s390/Kconfig: sort config S390 select list once again
[linux-2.6-microblaze.git] / drivers / scsi / scsi_lib.c
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 1999 Eric Youngdale
4  * Copyright (C) 2014 Christoph Hellwig
5  *
6  *  SCSI queueing library.
7  *      Initial versions: Eric Youngdale (eric@andante.org).
8  *                        Based upon conversations with large numbers
9  *                        of people at Linux Expo.
10  */
11
12 #include <linux/bio.h>
13 #include <linux/bitops.h>
14 #include <linux/blkdev.h>
15 #include <linux/completion.h>
16 #include <linux/kernel.h>
17 #include <linux/export.h>
18 #include <linux/init.h>
19 #include <linux/pci.h>
20 #include <linux/delay.h>
21 #include <linux/hardirq.h>
22 #include <linux/scatterlist.h>
23 #include <linux/blk-mq.h>
24 #include <linux/ratelimit.h>
25 #include <asm/unaligned.h>
26
27 #include <scsi/scsi.h>
28 #include <scsi/scsi_cmnd.h>
29 #include <scsi/scsi_dbg.h>
30 #include <scsi/scsi_device.h>
31 #include <scsi/scsi_driver.h>
32 #include <scsi/scsi_eh.h>
33 #include <scsi/scsi_host.h>
34 #include <scsi/scsi_transport.h> /* __scsi_init_queue() */
35 #include <scsi/scsi_dh.h>
36
37 #include <trace/events/scsi.h>
38
39 #include "scsi_debugfs.h"
40 #include "scsi_priv.h"
41 #include "scsi_logging.h"
42
43 /*
44  * Size of integrity metadata is usually small, 1 inline sg should
45  * cover normal cases.
46  */
47 #ifdef CONFIG_ARCH_NO_SG_CHAIN
48 #define  SCSI_INLINE_PROT_SG_CNT  0
49 #define  SCSI_INLINE_SG_CNT  0
50 #else
51 #define  SCSI_INLINE_PROT_SG_CNT  1
52 #define  SCSI_INLINE_SG_CNT  2
53 #endif
54
55 static struct kmem_cache *scsi_sense_cache;
56 static struct kmem_cache *scsi_sense_isadma_cache;
57 static DEFINE_MUTEX(scsi_sense_cache_mutex);
58
59 static void scsi_mq_uninit_cmd(struct scsi_cmnd *cmd);
60
61 static inline struct kmem_cache *
62 scsi_select_sense_cache(bool unchecked_isa_dma)
63 {
64         return unchecked_isa_dma ? scsi_sense_isadma_cache : scsi_sense_cache;
65 }
66
67 static void scsi_free_sense_buffer(bool unchecked_isa_dma,
68                                    unsigned char *sense_buffer)
69 {
70         kmem_cache_free(scsi_select_sense_cache(unchecked_isa_dma),
71                         sense_buffer);
72 }
73
74 static unsigned char *scsi_alloc_sense_buffer(bool unchecked_isa_dma,
75         gfp_t gfp_mask, int numa_node)
76 {
77         return kmem_cache_alloc_node(scsi_select_sense_cache(unchecked_isa_dma),
78                                      gfp_mask, numa_node);
79 }
80
81 int scsi_init_sense_cache(struct Scsi_Host *shost)
82 {
83         struct kmem_cache *cache;
84         int ret = 0;
85
86         mutex_lock(&scsi_sense_cache_mutex);
87         cache = scsi_select_sense_cache(shost->unchecked_isa_dma);
88         if (cache)
89                 goto exit;
90
91         if (shost->unchecked_isa_dma) {
92                 scsi_sense_isadma_cache =
93                         kmem_cache_create("scsi_sense_cache(DMA)",
94                                 SCSI_SENSE_BUFFERSIZE, 0,
95                                 SLAB_HWCACHE_ALIGN | SLAB_CACHE_DMA, NULL);
96                 if (!scsi_sense_isadma_cache)
97                         ret = -ENOMEM;
98         } else {
99                 scsi_sense_cache =
100                         kmem_cache_create_usercopy("scsi_sense_cache",
101                                 SCSI_SENSE_BUFFERSIZE, 0, SLAB_HWCACHE_ALIGN,
102                                 0, SCSI_SENSE_BUFFERSIZE, NULL);
103                 if (!scsi_sense_cache)
104                         ret = -ENOMEM;
105         }
106  exit:
107         mutex_unlock(&scsi_sense_cache_mutex);
108         return ret;
109 }
110
111 /*
112  * When to reinvoke queueing after a resource shortage. It's 3 msecs to
113  * not change behaviour from the previous unplug mechanism, experimentation
114  * may prove this needs changing.
115  */
116 #define SCSI_QUEUE_DELAY        3
117
118 static void
119 scsi_set_blocked(struct scsi_cmnd *cmd, int reason)
120 {
121         struct Scsi_Host *host = cmd->device->host;
122         struct scsi_device *device = cmd->device;
123         struct scsi_target *starget = scsi_target(device);
124
125         /*
126          * Set the appropriate busy bit for the device/host.
127          *
128          * If the host/device isn't busy, assume that something actually
129          * completed, and that we should be able to queue a command now.
130          *
131          * Note that the prior mid-layer assumption that any host could
132          * always queue at least one command is now broken.  The mid-layer
133          * will implement a user specifiable stall (see
134          * scsi_host.max_host_blocked and scsi_device.max_device_blocked)
135          * if a command is requeued with no other commands outstanding
136          * either for the device or for the host.
137          */
138         switch (reason) {
139         case SCSI_MLQUEUE_HOST_BUSY:
140                 atomic_set(&host->host_blocked, host->max_host_blocked);
141                 break;
142         case SCSI_MLQUEUE_DEVICE_BUSY:
143         case SCSI_MLQUEUE_EH_RETRY:
144                 atomic_set(&device->device_blocked,
145                            device->max_device_blocked);
146                 break;
147         case SCSI_MLQUEUE_TARGET_BUSY:
148                 atomic_set(&starget->target_blocked,
149                            starget->max_target_blocked);
150                 break;
151         }
152 }
153
154 static void scsi_mq_requeue_cmd(struct scsi_cmnd *cmd)
155 {
156         if (cmd->request->rq_flags & RQF_DONTPREP) {
157                 cmd->request->rq_flags &= ~RQF_DONTPREP;
158                 scsi_mq_uninit_cmd(cmd);
159         } else {
160                 WARN_ON_ONCE(true);
161         }
162         blk_mq_requeue_request(cmd->request, true);
163 }
164
165 /**
166  * __scsi_queue_insert - private queue insertion
167  * @cmd: The SCSI command being requeued
168  * @reason:  The reason for the requeue
169  * @unbusy: Whether the queue should be unbusied
170  *
171  * This is a private queue insertion.  The public interface
172  * scsi_queue_insert() always assumes the queue should be unbusied
173  * because it's always called before the completion.  This function is
174  * for a requeue after completion, which should only occur in this
175  * file.
176  */
177 static void __scsi_queue_insert(struct scsi_cmnd *cmd, int reason, bool unbusy)
178 {
179         struct scsi_device *device = cmd->device;
180
181         SCSI_LOG_MLQUEUE(1, scmd_printk(KERN_INFO, cmd,
182                 "Inserting command %p into mlqueue\n", cmd));
183
184         scsi_set_blocked(cmd, reason);
185
186         /*
187          * Decrement the counters, since these commands are no longer
188          * active on the host/device.
189          */
190         if (unbusy)
191                 scsi_device_unbusy(device, cmd);
192
193         /*
194          * Requeue this command.  It will go before all other commands
195          * that are already in the queue. Schedule requeue work under
196          * lock such that the kblockd_schedule_work() call happens
197          * before blk_cleanup_queue() finishes.
198          */
199         cmd->result = 0;
200
201         blk_mq_requeue_request(cmd->request, true);
202 }
203
204 /**
205  * scsi_queue_insert - Reinsert a command in the queue.
206  * @cmd:    command that we are adding to queue.
207  * @reason: why we are inserting command to queue.
208  *
209  * We do this for one of two cases. Either the host is busy and it cannot accept
210  * any more commands for the time being, or the device returned QUEUE_FULL and
211  * can accept no more commands.
212  *
213  * Context: This could be called either from an interrupt context or a normal
214  * process context.
215  */
216 void scsi_queue_insert(struct scsi_cmnd *cmd, int reason)
217 {
218         __scsi_queue_insert(cmd, reason, true);
219 }
220
221
222 /**
223  * __scsi_execute - insert request and wait for the result
224  * @sdev:       scsi device
225  * @cmd:        scsi command
226  * @data_direction: data direction
227  * @buffer:     data buffer
228  * @bufflen:    len of buffer
229  * @sense:      optional sense buffer
230  * @sshdr:      optional decoded sense header
231  * @timeout:    request timeout in seconds
232  * @retries:    number of times to retry request
233  * @flags:      flags for ->cmd_flags
234  * @rq_flags:   flags for ->rq_flags
235  * @resid:      optional residual length
236  *
237  * Returns the scsi_cmnd result field if a command was executed, or a negative
238  * Linux error code if we didn't get that far.
239  */
240 int __scsi_execute(struct scsi_device *sdev, const unsigned char *cmd,
241                  int data_direction, void *buffer, unsigned bufflen,
242                  unsigned char *sense, struct scsi_sense_hdr *sshdr,
243                  int timeout, int retries, u64 flags, req_flags_t rq_flags,
244                  int *resid)
245 {
246         struct request *req;
247         struct scsi_request *rq;
248         int ret = DRIVER_ERROR << 24;
249
250         req = blk_get_request(sdev->request_queue,
251                         data_direction == DMA_TO_DEVICE ?
252                         REQ_OP_SCSI_OUT : REQ_OP_SCSI_IN, BLK_MQ_REQ_PREEMPT);
253         if (IS_ERR(req))
254                 return ret;
255         rq = scsi_req(req);
256
257         if (bufflen &&  blk_rq_map_kern(sdev->request_queue, req,
258                                         buffer, bufflen, GFP_NOIO))
259                 goto out;
260
261         rq->cmd_len = COMMAND_SIZE(cmd[0]);
262         memcpy(rq->cmd, cmd, rq->cmd_len);
263         rq->retries = retries;
264         req->timeout = timeout;
265         req->cmd_flags |= flags;
266         req->rq_flags |= rq_flags | RQF_QUIET;
267
268         /*
269          * head injection *required* here otherwise quiesce won't work
270          */
271         blk_execute_rq(req->q, NULL, req, 1);
272
273         /*
274          * Some devices (USB mass-storage in particular) may transfer
275          * garbage data together with a residue indicating that the data
276          * is invalid.  Prevent the garbage from being misinterpreted
277          * and prevent security leaks by zeroing out the excess data.
278          */
279         if (unlikely(rq->resid_len > 0 && rq->resid_len <= bufflen))
280                 memset(buffer + (bufflen - rq->resid_len), 0, rq->resid_len);
281
282         if (resid)
283                 *resid = rq->resid_len;
284         if (sense && rq->sense_len)
285                 memcpy(sense, rq->sense, SCSI_SENSE_BUFFERSIZE);
286         if (sshdr)
287                 scsi_normalize_sense(rq->sense, rq->sense_len, sshdr);
288         ret = rq->result;
289  out:
290         blk_put_request(req);
291
292         return ret;
293 }
294 EXPORT_SYMBOL(__scsi_execute);
295
296 /*
297  * Wake up the error handler if necessary. Avoid as follows that the error
298  * handler is not woken up if host in-flight requests number ==
299  * shost->host_failed: use call_rcu() in scsi_eh_scmd_add() in combination
300  * with an RCU read lock in this function to ensure that this function in
301  * its entirety either finishes before scsi_eh_scmd_add() increases the
302  * host_failed counter or that it notices the shost state change made by
303  * scsi_eh_scmd_add().
304  */
305 static void scsi_dec_host_busy(struct Scsi_Host *shost, struct scsi_cmnd *cmd)
306 {
307         unsigned long flags;
308
309         rcu_read_lock();
310         __clear_bit(SCMD_STATE_INFLIGHT, &cmd->state);
311         if (unlikely(scsi_host_in_recovery(shost))) {
312                 spin_lock_irqsave(shost->host_lock, flags);
313                 if (shost->host_failed || shost->host_eh_scheduled)
314                         scsi_eh_wakeup(shost);
315                 spin_unlock_irqrestore(shost->host_lock, flags);
316         }
317         rcu_read_unlock();
318 }
319
320 void scsi_device_unbusy(struct scsi_device *sdev, struct scsi_cmnd *cmd)
321 {
322         struct Scsi_Host *shost = sdev->host;
323         struct scsi_target *starget = scsi_target(sdev);
324
325         scsi_dec_host_busy(shost, cmd);
326
327         if (starget->can_queue > 0)
328                 atomic_dec(&starget->target_busy);
329
330         atomic_dec(&sdev->device_busy);
331 }
332
333 static void scsi_kick_queue(struct request_queue *q)
334 {
335         blk_mq_run_hw_queues(q, false);
336 }
337
338 /*
339  * Called for single_lun devices on IO completion. Clear starget_sdev_user,
340  * and call blk_run_queue for all the scsi_devices on the target -
341  * including current_sdev first.
342  *
343  * Called with *no* scsi locks held.
344  */
345 static void scsi_single_lun_run(struct scsi_device *current_sdev)
346 {
347         struct Scsi_Host *shost = current_sdev->host;
348         struct scsi_device *sdev, *tmp;
349         struct scsi_target *starget = scsi_target(current_sdev);
350         unsigned long flags;
351
352         spin_lock_irqsave(shost->host_lock, flags);
353         starget->starget_sdev_user = NULL;
354         spin_unlock_irqrestore(shost->host_lock, flags);
355
356         /*
357          * Call blk_run_queue for all LUNs on the target, starting with
358          * current_sdev. We race with others (to set starget_sdev_user),
359          * but in most cases, we will be first. Ideally, each LU on the
360          * target would get some limited time or requests on the target.
361          */
362         scsi_kick_queue(current_sdev->request_queue);
363
364         spin_lock_irqsave(shost->host_lock, flags);
365         if (starget->starget_sdev_user)
366                 goto out;
367         list_for_each_entry_safe(sdev, tmp, &starget->devices,
368                         same_target_siblings) {
369                 if (sdev == current_sdev)
370                         continue;
371                 if (scsi_device_get(sdev))
372                         continue;
373
374                 spin_unlock_irqrestore(shost->host_lock, flags);
375                 scsi_kick_queue(sdev->request_queue);
376                 spin_lock_irqsave(shost->host_lock, flags);
377
378                 scsi_device_put(sdev);
379         }
380  out:
381         spin_unlock_irqrestore(shost->host_lock, flags);
382 }
383
384 static inline bool scsi_device_is_busy(struct scsi_device *sdev)
385 {
386         if (atomic_read(&sdev->device_busy) >= sdev->queue_depth)
387                 return true;
388         if (atomic_read(&sdev->device_blocked) > 0)
389                 return true;
390         return false;
391 }
392
393 static inline bool scsi_target_is_busy(struct scsi_target *starget)
394 {
395         if (starget->can_queue > 0) {
396                 if (atomic_read(&starget->target_busy) >= starget->can_queue)
397                         return true;
398                 if (atomic_read(&starget->target_blocked) > 0)
399                         return true;
400         }
401         return false;
402 }
403
404 static inline bool scsi_host_is_busy(struct Scsi_Host *shost)
405 {
406         if (atomic_read(&shost->host_blocked) > 0)
407                 return true;
408         if (shost->host_self_blocked)
409                 return true;
410         return false;
411 }
412
413 static void scsi_starved_list_run(struct Scsi_Host *shost)
414 {
415         LIST_HEAD(starved_list);
416         struct scsi_device *sdev;
417         unsigned long flags;
418
419         spin_lock_irqsave(shost->host_lock, flags);
420         list_splice_init(&shost->starved_list, &starved_list);
421
422         while (!list_empty(&starved_list)) {
423                 struct request_queue *slq;
424
425                 /*
426                  * As long as shost is accepting commands and we have
427                  * starved queues, call blk_run_queue. scsi_request_fn
428                  * drops the queue_lock and can add us back to the
429                  * starved_list.
