Merge branch 'kvm-basic-exit-reason' into HEAD
[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_sdb_cache;
56 static struct kmem_cache *scsi_sense_cache;
57 static struct kmem_cache *scsi_sense_isadma_cache;
58 static DEFINE_MUTEX(scsi_sense_cache_mutex);
59
60 static void scsi_mq_uninit_cmd(struct scsi_cmnd *cmd);
61
62 static inline struct kmem_cache *
63 scsi_select_sense_cache(bool unchecked_isa_dma)
64 {
65         return unchecked_isa_dma ? scsi_sense_isadma_cache : scsi_sense_cache;
66 }
67
68 static void scsi_free_sense_buffer(bool unchecked_isa_dma,
69                                    unsigned char *sense_buffer)
70 {
71         kmem_cache_free(scsi_select_sense_cache(unchecked_isa_dma),
72                         sense_buffer);
73 }
74
75 static unsigned char *scsi_alloc_sense_buffer(bool unchecked_isa_dma,
76         gfp_t gfp_mask, int numa_node)
77 {
78         return kmem_cache_alloc_node(scsi_select_sense_cache(unchecked_isa_dma),
79                                      gfp_mask, numa_node);
80 }
81
82 int scsi_init_sense_cache(struct Scsi_Host *shost)
83 {
84         struct kmem_cache *cache;
85         int ret = 0;
86
87         mutex_lock(&scsi_sense_cache_mutex);
88         cache = scsi_select_sense_cache(shost->unchecked_isa_dma);
89         if (cache)
90                 goto exit;
91
92         if (shost->unchecked_isa_dma) {
93                 scsi_sense_isadma_cache =
94                         kmem_cache_create("scsi_sense_cache(DMA)",
95                                 SCSI_SENSE_BUFFERSIZE, 0,
96                                 SLAB_HWCACHE_ALIGN | SLAB_CACHE_DMA, NULL);
97                 if (!scsi_sense_isadma_cache)
98                         ret = -ENOMEM;
99         } else {
100                 scsi_sense_cache =
101                         kmem_cache_create_usercopy("scsi_sense_cache",
102                                 SCSI_SENSE_BUFFERSIZE, 0, SLAB_HWCACHE_ALIGN,
103                                 0, SCSI_SENSE_BUFFERSIZE, NULL);
104                 if (!scsi_sense_cache)
105                         ret = -ENOMEM;
106         }
107  exit:
108         mutex_unlock(&scsi_sense_cache_mutex);
109         return ret;
110 }
111
112 /*
113  * When to reinvoke queueing after a resource shortage. It's 3 msecs to
114  * not change behaviour from the previous unplug mechanism, experimentation
115  * may prove this needs changing.
116  */
117 #define SCSI_QUEUE_DELAY        3
118
119 static void
120 scsi_set_blocked(struct scsi_cmnd *cmd, int reason)
121 {
122         struct Scsi_Host *host = cmd->device->host;
123         struct scsi_device *device = cmd->device;
124         struct scsi_target *starget = scsi_target(device);
125
126         /*
127          * Set the appropriate busy bit for the device/host.
128          *
129          * If the host/device isn't busy, assume that something actually
130          * completed, and that we should be able to queue a command now.
131          *
132          * Note that the prior mid-layer assumption that any host could
133          * always queue at least one command is now broken.  The mid-layer
134          * will implement a user specifiable stall (see
135          * scsi_host.max_host_blocked and scsi_device.max_device_blocked)
136          * if a command is requeued with no other commands outstanding
137          * either for the device or for the host.
138          */
139         switch (reason) {
140         case SCSI_MLQUEUE_HOST_BUSY:
141                 atomic_set(&host->host_blocked, host->max_host_blocked);
142                 break;
143         case SCSI_MLQUEUE_DEVICE_BUSY:
144         case SCSI_MLQUEUE_EH_RETRY:
145                 atomic_set(&device->device_blocked,
146                            device->max_device_blocked);
147                 break;
148         case SCSI_MLQUEUE_TARGET_BUSY:
149                 atomic_set(&starget->target_blocked,
150                            starget->max_target_blocked);
151                 break;
152         }
153 }
154
155 static void scsi_mq_requeue_cmd(struct scsi_cmnd *cmd)
156 {
157         if (cmd->request->rq_flags & RQF_DONTPREP) {
158                 cmd->request->rq_flags &= ~RQF_DONTPREP;
159                 scsi_mq_uninit_cmd(cmd);
160         } else {
161                 WARN_ON_ONCE(true);
162         }
163         blk_mq_requeue_request(cmd->request, true);
164 }
165
166 /**
167  * __scsi_queue_insert - private queue insertion
168  * @cmd: The SCSI command being requeued
169  * @reason:  The reason for the requeue
170  * @unbusy: Whether the queue should be unbusied
171  *
172  * This is a private queue insertion.  The public interface
173  * scsi_queue_insert() always assumes the queue should be unbusied
174  * because it's always called before the completion.  This function is
175  * for a requeue after completion, which should only occur in this
176  * file.
177  */
178 static void __scsi_queue_insert(struct scsi_cmnd *cmd, int reason, bool unbusy)
179 {
180         struct scsi_device *device = cmd->device;
181
182         SCSI_LOG_MLQUEUE(1, scmd_printk(KERN_INFO, cmd,
183                 "Inserting command %p into mlqueue\n", cmd));
184
185         scsi_set_blocked(cmd, reason);
186
187         /*
188          * Decrement the counters, since these commands are no longer
189          * active on the host/device.
190          */
191         if (unbusy)
192                 scsi_device_unbusy(device, cmd);
193
194         /*
195          * Requeue this command.  It will go before all other commands
196          * that are already in the queue. Schedule requeue work under
197          * lock such that the kblockd_schedule_work() call happens
198          * before blk_cleanup_queue() finishes.
199          */
200         cmd->result = 0;
201
202         blk_mq_requeue_request(cmd->request, true);
203 }
204
205 /*
206  * Function:    scsi_queue_insert()
207  *
208  * Purpose:     Insert a command in the midlevel queue.
209  *
210  * Arguments:   cmd    - command that we are adding to queue.
211  *              reason - why we are inserting command to queue.
212  *
213  * Lock status: Assumed that lock is not held upon entry.
214  *
215  * Returns:     Nothing.
216  *
217  * Notes:       We do this for one of two cases.  Either the host is busy
218  *              and it cannot accept any more commands for the time being,
219  *              or the device returned QUEUE_FULL and can accept no more
220  *              commands.
221  * Notes:       This could be called either from an interrupt context or a
222  *              normal process context.
223  */
224 void scsi_queue_insert(struct scsi_cmnd *cmd, int reason)
225 {
226         __scsi_queue_insert(cmd, reason, true);
227 }
228
229
230 /**
231  * __scsi_execute - insert request and wait for the result
232  * @sdev:       scsi device
233  * @cmd:        scsi command
234  * @data_direction: data direction
235  * @buffer:     data buffer
236  * @bufflen:    len of buffer
237  * @sense:      optional sense buffer
238  * @sshdr:      optional decoded sense header
239  * @timeout:    request timeout in seconds
240  * @retries:    number of times to retry request
241  * @flags:      flags for ->cmd_flags
242  * @rq_flags:   flags for ->rq_flags
243  * @resid:      optional residual length
244  *
245  * Returns the scsi_cmnd result field if a command was executed, or a negative
246  * Linux error code if we didn't get that far.
247  */
248 int __scsi_execute(struct scsi_device *sdev, const unsigned char *cmd,
249                  int data_direction, void *buffer, unsigned bufflen,
250                  unsigned char *sense, struct scsi_sense_hdr *sshdr,
251                  int timeout, int retries, u64 flags, req_flags_t rq_flags,
252                  int *resid)
253 {
254         struct request *req;
255         struct scsi_request *rq;
256         int ret = DRIVER_ERROR << 24;
257
258         req = blk_get_request(sdev->request_queue,
259                         data_direction == DMA_TO_DEVICE ?
260                         REQ_OP_SCSI_OUT : REQ_OP_SCSI_IN, BLK_MQ_REQ_PREEMPT);
261         if (IS_ERR(req))
262                 return ret;
263         rq = scsi_req(req);
264
265         if (bufflen &&  blk_rq_map_kern(sdev->request_queue, req,
266                                         buffer, bufflen, GFP_NOIO))
267                 goto out;
268
269         rq->cmd_len = COMMAND_SIZE(cmd[0]);
270         memcpy(rq->cmd, cmd, rq->cmd_len);
271         rq->retries = retries;
272         req->timeout = timeout;
273         req->cmd_flags |= flags;
274         req->rq_flags |= rq_flags | RQF_QUIET;
275
276         /*
277          * head injection *required* here otherwise quiesce won't work
278          */
279         blk_execute_rq(req->q, NULL, req, 1);
280
281         /*
282          * Some devices (USB mass-storage in particular) may transfer
283          * garbage data together with a residue indicating that the data
284          * is invalid.  Prevent the garbage from being misinterpreted
285          * and prevent security leaks by zeroing out the excess data.
286          */
287         if (unlikely(rq->resid_len > 0 && rq->resid_len <= bufflen))
288                 memset(buffer + (bufflen - rq->resid_len), 0, rq->resid_len);
289
290         if (resid)
291                 *resid = rq->resid_len;
292         if (sense && rq->sense_len)
293                 memcpy(sense, rq->sense, SCSI_SENSE_BUFFERSIZE);
294         if (sshdr)
295                 scsi_normalize_sense(rq->sense, rq->sense_len, sshdr);
296         ret = rq->result;
297  out:
298         blk_put_request(req);
299
300         return ret;
301 }
302 EXPORT_SYMBOL(__scsi_execute);
303
304 /*
305  * Function:    scsi_init_cmd_errh()
306  *
307  * Purpose:     Initialize cmd fields related to error handling.
308  *
309  * Arguments:   cmd     - command that is ready to be queued.
310  *
311  * Notes:       This function has the job of initializing a number of
312  *              fields related to error handling.   Typically this will
313  *              be called once for each command, as required.
314  */
315 static void scsi_init_cmd_errh(struct scsi_cmnd *cmd)
316 {
317         scsi_set_resid(cmd, 0);
318         memset(cmd->sense_buffer, 0, SCSI_SENSE_BUFFERSIZE);
319         if (cmd->cmd_len == 0)
320                 cmd->cmd_len = scsi_command_size(cmd->cmnd);
321 }
322
323 /*
324  * Wake up the error handler if necessary. Avoid as follows that the error
325  * handler is not woken up if host in-flight requests number ==
326  * shost->host_failed: use call_rcu() in scsi_eh_scmd_add() in combination
327  * with an RCU read lock in this function to ensure that this function in
328  * its entirety either finishes before scsi_eh_scmd_add() increases the
329  * host_failed counter or that it notices the shost state change made by
330  * scsi_eh_scmd_add().
