Merge branch 'pm-cpufreq'
[linux-2.6-microblaze.git] / drivers / scsi / scsi_lib.c
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 1999 Eric Youngdale
4  * Copyright (C) 2014 Christoph Hellwig
5  *
6  *  SCSI queueing library.
7  *      Initial versions: Eric Youngdale (eric@andante.org).
8  *                        Based upon conversations with large numbers
9  *                        of people at Linux Expo.
10  */
11
12 #include <linux/bio.h>
13 #include <linux/bitops.h>
14 #include <linux/blkdev.h>
15 #include <linux/completion.h>
16 #include <linux/kernel.h>
17 #include <linux/export.h>
18 #include <linux/init.h>
19 #include <linux/pci.h>
20 #include <linux/delay.h>
21 #include <linux/hardirq.h>
22 #include <linux/scatterlist.h>
23 #include <linux/blk-mq.h>
24 #include <linux/ratelimit.h>
25 #include <asm/unaligned.h>
26
27 #include <scsi/scsi.h>
28 #include <scsi/scsi_cmnd.h>
29 #include <scsi/scsi_dbg.h>
30 #include <scsi/scsi_device.h>
31 #include <scsi/scsi_driver.h>
32 #include <scsi/scsi_eh.h>
33 #include <scsi/scsi_host.h>
34 #include <scsi/scsi_transport.h> /* __scsi_init_queue() */
35 #include <scsi/scsi_dh.h>
36
37 #include <trace/events/scsi.h>
38
39 #include "scsi_debugfs.h"
40 #include "scsi_priv.h"
41 #include "scsi_logging.h"
42
43 /*
44  * Size of integrity metadata is usually small, 1 inline sg should
45  * cover normal cases.
46  */
47 #ifdef CONFIG_ARCH_NO_SG_CHAIN
48 #define  SCSI_INLINE_PROT_SG_CNT  0
49 #define  SCSI_INLINE_SG_CNT  0
50 #else
51 #define  SCSI_INLINE_PROT_SG_CNT  1
52 #define  SCSI_INLINE_SG_CNT  2
53 #endif
54
55 static struct kmem_cache *scsi_sense_cache;
56 static struct kmem_cache *scsi_sense_isadma_cache;
57 static DEFINE_MUTEX(scsi_sense_cache_mutex);
58
59 static void scsi_mq_uninit_cmd(struct scsi_cmnd *cmd);
60
61 static inline struct kmem_cache *
62 scsi_select_sense_cache(bool unchecked_isa_dma)
63 {
64         return unchecked_isa_dma ? scsi_sense_isadma_cache : scsi_sense_cache;
65 }
66
67 static void scsi_free_sense_buffer(bool unchecked_isa_dma,
68                                    unsigned char *sense_buffer)
69 {
70         kmem_cache_free(scsi_select_sense_cache(unchecked_isa_dma),
71                         sense_buffer);
72 }
73
74 static unsigned char *scsi_alloc_sense_buffer(bool unchecked_isa_dma,
75         gfp_t gfp_mask, int numa_node)
76 {
77         return kmem_cache_alloc_node(scsi_select_sense_cache(unchecked_isa_dma),
78                                      gfp_mask, numa_node);
79 }
80
81 int scsi_init_sense_cache(struct Scsi_Host *shost)
82 {
83         struct kmem_cache *cache;
84         int ret = 0;
85
86         mutex_lock(&scsi_sense_cache_mutex);
87         cache = scsi_select_sense_cache(shost->unchecked_isa_dma);
88         if (cache)
89                 goto exit;
90
91         if (shost->unchecked_isa_dma) {
92                 scsi_sense_isadma_cache =
93                         kmem_cache_create("scsi_sense_cache(DMA)",
94                                 SCSI_SENSE_BUFFERSIZE, 0,
95                                 SLAB_HWCACHE_ALIGN | SLAB_CACHE_DMA, NULL);
96                 if (!scsi_sense_isadma_cache)
97                         ret = -ENOMEM;
98         } else {
99                 scsi_sense_cache =
100                         kmem_cache_create_usercopy("scsi_sense_cache",
101                                 SCSI_SENSE_BUFFERSIZE, 0, SLAB_HWCACHE_ALIGN,
102                                 0, SCSI_SENSE_BUFFERSIZE, NULL);
103                 if (!scsi_sense_cache)
104                         ret = -ENOMEM;
105         }
106  exit:
107         mutex_unlock(&scsi_sense_cache_mutex);
108         return ret;
109 }
110
111 /*
112  * When to reinvoke queueing after a resource shortage. It's 3 msecs to
113  * not change behaviour from the previous unplug mechanism, experimentation
114  * may prove this needs changing.
115  */
116 #define SCSI_QUEUE_DELAY        3
117
118 static void
119 scsi_set_blocked(struct scsi_cmnd *cmd, int reason)
120 {
121         struct Scsi_Host *host = cmd->device->host;
122         struct scsi_device *device = cmd->device;
123         struct scsi_target *starget = scsi_target(device);
124
125         /*
126          * Set the appropriate busy bit for the device/host.
127          *
128          * If the host/device isn't busy, assume that something actually
129          * completed, and that we should be able to queue a command now.
130          *
131          * Note that the prior mid-layer assumption that any host could
132          * always queue at least one command is now broken.  The mid-layer
133          * will implement a user specifiable stall (see
134          * scsi_host.max_host_blocked and scsi_device.max_device_blocked)
135          * if a command is requeued with no other commands outstanding
136          * either for the device or for the host.
137          */
138         switch (reason) {
139         case SCSI_MLQUEUE_HOST_BUSY:
140                 atomic_set(&host->host_blocked, host->max_host_blocked);
141                 break;
142         case SCSI_MLQUEUE_DEVICE_BUSY:
143         case SCSI_MLQUEUE_EH_RETRY:
144                 atomic_set(&device->device_blocked,
145                            device->max_device_blocked);
146                 break;
147         case SCSI_MLQUEUE_TARGET_BUSY:
148                 atomic_set(&starget->target_blocked,
149                            starget->max_target_blocked);
150                 break;
151         }
152 }
153
154 static void scsi_mq_requeue_cmd(struct scsi_cmnd *cmd)
155 {
156         if (cmd->request->rq_flags & RQF_DONTPREP) {
157                 cmd->request->rq_flags &= ~RQF_DONTPREP;
158                 scsi_mq_uninit_cmd(cmd);
159         } else {
160                 WARN_ON_ONCE(true);
161         }
162         blk_mq_requeue_request(cmd->request, true);
163 }
164
165 /**
166  * __scsi_queue_insert - private queue insertion
167  * @cmd: The SCSI command being requeued
168  * @reason:  The reason for the requeue
169  * @unbusy: Whether the queue should be unbusied
170  *
171  * This is a private queue insertion.  The public interface
172  * scsi_queue_insert() always assumes the queue should be unbusied
173  * because it's always called before the completion.  This function is
174  * for a requeue after completion, which should only occur in this
175  * file.
176  */
177 static void __scsi_queue_insert(struct scsi_cmnd *cmd, int reason, bool unbusy)
178 {
179         struct scsi_device *device = cmd->device;
180
181         SCSI_LOG_MLQUEUE(1, scmd_printk(KERN_INFO, cmd,
182                 "Inserting command %p into mlqueue\n", cmd));
183
184         scsi_set_blocked(cmd, reason);
185
186         /*
187          * Decrement the counters, since these commands are no longer
188          * active on the host/device.
189          */
190         if (unbusy)
191                 scsi_device_unbusy(device, cmd);
192
193         /*
194          * Requeue this command.  It will go before all other commands
195          * that are already in the queue. Schedule requeue work under
196          * lock such that the kblockd_schedule_work() call happens
197          * before blk_cleanup_queue() finishes.
198          */
199         cmd->result = 0;
200
201         blk_mq_requeue_request(cmd->request, true);
202 }
203
204 /**
205  * scsi_queue_insert - Reinsert a command in the queue.
206  * @cmd:    command that we are adding to queue.
207  * @reason: why we are inserting command to queue.
208  *
209  * We do this for one of two cases. Either the host is busy and it cannot accept
210  * any more commands for the time being, or the device returned QUEUE_FULL and
211  * can accept no more commands.
212  *
213  * Context: This could be called either from an interrupt context or a normal
214  * process context.
215  */
216 void scsi_queue_insert(struct scsi_cmnd *cmd, int reason)
217 {
218         __scsi_queue_insert(cmd, reason, true);
219 }
220
221
222 /**
223  * __scsi_execute - insert request and wait for the result
224  * @sdev:       scsi device
225  * @cmd:        scsi command
226  * @data_direction: data direction
227  * @buffer:     data buffer
228  * @bufflen:    len of buffer
229  * @sense:      optional sense buffer
230  * @sshdr:      optional decoded sense header
231  * @timeout:    request timeout in seconds
232  * @retries:    number of times to retry request
233  * @flags:      flags for ->cmd_flags
234  * @rq_flags:   flags for ->rq_flags
235  * @resid:      optional residual length
236  *
237  * Returns the scsi_cmnd result field if a command was executed, or a negative
238  * Linux error code if we didn't get that far.
239  */
240 int __scsi_execute(struct scsi_device *sdev, const unsigned char *cmd,
241                  int data_direction, void *buffer, unsigned bufflen,
242                  unsigned char *sense, struct scsi_sense_hdr *sshdr,
243                  int timeout, int retries, u64 flags, req_flags_t rq_flags,
244                  int *resid)
245 {
246         struct request *req;
247         struct scsi_request *rq;
248         int ret = DRIVER_ERROR << 24;
249
250         req = blk_get_request(sdev->request_queue,
251                         data_direction == DMA_TO_DEVICE ?
252                         REQ_OP_SCSI_OUT : REQ_OP_SCSI_IN, BLK_MQ_REQ_PREEMPT);
253         if (IS_ERR(req))
254                 return ret;
255         rq = scsi_req(req);
256
257         if (bufflen &&  blk_rq_map_kern(sdev->request_queue, req,
258                                         buffer, bufflen, GFP_NOIO))
259                 goto out;
260
261         rq->cmd_len = COMMAND_SIZE(cmd[0]);
262         memcpy(rq->cmd, cmd, rq->cmd_len);
263         rq->retries = retries;
264         req->timeout = timeout;
265         req->cmd_flags |= flags;
266         req->rq_flags |= rq_flags | RQF_QUIET;
267
268         /*
269          * head injection *required* here otherwise quiesce won't work
270          */
271         blk_execute_rq(req->q, NULL, req, 1);
272
273         /*
274          * Some devices (USB mass-storage in particular) may transfer
275          * garbage data together with a residue indicating that the data
276          * is invalid.  Prevent the garbage from being misinterpreted
277          * and prevent security leaks by zeroing out the excess data.
278          */
279         if (unlikely(rq->resid_len > 0 && rq->resid_len <= bufflen))
280                 memset(buffer + (bufflen - rq->resid_len), 0, rq->resid_len);
281
282         if (resid)
283                 *resid = rq->resid_len;
284         if (sense && rq->sense_len)
285                 memcpy(sense, rq->sense, SCSI_SENSE_BUFFERSIZE);
286         if (sshdr)
287                 scsi_normalize_sense(rq->sense, rq->sense_len, sshdr);
288         ret = rq->result;
289  out:
290         blk_put_request(req);
291
292         return ret;
293 }
294 EXPORT_SYMBOL(__scsi_execute);
295
296 /*
297  * Wake up the error handler if necessary. Avoid as follows that the error
298  * handler is not woken up if host in-flight requests number ==
299  * shost->host_failed: use call_rcu() in scsi_eh_scmd_add() in combination
300  * with an RCU read lock in this function to ensure that this function in
301  * its entirety either finishes before scsi_eh_scmd_add() increases the
302  * host_failed counter or that it notices the shost state change made by
303  * scsi_eh_scmd_add().
304  */
305 static void scsi_dec_host_busy(struct Scsi_Host *shost, struct scsi_cmnd *cmd)
306 {
307         unsigned long flags;
308
309         rcu_read_lock();
310         __clear_bit(SCMD_STATE_INFLIGHT, &cmd->state);
311         if (unlikely(scsi_host_in_recovery(shost))) {
312                 spin_lock_irqsave(shost->host_lock, flags);
313                 if (shost->host_failed || shost->host_eh_scheduled)
314                         scsi_eh_wakeup(shost);
315                 spin_unlock_irqrestore(shost->host_lock, flags);
316         }
317         rcu_read_unlock();
318 }
319
320 void scsi_device_unbusy(struct scsi_device *sdev, struct scsi_cmnd *cmd)
321 {
322         struct Scsi_Host *shost = sdev->host;
323         struct scsi_target *starget = scsi_target(sdev);
324
325         scsi_dec_host_busy(shost, cmd);
326
327         if (starget->can_queue > 0)
328                 atomic_dec(&starget->target_busy);
329
330         atomic_dec(&sdev->device_busy);
331 }
332
333 static void scsi_kick_queue(struct request_queue *q)
334 {
335         blk_mq_run_hw_queues(q, false);
336 }
337
338 /*
339  * Called for single_lun devices on IO completion. Clear starget_sdev_user,
340  * and call blk_run_queue for all the scsi_devices on the target -
341  * including current_sdev first.
342  *
343  * Called with *no* scsi locks held.
344  */
345 static void scsi_single_lun_run(struct scsi_device *current_sdev)
346 {
347         struct Scsi_Host *shost = current_sdev->host;
348         struct scsi_device *sdev, *tmp;
349         struct scsi_target *starget = scsi_target(current_sdev);
350         unsigned long flags;
351
352         spin_lock_irqsave(shost->host_lock, flags);
353         starget->starget_sdev_user = NULL;
354         spin_unlock_irqrestore(shost->host_lock, flags);
355
356         /*
357          * Call blk_run_queue for all LUNs on the target, starting with
358          * current_sdev. We race with others (to set starget_sdev_user),
359          * but in most cases, we will be first. Ideally, each LU on the
360          * target would get some limited time or requests on the target.
361          */
362         scsi_kick_queue(current_sdev->request_queue);
363
364         spin_lock_irqsave(shost->host_lock, flags);
365         if (starget->starget_sdev_user)
366                 goto out;
367         list_for_each_entry_safe(sdev, tmp, &starget->devices,
368                         same_target_siblings) {
369                 if (sdev == current_sdev)
370                         continue;
371                 if (scsi_device_get(sdev))
372                         continue;
373
374                 spin_unlock_irqrestore(shost->host_lock, flags);
375                 scsi_kick_queue(sdev->request_queue);
376                 spin_lock_irqsave(shost->host_lock, flags);
377
378                 scsi_device_put(sdev);
379         }
380  out:
381         spin_unlock_irqrestore(shost->host_lock, flags);
382 }
383
384 static inline bool scsi_device_is_busy(struct scsi_device *sdev)
385 {
386         if (atomic_read(&sdev->device_busy) >= sdev->queue_depth)
387                 return true;
388         if (atomic_read(&sdev->device_blocked) > 0)
389                 return true;
390         return false;
391 }
392
393 static inline bool scsi_target_is_busy(struct scsi_target *starget)
394 {
395         if (starget->can_queue > 0) {
396                 if (atomic_read(&starget->target_busy) >= starget->can_queue)
397                         return true;
398                 if (atomic_read(&starget->target_blocked) > 0)
399                         return true;
400         }
401         return false;
402 }
403
404 static inline bool scsi_host_is_busy(struct Scsi_Host *shost)
405 {
406         if (atomic_read(&shost->host_blocked) > 0)
407                 return true;
408         if (shost->host_self_blocked)
409                 return true;
410         return false;
411 }
412
413 static void scsi_starved_list_run(struct Scsi_Host *shost)
414 {
415         LIST_HEAD(starved_list);
416         struct scsi_device *sdev;
417         unsigned long flags;
418
419         spin_lock_irqsave(shost->host_lock, flags);
420         list_splice_init(&shost->starved_list, &starved_list);
421
422         while (!list_empty(&starved_list)) {
423                 struct request_queue *slq;
424
425                 /*
426                  * As long as shost is accepting commands and we have
427                  * starved queues, call blk_run_queue. scsi_request_fn
428                  * drops the queue_lock and can add us back to the
429                  * starved_list.