430                  *
431                  * host_lock protects the starved_list and starved_entry.
432                  * scsi_request_fn must get the host_lock before checking
433                  * or modifying starved_list or starved_entry.
434                  */
435                 if (scsi_host_is_busy(shost))
436                         break;
437
438                 sdev = list_entry(starved_list.next,
439                                   struct scsi_device, starved_entry);
440                 list_del_init(&sdev->starved_entry);
441                 if (scsi_target_is_busy(scsi_target(sdev))) {
442                         list_move_tail(&sdev->starved_entry,
443                                        &shost->starved_list);
444                         continue;
445                 }
446
447                 /*
448                  * Once we drop the host lock, a racing scsi_remove_device()
449                  * call may remove the sdev from the starved list and destroy
450                  * it and the queue.  Mitigate by taking a reference to the
451                  * queue and never touching the sdev again after we drop the
452                  * host lock.  Note: if __scsi_remove_device() invokes
453                  * blk_cleanup_queue() before the queue is run from this
454                  * function then blk_run_queue() will return immediately since
455                  * blk_cleanup_queue() marks the queue with QUEUE_FLAG_DYING.
456                  */
457                 slq = sdev->request_queue;
458                 if (!blk_get_queue(slq))
459                         continue;
460                 spin_unlock_irqrestore(shost->host_lock, flags);
461
462                 scsi_kick_queue(slq);
463                 blk_put_queue(slq);
464
465                 spin_lock_irqsave(shost->host_lock, flags);
466         }
467         /* put any unprocessed entries back */
468         list_splice(&starved_list, &shost->starved_list);
469         spin_unlock_irqrestore(shost->host_lock, flags);
470 }
471
472 /**
473  * scsi_run_queue - Select a proper request queue to serve next.
474  * @q:  last request's queue
475  *
476  * The previous command was completely finished, start a new one if possible.
477  */
478 static void scsi_run_queue(struct request_queue *q)
479 {
480         struct scsi_device *sdev = q->queuedata;
481
482         if (scsi_target(sdev)->single_lun)
483                 scsi_single_lun_run(sdev);
484         if (!list_empty(&sdev->host->starved_list))
485                 scsi_starved_list_run(sdev->host);
486
487         blk_mq_run_hw_queues(q, false);
488 }
489
490 void scsi_requeue_run_queue(struct work_struct *work)
491 {
492         struct scsi_device *sdev;
493         struct request_queue *q;
494
495         sdev = container_of(work, struct scsi_device, requeue_work);
496         q = sdev->request_queue;
497         scsi_run_queue(q);
498 }
499
500 void scsi_run_host_queues(struct Scsi_Host *shost)
501 {
502         struct scsi_device *sdev;
503
504         shost_for_each_device(sdev, shost)
505                 scsi_run_queue(sdev->request_queue);
506 }
507
508 static void scsi_uninit_cmd(struct scsi_cmnd *cmd)
509 {
510         if (!blk_rq_is_passthrough(cmd->request)) {
511                 struct scsi_driver *drv = scsi_cmd_to_driver(cmd);
512
513                 if (drv->uninit_command)
514                         drv->uninit_command(cmd);
515         }
516 }
517
518 void scsi_free_sgtables(struct scsi_cmnd *cmd)
519 {
520         if (cmd->sdb.table.nents)
521                 sg_free_table_chained(&cmd->sdb.table,
522                                 SCSI_INLINE_SG_CNT);
523         if (scsi_prot_sg_count(cmd))
524                 sg_free_table_chained(&cmd->prot_sdb->table,
525                                 SCSI_INLINE_PROT_SG_CNT);
526 }
527 EXPORT_SYMBOL_GPL(scsi_free_sgtables);
528
529 static void scsi_mq_uninit_cmd(struct scsi_cmnd *cmd)
530 {
531         scsi_free_sgtables(cmd);
532         scsi_uninit_cmd(cmd);
533 }
534
535 static void scsi_run_queue_async(struct scsi_device *sdev)
536 {
537         if (scsi_target(sdev)->single_lun ||
538             !list_empty(&sdev->host->starved_list)) {
539                 kblockd_schedule_work(&sdev->requeue_work);
540         } else {
541                 /*
542                  * smp_mb() present in sbitmap_queue_clear() or implied in
543                  * .end_io is for ordering writing .device_busy in
544                  * scsi_device_unbusy() and reading sdev->restarts.
545                  */
546                 int old = atomic_read(&sdev->restarts);
547
548                 /*
549                  * ->restarts has to be kept as non-zero if new budget
550                  *  contention occurs.
551                  *
552                  *  No need to run queue when either another re-run
553                  *  queue wins in updating ->restarts or a new budget
554                  *  contention occurs.
555                  */
556                 if (old && atomic_cmpxchg(&sdev->restarts, old, 0) == old)
557                         blk_mq_run_hw_queues(sdev->request_queue, true);
558         }
559 }
560
561 /* Returns false when no more bytes to process, true if there are more */
562 static bool scsi_end_request(struct request *req, blk_status_t error,
563                 unsigned int bytes)
564 {
565         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
566         struct scsi_device *sdev = cmd->device;
567         struct request_queue *q = sdev->request_queue;
568
569         if (blk_update_request(req, error, bytes))
570                 return true;
571
572         if (blk_queue_add_random(q))
573                 add_disk_randomness(req->rq_disk);
574
575         if (!blk_rq_is_scsi(req)) {
576                 WARN_ON_ONCE(!(cmd->flags & SCMD_INITIALIZED));
577                 cmd->flags &= ~SCMD_INITIALIZED;
578         }
579
580         /*
581          * Calling rcu_barrier() is not necessary here because the
582          * SCSI error handler guarantees that the function called by
583          * call_rcu() has been called before scsi_end_request() is
584          * called.
585          */
586         destroy_rcu_head(&cmd->rcu);
587
588         /*
589          * In the MQ case the command gets freed by __blk_mq_end_request,
590          * so we have to do all cleanup that depends on it earlier.
591          *
592          * We also can't kick the queues from irq context, so we
593          * will have to defer it to a workqueue.
594          */
595         scsi_mq_uninit_cmd(cmd);
596
597         /*
598          * queue is still alive, so grab the ref for preventing it
599          * from being cleaned up during running queue.
600          */
601         percpu_ref_get(&q->q_usage_counter);
602
603         __blk_mq_end_request(req, error);
604
605         scsi_run_queue_async(sdev);
606
607         percpu_ref_put(&q->q_usage_counter);
608         return false;
609 }
610
611 /**
612  * scsi_result_to_blk_status - translate a SCSI result code into blk_status_t
613  * @cmd:        SCSI command
614  * @result:     scsi error code
615  *
616  * Translate a SCSI result code into a blk_status_t value. May reset the host
617  * byte of @cmd->result.
618  */
619 static blk_status_t scsi_result_to_blk_status(struct scsi_cmnd *cmd, int result)
620 {
621         switch (host_byte(result)) {
622         case DID_OK:
623                 /*
624                  * Also check the other bytes than the status byte in result
625                  * to handle the case when a SCSI LLD sets result to
626                  * DRIVER_SENSE << 24 without setting SAM_STAT_CHECK_CONDITION.
627                  */
628                 if (scsi_status_is_good(result) && (result & ~0xff) == 0)
629                         return BLK_STS_OK;
630                 return BLK_STS_IOERR;
631         case DID_TRANSPORT_FAILFAST:
632                 return BLK_STS_TRANSPORT;
633         case DID_TARGET_FAILURE:
634                 set_host_byte(cmd, DID_OK);
635                 return BLK_STS_TARGET;
636         case DID_NEXUS_FAILURE:
637                 set_host_byte(cmd, DID_OK);
638                 return BLK_STS_NEXUS;
639         case DID_ALLOC_FAILURE:
640                 set_host_byte(cmd, DID_OK);
641                 return BLK_STS_NOSPC;
642         case DID_MEDIUM_ERROR:
643                 set_host_byte(cmd, DID_OK);
644                 return BLK_STS_MEDIUM;
645         default:
646                 return BLK_STS_IOERR;
647         }
648 }
649
650 /* Helper for scsi_io_completion() when "reprep" action required. */
651 static void scsi_io_completion_reprep(struct scsi_cmnd *cmd,
652                                       struct request_queue *q)
653 {
654         /* A new command will be prepared and issued. */
655         scsi_mq_requeue_cmd(cmd);
656 }
657
658 static bool scsi_cmd_runtime_exceeced(struct scsi_cmnd *cmd)
659 {
660         struct request *req = cmd->request;
661         unsigned long wait_for;
662
663         if (cmd->allowed == SCSI_CMD_RETRIES_NO_LIMIT)
664                 return false;
665
666         wait_for = (cmd->allowed + 1) * req->timeout;
667         if (time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) {
668                 scmd_printk(KERN_ERR, cmd, "timing out command, waited %lus\n",
669                             wait_for/HZ);
670                 return true;
671         }
672         return false;
673 }
674
675 /* Helper for scsi_io_completion() when special action required. */
676 static void scsi_io_completion_action(struct scsi_cmnd *cmd, int result)
677 {
678         struct request_queue *q = cmd->device->request_queue;
679         struct request *req = cmd->request;
680         int level = 0;
681         enum {ACTION_FAIL, ACTION_REPREP, ACTION_RETRY,
682               ACTION_DELAYED_RETRY} action;
683         struct scsi_sense_hdr sshdr;
684         bool sense_valid;
685         bool sense_current = true;      /* false implies "deferred sense" */
686         blk_status_t blk_stat;
687
688         sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
689         if (sense_valid)
690                 sense_current = !scsi_sense_is_deferred(&sshdr);
691
692         blk_stat = scsi_result_to_blk_status(cmd, result);
693
694         if (host_byte(result) == DID_RESET) {
695                 /* Third party bus reset or reset for error recovery
696                  * reasons.  Just retry the command and see what
697                  * happens.
698                  */
699                 action = ACTION_RETRY;
700         } else if (sense_valid && sense_current) {
701                 switch (sshdr.sense_key) {
702                 case UNIT_ATTENTION:
703                         if (cmd->device->removable) {
704                                 /* Detected disc change.  Set a bit
705                                  * and quietly refuse further access.
706                                  */
707                                 cmd->device->changed = 1;
708                                 action = ACTION_FAIL;
709                         } else {
710                                 /* Must have been a power glitch, or a
711                                  * bus reset.  Could not have been a
712                                  * media change, so we just retry the
713                                  * command and see what happens.
714                                  */
715                                 action = ACTION_RETRY;
716                         }
717                         break;
718                 case ILLEGAL_REQUEST:
719                         /* If we had an ILLEGAL REQUEST returned, then
720                          * we may have performed an unsupported
721                          * command.  The only thing this should be
722                          * would be a ten byte read where only a six
723                          * byte read was supported.  Also, on a system
724                          * where READ CAPACITY failed, we may have
725                          * read past the end of the disk.
726                          */
727                         if ((cmd->device->use_10_for_rw &&
728                             sshdr.asc == 0x20 && sshdr.ascq == 0x00) &&
729                             (cmd->cmnd[0] == READ_10 ||
730                              cmd->cmnd[0] == WRITE_10)) {
731                                 /* This will issue a new 6-byte command. */
732                                 cmd->device->use_10_for_rw = 0;
733                                 action = ACTION_REPREP;
734                         } else if (sshdr.asc == 0x10) /* DIX */ {
735                                 action = ACTION_FAIL;
736                                 blk_stat = BLK_STS_PROTECTION;
737                         /* INVALID COMMAND OPCODE or INVALID FIELD IN CDB */
738                         } else if (sshdr.asc == 0x20 || sshdr.asc == 0x24) {
739                                 action = ACTION_FAIL;
740                                 blk_stat = BLK_STS_TARGET;
741                         } else
742                                 action = ACTION_FAIL;
743                         break;
744                 case ABORTED_COMMAND:
745                         action = ACTION_FAIL;
746                         if (sshdr.asc == 0x10) /* DIF */
747                                 blk_stat = BLK_STS_PROTECTION;
748                         break;
749                 case NOT_READY:
750                         /* If the device is in the process of becoming
751                          * ready, or has a temporary blockage, retry.
752                          */
753                         if (sshdr.asc == 0x04) {
754                                 switch (sshdr.ascq) {
755                                 case 0x01: /* becoming ready */
756                                 case 0x04: /* format in progress */
757                                 case 0x05: /* rebuild in progress */
758                                 case 0x06: /* recalculation in progress */
759                                 case 0x07: /* operation in progress */
760                                 case 0x08: /* Long write in progress */
761                                 case 0x09: /* self test in progress */
762                                 case 0x14: /* space allocation in progress */
763                                 case 0x1a: /* start stop unit in progress */
764                                 case 0x1b: /* sanitize in progress */
765                                 case 0x1d: /* configuration in progress */
766                                 case 0x24: /* depopulation in progress */
767                                         action = ACTION_DELAYED_RETRY;
768                                         break;
769                                 case 0x0a: /* ALUA state transition */
770                                         blk_stat = BLK_STS_AGAIN;
771                                         fallthrough;
772                                 default:
773                                         action = ACTION_FAIL;
774                                         break;
775                                 }
776                         } else
777                                 action = ACTION_FAIL;
778                         break;
779                 case VOLUME_OVERFLOW:
780                         /* See SSC3rXX or current. */
781                         action = ACTION_FAIL;
782                         break;
783                 case DATA_PROTECT:
784                         action = ACTION_FAIL;
785                         if ((sshdr.asc == 0x0C && sshdr.ascq == 0x12) ||
786                             (sshdr.asc == 0x55 &&
787                              (sshdr.ascq == 0x0E || sshdr.ascq == 0x0F))) {
788                                 /* Insufficient zone resources */
789                                 blk_stat = BLK_STS_ZONE_OPEN_RESOURCE;
790                         }
791                         break;
792                 default:
793                         action = ACTION_FAIL;
794                         break;
795                 }
796         } else
797                 action = ACTION_FAIL;
798
799         if (action != ACTION_FAIL && scsi_cmd_runtime_exceeced(cmd))
800                 action = ACTION_FAIL;
801
802         switch (action) {
803         case ACTION_FAIL:
804                 /* Give up and fail the remainder of the request */
805                 if (!(req->rq_flags & RQF_QUIET)) {
806                         static DEFINE_RATELIMIT_STATE(_rs,
807                                         DEFAULT_RATELIMIT_INTERVAL,
808                                         DEFAULT_RATELIMIT_BURST);
809
810                         if (unlikely(scsi_logging_level))
811                                 level =
812                                      SCSI_LOG_LEVEL(SCSI_LOG_MLCOMPLETE_SHIFT,
813                                                     SCSI_LOG_MLCOMPLETE_BITS);
814
815                         /*
816                          * if logging is enabled the failure will be printed
817                          * in scsi_log_completion(), so avoid duplicate messages
818                          */
819                         if (!level && __ratelimit(&_rs)) {
820                                 scsi_print_result(cmd, NULL, FAILED);
821                                 if (driver_byte(result) == DRIVER_SENSE)
822                                         scsi_print_sense(cmd);
823                                 scsi_print_command(cmd);
824                         }
825                 }
826                 if (!scsi_end_request(req, blk_stat, blk_rq_err_bytes(req)))
827                         return;
828                 fallthrough;
829         case ACTION_REPREP:
830                 scsi_io_completion_reprep(cmd, q);
831                 break;
832         case ACTION_RETRY:
833                 /* Retry the same command immediately */
834                 __scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY, false);
835                 break;
836         case ACTION_DELAYED_RETRY:
837                 /* Retry the same command after a delay */
838                 __scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY, false);
839                 break;
840         }
841 }
842
843 /*
844  * Helper for scsi_io_completion() when cmd->result is non-zero. Returns a
845  * new result that may suppress further error checking. Also modifies
846  * *blk_statp in some cases.