331  */
332 static void scsi_dec_host_busy(struct Scsi_Host *shost, struct scsi_cmnd *cmd)
333 {
334         unsigned long flags;
335
336         rcu_read_lock();
337         __clear_bit(SCMD_STATE_INFLIGHT, &cmd->state);
338         if (unlikely(scsi_host_in_recovery(shost))) {
339                 spin_lock_irqsave(shost->host_lock, flags);
340                 if (shost->host_failed || shost->host_eh_scheduled)
341                         scsi_eh_wakeup(shost);
342                 spin_unlock_irqrestore(shost->host_lock, flags);
343         }
344         rcu_read_unlock();
345 }
346
347 void scsi_device_unbusy(struct scsi_device *sdev, struct scsi_cmnd *cmd)
348 {
349         struct Scsi_Host *shost = sdev->host;
350         struct scsi_target *starget = scsi_target(sdev);
351
352         scsi_dec_host_busy(shost, cmd);
353
354         if (starget->can_queue > 0)
355                 atomic_dec(&starget->target_busy);
356
357         atomic_dec(&sdev->device_busy);
358 }
359
360 static void scsi_kick_queue(struct request_queue *q)
361 {
362         blk_mq_run_hw_queues(q, false);
363 }
364
365 /*
366  * Called for single_lun devices on IO completion. Clear starget_sdev_user,
367  * and call blk_run_queue for all the scsi_devices on the target -
368  * including current_sdev first.
369  *
370  * Called with *no* scsi locks held.
371  */
372 static void scsi_single_lun_run(struct scsi_device *current_sdev)
373 {
374         struct Scsi_Host *shost = current_sdev->host;
375         struct scsi_device *sdev, *tmp;
376         struct scsi_target *starget = scsi_target(current_sdev);
377         unsigned long flags;
378
379         spin_lock_irqsave(shost->host_lock, flags);
380         starget->starget_sdev_user = NULL;
381         spin_unlock_irqrestore(shost->host_lock, flags);
382
383         /*
384          * Call blk_run_queue for all LUNs on the target, starting with
385          * current_sdev. We race with others (to set starget_sdev_user),
386          * but in most cases, we will be first. Ideally, each LU on the
387          * target would get some limited time or requests on the target.
388          */
389         scsi_kick_queue(current_sdev->request_queue);
390
391         spin_lock_irqsave(shost->host_lock, flags);
392         if (starget->starget_sdev_user)
393                 goto out;
394         list_for_each_entry_safe(sdev, tmp, &starget->devices,
395                         same_target_siblings) {
396                 if (sdev == current_sdev)
397                         continue;
398                 if (scsi_device_get(sdev))
399                         continue;
400
401                 spin_unlock_irqrestore(shost->host_lock, flags);
402                 scsi_kick_queue(sdev->request_queue);
403                 spin_lock_irqsave(shost->host_lock, flags);
404         
405                 scsi_device_put(sdev);
406         }
407  out:
408         spin_unlock_irqrestore(shost->host_lock, flags);
409 }
410
411 static inline bool scsi_device_is_busy(struct scsi_device *sdev)
412 {
413         if (atomic_read(&sdev->device_busy) >= sdev->queue_depth)
414                 return true;
415         if (atomic_read(&sdev->device_blocked) > 0)
416                 return true;
417         return false;
418 }
419
420 static inline bool scsi_target_is_busy(struct scsi_target *starget)
421 {
422         if (starget->can_queue > 0) {
423                 if (atomic_read(&starget->target_busy) >= starget->can_queue)
424                         return true;
425                 if (atomic_read(&starget->target_blocked) > 0)
426                         return true;
427         }
428         return false;
429 }
430
431 static inline bool scsi_host_is_busy(struct Scsi_Host *shost)
432 {
433         if (atomic_read(&shost->host_blocked) > 0)
434                 return true;
435         if (shost->host_self_blocked)
436                 return true;
437         return false;
438 }
439
440 static void scsi_starved_list_run(struct Scsi_Host *shost)
441 {
442         LIST_HEAD(starved_list);
443         struct scsi_device *sdev;
444         unsigned long flags;
445
446         spin_lock_irqsave(shost->host_lock, flags);
447         list_splice_init(&shost->starved_list, &starved_list);
448
449         while (!list_empty(&starved_list)) {
450                 struct request_queue *slq;
451
452                 /*
453                  * As long as shost is accepting commands and we have
454                  * starved queues, call blk_run_queue. scsi_request_fn
455                  * drops the queue_lock and can add us back to the
456                  * starved_list.
457                  *
458                  * host_lock protects the starved_list and starved_entry.
459                  * scsi_request_fn must get the host_lock before checking
460                  * or modifying starved_list or starved_entry.
461                  */
462                 if (scsi_host_is_busy(shost))
463                         break;
464
465                 sdev = list_entry(starved_list.next,
466                                   struct scsi_device, starved_entry);
467                 list_del_init(&sdev->starved_entry);
468                 if (scsi_target_is_busy(scsi_target(sdev))) {
469                         list_move_tail(&sdev->starved_entry,
470                                        &shost->starved_list);
471                         continue;
472                 }
473
474                 /*
475                  * Once we drop the host lock, a racing scsi_remove_device()
476                  * call may remove the sdev from the starved list and destroy
477                  * it and the queue.  Mitigate by taking a reference to the
478                  * queue and never touching the sdev again after we drop the
479                  * host lock.  Note: if __scsi_remove_device() invokes
480                  * blk_cleanup_queue() before the queue is run from this
481                  * function then blk_run_queue() will return immediately since
482                  * blk_cleanup_queue() marks the queue with QUEUE_FLAG_DYING.
483                  */
484                 slq = sdev->request_queue;
485                 if (!blk_get_queue(slq))
486                         continue;
487                 spin_unlock_irqrestore(shost->host_lock, flags);
488
489                 scsi_kick_queue(slq);
490                 blk_put_queue(slq);
491
492                 spin_lock_irqsave(shost->host_lock, flags);
493         }
494         /* put any unprocessed entries back */
495         list_splice(&starved_list, &shost->starved_list);
496         spin_unlock_irqrestore(shost->host_lock, flags);
497 }
498
499 /*
500  * Function:   scsi_run_queue()
501  *
502  * Purpose:    Select a proper request queue to serve next
503  *
504  * Arguments:  q       - last request's queue
505  *
506  * Returns:     Nothing
507  *
508  * Notes:      The previous command was completely finished, start
509  *             a new one if possible.
510  */
511 static void scsi_run_queue(struct request_queue *q)
512 {
513         struct scsi_device *sdev = q->queuedata;
514
515         if (scsi_target(sdev)->single_lun)
516                 scsi_single_lun_run(sdev);
517         if (!list_empty(&sdev->host->starved_list))
518                 scsi_starved_list_run(sdev->host);
519
520         blk_mq_run_hw_queues(q, false);
521 }
522
523 void scsi_requeue_run_queue(struct work_struct *work)
524 {
525         struct scsi_device *sdev;
526         struct request_queue *q;
527
528         sdev = container_of(work, struct scsi_device, requeue_work);
529         q = sdev->request_queue;
530         scsi_run_queue(q);
531 }
532
533 void scsi_run_host_queues(struct Scsi_Host *shost)
534 {
535         struct scsi_device *sdev;
536
537         shost_for_each_device(sdev, shost)
538                 scsi_run_queue(sdev->request_queue);
539 }
540
541 static void scsi_uninit_cmd(struct scsi_cmnd *cmd)
542 {
543         if (!blk_rq_is_passthrough(cmd->request)) {
544                 struct scsi_driver *drv = scsi_cmd_to_driver(cmd);
545
546                 if (drv->uninit_command)
547                         drv->uninit_command(cmd);
548         }
549 }
550
551 static void scsi_mq_free_sgtables(struct scsi_cmnd *cmd)
552 {
553         if (cmd->sdb.table.nents)
554                 sg_free_table_chained(&cmd->sdb.table,
555                                 SCSI_INLINE_SG_CNT);
556         if (scsi_prot_sg_count(cmd))
557                 sg_free_table_chained(&cmd->prot_sdb->table,
558                                 SCSI_INLINE_PROT_SG_CNT);
559 }
560
561 static void scsi_mq_uninit_cmd(struct scsi_cmnd *cmd)
562 {
563         scsi_mq_free_sgtables(cmd);
564         scsi_uninit_cmd(cmd);
565 }
566
567 /* Returns false when no more bytes to process, true if there are more */
568 static bool scsi_end_request(struct request *req, blk_status_t error,
569                 unsigned int bytes)
570 {
571         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
572         struct scsi_device *sdev = cmd->device;
573         struct request_queue *q = sdev->request_queue;
574
575         if (blk_update_request(req, error, bytes))
576                 return true;
577
578         if (blk_queue_add_random(q))
579                 add_disk_randomness(req->rq_disk);
580
581         if (!blk_rq_is_scsi(req)) {
582                 WARN_ON_ONCE(!(cmd->flags & SCMD_INITIALIZED));
583                 cmd->flags &= ~SCMD_INITIALIZED;
584         }
585
586         /*
587          * Calling rcu_barrier() is not necessary here because the
588          * SCSI error handler guarantees that the function called by
589          * call_rcu() has been called before scsi_end_request() is
590          * called.
591          */
592         destroy_rcu_head(&cmd->rcu);
593
594         /*
595          * In the MQ case the command gets freed by __blk_mq_end_request,
596          * so we have to do all cleanup that depends on it earlier.
597          *
598          * We also can't kick the queues from irq context, so we
599          * will have to defer it to a workqueue.
600          */
601         scsi_mq_uninit_cmd(cmd);
602
603         /*
604          * queue is still alive, so grab the ref for preventing it
605          * from being cleaned up during running queue.
606          */
607         percpu_ref_get(&q->q_usage_counter);
608
609         __blk_mq_end_request(req, error);
610
611         if (scsi_target(sdev)->single_lun ||
612             !list_empty(&sdev->host->starved_list))
613                 kblockd_schedule_work(&sdev->requeue_work);
614         else
615                 blk_mq_run_hw_queues(q, true);
616
617         percpu_ref_put(&q->q_usage_counter);
618         return false;
619 }
620
621 /**
622  * scsi_result_to_blk_status - translate a SCSI result code into blk_status_t
623  * @cmd:        SCSI command
624  * @result:     scsi error code
625  *
626  * Translate a SCSI result code into a blk_status_t value. May reset the host
627  * byte of @cmd->result.
628  */
629 static blk_status_t scsi_result_to_blk_status(struct scsi_cmnd *cmd, int result)
630 {
631         switch (host_byte(result)) {
632         case DID_OK:
633                 /*
634                  * Also check the other bytes than the status byte in result
635                  * to handle the case when a SCSI LLD sets result to
636                  * DRIVER_SENSE << 24 without setting SAM_STAT_CHECK_CONDITION.
637                  */
638                 if (scsi_status_is_good(result) && (result & ~0xff) == 0)
639                         return BLK_STS_OK;
640                 return BLK_STS_IOERR;
641         case DID_TRANSPORT_FAILFAST:
642                 return BLK_STS_TRANSPORT;
643         case DID_TARGET_FAILURE:
644                 set_host_byte(cmd, DID_OK);
645                 return BLK_STS_TARGET;
646         case DID_NEXUS_FAILURE:
647                 set_host_byte(cmd, DID_OK);
648                 return BLK_STS_NEXUS;
649         case DID_ALLOC_FAILURE:
650                 set_host_byte(cmd, DID_OK);
651                 return BLK_STS_NOSPC;
652         case DID_MEDIUM_ERROR:
653                 set_host_byte(cmd, DID_OK);
654                 return BLK_STS_MEDIUM;
655         default:
656                 return BLK_STS_IOERR;
657         }
658 }
659
660 /* Helper for scsi_io_completion() when "reprep" action required. */
661 static void scsi_io_completion_reprep(struct scsi_cmnd *cmd,
662                                       struct request_queue *q)
663 {
664         /* A new command will be prepared and issued. */
665         scsi_mq_requeue_cmd(cmd);
666 }
667
668 /* Helper for scsi_io_completion() when special action required. */
669 static void scsi_io_completion_action(struct scsi_cmnd *cmd, int result)
670 {
671         struct request_queue *q = cmd->device->request_queue;
672         struct request *req = cmd->request;
673         int level = 0;
674         enum {ACTION_FAIL, ACTION_REPREP, ACTION_RETRY,
675               ACTION_DELAYED_RETRY} action;
676         unsigned long wait_for = (cmd->allowed + 1) * req->timeout;
677         struct scsi_sense_hdr sshdr;
678         bool sense_valid;
679         bool sense_current = true;      /* false implies "deferred sense" */
680         blk_status_t blk_stat;
681
682         sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
683         if (sense_valid)
684                 sense_current = !scsi_sense_is_deferred(&sshdr);
685
686         blk_stat = scsi_result_to_blk_status(cmd, result);
687
688         if (host_byte(result) == DID_RESET) {
689                 /* Third party bus reset or reset for error recovery
690                  * reasons.  Just retry the command and see what
691                  * happens.