430                  *
431                  * host_lock protects the starved_list and starved_entry.
432                  * scsi_request_fn must get the host_lock before checking
433                  * or modifying starved_list or starved_entry.
434                  */
435                 if (scsi_host_is_busy(shost))
436                         break;
437
438                 sdev = list_entry(starved_list.next,
439                                   struct scsi_device, starved_entry);
440                 list_del_init(&sdev->starved_entry);
441                 if (scsi_target_is_busy(scsi_target(sdev))) {
442                         list_move_tail(&sdev->starved_entry,
443                                        &shost->starved_list);
444                         continue;
445                 }
446
447                 /*
448                  * Once we drop the host lock, a racing scsi_remove_device()
449                  * call may remove the sdev from the starved list and destroy
450                  * it and the queue.  Mitigate by taking a reference to the
451                  * queue and never touching the sdev again after we drop the
452                  * host lock.  Note: if __scsi_remove_device() invokes
453                  * blk_cleanup_queue() before the queue is run from this
454                  * function then blk_run_queue() will return immediately since
455                  * blk_cleanup_queue() marks the queue with QUEUE_FLAG_DYING.
456                  */
457                 slq = sdev->request_queue;
458                 if (!blk_get_queue(slq))
459                         continue;
460                 spin_unlock_irqrestore(shost->host_lock, flags);
461
462                 scsi_kick_queue(slq);
463                 blk_put_queue(slq);
464
465                 spin_lock_irqsave(shost->host_lock, flags);
466         }
467         /* put any unprocessed entries back */
468         list_splice(&starved_list, &shost->starved_list);
469         spin_unlock_irqrestore(shost->host_lock, flags);
470 }
471
472 /**
473  * scsi_run_queue - Select a proper request queue to serve next.
474  * @q:  last request's queue
475  *
476  * The previous command was completely finished, start a new one if possible.
477  */
478 static void scsi_run_queue(struct request_queue *q)
479 {
480         struct scsi_device *sdev = q->queuedata;
481
482         if (scsi_target(sdev)->single_lun)
483                 scsi_single_lun_run(sdev);
484         if (!list_empty(&sdev->host->starved_list))
485                 scsi_starved_list_run(sdev->host);
486
487         blk_mq_run_hw_queues(q, false);
488 }
489
490 void scsi_requeue_run_queue(struct work_struct *work)
491 {
492         struct scsi_device *sdev;
493         struct request_queue *q;
494
495         sdev = container_of(work, struct scsi_device, requeue_work);
496         q = sdev->request_queue;
497         scsi_run_queue(q);
498 }
499
500 void scsi_run_host_queues(struct Scsi_Host *shost)
501 {
502         struct scsi_device *sdev;
503
504         shost_for_each_device(sdev, shost)
505                 scsi_run_queue(sdev->request_queue);
506 }
507
508 static void scsi_uninit_cmd(struct scsi_cmnd *cmd)
509 {
510         if (!blk_rq_is_passthrough(cmd->request)) {
511                 struct scsi_driver *drv = scsi_cmd_to_driver(cmd);
512
513                 if (drv->uninit_command)
514                         drv->uninit_command(cmd);
515         }
516 }
517
518 void scsi_free_sgtables(struct scsi_cmnd *cmd)
519 {
520         if (cmd->sdb.table.nents)
521                 sg_free_table_chained(&cmd->sdb.table,
522                                 SCSI_INLINE_SG_CNT);
523         if (scsi_prot_sg_count(cmd))
524                 sg_free_table_chained(&cmd->prot_sdb->table,
525                                 SCSI_INLINE_PROT_SG_CNT);
526 }
527 EXPORT_SYMBOL_GPL(scsi_free_sgtables);
528
529 static void scsi_mq_uninit_cmd(struct scsi_cmnd *cmd)
530 {
531         scsi_free_sgtables(cmd);
532         scsi_uninit_cmd(cmd);
533 }
534
535 static void scsi_run_queue_async(struct scsi_device *sdev)
536 {
537         if (scsi_target(sdev)->single_lun ||
538             !list_empty(&sdev->host->starved_list)) {
539                 kblockd_schedule_work(&sdev->requeue_work);
540         } else {
541                 /*
542                  * smp_mb() present in sbitmap_queue_clear() or implied in
543                  * .end_io is for ordering writing .device_busy in
544                  * scsi_device_unbusy() and reading sdev->restarts.
545                  */
546                 int old = atomic_read(&sdev->restarts);
547
548                 /*
549                  * ->restarts has to be kept as non-zero if new budget
550                  *  contention occurs.
551                  *
552                  *  No need to run queue when either another re-run
553                  *  queue wins in updating ->restarts or a new budget
554                  *  contention occurs.
555                  */
556                 if (old && atomic_cmpxchg(&sdev->restarts, old, 0) == old)
557                         blk_mq_run_hw_queues(sdev->request_queue, true);
558         }
559 }
560
561 /* Returns false when no more bytes to process, true if there are more */
562 static bool scsi_end_request(struct request *req, blk_status_t error,
563                 unsigned int bytes)
564 {
565         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
566         struct scsi_device *sdev = cmd->device;
567         struct request_queue *q = sdev->request_queue;
568
569         if (blk_update_request(req, error, bytes))
570                 return true;
571
572         if (blk_queue_add_random(q))
573                 add_disk_randomness(req->rq_disk);
574
575         if (!blk_rq_is_scsi(req)) {
576                 WARN_ON_ONCE(!(cmd->flags & SCMD_INITIALIZED));
577                 cmd->flags &= ~SCMD_INITIALIZED;
578         }
579
580         /*
581          * Calling rcu_barrier() is not necessary here because the
582          * SCSI error handler guarantees that the function called by
583          * call_rcu() has been called before scsi_end_request() is
584          * called.
585          */
586         destroy_rcu_head(&cmd->rcu);
587
588         /*
589          * In the MQ case the command gets freed by __blk_mq_end_request,
590          * so we have to do all cleanup that depends on it earlier.
591          *
592          * We also can't kick the queues from irq context, so we
593          * will have to defer it to a workqueue.
594          */
595         scsi_mq_uninit_cmd(cmd);
596
597         /*
598          * queue is still alive, so grab the ref for preventing it
599          * from being cleaned up during running queue.
600          */
601         percpu_ref_get(&q->q_usage_counter);
602
603         __blk_mq_end_request(req, error);
604
605         scsi_run_queue_async(sdev);
606
607         percpu_ref_put(&q->q_usage_counter);
608         return false;
609 }
610
611 /**
612  * scsi_result_to_blk_status - translate a SCSI result code into blk_status_t
613  * @cmd:        SCSI command
614  * @result:     scsi error code
615  *
616  * Translate a SCSI result code into a blk_status_t value. May reset the host
617  * byte of @cmd->result.
618  */
619 static blk_status_t scsi_result_to_blk_status(struct scsi_cmnd *cmd, int result)
620 {
621         switch (host_byte(result)) {
622         case DID_OK:
623                 /*
624                  * Also check the other bytes than the status byte in result
625                  * to handle the case when a SCSI LLD sets result to
626                  * DRIVER_SENSE << 24 without setting SAM_STAT_CHECK_CONDITION.
627                  */
628                 if (scsi_status_is_good(result) && (result & ~0xff) == 0)
629                         return BLK_STS_OK;
630                 return BLK_STS_IOERR;
631         case DID_TRANSPORT_FAILFAST:
632                 return BLK_STS_TRANSPORT;
633         case DID_TARGET_FAILURE:
634                 set_host_byte(cmd, DID_OK);
635                 return BLK_STS_TARGET;
636         case DID_NEXUS_FAILURE:
637                 set_host_byte(cmd, DID_OK);
638                 return BLK_STS_NEXUS;
639         case DID_ALLOC_FAILURE:
640                 set_host_byte(cmd, DID_OK);
641                 return BLK_STS_NOSPC;
642         case DID_MEDIUM_ERROR:
643                 set_host_byte(cmd, DID_OK);
644                 return BLK_STS_MEDIUM;
645         default:
646                 return BLK_STS_IOERR;
647         }
648 }
649
650 /* Helper for scsi_io_completion() when "reprep" action required. */
651 static void scsi_io_completion_reprep(struct scsi_cmnd *cmd,
652                                       struct request_queue *q)
653 {
654         /* A new command will be prepared and issued. */
655         scsi_mq_requeue_cmd(cmd);
656 }
657
658 static bool scsi_cmd_runtime_exceeced(struct scsi_cmnd *cmd)
659 {
660         struct request *req = cmd->request;
661         unsigned long wait_for;
662
663         if (cmd->allowed == SCSI_CMD_RETRIES_NO_LIMIT)
664                 return false;
665
666         wait_for = (cmd->allowed + 1) * req->timeout;
667         if (time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) {
668                 scmd_printk(KERN_ERR, cmd, "timing out command, waited %lus\n",
669                             wait_for/HZ);
670                 return true;
671         }
672         return false;
673 }
674
675 /* Helper for scsi_io_completion() when special action required. */
676 static void scsi_io_completion_action(struct scsi_cmnd *cmd, int result)
677 {
678         struct request_queue *q = cmd->device->request_queue;
679         struct request *req = cmd->request;
680         int level = 0;
681         enum {ACTION_FAIL, ACTION_REPREP, ACTION_RETRY,
682               ACTION_DELAYED_RETRY} action;
683         struct scsi_sense_hdr sshdr;
684         bool sense_valid;
685         bool sense_current = true;      /* false implies "deferred sense" */
686         blk_status_t blk_stat;
687
688         sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
689         if (sense_valid)
690                 sense_current = !scsi_sense_is_deferred(&sshdr);
691
692         blk_stat = scsi_result_to_blk_status(cmd, result);
693
694         if (host_byte(result) == DID_RESET) {
695                 /* Third party bus reset or reset for error recovery
696                  * reasons.  Just retry the command and see what
697                  * happens.
698                  */
699                 action = ACTION_RETRY;
700         } else if (sense_valid && sense_current) {
701                 switch (sshdr.sense_key) {
702                 case UNIT_ATTENTION:
703                         if (cmd->device->removable) {
704                                 /* Detected disc change.  Set a bit
705                                  * and quietly refuse further access.
706                                  */
707                                 cmd->device->changed = 1;
708                                 action = ACTION_FAIL;
709                         } else {
710                                 /* Must have been a power glitch, or a
711                                  * bus reset.  Could not have been a
712                                  * media change, so we just retry the
713                                  * command and see what happens.
714                                  */
715                                 action = ACTION_RETRY;
716                         }
717                         break;
718                 case ILLEGAL_REQUEST:
719                         /* If we had an ILLEGAL REQUEST returned, then
720                          * we may have performed an unsupported
721                          * command.  The only thing this should be
722                          * would be a ten byte read where only a six
723                          * byte read was supported.  Also, on a system
724                          * where READ CAPACITY failed, we may have
725                          * read past the end of the disk.
726                          */
727                         if ((cmd->device->use_10_for_rw &&
728                             sshdr.asc == 0x20 && sshdr.ascq == 0x00) &&
729                             (cmd->cmnd[0] == READ_10 ||
730                              cmd->cmnd[0] == WRITE_10)) {
731                                 /* This will issue a new 6-byte command. */
732                                 cmd->device->use_10_for_rw = 0;
733                                 action = ACTION_REPREP;
734                         } else if (sshdr.asc == 0x10) /* DIX */ {
735                                 action = ACTION_FAIL;
736                                 blk_stat = BLK_STS_PROTECTION;
737                         /* INVALID COMMAND OPCODE or INVALID FIELD IN CDB */
738                         } else if (sshdr.asc == 0x20 || sshdr.asc == 0x24) {
739                                 action = ACTION_FAIL;
740                                 blk_stat = BLK_STS_TARGET;
741                         } else
742                                 action = ACTION_FAIL;
743                         break;
744                 case ABORTED_COMMAND:
745                         action = ACTION_FAIL;
746                         if (sshdr.asc == 0x10) /* DIF */
747                                 blk_stat = BLK_STS_PROTECTION;
748                         break;
749                 case NOT_READY:
750                         /* If the device is in the process of becoming
751                          * ready, or has a temporary blockage, retry.
752                          */
753                         if (sshdr.asc == 0x04) {
754                                 switch (sshdr.ascq) {
755                                 case 0x01: /* becoming ready */
756                                 case 0x04: /* format in progress */
757                                 case 0x05: /* rebuild in progress */
758                                 case 0x06: /* recalculation in progress */
759                                 case 0x07: /* operation in progress */
760                                 case 0x08: /* Long write in progress */
761                                 case 0x09: /* self test in progress */
762                                 case 0x14: /* space allocation in progress */
763                                 case 0x1a: /* start stop unit in progress */
764                                 case 0x1b: /* sanitize in progress */
765                                 case 0x1d: /* configuration in progress */
766                                 case 0x24: /* depopulation in progress */
767                                         action = ACTION_DELAYED_RETRY;
768                                         break;
769                                 default:
770                                         action = ACTION_FAIL;
771                                         break;
772                                 }
773                         } else
774                                 action = ACTION_FAIL;
775                         break;
776                 case VOLUME_OVERFLOW:
777                         /* See SSC3rXX or current. */
778                         action = ACTION_FAIL;
779                         break;
780                 case DATA_PROTECT:
781                         action = ACTION_FAIL;
782                         if ((sshdr.asc == 0x0C && sshdr.ascq == 0x12) ||
783                             (sshdr.asc == 0x55 &&
784                              (sshdr.ascq == 0x0E || sshdr.ascq == 0x0F))) {
785                                 /* Insufficient zone resources */
786                                 blk_stat = BLK_STS_ZONE_OPEN_RESOURCE;
787                         }
788                         break;
789                 default:
790                         action = ACTION_FAIL;
791                         break;
792                 }
793         } else
794                 action = ACTION_FAIL;
795
796         if (action != ACTION_FAIL && scsi_cmd_runtime_exceeced(cmd))
797                 action = ACTION_FAIL;
798
799         switch (action) {
800         case ACTION_FAIL:
801                 /* Give up and fail the remainder of the request */
802                 if (!(req->rq_flags & RQF_QUIET)) {
803                         static DEFINE_RATELIMIT_STATE(_rs,
804                                         DEFAULT_RATELIMIT_INTERVAL,
805                                         DEFAULT_RATELIMIT_BURST);
806
807                         if (unlikely(scsi_logging_level))
808                                 level =
809                                      SCSI_LOG_LEVEL(SCSI_LOG_MLCOMPLETE_SHIFT,
810                                                     SCSI_LOG_MLCOMPLETE_BITS);
811
812                         /*
813                          * if logging is enabled the failure will be printed
814                          * in scsi_log_completion(), so avoid duplicate messages
815                          */
816                         if (!level && __ratelimit(&_rs)) {
817                                 scsi_print_result(cmd, NULL, FAILED);
818                                 if (driver_byte(result) == DRIVER_SENSE)
819                                         scsi_print_sense(cmd);
820                                 scsi_print_command(cmd);
821                         }
822                 }
823                 if (!scsi_end_request(req, blk_stat, blk_rq_err_bytes(req)))
824                         return;
825                 fallthrough;
826         case ACTION_REPREP:
827                 scsi_io_completion_reprep(cmd, q);
828                 break;
829         case ACTION_RETRY:
830                 /* Retry the same command immediately */
831                 __scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY, false);
832                 break;
833         case ACTION_DELAYED_RETRY:
834                 /* Retry the same command after a delay */
835                 __scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY, false);
836                 break;
837         }
838 }
839
840 /*
841  * Helper for scsi_io_completion() when cmd->result is non-zero. Returns a
842  * new result that may suppress further error checking. Also modifies
843  * *blk_statp in some cases.