847  */
848 static int scsi_io_completion_nz_result(struct scsi_cmnd *cmd, int result,
849                                         blk_status_t *blk_statp)
850 {
851         bool sense_valid;
852         bool sense_current = true;      /* false implies "deferred sense" */
853         struct request *req = cmd->request;
854         struct scsi_sense_hdr sshdr;
855
856         sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
857         if (sense_valid)
858                 sense_current = !scsi_sense_is_deferred(&sshdr);
859
860         if (blk_rq_is_passthrough(req)) {
861                 if (sense_valid) {
862                         /*
863                          * SG_IO wants current and deferred errors
864                          */
865                         scsi_req(req)->sense_len =
866                                 min(8 + cmd->sense_buffer[7],
867                                     SCSI_SENSE_BUFFERSIZE);
868                 }
869                 if (sense_current)
870                         *blk_statp = scsi_result_to_blk_status(cmd, result);
871         } else if (blk_rq_bytes(req) == 0 && sense_current) {
872                 /*
873                  * Flush commands do not transfers any data, and thus cannot use
874                  * good_bytes != blk_rq_bytes(req) as the signal for an error.
875                  * This sets *blk_statp explicitly for the problem case.
876                  */
877                 *blk_statp = scsi_result_to_blk_status(cmd, result);
878         }
879         /*
880          * Recovered errors need reporting, but they're always treated as
881          * success, so fiddle the result code here.  For passthrough requests
882          * we already took a copy of the original into sreq->result which
883          * is what gets returned to the user
884          */
885         if (sense_valid && (sshdr.sense_key == RECOVERED_ERROR)) {
886                 bool do_print = true;
887                 /*
888                  * if ATA PASS-THROUGH INFORMATION AVAILABLE [0x0, 0x1d]
889                  * skip print since caller wants ATA registers. Only occurs
890                  * on SCSI ATA PASS_THROUGH commands when CK_COND=1
891                  */
892                 if ((sshdr.asc == 0x0) && (sshdr.ascq == 0x1d))
893                         do_print = false;
894                 else if (req->rq_flags & RQF_QUIET)
895                         do_print = false;
896                 if (do_print)
897                         scsi_print_sense(cmd);
898                 result = 0;
899                 /* for passthrough, *blk_statp may be set */
900                 *blk_statp = BLK_STS_OK;
901         }
902         /*
903          * Another corner case: the SCSI status byte is non-zero but 'good'.
904          * Example: PRE-FETCH command returns SAM_STAT_CONDITION_MET when
905          * it is able to fit nominated LBs in its cache (and SAM_STAT_GOOD
906          * if it can't fit). Treat SAM_STAT_CONDITION_MET and the related
907          * intermediate statuses (both obsolete in SAM-4) as good.
908          */
909         if (status_byte(result) && scsi_status_is_good(result)) {
910                 result = 0;
911                 *blk_statp = BLK_STS_OK;
912         }
913         return result;
914 }
915
916 /**
917  * scsi_io_completion - Completion processing for SCSI commands.
918  * @cmd:        command that is finished.
919  * @good_bytes: number of processed bytes.
920  *
921  * We will finish off the specified number of sectors. If we are done, the
922  * command block will be released and the queue function will be goosed. If we
923  * are not done then we have to figure out what to do next:
924  *
925  *   a) We can call scsi_io_completion_reprep().  The request will be
926  *      unprepared and put back on the queue.  Then a new command will
927  *      be created for it.  This should be used if we made forward
928  *      progress, or if we want to switch from READ(10) to READ(6) for
929  *      example.
930  *
931  *   b) We can call scsi_io_completion_action().  The request will be
932  *      put back on the queue and retried using the same command as
933  *      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         }
954
955         /*
956          * Next deal with any sectors which we were able to correctly
957          * handle.
958          */
959         SCSI_LOG_HLCOMPLETE(1, scmd_printk(KERN_INFO, cmd,
960                 "%u sectors total, %d bytes done.\n",
961                 blk_rq_sectors(req), good_bytes));
962
963         /*
964          * Failed, zero length commands always need to drop down
965          * to retry code. Fast path should return in this block.
966          */
967         if (likely(blk_rq_bytes(req) > 0 || blk_stat == BLK_STS_OK)) {
968                 if (likely(!scsi_end_request(req, blk_stat, good_bytes)))
969                         return; /* no bytes remaining */
970         }
971
972         /* Kill remainder if no retries. */
973         if (unlikely(blk_stat && scsi_noretry_cmd(cmd))) {
974                 if (scsi_end_request(req, blk_stat, blk_rq_bytes(req)))
975                         WARN_ONCE(true,
976                             "Bytes remaining after failed, no-retry command");
977                 return;
978         }
979
980         /*
981          * If there had been no error, but we have leftover bytes in the
982          * requeues just queue the command up again.
983          */
984         if (likely(result == 0))
985                 scsi_io_completion_reprep(cmd, q);
986         else
987                 scsi_io_completion_action(cmd, result);
988 }
989
990 static inline bool scsi_cmd_needs_dma_drain(struct scsi_device *sdev,
991                 struct request *rq)
992 {
993         return sdev->dma_drain_len && blk_rq_is_passthrough(rq) &&
994                !op_is_write(req_op(rq)) &&
995                sdev->host->hostt->dma_need_drain(rq);
996 }
997
998 /**
999  * scsi_alloc_sgtables - allocate S/G tables for a command
1000  * @cmd:  command descriptor we wish to initialize
1001  *
1002  * Returns:
1003  * * BLK_STS_OK       - on success
1004  * * BLK_STS_RESOURCE - if the failure is retryable
1005  * * BLK_STS_IOERR    - if the failure is fatal
1006  */
1007 blk_status_t scsi_alloc_sgtables(struct scsi_cmnd *cmd)
1008 {
1009         struct scsi_device *sdev = cmd->device;
1010         struct request *rq = cmd->request;
1011         unsigned short nr_segs = blk_rq_nr_phys_segments(rq);
1012         struct scatterlist *last_sg = NULL;
1013         blk_status_t ret;
1014         bool need_drain = scsi_cmd_needs_dma_drain(sdev, rq);
1015         int count;
1016
1017         if (WARN_ON_ONCE(!nr_segs))
1018                 return BLK_STS_IOERR;
1019
1020         /*
1021          * Make sure there is space for the drain.  The driver must adjust
1022          * max_hw_segments to be prepared for this.
1023          */
1024         if (need_drain)
1025                 nr_segs++;
1026
1027         /*
1028          * If sg table allocation fails, requeue request later.
1029          */
1030         if (unlikely(sg_alloc_table_chained(&cmd->sdb.table, nr_segs,
1031                         cmd->sdb.table.sgl, SCSI_INLINE_SG_CNT)))
1032                 return BLK_STS_RESOURCE;
1033
1034         /*
1035          * Next, walk the list, and fill in the addresses and sizes of
1036          * each segment.
1037          */
1038         count = __blk_rq_map_sg(rq->q, rq, cmd->sdb.table.sgl, &last_sg);
1039
1040         if (blk_rq_bytes(rq) & rq->q->dma_pad_mask) {
1041                 unsigned int pad_len =
1042                         (rq->q->dma_pad_mask & ~blk_rq_bytes(rq)) + 1;
1043
1044                 last_sg->length += pad_len;
1045                 cmd->extra_len += pad_len;
1046         }
1047
1048         if (need_drain) {
1049                 sg_unmark_end(last_sg);
1050                 last_sg = sg_next(last_sg);
1051                 sg_set_buf(last_sg, sdev->dma_drain_buf, sdev->dma_drain_len);
1052                 sg_mark_end(last_sg);
1053
1054                 cmd->extra_len += sdev->dma_drain_len;
1055                 count++;
1056         }
1057
1058         BUG_ON(count > cmd->sdb.table.nents);
1059         cmd->sdb.table.nents = count;
1060         cmd->sdb.length = blk_rq_payload_bytes(rq);
1061
1062         if (blk_integrity_rq(rq)) {
1063                 struct scsi_data_buffer *prot_sdb = cmd->prot_sdb;
1064                 int ivecs;
1065
1066                 if (WARN_ON_ONCE(!prot_sdb)) {
1067                         /*
1068                          * This can happen if someone (e.g. multipath)
1069                          * queues a command to a device on an adapter
1070                          * that does not support DIX.
1071                          */
1072                         ret = BLK_STS_IOERR;
1073                         goto out_free_sgtables;
1074                 }
1075
1076                 ivecs = blk_rq_count_integrity_sg(rq->q, rq->bio);
1077
1078                 if (sg_alloc_table_chained(&prot_sdb->table, ivecs,
1079                                 prot_sdb->table.sgl,
1080                                 SCSI_INLINE_PROT_SG_CNT)) {
1081                         ret = BLK_STS_RESOURCE;
1082                         goto out_free_sgtables;
1083                 }
1084
1085                 count = blk_rq_map_integrity_sg(rq->q, rq->bio,
1086                                                 prot_sdb->table.sgl);
1087                 BUG_ON(count > ivecs);
1088                 BUG_ON(count > queue_max_integrity_segments(rq->q));
1089
1090                 cmd->prot_sdb = prot_sdb;
1091                 cmd->prot_sdb->table.nents = count;
1092         }
1093
1094         return BLK_STS_OK;
1095 out_free_sgtables:
1096         scsi_free_sgtables(cmd);
1097         return ret;
1098 }
1099 EXPORT_SYMBOL(scsi_alloc_sgtables);
1100
1101 /**
1102  * scsi_initialize_rq - initialize struct scsi_cmnd partially
1103  * @rq: Request associated with the SCSI command to be initialized.
1104  *
1105  * This function initializes the members of struct scsi_cmnd that must be
1106  * initialized before request processing starts and that won't be
1107  * reinitialized if a SCSI command is requeued.
1108  *
1109  * Called from inside blk_get_request() for pass-through requests and from
1110  * inside scsi_init_command() for filesystem requests.
1111  */
1112 static void scsi_initialize_rq(struct request *rq)
1113 {
1114         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1115
1116         scsi_req_init(&cmd->req);
1117         init_rcu_head(&cmd->rcu);
1118         cmd->jiffies_at_alloc = jiffies;
1119         cmd->retries = 0;
1120 }
1121
1122 /*
1123  * Only called when the request isn't completed by SCSI, and not freed by
1124  * SCSI
1125  */
1126 static void scsi_cleanup_rq(struct request *rq)
1127 {
1128         if (rq->rq_flags & RQF_DONTPREP) {
1129                 scsi_mq_uninit_cmd(blk_mq_rq_to_pdu(rq));
1130                 rq->rq_flags &= ~RQF_DONTPREP;
1131         }
1132 }
1133
1134 /* Called before a request is prepared. See also scsi_mq_prep_fn(). */
1135 void scsi_init_command(struct scsi_device *dev, struct scsi_cmnd *cmd)
1136 {
1137         void *buf = cmd->sense_buffer;
1138         void *prot = cmd->prot_sdb;
1139         struct request *rq = blk_mq_rq_from_pdu(cmd);
1140         unsigned int flags = cmd->flags & SCMD_PRESERVED_FLAGS;
1141         unsigned long jiffies_at_alloc;
1142         int retries, to_clear;
1143         bool in_flight;
1144
1145         if (!blk_rq_is_scsi(rq) && !(flags & SCMD_INITIALIZED)) {
1146                 flags |= SCMD_INITIALIZED;
1147                 scsi_initialize_rq(rq);
1148         }
1149
1150         jiffies_at_alloc = cmd->jiffies_at_alloc;
1151         retries = cmd->retries;
1152         in_flight = test_bit(SCMD_STATE_INFLIGHT, &cmd->state);
1153         /*
1154          * Zero out the cmd, except for the embedded scsi_request. Only clear
1155          * the driver-private command data if the LLD does not supply a
1156          * function to initialize that data.
1157          */
1158         to_clear = sizeof(*cmd) - sizeof(cmd->req);
1159         if (!dev->host->hostt->init_cmd_priv)
1160                 to_clear += dev->host->hostt->cmd_size;
1161         memset((char *)cmd + sizeof(cmd->req), 0, to_clear);
1162
1163         cmd->device = dev;
1164         cmd->sense_buffer = buf;
1165         cmd->prot_sdb = prot;
1166         cmd->flags = flags;
1167         INIT_DELAYED_WORK(&cmd->abort_work, scmd_eh_abort_handler);
1168         cmd->jiffies_at_alloc = jiffies_at_alloc;
1169         cmd->retries = retries;
1170         if (in_flight)
1171                 __set_bit(SCMD_STATE_INFLIGHT, &cmd->state);
1172
1173 }
1174
1175 static blk_status_t scsi_setup_scsi_cmnd(struct scsi_device *sdev,
1176                 struct request *req)
1177 {
1178         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1179
1180         /*
1181          * Passthrough requests may transfer data, in which case they must
1182          * a bio attached to them.  Or they might contain a SCSI command
1183          * that does not transfer data, in which case they may optionally
1184          * submit a request without an attached bio.
