scsi: core: Remove scsi_init_cmd_errh
[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 static 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
528 static void scsi_mq_uninit_cmd(struct scsi_cmnd *cmd)
529 {
530         scsi_free_sgtables(cmd);
531         scsi_uninit_cmd(cmd);
532 }
533
534 static void scsi_run_queue_async(struct scsi_device *sdev)
535 {
536         if (scsi_target(sdev)->single_lun ||
537             !list_empty(&sdev->host->starved_list)) {
538                 kblockd_schedule_work(&sdev->requeue_work);
539         } else {
540                 /*
541                  * smp_mb() present in sbitmap_queue_clear() or implied in
542                  * .end_io is for ordering writing .device_busy in
543                  * scsi_device_unbusy() and reading sdev->restarts.
544                  */
545                 int old = atomic_read(&sdev->restarts);
546
547                 /*
548                  * ->restarts has to be kept as non-zero if new budget
549                  *  contention occurs.
550                  *
551                  *  No need to run queue when either another re-run
552                  *  queue wins in updating ->restarts or a new budget
553                  *  contention occurs.
554                  */
555                 if (old && atomic_cmpxchg(&sdev->restarts, old, 0) == old)
556                         blk_mq_run_hw_queues(sdev->request_queue, true);
557         }
558 }
559
560 /* Returns false when no more bytes to process, true if there are more */
561 static bool scsi_end_request(struct request *req, blk_status_t error,
562                 unsigned int bytes)
563 {
564         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
565         struct scsi_device *sdev = cmd->device;
566         struct request_queue *q = sdev->request_queue;
567
568         if (blk_update_request(req, error, bytes))
569                 return true;
570
571         if (blk_queue_add_random(q))
572                 add_disk_randomness(req->rq_disk);
573
574         if (!blk_rq_is_scsi(req)) {
575                 WARN_ON_ONCE(!(cmd->flags & SCMD_INITIALIZED));
576                 cmd->flags &= ~SCMD_INITIALIZED;
577         }
578
579         /*
580          * Calling rcu_barrier() is not necessary here because the
581          * SCSI error handler guarantees that the function called by
582          * call_rcu() has been called before scsi_end_request() is
583          * called.
584          */
585         destroy_rcu_head(&cmd->rcu);
586
587         /*
588          * In the MQ case the command gets freed by __blk_mq_end_request,
589          * so we have to do all cleanup that depends on it earlier.
590          *
591          * We also can't kick the queues from irq context, so we
592          * will have to defer it to a workqueue.
593          */
594         scsi_mq_uninit_cmd(cmd);
595
596         /*
597          * queue is still alive, so grab the ref for preventing it
598          * from being cleaned up during running queue.
599          */
600         percpu_ref_get(&q->q_usage_counter);
601
602         __blk_mq_end_request(req, error);
603
604         scsi_run_queue_async(sdev);
605
606         percpu_ref_put(&q->q_usage_counter);
607         return false;
608 }
609
610 /**
611  * scsi_result_to_blk_status - translate a SCSI result code into blk_status_t
612  * @cmd:        SCSI command
613  * @result:     scsi error code
614  *
615  * Translate a SCSI result code into a blk_status_t value. May reset the host
616  * byte of @cmd->result.
617  */
618 static blk_status_t scsi_result_to_blk_status(struct scsi_cmnd *cmd, int result)
619 {
620         switch (host_byte(result)) {
621         case DID_OK:
622                 /*
623                  * Also check the other bytes than the status byte in result
624                  * to handle the case when a SCSI LLD sets result to
625                  * DRIVER_SENSE << 24 without setting SAM_STAT_CHECK_CONDITION.
626                  */
627                 if (scsi_status_is_good(result) && (result & ~0xff) == 0)
628                         return BLK_STS_OK;
629                 return BLK_STS_IOERR;
630         case DID_TRANSPORT_FAILFAST:
631                 return BLK_STS_TRANSPORT;
632         case DID_TARGET_FAILURE:
633                 set_host_byte(cmd, DID_OK);
634                 return BLK_STS_TARGET;
635         case DID_NEXUS_FAILURE:
636                 set_host_byte(cmd, DID_OK);
637                 return BLK_STS_NEXUS;
638         case DID_ALLOC_FAILURE:
639                 set_host_byte(cmd, DID_OK);
640                 return BLK_STS_NOSPC;
641         case DID_MEDIUM_ERROR:
642                 set_host_byte(cmd, DID_OK);
643                 return BLK_STS_MEDIUM;
644         default:
645                 return BLK_STS_IOERR;
646         }
647 }
648
649 /* Helper for scsi_io_completion() when "reprep" action required. */
650 static void scsi_io_completion_reprep(struct scsi_cmnd *cmd,
651                                       struct request_queue *q)
652 {
653         /* A new command will be prepared and issued. */
654         scsi_mq_requeue_cmd(cmd);
655 }
656
657 static bool scsi_cmd_runtime_exceeced(struct scsi_cmnd *cmd)
658 {
659         struct request *req = cmd->request;
660         unsigned long wait_for;
661
662         if (cmd->allowed == SCSI_CMD_RETRIES_NO_LIMIT)
663                 return false;
664
665         wait_for = (cmd->allowed + 1) * req->timeout;
666         if (time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) {
667                 scmd_printk(KERN_ERR, cmd, "timing out command, waited %lus\n",
668                             wait_for/HZ);
669                 return true;
670         }
671         return false;
672 }
673
674 /* Helper for scsi_io_completion() when special action required. */
675 static void scsi_io_completion_action(struct scsi_cmnd *cmd, int result)
676 {
677         struct request_queue *q = cmd->device->request_queue;
678         struct request *req = cmd->request;
679         int level = 0;
680         enum {ACTION_FAIL, ACTION_REPREP, ACTION_RETRY,
681               ACTION_DELAYED_RETRY} action;
682         struct scsi_sense_hdr sshdr;
683         bool sense_valid;
684         bool sense_current = true;      /* false implies "deferred sense" */
685         blk_status_t blk_stat;
686
687         sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
688         if (sense_valid)
689                 sense_current = !scsi_sense_is_deferred(&sshdr);
690
691         blk_stat = scsi_result_to_blk_status(cmd, result);
692
693         if (host_byte(result) == DID_RESET) {
694                 /* Third party bus reset or reset for error recovery
695                  * reasons.  Just retry the command and see what
696                  * happens.
697                  */
698                 action = ACTION_RETRY;
699         } else if (sense_valid && sense_current) {
700                 switch (sshdr.sense_key) {
701                 case UNIT_ATTENTION:
702                         if (cmd->device->removable) {
703                                 /* Detected disc change.  Set a bit
704                                  * and quietly refuse further access.
705                                  */
706                                 cmd->device->changed = 1;
707                                 action = ACTION_FAIL;
708                         } else {
709                                 /* Must have been a power glitch, or a
710                                  * bus reset.  Could not have been a
711                                  * media change, so we just retry the
712                                  * command and see what happens.
713                                  */
714                                 action = ACTION_RETRY;
715                         }
716                         break;
717                 case ILLEGAL_REQUEST:
718                         /* If we had an ILLEGAL REQUEST returned, then
719                          * we may have performed an unsupported
720                          * command.  The only thing this should be
721                          * would be a ten byte read where only a six
722                          * byte read was supported.  Also, on a system
723                          * where READ CAPACITY failed, we may have
724                          * read past the end of the disk.
725                          */
726                         if ((cmd->device->use_10_for_rw &&
727                             sshdr.asc == 0x20 && sshdr.ascq == 0x00) &&
728                             (cmd->cmnd[0] == READ_10 ||
729                              cmd->cmnd[0] == WRITE_10)) {
730                                 /* This will issue a new 6-byte command. */
731                                 cmd->device->use_10_for_rw = 0;
732                                 action = ACTION_REPREP;
733                         } else if (sshdr.asc == 0x10) /* DIX */ {
734                                 action = ACTION_FAIL;
735                                 blk_stat = BLK_STS_PROTECTION;
736                         /* INVALID COMMAND OPCODE or INVALID FIELD IN CDB */
737                         } else if (sshdr.asc == 0x20 || sshdr.asc == 0x24) {
738                                 action = ACTION_FAIL;
739                                 blk_stat = BLK_STS_TARGET;
740                         } else
741                                 action = ACTION_FAIL;
742                         break;
743                 case ABORTED_COMMAND:
744                         action = ACTION_FAIL;
745                         if (sshdr.asc == 0x10) /* DIF */
746                                 blk_stat = BLK_STS_PROTECTION;
747                         break;
748                 case NOT_READY:
749                         /* If the device is in the process of becoming
750                          * ready, or has a temporary blockage, retry.
751                          */
752                         if (sshdr.asc == 0x04) {
753                                 switch (sshdr.ascq) {
754                                 case 0x01: /* becoming ready */
755                                 case 0x04: /* format in progress */
756                                 case 0x05: /* rebuild in progress */
757                                 case 0x06: /* recalculation in progress */
758                                 case 0x07: /* operation in progress */
759                                 case 0x08: /* Long write in progress */
760                                 case 0x09: /* self test in progress */
761                                 case 0x14: /* space allocation in progress */
762                                 case 0x1a: /* start stop unit in progress */
763                                 case 0x1b: /* sanitize in progress */
764                                 case 0x1d: /* configuration in progress */
765                                 case 0x24: /* depopulation in progress */
766                                         action = ACTION_DELAYED_RETRY;
767                                         break;
768                                 default:
769                                         action = ACTION_FAIL;
770                                         break;
771                                 }
772                         } else
773                                 action = ACTION_FAIL;
774                         break;
775                 case VOLUME_OVERFLOW:
776                         /* See SSC3rXX or current. */
777                         action = ACTION_FAIL;
778                         break;
779                 default:
780                         action = ACTION_FAIL;
781                         break;
782                 }
783         } else
784                 action = ACTION_FAIL;
785
786         if (action != ACTION_FAIL && scsi_cmd_runtime_exceeced(cmd))
787                 action = ACTION_FAIL;
788
789         switch (action) {
790         case ACTION_FAIL:
791                 /* Give up and fail the remainder of the request */
792                 if (!(req->rq_flags & RQF_QUIET)) {
793                         static DEFINE_RATELIMIT_STATE(_rs,
794                                         DEFAULT_RATELIMIT_INTERVAL,
795                                         DEFAULT_RATELIMIT_BURST);
796
797                         if (unlikely(scsi_logging_level))
798                                 level =
799                                      SCSI_LOG_LEVEL(SCSI_LOG_MLCOMPLETE_SHIFT,
800                                                     SCSI_LOG_MLCOMPLETE_BITS);
801
802                         /*
803                          * if logging is enabled the failure will be printed
804                          * in scsi_log_completion(), so avoid duplicate messages
805                          */
806                         if (!level && __ratelimit(&_rs)) {
807                                 scsi_print_result(cmd, NULL, FAILED);
808                                 if (driver_byte(result) == DRIVER_SENSE)
809                                         scsi_print_sense(cmd);
810                                 scsi_print_command(cmd);
811                         }
812                 }
813                 if (!scsi_end_request(req, blk_stat, blk_rq_err_bytes(req)))
814                         return;
815                 /*FALLTHRU*/
816         case ACTION_REPREP:
817                 scsi_io_completion_reprep(cmd, q);
818                 break;
819         case ACTION_RETRY:
820                 /* Retry the same command immediately */
821                 __scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY, false);
822                 break;
823         case ACTION_DELAYED_RETRY:
824                 /* Retry the same command after a delay */
825                 __scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY, false);
826                 break;
827         }
828 }
829
830 /*
831  * Helper for scsi_io_completion() when cmd->result is non-zero. Returns a
832  * new result that may suppress further error checking. Also modifies
833  * *blk_statp in some cases.
834  */
835 static int scsi_io_completion_nz_result(struct scsi_cmnd *cmd, int result,
836                                         blk_status_t *blk_statp)
837 {
838         bool sense_valid;
839         bool sense_current = true;      /* false implies "deferred sense" */
840         struct request *req = cmd->request;
841         struct scsi_sense_hdr sshdr;
842
843         sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
844         if (sense_valid)
845                 sense_current = !scsi_sense_is_deferred(&sshdr);
846
847         if (blk_rq_is_passthrough(req)) {
848                 if (sense_valid) {
849                         /*
850                          * SG_IO wants current and deferred errors
851                          */
852                         scsi_req(req)->sense_len =
853                                 min(8 + cmd->sense_buffer[7],
854                                     SCSI_SENSE_BUFFERSIZE);
855                 }
856                 if (sense_current)
857                         *blk_statp = scsi_result_to_blk_status(cmd, result);
858         } else if (blk_rq_bytes(req) == 0 && sense_current) {
859                 /*
860                  * Flush commands do not transfers any data, and thus cannot use
861                  * good_bytes != blk_rq_bytes(req) as the signal for an error.