844  */
845 static int scsi_io_completion_nz_result(struct scsi_cmnd *cmd, int result,
846                                         blk_status_t *blk_statp)
847 {
848         bool sense_valid;
849         bool sense_current = true;      /* false implies "deferred sense" */
850         struct request *req = cmd->request;
851         struct scsi_sense_hdr sshdr;
852
853         sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
854         if (sense_valid)
855                 sense_current = !scsi_sense_is_deferred(&sshdr);
856
857         if (blk_rq_is_passthrough(req)) {
858                 if (sense_valid) {
859                         /*
860                          * SG_IO wants current and deferred errors
861                          */
862                         scsi_req(req)->sense_len =
863                                 min(8 + cmd->sense_buffer[7],
864                                     SCSI_SENSE_BUFFERSIZE);
865                 }
866                 if (sense_current)
867                         *blk_statp = scsi_result_to_blk_status(cmd, result);
868         } else if (blk_rq_bytes(req) == 0 && sense_current) {
869                 /*
870                  * Flush commands do not transfers any data, and thus cannot use
871                  * good_bytes != blk_rq_bytes(req) as the signal for an error.
872                  * This sets *blk_statp explicitly for the problem case.
873                  */
874                 *blk_statp = scsi_result_to_blk_status(cmd, result);
875         }
876         /*
877          * Recovered errors need reporting, but they're always treated as
878          * success, so fiddle the result code here.  For passthrough requests
879          * we already took a copy of the original into sreq->result which
880          * is what gets returned to the user
881          */
882         if (sense_valid && (sshdr.sense_key == RECOVERED_ERROR)) {
883                 bool do_print = true;
884                 /*
885                  * if ATA PASS-THROUGH INFORMATION AVAILABLE [0x0, 0x1d]
886                  * skip print since caller wants ATA registers. Only occurs
887                  * on SCSI ATA PASS_THROUGH commands when CK_COND=1
888                  */
889                 if ((sshdr.asc == 0x0) && (sshdr.ascq == 0x1d))
890                         do_print = false;
891                 else if (req->rq_flags & RQF_QUIET)
892                         do_print = false;
893                 if (do_print)
894                         scsi_print_sense(cmd);
895                 result = 0;
896                 /* for passthrough, *blk_statp may be set */
897                 *blk_statp = BLK_STS_OK;
898         }
899         /*
900          * Another corner case: the SCSI status byte is non-zero but 'good'.
901          * Example: PRE-FETCH command returns SAM_STAT_CONDITION_MET when
902          * it is able to fit nominated LBs in its cache (and SAM_STAT_GOOD
903          * if it can't fit). Treat SAM_STAT_CONDITION_MET and the related
904          * intermediate statuses (both obsolete in SAM-4) as good.
905          */
906         if (status_byte(result) && scsi_status_is_good(result)) {
907                 result = 0;
908                 *blk_statp = BLK_STS_OK;
909         }
910         return result;
911 }
912
913 /**
914  * scsi_io_completion - Completion processing for SCSI commands.
915  * @cmd:        command that is finished.
916  * @good_bytes: number of processed bytes.
917  *
918  * We will finish off the specified number of sectors. If we are done, the
919  * command block will be released and the queue function will be goosed. If we
920  * are not done then we have to figure out what to do next:
921  *
922  *   a) We can call scsi_io_completion_reprep().  The request will be
923  *      unprepared and put back on the queue.  Then a new command will
924  *      be created for it.  This should be used if we made forward
925  *      progress, or if we want to switch from READ(10) to READ(6) for
926  *      example.
927  *
928  *   b) We can call scsi_io_completion_action().  The request will be
929  *      put back on the queue and retried using the same command as
930  *      before, possibly after a delay.
931  *
932  *   c) We can call scsi_end_request() with blk_stat other than
933  *      BLK_STS_OK, to fail the remainder of the request.
934  */
935 void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes)
936 {
937         int result = cmd->result;
938         struct request_queue *q = cmd->device->request_queue;
939         struct request *req = cmd->request;
940         blk_status_t blk_stat = BLK_STS_OK;
941
942         if (unlikely(result))   /* a nz result may or may not be an error */
943                 result = scsi_io_completion_nz_result(cmd, result, &blk_stat);
944
945         if (unlikely(blk_rq_is_passthrough(req))) {
946                 /*
947                  * scsi_result_to_blk_status may have reset the host_byte
948                  */
949                 scsi_req(req)->result = cmd->result;
950         }
951
952         /*
953          * Next deal with any sectors which we were able to correctly
954          * handle.
955          */
956         SCSI_LOG_HLCOMPLETE(1, scmd_printk(KERN_INFO, cmd,
957                 "%u sectors total, %d bytes done.\n",
958                 blk_rq_sectors(req), good_bytes));
959
960         /*
961          * Failed, zero length commands always need to drop down
962          * to retry code. Fast path should return in this block.
963          */
964         if (likely(blk_rq_bytes(req) > 0 || blk_stat == BLK_STS_OK)) {
965                 if (likely(!scsi_end_request(req, blk_stat, good_bytes)))
966                         return; /* no bytes remaining */
967         }
968
969         /* Kill remainder if no retries. */
970         if (unlikely(blk_stat && scsi_noretry_cmd(cmd))) {
971                 if (scsi_end_request(req, blk_stat, blk_rq_bytes(req)))
972                         WARN_ONCE(true,
973                             "Bytes remaining after failed, no-retry command");
974                 return;
975         }
976
977         /*
978          * If there had been no error, but we have leftover bytes in the
979          * requeues just queue the command up again.
980          */
981         if (likely(result == 0))
982                 scsi_io_completion_reprep(cmd, q);
983         else
984                 scsi_io_completion_action(cmd, result);
985 }
986
987 static inline bool scsi_cmd_needs_dma_drain(struct scsi_device *sdev,
988                 struct request *rq)
989 {
990         return sdev->dma_drain_len && blk_rq_is_passthrough(rq) &&
991                !op_is_write(req_op(rq)) &&
992                sdev->host->hostt->dma_need_drain(rq);
993 }
994
995 /**
996  * scsi_alloc_sgtables - allocate S/G tables for a command
997  * @cmd:  command descriptor we wish to initialize
998  *
999  * Returns:
1000  * * BLK_STS_OK       - on success
1001  * * BLK_STS_RESOURCE - if the failure is retryable
1002  * * BLK_STS_IOERR    - if the failure is fatal
1003  */
1004 blk_status_t scsi_alloc_sgtables(struct scsi_cmnd *cmd)
1005 {
1006         struct scsi_device *sdev = cmd->device;
1007         struct request *rq = cmd->request;
1008         unsigned short nr_segs = blk_rq_nr_phys_segments(rq);
1009         struct scatterlist *last_sg = NULL;
1010         blk_status_t ret;
1011         bool need_drain = scsi_cmd_needs_dma_drain(sdev, rq);
1012         int count;
1013
1014         if (WARN_ON_ONCE(!nr_segs))
1015                 return BLK_STS_IOERR;
1016
1017         /*
1018          * Make sure there is space for the drain.  The driver must adjust
1019          * max_hw_segments to be prepared for this.
1020          */
1021         if (need_drain)
1022                 nr_segs++;
1023
1024         /*
1025          * If sg table allocation fails, requeue request later.
1026          */
1027         if (unlikely(sg_alloc_table_chained(&cmd->sdb.table, nr_segs,
1028                         cmd->sdb.table.sgl, SCSI_INLINE_SG_CNT)))
1029                 return BLK_STS_RESOURCE;
1030
1031         /*
1032          * Next, walk the list, and fill in the addresses and sizes of
1033          * each segment.
1034          */
1035         count = __blk_rq_map_sg(rq->q, rq, cmd->sdb.table.sgl, &last_sg);
1036
1037         if (blk_rq_bytes(rq) & rq->q->dma_pad_mask) {
1038                 unsigned int pad_len =
1039                         (rq->q->dma_pad_mask & ~blk_rq_bytes(rq)) + 1;
1040
1041                 last_sg->length += pad_len;
1042                 cmd->extra_len += pad_len;
1043         }
1044
1045         if (need_drain) {
1046                 sg_unmark_end(last_sg);
1047                 last_sg = sg_next(last_sg);
1048                 sg_set_buf(last_sg, sdev->dma_drain_buf, sdev->dma_drain_len);
1049                 sg_mark_end(last_sg);
1050
1051                 cmd->extra_len += sdev->dma_drain_len;
1052                 count++;
1053         }
1054
1055         BUG_ON(count > cmd->sdb.table.nents);
1056         cmd->sdb.table.nents = count;
1057         cmd->sdb.length = blk_rq_payload_bytes(rq);
1058
1059         if (blk_integrity_rq(rq)) {
1060                 struct scsi_data_buffer *prot_sdb = cmd->prot_sdb;
1061                 int ivecs;
1062
1063                 if (WARN_ON_ONCE(!prot_sdb)) {
1064                         /*
1065                          * This can happen if someone (e.g. multipath)
1066                          * queues a command to a device on an adapter
1067                          * that does not support DIX.
1068                          */
1069                         ret = BLK_STS_IOERR;
1070                         goto out_free_sgtables;
1071                 }
1072
1073                 ivecs = blk_rq_count_integrity_sg(rq->q, rq->bio);
1074
1075                 if (sg_alloc_table_chained(&prot_sdb->table, ivecs,
1076                                 prot_sdb->table.sgl,
1077                                 SCSI_INLINE_PROT_SG_CNT)) {
1078                         ret = BLK_STS_RESOURCE;
1079                         goto out_free_sgtables;
1080                 }
1081
1082                 count = blk_rq_map_integrity_sg(rq->q, rq->bio,
1083                                                 prot_sdb->table.sgl);
1084                 BUG_ON(count > ivecs);
1085                 BUG_ON(count > queue_max_integrity_segments(rq->q));
1086
1087                 cmd->prot_sdb = prot_sdb;
1088                 cmd->prot_sdb->table.nents = count;
1089         }
1090
1091         return BLK_STS_OK;
1092 out_free_sgtables:
1093         scsi_free_sgtables(cmd);
1094         return ret;
1095 }
1096 EXPORT_SYMBOL(scsi_alloc_sgtables);
1097
1098 /**
1099  * scsi_initialize_rq - initialize struct scsi_cmnd partially
1100  * @rq: Request associated with the SCSI command to be initialized.
1101  *
1102  * This function initializes the members of struct scsi_cmnd that must be
1103  * initialized before request processing starts and that won't be
1104  * reinitialized if a SCSI command is requeued.
1105  *
1106  * Called from inside blk_get_request() for pass-through requests and from
1107  * inside scsi_init_command() for filesystem requests.
1108  */
1109 static void scsi_initialize_rq(struct request *rq)
1110 {
1111         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1112
1113         scsi_req_init(&cmd->req);
1114         init_rcu_head(&cmd->rcu);
1115         cmd->jiffies_at_alloc = jiffies;
1116         cmd->retries = 0;
1117 }
1118
1119 /*
1120  * Only called when the request isn't completed by SCSI, and not freed by
1121  * SCSI
1122  */
1123 static void scsi_cleanup_rq(struct request *rq)
1124 {
1125         if (rq->rq_flags & RQF_DONTPREP) {
1126                 scsi_mq_uninit_cmd(blk_mq_rq_to_pdu(rq));
1127                 rq->rq_flags &= ~RQF_DONTPREP;
1128         }
1129 }
1130
1131 /* Called before a request is prepared. See also scsi_mq_prep_fn(). */
1132 void scsi_init_command(struct scsi_device *dev, struct scsi_cmnd *cmd)
1133 {
1134         void *buf = cmd->sense_buffer;
1135         void *prot = cmd->prot_sdb;
1136         struct request *rq = blk_mq_rq_from_pdu(cmd);
1137         unsigned int flags = cmd->flags & SCMD_PRESERVED_FLAGS;
1138         unsigned long jiffies_at_alloc;
1139         int retries, to_clear;
1140         bool in_flight;
1141
1142         if (!blk_rq_is_scsi(rq) && !(flags & SCMD_INITIALIZED)) {
1143                 flags |= SCMD_INITIALIZED;
1144                 scsi_initialize_rq(rq);
1145         }
1146
1147         jiffies_at_alloc = cmd->jiffies_at_alloc;
1148         retries = cmd->retries;
1149         in_flight = test_bit(SCMD_STATE_INFLIGHT, &cmd->state);
1150         /*
1151          * Zero out the cmd, except for the embedded scsi_request. Only clear
1152          * the driver-private command data if the LLD does not supply a
1153          * function to initialize that data.
1154          */
1155         to_clear = sizeof(*cmd) - sizeof(cmd->req);
1156         if (!dev->host->hostt->init_cmd_priv)
1157                 to_clear += dev->host->hostt->cmd_size;
1158         memset((char *)cmd + sizeof(cmd->req), 0, to_clear);
1159
1160         cmd->device = dev;
1161         cmd->sense_buffer = buf;
1162         cmd->prot_sdb = prot;
1163         cmd->flags = flags;
1164         INIT_DELAYED_WORK(&cmd->abort_work, scmd_eh_abort_handler);
1165         cmd->jiffies_at_alloc = jiffies_at_alloc;
1166         cmd->retries = retries;
1167         if (in_flight)
1168                 __set_bit(SCMD_STATE_INFLIGHT, &cmd->state);
1169
1170 }
1171
1172 static blk_status_t scsi_setup_scsi_cmnd(struct scsi_device *sdev,
1173                 struct request *req)
1174 {
1175         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1176
1177         /*
1178          * Passthrough requests may transfer data, in which case they must
1179          * a bio attached to them.  Or they might contain a SCSI command
1180          * that does not transfer data, in which case they may optionally
1181          * submit a request without an attached bio.