1185          */
1186         if (req->bio) {
1187                 blk_status_t ret = scsi_alloc_sgtables(cmd);
1188                 if (unlikely(ret != BLK_STS_OK))
1189                         return ret;
1190         } else {
1191                 BUG_ON(blk_rq_bytes(req));
1192
1193                 memset(&cmd->sdb, 0, sizeof(cmd->sdb));
1194         }
1195
1196         cmd->cmd_len = scsi_req(req)->cmd_len;
1197         if (cmd->cmd_len == 0)
1198                 cmd->cmd_len = scsi_command_size(cmd->cmnd);
1199         cmd->cmnd = scsi_req(req)->cmd;
1200         cmd->transfersize = blk_rq_bytes(req);
1201         cmd->allowed = scsi_req(req)->retries;
1202         return BLK_STS_OK;
1203 }
1204
1205 static blk_status_t
1206 scsi_device_state_check(struct scsi_device *sdev, struct request *req)
1207 {
1208         switch (sdev->sdev_state) {
1209         case SDEV_OFFLINE:
1210         case SDEV_TRANSPORT_OFFLINE:
1211                 /*
1212                  * If the device is offline we refuse to process any
1213                  * commands.  The device must be brought online
1214                  * before trying any recovery commands.
1215                  */
1216                 if (!sdev->offline_already) {
1217                         sdev->offline_already = true;
1218                         sdev_printk(KERN_ERR, sdev,
1219                                     "rejecting I/O to offline device\n");
1220                 }
1221                 return BLK_STS_IOERR;
1222         case SDEV_DEL:
1223                 /*
1224                  * If the device is fully deleted, we refuse to
1225                  * process any commands as well.
1226                  */
1227                 sdev_printk(KERN_ERR, sdev,
1228                             "rejecting I/O to dead device\n");
1229                 return BLK_STS_IOERR;
1230         case SDEV_BLOCK:
1231         case SDEV_CREATED_BLOCK:
1232                 return BLK_STS_RESOURCE;
1233         case SDEV_QUIESCE:
1234                 /*
1235                  * If the devices is blocked we defer normal commands.
1236                  */
1237                 if (req && !(req->rq_flags & RQF_PREEMPT))
1238                         return BLK_STS_RESOURCE;
1239                 return BLK_STS_OK;
1240         default:
1241                 /*
1242                  * For any other not fully online state we only allow
1243                  * special commands.  In particular any user initiated
1244                  * command is not allowed.
1245                  */
1246                 if (req && !(req->rq_flags & RQF_PREEMPT))
1247                         return BLK_STS_IOERR;
1248                 return BLK_STS_OK;
1249         }
1250 }
1251
1252 /*
1253  * scsi_dev_queue_ready: if we can send requests to sdev, return 1 else
1254  * return 0.
1255  *
1256  * Called with the queue_lock held.
1257  */
1258 static inline int scsi_dev_queue_ready(struct request_queue *q,
1259                                   struct scsi_device *sdev)
1260 {
1261         unsigned int busy;
1262
1263         busy = atomic_inc_return(&sdev->device_busy) - 1;
1264         if (atomic_read(&sdev->device_blocked)) {
1265                 if (busy)
1266                         goto out_dec;
1267
1268                 /*
1269                  * unblock after device_blocked iterates to zero
1270                  */
1271                 if (atomic_dec_return(&sdev->device_blocked) > 0)
1272                         goto out_dec;
1273                 SCSI_LOG_MLQUEUE(3, sdev_printk(KERN_INFO, sdev,
1274                                    "unblocking device at zero depth\n"));
1275         }
1276
1277         if (busy >= sdev->queue_depth)
1278                 goto out_dec;
1279
1280         return 1;
1281 out_dec:
1282         atomic_dec(&sdev->device_busy);
1283         return 0;
1284 }
1285
1286 /*
1287  * scsi_target_queue_ready: checks if there we can send commands to target
1288  * @sdev: scsi device on starget to check.
1289  */
1290 static inline int scsi_target_queue_ready(struct Scsi_Host *shost,
1291                                            struct scsi_device *sdev)
1292 {
1293         struct scsi_target *starget = scsi_target(sdev);
1294         unsigned int busy;
1295
1296         if (starget->single_lun) {
1297                 spin_lock_irq(shost->host_lock);
1298                 if (starget->starget_sdev_user &&
1299                     starget->starget_sdev_user != sdev) {
1300                         spin_unlock_irq(shost->host_lock);
1301                         return 0;
1302                 }
1303                 starget->starget_sdev_user = sdev;
1304                 spin_unlock_irq(shost->host_lock);
1305         }
1306
1307         if (starget->can_queue <= 0)
1308                 return 1;
1309
1310         busy = atomic_inc_return(&starget->target_busy) - 1;
1311         if (atomic_read(&starget->target_blocked) > 0) {
1312                 if (busy)
1313                         goto starved;
1314
1315                 /*
1316                  * unblock after target_blocked iterates to zero
1317                  */
1318                 if (atomic_dec_return(&starget->target_blocked) > 0)
1319                         goto out_dec;
1320
1321                 SCSI_LOG_MLQUEUE(3, starget_printk(KERN_INFO, starget,
1322                                  "unblocking target at zero depth\n"));
1323         }
1324
1325         if (busy >= starget->can_queue)
1326                 goto starved;
1327
1328         return 1;
1329
1330 starved:
1331         spin_lock_irq(shost->host_lock);
1332         list_move_tail(&sdev->starved_entry, &shost->starved_list);
1333         spin_unlock_irq(shost->host_lock);
1334 out_dec:
1335         if (starget->can_queue > 0)
1336                 atomic_dec(&starget->target_busy);
1337         return 0;
1338 }
1339
1340 /*
1341  * scsi_host_queue_ready: if we can send requests to shost, return 1 else
1342  * return 0. We must end up running the queue again whenever 0 is
1343  * returned, else IO can hang.
1344  */
1345 static inline int scsi_host_queue_ready(struct request_queue *q,
1346                                    struct Scsi_Host *shost,
1347                                    struct scsi_device *sdev,
1348                                    struct scsi_cmnd *cmd)
1349 {
1350         if (scsi_host_in_recovery(shost))
1351                 return 0;
1352
1353         if (atomic_read(&shost->host_blocked) > 0) {
1354                 if (scsi_host_busy(shost) > 0)
1355                         goto starved;
1356
1357                 /*
1358                  * unblock after host_blocked iterates to zero
1359                  */
1360                 if (atomic_dec_return(&shost->host_blocked) > 0)
1361                         goto out_dec;
1362
1363                 SCSI_LOG_MLQUEUE(3,
1364                         shost_printk(KERN_INFO, shost,
1365                                      "unblocking host at zero depth\n"));
1366         }
1367
1368         if (shost->host_self_blocked)
1369                 goto starved;
1370
1371         /* We're OK to process the command, so we can't be starved */
1372         if (!list_empty(&sdev->starved_entry)) {
1373                 spin_lock_irq(shost->host_lock);
1374                 if (!list_empty(&sdev->starved_entry))
1375                         list_del_init(&sdev->starved_entry);
1376                 spin_unlock_irq(shost->host_lock);
1377         }
1378
1379         __set_bit(SCMD_STATE_INFLIGHT, &cmd->state);
1380
1381         return 1;
1382
1383 starved:
1384         spin_lock_irq(shost->host_lock);
1385         if (list_empty(&sdev->starved_entry))
1386                 list_add_tail(&sdev->starved_entry, &shost->starved_list);
1387         spin_unlock_irq(shost->host_lock);
1388 out_dec:
1389         scsi_dec_host_busy(shost, cmd);
1390         return 0;
1391 }
1392
1393 /*
1394  * Busy state exporting function for request stacking drivers.
1395  *
1396  * For efficiency, no lock is taken to check the busy state of
1397  * shost/starget/sdev, since the returned value is not guaranteed and
1398  * may be changed after request stacking drivers call the function,
1399  * regardless of taking lock or not.
1400  *
1401  * When scsi can't dispatch I/Os anymore and needs to kill I/Os scsi
1402  * needs to return 'not busy'. Otherwise, request stacking drivers
1403  * may hold requests forever.
1404  */
1405 static bool scsi_mq_lld_busy(struct request_queue *q)
1406 {
1407         struct scsi_device *sdev = q->queuedata;
1408         struct Scsi_Host *shost;
1409
1410         if (blk_queue_dying(q))
1411                 return false;
1412
1413         shost = sdev->host;
1414
1415         /*
1416          * Ignore host/starget busy state.
1417          * Since block layer does not have a concept of fairness across
1418          * multiple queues, congestion of host/starget needs to be handled
1419          * in SCSI layer.
1420          */
1421         if (scsi_host_in_recovery(shost) || scsi_device_is_busy(sdev))
1422                 return true;
1423
1424         return false;
1425 }
1426
1427 static void scsi_softirq_done(struct request *rq)
1428 {
1429         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1430         int disposition;
1431
1432         INIT_LIST_HEAD(&cmd->eh_entry);
1433
1434         atomic_inc(&cmd->device->iodone_cnt);
1435         if (cmd->result)
1436                 atomic_inc(&cmd->device->ioerr_cnt);
1437
1438         disposition = scsi_decide_disposition(cmd);
1439         if (disposition != SUCCESS && scsi_cmd_runtime_exceeced(cmd))
1440                 disposition = SUCCESS;
1441
1442         scsi_log_completion(cmd, disposition);
1443
1444         switch (disposition) {
1445         case SUCCESS:
1446                 scsi_finish_command(cmd);
1447                 break;
1448         case NEEDS_RETRY:
1449                 scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY);
1450                 break;
1451         case ADD_TO_MLQUEUE:
1452                 scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY);
1453                 break;
1454         default:
1455                 scsi_eh_scmd_add(cmd);
1456                 break;
1457         }
1458 }
1459
1460 /**
1461  * scsi_dispatch_cmd - Dispatch a command to the low-level driver.
1462  * @cmd: command block we are dispatching.
1463  *
1464  * Return: nonzero return request was rejected and device's queue needs to be
1465  * plugged.
1466  */
1467 static int scsi_dispatch_cmd(struct scsi_cmnd *cmd)
1468 {
1469         struct Scsi_Host *host = cmd->device->host;
1470         int rtn = 0;
1471
1472         atomic_inc(&cmd->device->iorequest_cnt);
1473
1474         /* check if the device is still usable */
1475         if (unlikely(cmd->device->sdev_state == SDEV_DEL)) {
1476                 /* in SDEV_DEL we error all commands. DID_NO_CONNECT
1477                  * returns an immediate error upwards, and signals
1478                  * that the device is no longer present */
1479                 cmd->result = DID_NO_CONNECT << 16;
1480                 goto done;
1481         }
1482
1483         /* Check to see if the scsi lld made this device blocked. */
1484         if (unlikely(scsi_device_blocked(cmd->device))) {
1485                 /*
1486                  * in blocked state, the command is just put back on
1487                  * the device queue.  The suspend state has already
1488                  * blocked the queue so future requests should not
1489                  * occur until the device transitions out of the
1490                  * suspend state.
1491                  */
1492                 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1493                         "queuecommand : device blocked\n"));
1494                 return SCSI_MLQUEUE_DEVICE_BUSY;
1495         }
1496
1497         /* Store the LUN value in cmnd, if needed. */
1498         if (cmd->device->lun_in_cdb)
1499                 cmd->cmnd[1] = (cmd->cmnd[1] & 0x1f) |
1500                                (cmd->device->lun << 5 & 0xe0);
1501
1502         scsi_log_send(cmd);
1503
1504         /*
1505          * Before we queue this command, check if the command
1506          * length exceeds what the host adapter can handle.
1507          */
1508         if (cmd->cmd_len > cmd->device->host->max_cmd_len) {
1509                 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1510                                "queuecommand : command too long. "
1511                                "cdb_size=%d host->max_cmd_len=%d\n",
1512                                cmd->cmd_len, cmd->device->host->max_cmd_len));
1513                 cmd->result = (DID_ABORT << 16);
1514                 goto done;
1515         }
1516
1517         if (unlikely(host->shost_state == SHOST_DEL)) {
1518                 cmd->result = (DID_NO_CONNECT << 16);
1519                 goto done;
1520
1521         }
1522
1523         trace_scsi_dispatch_cmd_start(cmd);
1524         rtn = host->hostt->queuecommand(host, cmd);
1525         if (rtn) {
1526                 trace_scsi_dispatch_cmd_error(cmd, rtn);
1527                 if (rtn != SCSI_MLQUEUE_DEVICE_BUSY &&
1528                     rtn != SCSI_MLQUEUE_TARGET_BUSY)
1529                         rtn = SCSI_MLQUEUE_HOST_BUSY;
1530
1531                 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1532                         "queuecommand : request rejected\n"));
1533         }
1534
1535         return rtn;
1536  done:
1537         cmd->scsi_done(cmd);
1538         return 0;
1539 }
1540
1541 /* Size in bytes of the sg-list stored in the scsi-mq command-private data. */
1542 static unsigned int scsi_mq_inline_sgl_size(struct Scsi_Host *shost)
1543 {
1544         return min_t(unsigned int, shost->sg_tablesize, SCSI_INLINE_SG_CNT) *
1545                 sizeof(struct scatterlist);
1546 }
1547
1548 static blk_status_t scsi_prepare_cmd(struct request *req)
1549 {
1550         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1551         struct scsi_device *sdev = req->q->queuedata;
1552         struct Scsi_Host *shost = sdev->host;
1553         struct scatterlist *sg;
1554
1555         scsi_init_command(sdev, cmd);
1556
1557         cmd->request = req;
1558         cmd->tag = req->tag;
1559         cmd->prot_op = SCSI_PROT_NORMAL;
1560         if (blk_rq_bytes(req))
1561                 cmd->sc_data_direction = rq_dma_dir(req);
1562         else
1563                 cmd->sc_data_direction = DMA_NONE;
1564
1565         sg = (void *)cmd + sizeof(struct scsi_cmnd) + shost->hostt->cmd_size;
1566         cmd->sdb.table.sgl = sg;
1567
1568         if (scsi_host_get_prot(shost)) {
1569                 memset(cmd->prot_sdb, 0, sizeof(struct scsi_data_buffer));
1570
1571                 cmd->prot_sdb->table.sgl =
1572                         (struct scatterlist *)(cmd->prot_sdb + 1);
1573         }
1574
1575         /*
1576          * Special handling for passthrough commands, which don't go to the ULP
1577          * at all:
1578          */
1579         if (blk_rq_is_scsi(req))
1580                 return scsi_setup_scsi_cmnd(sdev, req);
1581
1582         if (sdev->handler && sdev->handler->prep_fn) {
1583                 blk_status_t ret = sdev->handler->prep_fn(sdev, req);
1584
1585                 if (ret != BLK_STS_OK)
1586                         return ret;
1587         }
1588
1589         cmd->cmnd = scsi_req(req)->cmd = scsi_req(req)->__cmd;
1590         memset(cmd->cmnd, 0, BLK_MAX_CDB);
1591         return scsi_cmd_to_driver(cmd)->init_command(cmd);
1592 }
1593
1594 static void scsi_mq_done(struct scsi_cmnd *cmd)
1595 {
1596         if (unlikely(blk_should_fake_timeout(cmd->request->q)))
1597                 return;
1598         if (unlikely(test_and_set_bit(SCMD_STATE_COMPLETE, &cmd->state)))
1599                 return;
1600         trace_scsi_dispatch_cmd_done(cmd);
1601         blk_mq_complete_request(cmd->request);
1602 }
1603
1604 static void scsi_mq_put_budget(struct request_queue *q)
1605 {
1606         struct scsi_device *sdev = q->queuedata;
1607
1608         atomic_dec(&sdev->device_busy);
1609 }
1610
1611 static bool scsi_mq_get_budget(struct request_queue *q)
1612 {
1613         struct scsi_device *sdev = q->queuedata;
1614
1615         if (scsi_dev_queue_ready(q, sdev))
1616                 return true;
1617
1618         atomic_inc(&sdev->restarts);
1619
1620         /*
1621          * Orders atomic_inc(&sdev->restarts) and atomic_read(&sdev->device_busy).