692                  */
693                 action = ACTION_RETRY;
694         } else if (sense_valid && sense_current) {
695                 switch (sshdr.sense_key) {
696                 case UNIT_ATTENTION:
697                         if (cmd->device->removable) {
698                                 /* Detected disc change.  Set a bit
699                                  * and quietly refuse further access.
700                                  */
701                                 cmd->device->changed = 1;
702                                 action = ACTION_FAIL;
703                         } else {
704                                 /* Must have been a power glitch, or a
705                                  * bus reset.  Could not have been a
706                                  * media change, so we just retry the
707                                  * command and see what happens.
708                                  */
709                                 action = ACTION_RETRY;
710                         }
711                         break;
712                 case ILLEGAL_REQUEST:
713                         /* If we had an ILLEGAL REQUEST returned, then
714                          * we may have performed an unsupported
715                          * command.  The only thing this should be
716                          * would be a ten byte read where only a six
717                          * byte read was supported.  Also, on a system
718                          * where READ CAPACITY failed, we may have
719                          * read past the end of the disk.
720                          */
721                         if ((cmd->device->use_10_for_rw &&
722                             sshdr.asc == 0x20 && sshdr.ascq == 0x00) &&
723                             (cmd->cmnd[0] == READ_10 ||
724                              cmd->cmnd[0] == WRITE_10)) {
725                                 /* This will issue a new 6-byte command. */
726                                 cmd->device->use_10_for_rw = 0;
727                                 action = ACTION_REPREP;
728                         } else if (sshdr.asc == 0x10) /* DIX */ {
729                                 action = ACTION_FAIL;
730                                 blk_stat = BLK_STS_PROTECTION;
731                         /* INVALID COMMAND OPCODE or INVALID FIELD IN CDB */
732                         } else if (sshdr.asc == 0x20 || sshdr.asc == 0x24) {
733                                 action = ACTION_FAIL;
734                                 blk_stat = BLK_STS_TARGET;
735                         } else
736                                 action = ACTION_FAIL;
737                         break;
738                 case ABORTED_COMMAND:
739                         action = ACTION_FAIL;
740                         if (sshdr.asc == 0x10) /* DIF */
741                                 blk_stat = BLK_STS_PROTECTION;
742                         break;
743                 case NOT_READY:
744                         /* If the device is in the process of becoming
745                          * ready, or has a temporary blockage, retry.
746                          */
747                         if (sshdr.asc == 0x04) {
748                                 switch (sshdr.ascq) {
749                                 case 0x01: /* becoming ready */
750                                 case 0x04: /* format in progress */
751                                 case 0x05: /* rebuild in progress */
752                                 case 0x06: /* recalculation in progress */
753                                 case 0x07: /* operation in progress */
754                                 case 0x08: /* Long write in progress */
755                                 case 0x09: /* self test in progress */
756                                 case 0x14: /* space allocation in progress */
757                                 case 0x1a: /* start stop unit in progress */
758                                 case 0x1b: /* sanitize in progress */
759                                 case 0x1d: /* configuration in progress */
760                                 case 0x24: /* depopulation in progress */
761                                         action = ACTION_DELAYED_RETRY;
762                                         break;
763                                 default:
764                                         action = ACTION_FAIL;
765                                         break;
766                                 }
767                         } else
768                                 action = ACTION_FAIL;
769                         break;
770                 case VOLUME_OVERFLOW:
771                         /* See SSC3rXX or current. */
772                         action = ACTION_FAIL;
773                         break;
774                 default:
775                         action = ACTION_FAIL;
776                         break;
777                 }
778         } else
779                 action = ACTION_FAIL;
780
781         if (action != ACTION_FAIL &&
782             time_before(cmd->jiffies_at_alloc + wait_for, jiffies))
783                 action = ACTION_FAIL;
784
785         switch (action) {
786         case ACTION_FAIL:
787                 /* Give up and fail the remainder of the request */
788                 if (!(req->rq_flags & RQF_QUIET)) {
789                         static DEFINE_RATELIMIT_STATE(_rs,
790                                         DEFAULT_RATELIMIT_INTERVAL,
791                                         DEFAULT_RATELIMIT_BURST);
792
793                         if (unlikely(scsi_logging_level))
794                                 level =
795                                      SCSI_LOG_LEVEL(SCSI_LOG_MLCOMPLETE_SHIFT,
796                                                     SCSI_LOG_MLCOMPLETE_BITS);
797
798                         /*
799                          * if logging is enabled the failure will be printed
800                          * in scsi_log_completion(), so avoid duplicate messages
801                          */
802                         if (!level && __ratelimit(&_rs)) {
803                                 scsi_print_result(cmd, NULL, FAILED);
804                                 if (driver_byte(result) == DRIVER_SENSE)
805                                         scsi_print_sense(cmd);
806                                 scsi_print_command(cmd);
807                         }
808                 }
809                 if (!scsi_end_request(req, blk_stat, blk_rq_err_bytes(req)))
810                         return;
811                 /*FALLTHRU*/
812         case ACTION_REPREP:
813                 scsi_io_completion_reprep(cmd, q);
814                 break;
815         case ACTION_RETRY:
816                 /* Retry the same command immediately */
817                 __scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY, false);
818                 break;
819         case ACTION_DELAYED_RETRY:
820                 /* Retry the same command after a delay */
821                 __scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY, false);
822                 break;
823         }
824 }
825
826 /*
827  * Helper for scsi_io_completion() when cmd->result is non-zero. Returns a
828  * new result that may suppress further error checking. Also modifies
829  * *blk_statp in some cases.
830  */
831 static int scsi_io_completion_nz_result(struct scsi_cmnd *cmd, int result,
832                                         blk_status_t *blk_statp)
833 {
834         bool sense_valid;
835         bool sense_current = true;      /* false implies "deferred sense" */
836         struct request *req = cmd->request;
837         struct scsi_sense_hdr sshdr;
838
839         sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
840         if (sense_valid)
841                 sense_current = !scsi_sense_is_deferred(&sshdr);
842
843         if (blk_rq_is_passthrough(req)) {
844                 if (sense_valid) {
845                         /*
846                          * SG_IO wants current and deferred errors
847                          */
848                         scsi_req(req)->sense_len =
849                                 min(8 + cmd->sense_buffer[7],
850                                     SCSI_SENSE_BUFFERSIZE);
851                 }
852                 if (sense_current)
853                         *blk_statp = scsi_result_to_blk_status(cmd, result);
854         } else if (blk_rq_bytes(req) == 0 && sense_current) {
855                 /*
856                  * Flush commands do not transfers any data, and thus cannot use
857                  * good_bytes != blk_rq_bytes(req) as the signal for an error.
858                  * This sets *blk_statp explicitly for the problem case.
859                  */
860                 *blk_statp = scsi_result_to_blk_status(cmd, result);
861         }
862         /*
863          * Recovered errors need reporting, but they're always treated as
864          * success, so fiddle the result code here.  For passthrough requests
865          * we already took a copy of the original into sreq->result which
866          * is what gets returned to the user
867          */
868         if (sense_valid && (sshdr.sense_key == RECOVERED_ERROR)) {
869                 bool do_print = true;
870                 /*
871                  * if ATA PASS-THROUGH INFORMATION AVAILABLE [0x0, 0x1d]
872                  * skip print since caller wants ATA registers. Only occurs
873                  * on SCSI ATA PASS_THROUGH commands when CK_COND=1
874                  */
875                 if ((sshdr.asc == 0x0) && (sshdr.ascq == 0x1d))
876                         do_print = false;
877                 else if (req->rq_flags & RQF_QUIET)
878                         do_print = false;
879                 if (do_print)
880                         scsi_print_sense(cmd);
881                 result = 0;
882                 /* for passthrough, *blk_statp may be set */
883                 *blk_statp = BLK_STS_OK;
884         }
885         /*
886          * Another corner case: the SCSI status byte is non-zero but 'good'.
887          * Example: PRE-FETCH command returns SAM_STAT_CONDITION_MET when
888          * it is able to fit nominated LBs in its cache (and SAM_STAT_GOOD
889          * if it can't fit). Treat SAM_STAT_CONDITION_MET and the related
890          * intermediate statuses (both obsolete in SAM-4) as good.
891          */
892         if (status_byte(result) && scsi_status_is_good(result)) {
893                 result = 0;
894                 *blk_statp = BLK_STS_OK;
895         }
896         return result;
897 }
898
899 /*
900  * Function:    scsi_io_completion()
901  *
902  * Purpose:     Completion processing for block device I/O requests.
903  *
904  * Arguments:   cmd   - command that is finished.
905  *
906  * Lock status: Assumed that no lock is held upon entry.
907  *
908  * Returns:     Nothing
909  *
910  * Notes:       We will finish off the specified number of sectors.  If we
911  *              are done, the command block will be released and the queue
912  *              function will be goosed.  If we are not done then we have to
913  *              figure out what to do next:
914  *
915  *              a) We can call scsi_requeue_command().  The request
916  *                 will be unprepared and put back on the queue.  Then
917  *                 a new command will be created for it.  This should
918  *                 be used if we made forward progress, or if we want
919  *                 to switch from READ(10) to READ(6) for example.
920  *
921  *              b) We can call __scsi_queue_insert().  The request will
922  *                 be put back on the queue and retried using the same
923  *                 command as before, possibly after a delay.
924  *
925  *              c) We can call scsi_end_request() with blk_stat other than
926  *                 BLK_STS_OK, to fail the remainder of the request.
927  */
928 void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes)
929 {
930         int result = cmd->result;
931         struct request_queue *q = cmd->device->request_queue;
932         struct request *req = cmd->request;
933         blk_status_t blk_stat = BLK_STS_OK;
934
935         if (unlikely(result))   /* a nz result may or may not be an error */
936                 result = scsi_io_completion_nz_result(cmd, result, &blk_stat);
937
938         if (unlikely(blk_rq_is_passthrough(req))) {
939                 /*
940                  * scsi_result_to_blk_status may have reset the host_byte
941                  */
942                 scsi_req(req)->result = cmd->result;
943         }
944
945         /*
946          * Next deal with any sectors which we were able to correctly
947          * handle.
948          */
949         SCSI_LOG_HLCOMPLETE(1, scmd_printk(KERN_INFO, cmd,
950                 "%u sectors total, %d bytes done.\n",
951                 blk_rq_sectors(req), good_bytes));
952
953         /*
954          * Next deal with any sectors which we were able to correctly
955          * handle. Failed, zero length commands always need to drop down
956          * to retry code. Fast path should return in this block.