1182          */
1183         if (req->bio) {
1184                 blk_status_t ret = scsi_alloc_sgtables(cmd);
1185                 if (unlikely(ret != BLK_STS_OK))
1186                         return ret;
1187         } else {
1188                 BUG_ON(blk_rq_bytes(req));
1189
1190                 memset(&cmd->sdb, 0, sizeof(cmd->sdb));
1191         }
1192
1193         cmd->cmd_len = scsi_req(req)->cmd_len;
1194         if (cmd->cmd_len == 0)
1195                 cmd->cmd_len = scsi_command_size(cmd->cmnd);
1196         cmd->cmnd = scsi_req(req)->cmd;
1197         cmd->transfersize = blk_rq_bytes(req);
1198         cmd->allowed = scsi_req(req)->retries;
1199         return BLK_STS_OK;
1200 }
1201
1202 static blk_status_t
1203 scsi_device_state_check(struct scsi_device *sdev, struct request *req)
1204 {
1205         switch (sdev->sdev_state) {
1206         case SDEV_OFFLINE:
1207         case SDEV_TRANSPORT_OFFLINE:
1208                 /*
1209                  * If the device is offline we refuse to process any
1210                  * commands.  The device must be brought online
1211                  * before trying any recovery commands.
1212                  */
1213                 if (!sdev->offline_already) {
1214                         sdev->offline_already = true;
1215                         sdev_printk(KERN_ERR, sdev,
1216                                     "rejecting I/O to offline device\n");
1217                 }
1218                 return BLK_STS_IOERR;
1219         case SDEV_DEL:
1220                 /*
1221                  * If the device is fully deleted, we refuse to
1222                  * process any commands as well.
1223                  */
1224                 sdev_printk(KERN_ERR, sdev,
1225                             "rejecting I/O to dead device\n");
1226                 return BLK_STS_IOERR;
1227         case SDEV_BLOCK:
1228         case SDEV_CREATED_BLOCK:
1229                 return BLK_STS_RESOURCE;
1230         case SDEV_QUIESCE:
1231                 /*
1232                  * If the devices is blocked we defer normal commands.
1233                  */
1234                 if (req && !(req->rq_flags & RQF_PREEMPT))
1235                         return BLK_STS_RESOURCE;
1236                 return BLK_STS_OK;
1237         default:
1238                 /*
1239                  * For any other not fully online state we only allow
1240                  * special commands.  In particular any user initiated
1241                  * command is not allowed.
1242                  */
1243                 if (req && !(req->rq_flags & RQF_PREEMPT))
1244                         return BLK_STS_IOERR;
1245                 return BLK_STS_OK;
1246         }
1247 }
1248
1249 /*
1250  * scsi_dev_queue_ready: if we can send requests to sdev, return 1 else
1251  * return 0.
1252  *
1253  * Called with the queue_lock held.
1254  */
1255 static inline int scsi_dev_queue_ready(struct request_queue *q,
1256                                   struct scsi_device *sdev)
1257 {
1258         unsigned int busy;
1259
1260         busy = atomic_inc_return(&sdev->device_busy) - 1;
1261         if (atomic_read(&sdev->device_blocked)) {
1262                 if (busy)
1263                         goto out_dec;
1264
1265                 /*
1266                  * unblock after device_blocked iterates to zero
1267                  */
1268                 if (atomic_dec_return(&sdev->device_blocked) > 0)
1269                         goto out_dec;
1270                 SCSI_LOG_MLQUEUE(3, sdev_printk(KERN_INFO, sdev,
1271                                    "unblocking device at zero depth\n"));
1272         }
1273
1274         if (busy >= sdev->queue_depth)
1275                 goto out_dec;
1276
1277         return 1;
1278 out_dec:
1279         atomic_dec(&sdev->device_busy);
1280         return 0;
1281 }
1282
1283 /*
1284  * scsi_target_queue_ready: checks if there we can send commands to target
1285  * @sdev: scsi device on starget to check.
1286  */
1287 static inline int scsi_target_queue_ready(struct Scsi_Host *shost,
1288                                            struct scsi_device *sdev)
1289 {
1290         struct scsi_target *starget = scsi_target(sdev);
1291         unsigned int busy;
1292
1293         if (starget->single_lun) {
1294                 spin_lock_irq(shost->host_lock);
1295                 if (starget->starget_sdev_user &&
1296                     starget->starget_sdev_user != sdev) {
1297                         spin_unlock_irq(shost->host_lock);
1298                         return 0;
1299                 }
1300                 starget->starget_sdev_user = sdev;
1301                 spin_unlock_irq(shost->host_lock);
1302         }
1303
1304         if (starget->can_queue <= 0)
1305                 return 1;
1306
1307         busy = atomic_inc_return(&starget->target_busy) - 1;
1308         if (atomic_read(&starget->target_blocked) > 0) {
1309                 if (busy)
1310                         goto starved;
1311
1312                 /*
1313                  * unblock after target_blocked iterates to zero
1314                  */
1315                 if (atomic_dec_return(&starget->target_blocked) > 0)
1316                         goto out_dec;
1317
1318                 SCSI_LOG_MLQUEUE(3, starget_printk(KERN_INFO, starget,
1319                                  "unblocking target at zero depth\n"));
1320         }
1321
1322         if (busy >= starget->can_queue)
1323                 goto starved;
1324
1325         return 1;
1326
1327 starved:
1328         spin_lock_irq(shost->host_lock);
1329         list_move_tail(&sdev->starved_entry, &shost->starved_list);
1330         spin_unlock_irq(shost->host_lock);
1331 out_dec:
1332         if (starget->can_queue > 0)
1333                 atomic_dec(&starget->target_busy);
1334         return 0;
1335 }
1336
1337 /*
1338  * scsi_host_queue_ready: if we can send requests to shost, return 1 else
1339  * return 0. We must end up running the queue again whenever 0 is
1340  * returned, else IO can hang.
1341  */
1342 static inline int scsi_host_queue_ready(struct request_queue *q,
1343                                    struct Scsi_Host *shost,
1344                                    struct scsi_device *sdev,
1345                                    struct scsi_cmnd *cmd)
1346 {
1347         if (scsi_host_in_recovery(shost))
1348                 return 0;
1349
1350         if (atomic_read(&shost->host_blocked) > 0) {
1351                 if (scsi_host_busy(shost) > 0)
1352                         goto starved;
1353
1354                 /*
1355                  * unblock after host_blocked iterates to zero
1356                  */
1357                 if (atomic_dec_return(&shost->host_blocked) > 0)
1358                         goto out_dec;
1359
1360                 SCSI_LOG_MLQUEUE(3,
1361                         shost_printk(KERN_INFO, shost,
1362                                      "unblocking host at zero depth\n"));
1363         }
1364
1365         if (shost->host_self_blocked)
1366                 goto starved;
1367
1368         /* We're OK to process the command, so we can't be starved */
1369         if (!list_empty(&sdev->starved_entry)) {
1370                 spin_lock_irq(shost->host_lock);
1371                 if (!list_empty(&sdev->starved_entry))
1372                         list_del_init(&sdev->starved_entry);
1373                 spin_unlock_irq(shost->host_lock);
1374         }
1375
1376         __set_bit(SCMD_STATE_INFLIGHT, &cmd->state);
1377
1378         return 1;
1379
1380 starved:
1381         spin_lock_irq(shost->host_lock);
1382         if (list_empty(&sdev->starved_entry))
1383                 list_add_tail(&sdev->starved_entry, &shost->starved_list);
1384         spin_unlock_irq(shost->host_lock);
1385 out_dec:
1386         scsi_dec_host_busy(shost, cmd);
1387         return 0;
1388 }
1389
1390 /*
1391  * Busy state exporting function for request stacking drivers.
1392  *
1393  * For efficiency, no lock is taken to check the busy state of
1394  * shost/starget/sdev, since the returned value is not guaranteed and
1395  * may be changed after request stacking drivers call the function,
1396  * regardless of taking lock or not.
1397  *
1398  * When scsi can't dispatch I/Os anymore and needs to kill I/Os scsi
1399  * needs to return 'not busy'. Otherwise, request stacking drivers
1400  * may hold requests forever.
1401  */
1402 static bool scsi_mq_lld_busy(struct request_queue *q)
1403 {
1404         struct scsi_device *sdev = q->queuedata;
1405         struct Scsi_Host *shost;
1406
1407         if (blk_queue_dying(q))
1408                 return false;
1409
1410         shost = sdev->host;
1411
1412         /*
1413          * Ignore host/starget busy state.
1414          * Since block layer does not have a concept of fairness across
1415          * multiple queues, congestion of host/starget needs to be handled
1416          * in SCSI layer.
1417          */
1418         if (scsi_host_in_recovery(shost) || scsi_device_is_busy(sdev))
1419                 return true;
1420
1421         return false;
1422 }
1423
1424 static void scsi_softirq_done(struct request *rq)
1425 {
1426         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1427         int disposition;
1428
1429         INIT_LIST_HEAD(&cmd->eh_entry);
1430
1431         atomic_inc(&cmd->device->iodone_cnt);
1432         if (cmd->result)
1433                 atomic_inc(&cmd->device->ioerr_cnt);
1434
1435         disposition = scsi_decide_disposition(cmd);
1436         if (disposition != SUCCESS && scsi_cmd_runtime_exceeced(cmd))
1437                 disposition = SUCCESS;
1438
1439         scsi_log_completion(cmd, disposition);
1440
1441         switch (disposition) {
1442         case SUCCESS:
1443                 scsi_finish_command(cmd);
1444                 break;
1445         case NEEDS_RETRY:
1446                 scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY);
1447                 break;
1448         case ADD_TO_MLQUEUE:
1449                 scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY);
1450                 break;
1451         default:
1452                 scsi_eh_scmd_add(cmd);
1453                 break;
1454         }
1455 }
1456
1457 /**
1458  * scsi_dispatch_command - Dispatch a command to the low-level driver.
1459  * @cmd: command block we are dispatching.
1460  *
1461  * Return: nonzero return request was rejected and device's queue needs to be
1462  * plugged.
1463  */
1464 static int scsi_dispatch_cmd(struct scsi_cmnd *cmd)
1465 {
1466         struct Scsi_Host *host = cmd->device->host;
1467         int rtn = 0;
1468
1469         atomic_inc(&cmd->device->iorequest_cnt);
1470
1471         /* check if the device is still usable */
1472         if (unlikely(cmd->device->sdev_state == SDEV_DEL)) {
1473                 /* in SDEV_DEL we error all commands. DID_NO_CONNECT
1474                  * returns an immediate error upwards, and signals
1475                  * that the device is no longer present */
1476                 cmd->result = DID_NO_CONNECT << 16;
1477                 goto done;
1478         }
1479
1480         /* Check to see if the scsi lld made this device blocked. */
1481         if (unlikely(scsi_device_blocked(cmd->device))) {
1482                 /*
1483                  * in blocked state, the command is just put back on
1484                  * the device queue.  The suspend state has already
1485                  * blocked the queue so future requests should not
1486                  * occur until the device transitions out of the
1487                  * suspend state.
1488                  */
1489                 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1490                         "queuecommand : device blocked\n"));
1491                 return SCSI_MLQUEUE_DEVICE_BUSY;
1492         }
1493
1494         /* Store the LUN value in cmnd, if needed. */
1495         if (cmd->device->lun_in_cdb)
1496                 cmd->cmnd[1] = (cmd->cmnd[1] & 0x1f) |
1497                                (cmd->device->lun << 5 & 0xe0);
1498
1499         scsi_log_send(cmd);
1500
1501         /*
1502          * Before we queue this command, check if the command
1503          * length exceeds what the host adapter can handle.
1504          */
1505         if (cmd->cmd_len > cmd->device->host->max_cmd_len) {
1506                 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1507                                "queuecommand : command too long. "
1508                                "cdb_size=%d host->max_cmd_len=%d\n",
1509                                cmd->cmd_len, cmd->device->host->max_cmd_len));
1510                 cmd->result = (DID_ABORT << 16);
1511                 goto done;
1512         }
1513
1514         if (unlikely(host->shost_state == SHOST_DEL)) {
1515                 cmd->result = (DID_NO_CONNECT << 16);
1516                 goto done;
1517
1518         }
1519
1520         trace_scsi_dispatch_cmd_start(cmd);
1521         rtn = host->hostt->queuecommand(host, cmd);
1522         if (rtn) {
1523                 trace_scsi_dispatch_cmd_error(cmd, rtn);
1524                 if (rtn != SCSI_MLQUEUE_DEVICE_BUSY &&
1525                     rtn != SCSI_MLQUEUE_TARGET_BUSY)
1526                         rtn = SCSI_MLQUEUE_HOST_BUSY;
1527
1528                 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1529                         "queuecommand : request rejected\n"));
1530         }
1531
1532         return rtn;
1533  done:
1534         cmd->scsi_done(cmd);
1535         return 0;
1536 }
1537
1538 /* Size in bytes of the sg-list stored in the scsi-mq command-private data. */
1539 static unsigned int scsi_mq_inline_sgl_size(struct Scsi_Host *shost)
1540 {
1541         return min_t(unsigned int, shost->sg_tablesize, SCSI_INLINE_SG_CNT) *
1542                 sizeof(struct scatterlist);
1543 }
1544
1545 static blk_status_t scsi_prepare_cmd(struct request *req)
1546 {
1547         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1548         struct scsi_device *sdev = req->q->queuedata;
1549         struct Scsi_Host *shost = sdev->host;
1550         struct scatterlist *sg;
1551
1552         scsi_init_command(sdev, cmd);
1553
1554         cmd->request = req;
1555         cmd->tag = req->tag;
1556         cmd->prot_op = SCSI_PROT_NORMAL;
1557         if (blk_rq_bytes(req))
1558                 cmd->sc_data_direction = rq_dma_dir(req);
1559         else
1560                 cmd->sc_data_direction = DMA_NONE;
1561
1562         sg = (void *)cmd + sizeof(struct scsi_cmnd) + shost->hostt->cmd_size;
1563         cmd->sdb.table.sgl = sg;
1564
1565         if (scsi_host_get_prot(shost)) {
1566                 memset(cmd->prot_sdb, 0, sizeof(struct scsi_data_buffer));
1567
1568                 cmd->prot_sdb->table.sgl =
1569                         (struct scatterlist *)(cmd->prot_sdb + 1);
1570         }
1571
1572         /*
1573          * Special handling for passthrough commands, which don't go to the ULP
1574          * at all:
1575          */
1576         if (blk_rq_is_scsi(req))
1577                 return scsi_setup_scsi_cmnd(sdev, req);
1578
1579         if (sdev->handler && sdev->handler->prep_fn) {
1580                 blk_status_t ret = sdev->handler->prep_fn(sdev, req);
1581
1582                 if (ret != BLK_STS_OK)
1583                         return ret;
1584         }
1585
1586         cmd->cmnd = scsi_req(req)->cmd = scsi_req(req)->__cmd;
1587         memset(cmd->cmnd, 0, BLK_MAX_CDB);
1588         return scsi_cmd_to_driver(cmd)->init_command(cmd);
1589 }
1590
1591 static void scsi_mq_done(struct scsi_cmnd *cmd)
1592 {
1593         if (unlikely(blk_should_fake_timeout(cmd->request->q)))
1594                 return;
1595         if (unlikely(test_and_set_bit(SCMD_STATE_COMPLETE, &cmd->state)))
1596                 return;
1597         trace_scsi_dispatch_cmd_done(cmd);
1598         blk_mq_complete_request(cmd->request);
1599 }
1600
1601 static void scsi_mq_put_budget(struct request_queue *q)
1602 {
1603         struct scsi_device *sdev = q->queuedata;
1604
1605         atomic_dec(&sdev->device_busy);
1606 }
1607
1608 static bool scsi_mq_get_budget(struct request_queue *q)
1609 {
1610         struct scsi_device *sdev = q->queuedata;
1611
1612         if (scsi_dev_queue_ready(q, sdev))
1613                 return true;
1614
1615         atomic_inc(&sdev->restarts);
1616
1617         /*
1618          * Orders atomic_inc(&sdev->restarts) and atomic_read(&sdev->device_busy).