862                  * This sets *blk_statp explicitly for the problem case.
863                  */
864                 *blk_statp = scsi_result_to_blk_status(cmd, result);
865         }
866         /*
867          * Recovered errors need reporting, but they're always treated as
868          * success, so fiddle the result code here.  For passthrough requests
869          * we already took a copy of the original into sreq->result which
870          * is what gets returned to the user
871          */
872         if (sense_valid && (sshdr.sense_key == RECOVERED_ERROR)) {
873                 bool do_print = true;
874                 /*
875                  * if ATA PASS-THROUGH INFORMATION AVAILABLE [0x0, 0x1d]
876                  * skip print since caller wants ATA registers. Only occurs
877                  * on SCSI ATA PASS_THROUGH commands when CK_COND=1
878                  */
879                 if ((sshdr.asc == 0x0) && (sshdr.ascq == 0x1d))
880                         do_print = false;
881                 else if (req->rq_flags & RQF_QUIET)
882                         do_print = false;
883                 if (do_print)
884                         scsi_print_sense(cmd);
885                 result = 0;
886                 /* for passthrough, *blk_statp may be set */
887                 *blk_statp = BLK_STS_OK;
888         }
889         /*
890          * Another corner case: the SCSI status byte is non-zero but 'good'.
891          * Example: PRE-FETCH command returns SAM_STAT_CONDITION_MET when
892          * it is able to fit nominated LBs in its cache (and SAM_STAT_GOOD
893          * if it can't fit). Treat SAM_STAT_CONDITION_MET and the related
894          * intermediate statuses (both obsolete in SAM-4) as good.
895          */
896         if (status_byte(result) && scsi_status_is_good(result)) {
897                 result = 0;
898                 *blk_statp = BLK_STS_OK;
899         }
900         return result;
901 }
902
903 /**
904  * scsi_io_completion - Completion processing for SCSI commands.
905  * @cmd:        command that is finished.
906  * @good_bytes: number of processed bytes.
907  *
908  * We will finish off the specified number of sectors. If we are done, the
909  * command block will be released and the queue function will be goosed. If we
910  * are not done then we have to figure out what to do next:
911  *
912  *   a) We can call scsi_io_completion_reprep().  The request will be
913  *      unprepared and put back on the queue.  Then a new command will
914  *      be created for it.  This should be used if we made forward
915  *      progress, or if we want to switch from READ(10) to READ(6) for
916  *      example.
917  *
918  *   b) We can call scsi_io_completion_action().  The request will be
919  *      put back on the queue and retried using the same command as
920  *      before, possibly after a delay.
921  *
922  *   c) We can call scsi_end_request() with blk_stat other than
923  *      BLK_STS_OK, to fail the remainder of the request.
924  */
925 void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes)
926 {
927         int result = cmd->result;
928         struct request_queue *q = cmd->device->request_queue;
929         struct request *req = cmd->request;
930         blk_status_t blk_stat = BLK_STS_OK;
931
932         if (unlikely(result))   /* a nz result may or may not be an error */
933                 result = scsi_io_completion_nz_result(cmd, result, &blk_stat);
934
935         if (unlikely(blk_rq_is_passthrough(req))) {
936                 /*
937                  * scsi_result_to_blk_status may have reset the host_byte
938                  */
939                 scsi_req(req)->result = cmd->result;
940         }
941
942         /*
943          * Next deal with any sectors which we were able to correctly
944          * handle.
945          */
946         SCSI_LOG_HLCOMPLETE(1, scmd_printk(KERN_INFO, cmd,
947                 "%u sectors total, %d bytes done.\n",
948                 blk_rq_sectors(req), good_bytes));
949
950         /*
951          * Failed, zero length commands always need to drop down
952          * to retry code. Fast path should return in this block.
953          */
954         if (likely(blk_rq_bytes(req) > 0 || blk_stat == BLK_STS_OK)) {
955                 if (likely(!scsi_end_request(req, blk_stat, good_bytes)))
956                         return; /* no bytes remaining */
957         }
958
959         /* Kill remainder if no retries. */
960         if (unlikely(blk_stat && scsi_noretry_cmd(cmd))) {
961                 if (scsi_end_request(req, blk_stat, blk_rq_bytes(req)))
962                         WARN_ONCE(true,
963                             "Bytes remaining after failed, no-retry command");
964                 return;
965         }
966
967         /*
968          * If there had been no error, but we have leftover bytes in the
969          * requeues just queue the command up again.
970          */
971         if (likely(result == 0))
972                 scsi_io_completion_reprep(cmd, q);
973         else
974                 scsi_io_completion_action(cmd, result);
975 }
976
977 static inline bool scsi_cmd_needs_dma_drain(struct scsi_device *sdev,
978                 struct request *rq)
979 {
980         return sdev->dma_drain_len && blk_rq_is_passthrough(rq) &&
981                !op_is_write(req_op(rq)) &&
982                sdev->host->hostt->dma_need_drain(rq);
983 }
984
985 /**
986  * scsi_init_io - SCSI I/O initialization function.
987  * @cmd:  command descriptor we wish to initialize
988  *
989  * Returns:
990  * * BLK_STS_OK       - on success
991  * * BLK_STS_RESOURCE - if the failure is retryable
992  * * BLK_STS_IOERR    - if the failure is fatal
993  */
994 blk_status_t scsi_init_io(struct scsi_cmnd *cmd)
995 {
996         struct scsi_device *sdev = cmd->device;
997         struct request *rq = cmd->request;
998         unsigned short nr_segs = blk_rq_nr_phys_segments(rq);
999         struct scatterlist *last_sg = NULL;
1000         blk_status_t ret;
1001         bool need_drain = scsi_cmd_needs_dma_drain(sdev, rq);
1002         int count;
1003
1004         if (WARN_ON_ONCE(!nr_segs))
1005                 return BLK_STS_IOERR;
1006
1007         /*
1008          * Make sure there is space for the drain.  The driver must adjust
1009          * max_hw_segments to be prepared for this.
1010          */
1011         if (need_drain)
1012                 nr_segs++;
1013
1014         /*
1015          * If sg table allocation fails, requeue request later.
1016          */
1017         if (unlikely(sg_alloc_table_chained(&cmd->sdb.table, nr_segs,
1018                         cmd->sdb.table.sgl, SCSI_INLINE_SG_CNT)))
1019                 return BLK_STS_RESOURCE;
1020
1021         /*
1022          * Next, walk the list, and fill in the addresses and sizes of
1023          * each segment.
1024          */
1025         count = __blk_rq_map_sg(rq->q, rq, cmd->sdb.table.sgl, &last_sg);
1026
1027         if (blk_rq_bytes(rq) & rq->q->dma_pad_mask) {
1028                 unsigned int pad_len =
1029                         (rq->q->dma_pad_mask & ~blk_rq_bytes(rq)) + 1;
1030
1031                 last_sg->length += pad_len;
1032                 cmd->extra_len += pad_len;
1033         }
1034
1035         if (need_drain) {
1036                 sg_unmark_end(last_sg);
1037                 last_sg = sg_next(last_sg);
1038                 sg_set_buf(last_sg, sdev->dma_drain_buf, sdev->dma_drain_len);
1039                 sg_mark_end(last_sg);
1040
1041                 cmd->extra_len += sdev->dma_drain_len;
1042                 count++;
1043         }
1044
1045         BUG_ON(count > cmd->sdb.table.nents);
1046         cmd->sdb.table.nents = count;
1047         cmd->sdb.length = blk_rq_payload_bytes(rq);
1048
1049         if (blk_integrity_rq(rq)) {
1050                 struct scsi_data_buffer *prot_sdb = cmd->prot_sdb;
1051                 int ivecs;
1052
1053                 if (WARN_ON_ONCE(!prot_sdb)) {
1054                         /*
1055                          * This can happen if someone (e.g. multipath)
1056                          * queues a command to a device on an adapter
1057                          * that does not support DIX.
1058                          */
1059                         ret = BLK_STS_IOERR;
1060                         goto out_free_sgtables;
1061                 }
1062
1063                 ivecs = blk_rq_count_integrity_sg(rq->q, rq->bio);
1064
1065                 if (sg_alloc_table_chained(&prot_sdb->table, ivecs,
1066                                 prot_sdb->table.sgl,
1067                                 SCSI_INLINE_PROT_SG_CNT)) {
1068                         ret = BLK_STS_RESOURCE;
1069                         goto out_free_sgtables;
1070                 }
1071
1072                 count = blk_rq_map_integrity_sg(rq->q, rq->bio,
1073                                                 prot_sdb->table.sgl);
1074                 BUG_ON(count > ivecs);
1075                 BUG_ON(count > queue_max_integrity_segments(rq->q));
1076
1077                 cmd->prot_sdb = prot_sdb;
1078                 cmd->prot_sdb->table.nents = count;
1079         }
1080
1081         return BLK_STS_OK;
1082 out_free_sgtables:
1083         scsi_free_sgtables(cmd);
1084         return ret;
1085 }
1086 EXPORT_SYMBOL(scsi_init_io);
1087
1088 /**
1089  * scsi_initialize_rq - initialize struct scsi_cmnd partially
1090  * @rq: Request associated with the SCSI command to be initialized.
1091  *
1092  * This function initializes the members of struct scsi_cmnd that must be
1093  * initialized before request processing starts and that won't be
1094  * reinitialized if a SCSI command is requeued.
1095  *
1096  * Called from inside blk_get_request() for pass-through requests and from
1097  * inside scsi_init_command() for filesystem requests.
1098  */
1099 static void scsi_initialize_rq(struct request *rq)
1100 {
1101         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1102
1103         scsi_req_init(&cmd->req);
1104         init_rcu_head(&cmd->rcu);
1105         cmd->jiffies_at_alloc = jiffies;
1106         cmd->retries = 0;
1107 }
1108
1109 /*
1110  * Only called when the request isn't completed by SCSI, and not freed by
1111  * SCSI
1112  */
1113 static void scsi_cleanup_rq(struct request *rq)
1114 {
1115         if (rq->rq_flags & RQF_DONTPREP) {
1116                 scsi_mq_uninit_cmd(blk_mq_rq_to_pdu(rq));
1117                 rq->rq_flags &= ~RQF_DONTPREP;
1118         }
1119 }
1120
1121 /* Called before a request is prepared. See also scsi_mq_prep_fn(). */
1122 void scsi_init_command(struct scsi_device *dev, struct scsi_cmnd *cmd)
1123 {
1124         void *buf = cmd->sense_buffer;
1125         void *prot = cmd->prot_sdb;
1126         struct request *rq = blk_mq_rq_from_pdu(cmd);
1127         unsigned int flags = cmd->flags & SCMD_PRESERVED_FLAGS;
1128         unsigned long jiffies_at_alloc;
1129         int retries, to_clear;
1130         bool in_flight;
1131
1132         if (!blk_rq_is_scsi(rq) && !(flags & SCMD_INITIALIZED)) {
1133                 flags |= SCMD_INITIALIZED;
1134                 scsi_initialize_rq(rq);
1135         }
1136
1137         jiffies_at_alloc = cmd->jiffies_at_alloc;
1138         retries = cmd->retries;
1139         in_flight = test_bit(SCMD_STATE_INFLIGHT, &cmd->state);
1140         /*
1141          * Zero out the cmd, except for the embedded scsi_request. Only clear
1142          * the driver-private command data if the LLD does not supply a
1143          * function to initialize that data.
1144          */
1145         to_clear = sizeof(*cmd) - sizeof(cmd->req);
1146         if (!dev->host->hostt->init_cmd_priv)
1147                 to_clear += dev->host->hostt->cmd_size;
1148         memset((char *)cmd + sizeof(cmd->req), 0, to_clear);
1149
1150         cmd->device = dev;
1151         cmd->sense_buffer = buf;
1152         cmd->prot_sdb = prot;
1153         cmd->flags = flags;
1154         INIT_DELAYED_WORK(&cmd->abort_work, scmd_eh_abort_handler);
1155         cmd->jiffies_at_alloc = jiffies_at_alloc;
1156         cmd->retries = retries;
1157         if (in_flight)
1158                 __set_bit(SCMD_STATE_INFLIGHT, &cmd->state);
1159
1160 }
1161
1162 static blk_status_t scsi_setup_scsi_cmnd(struct scsi_device *sdev,
1163                 struct request *req)
1164 {
1165         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1166
1167         /*
1168          * Passthrough requests may transfer data, in which case they must
1169          * a bio attached to them.  Or they might contain a SCSI command
1170          * that does not transfer data, in which case they may optionally
1171          * submit a request without an attached bio.