957          */
958         if (likely(blk_rq_bytes(req) > 0 || blk_stat == BLK_STS_OK)) {
959                 if (likely(!scsi_end_request(req, blk_stat, good_bytes)))
960                         return; /* no bytes remaining */
961         }
962
963         /* Kill remainder if no retries. */
964         if (unlikely(blk_stat && scsi_noretry_cmd(cmd))) {
965                 if (scsi_end_request(req, blk_stat, blk_rq_bytes(req)))
966                         WARN_ONCE(true,
967                             "Bytes remaining after failed, no-retry command");
968                 return;
969         }
970
971         /*
972          * If there had been no error, but we have leftover bytes in the
973          * requeues just queue the command up again.
974          */
975         if (likely(result == 0))
976                 scsi_io_completion_reprep(cmd, q);
977         else
978                 scsi_io_completion_action(cmd, result);
979 }
980
981 static inline bool scsi_cmd_needs_dma_drain(struct scsi_device *sdev,
982                 struct request *rq)
983 {
984         return sdev->dma_drain_len && blk_rq_is_passthrough(rq) &&
985                !op_is_write(req_op(rq)) &&
986                sdev->host->hostt->dma_need_drain(rq);
987 }
988
989 /*
990  * Function:    scsi_init_io()
991  *
992  * Purpose:     SCSI I/O initialize function.
993  *
994  * Arguments:   cmd   - Command descriptor we wish to initialize
995  *
996  * Returns:     BLK_STS_OK on success
997  *              BLK_STS_RESOURCE if the failure is retryable
998  *              BLK_STS_IOERR if the failure is fatal
999  */
1000 blk_status_t scsi_init_io(struct scsi_cmnd *cmd)
1001 {
1002         struct scsi_device *sdev = cmd->device;
1003         struct request *rq = cmd->request;
1004         unsigned short nr_segs = blk_rq_nr_phys_segments(rq);
1005         struct scatterlist *last_sg = NULL;
1006         blk_status_t ret;
1007         bool need_drain = scsi_cmd_needs_dma_drain(sdev, rq);
1008         int count;
1009
1010         if (WARN_ON_ONCE(!nr_segs))
1011                 return BLK_STS_IOERR;
1012
1013         /*
1014          * Make sure there is space for the drain.  The driver must adjust
1015          * max_hw_segments to be prepared for this.
1016          */
1017         if (need_drain)
1018                 nr_segs++;
1019
1020         /*
1021          * If sg table allocation fails, requeue request later.
1022          */
1023         if (unlikely(sg_alloc_table_chained(&cmd->sdb.table, nr_segs,
1024                         cmd->sdb.table.sgl, SCSI_INLINE_SG_CNT)))
1025                 return BLK_STS_RESOURCE;
1026
1027         /*
1028          * Next, walk the list, and fill in the addresses and sizes of
1029          * each segment.
1030          */
1031         count = __blk_rq_map_sg(rq->q, rq, cmd->sdb.table.sgl, &last_sg);
1032
1033         if (blk_rq_bytes(rq) & rq->q->dma_pad_mask) {
1034                 unsigned int pad_len =
1035                         (rq->q->dma_pad_mask & ~blk_rq_bytes(rq)) + 1;
1036
1037                 last_sg->length += pad_len;
1038                 cmd->extra_len += pad_len;
1039         }
1040
1041         if (need_drain) {
1042                 sg_unmark_end(last_sg);
1043                 last_sg = sg_next(last_sg);
1044                 sg_set_buf(last_sg, sdev->dma_drain_buf, sdev->dma_drain_len);
1045                 sg_mark_end(last_sg);
1046
1047                 cmd->extra_len += sdev->dma_drain_len;
1048                 count++;
1049         }
1050
1051         BUG_ON(count > cmd->sdb.table.nents);
1052         cmd->sdb.table.nents = count;
1053         cmd->sdb.length = blk_rq_payload_bytes(rq);
1054
1055         if (blk_integrity_rq(rq)) {
1056                 struct scsi_data_buffer *prot_sdb = cmd->prot_sdb;
1057                 int ivecs;
1058
1059                 if (WARN_ON_ONCE(!prot_sdb)) {
1060                         /*
1061                          * This can happen if someone (e.g. multipath)
1062                          * queues a command to a device on an adapter
1063                          * that does not support DIX.
1064                          */
1065                         ret = BLK_STS_IOERR;
1066                         goto out_free_sgtables;
1067                 }
1068
1069                 ivecs = blk_rq_count_integrity_sg(rq->q, rq->bio);
1070
1071                 if (sg_alloc_table_chained(&prot_sdb->table, ivecs,
1072                                 prot_sdb->table.sgl,
1073                                 SCSI_INLINE_PROT_SG_CNT)) {
1074                         ret = BLK_STS_RESOURCE;
1075                         goto out_free_sgtables;
1076                 }
1077
1078                 count = blk_rq_map_integrity_sg(rq->q, rq->bio,
1079                                                 prot_sdb->table.sgl);
1080                 BUG_ON(count > ivecs);
1081                 BUG_ON(count > queue_max_integrity_segments(rq->q));
1082
1083                 cmd->prot_sdb = prot_sdb;
1084                 cmd->prot_sdb->table.nents = count;
1085         }
1086
1087         return BLK_STS_OK;
1088 out_free_sgtables:
1089         scsi_mq_free_sgtables(cmd);
1090         return ret;
1091 }
1092 EXPORT_SYMBOL(scsi_init_io);
1093
1094 /**
1095  * scsi_initialize_rq - initialize struct scsi_cmnd partially
1096  * @rq: Request associated with the SCSI command to be initialized.
1097  *
1098  * This function initializes the members of struct scsi_cmnd that must be
1099  * initialized before request processing starts and that won't be
1100  * reinitialized if a SCSI command is requeued.
1101  *
1102  * Called from inside blk_get_request() for pass-through requests and from
1103  * inside scsi_init_command() for filesystem requests.
1104  */
1105 static void scsi_initialize_rq(struct request *rq)
1106 {
1107         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1108
1109         scsi_req_init(&cmd->req);
1110         init_rcu_head(&cmd->rcu);
1111         cmd->jiffies_at_alloc = jiffies;
1112         cmd->retries = 0;
1113 }
1114
1115 /*
1116  * Only called when the request isn't completed by SCSI, and not freed by
1117  * SCSI
1118  */
1119 static void scsi_cleanup_rq(struct request *rq)
1120 {
1121         if (rq->rq_flags & RQF_DONTPREP) {
1122                 scsi_mq_uninit_cmd(blk_mq_rq_to_pdu(rq));
1123                 rq->rq_flags &= ~RQF_DONTPREP;
1124         }
1125 }
1126
1127 /* Called before a request is prepared. See also scsi_mq_prep_fn(). */
1128 void scsi_init_command(struct scsi_device *dev, struct scsi_cmnd *cmd)
1129 {
1130         void *buf = cmd->sense_buffer;
1131         void *prot = cmd->prot_sdb;
1132         struct request *rq = blk_mq_rq_from_pdu(cmd);
1133         unsigned int flags = cmd->flags & SCMD_PRESERVED_FLAGS;
1134         unsigned long jiffies_at_alloc;
1135         int retries, to_clear;
1136         bool in_flight;
1137
1138         if (!blk_rq_is_scsi(rq) && !(flags & SCMD_INITIALIZED)) {
1139                 flags |= SCMD_INITIALIZED;
1140                 scsi_initialize_rq(rq);
1141         }
1142
1143         jiffies_at_alloc = cmd->jiffies_at_alloc;
1144         retries = cmd->retries;
1145         in_flight = test_bit(SCMD_STATE_INFLIGHT, &cmd->state);
1146         /*
1147          * Zero out the cmd, except for the embedded scsi_request. Only clear
1148          * the driver-private command data if the LLD does not supply a
1149          * function to initialize that data.
1150          */
1151         to_clear = sizeof(*cmd) - sizeof(cmd->req);
1152         if (!dev->host->hostt->init_cmd_priv)
1153                 to_clear += dev->host->hostt->cmd_size;
1154         memset((char *)cmd + sizeof(cmd->req), 0, to_clear);
1155
1156         cmd->device = dev;
1157         cmd->sense_buffer = buf;
1158         cmd->prot_sdb = prot;
1159         cmd->flags = flags;
1160         INIT_DELAYED_WORK(&cmd->abort_work, scmd_eh_abort_handler);
1161         cmd->jiffies_at_alloc = jiffies_at_alloc;
1162         cmd->retries = retries;
1163         if (in_flight)
1164                 __set_bit(SCMD_STATE_INFLIGHT, &cmd->state);
1165
1166 }
1167
1168 static blk_status_t scsi_setup_scsi_cmnd(struct scsi_device *sdev,
1169                 struct request *req)
1170 {
1171         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1172
1173         /*
1174          * Passthrough requests may transfer data, in which case they must
1175          * a bio attached to them.  Or they might contain a SCSI command
1176          * that does not transfer data, in which case they may optionally
1177          * submit a request without an attached bio.
1178          */
1179         if (req->bio) {
1180                 blk_status_t ret = scsi_init_io(cmd);
1181                 if (unlikely(ret != BLK_STS_OK))
1182                         return ret;
1183         } else {
1184                 BUG_ON(blk_rq_bytes(req));
1185
1186                 memset(&cmd->sdb, 0, sizeof(cmd->sdb));
1187         }
1188
1189         cmd->cmd_len = scsi_req(req)->cmd_len;
1190         cmd->cmnd = scsi_req(req)->cmd;
1191         cmd->transfersize = blk_rq_bytes(req);
1192         cmd->allowed = scsi_req(req)->retries;
1193         return BLK_STS_OK;
1194 }
1195
1196 /*
1197  * Setup a normal block command.  These are simple request from filesystems
1198  * that still need to be translated to SCSI CDBs from the ULD.
1199  */
1200 static blk_status_t scsi_setup_fs_cmnd(struct scsi_device *sdev,
1201                 struct request *req)
1202 {
1203         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1204
1205         if (unlikely(sdev->handler && sdev->handler->prep_fn)) {
1206                 blk_status_t ret = sdev->handler->prep_fn(sdev, req);
1207                 if (ret != BLK_STS_OK)
1208                         return ret;
1209         }
1210
1211         cmd->cmnd = scsi_req(req)->cmd = scsi_req(req)->__cmd;
1212         memset(cmd->cmnd, 0, BLK_MAX_CDB);
1213         return scsi_cmd_to_driver(cmd)->init_command(cmd);
1214 }
1215
1216 static blk_status_t scsi_setup_cmnd(struct scsi_device *sdev,
1217                 struct request *req)
1218 {
1219         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1220
1221         if (!blk_rq_bytes(req))
1222                 cmd->sc_data_direction = DMA_NONE;
1223         else if (rq_data_dir(req) == WRITE)
1224                 cmd->sc_data_direction = DMA_TO_DEVICE;
1225         else
1226                 cmd->sc_data_direction = DMA_FROM_DEVICE;
1227
1228         if (blk_rq_is_scsi(req))
1229                 return scsi_setup_scsi_cmnd(sdev, req);
1230         else
1231                 return scsi_setup_fs_cmnd(sdev, req);
1232 }
1233
1234 static blk_status_t
1235 scsi_prep_state_check(struct scsi_device *sdev, struct request *req)
1236 {
1237         switch (sdev->sdev_state) {
1238         case SDEV_OFFLINE:
1239         case SDEV_TRANSPORT_OFFLINE:
1240                 /*
1241                  * If the device is offline we refuse to process any
1242                  * commands.  The device must be brought online
1243                  * before trying any recovery commands.