1619          * .restarts must be incremented before .device_busy is read because the
1620          * code in scsi_run_queue_async() depends on the order of these operations.
1621          */
1622         smp_mb__after_atomic();
1623
1624         /*
1625          * If all in-flight requests originated from this LUN are completed
1626          * before reading .device_busy, sdev->device_busy will be observed as
1627          * zero, then blk_mq_delay_run_hw_queues() will dispatch this request
1628          * soon. Otherwise, completion of one of these requests will observe
1629          * the .restarts flag, and the request queue will be run for handling
1630          * this request, see scsi_end_request().
1631          */
1632         if (unlikely(atomic_read(&sdev->device_busy) == 0 &&
1633                                 !scsi_device_blocked(sdev)))
1634                 blk_mq_delay_run_hw_queues(sdev->request_queue, SCSI_QUEUE_DELAY);
1635         return false;
1636 }
1637
1638 static blk_status_t scsi_queue_rq(struct blk_mq_hw_ctx *hctx,
1639                          const struct blk_mq_queue_data *bd)
1640 {
1641         struct request *req = bd->rq;
1642         struct request_queue *q = req->q;
1643         struct scsi_device *sdev = q->queuedata;
1644         struct Scsi_Host *shost = sdev->host;
1645         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1646         blk_status_t ret;
1647         int reason;
1648
1649         /*
1650          * If the device is not in running state we will reject some or all
1651          * commands.
1652          */
1653         if (unlikely(sdev->sdev_state != SDEV_RUNNING)) {
1654                 ret = scsi_device_state_check(sdev, req);
1655                 if (ret != BLK_STS_OK)
1656                         goto out_put_budget;
1657         }
1658
1659         ret = BLK_STS_RESOURCE;
1660         if (!scsi_target_queue_ready(shost, sdev))
1661                 goto out_put_budget;
1662         if (!scsi_host_queue_ready(q, shost, sdev, cmd))
1663                 goto out_dec_target_busy;
1664
1665         if (!(req->rq_flags & RQF_DONTPREP)) {
1666                 ret = scsi_prepare_cmd(req);
1667                 if (ret != BLK_STS_OK)
1668                         goto out_dec_host_busy;
1669                 req->rq_flags |= RQF_DONTPREP;
1670         } else {
1671                 clear_bit(SCMD_STATE_COMPLETE, &cmd->state);
1672         }
1673
1674         cmd->flags &= SCMD_PRESERVED_FLAGS;
1675         if (sdev->simple_tags)
1676                 cmd->flags |= SCMD_TAGGED;
1677         if (bd->last)
1678                 cmd->flags |= SCMD_LAST;
1679
1680         scsi_set_resid(cmd, 0);
1681         memset(cmd->sense_buffer, 0, SCSI_SENSE_BUFFERSIZE);
1682         cmd->scsi_done = scsi_mq_done;
1683
1684         blk_mq_start_request(req);
1685         reason = scsi_dispatch_cmd(cmd);
1686         if (reason) {
1687                 scsi_set_blocked(cmd, reason);
1688                 ret = BLK_STS_RESOURCE;
1689                 goto out_dec_host_busy;
1690         }
1691
1692         return BLK_STS_OK;
1693
1694 out_dec_host_busy:
1695         scsi_dec_host_busy(shost, cmd);
1696 out_dec_target_busy:
1697         if (scsi_target(sdev)->can_queue > 0)
1698                 atomic_dec(&scsi_target(sdev)->target_busy);
1699 out_put_budget:
1700         scsi_mq_put_budget(q);
1701         switch (ret) {
1702         case BLK_STS_OK:
1703                 break;
1704         case BLK_STS_RESOURCE:
1705         case BLK_STS_ZONE_RESOURCE:
1706                 if (scsi_device_blocked(sdev))
1707                         ret = BLK_STS_DEV_RESOURCE;
1708                 break;
1709         default:
1710                 if (unlikely(!scsi_device_online(sdev)))
1711                         scsi_req(req)->result = DID_NO_CONNECT << 16;
1712                 else
1713                         scsi_req(req)->result = DID_ERROR << 16;
1714                 /*
1715                  * Make sure to release all allocated resources when
1716                  * we hit an error, as we will never see this command
1717                  * again.
1718                  */
1719                 if (req->rq_flags & RQF_DONTPREP)
1720                         scsi_mq_uninit_cmd(cmd);
1721                 scsi_run_queue_async(sdev);
1722                 break;
1723         }
1724         return ret;
1725 }
1726
1727 static enum blk_eh_timer_return scsi_timeout(struct request *req,
1728                 bool reserved)
1729 {
1730         if (reserved)
1731                 return BLK_EH_RESET_TIMER;
1732         return scsi_times_out(req);
1733 }
1734
1735 static int scsi_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
1736                                 unsigned int hctx_idx, unsigned int numa_node)
1737 {
1738         struct Scsi_Host *shost = set->driver_data;
1739         const bool unchecked_isa_dma = shost->unchecked_isa_dma;
1740         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1741         struct scatterlist *sg;
1742         int ret = 0;
1743
1744         if (unchecked_isa_dma)
1745                 cmd->flags |= SCMD_UNCHECKED_ISA_DMA;
1746         cmd->sense_buffer = scsi_alloc_sense_buffer(unchecked_isa_dma,
1747                                                     GFP_KERNEL, numa_node);
1748         if (!cmd->sense_buffer)
1749                 return -ENOMEM;
1750         cmd->req.sense = cmd->sense_buffer;
1751
1752         if (scsi_host_get_prot(shost)) {
1753                 sg = (void *)cmd + sizeof(struct scsi_cmnd) +
1754                         shost->hostt->cmd_size;
1755                 cmd->prot_sdb = (void *)sg + scsi_mq_inline_sgl_size(shost);
1756         }
1757
1758         if (shost->hostt->init_cmd_priv) {
1759                 ret = shost->hostt->init_cmd_priv(shost, cmd);
1760                 if (ret < 0)
1761                         scsi_free_sense_buffer(unchecked_isa_dma,
1762                                                cmd->sense_buffer);
1763         }
1764
1765         return ret;
1766 }
1767
1768 static void scsi_mq_exit_request(struct blk_mq_tag_set *set, struct request *rq,
1769                                  unsigned int hctx_idx)
1770 {
1771         struct Scsi_Host *shost = set->driver_data;
1772         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1773
1774         if (shost->hostt->exit_cmd_priv)
1775                 shost->hostt->exit_cmd_priv(shost, cmd);
1776         scsi_free_sense_buffer(cmd->flags & SCMD_UNCHECKED_ISA_DMA,
1777                                cmd->sense_buffer);
1778 }
1779
1780 static int scsi_map_queues(struct blk_mq_tag_set *set)
1781 {
1782         struct Scsi_Host *shost = container_of(set, struct Scsi_Host, tag_set);
1783
1784         if (shost->hostt->map_queues)
1785                 return shost->hostt->map_queues(shost);
1786         return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
1787 }
1788
1789 void __scsi_init_queue(struct Scsi_Host *shost, struct request_queue *q)
1790 {
1791         struct device *dev = shost->dma_dev;
1792
1793         /*
1794          * this limit is imposed by hardware restrictions
1795          */
1796         blk_queue_max_segments(q, min_t(unsigned short, shost->sg_tablesize,
1797                                         SG_MAX_SEGMENTS));
1798
1799         if (scsi_host_prot_dma(shost)) {
1800                 shost->sg_prot_tablesize =
1801                         min_not_zero(shost->sg_prot_tablesize,
1802                                      (unsigned short)SCSI_MAX_PROT_SG_SEGMENTS);
1803                 BUG_ON(shost->sg_prot_tablesize < shost->sg_tablesize);
1804                 blk_queue_max_integrity_segments(q, shost->sg_prot_tablesize);
1805         }
1806
1807         if (dev->dma_mask) {
1808                 shost->max_sectors = min_t(unsigned int, shost->max_sectors,
1809                                 dma_max_mapping_size(dev) >> SECTOR_SHIFT);
1810         }
1811         blk_queue_max_hw_sectors(q, shost->max_sectors);
1812         if (shost->unchecked_isa_dma)
1813                 blk_queue_bounce_limit(q, BLK_BOUNCE_ISA);
1814         blk_queue_segment_boundary(q, shost->dma_boundary);
1815         dma_set_seg_boundary(dev, shost->dma_boundary);
1816
1817         blk_queue_max_segment_size(q, shost->max_segment_size);
1818         blk_queue_virt_boundary(q, shost->virt_boundary_mask);
1819         dma_set_max_seg_size(dev, queue_max_segment_size(q));
1820
1821         /*
1822          * Set a reasonable default alignment:  The larger of 32-byte (dword),
1823          * which is a common minimum for HBAs, and the minimum DMA alignment,
1824          * which is set by the platform.
1825          *
1826          * Devices that require a bigger alignment can increase it later.
1827          */
1828         blk_queue_dma_alignment(q, max(4, dma_get_cache_alignment()) - 1);
1829 }
1830 EXPORT_SYMBOL_GPL(__scsi_init_queue);
1831
1832 static const struct blk_mq_ops scsi_mq_ops_no_commit = {
1833         .get_budget     = scsi_mq_get_budget,
1834         .put_budget     = scsi_mq_put_budget,
1835         .queue_rq       = scsi_queue_rq,
1836         .complete       = scsi_softirq_done,
1837         .timeout        = scsi_timeout,
1838 #ifdef CONFIG_BLK_DEBUG_FS
1839         .show_rq        = scsi_show_rq,
1840 #endif
1841         .init_request   = scsi_mq_init_request,
1842         .exit_request   = scsi_mq_exit_request,
1843         .initialize_rq_fn = scsi_initialize_rq,
1844         .cleanup_rq     = scsi_cleanup_rq,
1845         .busy           = scsi_mq_lld_busy,
1846         .map_queues     = scsi_map_queues,
1847 };
1848
1849
1850 static void scsi_commit_rqs(struct blk_mq_hw_ctx *hctx)
1851 {
1852         struct request_queue *q = hctx->queue;
1853         struct scsi_device *sdev = q->queuedata;
1854         struct Scsi_Host *shost = sdev->host;
1855
1856         shost->hostt->commit_rqs(shost, hctx->queue_num);
1857 }
1858
1859 static const struct blk_mq_ops scsi_mq_ops = {
1860         .get_budget     = scsi_mq_get_budget,
1861         .put_budget     = scsi_mq_put_budget,
1862         .queue_rq       = scsi_queue_rq,
1863         .commit_rqs     = scsi_commit_rqs,
1864         .complete       = scsi_softirq_done,
1865         .timeout        = scsi_timeout,
1866 #ifdef CONFIG_BLK_DEBUG_FS
1867         .show_rq        = scsi_show_rq,
1868 #endif
1869         .init_request   = scsi_mq_init_request,
1870         .exit_request   = scsi_mq_exit_request,
1871         .initialize_rq_fn = scsi_initialize_rq,
1872         .cleanup_rq     = scsi_cleanup_rq,
1873         .busy           = scsi_mq_lld_busy,
1874         .map_queues     = scsi_map_queues,
1875 };
1876
1877 struct request_queue *scsi_mq_alloc_queue(struct scsi_device *sdev)
1878 {
1879         sdev->request_queue = blk_mq_init_queue(&sdev->host->tag_set);
1880         if (IS_ERR(sdev->request_queue))
1881                 return NULL;
1882
1883         sdev->request_queue->queuedata = sdev;
1884         __scsi_init_queue(sdev->host, sdev->request_queue);
1885         blk_queue_flag_set(QUEUE_FLAG_SCSI_PASSTHROUGH, sdev->request_queue);
1886         return sdev->request_queue;
1887 }
1888
1889 int scsi_mq_setup_tags(struct Scsi_Host *shost)
1890 {
1891         unsigned int cmd_size, sgl_size;
1892         struct blk_mq_tag_set *tag_set = &shost->tag_set;
1893
1894         sgl_size = max_t(unsigned int, sizeof(struct scatterlist),
1895                                 scsi_mq_inline_sgl_size(shost));
1896         cmd_size = sizeof(struct scsi_cmnd) + shost->hostt->cmd_size + sgl_size;
1897         if (scsi_host_get_prot(shost))
1898                 cmd_size += sizeof(struct scsi_data_buffer) +
1899                         sizeof(struct scatterlist) * SCSI_INLINE_PROT_SG_CNT;
1900
1901         memset(tag_set, 0, sizeof(*tag_set));
1902         if (shost->hostt->commit_rqs)
1903                 tag_set->ops = &scsi_mq_ops;
1904         else
1905                 tag_set->ops = &scsi_mq_ops_no_commit;
1906         tag_set->nr_hw_queues = shost->nr_hw_queues ? : 1;
1907         tag_set->queue_depth = shost->can_queue;
1908         tag_set->cmd_size = cmd_size;
1909         tag_set->numa_node = NUMA_NO_NODE;
1910         tag_set->flags = BLK_MQ_F_SHOULD_MERGE;
1911         tag_set->flags |=
1912                 BLK_ALLOC_POLICY_TO_MQ_FLAG(shost->hostt->tag_alloc_policy);
1913         tag_set->driver_data = shost;
1914         if (shost->host_tagset)
1915                 tag_set->flags |= BLK_MQ_F_TAG_HCTX_SHARED;
1916
1917         return blk_mq_alloc_tag_set(tag_set);
1918 }
1919
1920 void scsi_mq_destroy_tags(struct Scsi_Host *shost)
1921 {
1922         blk_mq_free_tag_set(&shost->tag_set);
1923 }
1924
1925 /**
1926  * scsi_device_from_queue - return sdev associated with a request_queue
1927  * @q: The request queue to return the sdev from
1928  *
1929  * Return the sdev associated with a request queue or NULL if the
1930  * request_queue does not reference a SCSI device.