1172          */
1173         if (req->bio) {
1174                 blk_status_t ret = scsi_init_io(cmd);
1175                 if (unlikely(ret != BLK_STS_OK))
1176                         return ret;
1177         } else {
1178                 BUG_ON(blk_rq_bytes(req));
1179
1180                 memset(&cmd->sdb, 0, sizeof(cmd->sdb));
1181         }
1182
1183         cmd->cmd_len = scsi_req(req)->cmd_len;
1184         cmd->cmnd = scsi_req(req)->cmd;
1185         cmd->transfersize = blk_rq_bytes(req);
1186         cmd->allowed = scsi_req(req)->retries;
1187         return BLK_STS_OK;
1188 }
1189
1190 /*
1191  * Setup a normal block command.  These are simple request from filesystems
1192  * that still need to be translated to SCSI CDBs from the ULD.
1193  */
1194 static blk_status_t scsi_setup_fs_cmnd(struct scsi_device *sdev,
1195                 struct request *req)
1196 {
1197         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1198
1199         if (unlikely(sdev->handler && sdev->handler->prep_fn)) {
1200                 blk_status_t ret = sdev->handler->prep_fn(sdev, req);
1201                 if (ret != BLK_STS_OK)
1202                         return ret;
1203         }
1204
1205         cmd->cmnd = scsi_req(req)->cmd = scsi_req(req)->__cmd;
1206         memset(cmd->cmnd, 0, BLK_MAX_CDB);
1207         return scsi_cmd_to_driver(cmd)->init_command(cmd);
1208 }
1209
1210 static blk_status_t scsi_setup_cmnd(struct scsi_device *sdev,
1211                 struct request *req)
1212 {
1213         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1214         blk_status_t ret;
1215
1216         if (!blk_rq_bytes(req))
1217                 cmd->sc_data_direction = DMA_NONE;
1218         else if (rq_data_dir(req) == WRITE)
1219                 cmd->sc_data_direction = DMA_TO_DEVICE;
1220         else
1221                 cmd->sc_data_direction = DMA_FROM_DEVICE;
1222
1223         if (blk_rq_is_scsi(req))
1224                 ret = scsi_setup_scsi_cmnd(sdev, req);
1225         else
1226                 ret = scsi_setup_fs_cmnd(sdev, req);
1227
1228         if (ret != BLK_STS_OK)
1229                 scsi_free_sgtables(cmd);
1230
1231         return ret;
1232 }
1233
1234 static blk_status_t
1235 scsi_prep_state_check(struct scsi_device *sdev, struct request *req)
1236 {
1237         switch (sdev->sdev_state) {
1238         case SDEV_OFFLINE:
1239         case SDEV_TRANSPORT_OFFLINE:
1240                 /*
1241                  * If the device is offline we refuse to process any
1242                  * commands.  The device must be brought online
1243                  * before trying any recovery commands.
1244                  */
1245                 if (!sdev->offline_already) {
1246                         sdev->offline_already = true;
1247                         sdev_printk(KERN_ERR, sdev,
1248                                     "rejecting I/O to offline device\n");
1249                 }
1250                 return BLK_STS_IOERR;
1251         case SDEV_DEL:
1252                 /*
1253                  * If the device is fully deleted, we refuse to
1254                  * process any commands as well.
1255                  */
1256                 sdev_printk(KERN_ERR, sdev,
1257                             "rejecting I/O to dead device\n");
1258                 return BLK_STS_IOERR;
1259         case SDEV_BLOCK:
1260         case SDEV_CREATED_BLOCK:
1261                 return BLK_STS_RESOURCE;
1262         case SDEV_QUIESCE:
1263                 /*
1264                  * If the devices is blocked we defer normal commands.
1265                  */
1266                 if (req && !(req->rq_flags & RQF_PREEMPT))
1267                         return BLK_STS_RESOURCE;
1268                 return BLK_STS_OK;
1269         default:
1270                 /*
1271                  * For any other not fully online state we only allow
1272                  * special commands.  In particular any user initiated
1273                  * command is not allowed.
1274                  */
1275                 if (req && !(req->rq_flags & RQF_PREEMPT))
1276                         return BLK_STS_IOERR;
1277                 return BLK_STS_OK;
1278         }
1279 }
1280
1281 /*
1282  * scsi_dev_queue_ready: if we can send requests to sdev, return 1 else
1283  * return 0.
1284  *
1285  * Called with the queue_lock held.
1286  */
1287 static inline int scsi_dev_queue_ready(struct request_queue *q,
1288                                   struct scsi_device *sdev)
1289 {
1290         unsigned int busy;
1291
1292         busy = atomic_inc_return(&sdev->device_busy) - 1;
1293         if (atomic_read(&sdev->device_blocked)) {
1294                 if (busy)
1295                         goto out_dec;
1296
1297                 /*
1298                  * unblock after device_blocked iterates to zero
1299                  */
1300                 if (atomic_dec_return(&sdev->device_blocked) > 0)
1301                         goto out_dec;
1302                 SCSI_LOG_MLQUEUE(3, sdev_printk(KERN_INFO, sdev,
1303                                    "unblocking device at zero depth\n"));
1304         }
1305
1306         if (busy >= sdev->queue_depth)
1307                 goto out_dec;
1308
1309         return 1;
1310 out_dec:
1311         atomic_dec(&sdev->device_busy);
1312         return 0;
1313 }
1314
1315 /*
1316  * scsi_target_queue_ready: checks if there we can send commands to target
1317  * @sdev: scsi device on starget to check.
1318  */
1319 static inline int scsi_target_queue_ready(struct Scsi_Host *shost,
1320                                            struct scsi_device *sdev)
1321 {
1322         struct scsi_target *starget = scsi_target(sdev);
1323         unsigned int busy;
1324
1325         if (starget->single_lun) {
1326                 spin_lock_irq(shost->host_lock);
1327                 if (starget->starget_sdev_user &&
1328                     starget->starget_sdev_user != sdev) {
1329                         spin_unlock_irq(shost->host_lock);
1330                         return 0;
1331                 }
1332                 starget->starget_sdev_user = sdev;
1333                 spin_unlock_irq(shost->host_lock);
1334         }
1335
1336         if (starget->can_queue <= 0)
1337                 return 1;
1338
1339         busy = atomic_inc_return(&starget->target_busy) - 1;
1340         if (atomic_read(&starget->target_blocked) > 0) {
1341                 if (busy)
1342                         goto starved;
1343
1344                 /*
1345                  * unblock after target_blocked iterates to zero
1346                  */
1347                 if (atomic_dec_return(&starget->target_blocked) > 0)
1348                         goto out_dec;
1349
1350                 SCSI_LOG_MLQUEUE(3, starget_printk(KERN_INFO, starget,
1351                                  "unblocking target at zero depth\n"));
1352         }
1353
1354         if (busy >= starget->can_queue)
1355                 goto starved;
1356
1357         return 1;
1358
1359 starved:
1360         spin_lock_irq(shost->host_lock);
1361         list_move_tail(&sdev->starved_entry, &shost->starved_list);
1362         spin_unlock_irq(shost->host_lock);
1363 out_dec:
1364         if (starget->can_queue > 0)
1365                 atomic_dec(&starget->target_busy);
1366         return 0;
1367 }
1368
1369 /*
1370  * scsi_host_queue_ready: if we can send requests to shost, return 1 else
1371  * return 0. We must end up running the queue again whenever 0 is
1372  * returned, else IO can hang.
1373  */
1374 static inline int scsi_host_queue_ready(struct request_queue *q,
1375                                    struct Scsi_Host *shost,
1376                                    struct scsi_device *sdev,
1377                                    struct scsi_cmnd *cmd)
1378 {
1379         if (scsi_host_in_recovery(shost))
1380                 return 0;
1381
1382         if (atomic_read(&shost->host_blocked) > 0) {
1383                 if (scsi_host_busy(shost) > 0)
1384                         goto starved;
1385
1386                 /*
1387                  * unblock after host_blocked iterates to zero
1388                  */
1389                 if (atomic_dec_return(&shost->host_blocked) > 0)
1390                         goto out_dec;
1391
1392                 SCSI_LOG_MLQUEUE(3,
1393                         shost_printk(KERN_INFO, shost,
1394                                      "unblocking host at zero depth\n"));
1395         }
1396
1397         if (shost->host_self_blocked)
1398                 goto starved;
1399
1400         /* We're OK to process the command, so we can't be starved */
1401         if (!list_empty(&sdev->starved_entry)) {
1402                 spin_lock_irq(shost->host_lock);
1403                 if (!list_empty(&sdev->starved_entry))
1404                         list_del_init(&sdev->starved_entry);
1405                 spin_unlock_irq(shost->host_lock);
1406         }
1407
1408         __set_bit(SCMD_STATE_INFLIGHT, &cmd->state);
1409
1410         return 1;
1411
1412 starved:
1413         spin_lock_irq(shost->host_lock);
1414         if (list_empty(&sdev->starved_entry))
1415                 list_add_tail(&sdev->starved_entry, &shost->starved_list);
1416         spin_unlock_irq(shost->host_lock);
1417 out_dec:
1418         scsi_dec_host_busy(shost, cmd);
1419         return 0;
1420 }
1421
1422 /*
1423  * Busy state exporting function for request stacking drivers.
1424  *
1425  * For efficiency, no lock is taken to check the busy state of
1426  * shost/starget/sdev, since the returned value is not guaranteed and
1427  * may be changed after request stacking drivers call the function,
1428  * regardless of taking lock or not.
1429  *
1430  * When scsi can't dispatch I/Os anymore and needs to kill I/Os scsi
1431  * needs to return 'not busy'. Otherwise, request stacking drivers
1432  * may hold requests forever.
1433  */
1434 static bool scsi_mq_lld_busy(struct request_queue *q)
1435 {
1436         struct scsi_device *sdev = q->queuedata;
1437         struct Scsi_Host *shost;
1438
1439         if (blk_queue_dying(q))
1440                 return false;
1441
1442         shost = sdev->host;
1443
1444         /*
1445          * Ignore host/starget busy state.
1446          * Since block layer does not have a concept of fairness across
1447          * multiple queues, congestion of host/starget needs to be handled
1448          * in SCSI layer.
1449          */
1450         if (scsi_host_in_recovery(shost) || scsi_device_is_busy(sdev))
1451                 return true;
1452
1453         return false;
1454 }
1455
1456 static void scsi_softirq_done(struct request *rq)
1457 {
1458         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1459         int disposition;
1460
1461         INIT_LIST_HEAD(&cmd->eh_entry);
1462
1463         atomic_inc(&cmd->device->iodone_cnt);
1464         if (cmd->result)
1465                 atomic_inc(&cmd->device->ioerr_cnt);
1466
1467         disposition = scsi_decide_disposition(cmd);
1468         if (disposition != SUCCESS && scsi_cmd_runtime_exceeced(cmd))
1469                 disposition = SUCCESS;
1470
1471         scsi_log_completion(cmd, disposition);
1472
1473         switch (disposition) {
1474         case SUCCESS:
1475                 scsi_finish_command(cmd);
1476                 break;
1477         case NEEDS_RETRY:
1478                 scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY);
1479                 break;
1480         case ADD_TO_MLQUEUE:
1481                 scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY);
1482                 break;
1483         default:
1484                 scsi_eh_scmd_add(cmd);
1485                 break;
1486         }
1487 }
1488
1489 /**
1490  * scsi_dispatch_command - Dispatch a command to the low-level driver.
1491  * @cmd: command block we are dispatching.
1492  *
1493  * Return: nonzero return request was rejected and device's queue needs to be
1494  * plugged.
1495  */
1496 static int scsi_dispatch_cmd(struct scsi_cmnd *cmd)
1497 {
1498         struct Scsi_Host *host = cmd->device->host;
1499         int rtn = 0;
1500
1501         atomic_inc(&cmd->device->iorequest_cnt);
1502
1503         /* check if the device is still usable */
1504         if (unlikely(cmd->device->sdev_state == SDEV_DEL)) {
1505                 /* in SDEV_DEL we error all commands. DID_NO_CONNECT
1506                  * returns an immediate error upwards, and signals
1507                  * that the device is no longer present */
1508                 cmd->result = DID_NO_CONNECT << 16;
1509                 goto done;
1510         }
1511
1512         /* Check to see if the scsi lld made this device blocked. */
1513         if (unlikely(scsi_device_blocked(cmd->device))) {
1514                 /*
1515                  * in blocked state, the command is just put back on
1516                  * the device queue.  The suspend state has already
1517                  * blocked the queue so future requests should not
1518                  * occur until the device transitions out of the
1519                  * suspend state.