1244                  */
1245                 if (!sdev->offline_already) {
1246                         sdev->offline_already = true;
1247                         sdev_printk(KERN_ERR, sdev,
1248                                     "rejecting I/O to offline device\n");
1249                 }
1250                 return BLK_STS_IOERR;
1251         case SDEV_DEL:
1252                 /*
1253                  * If the device is fully deleted, we refuse to
1254                  * process any commands as well.
1255                  */
1256                 sdev_printk(KERN_ERR, sdev,
1257                             "rejecting I/O to dead device\n");
1258                 return BLK_STS_IOERR;
1259         case SDEV_BLOCK:
1260         case SDEV_CREATED_BLOCK:
1261                 return BLK_STS_RESOURCE;
1262         case SDEV_QUIESCE:
1263                 /*
1264                  * If the devices is blocked we defer normal commands.
1265                  */
1266                 if (req && !(req->rq_flags & RQF_PREEMPT))
1267                         return BLK_STS_RESOURCE;
1268                 return BLK_STS_OK;
1269         default:
1270                 /*
1271                  * For any other not fully online state we only allow
1272                  * special commands.  In particular any user initiated
1273                  * command is not allowed.
1274                  */
1275                 if (req && !(req->rq_flags & RQF_PREEMPT))
1276                         return BLK_STS_IOERR;
1277                 return BLK_STS_OK;
1278         }
1279 }
1280
1281 /*
1282  * scsi_dev_queue_ready: if we can send requests to sdev, return 1 else
1283  * return 0.
1284  *
1285  * Called with the queue_lock held.
1286  */
1287 static inline int scsi_dev_queue_ready(struct request_queue *q,
1288                                   struct scsi_device *sdev)
1289 {
1290         unsigned int busy;
1291
1292         busy = atomic_inc_return(&sdev->device_busy) - 1;
1293         if (atomic_read(&sdev->device_blocked)) {
1294                 if (busy)
1295                         goto out_dec;
1296
1297                 /*
1298                  * unblock after device_blocked iterates to zero
1299                  */
1300                 if (atomic_dec_return(&sdev->device_blocked) > 0)
1301                         goto out_dec;
1302                 SCSI_LOG_MLQUEUE(3, sdev_printk(KERN_INFO, sdev,
1303                                    "unblocking device at zero depth\n"));
1304         }
1305
1306         if (busy >= sdev->queue_depth)
1307                 goto out_dec;
1308
1309         return 1;
1310 out_dec:
1311         atomic_dec(&sdev->device_busy);
1312         return 0;
1313 }
1314
1315 /*
1316  * scsi_target_queue_ready: checks if there we can send commands to target
1317  * @sdev: scsi device on starget to check.
1318  */
1319 static inline int scsi_target_queue_ready(struct Scsi_Host *shost,
1320                                            struct scsi_device *sdev)
1321 {
1322         struct scsi_target *starget = scsi_target(sdev);
1323         unsigned int busy;
1324
1325         if (starget->single_lun) {
1326                 spin_lock_irq(shost->host_lock);
1327                 if (starget->starget_sdev_user &&
1328                     starget->starget_sdev_user != sdev) {
1329                         spin_unlock_irq(shost->host_lock);
1330                         return 0;
1331                 }
1332                 starget->starget_sdev_user = sdev;
1333                 spin_unlock_irq(shost->host_lock);
1334         }
1335
1336         if (starget->can_queue <= 0)
1337                 return 1;
1338
1339         busy = atomic_inc_return(&starget->target_busy) - 1;
1340         if (atomic_read(&starget->target_blocked) > 0) {
1341                 if (busy)
1342                         goto starved;
1343
1344                 /*
1345                  * unblock after target_blocked iterates to zero
1346                  */
1347                 if (atomic_dec_return(&starget->target_blocked) > 0)
1348                         goto out_dec;
1349
1350                 SCSI_LOG_MLQUEUE(3, starget_printk(KERN_INFO, starget,
1351                                  "unblocking target at zero depth\n"));
1352         }
1353
1354         if (busy >= starget->can_queue)
1355                 goto starved;
1356
1357         return 1;
1358
1359 starved:
1360         spin_lock_irq(shost->host_lock);
1361         list_move_tail(&sdev->starved_entry, &shost->starved_list);
1362         spin_unlock_irq(shost->host_lock);
1363 out_dec:
1364         if (starget->can_queue > 0)
1365                 atomic_dec(&starget->target_busy);
1366         return 0;
1367 }
1368
1369 /*
1370  * scsi_host_queue_ready: if we can send requests to shost, return 1 else
1371  * return 0. We must end up running the queue again whenever 0 is
1372  * returned, else IO can hang.
1373  */
1374 static inline int scsi_host_queue_ready(struct request_queue *q,
1375                                    struct Scsi_Host *shost,
1376                                    struct scsi_device *sdev,
1377                                    struct scsi_cmnd *cmd)
1378 {
1379         if (scsi_host_in_recovery(shost))
1380                 return 0;
1381
1382         if (atomic_read(&shost->host_blocked) > 0) {
1383                 if (scsi_host_busy(shost) > 0)
1384                         goto starved;
1385
1386                 /*
1387                  * unblock after host_blocked iterates to zero
1388                  */
1389                 if (atomic_dec_return(&shost->host_blocked) > 0)
1390                         goto out_dec;
1391
1392                 SCSI_LOG_MLQUEUE(3,
1393                         shost_printk(KERN_INFO, shost,
1394                                      "unblocking host at zero depth\n"));
1395         }
1396
1397         if (shost->host_self_blocked)
1398                 goto starved;
1399
1400         /* We're OK to process the command, so we can't be starved */
1401         if (!list_empty(&sdev->starved_entry)) {
1402                 spin_lock_irq(shost->host_lock);
1403                 if (!list_empty(&sdev->starved_entry))
1404                         list_del_init(&sdev->starved_entry);
1405                 spin_unlock_irq(shost->host_lock);
1406         }
1407
1408         __set_bit(SCMD_STATE_INFLIGHT, &cmd->state);
1409
1410         return 1;
1411
1412 starved:
1413         spin_lock_irq(shost->host_lock);
1414         if (list_empty(&sdev->starved_entry))
1415                 list_add_tail(&sdev->starved_entry, &shost->starved_list);
1416         spin_unlock_irq(shost->host_lock);
1417 out_dec:
1418         scsi_dec_host_busy(shost, cmd);
1419         return 0;
1420 }
1421
1422 /*
1423  * Busy state exporting function for request stacking drivers.
1424  *
1425  * For efficiency, no lock is taken to check the busy state of
1426  * shost/starget/sdev, since the returned value is not guaranteed and
1427  * may be changed after request stacking drivers call the function,
1428  * regardless of taking lock or not.
1429  *
1430  * When scsi can't dispatch I/Os anymore and needs to kill I/Os scsi
1431  * needs to return 'not busy'. Otherwise, request stacking drivers
1432  * may hold requests forever.
1433  */
1434 static bool scsi_mq_lld_busy(struct request_queue *q)
1435 {
1436         struct scsi_device *sdev = q->queuedata;
1437         struct Scsi_Host *shost;
1438
1439         if (blk_queue_dying(q))
1440                 return false;
1441
1442         shost = sdev->host;
1443
1444         /*
1445          * Ignore host/starget busy state.
1446          * Since block layer does not have a concept of fairness across
1447          * multiple queues, congestion of host/starget needs to be handled
1448          * in SCSI layer.
1449          */
1450         if (scsi_host_in_recovery(shost) || scsi_device_is_busy(sdev))
1451                 return true;
1452
1453         return false;
1454 }
1455
1456 static void scsi_softirq_done(struct request *rq)
1457 {
1458         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1459         unsigned long wait_for = (cmd->allowed + 1) * rq->timeout;
1460         int disposition;
1461
1462         INIT_LIST_HEAD(&cmd->eh_entry);
1463
1464         atomic_inc(&cmd->device->iodone_cnt);
1465         if (cmd->result)
1466                 atomic_inc(&cmd->device->ioerr_cnt);
1467
1468         disposition = scsi_decide_disposition(cmd);
1469         if (disposition != SUCCESS &&
1470             time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) {
1471                 scmd_printk(KERN_ERR, cmd,
1472                             "timing out command, waited %lus\n",
1473                             wait_for/HZ);
1474                 disposition = SUCCESS;
1475         }
1476
1477         scsi_log_completion(cmd, disposition);
1478
1479         switch (disposition) {
1480                 case SUCCESS:
1481                         scsi_finish_command(cmd);
1482                         break;
1483                 case NEEDS_RETRY:
1484                         scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY);
1485                         break;
1486                 case ADD_TO_MLQUEUE:
1487                         scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY);
1488                         break;
1489                 default:
1490                         scsi_eh_scmd_add(cmd);
1491                         break;
1492         }
1493 }
1494
1495 /**
1496  * scsi_dispatch_command - Dispatch a command to the low-level driver.
1497  * @cmd: command block we are dispatching.
1498  *
1499  * Return: nonzero return request was rejected and device's queue needs to be
1500  * plugged.
1501  */
1502 static int scsi_dispatch_cmd(struct scsi_cmnd *cmd)
1503 {
1504         struct Scsi_Host *host = cmd->device->host;
1505         int rtn = 0;
1506
1507         atomic_inc(&cmd->device->iorequest_cnt);
1508
1509         /* check if the device is still usable */
1510         if (unlikely(cmd->device->sdev_state == SDEV_DEL)) {
1511                 /* in SDEV_DEL we error all commands. DID_NO_CONNECT
1512                  * returns an immediate error upwards, and signals
1513                  * that the device is no longer present */
1514                 cmd->result = DID_NO_CONNECT << 16;
1515                 goto done;
1516         }
1517
1518         /* Check to see if the scsi lld made this device blocked. */
1519         if (unlikely(scsi_device_blocked(cmd->device))) {
1520                 /*
1521                  * in blocked state, the command is just put back on
1522                  * the device queue.  The suspend state has already
1523                  * blocked the queue so future requests should not
1524                  * occur until the device transitions out of the
1525                  * suspend state.
1526                  */
1527                 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1528                         "queuecommand : device blocked\n"));
1529                 return SCSI_MLQUEUE_DEVICE_BUSY;
1530         }
1531
1532         /* Store the LUN value in cmnd, if needed. */
1533         if (cmd->device->lun_in_cdb)
1534                 cmd->cmnd[1] = (cmd->cmnd[1] & 0x1f) |
1535                                (cmd->device->lun << 5 & 0xe0);
1536
1537         scsi_log_send(cmd);
1538
1539         /*
1540          * Before we queue this command, check if the command
1541          * length exceeds what the host adapter can handle.