1622          * .restarts must be incremented before .device_busy is read because the
1623          * code in scsi_run_queue_async() depends on the order of these operations.
1624          */
1625         smp_mb__after_atomic();
1626
1627         /*
1628          * If all in-flight requests originated from this LUN are completed
1629          * before reading .device_busy, sdev->device_busy will be observed as
1630          * zero, then blk_mq_delay_run_hw_queues() will dispatch this request
1631          * soon. Otherwise, completion of one of these requests will observe
1632          * the .restarts flag, and the request queue will be run for handling
1633          * this request, see scsi_end_request().
1634          */
1635         if (unlikely(atomic_read(&sdev->device_busy) == 0 &&
1636                                 !scsi_device_blocked(sdev)))
1637                 blk_mq_delay_run_hw_queues(sdev->request_queue, SCSI_QUEUE_DELAY);
1638         return false;
1639 }
1640
1641 static blk_status_t scsi_queue_rq(struct blk_mq_hw_ctx *hctx,
1642                          const struct blk_mq_queue_data *bd)
1643 {
1644         struct request *req = bd->rq;
1645         struct request_queue *q = req->q;
1646         struct scsi_device *sdev = q->queuedata;
1647         struct Scsi_Host *shost = sdev->host;
1648         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1649         blk_status_t ret;
1650         int reason;
1651
1652         /*
1653          * If the device is not in running state we will reject some or all
1654          * commands.
1655          */
1656         if (unlikely(sdev->sdev_state != SDEV_RUNNING)) {
1657                 ret = scsi_device_state_check(sdev, req);
1658                 if (ret != BLK_STS_OK)
1659                         goto out_put_budget;
1660         }
1661
1662         ret = BLK_STS_RESOURCE;
1663         if (!scsi_target_queue_ready(shost, sdev))
1664                 goto out_put_budget;
1665         if (!scsi_host_queue_ready(q, shost, sdev, cmd))
1666                 goto out_dec_target_busy;
1667
1668         if (!(req->rq_flags & RQF_DONTPREP)) {
1669                 ret = scsi_prepare_cmd(req);
1670                 if (ret != BLK_STS_OK)
1671                         goto out_dec_host_busy;
1672                 req->rq_flags |= RQF_DONTPREP;
1673         } else {
1674                 clear_bit(SCMD_STATE_COMPLETE, &cmd->state);
1675         }
1676
1677         cmd->flags &= SCMD_PRESERVED_FLAGS;
1678         if (sdev->simple_tags)
1679                 cmd->flags |= SCMD_TAGGED;
1680         if (bd->last)
1681                 cmd->flags |= SCMD_LAST;
1682
1683         scsi_set_resid(cmd, 0);
1684         memset(cmd->sense_buffer, 0, SCSI_SENSE_BUFFERSIZE);
1685         cmd->scsi_done = scsi_mq_done;
1686
1687         blk_mq_start_request(req);
1688         reason = scsi_dispatch_cmd(cmd);
1689         if (reason) {
1690                 scsi_set_blocked(cmd, reason);
1691                 ret = BLK_STS_RESOURCE;
1692                 goto out_dec_host_busy;
1693         }
1694
1695         return BLK_STS_OK;
1696
1697 out_dec_host_busy:
1698         scsi_dec_host_busy(shost, cmd);
1699 out_dec_target_busy:
1700         if (scsi_target(sdev)->can_queue > 0)
1701                 atomic_dec(&scsi_target(sdev)->target_busy);
1702 out_put_budget:
1703         scsi_mq_put_budget(q);
1704         switch (ret) {
1705         case BLK_STS_OK:
1706                 break;
1707         case BLK_STS_RESOURCE:
1708         case BLK_STS_ZONE_RESOURCE:
1709                 if (scsi_device_blocked(sdev))
1710                         ret = BLK_STS_DEV_RESOURCE;
1711                 break;
1712         case BLK_STS_AGAIN:
1713                 scsi_req(req)->result = DID_BUS_BUSY << 16;
1714                 if (req->rq_flags & RQF_DONTPREP)
1715                         scsi_mq_uninit_cmd(cmd);
1716                 break;
1717         default:
1718                 if (unlikely(!scsi_device_online(sdev)))
1719                         scsi_req(req)->result = DID_NO_CONNECT << 16;
1720                 else
1721                         scsi_req(req)->result = DID_ERROR << 16;
1722                 /*
1723                  * Make sure to release all allocated resources when
1724                  * we hit an error, as we will never see this command
1725                  * again.
1726                  */
1727                 if (req->rq_flags & RQF_DONTPREP)
1728                         scsi_mq_uninit_cmd(cmd);
1729                 scsi_run_queue_async(sdev);
1730                 break;
1731         }
1732         return ret;
1733 }
1734
1735 static enum blk_eh_timer_return scsi_timeout(struct request *req,
1736                 bool reserved)
1737 {
1738         if (reserved)
1739                 return BLK_EH_RESET_TIMER;
1740         return scsi_times_out(req);
1741 }
1742
1743 static int scsi_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
1744                                 unsigned int hctx_idx, unsigned int numa_node)
1745 {
1746         struct Scsi_Host *shost = set->driver_data;
1747         const bool unchecked_isa_dma = shost->unchecked_isa_dma;
1748         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1749         struct scatterlist *sg;
1750         int ret = 0;
1751
1752         if (unchecked_isa_dma)
1753                 cmd->flags |= SCMD_UNCHECKED_ISA_DMA;
1754         cmd->sense_buffer = scsi_alloc_sense_buffer(unchecked_isa_dma,
1755                                                     GFP_KERNEL, numa_node);
1756         if (!cmd->sense_buffer)
1757                 return -ENOMEM;
1758         cmd->req.sense = cmd->sense_buffer;
1759
1760         if (scsi_host_get_prot(shost)) {
1761                 sg = (void *)cmd + sizeof(struct scsi_cmnd) +
1762                         shost->hostt->cmd_size;
1763                 cmd->prot_sdb = (void *)sg + scsi_mq_inline_sgl_size(shost);
1764         }
1765
1766         if (shost->hostt->init_cmd_priv) {
1767                 ret = shost->hostt->init_cmd_priv(shost, cmd);
1768                 if (ret < 0)
1769                         scsi_free_sense_buffer(unchecked_isa_dma,
1770                                                cmd->sense_buffer);
1771         }
1772
1773         return ret;
1774 }
1775
1776 static void scsi_mq_exit_request(struct blk_mq_tag_set *set, struct request *rq,
1777                                  unsigned int hctx_idx)
1778 {
1779         struct Scsi_Host *shost = set->driver_data;
1780         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1781
1782         if (shost->hostt->exit_cmd_priv)
1783                 shost->hostt->exit_cmd_priv(shost, cmd);
1784         scsi_free_sense_buffer(cmd->flags & SCMD_UNCHECKED_ISA_DMA,
1785                                cmd->sense_buffer);
1786 }
1787
1788 static int scsi_map_queues(struct blk_mq_tag_set *set)
1789 {
1790         struct Scsi_Host *shost = container_of(set, struct Scsi_Host, tag_set);
1791
1792         if (shost->hostt->map_queues)
1793                 return shost->hostt->map_queues(shost);
1794         return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
1795 }
1796
1797 void __scsi_init_queue(struct Scsi_Host *shost, struct request_queue *q)
1798 {
1799         struct device *dev = shost->dma_dev;
1800
1801         /*
1802          * this limit is imposed by hardware restrictions
1803          */
1804         blk_queue_max_segments(q, min_t(unsigned short, shost->sg_tablesize,
1805                                         SG_MAX_SEGMENTS));
1806
1807         if (scsi_host_prot_dma(shost)) {
1808                 shost->sg_prot_tablesize =
1809                         min_not_zero(shost->sg_prot_tablesize,
1810                                      (unsigned short)SCSI_MAX_PROT_SG_SEGMENTS);
1811                 BUG_ON(shost->sg_prot_tablesize < shost->sg_tablesize);
1812                 blk_queue_max_integrity_segments(q, shost->sg_prot_tablesize);
1813         }
1814
1815         if (dev->dma_mask) {
1816                 shost->max_sectors = min_t(unsigned int, shost->max_sectors,
1817                                 dma_max_mapping_size(dev) >> SECTOR_SHIFT);
1818         }
1819         blk_queue_max_hw_sectors(q, shost->max_sectors);
1820         if (shost->unchecked_isa_dma)
1821                 blk_queue_bounce_limit(q, BLK_BOUNCE_ISA);
1822         blk_queue_segment_boundary(q, shost->dma_boundary);
1823         dma_set_seg_boundary(dev, shost->dma_boundary);
1824
1825         blk_queue_max_segment_size(q, shost->max_segment_size);
1826         blk_queue_virt_boundary(q, shost->virt_boundary_mask);
1827         dma_set_max_seg_size(dev, queue_max_segment_size(q));
1828
1829         /*
1830          * Set a reasonable default alignment:  The larger of 32-byte (dword),
1831          * which is a common minimum for HBAs, and the minimum DMA alignment,
1832          * which is set by the platform.
1833          *
1834          * Devices that require a bigger alignment can increase it later.
1835          */
1836         blk_queue_dma_alignment(q, max(4, dma_get_cache_alignment()) - 1);
1837 }
1838 EXPORT_SYMBOL_GPL(__scsi_init_queue);
1839
1840 static const struct blk_mq_ops scsi_mq_ops_no_commit = {
1841         .get_budget     = scsi_mq_get_budget,
1842         .put_budget     = scsi_mq_put_budget,
1843         .queue_rq       = scsi_queue_rq,
1844         .complete       = scsi_softirq_done,
1845         .timeout        = scsi_timeout,
1846 #ifdef CONFIG_BLK_DEBUG_FS
1847         .show_rq        = scsi_show_rq,
1848 #endif
1849         .init_request   = scsi_mq_init_request,
1850         .exit_request   = scsi_mq_exit_request,
1851         .initialize_rq_fn = scsi_initialize_rq,
1852         .cleanup_rq     = scsi_cleanup_rq,
1853         .busy           = scsi_mq_lld_busy,
1854         .map_queues     = scsi_map_queues,
1855 };
1856
1857
1858 static void scsi_commit_rqs(struct blk_mq_hw_ctx *hctx)
1859 {
1860         struct request_queue *q = hctx->queue;
1861         struct scsi_device *sdev = q->queuedata;
1862         struct Scsi_Host *shost = sdev->host;
1863
1864         shost->hostt->commit_rqs(shost, hctx->queue_num);
1865 }
1866
1867 static const struct blk_mq_ops scsi_mq_ops = {
1868         .get_budget     = scsi_mq_get_budget,
1869         .put_budget     = scsi_mq_put_budget,
1870         .queue_rq       = scsi_queue_rq,
1871         .commit_rqs     = scsi_commit_rqs,
1872         .complete       = scsi_softirq_done,
1873         .timeout        = scsi_timeout,
1874 #ifdef CONFIG_BLK_DEBUG_FS
1875         .show_rq        = scsi_show_rq,
1876 #endif
1877         .init_request   = scsi_mq_init_request,
1878         .exit_request   = scsi_mq_exit_request,
1879         .initialize_rq_fn = scsi_initialize_rq,
1880         .cleanup_rq     = scsi_cleanup_rq,
1881         .busy           = scsi_mq_lld_busy,
1882         .map_queues     = scsi_map_queues,
1883 };
1884
1885 struct request_queue *scsi_mq_alloc_queue(struct scsi_device *sdev)
1886 {
1887         sdev->request_queue = blk_mq_init_queue(&sdev->host->tag_set);
1888         if (IS_ERR(sdev->request_queue))
1889                 return NULL;
1890
1891         sdev->request_queue->queuedata = sdev;
1892         __scsi_init_queue(sdev->host, sdev->request_queue);
1893         blk_queue_flag_set(QUEUE_FLAG_SCSI_PASSTHROUGH, sdev->request_queue);
1894         return sdev->request_queue;
1895 }
1896
1897 int scsi_mq_setup_tags(struct Scsi_Host *shost)
1898 {
1899         unsigned int cmd_size, sgl_size;
1900         struct blk_mq_tag_set *tag_set = &shost->tag_set;
1901
1902         sgl_size = max_t(unsigned int, sizeof(struct scatterlist),
1903                                 scsi_mq_inline_sgl_size(shost));
1904         cmd_size = sizeof(struct scsi_cmnd) + shost->hostt->cmd_size + sgl_size;
1905         if (scsi_host_get_prot(shost))
1906                 cmd_size += sizeof(struct scsi_data_buffer) +
1907                         sizeof(struct scatterlist) * SCSI_INLINE_PROT_SG_CNT;
1908
1909         memset(tag_set, 0, sizeof(*tag_set));
1910         if (shost->hostt->commit_rqs)
1911                 tag_set->ops = &scsi_mq_ops;
1912         else
1913                 tag_set->ops = &scsi_mq_ops_no_commit;
1914         tag_set->nr_hw_queues = shost->nr_hw_queues ? : 1;
1915         tag_set->queue_depth = shost->can_queue;
1916         tag_set->cmd_size = cmd_size;
1917         tag_set->numa_node = NUMA_NO_NODE;
1918         tag_set->flags = BLK_MQ_F_SHOULD_MERGE;
1919         tag_set->flags |=
1920                 BLK_ALLOC_POLICY_TO_MQ_FLAG(shost->hostt->tag_alloc_policy);
1921         tag_set->driver_data = shost;
1922         if (shost->host_tagset)
1923                 tag_set->flags |= BLK_MQ_F_TAG_HCTX_SHARED;
1924
1925         return blk_mq_alloc_tag_set(tag_set);
1926 }
1927
1928 void scsi_mq_destroy_tags(struct Scsi_Host *shost)
1929 {
1930         blk_mq_free_tag_set(&shost->tag_set);
1931 }
1932
1933 /**
1934  * scsi_device_from_queue - return sdev associated with a request_queue
1935  * @q: The request queue to return the sdev from
1936  *
1937  * Return the sdev associated with a request queue or NULL if the
1938  * request_queue does not reference a SCSI device.
1939  */
1940 struct scsi_device *scsi_device_from_queue(struct request_queue *q)
1941 {
1942         struct scsi_device *sdev = NULL;
1943
1944         if (q->mq_ops == &scsi_mq_ops_no_commit ||
1945             q->mq_ops == &scsi_mq_ops)
1946                 sdev = q->queuedata;
1947         if (!sdev || !get_device(&sdev->sdev_gendev))
1948                 sdev = NULL;
1949
1950         return sdev;
1951 }
1952
1953 /**
1954  * scsi_block_requests - Utility function used by low-level drivers to prevent
1955  * further commands from being queued to the device.