1520                  */
1521                 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1522                         "queuecommand : device blocked\n"));
1523                 return SCSI_MLQUEUE_DEVICE_BUSY;
1524         }
1525
1526         /* Store the LUN value in cmnd, if needed. */
1527         if (cmd->device->lun_in_cdb)
1528                 cmd->cmnd[1] = (cmd->cmnd[1] & 0x1f) |
1529                                (cmd->device->lun << 5 & 0xe0);
1530
1531         scsi_log_send(cmd);
1532
1533         /*
1534          * Before we queue this command, check if the command
1535          * length exceeds what the host adapter can handle.
1536          */
1537         if (cmd->cmd_len > cmd->device->host->max_cmd_len) {
1538                 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1539                                "queuecommand : command too long. "
1540                                "cdb_size=%d host->max_cmd_len=%d\n",
1541                                cmd->cmd_len, cmd->device->host->max_cmd_len));
1542                 cmd->result = (DID_ABORT << 16);
1543                 goto done;
1544         }
1545
1546         if (unlikely(host->shost_state == SHOST_DEL)) {
1547                 cmd->result = (DID_NO_CONNECT << 16);
1548                 goto done;
1549
1550         }
1551
1552         trace_scsi_dispatch_cmd_start(cmd);
1553         rtn = host->hostt->queuecommand(host, cmd);
1554         if (rtn) {
1555                 trace_scsi_dispatch_cmd_error(cmd, rtn);
1556                 if (rtn != SCSI_MLQUEUE_DEVICE_BUSY &&
1557                     rtn != SCSI_MLQUEUE_TARGET_BUSY)
1558                         rtn = SCSI_MLQUEUE_HOST_BUSY;
1559
1560                 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1561                         "queuecommand : request rejected\n"));
1562         }
1563
1564         return rtn;
1565  done:
1566         cmd->scsi_done(cmd);
1567         return 0;
1568 }
1569
1570 /* Size in bytes of the sg-list stored in the scsi-mq command-private data. */
1571 static unsigned int scsi_mq_inline_sgl_size(struct Scsi_Host *shost)
1572 {
1573         return min_t(unsigned int, shost->sg_tablesize, SCSI_INLINE_SG_CNT) *
1574                 sizeof(struct scatterlist);
1575 }
1576
1577 static blk_status_t scsi_mq_prep_fn(struct request *req)
1578 {
1579         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1580         struct scsi_device *sdev = req->q->queuedata;
1581         struct Scsi_Host *shost = sdev->host;
1582         struct scatterlist *sg;
1583
1584         scsi_init_command(sdev, cmd);
1585
1586         cmd->request = req;
1587         cmd->tag = req->tag;
1588         cmd->prot_op = SCSI_PROT_NORMAL;
1589
1590         sg = (void *)cmd + sizeof(struct scsi_cmnd) + shost->hostt->cmd_size;
1591         cmd->sdb.table.sgl = sg;
1592
1593         if (scsi_host_get_prot(shost)) {
1594                 memset(cmd->prot_sdb, 0, sizeof(struct scsi_data_buffer));
1595
1596                 cmd->prot_sdb->table.sgl =
1597                         (struct scatterlist *)(cmd->prot_sdb + 1);
1598         }
1599
1600         blk_mq_start_request(req);
1601
1602         return scsi_setup_cmnd(sdev, req);
1603 }
1604
1605 static void scsi_mq_done(struct scsi_cmnd *cmd)
1606 {
1607         if (unlikely(blk_should_fake_timeout(cmd->request->q)))
1608                 return;
1609         if (unlikely(test_and_set_bit(SCMD_STATE_COMPLETE, &cmd->state)))
1610                 return;
1611         trace_scsi_dispatch_cmd_done(cmd);
1612         blk_mq_complete_request(cmd->request);
1613 }
1614
1615 static void scsi_mq_put_budget(struct request_queue *q)
1616 {
1617         struct scsi_device *sdev = q->queuedata;
1618
1619         atomic_dec(&sdev->device_busy);
1620 }
1621
1622 static bool scsi_mq_get_budget(struct request_queue *q)
1623 {
1624         struct scsi_device *sdev = q->queuedata;
1625
1626         if (scsi_dev_queue_ready(q, sdev))
1627                 return true;
1628
1629         atomic_inc(&sdev->restarts);
1630
1631         /*
1632          * Orders atomic_inc(&sdev->restarts) and atomic_read(&sdev->device_busy).
1633          * .restarts must be incremented before .device_busy is read because the
1634          * code in scsi_run_queue_async() depends on the order of these operations.
1635          */
1636         smp_mb__after_atomic();
1637
1638         /*
1639          * If all in-flight requests originated from this LUN are completed
1640          * before reading .device_busy, sdev->device_busy will be observed as
1641          * zero, then blk_mq_delay_run_hw_queues() will dispatch this request
1642          * soon. Otherwise, completion of one of these requests will observe
1643          * the .restarts flag, and the request queue will be run for handling
1644          * this request, see scsi_end_request().
1645          */
1646         if (unlikely(atomic_read(&sdev->device_busy) == 0 &&
1647                                 !scsi_device_blocked(sdev)))
1648                 blk_mq_delay_run_hw_queues(sdev->request_queue, SCSI_QUEUE_DELAY);
1649         return false;
1650 }
1651
1652 static blk_status_t scsi_queue_rq(struct blk_mq_hw_ctx *hctx,
1653                          const struct blk_mq_queue_data *bd)
1654 {
1655         struct request *req = bd->rq;
1656         struct request_queue *q = req->q;
1657         struct scsi_device *sdev = q->queuedata;
1658         struct Scsi_Host *shost = sdev->host;
1659         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1660         blk_status_t ret;
1661         int reason;
1662
1663         /*
1664          * If the device is not in running state we will reject some or all
1665          * commands.
1666          */
1667         if (unlikely(sdev->sdev_state != SDEV_RUNNING)) {
1668                 ret = scsi_prep_state_check(sdev, req);
1669                 if (ret != BLK_STS_OK)
1670                         goto out_put_budget;
1671         }
1672
1673         ret = BLK_STS_RESOURCE;
1674         if (!scsi_target_queue_ready(shost, sdev))
1675                 goto out_put_budget;
1676         if (!scsi_host_queue_ready(q, shost, sdev, cmd))
1677                 goto out_dec_target_busy;
1678
1679         if (!(req->rq_flags & RQF_DONTPREP)) {
1680                 ret = scsi_mq_prep_fn(req);
1681                 if (ret != BLK_STS_OK)
1682                         goto out_dec_host_busy;
1683                 req->rq_flags |= RQF_DONTPREP;
1684         } else {
1685                 clear_bit(SCMD_STATE_COMPLETE, &cmd->state);
1686                 blk_mq_start_request(req);
1687         }
1688
1689         cmd->flags &= SCMD_PRESERVED_FLAGS;
1690         if (sdev->simple_tags)
1691                 cmd->flags |= SCMD_TAGGED;
1692         if (bd->last)
1693                 cmd->flags |= SCMD_LAST;
1694
1695         scsi_set_resid(cmd, 0);
1696         memset(cmd->sense_buffer, 0, SCSI_SENSE_BUFFERSIZE);
1697         if (cmd->cmd_len == 0)
1698                 cmd->cmd_len = scsi_command_size(cmd->cmnd);
1699         cmd->scsi_done = scsi_mq_done;
1700
1701         reason = scsi_dispatch_cmd(cmd);
1702         if (reason) {
1703                 scsi_set_blocked(cmd, reason);
1704                 ret = BLK_STS_RESOURCE;
1705                 goto out_dec_host_busy;
1706         }
1707
1708         return BLK_STS_OK;
1709
1710 out_dec_host_busy:
1711         scsi_dec_host_busy(shost, cmd);
1712 out_dec_target_busy:
1713         if (scsi_target(sdev)->can_queue > 0)
1714                 atomic_dec(&scsi_target(sdev)->target_busy);
1715 out_put_budget:
1716         scsi_mq_put_budget(q);
1717         switch (ret) {
1718         case BLK_STS_OK:
1719                 break;
1720         case BLK_STS_RESOURCE:
1721         case BLK_STS_ZONE_RESOURCE:
1722                 if (atomic_read(&sdev->device_busy) ||
1723                     scsi_device_blocked(sdev))
1724                         ret = BLK_STS_DEV_RESOURCE;
1725                 break;
1726         default:
1727                 if (unlikely(!scsi_device_online(sdev)))
1728                         scsi_req(req)->result = DID_NO_CONNECT << 16;
1729                 else
1730                         scsi_req(req)->result = DID_ERROR << 16;
1731                 /*
1732                  * Make sure to release all allocated resources when
1733                  * we hit an error, as we will never see this command
1734                  * again.
1735                  */
1736                 if (req->rq_flags & RQF_DONTPREP)
1737                         scsi_mq_uninit_cmd(cmd);
1738                 scsi_run_queue_async(sdev);
1739                 break;
1740         }
1741         return ret;
1742 }
1743
1744 static enum blk_eh_timer_return scsi_timeout(struct request *req,
1745                 bool reserved)
1746 {
1747         if (reserved)
1748                 return BLK_EH_RESET_TIMER;
1749         return scsi_times_out(req);
1750 }
1751
1752 static int scsi_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
1753                                 unsigned int hctx_idx, unsigned int numa_node)
1754 {
1755         struct Scsi_Host *shost = set->driver_data;
1756         const bool unchecked_isa_dma = shost->unchecked_isa_dma;
1757         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1758         struct scatterlist *sg;
1759         int ret = 0;
1760
1761         if (unchecked_isa_dma)
1762                 cmd->flags |= SCMD_UNCHECKED_ISA_DMA;
1763         cmd->sense_buffer = scsi_alloc_sense_buffer(unchecked_isa_dma,
1764                                                     GFP_KERNEL, numa_node);
1765         if (!cmd->sense_buffer)
1766                 return -ENOMEM;
1767         cmd->req.sense = cmd->sense_buffer;
1768
1769         if (scsi_host_get_prot(shost)) {
1770                 sg = (void *)cmd + sizeof(struct scsi_cmnd) +
1771                         shost->hostt->cmd_size;
1772                 cmd->prot_sdb = (void *)sg + scsi_mq_inline_sgl_size(shost);
1773         }
1774
1775         if (shost->hostt->init_cmd_priv) {
1776                 ret = shost->hostt->init_cmd_priv(shost, cmd);
1777                 if (ret < 0)
1778                         scsi_free_sense_buffer(unchecked_isa_dma,
1779                                                cmd->sense_buffer);
1780         }
1781
1782         return ret;
1783 }
1784
1785 static void scsi_mq_exit_request(struct blk_mq_tag_set *set, struct request *rq,
1786                                  unsigned int hctx_idx)
1787 {
1788         struct Scsi_Host *shost = set->driver_data;
1789         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1790
1791         if (shost->hostt->exit_cmd_priv)
1792                 shost->hostt->exit_cmd_priv(shost, cmd);
1793         scsi_free_sense_buffer(cmd->flags & SCMD_UNCHECKED_ISA_DMA,
1794                                cmd->sense_buffer);
1795 }
1796
1797 static int scsi_map_queues(struct blk_mq_tag_set *set)
1798 {
1799         struct Scsi_Host *shost = container_of(set, struct Scsi_Host, tag_set);
1800
1801         if (shost->hostt->map_queues)
1802                 return shost->hostt->map_queues(shost);
1803         return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
1804 }
1805
1806 void __scsi_init_queue(struct Scsi_Host *shost, struct request_queue *q)
1807 {
1808         struct device *dev = shost->dma_dev;
1809
1810         /*
1811          * this limit is imposed by hardware restrictions
1812          */
1813         blk_queue_max_segments(q, min_t(unsigned short, shost->sg_tablesize,
1814                                         SG_MAX_SEGMENTS));
1815
1816         if (scsi_host_prot_dma(shost)) {
1817                 shost->sg_prot_tablesize =
1818                         min_not_zero(shost->sg_prot_tablesize,
1819                                      (unsigned short)SCSI_MAX_PROT_SG_SEGMENTS);
1820                 BUG_ON(shost->sg_prot_tablesize < shost->sg_tablesize);
1821                 blk_queue_max_integrity_segments(q, shost->sg_prot_tablesize);
1822         }
1823
1824         if (dev->dma_mask) {
1825                 shost->max_sectors = min_t(unsigned int, shost->max_sectors,
1826                                 dma_max_mapping_size(dev) >> SECTOR_SHIFT);
1827         }
1828         blk_queue_max_hw_sectors(q, shost->max_sectors);
1829         if (shost->unchecked_isa_dma)
1830                 blk_queue_bounce_limit(q, BLK_BOUNCE_ISA);
1831         blk_queue_segment_boundary(q, shost->dma_boundary);
1832         dma_set_seg_boundary(dev, shost->dma_boundary);
1833
1834         blk_queue_max_segment_size(q, shost->max_segment_size);
1835         blk_queue_virt_boundary(q, shost->virt_boundary_mask);
1836         dma_set_max_seg_size(dev, queue_max_segment_size(q));
1837
1838         /*
1839          * Set a reasonable default alignment:  The larger of 32-byte (dword),
1840          * which is a common minimum for HBAs, and the minimum DMA alignment,
1841          * which is set by the platform.