1931  */
1932 struct scsi_device *scsi_device_from_queue(struct request_queue *q)
1933 {
1934         struct scsi_device *sdev = NULL;
1935
1936         if (q->mq_ops == &scsi_mq_ops_no_commit ||
1937             q->mq_ops == &scsi_mq_ops)
1938                 sdev = q->queuedata;
1939         if (!sdev || !get_device(&sdev->sdev_gendev))
1940                 sdev = NULL;
1941
1942         return sdev;
1943 }
1944
1945 /**
1946  * scsi_block_requests - Utility function used by low-level drivers to prevent
1947  * further commands from being queued to the device.
1948  * @shost:  host in question
1949  *
1950  * There is no timer nor any other means by which the requests get unblocked
1951  * other than the low-level driver calling scsi_unblock_requests().
1952  */
1953 void scsi_block_requests(struct Scsi_Host *shost)
1954 {
1955         shost->host_self_blocked = 1;
1956 }
1957 EXPORT_SYMBOL(scsi_block_requests);
1958
1959 /**
1960  * scsi_unblock_requests - Utility function used by low-level drivers to allow
1961  * further commands to be queued to the device.
1962  * @shost:  host in question
1963  *
1964  * There is no timer nor any other means by which the requests get unblocked
1965  * other than the low-level driver calling scsi_unblock_requests(). This is done
1966  * as an API function so that changes to the internals of the scsi mid-layer
1967  * won't require wholesale changes to drivers that use this feature.
1968  */
1969 void scsi_unblock_requests(struct Scsi_Host *shost)
1970 {
1971         shost->host_self_blocked = 0;
1972         scsi_run_host_queues(shost);
1973 }
1974 EXPORT_SYMBOL(scsi_unblock_requests);
1975
1976 void scsi_exit_queue(void)
1977 {
1978         kmem_cache_destroy(scsi_sense_cache);
1979         kmem_cache_destroy(scsi_sense_isadma_cache);
1980 }
1981
1982 /**
1983  *      scsi_mode_select - issue a mode select
1984  *      @sdev:  SCSI device to be queried
1985  *      @pf:    Page format bit (1 == standard, 0 == vendor specific)
1986  *      @sp:    Save page bit (0 == don't save, 1 == save)
1987  *      @modepage: mode page being requested
1988  *      @buffer: request buffer (may not be smaller than eight bytes)
1989  *      @len:   length of request buffer.
1990  *      @timeout: command timeout
1991  *      @retries: number of retries before failing
1992  *      @data: returns a structure abstracting the mode header data
1993  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
1994  *              must be SCSI_SENSE_BUFFERSIZE big.
1995  *
1996  *      Returns zero if successful; negative error number or scsi
1997  *      status on error
1998  *
1999  */
2000 int
2001 scsi_mode_select(struct scsi_device *sdev, int pf, int sp, int modepage,
2002                  unsigned char *buffer, int len, int timeout, int retries,
2003                  struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
2004 {
2005         unsigned char cmd[10];
2006         unsigned char *real_buffer;
2007         int ret;
2008
2009         memset(cmd, 0, sizeof(cmd));
2010         cmd[1] = (pf ? 0x10 : 0) | (sp ? 0x01 : 0);
2011
2012         if (sdev->use_10_for_ms) {
2013                 if (len > 65535)
2014                         return -EINVAL;
2015                 real_buffer = kmalloc(8 + len, GFP_KERNEL);
2016                 if (!real_buffer)
2017                         return -ENOMEM;
2018                 memcpy(real_buffer + 8, buffer, len);
2019                 len += 8;
2020                 real_buffer[0] = 0;
2021                 real_buffer[1] = 0;
2022                 real_buffer[2] = data->medium_type;
2023                 real_buffer[3] = data->device_specific;
2024                 real_buffer[4] = data->longlba ? 0x01 : 0;
2025                 real_buffer[5] = 0;
2026                 real_buffer[6] = data->block_descriptor_length >> 8;
2027                 real_buffer[7] = data->block_descriptor_length;
2028
2029                 cmd[0] = MODE_SELECT_10;
2030                 cmd[7] = len >> 8;
2031                 cmd[8] = len;
2032         } else {
2033                 if (len > 255 || data->block_descriptor_length > 255 ||
2034                     data->longlba)
2035                         return -EINVAL;
2036
2037                 real_buffer = kmalloc(4 + len, GFP_KERNEL);
2038                 if (!real_buffer)
2039                         return -ENOMEM;
2040                 memcpy(real_buffer + 4, buffer, len);
2041                 len += 4;
2042                 real_buffer[0] = 0;
2043                 real_buffer[1] = data->medium_type;
2044                 real_buffer[2] = data->device_specific;
2045                 real_buffer[3] = data->block_descriptor_length;
2046
2047                 cmd[0] = MODE_SELECT;
2048                 cmd[4] = len;
2049         }
2050
2051         ret = scsi_execute_req(sdev, cmd, DMA_TO_DEVICE, real_buffer, len,
2052                                sshdr, timeout, retries, NULL);
2053         kfree(real_buffer);
2054         return ret;
2055 }
2056 EXPORT_SYMBOL_GPL(scsi_mode_select);
2057
2058 /**
2059  *      scsi_mode_sense - issue a mode sense, falling back from 10 to six bytes if necessary.
2060  *      @sdev:  SCSI device to be queried
2061  *      @dbd:   set if mode sense will allow block descriptors to be returned
2062  *      @modepage: mode page being requested
2063  *      @buffer: request buffer (may not be smaller than eight bytes)
2064  *      @len:   length of request buffer.
2065  *      @timeout: command timeout
2066  *      @retries: number of retries before failing
2067  *      @data: returns a structure abstracting the mode header data
2068  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
2069  *              must be SCSI_SENSE_BUFFERSIZE big.
2070  *
2071  *      Returns zero if unsuccessful, or the header offset (either 4
2072  *      or 8 depending on whether a six or ten byte command was
2073  *      issued) if successful.
2074  */
2075 int
2076 scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
2077                   unsigned char *buffer, int len, int timeout, int retries,
2078                   struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
2079 {
2080         unsigned char cmd[12];
2081         int use_10_for_ms;
2082         int header_length;
2083         int result, retry_count = retries;
2084         struct scsi_sense_hdr my_sshdr;
2085
2086         memset(data, 0, sizeof(*data));
2087         memset(&cmd[0], 0, 12);
2088
2089         dbd = sdev->set_dbd_for_ms ? 8 : dbd;
2090         cmd[1] = dbd & 0x18;    /* allows DBD and LLBA bits */
2091         cmd[2] = modepage;
2092
2093         /* caller might not be interested in sense, but we need it */
2094         if (!sshdr)
2095                 sshdr = &my_sshdr;
2096
2097  retry:
2098         use_10_for_ms = sdev->use_10_for_ms;
2099
2100         if (use_10_for_ms) {
2101                 if (len < 8)
2102                         len = 8;
2103
2104                 cmd[0] = MODE_SENSE_10;
2105                 cmd[8] = len;
2106                 header_length = 8;
2107         } else {
2108                 if (len < 4)
2109                         len = 4;
2110
2111                 cmd[0] = MODE_SENSE;
2112                 cmd[4] = len;
2113                 header_length = 4;
2114         }
2115
2116         memset(buffer, 0, len);
2117
2118         result = scsi_execute_req(sdev, cmd, DMA_FROM_DEVICE, buffer, len,
2119                                   sshdr, timeout, retries, NULL);
2120
2121         /* This code looks awful: what it's doing is making sure an
2122          * ILLEGAL REQUEST sense return identifies the actual command
2123          * byte as the problem.  MODE_SENSE commands can return
2124          * ILLEGAL REQUEST if the code page isn't supported */
2125
2126         if (use_10_for_ms && !scsi_status_is_good(result) &&
2127             driver_byte(result) == DRIVER_SENSE) {
2128                 if (scsi_sense_valid(sshdr)) {
2129                         if ((sshdr->sense_key == ILLEGAL_REQUEST) &&
2130                             (sshdr->asc == 0x20) && (sshdr->ascq == 0)) {
2131                                 /*
2132                                  * Invalid command operation code
2133                                  */
2134                                 sdev->use_10_for_ms = 0;
2135                                 goto retry;
2136                         }
2137                 }
2138         }
2139
2140         if (scsi_status_is_good(result)) {
2141                 if (unlikely(buffer[0] == 0x86 && buffer[1] == 0x0b &&
2142                              (modepage == 6 || modepage == 8))) {
2143                         /* Initio breakage? */
2144                         header_length = 0;
2145                         data->length = 13;
2146                         data->medium_type = 0;
2147                         data->device_specific = 0;
2148                         data->longlba = 0;
2149                         data->block_descriptor_length = 0;
2150                 } else if (use_10_for_ms) {
2151                         data->length = buffer[0]*256 + buffer[1] + 2;
2152                         data->medium_type = buffer[2];
2153                         data->device_specific = buffer[3];
2154                         data->longlba = buffer[4] & 0x01;
2155                         data->block_descriptor_length = buffer[6]*256
2156                                 + buffer[7];
2157                 } else {
2158                         data->length = buffer[0] + 1;
2159                         data->medium_type = buffer[1];
2160                         data->device_specific = buffer[2];
2161                         data->block_descriptor_length = buffer[3];
2162                 }
2163                 data->header_length = header_length;
2164         } else if ((status_byte(result) == CHECK_CONDITION) &&
2165                    scsi_sense_valid(sshdr) &&
2166                    sshdr->sense_key == UNIT_ATTENTION && retry_count) {
2167                 retry_count--;
2168                 goto retry;
2169         }
2170
2171         return result;
2172 }
2173 EXPORT_SYMBOL(scsi_mode_sense);
2174
2175 /**
2176  *      scsi_test_unit_ready - test if unit is ready
2177  *      @sdev:  scsi device to change the state of.
2178  *      @timeout: command timeout
2179  *      @retries: number of retries before failing
2180  *      @sshdr: outpout pointer for decoded sense information.
2181  *
2182  *      Returns zero if unsuccessful or an error if TUR failed.  For
2183  *      removable media, UNIT_ATTENTION sets ->changed flag.
2184  **/
2185 int
2186 scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries,
2187                      struct scsi_sense_hdr *sshdr)
2188 {
2189         char cmd[] = {
2190                 TEST_UNIT_READY, 0, 0, 0, 0, 0,
2191         };
2192         int result;
2193
2194         /* try to eat the UNIT_ATTENTION if there are enough retries */
2195         do {
2196                 result = scsi_execute_req(sdev, cmd, DMA_NONE, NULL, 0, sshdr,
2197                                           timeout, 1, NULL);
2198                 if (sdev->removable && scsi_sense_valid(sshdr) &&
2199                     sshdr->sense_key == UNIT_ATTENTION)
2200                         sdev->changed = 1;
2201         } while (scsi_sense_valid(sshdr) &&
2202                  sshdr->sense_key == UNIT_ATTENTION && --retries);
2203
2204         return result;
2205 }
2206 EXPORT_SYMBOL(scsi_test_unit_ready);
2207
2208 /**
2209  *      scsi_device_set_state - Take the given device through the device state model.
2210  *      @sdev:  scsi device to change the state of.
2211  *      @state: state to change to.
2212  *
2213  *      Returns zero if successful or an error if the requested
2214  *      transition is illegal.
2215  */
2216 int
2217 scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
2218 {
2219         enum scsi_device_state oldstate = sdev->sdev_state;
2220
2221         if (state == oldstate)
2222                 return 0;
2223
2224         switch (state) {
2225         case SDEV_CREATED:
2226                 switch (oldstate) {
2227                 case SDEV_CREATED_BLOCK:
2228                         break;
2229                 default:
2230                         goto illegal;
2231                 }
2232                 break;
2233
2234         case SDEV_RUNNING:
2235                 switch (oldstate) {
2236                 case SDEV_CREATED:
2237                 case SDEV_OFFLINE:
2238                 case SDEV_TRANSPORT_OFFLINE:
2239                 case SDEV_QUIESCE:
2240                 case SDEV_BLOCK:
2241                         break;
2242                 default:
2243                         goto illegal;
2244                 }
2245                 break;
2246
2247         case SDEV_QUIESCE:
2248                 switch (oldstate) {
2249                 case SDEV_RUNNING:
2250                 case SDEV_OFFLINE:
2251                 case SDEV_TRANSPORT_OFFLINE:
2252                         break;
2253                 default:
2254                         goto illegal;
2255                 }
2256                 break;
2257
2258         case SDEV_OFFLINE:
2259         case SDEV_TRANSPORT_OFFLINE:
2260                 switch (oldstate) {
2261                 case SDEV_CREATED:
2262                 case SDEV_RUNNING:
2263                 case SDEV_QUIESCE:
2264                 case SDEV_BLOCK:
2265                         break;
2266                 default:
2267                         goto illegal;
2268                 }
2269                 break;
2270
2271         case SDEV_BLOCK:
2272                 switch (oldstate) {
2273                 case SDEV_RUNNING:
2274                 case SDEV_CREATED_BLOCK:
2275                 case SDEV_QUIESCE:
2276                 case SDEV_OFFLINE:
2277                         break;
2278                 default:
2279                         goto illegal;
2280                 }
2281                 break;
2282
2283         case SDEV_CREATED_BLOCK:
2284                 switch (oldstate) {
2285                 case SDEV_CREATED:
2286                         break;
2287                 default:
2288                         goto illegal;
2289                 }
2290                 break;
2291
2292         case SDEV_CANCEL:
2293                 switch (oldstate) {
2294                 case SDEV_CREATED:
2295                 case SDEV_RUNNING:
2296                 case SDEV_QUIESCE:
2297                 case SDEV_OFFLINE:
2298                 case SDEV_TRANSPORT_OFFLINE:
2299                         break;
2300                 default:
2301                         goto illegal;
2302                 }
2303                 break;
2304
2305         case SDEV_DEL:
2306                 switch (oldstate) {
2307                 case SDEV_CREATED:
2308                 case SDEV_RUNNING:
2309                 case SDEV_OFFLINE:
2310                 case SDEV_TRANSPORT_OFFLINE:
2311                 case SDEV_CANCEL:
2312                 case SDEV_BLOCK:
2313                 case SDEV_CREATED_BLOCK:
2314                         break;
2315                 default:
2316                         goto illegal;
2317                 }
2318                 break;
2319
2320         }
2321         sdev->offline_already = false;
2322         sdev->sdev_state = state;
2323         return 0;
2324
2325  illegal:
2326         SCSI_LOG_ERROR_RECOVERY(1,
2327                                 sdev_printk(KERN_ERR, sdev,
2328                                             "Illegal state transition %s->%s",
2329                                             scsi_device_state_name(oldstate),
2330                                             scsi_device_state_name(state))
2331                                 );
2332         return -EINVAL;
2333 }
2334 EXPORT_SYMBOL(scsi_device_set_state);
2335
2336 /**
2337  *      sdev_evt_emit - emit a single SCSI device uevent
2338  *      @sdev: associated SCSI device
2339  *      @evt: event to emit
2340  *
2341  *      Send a single uevent (scsi_event) to the associated scsi_device.