1956  * @shost:  host in question
1957  *
1958  * There is no timer nor any other means by which the requests get unblocked
1959  * other than the low-level driver calling scsi_unblock_requests().
1960  */
1961 void scsi_block_requests(struct Scsi_Host *shost)
1962 {
1963         shost->host_self_blocked = 1;
1964 }
1965 EXPORT_SYMBOL(scsi_block_requests);
1966
1967 /**
1968  * scsi_unblock_requests - Utility function used by low-level drivers to allow
1969  * further commands to be queued to the device.
1970  * @shost:  host in question
1971  *
1972  * There is no timer nor any other means by which the requests get unblocked
1973  * other than the low-level driver calling scsi_unblock_requests(). This is done
1974  * as an API function so that changes to the internals of the scsi mid-layer
1975  * won't require wholesale 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 void scsi_exit_queue(void)
1985 {
1986         kmem_cache_destroy(scsi_sense_cache);
1987         kmem_cache_destroy(scsi_sense_isadma_cache);
1988 }
1989
1990 /**
1991  *      scsi_mode_select - issue a mode select
1992  *      @sdev:  SCSI device to be queried
1993  *      @pf:    Page format bit (1 == standard, 0 == vendor specific)
1994  *      @sp:    Save page bit (0 == don't save, 1 == save)
1995  *      @modepage: mode page being requested
1996  *      @buffer: request buffer (may not be smaller than eight bytes)
1997  *      @len:   length of request buffer.
1998  *      @timeout: command timeout
1999  *      @retries: number of retries before failing
2000  *      @data: returns a structure abstracting the mode header data
2001  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
2002  *              must be SCSI_SENSE_BUFFERSIZE big.
2003  *
2004  *      Returns zero if successful; negative error number or scsi
2005  *      status on error
2006  *
2007  */
2008 int
2009 scsi_mode_select(struct scsi_device *sdev, int pf, int sp, int modepage,
2010                  unsigned char *buffer, int len, int timeout, int retries,
2011                  struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
2012 {
2013         unsigned char cmd[10];
2014         unsigned char *real_buffer;
2015         int ret;
2016
2017         memset(cmd, 0, sizeof(cmd));
2018         cmd[1] = (pf ? 0x10 : 0) | (sp ? 0x01 : 0);
2019
2020         if (sdev->use_10_for_ms) {
2021                 if (len > 65535)
2022                         return -EINVAL;
2023                 real_buffer = kmalloc(8 + len, GFP_KERNEL);
2024                 if (!real_buffer)
2025                         return -ENOMEM;
2026                 memcpy(real_buffer + 8, buffer, len);
2027                 len += 8;
2028                 real_buffer[0] = 0;
2029                 real_buffer[1] = 0;
2030                 real_buffer[2] = data->medium_type;
2031                 real_buffer[3] = data->device_specific;
2032                 real_buffer[4] = data->longlba ? 0x01 : 0;
2033                 real_buffer[5] = 0;
2034                 real_buffer[6] = data->block_descriptor_length >> 8;
2035                 real_buffer[7] = data->block_descriptor_length;
2036
2037                 cmd[0] = MODE_SELECT_10;
2038                 cmd[7] = len >> 8;
2039                 cmd[8] = len;
2040         } else {
2041                 if (len > 255 || data->block_descriptor_length > 255 ||
2042                     data->longlba)
2043                         return -EINVAL;
2044
2045                 real_buffer = kmalloc(4 + len, GFP_KERNEL);
2046                 if (!real_buffer)
2047                         return -ENOMEM;
2048                 memcpy(real_buffer + 4, buffer, len);
2049                 len += 4;
2050                 real_buffer[0] = 0;
2051                 real_buffer[1] = data->medium_type;
2052                 real_buffer[2] = data->device_specific;
2053                 real_buffer[3] = data->block_descriptor_length;
2054
2055                 cmd[0] = MODE_SELECT;
2056                 cmd[4] = len;
2057         }
2058
2059         ret = scsi_execute_req(sdev, cmd, DMA_TO_DEVICE, real_buffer, len,
2060                                sshdr, timeout, retries, NULL);
2061         kfree(real_buffer);
2062         return ret;
2063 }
2064 EXPORT_SYMBOL_GPL(scsi_mode_select);
2065
2066 /**
2067  *      scsi_mode_sense - issue a mode sense, falling back from 10 to six bytes if necessary.
2068  *      @sdev:  SCSI device to be queried
2069  *      @dbd:   set if mode sense will allow block descriptors to be returned
2070  *      @modepage: mode page being requested
2071  *      @buffer: request buffer (may not be smaller than eight bytes)
2072  *      @len:   length of request buffer.
2073  *      @timeout: command timeout
2074  *      @retries: number of retries before failing
2075  *      @data: returns a structure abstracting the mode header data
2076  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
2077  *              must be SCSI_SENSE_BUFFERSIZE big.
2078  *
2079  *      Returns zero if unsuccessful, or the header offset (either 4
2080  *      or 8 depending on whether a six or ten byte command was
2081  *      issued) if successful.
2082  */
2083 int
2084 scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
2085                   unsigned char *buffer, int len, int timeout, int retries,
2086                   struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
2087 {
2088         unsigned char cmd[12];
2089         int use_10_for_ms;
2090         int header_length;
2091         int result, retry_count = retries;
2092         struct scsi_sense_hdr my_sshdr;
2093
2094         memset(data, 0, sizeof(*data));
2095         memset(&cmd[0], 0, 12);
2096
2097         dbd = sdev->set_dbd_for_ms ? 8 : dbd;
2098         cmd[1] = dbd & 0x18;    /* allows DBD and LLBA bits */
2099         cmd[2] = modepage;
2100
2101         /* caller might not be interested in sense, but we need it */
2102         if (!sshdr)
2103                 sshdr = &my_sshdr;
2104
2105  retry:
2106         use_10_for_ms = sdev->use_10_for_ms;
2107
2108         if (use_10_for_ms) {
2109                 if (len < 8)
2110                         len = 8;
2111
2112                 cmd[0] = MODE_SENSE_10;
2113                 cmd[8] = len;
2114                 header_length = 8;
2115         } else {
2116                 if (len < 4)
2117                         len = 4;
2118
2119                 cmd[0] = MODE_SENSE;
2120                 cmd[4] = len;
2121                 header_length = 4;
2122         }
2123
2124         memset(buffer, 0, len);
2125
2126         result = scsi_execute_req(sdev, cmd, DMA_FROM_DEVICE, buffer, len,
2127                                   sshdr, timeout, retries, NULL);
2128
2129         /* This code looks awful: what it's doing is making sure an
2130          * ILLEGAL REQUEST sense return identifies the actual command
2131          * byte as the problem.  MODE_SENSE commands can return
2132          * ILLEGAL REQUEST if the code page isn't supported */
2133
2134         if (use_10_for_ms && !scsi_status_is_good(result) &&
2135             driver_byte(result) == DRIVER_SENSE) {
2136                 if (scsi_sense_valid(sshdr)) {
2137                         if ((sshdr->sense_key == ILLEGAL_REQUEST) &&
2138                             (sshdr->asc == 0x20) && (sshdr->ascq == 0)) {
2139                                 /*
2140                                  * Invalid command operation code
2141                                  */
2142                                 sdev->use_10_for_ms = 0;
2143                                 goto retry;
2144                         }
2145                 }
2146         }
2147
2148         if (scsi_status_is_good(result)) {
2149                 if (unlikely(buffer[0] == 0x86 && buffer[1] == 0x0b &&
2150                              (modepage == 6 || modepage == 8))) {
2151                         /* Initio breakage? */
2152                         header_length = 0;
2153                         data->length = 13;
2154                         data->medium_type = 0;
2155                         data->device_specific = 0;
2156                         data->longlba = 0;
2157                         data->block_descriptor_length = 0;
2158                 } else if (use_10_for_ms) {
2159                         data->length = buffer[0]*256 + buffer[1] + 2;
2160                         data->medium_type = buffer[2];
2161                         data->device_specific = buffer[3];
2162                         data->longlba = buffer[4] & 0x01;
2163                         data->block_descriptor_length = buffer[6]*256
2164                                 + buffer[7];
2165                 } else {
2166                         data->length = buffer[0] + 1;
2167                         data->medium_type = buffer[1];
2168                         data->device_specific = buffer[2];
2169                         data->block_descriptor_length = buffer[3];
2170                 }
2171                 data->header_length = header_length;
2172         } else if ((status_byte(result) == CHECK_CONDITION) &&
2173                    scsi_sense_valid(sshdr) &&
2174                    sshdr->sense_key == UNIT_ATTENTION && retry_count) {
2175                 retry_count--;
2176                 goto retry;
2177         }
2178
2179         return result;
2180 }
2181 EXPORT_SYMBOL(scsi_mode_sense);
2182
2183 /**
2184  *      scsi_test_unit_ready - test if unit is ready
2185  *      @sdev:  scsi device to change the state of.
2186  *      @timeout: command timeout
2187  *      @retries: number of retries before failing
2188  *      @sshdr: outpout pointer for decoded sense information.
2189  *
2190  *      Returns zero if unsuccessful or an error if TUR failed.  For
2191  *      removable media, UNIT_ATTENTION sets ->changed flag.
2192  **/
2193 int
2194 scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries,
2195                      struct scsi_sense_hdr *sshdr)
2196 {
2197         char cmd[] = {
2198                 TEST_UNIT_READY, 0, 0, 0, 0, 0,
2199         };
2200         int result;
2201
2202         /* try to eat the UNIT_ATTENTION if there are enough retries */
2203         do {
2204                 result = scsi_execute_req(sdev, cmd, DMA_NONE, NULL, 0, sshdr,
2205                                           timeout, 1, NULL);
2206                 if (sdev->removable && scsi_sense_valid(sshdr) &&
2207                     sshdr->sense_key == UNIT_ATTENTION)
2208                         sdev->changed = 1;
2209         } while (scsi_sense_valid(sshdr) &&
2210                  sshdr->sense_key == UNIT_ATTENTION && --retries);
2211
2212         return result;
2213 }
2214 EXPORT_SYMBOL(scsi_test_unit_ready);
2215
2216 /**
2217  *      scsi_device_set_state - Take the given device through the device state model.
2218  *      @sdev:  scsi device to change the state of.
2219  *      @state: state to change to.
2220  *
2221  *      Returns zero if successful or an error if the requested
2222  *      transition is illegal.
2223  */
2224 int
2225 scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
2226 {
2227         enum scsi_device_state oldstate = sdev->sdev_state;
2228
2229         if (state == oldstate)
2230                 return 0;
2231
2232         switch (state) {
2233         case SDEV_CREATED:
2234                 switch (oldstate) {
2235                 case SDEV_CREATED_BLOCK:
2236                         break;
2237                 default:
2238                         goto illegal;
2239                 }
2240                 break;
2241
2242         case SDEV_RUNNING:
2243                 switch (oldstate) {
2244                 case SDEV_CREATED:
2245                 case SDEV_OFFLINE:
2246                 case SDEV_TRANSPORT_OFFLINE:
2247                 case SDEV_QUIESCE:
2248                 case SDEV_BLOCK:
2249                         break;
2250                 default:
2251                         goto illegal;
2252                 }
2253                 break;
2254
2255         case SDEV_QUIESCE:
2256                 switch (oldstate) {
2257                 case SDEV_RUNNING:
2258                 case SDEV_OFFLINE:
2259                 case SDEV_TRANSPORT_OFFLINE:
2260                         break;
2261                 default:
2262                         goto illegal;
2263                 }
2264                 break;
2265
2266         case SDEV_OFFLINE:
2267         case SDEV_TRANSPORT_OFFLINE:
2268                 switch (oldstate) {
2269                 case SDEV_CREATED:
2270                 case SDEV_RUNNING:
2271                 case SDEV_QUIESCE:
2272                 case SDEV_BLOCK:
2273                         break;
2274                 default:
2275                         goto illegal;
2276                 }
2277                 break;
2278
2279         case SDEV_BLOCK:
2280                 switch (oldstate) {
2281                 case SDEV_RUNNING:
2282                 case SDEV_CREATED_BLOCK:
2283                 case SDEV_QUIESCE:
2284                 case SDEV_OFFLINE:
2285                         break;
2286                 default:
2287                         goto illegal;
2288                 }
2289                 break;
2290
2291         case SDEV_CREATED_BLOCK:
2292                 switch (oldstate) {
2293                 case SDEV_CREATED:
2294                         break;
2295                 default:
2296                         goto illegal;
2297                 }
2298                 break;
2299
2300         case SDEV_CANCEL:
2301                 switch (oldstate) {
2302                 case SDEV_CREATED:
2303                 case SDEV_RUNNING:
2304                 case SDEV_QUIESCE:
2305                 case SDEV_OFFLINE:
2306                 case SDEV_TRANSPORT_OFFLINE:
2307                         break;
2308                 default:
2309                         goto illegal;
2310                 }
2311                 break;
2312
2313         case SDEV_DEL:
2314                 switch (oldstate) {
2315                 case SDEV_CREATED:
2316                 case SDEV_RUNNING:
2317                 case SDEV_OFFLINE:
2318                 case SDEV_TRANSPORT_OFFLINE:
2319                 case SDEV_CANCEL:
2320                 case SDEV_BLOCK:
2321                 case SDEV_CREATED_BLOCK:
2322                         break;
2323                 default:
2324                         goto illegal;
2325                 }
2326                 break;
2327
2328         }
2329         sdev->offline_already = false;
2330         sdev->sdev_state = state;
2331         return 0;
2332
2333  illegal:
2334         SCSI_LOG_ERROR_RECOVERY(1,
2335                                 sdev_printk(KERN_ERR, sdev,
2336                                             "Illegal state transition %s->%s",
2337                                             scsi_device_state_name(oldstate),
2338                                             scsi_device_state_name(state))
2339                                 );
2340         return -EINVAL;
2341 }
2342 EXPORT_SYMBOL(scsi_device_set_state);
2343
2344 /**
2345  *      scsi_evt_emit - emit a single SCSI device uevent
2346  *      @sdev: associated SCSI device
2347  *      @evt: event to emit
2348  *
2349  *      Send a single uevent (scsi_event) to the associated scsi_device.