1542          */
1543         if (cmd->cmd_len > cmd->device->host->max_cmd_len) {
1544                 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1545                                "queuecommand : command too long. "
1546                                "cdb_size=%d host->max_cmd_len=%d\n",
1547                                cmd->cmd_len, cmd->device->host->max_cmd_len));
1548                 cmd->result = (DID_ABORT << 16);
1549                 goto done;
1550         }
1551
1552         if (unlikely(host->shost_state == SHOST_DEL)) {
1553                 cmd->result = (DID_NO_CONNECT << 16);
1554                 goto done;
1555
1556         }
1557
1558         trace_scsi_dispatch_cmd_start(cmd);
1559         rtn = host->hostt->queuecommand(host, cmd);
1560         if (rtn) {
1561                 trace_scsi_dispatch_cmd_error(cmd, rtn);
1562                 if (rtn != SCSI_MLQUEUE_DEVICE_BUSY &&
1563                     rtn != SCSI_MLQUEUE_TARGET_BUSY)
1564                         rtn = SCSI_MLQUEUE_HOST_BUSY;
1565
1566                 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1567                         "queuecommand : request rejected\n"));
1568         }
1569
1570         return rtn;
1571  done:
1572         cmd->scsi_done(cmd);
1573         return 0;
1574 }
1575
1576 /* Size in bytes of the sg-list stored in the scsi-mq command-private data. */
1577 static unsigned int scsi_mq_inline_sgl_size(struct Scsi_Host *shost)
1578 {
1579         return min_t(unsigned int, shost->sg_tablesize, SCSI_INLINE_SG_CNT) *
1580                 sizeof(struct scatterlist);
1581 }
1582
1583 static blk_status_t scsi_mq_prep_fn(struct request *req)
1584 {
1585         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1586         struct scsi_device *sdev = req->q->queuedata;
1587         struct Scsi_Host *shost = sdev->host;
1588         struct scatterlist *sg;
1589
1590         scsi_init_command(sdev, cmd);
1591
1592         cmd->request = req;
1593         cmd->tag = req->tag;
1594         cmd->prot_op = SCSI_PROT_NORMAL;
1595
1596         sg = (void *)cmd + sizeof(struct scsi_cmnd) + shost->hostt->cmd_size;
1597         cmd->sdb.table.sgl = sg;
1598
1599         if (scsi_host_get_prot(shost)) {
1600                 memset(cmd->prot_sdb, 0, sizeof(struct scsi_data_buffer));
1601
1602                 cmd->prot_sdb->table.sgl =
1603                         (struct scatterlist *)(cmd->prot_sdb + 1);
1604         }
1605
1606         blk_mq_start_request(req);
1607
1608         return scsi_setup_cmnd(sdev, req);
1609 }
1610
1611 static void scsi_mq_done(struct scsi_cmnd *cmd)
1612 {
1613         if (unlikely(test_and_set_bit(SCMD_STATE_COMPLETE, &cmd->state)))
1614                 return;
1615         trace_scsi_dispatch_cmd_done(cmd);
1616
1617         /*
1618          * If the block layer didn't complete the request due to a timeout
1619          * injection, scsi must clear its internal completed state so that the
1620          * timeout handler will see it needs to escalate its own error
1621          * recovery.
1622          */
1623         if (unlikely(!blk_mq_complete_request(cmd->request)))
1624                 clear_bit(SCMD_STATE_COMPLETE, &cmd->state);
1625 }
1626
1627 static void scsi_mq_put_budget(struct blk_mq_hw_ctx *hctx)
1628 {
1629         struct request_queue *q = hctx->queue;
1630         struct scsi_device *sdev = q->queuedata;
1631
1632         atomic_dec(&sdev->device_busy);
1633 }
1634
1635 static bool scsi_mq_get_budget(struct blk_mq_hw_ctx *hctx)
1636 {
1637         struct request_queue *q = hctx->queue;
1638         struct scsi_device *sdev = q->queuedata;
1639
1640         return scsi_dev_queue_ready(q, sdev);
1641 }
1642
1643 static blk_status_t scsi_queue_rq(struct blk_mq_hw_ctx *hctx,
1644                          const struct blk_mq_queue_data *bd)
1645 {
1646         struct request *req = bd->rq;
1647         struct request_queue *q = req->q;
1648         struct scsi_device *sdev = q->queuedata;
1649         struct Scsi_Host *shost = sdev->host;
1650         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1651         blk_status_t ret;
1652         int reason;
1653
1654         /*
1655          * If the device is not in running state we will reject some or all
1656          * commands.
1657          */
1658         if (unlikely(sdev->sdev_state != SDEV_RUNNING)) {
1659                 ret = scsi_prep_state_check(sdev, req);
1660                 if (ret != BLK_STS_OK)
1661                         goto out_put_budget;
1662         }
1663
1664         ret = BLK_STS_RESOURCE;
1665         if (!scsi_target_queue_ready(shost, sdev))
1666                 goto out_put_budget;
1667         if (!scsi_host_queue_ready(q, shost, sdev, cmd))
1668                 goto out_dec_target_busy;
1669
1670         if (!(req->rq_flags & RQF_DONTPREP)) {
1671                 ret = scsi_mq_prep_fn(req);
1672                 if (ret != BLK_STS_OK)
1673                         goto out_dec_host_busy;
1674                 req->rq_flags |= RQF_DONTPREP;
1675         } else {
1676                 clear_bit(SCMD_STATE_COMPLETE, &cmd->state);
1677                 blk_mq_start_request(req);
1678         }
1679
1680         cmd->flags &= SCMD_PRESERVED_FLAGS;
1681         if (sdev->simple_tags)
1682                 cmd->flags |= SCMD_TAGGED;
1683         if (bd->last)
1684                 cmd->flags |= SCMD_LAST;
1685
1686         scsi_init_cmd_errh(cmd);
1687         cmd->scsi_done = scsi_mq_done;
1688
1689         reason = scsi_dispatch_cmd(cmd);
1690         if (reason) {
1691                 scsi_set_blocked(cmd, reason);
1692                 ret = BLK_STS_RESOURCE;
1693                 goto out_dec_host_busy;
1694         }
1695
1696         return BLK_STS_OK;
1697
1698 out_dec_host_busy:
1699         scsi_dec_host_busy(shost, cmd);
1700 out_dec_target_busy:
1701         if (scsi_target(sdev)->can_queue > 0)
1702                 atomic_dec(&scsi_target(sdev)->target_busy);
1703 out_put_budget:
1704         scsi_mq_put_budget(hctx);
1705         switch (ret) {
1706         case BLK_STS_OK:
1707                 break;
1708         case BLK_STS_RESOURCE:
1709         case BLK_STS_ZONE_RESOURCE:
1710                 if (atomic_read(&sdev->device_busy) ||
1711                     scsi_device_blocked(sdev))
1712                         ret = BLK_STS_DEV_RESOURCE;
1713                 break;
1714         default:
1715                 if (unlikely(!scsi_device_online(sdev)))
1716                         scsi_req(req)->result = DID_NO_CONNECT << 16;
1717                 else
1718                         scsi_req(req)->result = DID_ERROR << 16;
1719                 /*
1720                  * Make sure to release all allocated resources when
1721                  * we hit an error, as we will never see this command
1722                  * again.
1723                  */
1724                 if (req->rq_flags & RQF_DONTPREP)
1725                         scsi_mq_uninit_cmd(cmd);
1726                 break;
1727         }
1728         return ret;
1729 }
1730
1731 static enum blk_eh_timer_return scsi_timeout(struct request *req,
1732                 bool reserved)
1733 {
1734         if (reserved)
1735                 return BLK_EH_RESET_TIMER;
1736         return scsi_times_out(req);
1737 }
1738
1739 static int scsi_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
1740                                 unsigned int hctx_idx, unsigned int numa_node)
1741 {
1742         struct Scsi_Host *shost = set->driver_data;
1743         const bool unchecked_isa_dma = shost->unchecked_isa_dma;
1744         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1745         struct scatterlist *sg;
1746         int ret = 0;
1747
1748         if (unchecked_isa_dma)
1749                 cmd->flags |= SCMD_UNCHECKED_ISA_DMA;
1750         cmd->sense_buffer = scsi_alloc_sense_buffer(unchecked_isa_dma,
1751                                                     GFP_KERNEL, numa_node);
1752         if (!cmd->sense_buffer)
1753                 return -ENOMEM;
1754         cmd->req.sense = cmd->sense_buffer;
1755
1756         if (scsi_host_get_prot(shost)) {
1757                 sg = (void *)cmd + sizeof(struct scsi_cmnd) +
1758                         shost->hostt->cmd_size;
1759                 cmd->prot_sdb = (void *)sg + scsi_mq_inline_sgl_size(shost);
1760         }
1761
1762         if (shost->hostt->init_cmd_priv) {
1763                 ret = shost->hostt->init_cmd_priv(shost, cmd);
1764                 if (ret < 0)
1765                         scsi_free_sense_buffer(unchecked_isa_dma,
1766                                                cmd->sense_buffer);
1767         }
1768
1769         return ret;
1770 }
1771
1772 static void scsi_mq_exit_request(struct blk_mq_tag_set *set, struct request *rq,
1773                                  unsigned int hctx_idx)
1774 {
1775         struct Scsi_Host *shost = set->driver_data;
1776         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1777
1778         if (shost->hostt->exit_cmd_priv)
1779                 shost->hostt->exit_cmd_priv(shost, cmd);
1780         scsi_free_sense_buffer(cmd->flags & SCMD_UNCHECKED_ISA_DMA,
1781                                cmd->sense_buffer);
1782 }
1783
1784 static int scsi_map_queues(struct blk_mq_tag_set *set)
1785 {
1786         struct Scsi_Host *shost = container_of(set, struct Scsi_Host, tag_set);
1787
1788         if (shost->hostt->map_queues)
1789                 return shost->hostt->map_queues(shost);
1790         return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
1791 }
1792
1793 void __scsi_init_queue(struct Scsi_Host *shost, struct request_queue *q)
1794 {
1795         struct device *dev = shost->dma_dev;
1796
1797         /*
1798          * this limit is imposed by hardware restrictions
1799          */
1800         blk_queue_max_segments(q, min_t(unsigned short, shost->sg_tablesize,
1801                                         SG_MAX_SEGMENTS));
1802
1803         if (scsi_host_prot_dma(shost)) {
1804                 shost->sg_prot_tablesize =
1805                         min_not_zero(shost->sg_prot_tablesize,
1806                                      (unsigned short)SCSI_MAX_PROT_SG_SEGMENTS);
1807                 BUG_ON(shost->sg_prot_tablesize < shost->sg_tablesize);
1808                 blk_queue_max_integrity_segments(q, shost->sg_prot_tablesize);
1809         }
1810
1811         if (dev->dma_mask) {
1812                 shost->max_sectors = min_t(unsigned int, shost->max_sectors,
1813                                 dma_max_mapping_size(dev) >> SECTOR_SHIFT);
1814         }
1815         blk_queue_max_hw_sectors(q, shost->max_sectors);
1816         if (shost->unchecked_isa_dma)
1817                 blk_queue_bounce_limit(q, BLK_BOUNCE_ISA);
1818         blk_queue_segment_boundary(q, shost->dma_boundary);
1819         dma_set_seg_boundary(dev, shost->dma_boundary);
1820
1821         blk_queue_max_segment_size(q, shost->max_segment_size);
1822         blk_queue_virt_boundary(q, shost->virt_boundary_mask);
1823         dma_set_max_seg_size(dev, queue_max_segment_size(q));
1824
1825         /*
1826          * Set a reasonable default alignment:  The larger of 32-byte (dword),
1827          * which is a common minimum for HBAs, and the minimum DMA alignment,
1828          * which is set by the platform.
1829          *
1830          * Devices that require a bigger alignment can increase it later.