2342  */
2343 static void scsi_evt_emit(struct scsi_device *sdev, struct scsi_event *evt)
2344 {
2345         int idx = 0;
2346         char *envp[3];
2347
2348         switch (evt->evt_type) {
2349         case SDEV_EVT_MEDIA_CHANGE:
2350                 envp[idx++] = "SDEV_MEDIA_CHANGE=1";
2351                 break;
2352         case SDEV_EVT_INQUIRY_CHANGE_REPORTED:
2353                 scsi_rescan_device(&sdev->sdev_gendev);
2354                 envp[idx++] = "SDEV_UA=INQUIRY_DATA_HAS_CHANGED";
2355                 break;
2356         case SDEV_EVT_CAPACITY_CHANGE_REPORTED:
2357                 envp[idx++] = "SDEV_UA=CAPACITY_DATA_HAS_CHANGED";
2358                 break;
2359         case SDEV_EVT_SOFT_THRESHOLD_REACHED_REPORTED:
2360                envp[idx++] = "SDEV_UA=THIN_PROVISIONING_SOFT_THRESHOLD_REACHED";
2361                 break;
2362         case SDEV_EVT_MODE_PARAMETER_CHANGE_REPORTED:
2363                 envp[idx++] = "SDEV_UA=MODE_PARAMETERS_CHANGED";
2364                 break;
2365         case SDEV_EVT_LUN_CHANGE_REPORTED:
2366                 envp[idx++] = "SDEV_UA=REPORTED_LUNS_DATA_HAS_CHANGED";
2367                 break;
2368         case SDEV_EVT_ALUA_STATE_CHANGE_REPORTED:
2369                 envp[idx++] = "SDEV_UA=ASYMMETRIC_ACCESS_STATE_CHANGED";
2370                 break;
2371         case SDEV_EVT_POWER_ON_RESET_OCCURRED:
2372                 envp[idx++] = "SDEV_UA=POWER_ON_RESET_OCCURRED";
2373                 break;
2374         default:
2375                 /* do nothing */
2376                 break;
2377         }
2378
2379         envp[idx++] = NULL;
2380
2381         kobject_uevent_env(&sdev->sdev_gendev.kobj, KOBJ_CHANGE, envp);
2382 }
2383
2384 /**
2385  *      sdev_evt_thread - send a uevent for each scsi event
2386  *      @work: work struct for scsi_device
2387  *
2388  *      Dispatch queued events to their associated scsi_device kobjects
2389  *      as uevents.
2390  */
2391 void scsi_evt_thread(struct work_struct *work)
2392 {
2393         struct scsi_device *sdev;
2394         enum scsi_device_event evt_type;
2395         LIST_HEAD(event_list);
2396
2397         sdev = container_of(work, struct scsi_device, event_work);
2398
2399         for (evt_type = SDEV_EVT_FIRST; evt_type <= SDEV_EVT_LAST; evt_type++)
2400                 if (test_and_clear_bit(evt_type, sdev->pending_events))
2401                         sdev_evt_send_simple(sdev, evt_type, GFP_KERNEL);
2402
2403         while (1) {
2404                 struct scsi_event *evt;
2405                 struct list_head *this, *tmp;
2406                 unsigned long flags;
2407
2408                 spin_lock_irqsave(&sdev->list_lock, flags);
2409                 list_splice_init(&sdev->event_list, &event_list);
2410                 spin_unlock_irqrestore(&sdev->list_lock, flags);
2411
2412                 if (list_empty(&event_list))
2413                         break;
2414
2415                 list_for_each_safe(this, tmp, &event_list) {
2416                         evt = list_entry(this, struct scsi_event, node);
2417                         list_del(&evt->node);
2418                         scsi_evt_emit(sdev, evt);
2419                         kfree(evt);
2420                 }
2421         }
2422 }
2423
2424 /**
2425  *      sdev_evt_send - send asserted event to uevent thread
2426  *      @sdev: scsi_device event occurred on
2427  *      @evt: event to send
2428  *
2429  *      Assert scsi device event asynchronously.
2430  */
2431 void sdev_evt_send(struct scsi_device *sdev, struct scsi_event *evt)
2432 {
2433         unsigned long flags;
2434
2435 #if 0
2436         /* FIXME: currently this check eliminates all media change events
2437          * for polled devices.  Need to update to discriminate between AN
2438          * and polled events */
2439         if (!test_bit(evt->evt_type, sdev->supported_events)) {
2440                 kfree(evt);
2441                 return;
2442         }
2443 #endif
2444
2445         spin_lock_irqsave(&sdev->list_lock, flags);
2446         list_add_tail(&evt->node, &sdev->event_list);
2447         schedule_work(&sdev->event_work);
2448         spin_unlock_irqrestore(&sdev->list_lock, flags);
2449 }
2450 EXPORT_SYMBOL_GPL(sdev_evt_send);
2451
2452 /**
2453  *      sdev_evt_alloc - allocate a new scsi event
2454  *      @evt_type: type of event to allocate
2455  *      @gfpflags: GFP flags for allocation
2456  *
2457  *      Allocates and returns a new scsi_event.
2458  */
2459 struct scsi_event *sdev_evt_alloc(enum scsi_device_event evt_type,
2460                                   gfp_t gfpflags)
2461 {
2462         struct scsi_event *evt = kzalloc(sizeof(struct scsi_event), gfpflags);
2463         if (!evt)
2464                 return NULL;
2465
2466         evt->evt_type = evt_type;
2467         INIT_LIST_HEAD(&evt->node);
2468
2469         /* evt_type-specific initialization, if any */
2470         switch (evt_type) {
2471         case SDEV_EVT_MEDIA_CHANGE:
2472         case SDEV_EVT_INQUIRY_CHANGE_REPORTED:
2473         case SDEV_EVT_CAPACITY_CHANGE_REPORTED:
2474         case SDEV_EVT_SOFT_THRESHOLD_REACHED_REPORTED:
2475         case SDEV_EVT_MODE_PARAMETER_CHANGE_REPORTED:
2476         case SDEV_EVT_LUN_CHANGE_REPORTED:
2477         case SDEV_EVT_ALUA_STATE_CHANGE_REPORTED:
2478         case SDEV_EVT_POWER_ON_RESET_OCCURRED:
2479         default:
2480                 /* do nothing */
2481                 break;
2482         }
2483
2484         return evt;
2485 }
2486 EXPORT_SYMBOL_GPL(sdev_evt_alloc);
2487
2488 /**
2489  *      sdev_evt_send_simple - send asserted event to uevent thread
2490  *      @sdev: scsi_device event occurred on
2491  *      @evt_type: type of event to send
2492  *      @gfpflags: GFP flags for allocation
2493  *
2494  *      Assert scsi device event asynchronously, given an event type.
2495  */
2496 void sdev_evt_send_simple(struct scsi_device *sdev,
2497                           enum scsi_device_event evt_type, gfp_t gfpflags)
2498 {
2499         struct scsi_event *evt = sdev_evt_alloc(evt_type, gfpflags);
2500         if (!evt) {
2501                 sdev_printk(KERN_ERR, sdev, "event %d eaten due to OOM\n",
2502                             evt_type);
2503                 return;
2504         }
2505
2506         sdev_evt_send(sdev, evt);
2507 }
2508 EXPORT_SYMBOL_GPL(sdev_evt_send_simple);
2509
2510 /**
2511  *      scsi_device_quiesce - Block user issued commands.
2512  *      @sdev:  scsi device to quiesce.
2513  *
2514  *      This works by trying to transition to the SDEV_QUIESCE state
2515  *      (which must be a legal transition).  When the device is in this
2516  *      state, only special requests will be accepted, all others will
2517  *      be deferred.  Since special requests may also be requeued requests,
2518  *      a successful return doesn't guarantee the device will be
2519  *      totally quiescent.
2520  *
2521  *      Must be called with user context, may sleep.
2522  *
2523  *      Returns zero if unsuccessful or an error if not.
2524  */
2525 int
2526 scsi_device_quiesce(struct scsi_device *sdev)
2527 {
2528         struct request_queue *q = sdev->request_queue;
2529         int err;
2530
2531         /*
2532          * It is allowed to call scsi_device_quiesce() multiple times from
2533          * the same context but concurrent scsi_device_quiesce() calls are
2534          * not allowed.
2535          */
2536         WARN_ON_ONCE(sdev->quiesced_by && sdev->quiesced_by != current);
2537
2538         if (sdev->quiesced_by == current)
2539                 return 0;
2540
2541         blk_set_pm_only(q);
2542
2543         blk_mq_freeze_queue(q);
2544         /*
2545          * Ensure that the effect of blk_set_pm_only() will be visible
2546          * for percpu_ref_tryget() callers that occur after the queue
2547          * unfreeze even if the queue was already frozen before this function
2548          * was called. See also https://lwn.net/Articles/573497/.
2549          */
2550         synchronize_rcu();
2551         blk_mq_unfreeze_queue(q);
2552
2553         mutex_lock(&sdev->state_mutex);
2554         err = scsi_device_set_state(sdev, SDEV_QUIESCE);
2555         if (err == 0)
2556                 sdev->quiesced_by = current;
2557         else
2558                 blk_clear_pm_only(q);
2559         mutex_unlock(&sdev->state_mutex);
2560
2561         return err;
2562 }
2563 EXPORT_SYMBOL(scsi_device_quiesce);
2564
2565 /**
2566  *      scsi_device_resume - Restart user issued commands to a quiesced device.
2567  *      @sdev:  scsi device to resume.
2568  *
2569  *      Moves the device from quiesced back to running and restarts the
2570  *      queues.
2571  *
2572  *      Must be called with user context, may sleep.
2573  */
2574 void scsi_device_resume(struct scsi_device *sdev)
2575 {
2576         /* check if the device state was mutated prior to resume, and if
2577          * so assume the state is being managed elsewhere (for example
2578          * device deleted during suspend)
2579          */
2580         mutex_lock(&sdev->state_mutex);
2581         if (sdev->quiesced_by) {
2582                 sdev->quiesced_by = NULL;
2583                 blk_clear_pm_only(sdev->request_queue);
2584         }
2585         if (sdev->sdev_state == SDEV_QUIESCE)
2586                 scsi_device_set_state(sdev, SDEV_RUNNING);
2587         mutex_unlock(&sdev->state_mutex);
2588 }
2589 EXPORT_SYMBOL(scsi_device_resume);
2590
2591 static void
2592 device_quiesce_fn(struct scsi_device *sdev, void *data)
2593 {
2594         scsi_device_quiesce(sdev);
2595 }
2596
2597 void
2598 scsi_target_quiesce(struct scsi_target *starget)
2599 {
2600         starget_for_each_device(starget, NULL, device_quiesce_fn);
2601 }
2602 EXPORT_SYMBOL(scsi_target_quiesce);
2603
2604 static void
2605 device_resume_fn(struct scsi_device *sdev, void *data)
2606 {
2607         scsi_device_resume(sdev);
2608 }
2609
2610 void
2611 scsi_target_resume(struct scsi_target *starget)
2612 {
2613         starget_for_each_device(starget, NULL, device_resume_fn);
2614 }
2615 EXPORT_SYMBOL(scsi_target_resume);
2616
2617 /**
2618  * scsi_internal_device_block_nowait - try to transition to the SDEV_BLOCK state
2619  * @sdev: device to block
2620  *
2621  * Pause SCSI command processing on the specified device. Does not sleep.
2622  *
2623  * Returns zero if successful or a negative error code upon failure.
2624  *
2625  * Notes:
2626  * This routine transitions the device to the SDEV_BLOCK state (which must be
2627  * a legal transition). When the device is in this state, command processing
2628  * is paused until the device leaves the SDEV_BLOCK state. See also
2629  * scsi_internal_device_unblock_nowait().
2630  */
2631 int scsi_internal_device_block_nowait(struct scsi_device *sdev)
2632 {
2633         struct request_queue *q = sdev->request_queue;
2634         int err = 0;
2635
2636         err = scsi_device_set_state(sdev, SDEV_BLOCK);
2637         if (err) {
2638                 err = scsi_device_set_state(sdev, SDEV_CREATED_BLOCK);
2639
2640                 if (err)
2641                         return err;
2642         }
2643
2644         /*
2645          * The device has transitioned to SDEV_BLOCK.  Stop the
2646          * block layer from calling the midlayer with this device's
2647          * request queue.
2648          */
2649         blk_mq_quiesce_queue_nowait(q);
2650         return 0;
2651 }
2652 EXPORT_SYMBOL_GPL(scsi_internal_device_block_nowait);
2653
2654 /**
2655  * scsi_internal_device_block - try to transition to the SDEV_BLOCK state
2656  * @sdev: device to block
2657  *
2658  * Pause SCSI command processing on the specified device and wait until all
2659  * ongoing scsi_request_fn() / scsi_queue_rq() calls have finished. May sleep.
2660  *
2661  * Returns zero if successful or a negative error code upon failure.
2662  *
2663  * Note:
2664  * This routine transitions the device to the SDEV_BLOCK state (which must be
2665  * a legal transition). When the device is in this state, command processing
2666  * is paused until the device leaves the SDEV_BLOCK state. See also
2667  * scsi_internal_device_unblock().
2668  */
2669 static int scsi_internal_device_block(struct scsi_device *sdev)
2670 {
2671         struct request_queue *q = sdev->request_queue;
2672         int err;
2673
2674         mutex_lock(&sdev->state_mutex);
2675         err = scsi_internal_device_block_nowait(sdev);
2676         if (err == 0)
2677                 blk_mq_quiesce_queue(q);
2678         mutex_unlock(&sdev->state_mutex);
2679
2680         return err;
2681 }
2682
2683 void scsi_start_queue(struct scsi_device *sdev)
2684 {
2685         struct request_queue *q = sdev->request_queue;
2686
2687         blk_mq_unquiesce_queue(q);
2688 }
2689
2690 /**
2691  * scsi_internal_device_unblock_nowait - resume a device after a block request
2692  * @sdev:       device to resume
2693  * @new_state:  state to set the device to after unblocking
2694  *
2695  * Restart the device queue for a previously suspended SCSI device. Does not
2696  * sleep.
2697  *
2698  * Returns zero if successful or a negative error code upon failure.
2699  *
2700  * Notes:
2701  * This routine transitions the device to the SDEV_RUNNING state or to one of
2702  * the offline states (which must be a legal transition) allowing the midlayer
2703  * to goose the queue for this device.
2704  */
2705 int scsi_internal_device_unblock_nowait(struct scsi_device *sdev,
2706                                         enum scsi_device_state new_state)
2707 {
2708         switch (new_state) {
2709         case SDEV_RUNNING:
2710         case SDEV_TRANSPORT_OFFLINE:
2711                 break;
2712         default:
2713                 return -EINVAL;
2714         }
2715
2716         /*
2717          * Try to transition the scsi device to SDEV_RUNNING or one of the
2718          * offlined states and goose the device queue if successful.