2350  */
2351 static void scsi_evt_emit(struct scsi_device *sdev, struct scsi_event *evt)
2352 {
2353         int idx = 0;
2354         char *envp[3];
2355
2356         switch (evt->evt_type) {
2357         case SDEV_EVT_MEDIA_CHANGE:
2358                 envp[idx++] = "SDEV_MEDIA_CHANGE=1";
2359                 break;
2360         case SDEV_EVT_INQUIRY_CHANGE_REPORTED:
2361                 scsi_rescan_device(&sdev->sdev_gendev);
2362                 envp[idx++] = "SDEV_UA=INQUIRY_DATA_HAS_CHANGED";
2363                 break;
2364         case SDEV_EVT_CAPACITY_CHANGE_REPORTED:
2365                 envp[idx++] = "SDEV_UA=CAPACITY_DATA_HAS_CHANGED";
2366                 break;
2367         case SDEV_EVT_SOFT_THRESHOLD_REACHED_REPORTED:
2368                envp[idx++] = "SDEV_UA=THIN_PROVISIONING_SOFT_THRESHOLD_REACHED";
2369                 break;
2370         case SDEV_EVT_MODE_PARAMETER_CHANGE_REPORTED:
2371                 envp[idx++] = "SDEV_UA=MODE_PARAMETERS_CHANGED";
2372                 break;
2373         case SDEV_EVT_LUN_CHANGE_REPORTED:
2374                 envp[idx++] = "SDEV_UA=REPORTED_LUNS_DATA_HAS_CHANGED";
2375                 break;
2376         case SDEV_EVT_ALUA_STATE_CHANGE_REPORTED:
2377                 envp[idx++] = "SDEV_UA=ASYMMETRIC_ACCESS_STATE_CHANGED";
2378                 break;
2379         case SDEV_EVT_POWER_ON_RESET_OCCURRED:
2380                 envp[idx++] = "SDEV_UA=POWER_ON_RESET_OCCURRED";
2381                 break;
2382         default:
2383                 /* do nothing */
2384                 break;
2385         }
2386
2387         envp[idx++] = NULL;
2388
2389         kobject_uevent_env(&sdev->sdev_gendev.kobj, KOBJ_CHANGE, envp);
2390 }
2391
2392 /**
2393  *      scsi_evt_thread - send a uevent for each scsi event
2394  *      @work: work struct for scsi_device
2395  *
2396  *      Dispatch queued events to their associated scsi_device kobjects
2397  *      as uevents.
2398  */
2399 void scsi_evt_thread(struct work_struct *work)
2400 {
2401         struct scsi_device *sdev;
2402         enum scsi_device_event evt_type;
2403         LIST_HEAD(event_list);
2404
2405         sdev = container_of(work, struct scsi_device, event_work);
2406
2407         for (evt_type = SDEV_EVT_FIRST; evt_type <= SDEV_EVT_LAST; evt_type++)
2408                 if (test_and_clear_bit(evt_type, sdev->pending_events))
2409                         sdev_evt_send_simple(sdev, evt_type, GFP_KERNEL);
2410
2411         while (1) {
2412                 struct scsi_event *evt;
2413                 struct list_head *this, *tmp;
2414                 unsigned long flags;
2415
2416                 spin_lock_irqsave(&sdev->list_lock, flags);
2417                 list_splice_init(&sdev->event_list, &event_list);
2418                 spin_unlock_irqrestore(&sdev->list_lock, flags);
2419
2420                 if (list_empty(&event_list))
2421                         break;
2422
2423                 list_for_each_safe(this, tmp, &event_list) {
2424                         evt = list_entry(this, struct scsi_event, node);
2425                         list_del(&evt->node);
2426                         scsi_evt_emit(sdev, evt);
2427                         kfree(evt);
2428                 }
2429         }
2430 }
2431
2432 /**
2433  *      sdev_evt_send - send asserted event to uevent thread
2434  *      @sdev: scsi_device event occurred on
2435  *      @evt: event to send
2436  *
2437  *      Assert scsi device event asynchronously.
2438  */
2439 void sdev_evt_send(struct scsi_device *sdev, struct scsi_event *evt)
2440 {
2441         unsigned long flags;
2442
2443 #if 0
2444         /* FIXME: currently this check eliminates all media change events
2445          * for polled devices.  Need to update to discriminate between AN
2446          * and polled events */
2447         if (!test_bit(evt->evt_type, sdev->supported_events)) {
2448                 kfree(evt);
2449                 return;
2450         }
2451 #endif
2452
2453         spin_lock_irqsave(&sdev->list_lock, flags);
2454         list_add_tail(&evt->node, &sdev->event_list);
2455         schedule_work(&sdev->event_work);
2456         spin_unlock_irqrestore(&sdev->list_lock, flags);
2457 }
2458 EXPORT_SYMBOL_GPL(sdev_evt_send);
2459
2460 /**
2461  *      sdev_evt_alloc - allocate a new scsi event
2462  *      @evt_type: type of event to allocate
2463  *      @gfpflags: GFP flags for allocation
2464  *
2465  *      Allocates and returns a new scsi_event.
2466  */
2467 struct scsi_event *sdev_evt_alloc(enum scsi_device_event evt_type,
2468                                   gfp_t gfpflags)
2469 {
2470         struct scsi_event *evt = kzalloc(sizeof(struct scsi_event), gfpflags);
2471         if (!evt)
2472                 return NULL;
2473
2474         evt->evt_type = evt_type;
2475         INIT_LIST_HEAD(&evt->node);
2476
2477         /* evt_type-specific initialization, if any */
2478         switch (evt_type) {
2479         case SDEV_EVT_MEDIA_CHANGE:
2480         case SDEV_EVT_INQUIRY_CHANGE_REPORTED:
2481         case SDEV_EVT_CAPACITY_CHANGE_REPORTED:
2482         case SDEV_EVT_SOFT_THRESHOLD_REACHED_REPORTED:
2483         case SDEV_EVT_MODE_PARAMETER_CHANGE_REPORTED:
2484         case SDEV_EVT_LUN_CHANGE_REPORTED:
2485         case SDEV_EVT_ALUA_STATE_CHANGE_REPORTED:
2486         case SDEV_EVT_POWER_ON_RESET_OCCURRED:
2487         default:
2488                 /* do nothing */
2489                 break;
2490         }
2491
2492         return evt;
2493 }
2494 EXPORT_SYMBOL_GPL(sdev_evt_alloc);
2495
2496 /**
2497  *      sdev_evt_send_simple - send asserted event to uevent thread
2498  *      @sdev: scsi_device event occurred on
2499  *      @evt_type: type of event to send
2500  *      @gfpflags: GFP flags for allocation
2501  *
2502  *      Assert scsi device event asynchronously, given an event type.
2503  */
2504 void sdev_evt_send_simple(struct scsi_device *sdev,
2505                           enum scsi_device_event evt_type, gfp_t gfpflags)
2506 {
2507         struct scsi_event *evt = sdev_evt_alloc(evt_type, gfpflags);
2508         if (!evt) {
2509                 sdev_printk(KERN_ERR, sdev, "event %d eaten due to OOM\n",
2510                             evt_type);
2511                 return;
2512         }
2513
2514         sdev_evt_send(sdev, evt);
2515 }
2516 EXPORT_SYMBOL_GPL(sdev_evt_send_simple);
2517
2518 /**
2519  *      scsi_device_quiesce - Block user issued commands.
2520  *      @sdev:  scsi device to quiesce.
2521  *
2522  *      This works by trying to transition to the SDEV_QUIESCE state
2523  *      (which must be a legal transition).  When the device is in this
2524  *      state, only special requests will be accepted, all others will
2525  *      be deferred.  Since special requests may also be requeued requests,
2526  *      a successful return doesn't guarantee the device will be
2527  *      totally quiescent.
2528  *
2529  *      Must be called with user context, may sleep.
2530  *
2531  *      Returns zero if unsuccessful or an error if not.
2532  */
2533 int
2534 scsi_device_quiesce(struct scsi_device *sdev)
2535 {
2536         struct request_queue *q = sdev->request_queue;
2537         int err;
2538
2539         /*
2540          * It is allowed to call scsi_device_quiesce() multiple times from
2541          * the same context but concurrent scsi_device_quiesce() calls are
2542          * not allowed.
2543          */
2544         WARN_ON_ONCE(sdev->quiesced_by && sdev->quiesced_by != current);
2545
2546         if (sdev->quiesced_by == current)
2547                 return 0;
2548
2549         blk_set_pm_only(q);
2550
2551         blk_mq_freeze_queue(q);
2552         /*
2553          * Ensure that the effect of blk_set_pm_only() will be visible
2554          * for percpu_ref_tryget() callers that occur after the queue
2555          * unfreeze even if the queue was already frozen before this function
2556          * was called. See also https://lwn.net/Articles/573497/.
2557          */
2558         synchronize_rcu();
2559         blk_mq_unfreeze_queue(q);
2560
2561         mutex_lock(&sdev->state_mutex);
2562         err = scsi_device_set_state(sdev, SDEV_QUIESCE);
2563         if (err == 0)
2564                 sdev->quiesced_by = current;
2565         else
2566                 blk_clear_pm_only(q);
2567         mutex_unlock(&sdev->state_mutex);
2568
2569         return err;
2570 }
2571 EXPORT_SYMBOL(scsi_device_quiesce);
2572
2573 /**
2574  *      scsi_device_resume - Restart user issued commands to a quiesced device.
2575  *      @sdev:  scsi device to resume.
2576  *
2577  *      Moves the device from quiesced back to running and restarts the
2578  *      queues.
2579  *
2580  *      Must be called with user context, may sleep.
2581  */
2582 void scsi_device_resume(struct scsi_device *sdev)
2583 {
2584         /* check if the device state was mutated prior to resume, and if
2585          * so assume the state is being managed elsewhere (for example
2586          * device deleted during suspend)
2587          */
2588         mutex_lock(&sdev->state_mutex);
2589         if (sdev->quiesced_by) {
2590                 sdev->quiesced_by = NULL;
2591                 blk_clear_pm_only(sdev->request_queue);
2592         }
2593         if (sdev->sdev_state == SDEV_QUIESCE)
2594                 scsi_device_set_state(sdev, SDEV_RUNNING);
2595         mutex_unlock(&sdev->state_mutex);
2596 }
2597 EXPORT_SYMBOL(scsi_device_resume);
2598
2599 static void
2600 device_quiesce_fn(struct scsi_device *sdev, void *data)
2601 {
2602         scsi_device_quiesce(sdev);
2603 }
2604
2605 void
2606 scsi_target_quiesce(struct scsi_target *starget)
2607 {
2608         starget_for_each_device(starget, NULL, device_quiesce_fn);
2609 }
2610 EXPORT_SYMBOL(scsi_target_quiesce);
2611
2612 static void
2613 device_resume_fn(struct scsi_device *sdev, void *data)
2614 {
2615         scsi_device_resume(sdev);
2616 }
2617
2618 void
2619 scsi_target_resume(struct scsi_target *starget)
2620 {
2621         starget_for_each_device(starget, NULL, device_resume_fn);
2622 }
2623 EXPORT_SYMBOL(scsi_target_resume);
2624
2625 /**
2626  * scsi_internal_device_block_nowait - try to transition to the SDEV_BLOCK state
2627  * @sdev: device to block
2628  *
2629  * Pause SCSI command processing on the specified device. Does not sleep.
2630  *
2631  * Returns zero if successful or a negative error code upon failure.
2632  *
2633  * Notes:
2634  * This routine transitions the device to the SDEV_BLOCK state (which must be
2635  * a legal transition). When the device is in this state, command processing
2636  * is paused until the device leaves the SDEV_BLOCK state. See also
2637  * scsi_internal_device_unblock_nowait().
2638  */
2639 int scsi_internal_device_block_nowait(struct scsi_device *sdev)
2640 {
2641         struct request_queue *q = sdev->request_queue;
2642         int err = 0;
2643
2644         err = scsi_device_set_state(sdev, SDEV_BLOCK);
2645         if (err) {
2646                 err = scsi_device_set_state(sdev, SDEV_CREATED_BLOCK);
2647
2648                 if (err)
2649                         return err;
2650         }
2651
2652         /*
2653          * The device has transitioned to SDEV_BLOCK.  Stop the
2654          * block layer from calling the midlayer with this device's
2655          * request queue.
2656          */
2657         blk_mq_quiesce_queue_nowait(q);
2658         return 0;
2659 }
2660 EXPORT_SYMBOL_GPL(scsi_internal_device_block_nowait);
2661
2662 /**
2663  * scsi_internal_device_block - try to transition to the SDEV_BLOCK state
2664  * @sdev: device to block
2665  *
2666  * Pause SCSI command processing on the specified device and wait until all
2667  * ongoing scsi_request_fn() / scsi_queue_rq() calls have finished. May sleep.
2668  *
2669  * Returns zero if successful or a negative error code upon failure.
2670  *
2671  * Note:
2672  * This routine transitions the device to the SDEV_BLOCK state (which must be
2673  * a legal transition). When the device is in this state, command processing
2674  * is paused until the device leaves the SDEV_BLOCK state. See also
2675  * scsi_internal_device_unblock().
2676  */
2677 static int scsi_internal_device_block(struct scsi_device *sdev)
2678 {
2679         struct request_queue *q = sdev->request_queue;
2680         int err;
2681
2682         mutex_lock(&sdev->state_mutex);
2683         err = scsi_internal_device_block_nowait(sdev);
2684         if (err == 0)
2685                 blk_mq_quiesce_queue(q);
2686         mutex_unlock(&sdev->state_mutex);
2687
2688         return err;
2689 }
2690
2691 void scsi_start_queue(struct scsi_device *sdev)
2692 {
2693         struct request_queue *q = sdev->request_queue;
2694
2695         blk_mq_unquiesce_queue(q);
2696 }
2697
2698 /**
2699  * scsi_internal_device_unblock_nowait - resume a device after a block request
2700  * @sdev:       device to resume
2701  * @new_state:  state to set the device to after unblocking
2702  *
2703  * Restart the device queue for a previously suspended SCSI device. Does not
2704  * sleep.
2705  *
2706  * Returns zero if successful or a negative error code upon failure.
2707  *
2708  * Notes:
2709  * This routine transitions the device to the SDEV_RUNNING state or to one of
2710  * the offline states (which must be a legal transition) allowing the midlayer
2711  * to goose the queue for this device.
2712  */
2713 int scsi_internal_device_unblock_nowait(struct scsi_device *sdev,
2714                                         enum scsi_device_state new_state)
2715 {
2716         switch (new_state) {
2717         case SDEV_RUNNING:
2718         case SDEV_TRANSPORT_OFFLINE:
2719                 break;
2720         default:
2721                 return -EINVAL;
2722         }
2723
2724         /*
2725          * Try to transition the scsi device to SDEV_RUNNING or one of the
2726          * offlined states and goose the device queue if successful.
2727          */
2728         switch (sdev->sdev_state) {
2729         case SDEV_BLOCK:
2730         case SDEV_TRANSPORT_OFFLINE:
2731                 sdev->sdev_state = new_state;
2732                 break;
2733         case SDEV_CREATED_BLOCK:
2734                 if (new_state == SDEV_TRANSPORT_OFFLINE ||
2735                     new_state == SDEV_OFFLINE)
2736                         sdev->sdev_state = new_state;
2737                 else
2738                         sdev->sdev_state = SDEV_CREATED;
2739                 break;
2740         case SDEV_CANCEL:
2741         case SDEV_OFFLINE:
2742                 break;
2743         default:
2744                 return -EINVAL;
2745         }
2746         scsi_start_queue(sdev);
2747
2748         return 0;
2749 }
2750 EXPORT_SYMBOL_GPL(scsi_internal_device_unblock_nowait);
2751
2752 /**
2753  * scsi_internal_device_unblock - resume a device after a block request
2754  * @sdev:       device to resume
2755  * @new_state:  state to set the device to after unblocking
2756  *
2757  * Restart the device queue for a previously suspended SCSI device. May sleep.
2758  *
2759  * Returns zero if successful or a negative error code upon failure.