1831          */
1832         blk_queue_dma_alignment(q, max(4, dma_get_cache_alignment()) - 1);
1833 }
1834 EXPORT_SYMBOL_GPL(__scsi_init_queue);
1835
1836 static const struct blk_mq_ops scsi_mq_ops_no_commit = {
1837         .get_budget     = scsi_mq_get_budget,
1838         .put_budget     = scsi_mq_put_budget,
1839         .queue_rq       = scsi_queue_rq,
1840         .complete       = scsi_softirq_done,
1841         .timeout        = scsi_timeout,
1842 #ifdef CONFIG_BLK_DEBUG_FS
1843         .show_rq        = scsi_show_rq,
1844 #endif
1845         .init_request   = scsi_mq_init_request,
1846         .exit_request   = scsi_mq_exit_request,
1847         .initialize_rq_fn = scsi_initialize_rq,
1848         .cleanup_rq     = scsi_cleanup_rq,
1849         .busy           = scsi_mq_lld_busy,
1850         .map_queues     = scsi_map_queues,
1851 };
1852
1853
1854 static void scsi_commit_rqs(struct blk_mq_hw_ctx *hctx)
1855 {
1856         struct request_queue *q = hctx->queue;
1857         struct scsi_device *sdev = q->queuedata;
1858         struct Scsi_Host *shost = sdev->host;
1859
1860         shost->hostt->commit_rqs(shost, hctx->queue_num);
1861 }
1862
1863 static const struct blk_mq_ops scsi_mq_ops = {
1864         .get_budget     = scsi_mq_get_budget,
1865         .put_budget     = scsi_mq_put_budget,
1866         .queue_rq       = scsi_queue_rq,
1867         .commit_rqs     = scsi_commit_rqs,
1868         .complete       = scsi_softirq_done,
1869         .timeout        = scsi_timeout,
1870 #ifdef CONFIG_BLK_DEBUG_FS
1871         .show_rq        = scsi_show_rq,
1872 #endif
1873         .init_request   = scsi_mq_init_request,
1874         .exit_request   = scsi_mq_exit_request,
1875         .initialize_rq_fn = scsi_initialize_rq,
1876         .cleanup_rq     = scsi_cleanup_rq,
1877         .busy           = scsi_mq_lld_busy,
1878         .map_queues     = scsi_map_queues,
1879 };
1880
1881 struct request_queue *scsi_mq_alloc_queue(struct scsi_device *sdev)
1882 {
1883         sdev->request_queue = blk_mq_init_queue(&sdev->host->tag_set);
1884         if (IS_ERR(sdev->request_queue))
1885                 return NULL;
1886
1887         sdev->request_queue->queuedata = sdev;
1888         __scsi_init_queue(sdev->host, sdev->request_queue);
1889         blk_queue_flag_set(QUEUE_FLAG_SCSI_PASSTHROUGH, sdev->request_queue);
1890         return sdev->request_queue;
1891 }
1892
1893 int scsi_mq_setup_tags(struct Scsi_Host *shost)
1894 {
1895         unsigned int cmd_size, sgl_size;
1896
1897         sgl_size = max_t(unsigned int, sizeof(struct scatterlist),
1898                                 scsi_mq_inline_sgl_size(shost));
1899         cmd_size = sizeof(struct scsi_cmnd) + shost->hostt->cmd_size + sgl_size;
1900         if (scsi_host_get_prot(shost))
1901                 cmd_size += sizeof(struct scsi_data_buffer) +
1902                         sizeof(struct scatterlist) * SCSI_INLINE_PROT_SG_CNT;
1903
1904         memset(&shost->tag_set, 0, sizeof(shost->tag_set));
1905         if (shost->hostt->commit_rqs)
1906                 shost->tag_set.ops = &scsi_mq_ops;
1907         else
1908                 shost->tag_set.ops = &scsi_mq_ops_no_commit;
1909         shost->tag_set.nr_hw_queues = shost->nr_hw_queues ? : 1;
1910         shost->tag_set.queue_depth = shost->can_queue;
1911         shost->tag_set.cmd_size = cmd_size;
1912         shost->tag_set.numa_node = NUMA_NO_NODE;
1913         shost->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
1914         shost->tag_set.flags |=
1915                 BLK_ALLOC_POLICY_TO_MQ_FLAG(shost->hostt->tag_alloc_policy);
1916         shost->tag_set.driver_data = shost;
1917
1918         return blk_mq_alloc_tag_set(&shost->tag_set);
1919 }
1920
1921 void scsi_mq_destroy_tags(struct Scsi_Host *shost)
1922 {
1923         blk_mq_free_tag_set(&shost->tag_set);
1924 }
1925
1926 /**
1927  * scsi_device_from_queue - return sdev associated with a request_queue
1928  * @q: The request queue to return the sdev from
1929  *
1930  * Return the sdev associated with a request queue or NULL if the
1931  * request_queue does not reference a SCSI device.
1932  */
1933 struct scsi_device *scsi_device_from_queue(struct request_queue *q)
1934 {
1935         struct scsi_device *sdev = NULL;
1936
1937         if (q->mq_ops == &scsi_mq_ops_no_commit ||
1938             q->mq_ops == &scsi_mq_ops)
1939                 sdev = q->queuedata;
1940         if (!sdev || !get_device(&sdev->sdev_gendev))
1941                 sdev = NULL;
1942
1943         return sdev;
1944 }
1945 EXPORT_SYMBOL_GPL(scsi_device_from_queue);
1946
1947 /*
1948  * Function:    scsi_block_requests()
1949  *
1950  * Purpose:     Utility function used by low-level drivers to prevent further
1951  *              commands from being queued to the device.
1952  *
1953  * Arguments:   shost       - Host in question
1954  *
1955  * Returns:     Nothing
1956  *
1957  * Lock status: No locks are assumed held.
1958  *
1959  * Notes:       There is no timer nor any other means by which the requests
1960  *              get unblocked other than the low-level driver calling
1961  *              scsi_unblock_requests().
1962  */
1963 void scsi_block_requests(struct Scsi_Host *shost)
1964 {
1965         shost->host_self_blocked = 1;
1966 }
1967 EXPORT_SYMBOL(scsi_block_requests);
1968
1969 /*
1970  * Function:    scsi_unblock_requests()
1971  *
1972  * Purpose:     Utility function used by low-level drivers to allow further
1973  *              commands from being queued to the device.
1974  *
1975  * Arguments:   shost       - Host in question
1976  *
1977  * Returns:     Nothing
1978  *
1979  * Lock status: No locks are assumed held.
1980  *
1981  * Notes:       There is no timer nor any other means by which the requests
1982  *              get unblocked other than the low-level driver calling
1983  *              scsi_unblock_requests().
1984  *
1985  *              This is done as an API function so that changes to the
1986  *              internals of the scsi mid-layer won't require wholesale
1987  *              changes to drivers that use this feature.
1988  */
1989 void scsi_unblock_requests(struct Scsi_Host *shost)
1990 {
1991         shost->host_self_blocked = 0;
1992         scsi_run_host_queues(shost);
1993 }
1994 EXPORT_SYMBOL(scsi_unblock_requests);
1995
1996 int __init scsi_init_queue(void)
1997 {
1998         scsi_sdb_cache = kmem_cache_create("scsi_data_buffer",
1999                                            sizeof(struct scsi_data_buffer),
2000                                            0, 0, NULL);
2001         if (!scsi_sdb_cache) {
2002                 printk(KERN_ERR "SCSI: can't init scsi sdb cache\n");
2003                 return -ENOMEM;
2004         }
2005
2006         return 0;
2007 }
2008
2009 void scsi_exit_queue(void)
2010 {
2011         kmem_cache_destroy(scsi_sense_cache);
2012         kmem_cache_destroy(scsi_sense_isadma_cache);
2013         kmem_cache_destroy(scsi_sdb_cache);
2014 }
2015
2016 /**
2017  *      scsi_mode_select - issue a mode select
2018  *      @sdev:  SCSI device to be queried
2019  *      @pf:    Page format bit (1 == standard, 0 == vendor specific)
2020  *      @sp:    Save page bit (0 == don't save, 1 == save)
2021  *      @modepage: mode page being requested
2022  *      @buffer: request buffer (may not be smaller than eight bytes)
2023  *      @len:   length of request buffer.
2024  *      @timeout: command timeout
2025  *      @retries: number of retries before failing
2026  *      @data: returns a structure abstracting the mode header data
2027  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
2028  *              must be SCSI_SENSE_BUFFERSIZE big.
2029  *
2030  *      Returns zero if successful; negative error number or scsi
2031  *      status on error
2032  *
2033  */
2034 int
2035 scsi_mode_select(struct scsi_device *sdev, int pf, int sp, int modepage,
2036                  unsigned char *buffer, int len, int timeout, int retries,
2037                  struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
2038 {
2039         unsigned char cmd[10];
2040         unsigned char *real_buffer;
2041         int ret;
2042
2043         memset(cmd, 0, sizeof(cmd));
2044         cmd[1] = (pf ? 0x10 : 0) | (sp ? 0x01 : 0);
2045
2046         if (sdev->use_10_for_ms) {
2047                 if (len > 65535)
2048                         return -EINVAL;
2049                 real_buffer = kmalloc(8 + len, GFP_KERNEL);
2050                 if (!real_buffer)
2051                         return -ENOMEM;
2052                 memcpy(real_buffer + 8, buffer, len);
2053                 len += 8;
2054                 real_buffer[0] = 0;
2055                 real_buffer[1] = 0;
2056                 real_buffer[2] = data->medium_type;
2057                 real_buffer[3] = data->device_specific;
2058                 real_buffer[4] = data->longlba ? 0x01 : 0;
2059                 real_buffer[5] = 0;
2060                 real_buffer[6] = data->block_descriptor_length >> 8;
2061                 real_buffer[7] = data->block_descriptor_length;
2062
2063                 cmd[0] = MODE_SELECT_10;
2064                 cmd[7] = len >> 8;
2065                 cmd[8] = len;
2066         } else {
2067                 if (len > 255 || data->block_descriptor_length > 255 ||
2068                     data->longlba)
2069                         return -EINVAL;
2070
2071                 real_buffer = kmalloc(4 + len, GFP_KERNEL);
2072                 if (!real_buffer)
2073                         return -ENOMEM;
2074                 memcpy(real_buffer + 4, buffer, len);
2075                 len += 4;
2076                 real_buffer[0] = 0;
2077                 real_buffer[1] = data->medium_type;
2078                 real_buffer[2] = data->device_specific;
2079                 real_buffer[3] = data->block_descriptor_length;
2080                 
2081
2082                 cmd[0] = MODE_SELECT;
2083                 cmd[4] = len;
2084         }
2085
2086         ret = scsi_execute_req(sdev, cmd, DMA_TO_DEVICE, real_buffer, len,
2087                                sshdr, timeout, retries, NULL);
2088         kfree(real_buffer);
2089         return ret;
2090 }
2091 EXPORT_SYMBOL_GPL(scsi_mode_select);
2092
2093 /**
2094  *      scsi_mode_sense - issue a mode sense, falling back from 10 to six bytes if necessary.
2095  *      @sdev:  SCSI device to be queried
2096  *      @dbd:   set if mode sense will allow block descriptors to be returned
2097  *      @modepage: mode page being requested
2098  *      @buffer: request buffer (may not be smaller than eight bytes)
2099  *      @len:   length of request buffer.
2100  *      @timeout: command timeout
2101  *      @retries: number of retries before failing
2102  *      @data: returns a structure abstracting the mode header data
2103  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
2104  *              must be SCSI_SENSE_BUFFERSIZE big.
2105  *
2106  *      Returns zero if unsuccessful, or the header offset (either 4
2107  *      or 8 depending on whether a six or ten byte command was
2108  *      issued) if successful.