2719          */
2720         switch (sdev->sdev_state) {
2721         case SDEV_BLOCK:
2722         case SDEV_TRANSPORT_OFFLINE:
2723                 sdev->sdev_state = new_state;
2724                 break;
2725         case SDEV_CREATED_BLOCK:
2726                 if (new_state == SDEV_TRANSPORT_OFFLINE ||
2727                     new_state == SDEV_OFFLINE)
2728                         sdev->sdev_state = new_state;
2729                 else
2730                         sdev->sdev_state = SDEV_CREATED;
2731                 break;
2732         case SDEV_CANCEL:
2733         case SDEV_OFFLINE:
2734                 break;
2735         default:
2736                 return -EINVAL;
2737         }
2738         scsi_start_queue(sdev);
2739
2740         return 0;
2741 }
2742 EXPORT_SYMBOL_GPL(scsi_internal_device_unblock_nowait);
2743
2744 /**
2745  * scsi_internal_device_unblock - resume a device after a block request
2746  * @sdev:       device to resume
2747  * @new_state:  state to set the device to after unblocking
2748  *
2749  * Restart the device queue for a previously suspended SCSI device. May sleep.
2750  *
2751  * Returns zero if successful or a negative error code upon failure.
2752  *
2753  * Notes:
2754  * This routine transitions the device to the SDEV_RUNNING state or to one of
2755  * the offline states (which must be a legal transition) allowing the midlayer
2756  * to goose the queue for this device.
2757  */
2758 static int scsi_internal_device_unblock(struct scsi_device *sdev,
2759                                         enum scsi_device_state new_state)
2760 {
2761         int ret;
2762
2763         mutex_lock(&sdev->state_mutex);
2764         ret = scsi_internal_device_unblock_nowait(sdev, new_state);
2765         mutex_unlock(&sdev->state_mutex);
2766
2767         return ret;
2768 }
2769
2770 static void
2771 device_block(struct scsi_device *sdev, void *data)
2772 {
2773         int ret;
2774
2775         ret = scsi_internal_device_block(sdev);
2776
2777         WARN_ONCE(ret, "scsi_internal_device_block(%s) failed: ret = %d\n",
2778                   dev_name(&sdev->sdev_gendev), ret);
2779 }
2780
2781 static int
2782 target_block(struct device *dev, void *data)
2783 {
2784         if (scsi_is_target_device(dev))
2785                 starget_for_each_device(to_scsi_target(dev), NULL,
2786                                         device_block);
2787         return 0;
2788 }
2789
2790 void
2791 scsi_target_block(struct device *dev)
2792 {
2793         if (scsi_is_target_device(dev))
2794                 starget_for_each_device(to_scsi_target(dev), NULL,
2795                                         device_block);
2796         else
2797                 device_for_each_child(dev, NULL, target_block);
2798 }
2799 EXPORT_SYMBOL_GPL(scsi_target_block);
2800
2801 static void
2802 device_unblock(struct scsi_device *sdev, void *data)
2803 {
2804         scsi_internal_device_unblock(sdev, *(enum scsi_device_state *)data);
2805 }
2806
2807 static int
2808 target_unblock(struct device *dev, void *data)
2809 {
2810         if (scsi_is_target_device(dev))
2811                 starget_for_each_device(to_scsi_target(dev), data,
2812                                         device_unblock);
2813         return 0;
2814 }
2815
2816 void
2817 scsi_target_unblock(struct device *dev, enum scsi_device_state new_state)
2818 {
2819         if (scsi_is_target_device(dev))
2820                 starget_for_each_device(to_scsi_target(dev), &new_state,
2821                                         device_unblock);
2822         else
2823                 device_for_each_child(dev, &new_state, target_unblock);
2824 }
2825 EXPORT_SYMBOL_GPL(scsi_target_unblock);
2826
2827 int
2828 scsi_host_block(struct Scsi_Host *shost)
2829 {
2830         struct scsi_device *sdev;
2831         int ret = 0;
2832
2833         /*
2834          * Call scsi_internal_device_block_nowait so we can avoid
2835          * calling synchronize_rcu() for each LUN.
2836          */
2837         shost_for_each_device(sdev, shost) {
2838                 mutex_lock(&sdev->state_mutex);
2839                 ret = scsi_internal_device_block_nowait(sdev);
2840                 mutex_unlock(&sdev->state_mutex);
2841                 if (ret) {
2842                         scsi_device_put(sdev);
2843                         break;
2844                 }
2845         }
2846
2847         /*
2848          * SCSI never enables blk-mq's BLK_MQ_F_BLOCKING flag so
2849          * calling synchronize_rcu() once is enough.
2850          */
2851         WARN_ON_ONCE(shost->tag_set.flags & BLK_MQ_F_BLOCKING);
2852
2853         if (!ret)
2854                 synchronize_rcu();
2855
2856         return ret;
2857 }
2858 EXPORT_SYMBOL_GPL(scsi_host_block);
2859
2860 int
2861 scsi_host_unblock(struct Scsi_Host *shost, int new_state)
2862 {
2863         struct scsi_device *sdev;
2864         int ret = 0;
2865
2866         shost_for_each_device(sdev, shost) {
2867                 ret = scsi_internal_device_unblock(sdev, new_state);
2868                 if (ret) {
2869                         scsi_device_put(sdev);
2870                         break;
2871                 }
2872         }
2873         return ret;
2874 }
2875 EXPORT_SYMBOL_GPL(scsi_host_unblock);
2876
2877 /**
2878  * scsi_kmap_atomic_sg - find and atomically map an sg-elemnt
2879  * @sgl:        scatter-gather list
2880  * @sg_count:   number of segments in sg
2881  * @offset:     offset in bytes into sg, on return offset into the mapped area
2882  * @len:        bytes to map, on return number of bytes mapped
2883  *
2884  * Returns virtual address of the start of the mapped page
2885  */
2886 void *scsi_kmap_atomic_sg(struct scatterlist *sgl, int sg_count,
2887                           size_t *offset, size_t *len)
2888 {
2889         int i;
2890         size_t sg_len = 0, len_complete = 0;
2891         struct scatterlist *sg;
2892         struct page *page;
2893
2894         WARN_ON(!irqs_disabled());
2895
2896         for_each_sg(sgl, sg, sg_count, i) {
2897                 len_complete = sg_len; /* Complete sg-entries */
2898                 sg_len += sg->length;
2899                 if (sg_len > *offset)
2900                         break;
2901         }
2902
2903         if (unlikely(i == sg_count)) {
2904                 printk(KERN_ERR "%s: Bytes in sg: %zu, requested offset %zu, "
2905                         "elements %d\n",
2906                        __func__, sg_len, *offset, sg_count);
2907                 WARN_ON(1);
2908                 return NULL;
2909         }
2910
2911         /* Offset starting from the beginning of first page in this sg-entry */
2912         *offset = *offset - len_complete + sg->offset;
2913
2914         /* Assumption: contiguous pages can be accessed as "page + i" */
2915         page = nth_page(sg_page(sg), (*offset >> PAGE_SHIFT));
2916         *offset &= ~PAGE_MASK;
2917
2918         /* Bytes in this sg-entry from *offset to the end of the page */
2919         sg_len = PAGE_SIZE - *offset;
2920         if (*len > sg_len)
2921                 *len = sg_len;
2922
2923         return kmap_atomic(page);
2924 }
2925 EXPORT_SYMBOL(scsi_kmap_atomic_sg);
2926
2927 /**
2928  * scsi_kunmap_atomic_sg - atomically unmap a virtual address, previously mapped with scsi_kmap_atomic_sg
2929  * @virt:       virtual address to be unmapped
2930  */
2931 void scsi_kunmap_atomic_sg(void *virt)
2932 {
2933         kunmap_atomic(virt);
2934 }
2935 EXPORT_SYMBOL(scsi_kunmap_atomic_sg);
2936
2937 void sdev_disable_disk_events(struct scsi_device *sdev)
2938 {
2939         atomic_inc(&sdev->disk_events_disable_depth);
2940 }
2941 EXPORT_SYMBOL(sdev_disable_disk_events);
2942
2943 void sdev_enable_disk_events(struct scsi_device *sdev)
2944 {
2945         if (WARN_ON_ONCE(atomic_read(&sdev->disk_events_disable_depth) <= 0))
2946                 return;
2947         atomic_dec(&sdev->disk_events_disable_depth);
2948 }
2949 EXPORT_SYMBOL(sdev_enable_disk_events);
2950
2951 /**
2952  * scsi_vpd_lun_id - return a unique device identification
2953  * @sdev: SCSI device
2954  * @id:   buffer for the identification
2955  * @id_len:  length of the buffer
2956  *
2957  * Copies a unique device identification into @id based
2958  * on the information in the VPD page 0x83 of the device.
2959  * The string will be formatted as a SCSI name string.
2960  *
2961  * Returns the length of the identification or error on failure.
2962  * If the identifier is longer than the supplied buffer the actual
2963  * identifier length is returned and the buffer is not zero-padded.
2964  */
2965 int scsi_vpd_lun_id(struct scsi_device *sdev, char *id, size_t id_len)
2966 {
2967         u8 cur_id_type = 0xff;
2968         u8 cur_id_size = 0;
2969         const unsigned char *d, *cur_id_str;
2970         const struct scsi_vpd *vpd_pg83;
2971         int id_size = -EINVAL;
2972
2973         rcu_read_lock();
2974         vpd_pg83 = rcu_dereference(sdev->vpd_pg83);
2975         if (!vpd_pg83) {
2976                 rcu_read_unlock();
2977                 return -ENXIO;
2978         }
2979
2980         /*
2981          * Look for the correct descriptor.
2982          * Order of preference for lun descriptor:
2983          * - SCSI name string
2984          * - NAA IEEE Registered Extended
2985          * - EUI-64 based 16-byte
2986          * - EUI-64 based 12-byte
2987          * - NAA IEEE Registered
2988          * - NAA IEEE Extended
2989          * - T10 Vendor ID
2990          * as longer descriptors reduce the likelyhood
2991          * of identification clashes.
2992          */
2993
2994         /* The id string must be at least 20 bytes + terminating NULL byte */
2995         if (id_len < 21) {
2996                 rcu_read_unlock();
2997                 return -EINVAL;
2998         }
2999
3000         memset(id, 0, id_len);
3001         d = vpd_pg83->data + 4;
3002         while (d < vpd_pg83->data + vpd_pg83->len) {
3003                 /* Skip designators not referring to the LUN */
3004                 if ((d[1] & 0x30) != 0x00)
3005                         goto next_desig;
3006
3007                 switch (d[1] & 0xf) {
3008                 case 0x1:
3009                         /* T10 Vendor ID */
3010                         if (cur_id_size > d[3])
3011                                 break;
3012                         /* Prefer anything */
3013                         if (cur_id_type > 0x01 && cur_id_type != 0xff)
3014                                 break;
3015                         cur_id_size = d[3];
3016                         if (cur_id_size + 4 > id_len)
3017                                 cur_id_size = id_len - 4;
3018                         cur_id_str = d + 4;
3019                         cur_id_type = d[1] & 0xf;
3020                         id_size = snprintf(id, id_len, "t10.%*pE",
3021                                            cur_id_size, cur_id_str);
3022                         break;
3023                 case 0x2:
3024                         /* EUI-64 */
3025                         if (cur_id_size > d[3])
3026                                 break;
3027                         /* Prefer NAA IEEE Registered Extended */
3028                         if (cur_id_type == 0x3 &&
3029                             cur_id_size == d[3])
3030                                 break;
3031                         cur_id_size = d[3];
3032                         cur_id_str = d + 4;
3033                         cur_id_type = d[1] & 0xf;
3034                         switch (cur_id_size) {
3035                         case 8:
3036                                 id_size = snprintf(id, id_len,
3037                                                    "eui.%8phN",
3038                                                    cur_id_str);
3039                                 break;
3040                         case 12:
3041                                 id_size = snprintf(id, id_len,
3042                                                    "eui.%12phN",
3043                                                    cur_id_str);
3044                                 break;
3045                         case 16:
3046                                 id_size = snprintf(id, id_len,
3047                                                    "eui.%16phN",
3048                                                    cur_id_str);
3049                                 break;
3050                         default:
3051                                 cur_id_size = 0;
3052                                 break;
3053                         }
3054                         break;
3055                 case 0x3:
3056                         /* NAA */
3057                         if (cur_id_size > d[3])
3058                                 break;
3059                         cur_id_size = d[3];
3060                         cur_id_str = d + 4;
3061                         cur_id_type = d[1] & 0xf;
3062                         switch (cur_id_size) {
3063                         case 8:
3064                                 id_size = snprintf(id, id_len,
3065                                                    "naa.%8phN",
3066                                                    cur_id_str);
3067                                 break;
3068                         case 16:
3069                                 id_size = snprintf(id, id_len,
3070                                                    "naa.%16phN",
3071                                                    cur_id_str);
3072                                 break;
3073                         default:
3074                                 cur_id_size = 0;
3075                                 break;
3076                         }
3077                         break;
3078                 case 0x8:
3079                         /* SCSI name string */
3080                         if (cur_id_size + 4 > d[3])
3081                                 break;
3082                         /* Prefer others for truncated descriptor */
3083                         if (cur_id_size && d[3] > id_len)
3084                                 break;
3085                         cur_id_size = id_size = d[3];
3086                         cur_id_str = d + 4;
3087                         cur_id_type = d[1] & 0xf;
3088                         if (cur_id_size >= id_len)
3089                                 cur_id_size = id_len - 1;
3090                         memcpy(id, cur_id_str, cur_id_size);
3091                         /* Decrease priority for truncated descriptor */
3092                         if (cur_id_size != id_size)
3093                                 cur_id_size = 6;
3094                         break;
3095                 default:
3096                         break;
3097                 }
3098 next_desig:
3099                 d += d[3] + 4;
3100         }
3101         rcu_read_unlock();
3102
3103         return id_size;
3104 }
3105 EXPORT_SYMBOL(scsi_vpd_lun_id);
3106
3107 /*
3108  * scsi_vpd_tpg_id - return a target port group identifier
3109  * @sdev: SCSI device
3110  *
3111  * Returns the Target Port Group identifier from the information
3112  * froom VPD page 0x83 of the device.
3113  *
3114  * Returns the identifier or error on failure.
3115  */
3116 int scsi_vpd_tpg_id(struct scsi_device *sdev, int *rel_id)
3117 {
3118         const unsigned char *d;
3119         const struct scsi_vpd *vpd_pg83;
3120         int group_id = -EAGAIN, rel_port = -1;
3121
3122         rcu_read_lock();
3123         vpd_pg83 = rcu_dereference(sdev->vpd_pg83);
3124         if (!vpd_pg83) {
3125                 rcu_read_unlock();
3126                 return -ENXIO;
3127         }
3128
3129         d = vpd_pg83->data + 4;
3130         while (d < vpd_pg83->data + vpd_pg83->len) {
3131                 switch (d[1] & 0xf) {
3132                 case 0x4:
3133                         /* Relative target port */
3134                         rel_port = get_unaligned_be16(&d[6]);
3135                         break;
3136                 case 0x5:
3137                         /* Target port group */
3138                         group_id = get_unaligned_be16(&d[6]);
3139                         break;
3140                 default:
3141                         break;
3142                 }
3143                 d += d[3] + 4;
3144         }
3145         rcu_read_unlock();
3146
3147         if (group_id >= 0 && rel_id && rel_port != -1)
3148                 *rel_id = rel_port;
3149
3150         return group_id;
3151 }
3152 EXPORT_SYMBOL(scsi_vpd_tpg_id);