2760  *
2761  * Notes:
2762  * This routine transitions the device to the SDEV_RUNNING state or to one of
2763  * the offline states (which must be a legal transition) allowing the midlayer
2764  * to goose the queue for this device.
2765  */
2766 static int scsi_internal_device_unblock(struct scsi_device *sdev,
2767                                         enum scsi_device_state new_state)
2768 {
2769         int ret;
2770
2771         mutex_lock(&sdev->state_mutex);
2772         ret = scsi_internal_device_unblock_nowait(sdev, new_state);
2773         mutex_unlock(&sdev->state_mutex);
2774
2775         return ret;
2776 }
2777
2778 static void
2779 device_block(struct scsi_device *sdev, void *data)
2780 {
2781         int ret;
2782
2783         ret = scsi_internal_device_block(sdev);
2784
2785         WARN_ONCE(ret, "scsi_internal_device_block(%s) failed: ret = %d\n",
2786                   dev_name(&sdev->sdev_gendev), ret);
2787 }
2788
2789 static int
2790 target_block(struct device *dev, void *data)
2791 {
2792         if (scsi_is_target_device(dev))
2793                 starget_for_each_device(to_scsi_target(dev), NULL,
2794                                         device_block);
2795         return 0;
2796 }
2797
2798 void
2799 scsi_target_block(struct device *dev)
2800 {
2801         if (scsi_is_target_device(dev))
2802                 starget_for_each_device(to_scsi_target(dev), NULL,
2803                                         device_block);
2804         else
2805                 device_for_each_child(dev, NULL, target_block);
2806 }
2807 EXPORT_SYMBOL_GPL(scsi_target_block);
2808
2809 static void
2810 device_unblock(struct scsi_device *sdev, void *data)
2811 {
2812         scsi_internal_device_unblock(sdev, *(enum scsi_device_state *)data);
2813 }
2814
2815 static int
2816 target_unblock(struct device *dev, void *data)
2817 {
2818         if (scsi_is_target_device(dev))
2819                 starget_for_each_device(to_scsi_target(dev), data,
2820                                         device_unblock);
2821         return 0;
2822 }
2823
2824 void
2825 scsi_target_unblock(struct device *dev, enum scsi_device_state new_state)
2826 {
2827         if (scsi_is_target_device(dev))
2828                 starget_for_each_device(to_scsi_target(dev), &new_state,
2829                                         device_unblock);
2830         else
2831                 device_for_each_child(dev, &new_state, target_unblock);
2832 }
2833 EXPORT_SYMBOL_GPL(scsi_target_unblock);
2834
2835 int
2836 scsi_host_block(struct Scsi_Host *shost)
2837 {
2838         struct scsi_device *sdev;
2839         int ret = 0;
2840
2841         /*
2842          * Call scsi_internal_device_block_nowait so we can avoid
2843          * calling synchronize_rcu() for each LUN.
2844          */
2845         shost_for_each_device(sdev, shost) {
2846                 mutex_lock(&sdev->state_mutex);
2847                 ret = scsi_internal_device_block_nowait(sdev);
2848                 mutex_unlock(&sdev->state_mutex);
2849                 if (ret) {
2850                         scsi_device_put(sdev);
2851                         break;
2852                 }
2853         }
2854
2855         /*
2856          * SCSI never enables blk-mq's BLK_MQ_F_BLOCKING flag so
2857          * calling synchronize_rcu() once is enough.
2858          */
2859         WARN_ON_ONCE(shost->tag_set.flags & BLK_MQ_F_BLOCKING);
2860
2861         if (!ret)
2862                 synchronize_rcu();
2863
2864         return ret;
2865 }
2866 EXPORT_SYMBOL_GPL(scsi_host_block);
2867
2868 int
2869 scsi_host_unblock(struct Scsi_Host *shost, int new_state)
2870 {
2871         struct scsi_device *sdev;
2872         int ret = 0;
2873
2874         shost_for_each_device(sdev, shost) {
2875                 ret = scsi_internal_device_unblock(sdev, new_state);
2876                 if (ret) {
2877                         scsi_device_put(sdev);
2878                         break;
2879                 }
2880         }
2881         return ret;
2882 }
2883 EXPORT_SYMBOL_GPL(scsi_host_unblock);
2884
2885 /**
2886  * scsi_kmap_atomic_sg - find and atomically map an sg-elemnt
2887  * @sgl:        scatter-gather list
2888  * @sg_count:   number of segments in sg
2889  * @offset:     offset in bytes into sg, on return offset into the mapped area
2890  * @len:        bytes to map, on return number of bytes mapped
2891  *
2892  * Returns virtual address of the start of the mapped page
2893  */
2894 void *scsi_kmap_atomic_sg(struct scatterlist *sgl, int sg_count,
2895                           size_t *offset, size_t *len)
2896 {
2897         int i;
2898         size_t sg_len = 0, len_complete = 0;
2899         struct scatterlist *sg;
2900         struct page *page;
2901
2902         WARN_ON(!irqs_disabled());
2903
2904         for_each_sg(sgl, sg, sg_count, i) {
2905                 len_complete = sg_len; /* Complete sg-entries */
2906                 sg_len += sg->length;
2907                 if (sg_len > *offset)
2908                         break;
2909         }
2910
2911         if (unlikely(i == sg_count)) {
2912                 printk(KERN_ERR "%s: Bytes in sg: %zu, requested offset %zu, "
2913                         "elements %d\n",
2914                        __func__, sg_len, *offset, sg_count);
2915                 WARN_ON(1);
2916                 return NULL;
2917         }
2918
2919         /* Offset starting from the beginning of first page in this sg-entry */
2920         *offset = *offset - len_complete + sg->offset;
2921
2922         /* Assumption: contiguous pages can be accessed as "page + i" */
2923         page = nth_page(sg_page(sg), (*offset >> PAGE_SHIFT));
2924         *offset &= ~PAGE_MASK;
2925
2926         /* Bytes in this sg-entry from *offset to the end of the page */
2927         sg_len = PAGE_SIZE - *offset;
2928         if (*len > sg_len)
2929                 *len = sg_len;
2930
2931         return kmap_atomic(page);
2932 }
2933 EXPORT_SYMBOL(scsi_kmap_atomic_sg);
2934
2935 /**
2936  * scsi_kunmap_atomic_sg - atomically unmap a virtual address, previously mapped with scsi_kmap_atomic_sg
2937  * @virt:       virtual address to be unmapped
2938  */
2939 void scsi_kunmap_atomic_sg(void *virt)
2940 {
2941         kunmap_atomic(virt);
2942 }
2943 EXPORT_SYMBOL(scsi_kunmap_atomic_sg);
2944
2945 void sdev_disable_disk_events(struct scsi_device *sdev)
2946 {
2947         atomic_inc(&sdev->disk_events_disable_depth);
2948 }
2949 EXPORT_SYMBOL(sdev_disable_disk_events);
2950
2951 void sdev_enable_disk_events(struct scsi_device *sdev)
2952 {
2953         if (WARN_ON_ONCE(atomic_read(&sdev->disk_events_disable_depth) <= 0))
2954                 return;
2955         atomic_dec(&sdev->disk_events_disable_depth);
2956 }
2957 EXPORT_SYMBOL(sdev_enable_disk_events);
2958
2959 static unsigned char designator_prio(const unsigned char *d)
2960 {
2961         if (d[1] & 0x30)
2962                 /* not associated with LUN */
2963                 return 0;
2964
2965         if (d[3] == 0)
2966                 /* invalid length */
2967                 return 0;
2968
2969         /*
2970          * Order of preference for lun descriptor:
2971          * - SCSI name string
2972          * - NAA IEEE Registered Extended
2973          * - EUI-64 based 16-byte
2974          * - EUI-64 based 12-byte
2975          * - NAA IEEE Registered
2976          * - NAA IEEE Extended
2977          * - EUI-64 based 8-byte
2978          * - SCSI name string (truncated)
2979          * - T10 Vendor ID
2980          * as longer descriptors reduce the likelyhood
2981          * of identification clashes.
2982          */
2983
2984         switch (d[1] & 0xf) {
2985         case 8:
2986                 /* SCSI name string, variable-length UTF-8 */
2987                 return 9;
2988         case 3:
2989                 switch (d[4] >> 4) {
2990                 case 6:
2991                         /* NAA registered extended */
2992                         return 8;
2993                 case 5:
2994                         /* NAA registered */
2995                         return 5;
2996                 case 4:
2997                         /* NAA extended */
2998                         return 4;
2999                 case 3:
3000                         /* NAA locally assigned */
3001                         return 1;
3002                 default:
3003                         break;
3004                 }
3005                 break;
3006         case 2:
3007                 switch (d[3]) {
3008                 case 16:
3009                         /* EUI64-based, 16 byte */
3010                         return 7;
3011                 case 12:
3012                         /* EUI64-based, 12 byte */
3013                         return 6;
3014                 case 8:
3015                         /* EUI64-based, 8 byte */
3016                         return 3;
3017                 default:
3018                         break;
3019                 }
3020                 break;
3021         case 1:
3022                 /* T10 vendor ID */
3023                 return 1;
3024         default:
3025                 break;
3026         }
3027
3028         return 0;
3029 }
3030
3031 /**
3032  * scsi_vpd_lun_id - return a unique device identification
3033  * @sdev: SCSI device
3034  * @id:   buffer for the identification
3035  * @id_len:  length of the buffer
3036  *
3037  * Copies a unique device identification into @id based
3038  * on the information in the VPD page 0x83 of the device.
3039  * The string will be formatted as a SCSI name string.
3040  *
3041  * Returns the length of the identification or error on failure.
3042  * If the identifier is longer than the supplied buffer the actual
3043  * identifier length is returned and the buffer is not zero-padded.
3044  */
3045 int scsi_vpd_lun_id(struct scsi_device *sdev, char *id, size_t id_len)
3046 {
3047         u8 cur_id_prio = 0;
3048         u8 cur_id_size = 0;
3049         const unsigned char *d, *cur_id_str;
3050         const struct scsi_vpd *vpd_pg83;
3051         int id_size = -EINVAL;
3052
3053         rcu_read_lock();
3054         vpd_pg83 = rcu_dereference(sdev->vpd_pg83);
3055         if (!vpd_pg83) {
3056                 rcu_read_unlock();
3057                 return -ENXIO;
3058         }
3059
3060         /* The id string must be at least 20 bytes + terminating NULL byte */
3061         if (id_len < 21) {
3062                 rcu_read_unlock();
3063                 return -EINVAL;
3064         }
3065
3066         memset(id, 0, id_len);
3067         for (d = vpd_pg83->data + 4;
3068              d < vpd_pg83->data + vpd_pg83->len;
3069              d += d[3] + 4) {
3070                 u8 prio = designator_prio(d);
3071
3072                 if (prio == 0 || cur_id_prio > prio)
3073                         continue;
3074
3075                 switch (d[1] & 0xf) {
3076                 case 0x1:
3077                         /* T10 Vendor ID */
3078                         if (cur_id_size > d[3])
3079                                 break;
3080                         cur_id_prio = prio;
3081                         cur_id_size = d[3];
3082                         if (cur_id_size + 4 > id_len)
3083                                 cur_id_size = id_len - 4;
3084                         cur_id_str = d + 4;
3085                         id_size = snprintf(id, id_len, "t10.%*pE",
3086                                            cur_id_size, cur_id_str);
3087                         break;
3088                 case 0x2:
3089                         /* EUI-64 */
3090                         cur_id_prio = prio;
3091                         cur_id_size = d[3];
3092                         cur_id_str = d + 4;
3093                         switch (cur_id_size) {
3094                         case 8:
3095                                 id_size = snprintf(id, id_len,
3096                                                    "eui.%8phN",
3097                                                    cur_id_str);
3098                                 break;
3099                         case 12:
3100                                 id_size = snprintf(id, id_len,
3101                                                    "eui.%12phN",
3102                                                    cur_id_str);
3103                                 break;
3104                         case 16:
3105                                 id_size = snprintf(id, id_len,
3106                                                    "eui.%16phN",
3107                                                    cur_id_str);
3108                                 break;
3109                         default:
3110                                 break;
3111                         }
3112                         break;
3113                 case 0x3:
3114                         /* NAA */
3115                         cur_id_prio = prio;
3116                         cur_id_size = d[3];
3117                         cur_id_str = d + 4;
3118                         switch (cur_id_size) {
3119                         case 8:
3120                                 id_size = snprintf(id, id_len,
3121                                                    "naa.%8phN",
3122                                                    cur_id_str);
3123                                 break;
3124                         case 16:
3125                                 id_size = snprintf(id, id_len,
3126                                                    "naa.%16phN",
3127                                                    cur_id_str);
3128                                 break;
3129                         default:
3130                                 break;
3131                         }
3132                         break;
3133                 case 0x8:
3134                         /* SCSI name string */
3135                         if (cur_id_size > d[3])
3136                                 break;
3137                         /* Prefer others for truncated descriptor */
3138                         if (d[3] > id_len) {
3139                                 prio = 2;
3140                                 if (cur_id_prio > prio)
3141                                         break;
3142                         }
3143                         cur_id_prio = prio;
3144                         cur_id_size = id_size = d[3];
3145                         cur_id_str = d + 4;
3146                         if (cur_id_size >= id_len)
3147                                 cur_id_size = id_len - 1;
3148                         memcpy(id, cur_id_str, cur_id_size);
3149                         break;
3150                 default:
3151                         break;
3152                 }
3153         }
3154         rcu_read_unlock();
3155
3156         return id_size;
3157 }
3158 EXPORT_SYMBOL(scsi_vpd_lun_id);
3159
3160 /*
3161  * scsi_vpd_tpg_id - return a target port group identifier
3162  * @sdev: SCSI device
3163  *
3164  * Returns the Target Port Group identifier from the information
3165  * froom VPD page 0x83 of the device.
3166  *
3167  * Returns the identifier or error on failure.
3168  */
3169 int scsi_vpd_tpg_id(struct scsi_device *sdev, int *rel_id)
3170 {
3171         const unsigned char *d;
3172         const struct scsi_vpd *vpd_pg83;
3173         int group_id = -EAGAIN, rel_port = -1;
3174
3175         rcu_read_lock();
3176         vpd_pg83 = rcu_dereference(sdev->vpd_pg83);
3177         if (!vpd_pg83) {
3178                 rcu_read_unlock();
3179                 return -ENXIO;
3180         }
3181
3182         d = vpd_pg83->data + 4;
3183         while (d < vpd_pg83->data + vpd_pg83->len) {
3184                 switch (d[1] & 0xf) {
3185                 case 0x4:
3186                         /* Relative target port */
3187                         rel_port = get_unaligned_be16(&d[6]);
3188                         break;
3189                 case 0x5:
3190                         /* Target port group */
3191                         group_id = get_unaligned_be16(&d[6]);
3192                         break;
3193                 default:
3194                         break;
3195                 }
3196                 d += d[3] + 4;
3197         }
3198         rcu_read_unlock();
3199
3200         if (group_id >= 0 && rel_id && rel_port != -1)
3201                 *rel_id = rel_port;
3202
3203         return group_id;
3204 }
3205 EXPORT_SYMBOL(scsi_vpd_tpg_id);