1842          *
1843          * Devices that require a bigger alignment can increase it later.
1844          */
1845         blk_queue_dma_alignment(q, max(4, dma_get_cache_alignment()) - 1);
1846 }
1847 EXPORT_SYMBOL_GPL(__scsi_init_queue);
1848
1849 static const struct blk_mq_ops scsi_mq_ops_no_commit = {
1850         .get_budget     = scsi_mq_get_budget,
1851         .put_budget     = scsi_mq_put_budget,
1852         .queue_rq       = scsi_queue_rq,
1853         .complete       = scsi_softirq_done,
1854         .timeout        = scsi_timeout,
1855 #ifdef CONFIG_BLK_DEBUG_FS
1856         .show_rq        = scsi_show_rq,
1857 #endif
1858         .init_request   = scsi_mq_init_request,
1859         .exit_request   = scsi_mq_exit_request,
1860         .initialize_rq_fn = scsi_initialize_rq,
1861         .cleanup_rq     = scsi_cleanup_rq,
1862         .busy           = scsi_mq_lld_busy,
1863         .map_queues     = scsi_map_queues,
1864 };
1865
1866
1867 static void scsi_commit_rqs(struct blk_mq_hw_ctx *hctx)
1868 {
1869         struct request_queue *q = hctx->queue;
1870         struct scsi_device *sdev = q->queuedata;
1871         struct Scsi_Host *shost = sdev->host;
1872
1873         shost->hostt->commit_rqs(shost, hctx->queue_num);
1874 }
1875
1876 static const struct blk_mq_ops scsi_mq_ops = {
1877         .get_budget     = scsi_mq_get_budget,
1878         .put_budget     = scsi_mq_put_budget,
1879         .queue_rq       = scsi_queue_rq,
1880         .commit_rqs     = scsi_commit_rqs,
1881         .complete       = scsi_softirq_done,
1882         .timeout        = scsi_timeout,
1883 #ifdef CONFIG_BLK_DEBUG_FS
1884         .show_rq        = scsi_show_rq,
1885 #endif
1886         .init_request   = scsi_mq_init_request,
1887         .exit_request   = scsi_mq_exit_request,
1888         .initialize_rq_fn = scsi_initialize_rq,
1889         .cleanup_rq     = scsi_cleanup_rq,
1890         .busy           = scsi_mq_lld_busy,
1891         .map_queues     = scsi_map_queues,
1892 };
1893
1894 struct request_queue *scsi_mq_alloc_queue(struct scsi_device *sdev)
1895 {
1896         sdev->request_queue = blk_mq_init_queue(&sdev->host->tag_set);
1897         if (IS_ERR(sdev->request_queue))
1898                 return NULL;
1899
1900         sdev->request_queue->queuedata = sdev;
1901         __scsi_init_queue(sdev->host, sdev->request_queue);
1902         blk_queue_flag_set(QUEUE_FLAG_SCSI_PASSTHROUGH, sdev->request_queue);
1903         return sdev->request_queue;
1904 }
1905
1906 int scsi_mq_setup_tags(struct Scsi_Host *shost)
1907 {
1908         unsigned int cmd_size, sgl_size;
1909         struct blk_mq_tag_set *tag_set = &shost->tag_set;
1910
1911         sgl_size = max_t(unsigned int, sizeof(struct scatterlist),
1912                                 scsi_mq_inline_sgl_size(shost));
1913         cmd_size = sizeof(struct scsi_cmnd) + shost->hostt->cmd_size + sgl_size;
1914         if (scsi_host_get_prot(shost))
1915                 cmd_size += sizeof(struct scsi_data_buffer) +
1916                         sizeof(struct scatterlist) * SCSI_INLINE_PROT_SG_CNT;
1917
1918         memset(tag_set, 0, sizeof(*tag_set));
1919         if (shost->hostt->commit_rqs)
1920                 tag_set->ops = &scsi_mq_ops;
1921         else
1922                 tag_set->ops = &scsi_mq_ops_no_commit;
1923         tag_set->nr_hw_queues = shost->nr_hw_queues ? : 1;
1924         tag_set->queue_depth = shost->can_queue;
1925         tag_set->cmd_size = cmd_size;
1926         tag_set->numa_node = NUMA_NO_NODE;
1927         tag_set->flags = BLK_MQ_F_SHOULD_MERGE;
1928         tag_set->flags |=
1929                 BLK_ALLOC_POLICY_TO_MQ_FLAG(shost->hostt->tag_alloc_policy);
1930         tag_set->driver_data = shost;
1931
1932         return blk_mq_alloc_tag_set(tag_set);
1933 }
1934
1935 void scsi_mq_destroy_tags(struct Scsi_Host *shost)
1936 {
1937         blk_mq_free_tag_set(&shost->tag_set);
1938 }
1939
1940 /**
1941  * scsi_device_from_queue - return sdev associated with a request_queue
1942  * @q: The request queue to return the sdev from
1943  *
1944  * Return the sdev associated with a request queue or NULL if the
1945  * request_queue does not reference a SCSI device.
1946  */
1947 struct scsi_device *scsi_device_from_queue(struct request_queue *q)
1948 {
1949         struct scsi_device *sdev = NULL;
1950
1951         if (q->mq_ops == &scsi_mq_ops_no_commit ||
1952             q->mq_ops == &scsi_mq_ops)
1953                 sdev = q->queuedata;
1954         if (!sdev || !get_device(&sdev->sdev_gendev))
1955                 sdev = NULL;
1956
1957         return sdev;
1958 }
1959
1960 /**
1961  * scsi_block_requests - Utility function used by low-level drivers to prevent
1962  * further commands from being queued to the device.
1963  * @shost:  host in question
1964  *
1965  * There is no timer nor any other means by which the requests get unblocked
1966  * other than the low-level driver calling scsi_unblock_requests().
1967  */
1968 void scsi_block_requests(struct Scsi_Host *shost)
1969 {
1970         shost->host_self_blocked = 1;
1971 }
1972 EXPORT_SYMBOL(scsi_block_requests);
1973
1974 /**
1975  * scsi_unblock_requests - Utility function used by low-level drivers to allow
1976  * further commands to be queued to the device.
1977  * @shost:  host in question
1978  *
1979  * There is no timer nor any other means by which the requests get unblocked
1980  * other than the low-level driver calling scsi_unblock_requests(). This is done
1981  * as an API function so that changes to the internals of the scsi mid-layer
1982  * won't require wholesale changes to drivers that use this feature.
1983  */
1984 void scsi_unblock_requests(struct Scsi_Host *shost)
1985 {
1986         shost->host_self_blocked = 0;
1987         scsi_run_host_queues(shost);
1988 }
1989 EXPORT_SYMBOL(scsi_unblock_requests);
1990
1991 void scsi_exit_queue(void)
1992 {
1993         kmem_cache_destroy(scsi_sense_cache);
1994         kmem_cache_destroy(scsi_sense_isadma_cache);
1995 }
1996
1997 /**
1998  *      scsi_mode_select - issue a mode select
1999  *      @sdev:  SCSI device to be queried
2000  *      @pf:    Page format bit (1 == standard, 0 == vendor specific)
2001  *      @sp:    Save page bit (0 == don't save, 1 == save)
2002  *      @modepage: mode page being requested
2003  *      @buffer: request buffer (may not be smaller than eight bytes)
2004  *      @len:   length of request buffer.
2005  *      @timeout: command timeout
2006  *      @retries: number of retries before failing
2007  *      @data: returns a structure abstracting the mode header data
2008  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
2009  *              must be SCSI_SENSE_BUFFERSIZE big.
2010  *
2011  *      Returns zero if successful; negative error number or scsi
2012  *      status on error
2013  *
2014  */
2015 int
2016 scsi_mode_select(struct scsi_device *sdev, int pf, int sp, int modepage,
2017                  unsigned char *buffer, int len, int timeout, int retries,
2018                  struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
2019 {
2020         unsigned char cmd[10];
2021         unsigned char *real_buffer;
2022         int ret;
2023
2024         memset(cmd, 0, sizeof(cmd));
2025         cmd[1] = (pf ? 0x10 : 0) | (sp ? 0x01 : 0);
2026
2027         if (sdev->use_10_for_ms) {
2028                 if (len > 65535)
2029                         return -EINVAL;
2030                 real_buffer = kmalloc(8 + len, GFP_KERNEL);
2031                 if (!real_buffer)
2032                         return -ENOMEM;
2033                 memcpy(real_buffer + 8, buffer, len);
2034                 len += 8;
2035                 real_buffer[0] = 0;
2036                 real_buffer[1] = 0;
2037                 real_buffer[2] = data->medium_type;
2038                 real_buffer[3] = data->device_specific;
2039                 real_buffer[4] = data->longlba ? 0x01 : 0;
2040                 real_buffer[5] = 0;
2041                 real_buffer[6] = data->block_descriptor_length >> 8;
2042                 real_buffer[7] = data->block_descriptor_length;
2043
2044                 cmd[0] = MODE_SELECT_10;
2045                 cmd[7] = len >> 8;
2046                 cmd[8] = len;
2047         } else {
2048                 if (len > 255 || data->block_descriptor_length > 255 ||
2049                     data->longlba)
2050                         return -EINVAL;
2051
2052                 real_buffer = kmalloc(4 + len, GFP_KERNEL);
2053                 if (!real_buffer)
2054                         return -ENOMEM;
2055                 memcpy(real_buffer + 4, buffer, len);
2056                 len += 4;
2057                 real_buffer[0] = 0;
2058                 real_buffer[1] = data->medium_type;
2059                 real_buffer[2] = data->device_specific;
2060                 real_buffer[3] = data->block_descriptor_length;
2061
2062                 cmd[0] = MODE_SELECT;
2063                 cmd[4] = len;
2064         }
2065
2066         ret = scsi_execute_req(sdev, cmd, DMA_TO_DEVICE, real_buffer, len,
2067                                sshdr, timeout, retries, NULL);
2068         kfree(real_buffer);
2069         return ret;
2070 }
2071 EXPORT_SYMBOL_GPL(scsi_mode_select);
2072
2073 /**
2074  *      scsi_mode_sense - issue a mode sense, falling back from 10 to six bytes if necessary.
2075  *      @sdev:  SCSI device to be queried
2076  *      @dbd:   set if mode sense will allow block descriptors to be returned
2077  *      @modepage: mode page being requested
2078  *      @buffer: request buffer (may not be smaller than eight bytes)
2079  *      @len:   length of request buffer.
2080  *      @timeout: command timeout
2081  *      @retries: number of retries before failing
2082  *      @data: returns a structure abstracting the mode header data
2083  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
2084  *              must be SCSI_SENSE_BUFFERSIZE big.
2085  *
2086  *      Returns zero if unsuccessful, or the header offset (either 4
2087  *      or 8 depending on whether a six or ten byte command was
2088  *      issued) if successful.
