x83 of the device.
3072  *
3073  * Returns the identifier or error on failure.
3074  */
3075 int scsi_vpd_tpg_id(struct scsi_device *sdev, int *rel_id)
3076 {
3077         const unsigned char *d;
3078         const struct scsi_vpd *vpd_pg83;
3079         int group_id = -EAGAIN, rel_port = -1;
3080
3081         rcu_read_lock();
3082         vpd_pg83 = rcu_dereference(sdev->vpd_pg83);
3083         if (!vpd_pg83) {
3084                 rcu_read_unlock();
3085                 return -ENXIO;
3086         }
3087
3088         d = vpd_pg83->data + 4;
3089         while (d < vpd_pg83->data + vpd_pg83->len) {
3090                 switch (d[1] & 0xf) {
3091                 case 0x4:
3092                         /* Relative target port */
3093                         rel_port = get_unaligned_be16(&d[6]);
3094                         break;
3095                 case 0x5:
3096                         /* Target port group */
3097                         group_id = get_unaligned_be16(&d[6]);
3098                         break;
3099                 default:
3100                         break;
3101                 }
3102                 d += d[3] + 4;
3103         }
3104         rcu_read_unlock();
3105
3106         if (group_id >= 0 && rel_id && rel_port != -1)
3107                 *rel_id = rel_port;
3108
3109         return group_id;
3110 }
3111 EXPORT_SYMBOL(scsi_vpd_tpg_id);