2089  */
2090 int
2091 scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
2092                   unsigned char *buffer, int len, int timeout, int retries,
2093                   struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
2094 {
2095         unsigned char cmd[12];
2096         int use_10_for_ms;
2097         int header_length;
2098         int result, retry_count = retries;
2099         struct scsi_sense_hdr my_sshdr;
2100
2101         memset(data, 0, sizeof(*data));
2102         memset(&cmd[0], 0, 12);
2103
2104         dbd = sdev->set_dbd_for_ms ? 8 : dbd;
2105         cmd[1] = dbd & 0x18;    /* allows DBD and LLBA bits */
2106         cmd[2] = modepage;
2107
2108         /* caller might not be interested in sense, but we need it */
2109         if (!sshdr)
2110                 sshdr = &my_sshdr;
2111
2112  retry:
2113         use_10_for_ms = sdev->use_10_for_ms;
2114
2115         if (use_10_for_ms) {
2116                 if (len < 8)
2117                         len = 8;
2118
2119                 cmd[0] = MODE_SENSE_10;
2120                 cmd[8] = len;
2121                 header_length = 8;
2122         } else {
2123                 if (len < 4)
2124                         len = 4;
2125
2126                 cmd[0] = MODE_SENSE;
2127                 cmd[4] = len;
2128                 header_length = 4;
2129         }
2130
2131         memset(buffer, 0, len);
2132
2133         result = scsi_execute_req(sdev, cmd, DMA_FROM_DEVICE, buffer, len,
2134                                   sshdr, timeout, retries, NULL);
2135
2136         /* This code looks awful: what it's doing is making sure an
2137          * ILLEGAL REQUEST sense return identifies the actual command
2138          * byte as the problem.  MODE_SENSE commands can return
2139          * ILLEGAL REQUEST if the code page isn't supported */
2140
2141         if (use_10_for_ms && !scsi_status_is_good(result) &&
2142             driver_byte(result) == DRIVER_SENSE) {
2143                 if (scsi_sense_valid(sshdr)) {
2144                         if ((sshdr->sense_key == ILLEGAL_REQUEST) &&
2145                             (sshdr->asc == 0x20) && (sshdr->ascq == 0)) {
2146                                 /*
2147                                  * Invalid command operation code
2148                                  */
2149                                 sdev->use_10_for_ms = 0;
2150                                 goto retry;
2151                         }
2152                 }
2153         }
2154
2155         if (scsi_status_is_good(result)) {
2156                 if (unlikely(buffer[0] == 0x86 && buffer[1] == 0x0b &&
2157                              (modepage == 6 || modepage == 8))) {
2158                         /* Initio breakage? */
2159                         header_length = 0;
2160                         data->length = 13;
2161                         data->medium_type = 0;
2162                         data->device_specific = 0;
2163                         data->longlba = 0;
2164                         data->block_descriptor_length = 0;
2165                 } else if (use_10_for_ms) {
2166                         data->length = buffer[0]*256 + buffer[1] + 2;
2167                         data->medium_type = buffer[2];
2168                         data->device_specific = buffer[3];
2169                         data->longlba = buffer[4] & 0x01;
2170                         data->block_descriptor_length = buffer[6]*256
2171                                 + buffer[7];
2172                 } else {
2173                         data->length = buffer[0] + 1;
2174                         data->medium_type = buffer[1];
2175                         data->device_specific = buffer[2];
2176                         data->block_descriptor_length = buffer[3];
2177                 }
2178                 data->header_length = header_length;
2179         } else if ((status_byte(result) == CHECK_CONDITION) &&
2180                    scsi_sense_valid(sshdr) &&
2181                    sshdr->sense_key == UNIT_ATTENTION && retry_count) {
2182                 retry_count--;
2183                 goto retry;
2184         }
2185
2186         return result;
2187 }
2188 EXPORT_SYMBOL(scsi_mode_sense);
2189
2190 /**
2191  *      scsi_test_unit_ready - test if unit is ready
2192  *      @sdev:  scsi device to change the state of.
2193  *      @timeout: command timeout
2194  *      @retries: number of retries before failing
2195  *      @sshdr: outpout pointer for decoded sense information.
2196  *
2197  *      Returns zero if unsuccessful or an error if TUR failed.  For
2198  *      removable media, UNIT_ATTENTION sets ->changed flag.
2199  **/
2200 int
2201 scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries,
2202                      struct scsi_sense_hdr *sshdr)
2203 {
2204         char cmd[] = {
2205                 TEST_UNIT_READY, 0, 0, 0, 0, 0,
2206         };
2207         int result;
2208
2209         /* try to eat the UNIT_ATTENTION if there are enough retries */
2210         do {
2211                 result = scsi_execute_req(sdev, cmd, DMA_NONE, NULL, 0, sshdr,
2212                                           timeout, 1, NULL);
2213                 if (sdev->removable && scsi_sense_valid(sshdr) &&
2214                     sshdr->sense_key == UNIT_ATTENTION)
2215                         sdev->changed = 1;
2216         } while (scsi_sense_valid(sshdr) &&
2217                  sshdr->sense_key == UNIT_ATTENTION && --retries);
2218
2219         return result;
2220 }
2221 EXPORT_SYMBOL(scsi_test_unit_ready);
2222
2223 /**
2224  *      scsi_device_set_state - Take the given device through the device state model.
2225  *      @sdev:  scsi device to change the state of.
2226  *      @state: state to change to.
2227  *
2228  *      Returns zero if successful or an error if the requested
2229  *      transition is illegal.
2230  */
2231 int
2232 scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
2233 {
2234         enum scsi_device_state oldstate = sdev->sdev_state;
2235
2236         if (state == oldstate)
2237                 return 0;
2238
2239         switch (state) {
2240         case SDEV_CREATED:
2241                 switch (oldstate) {
2242                 case SDEV_CREATED_BLOCK:
2243                         break;
2244                 default:
2245                         goto illegal;
2246                 }
2247                 break;
2248
2249         case SDEV_RUNNING:
2250                 switch (oldstate) {
2251                 case SDEV_CREATED:
2252                 case SDEV_OFFLINE:
2253                 case SDEV_TRANSPORT_OFFLINE:
2254                 case SDEV_QUIESCE:
2255                 case SDEV_BLOCK:
2256                         break;
2257                 default:
2258                         goto illegal;
2259                 }
2260                 break;
2261
2262         case SDEV_QUIESCE:
2263                 switch (oldstate) {
2264                 case SDEV_RUNNING:
2265                 case SDEV_OFFLINE:
2266                 case SDEV_TRANSPORT_OFFLINE:
2267                         break;
2268                 default:
2269                         goto illegal;
2270                 }
2271                 break;
2272
2273         case SDEV_OFFLINE:
2274         case SDEV_TRANSPORT_OFFLINE:
2275                 switch (oldstate) {
2276                 case SDEV_CREATED:
2277                 case SDEV_RUNNING:
2278                 case SDEV_QUIESCE:
2279                 case SDEV_BLOCK:
2280                         break;
2281                 default:
2282                         goto illegal;
2283                 }
2284                 break;
2285
2286         case SDEV_BLOCK:
2287                 switch (oldstate) {
2288                 case SDEV_RUNNING:
2289                 case SDEV_CREATED_BLOCK:
2290                 case SDEV_QUIESCE:
2291                 case SDEV_OFFLINE:
2292                         break;
2293                 default:
2294                         goto illegal;
2295                 }
2296                 break;
2297
2298         case SDEV_CREATED_BLOCK:
2299                 switch (oldstate) {
2300                 case SDEV_CREATED:
2301                         break;
2302                 default:
2303                         goto illegal;
2304                 }
2305                 break;
2306
2307         case SDEV_CANCEL:
2308                 switch (oldstate) {
2309                 case SDEV_CREATED:
2310                 case SDEV_RUNNING:
2311                 case SDEV_QUIESCE:
2312                 case SDEV_OFFLINE:
2313                 case SDEV_TRANSPORT_OFFLINE:
2314                         break;
2315                 default:
2316                         goto illegal;
2317                 }
2318                 break;
2319
2320         case SDEV_DEL:
2321                 switch (oldstate) {
2322                 case SDEV_CREATED:
2323                 case SDEV_RUNNING:
2324                 case SDEV_OFFLINE:
2325                 case SDEV_TRANSPORT_OFFLINE:
2326                 case SDEV_CANCEL:
2327                 case SDEV_BLOCK:
2328                 case SDEV_CREATED_BLOCK:
2329                         break;
2330                 default:
2331                         goto illegal;
2332                 }
2333                 break;
2334
2335         }
2336         sdev->offline_already = false;
2337         sdev->sdev_state = state;
2338         return 0;
2339
2340  illegal:
2341         SCSI_LOG_ERROR_RECOVERY(1,
2342                                 sdev_printk(KERN_ERR, sdev,
2343                                             "Illegal state transition %s->%s",
2344                                             scsi_device_state_name(oldstate),
2345                                             scsi_device_state_name(state))
2346                                 );
2347         return -EINVAL;
2348 }
2349 EXPORT_SYMBOL(scsi_device_set_state);
2350
2351 /**
2352  *      sdev_evt_emit - emit a single SCSI device uevent
2353  *      @sdev: associated SCSI device
2354  *      @evt: event to emit
2355  *
2356  *      Send a single uevent (scsi_event) to the associated scsi_device.
2357  */
2358 static void scsi_evt_emit(struct scsi_device *sdev, struct scsi_event *evt)
2359 {
2360         int idx = 0;
2361         char *envp[3];
2362
2363         switch (evt->evt_type) {
2364         case SDEV_EVT_MEDIA_CHANGE:
2365                 envp[idx++] = "SDEV_MEDIA_CHANGE=1";
2366                 break;
2367         case SDEV_EVT_INQUIRY_CHANGE_REPORTED:
2368                 scsi_rescan_device(&sdev->sdev_gendev);
2369                 envp[idx++] = "SDEV_UA=INQUIRY_DATA_HAS_CHANGED";
2370                 break;
2371         case SDEV_EVT_CAPACITY_CHANGE_REPORTED:
2372                 envp[idx++] = "SDEV_UA=CAPACITY_DATA_HAS_CHANGED";
2373                 break;
2374         case SDEV_EVT_SOFT_THRESHOLD_REACHED_REPORTED:
2375                envp[idx++] = "SDEV_UA=THIN_PROVISIONING_SOFT_THRESHOLD_REACHED";
2376                 break;
2377         case SDEV_EVT_MODE_PARAMETER_CHANGE_REPORTED:
2378                 envp[idx++] = "SDEV_UA=MODE_PARAMETERS_CHANGED";
2379                 break;
2380         case SDEV_EVT_LUN_CHANGE_REPORTED:
2381                 envp[idx++] = "SDEV_UA=REPORTED_LUNS_DATA_HAS_CHANGED";
2382                 break;
2383         case SDEV_EVT_ALUA_STATE_CHANGE_REPORTED:
2384                 envp[idx++] = "SDEV_UA=ASYMMETRIC_ACCESS_STATE_CHANGED";
2385                 break;
2386         case SDEV_EVT_POWER_ON_RESET_OCCURRED:
2387                 envp[idx++] = "SDEV_UA=POWER_ON_RESET_OCCURRED";
2388                 break;
2389         default:
2390                 /* do nothing */
2391                 break;
2392         }
2393
2394         envp[idx++] = NULL;
2395
2396         kobject_uevent_env(&sdev->sdev_gendev.kobj, KOBJ_CHANGE, envp);
2397 }
2398
2399 /**
2400  *      sdev_evt_thread - send a uevent for each scsi event
2401  *      @work: work struct for scsi_device
2402  *
2403  *      Dispatch queued events to their associated scsi_device kobjects
2404  *      as uevents.
2405  */
2406 void scsi_evt_thread(struct work_struct *work)
2407 {
2408         struct scsi_device *sdev;
2409         enum scsi_device_event evt_type;
2410         LIST_HEAD(event_list);
2411
2412         sdev = container_of(work, struct scsi_device, event_work);
2413
2414         for (evt_type = SDEV_EVT_FIRST; evt_type <= SDEV_EVT_LAST; evt_type++)
2415                 if (test_and_clear_bit(evt_type, sdev->pending_events))
2416                         sdev_evt_send_simple(sdev, evt_type, GFP_KERNEL);
2417
2418         while (1) {
2419                 struct scsi_event *evt;
2420                 struct list_head *this, *tmp;
2421                 unsigned long flags;
2422
2423                 spin_lock_irqsave(&sdev->list_lock, flags);
2424                 list_splice_init(&sdev->event_list, &event_list);
2425                 spin_unlock_irqrestore(&sdev->list_lock, flags);
2426
2427                 if (list_empty(&event_list))
2428                         break;
2429
2430                 list_for_each_safe(this, tmp, &event_list) {
2431                         evt = list_entry(this, struct scsi_event, node);
2432                         list_del(&evt->node);
2433                         scsi_evt_emit(sdev, evt);
2434                         kfree(evt);
2435                 }
2436         }
2437 }
2438
2439 /**
2440  *      sdev_evt_send - send asserted event to uevent thread
2441  *      @sdev: scsi_device event occurred on
2442  *      @evt: event to send
2443  *
2444  *      Assert scsi device event asynchronously.
2445  */
2446 void sdev_evt_send(struct scsi_device *sdev, struct scsi_event *evt)
2447 {
2448         unsigned long flags;
2449
2450 #if 0
2451         /* FIXME: currently this check eliminates all media change events
2452          * for polled devices.  Need to update to discriminate between AN
2453          * and polled events */
2454         if (!test_bit(evt->evt_type, sdev->supported_events)) {
2455                 kfree(evt);
2456                 return;
2457         }
2458 #endif
2459
2460         spin_lock_irqsave(&sdev->list_lock, flags);
2461         list_add_tail(&evt->node, &sdev->event_list);
2462         schedule_work(&sdev->event_work);
2463         spin_unlock_irqrestore(&sdev->list_lock, flags);
2464 }
2465 EXPORT_SYMBOL_GPL(sdev_evt_send);
2466
2467 /**
2468  *      sdev_evt_alloc - allocate a new scsi event
2469  *      @evt_type: type of event to allocate
2470  *      @gfpflags: GFP flags for allocation
2471  *
2472  *      Allocates and returns a new scsi_event.
2473  */
2474 struct scsi_event *sdev_evt_alloc(enum scsi_device_event evt_type,
2475                                   gfp_t gfpflags)
2476 {
2477         struct scsi_event *evt = kzalloc(sizeof(struct scsi_event), gfpflags);
2478         if (!evt)
2479                 return NULL;
2480
2481         evt->evt_type = evt_type;
2482         INIT_LIST_HEAD(&evt->node);
2483
2484         /* evt_type-specific initialization, if any */
2485         switch (evt_type) {
2486         case SDEV_EVT_MEDIA_CHANGE:
2487         case SDEV_EVT_INQUIRY_CHANGE_REPORTED:
2488         case SDEV_EVT_CAPACITY_CHANGE_REPORTED:
2489         case SDEV_EVT_SOFT_THRESHOLD_REACHED_REPORTED:
2490         case SDEV_EVT_MODE_PARAMETER_CHANGE_REPORTED:
2491         case SDEV_EVT_LUN_CHANGE_REPORTED:
2492         case SDEV_EVT_ALUA_STATE_CHANGE_REPORTED:
2493         case SDEV_EVT_POWER_ON_RESET_OCCURRED:
2494         default:
2495                 /* do nothing */
2496                 break;
2497         }
2498
2499         return evt;
2500 }
2501 EXPORT_SYMBOL_GPL(sdev_evt_alloc);
2502
2503 /**
2504  *      sdev_evt_send_simple - send asserted event to uevent thread
2505  *      @sdev: scsi_device event occurred on
2506  *      @evt_type: type of event to send
2507  *      @gfpflags: GFP flags for allocation
2508  *
2509  *      Assert scsi device event asynchronously, given an event type.
2510  */
2511 void sdev_evt_send_simple(struct scsi_device *sdev,
2512                           enum scsi_device_event evt_type, gfp_t gfpflags)
2513 {
2514         struct scsi_event *evt = sdev_evt_alloc(evt_type, gfpflags);
2515         if (!evt) {
2516                 sdev_printk(KERN_ERR, sdev, "event %d eaten due to OOM\n",
2517                             evt_type);
2518                 return;
2519         }
2520
2521         sdev_evt_send(sdev, evt);
2522 }
2523 EXPORT_SYMBOL_GPL(sdev_evt_send_simple);
2524
2525 /**
2526  *      scsi_device_quiesce - Block user issued commands.
2527  *      @sdev:  scsi device to quiesce.
2528  *
2529  *      This works by trying to transition to the SDEV_QUIESCE state
2530  *      (which must be a legal transition).  When the device is in this
2531  *      state, only special requests will be accepted, all others will
2532  *      be deferred.  Since special requests may also be requeued requests,
2533  *      a successful return doesn't guarantee the device will be
2534  *      totally quiescent.
2535  *
2536  *      Must be called with user context, may sleep.
2537  *
2538  *      Returns zero if unsuccessful or an error if not.
2539  */
2540 int
2541 scsi_device_quiesce(struct scsi_device *sdev)
2542 {
2543         struct request_queue *q = sdev->request_queue;
2544         int err;
2545
2546         /*
2547          * It is allowed to call scsi_device_quiesce() multiple times from
2548          * the same context but concurrent scsi_device_quiesce() calls are
2549          * not allowed.
2550          */
2551         WARN_ON_ONCE(sdev->quiesced_by && sdev->quiesced_by != current);
2552
2553         if (sdev->quiesced_by == current)
2554                 return 0;
2555
2556         blk_set_pm_only(q);
2557
2558         blk_mq_freeze_queue(q);
2559         /*
2560          * Ensure that the effect of blk_set_pm_only() will be visible
2561          * for percpu_ref_tryget() callers that occur after the queue
2562          * unfreeze even if the queue was already frozen before this function
2563          * was called. See also https://lwn.net/Articles/573497/.
2564          */
2565         synchronize_rcu();
2566         blk_mq_unfreeze_queue(q);
2567
2568         mutex_lock(&sdev->state_mutex);
2569         err = scsi_device_set_state(sdev, SDEV_QUIESCE);
2570         if (err == 0)
2571                 sdev->quiesced_by = current;
2572         else
2573                 blk_clear_pm_only(q);
2574         mutex_unlock(&sdev->state_mutex);
2575
2576         return err;
2577 }
2578 EXPORT_SYMBOL(scsi_device_quiesce);
2579
2580 /**
2581  *      scsi_device_resume - Restart user issued commands to a quiesced device.
2582  *      @sdev:  scsi device to resume.
2583  *
2584  *      Moves the device from quiesced back to running and restarts the
2585  *      queues.
2586  *
2587  *      Must be called with user context, may sleep.
2588  */
2589 void scsi_device_resume(struct scsi_device *sdev)
2590 {
2591         /* check if the device state was mutated prior to resume, and if
2592          * so assume the state is being managed elsewhere (for example
2593          * device deleted during suspend)
2594          */
2595         mutex_lock(&sdev->state_mutex);
2596         if (sdev->quiesced_by) {
2597                 sdev->quiesced_by = NULL;
2598                 blk_clear_pm_only(sdev->request_queue);
2599         }
2600         if (sdev->sdev_state == SDEV_QUIESCE)
2601                 scsi_device_set_state(sdev, SDEV_RUNNING);
2602         mutex_unlock(&sdev->state_mutex);
2603 }
2604 EXPORT_SYMBOL(scsi_device_resume);
2605
2606 static void
2607 device_quiesce_fn(struct scsi_device *sdev, void *data)
2608 {
2609         scsi_device_quiesce(sdev);
2610 }
2611
2612 void
2613 scsi_target_quiesce(struct scsi_target *starget)
2614 {
2615         starget_for_each_device(starget, NULL, device_quiesce_fn);
2616 }
2617 EXPORT_SYMBOL(scsi_target_quiesce);
2618
2619 static void
2620 device_resume_fn(struct scsi_device *sdev, void *data)
2621 {
2622         scsi_device_resume(sdev);
2623 }
2624
2625 void
2626 scsi_target_resume(struct scsi_target *starget)
2627 {
2628         starget_for_each_device(starget, NULL, device_resume_fn);
2629 }
2630 EXPORT_SYMBOL(scsi_target_resume);
2631
2632 /**
2633  * scsi_internal_device_block_nowait - try to transition to the SDEV_BLOCK state
2634  * @sdev: device to block
2635  *
2636  * Pause SCSI command processing on the specified device. Does not sleep.
2637  *
2638  * Returns zero if successful or a negative error code upon failure.
2639  *
2640  * Notes:
2641  * This routine transitions the device to the SDEV_BLOCK state (which must be
2642  * a legal transition). When the device is in this state, command processing
2643  * is paused until the device leaves the SDEV_BLOCK state. See also
2644  * scsi_internal_device_unblock_nowait().
2645  */
2646 int scsi_internal_device_block_nowait(struct scsi_device *sdev)
2647 {
2648         struct request_queue *q = sdev->request_queue;
2649         int err = 0;
2650
2651         err = scsi_device_set_state(sdev, SDEV_BLOCK);
2652         if (err) {
2653                 err = scsi_device_set_state(sdev, SDEV_CREATED_BLOCK);
2654
2655                 if (err)
2656                         return err;
2657         }
2658
2659         /*
2660          * The device has transitioned to SDEV_BLOCK.  Stop the
2661          * block layer from calling the midlayer with this device's
2662          * request queue.
2663          */
2664         blk_mq_quiesce_queue_nowait(q);
2665         return 0;
2666 }
2667 EXPORT_SYMBOL_GPL(scsi_internal_device_block_nowait);
2668
2669 /**
2670  * scsi_internal_device_block - try to transition to the SDEV_BLOCK state
2671  * @sdev: device to block
2672  *
2673  * Pause SCSI command processing on the specified device and wait until all
2674  * ongoing scsi_request_fn() / scsi_queue_rq() calls have finished. May sleep.
2675  *
2676  * Returns zero if successful or a negative error code upon failure.
2677  *
2678  * Note:
2679  * This routine transitions the device to the SDEV_BLOCK state (which must be
2680  * a legal transition). When the device is in this state, command processing
2681  * is paused until the device leaves the SDEV_BLOCK state. See also
2682  * scsi_internal_device_unblock().
2683  */
2684 static int scsi_internal_device_block(struct scsi_device *sdev)
2685 {
2686         struct request_queue *q = sdev->request_queue;
2687         int err;
2688
2689         mutex_lock(&sdev->state_mutex);
2690         err = scsi_internal_device_block_nowait(sdev);
2691         if (err == 0)
2692                 blk_mq_quiesce_queue(q);
2693         mutex_unlock(&sdev->state_mutex);
2694
2695         return err;
2696 }
2697
2698 void scsi_start_queue(struct scsi_device *sdev)
2699 {
2700         struct request_queue *q = sdev->request_queue;
2701
2702         blk_mq_unquiesce_queue(q);
2703 }
2704
2705 /**
2706  * scsi_internal_device_unblock_nowait - resume a device after a block request
2707  * @sdev:       device to resume
2708  * @new_state:  state to set the device to after unblocking
2709  *
2710  * Restart the device queue for a previously suspended SCSI device. Does not
2711  * sleep.
2712  *
2713  * Returns zero if successful or a negative error code upon failure.
2714  *
2715  * Notes:
2716  * This routine transitions the device to the SDEV_RUNNING state or to one of
2717  * the offline states (which must be a legal transition) allowing the midlayer
2718  * to goose the queue for this device.
2719  */
2720 int scsi_internal_device_unblock_nowait(struct scsi_device *sdev,
2721                                         enum scsi_device_state new_state)
2722 {
2723         switch (new_state) {
2724         case SDEV_RUNNING:
2725         case SDEV_TRANSPORT_OFFLINE:
2726                 break;
2727         default:
2728                 return -EINVAL;
2729         }
2730
2731         /*
2732          * Try to transition the scsi device to SDEV_RUNNING or one of the
2733          * offlined states and goose the device queue if successful.
2734          */
2735         switch (sdev->sdev_state) {
2736         case SDEV_BLOCK:
2737         case SDEV_TRANSPORT_OFFLINE:
2738                 sdev->sdev_state = new_state;
2739                 break;
2740         case SDEV_CREATED_BLOCK:
2741                 if (new_state == SDEV_TRANSPORT_OFFLINE ||
2742                     new_state == SDEV_OFFLINE)
2743                         sdev->sdev_state = new_state;
2744                 else
2745                         sdev->sdev_state = SDEV_CREATED;
2746                 break;
2747         case SDEV_CANCEL:
2748         case SDEV_OFFLINE:
2749                 break;
2750         default:
2751                 return -EINVAL;
2752         }
2753         scsi_start_queue(sdev);
2754
2755         return 0;
2756 }
2757 EXPORT_SYMBOL_GPL(scsi_internal_device_unblock_nowait);
2758
2759 /**
2760  * scsi_internal_device_unblock - resume a device after a block request
2761  * @sdev:       device to resume
2762  * @new_state:  state to set the device to after unblocking
2763  *
2764  * Restart the device queue for a previously suspended SCSI device. May sleep.
2765  *
2766  * Returns zero if successful or a negative error code upon failure.
2767  *
2768  * Notes:
2769  * This routine transitions the device to the SDEV_RUNNING state or to one of
2770  * the offline states (which must be a legal transition) allowing the midlayer
2771  * to goose the queue for this device.
2772  */
2773 static int scsi_internal_device_unblock(struct scsi_device *sdev,
2774                                         enum scsi_device_state new_state)
2775 {
2776         int ret;
2777
2778         mutex_lock(&sdev->state_mutex);
2779         ret = scsi_internal_device_unblock_nowait(sdev, new_state);
2780         mutex_unlock(&sdev->state_mutex);
2781
2782         return ret;
2783 }
2784
2785 static void
2786 device_block(struct scsi_device *sdev, void *data)
2787 {
2788         int ret;
2789
2790         ret = scsi_internal_device_block(sdev);
2791
2792         WARN_ONCE(ret, "scsi_internal_device_block(%s) failed: ret = %d\n",
2793                   dev_name(&sdev->sdev_gendev), ret);
2794 }
2795
2796 static int
2797 target_block(struct device *dev, void *data)
2798 {
2799         if (scsi_is_target_device(dev))
2800                 starget_for_each_device(to_scsi_target(dev), NULL,
2801                                         device_block);
2802         return 0;
2803 }
2804
2805 void
2806 scsi_target_block(struct device *dev)
2807 {
2808         if (scsi_is_target_device(dev))
2809                 starget_for_each_device(to_scsi_target(dev), NULL,
2810                                         device_block);
2811         else
2812                 device_for_each_child(dev, NULL, target_block);
2813 }
2814 EXPORT_SYMBOL_GPL(scsi_target_block);
2815
2816 static void
2817 device_unblock(struct scsi_device *sdev, void *data)
2818 {
2819         scsi_internal_device_unblock(sdev, *(enum scsi_device_state *)data);
2820 }
2821
2822 static int
2823 target_unblock(struct device *dev, void *data)
2824 {
2825         if (scsi_is_target_device(dev))
2826                 starget_for_each_device(to_scsi_target(dev), data,
2827                                         device_unblock);
2828         return 0;
2829 }
2830
2831 void
2832 scsi_target_unblock(struct device *dev, enum scsi_device_state new_state)
2833 {
2834         if (scsi_is_target_device(dev))
2835                 starget_for_each_device(to_scsi_target(dev), &new_state,
2836                                         device_unblock);
2837         else
2838                 device_for_each_child(dev, &new_state, target_unblock);
2839 }
2840 EXPORT_SYMBOL_GPL(scsi_target_unblock);
2841
2842 int
2843 scsi_host_block(struct Scsi_Host *shost)
2844 {
2845         struct scsi_device *sdev;
2846         int ret = 0;
2847
2848         /*
2849          * Call scsi_internal_device_block_nowait so we can avoid
2850          * calling synchronize_rcu() for each LUN.
2851          */
2852         shost_for_each_device(sdev, shost) {
2853                 mutex_lock(&sdev->state_mutex);
2854                 ret = scsi_internal_device_block_nowait(sdev);
2855                 mutex_unlock(&sdev->state_mutex);
2856                 if (ret) {
2857                         scsi_device_put(sdev);
2858                         break;
2859                 }
2860         }
2861
2862         /*
2863          * SCSI never enables blk-mq's BLK_MQ_F_BLOCKING flag so
2864          * calling synchronize_rcu() once is enough.
2865          */
2866         WARN_ON_ONCE(shost->tag_set.flags & BLK_MQ_F_BLOCKING);
2867
2868         if (!ret)
2869                 synchronize_rcu();
2870
2871         return ret;
2872 }
2873 EXPORT_SYMBOL_GPL(scsi_host_block);
2874
2875 int
2876 scsi_host_unblock(struct Scsi_Host *shost, int new_state)
2877 {
2878         struct scsi_device *sdev;
2879         int ret = 0;
2880
2881         shost_for_each_device(sdev, shost) {
2882                 ret = scsi_internal_device_unblock(sdev, new_state);
2883                 if (ret) {
2884                         scsi_device_put(sdev);
2885                         break;
2886                 }
2887         }
2888         return ret;
2889 }
2890 EXPORT_SYMBOL_GPL(scsi_host_unblock);
2891
2892 /**
2893  * scsi_kmap_atomic_sg - find and atomically map an sg-elemnt
2894  * @sgl:        scatter-gather list
2895  * @sg_count:   number of segments in sg
2896  * @offset:     offset in bytes into sg, on return offset into the mapped area
2897  * @len:        bytes to map, on return number of bytes mapped
2898  *
2899  * Returns virtual address of the start of the mapped page
2900  */
2901 void *scsi_kmap_atomic_sg(struct scatterlist *sgl, int sg_count,
2902                           size_t *offset, size_t *len)
2903 {
2904         int i;
2905         size_t sg_len = 0, len_complete = 0;
2906         struct scatterlist *sg;
2907         struct page *page;
2908
2909         WARN_ON(!irqs_disabled());
2910
2911         for_each_sg(sgl, sg, sg_count, i) {
2912                 len_complete = sg_len; /* Complete sg-entries */
2913                 sg_len += sg->length;
2914                 if (sg_len > *offset)
2915                         break;
2916         }
2917
2918         if (unlikely(i == sg_count)) {
2919                 printk(KERN_ERR "%s: Bytes in sg: %zu, requested offset %zu, "
2920                         "elements %d\n",
2921                        __func__, sg_len, *offset, sg_count);
2922                 WARN_ON(1);
2923                 return NULL;
2924         }
2925
2926         /* Offset starting from the beginning of first page in this sg-entry */
2927         *offset = *offset - len_complete + sg->offset;
2928
2929         /* Assumption: contiguous pages can be accessed as "page + i" */
2930         page = nth_page(sg_page(sg), (*offset >> PAGE_SHIFT));
2931         *offset &= ~PAGE_MASK;
2932
2933         /* Bytes in this sg-entry from *offset to the end of the page */
2934         sg_len = PAGE_SIZE - *offset;
2935         if (*len > sg_len)
2936                 *len = sg_len;
2937
2938         return kmap_atomic(page);
2939 }
2940 EXPORT_SYMBOL(scsi_kmap_atomic_sg);
2941
2942 /**
2943  * scsi_kunmap_atomic_sg - atomically unmap a virtual address, previously mapped with scsi_kmap_atomic_sg
2944  * @virt:       virtual address to be unmapped
2945  */
2946 void scsi_kunmap_atomic_sg(void *virt)
2947 {
2948         kunmap_atomic(virt);
2949 }
2950 EXPORT_SYMBOL(scsi_kunmap_atomic_sg);
2951
2952 void sdev_disable_disk_events(struct scsi_device *sdev)
2953 {
2954         atomic_inc(&sdev->disk_events_disable_depth);
2955 }
2956 EXPORT_SYMBOL(sdev_disable_disk_events);
2957
2958 void sdev_enable_disk_events(struct scsi_device *sdev)
2959 {
2960         if (WARN_ON_ONCE(atomic_read(&sdev->disk_events_disable_depth) <= 0))
2961                 return;
2962         atomic_dec(&sdev->disk_events_disable_depth);
2963 }
2964 EXPORT_SYMBOL(sdev_enable_disk_events);
2965
2966 /**
2967  * scsi_vpd_lun_id - return a unique device identification
2968  * @sdev: SCSI device
2969  * @id:   buffer for the identification
2970  * @id_len:  length of the buffer
2971  *
2972  * Copies a unique device identification into @id based
2973  * on the information in the VPD page 0x83 of the device.
2974  * The string will be formatted as a SCSI name string.
2975  *
2976  * Returns the length of the identification or error on failure.
2977  * If the identifier is longer than the supplied buffer the actual
2978  * identifier length is returned and the buffer is not zero-padded.
2979  */
2980 int scsi_vpd_lun_id(struct scsi_device *sdev, char *id, size_t id_len)
2981 {
2982         u8 cur_id_type = 0xff;
2983         u8 cur_id_size = 0;
2984         const unsigned char *d, *cur_id_str;
2985         const struct scsi_vpd *vpd_pg83;
2986         int id_size = -EINVAL;
2987
2988         rcu_read_lock();
2989         vpd_pg83 = rcu_dereference(sdev->vpd_pg83);
2990         if (!vpd_pg83) {
2991                 rcu_read_unlock();
2992                 return -ENXIO;
2993         }
2994
2995         /*
2996          * Look for the correct descriptor.
2997          * Order of preference for lun descriptor:
2998          * - SCSI name string
2999          * - NAA IEEE Registered Extended
3000          * - EUI-64 based 16-byte
3001          * - EUI-64 based 12-byte
3002          * - NAA IEEE Registered
3003          * - NAA IEEE Extended
3004          * - T10 Vendor ID
3005          * as longer descriptors reduce the likelyhood
3006          * of identification clashes.
3007          */
3008
3009         /* The id string must be at least 20 bytes + terminating NULL byte */
3010         if (id_len < 21) {
3011                 rcu_read_unlock();
3012                 return -EINVAL;
3013         }
3014
3015         memset(id, 0, id_len);
3016         d = vpd_pg83->data + 4;
3017         while (d < vpd_pg83->data + vpd_pg83->len) {
3018                 /* Skip designators not referring to the LUN */
3019                 if ((d[1] & 0x30) != 0x00)
3020                         goto next_desig;
3021
3022                 switch (d[1] & 0xf) {
3023                 case 0x1:
3024                         /* T10 Vendor ID */
3025                         if (cur_id_size > d[3])
3026                                 break;
3027                         /* Prefer anything */
3028                         if (cur_id_type > 0x01 && cur_id_type != 0xff)
3029                                 break;
3030                         cur_id_size = d[3];
3031                         if (cur_id_size + 4 > id_len)
3032                                 cur_id_size = id_len - 4;
3033                         cur_id_str = d + 4;
3034                         cur_id_type = d[1] & 0xf;
3035                         id_size = snprintf(id, id_len, "t10.%*pE",
3036                                            cur_id_size, cur_id_str);
3037                         break;
3038                 case 0x2:
3039                         /* EUI-64 */
3040                         if (cur_id_size > d[3])
3041                                 break;
3042                         /* Prefer NAA IEEE Registered Extended */
3043                         if (cur_id_type == 0x3 &&
3044                             cur_id_size == d[3])
3045                                 break;
3046                         cur_id_size = d[3];
3047                         cur_id_str = d + 4;
3048                         cur_id_type = d[1] & 0xf;
3049                         switch (cur_id_size) {
3050                         case 8:
3051                                 id_size = snprintf(id, id_len,
3052                                                    "eui.%8phN",
3053                                                    cur_id_str);
3054                                 break;
3055                         case 12:
3056                                 id_size = snprintf(id, id_len,
3057                                                    "eui.%12phN",
3058                                                    cur_id_str);
3059                                 break;
3060                         case 16:
3061                                 id_size = snprintf(id, id_len,
3062                                                    "eui.%16phN",
3063                                                    cur_id_str);
3064                                 break;
3065                         default:
3066                                 cur_id_size = 0;
3067                                 break;
3068                         }
3069                         break;
3070                 case 0x3:
3071                         /* NAA */
3072                         if (cur_id_size > d[3])
3073                                 break;
3074                         cur_id_size = d[3];
3075                         cur_id_str = d + 4;
3076                         cur_id_type = d[1] & 0xf;
3077                         switch (cur_id_size) {
3078                         case 8:
3079                                 id_size = snprintf(id, id_len,
3080                                                    "naa.%8phN",
3081                                                    cur_id_str);
3082                                 break;
3083                         case 16:
3084                                 id_size = snprintf(id, id_len,
3085                                                    "naa.%16phN",
3086                                                    cur_id_str);
3087                                 break;
3088                         default:
3089                                 cur_id_size = 0;
3090                                 break;
3091                         }
3092                         break;
3093                 case 0x8:
3094                         /* SCSI name string */
3095                         if (cur_id_size + 4 > d[3])
3096                                 break;
3097                         /* Prefer others for truncated descriptor */
3098                         if (cur_id_size && d[3] > id_len)
3099                                 break;
3100                         cur_id_size = id_size = d[3];
3101                         cur_id_str = d + 4;
3102                         cur_id_type = d[1] & 0xf;
3103                         if (cur_id_size >= id_len)
3104                                 cur_id_size = id_len - 1;
3105                         memcpy(id, cur_id_str, cur_id_size);
3106                         /* Decrease priority for truncated descriptor */
3107                         if (cur_id_size != id_size)
3108                                 cur_id_size = 6;
3109                         break;
3110                 default:
3111                         break;
3112                 }
3113 next_desig:
3114                 d += d[3] + 4;
3115         }
3116         rcu_read_unlock();
3117
3118         return id_size;
3119 }
3120 EXPORT_SYMBOL(scsi_vpd_lun_id);
3121
3122 /*
3123  * scsi_vpd_tpg_id - return a target port group identifier
3124  * @sdev: SCSI device
3125  *
3126  * Returns the Target Port Group identifier from the information
3127  * froom VPD page 0x83 of the device.
3128  *
3129  * Returns the identifier or error on failure.
3130  */
3131 int scsi_vpd_tpg_id(struct scsi_device *sdev, int *rel_id)
3132 {
3133         const unsigned char *d;
3134         const struct scsi_vpd *vpd_pg83;
3135         int group_id = -EAGAIN, rel_port = -1;
3136
3137         rcu_read_lock();
3138         vpd_pg83 = rcu_dereference(sdev->vpd_pg83);
3139         if (!vpd_pg83) {
3140                 rcu_read_unlock();
3141                 return -ENXIO;
3142         }
3143
3144         d = vpd_pg83->data + 4;
3145         while (d < vpd_pg83->data + vpd_pg83->len) {
3146                 switch (d[1] & 0xf) {
3147                 case 0x4:
3148                         /* Relative target port */
3149                         rel_port = get_unaligned_be16(&d[6]);
3150                         break;
3151                 case 0x5:
3152                         /* Target port group */
3153                         group_id = get_unaligned_be16(&d[6]);
3154                         break;
3155                 default:
3156                         break;
3157                 }
3158                 d += d[3] + 4;
3159         }
3160         rcu_read_unlock();
3161
3162         if (group_id >= 0 && rel_id && rel_port != -1)
3163                 *rel_id = rel_port;
3164
3165         return group_id;
3166 }
3167 EXPORT_SYMBOL(scsi_vpd_tpg_id);