2109  */
2110 int
2111 scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
2112                   unsigned char *buffer, int len, int timeout, int retries,
2113                   struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
2114 {
2115         unsigned char cmd[12];
2116         int use_10_for_ms;
2117         int header_length;
2118         int result, retry_count = retries;
2119         struct scsi_sense_hdr my_sshdr;
2120
2121         memset(data, 0, sizeof(*data));
2122         memset(&cmd[0], 0, 12);
2123
2124         dbd = sdev->set_dbd_for_ms ? 8 : dbd;
2125         cmd[1] = dbd & 0x18;    /* allows DBD and LLBA bits */
2126         cmd[2] = modepage;
2127
2128         /* caller might not be interested in sense, but we need it */
2129         if (!sshdr)
2130                 sshdr = &my_sshdr;
2131
2132  retry:
2133         use_10_for_ms = sdev->use_10_for_ms;
2134
2135         if (use_10_for_ms) {
2136                 if (len < 8)
2137                         len = 8;
2138
2139                 cmd[0] = MODE_SENSE_10;
2140                 cmd[8] = len;
2141                 header_length = 8;
2142         } else {
2143                 if (len < 4)
2144                         len = 4;
2145
2146                 cmd[0] = MODE_SENSE;
2147                 cmd[4] = len;
2148                 header_length = 4;
2149         }
2150
2151         memset(buffer, 0, len);
2152
2153         result = scsi_execute_req(sdev, cmd, DMA_FROM_DEVICE, buffer, len,
2154                                   sshdr, timeout, retries, NULL);
2155
2156         /* This code looks awful: what it's doing is making sure an
2157          * ILLEGAL REQUEST sense return identifies the actual command
2158          * byte as the problem.  MODE_SENSE commands can return
2159          * ILLEGAL REQUEST if the code page isn't supported */
2160
2161         if (use_10_for_ms && !scsi_status_is_good(result) &&
2162             driver_byte(result) == DRIVER_SENSE) {
2163                 if (scsi_sense_valid(sshdr)) {
2164                         if ((sshdr->sense_key == ILLEGAL_REQUEST) &&
2165                             (sshdr->asc == 0x20) && (sshdr->ascq == 0)) {
2166                                 /* 
2167                                  * Invalid command operation code
2168                                  */
2169                                 sdev->use_10_for_ms = 0;
2170                                 goto retry;
2171                         }
2172                 }
2173         }
2174
2175         if(scsi_status_is_good(result)) {
2176                 if (unlikely(buffer[0] == 0x86 && buffer[1] == 0x0b &&
2177                              (modepage == 6 || modepage == 8))) {
2178                         /* Initio breakage? */
2179                         header_length = 0;
2180                         data->length = 13;
2181                         data->medium_type = 0;
2182                         data->device_specific = 0;
2183                         data->longlba = 0;
2184                         data->block_descriptor_length = 0;
2185                 } else if(use_10_for_ms) {
2186                         data->length = buffer[0]*256 + buffer[1] + 2;
2187                         data->medium_type = buffer[2];
2188                         data->device_specific = buffer[3];
2189                         data->longlba = buffer[4] & 0x01;
2190                         data->block_descriptor_length = buffer[6]*256
2191                                 + buffer[7];
2192                 } else {
2193                         data->length = buffer[0] + 1;
2194                         data->medium_type = buffer[1];
2195                         data->device_specific = buffer[2];
2196                         data->block_descriptor_length = buffer[3];
2197                 }
2198                 data->header_length = header_length;
2199         } else if ((status_byte(result) == CHECK_CONDITION) &&
2200                    scsi_sense_valid(sshdr) &&
2201                    sshdr->sense_key == UNIT_ATTENTION && retry_count) {
2202                 retry_count--;
2203                 goto retry;
2204         }
2205
2206         return result;
2207 }
2208 EXPORT_SYMBOL(scsi_mode_sense);
2209
2210 /**
2211  *      scsi_test_unit_ready - test if unit is ready
2212  *      @sdev:  scsi device to change the state of.
2213  *      @timeout: command timeout
2214  *      @retries: number of retries before failing
2215  *      @sshdr: outpout pointer for decoded sense information.
2216  *
2217  *      Returns zero if unsuccessful or an error if TUR failed.  For
2218  *      removable media, UNIT_ATTENTION sets ->changed flag.
2219  **/
2220 int
2221 scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries,
2222                      struct scsi_sense_hdr *sshdr)
2223 {
2224         char cmd[] = {
2225                 TEST_UNIT_READY, 0, 0, 0, 0, 0,
2226         };
2227         int result;
2228
2229         /* try to eat the UNIT_ATTENTION if there are enough retries */
2230         do {
2231                 result = scsi_execute_req(sdev, cmd, DMA_NONE, NULL, 0, sshdr,
2232                                           timeout, 1, NULL);
2233                 if (sdev->removable && scsi_sense_valid(sshdr) &&
2234                     sshdr->sense_key == UNIT_ATTENTION)
2235                         sdev->changed = 1;
2236         } while (scsi_sense_valid(sshdr) &&
2237                  sshdr->sense_key == UNIT_ATTENTION && --retries);
2238
2239         return result;
2240 }
2241 EXPORT_SYMBOL(scsi_test_unit_ready);
2242
2243 /**
2244  *      scsi_device_set_state - Take the given device through the device state model.
2245  *      @sdev:  scsi device to change the state of.
2246  *      @state: state to change to.
2247  *
2248  *      Returns zero if successful or an error if the requested
2249  *      transition is illegal.
2250  */
2251 int
2252 scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
2253 {
2254         enum scsi_device_state oldstate = sdev->sdev_state;
2255
2256         if (state == oldstate)
2257                 return 0;
2258
2259         switch (state) {
2260         case SDEV_CREATED:
2261                 switch (oldstate) {
2262                 case SDEV_CREATED_BLOCK:
2263                         break;
2264                 default:
2265                         goto illegal;
2266                 }
2267                 break;
2268                         
2269         case SDEV_RUNNING:
2270                 switch (oldstate) {
2271                 case SDEV_CREATED:
2272                 case SDEV_OFFLINE:
2273                 case SDEV_TRANSPORT_OFFLINE:
2274                 case SDEV_QUIESCE:
2275                 case SDEV_BLOCK:
2276                         break;
2277                 default:
2278                         goto illegal;
2279                 }
2280                 break;
2281
2282         case SDEV_QUIESCE:
2283                 switch (oldstate) {
2284                 case SDEV_RUNNING:
2285                 case SDEV_OFFLINE:
2286                 case SDEV_TRANSPORT_OFFLINE:
2287                         break;
2288                 default:
2289                         goto illegal;
2290                 }
2291                 break;
2292
2293         case SDEV_OFFLINE:
2294         case SDEV_TRANSPORT_OFFLINE:
2295                 switch (oldstate) {
2296                 case SDEV_CREATED:
2297                 case SDEV_RUNNING:
2298                 case SDEV_QUIESCE:
2299                 case SDEV_BLOCK:
2300                         break;
2301                 default:
2302                         goto illegal;
2303                 }
2304                 break;
2305
2306         case SDEV_BLOCK:
2307                 switch (oldstate) {
2308                 case SDEV_RUNNING:
2309                 case SDEV_CREATED_BLOCK:
2310                 case SDEV_QUIESCE:
2311                 case SDEV_OFFLINE:
2312                         break;
2313                 default:
2314                         goto illegal;
2315                 }
2316                 break;
2317
2318         case SDEV_CREATED_BLOCK:
2319                 switch (oldstate) {
2320                 case SDEV_CREATED:
2321                         break;
2322                 default:
2323                         goto illegal;
2324                 }
2325                 break;
2326
2327         case SDEV_CANCEL:
2328                 switch (oldstate) {
2329                 case SDEV_CREATED:
2330                 case SDEV_RUNNING:
2331                 case SDEV_QUIESCE:
2332                 case SDEV_OFFLINE:
2333                 case SDEV_TRANSPORT_OFFLINE:
2334                         break;
2335                 default:
2336                         goto illegal;
2337                 }
2338                 break;
2339
2340         case SDEV_DEL:
2341                 switch (oldstate) {
2342                 case SDEV_CREATED:
2343                 case SDEV_RUNNING:
2344                 case SDEV_OFFLINE:
2345                 case SDEV_TRANSPORT_OFFLINE:
2346                 case SDEV_CANCEL:
2347                 case SDEV_BLOCK:
2348                 case SDEV_CREATED_BLOCK:
2349                         break;
2350                 default:
2351                         goto illegal;
2352                 }
2353                 break;
2354
2355         }
2356         sdev->offline_already = false;
2357         sdev->sdev_state = state;
2358         return 0;
2359
2360  illegal:
2361         SCSI_LOG_ERROR_RECOVERY(1,
2362                                 sdev_printk(KERN_ERR, sdev,
2363                                             "Illegal state transition %s->%s",
2364                                             scsi_device_state_name(oldstate),
2365                                             scsi_device_state_name(state))
2366                                 );
2367         return -EINVAL;
2368 }
2369 EXPORT_SYMBOL(scsi_device_set_state);
2370
2371 /**
2372  *      sdev_evt_emit - emit a single SCSI device uevent
2373  *      @sdev: associated SCSI device
2374  *      @evt: event to emit
2375  *
2376  *      Send a single uevent (scsi_event) to the associated scsi_device.
2377  */
2378 static void scsi_evt_emit(struct scsi_device *sdev, struct scsi_event *evt)
2379 {
2380         int idx = 0;
2381         char *envp[3];
2382
2383         switch (evt->evt_type) {
2384         case SDEV_EVT_MEDIA_CHANGE:
2385                 envp[idx++] = "SDEV_MEDIA_CHANGE=1";
2386                 break;
2387         case SDEV_EVT_INQUIRY_CHANGE_REPORTED:
2388                 scsi_rescan_device(&sdev->sdev_gendev);
2389                 envp[idx++] = "SDEV_UA=INQUIRY_DATA_HAS_CHANGED";
2390                 break;
2391         case SDEV_EVT_CAPACITY_CHANGE_REPORTED:
2392                 envp[idx++] = "SDEV_UA=CAPACITY_DATA_HAS_CHANGED";
2393                 break;
2394         case SDEV_EVT_SOFT_THRESHOLD_REACHED_REPORTED:
2395                envp[idx++] = "SDEV_UA=THIN_PROVISIONING_SOFT_THRESHOLD_REACHED";
2396                 break;
2397         case SDEV_EVT_MODE_PARAMETER_CHANGE_REPORTED:
2398                 envp[idx++] = "SDEV_UA=MODE_PARAMETERS_CHANGED";
2399                 break;
2400         case SDEV_EVT_LUN_CHANGE_REPORTED:
2401                 envp[idx++] = "SDEV_UA=REPORTED_LUNS_DATA_HAS_CHANGED";
2402                 break;
2403         case SDEV_EVT_ALUA_STATE_CHANGE_REPORTED:
2404                 envp[idx++] = "SDEV_UA=ASYMMETRIC_ACCESS_STATE_CHANGED";
2405                 break;
2406         case SDEV_EVT_POWER_ON_RESET_OCCURRED:
2407                 envp[idx++] = "SDEV_UA=POWER_ON_RESET_OCCURRED";
2408                 break;
2409         default:
2410                 /* do nothing */
2411                 break;
2412         }
2413
2414         envp[idx++] = NULL;
2415
2416         kobject_uevent_env(&sdev->sdev_gendev.kobj, KOBJ_CHANGE, envp);
2417 }
2418
2419 /**
2420  *      sdev_evt_thread - send a uevent for each scsi event
2421  *      @work: work struct for scsi_device
2422  *
2423  *      Dispatch queued events to their associated scsi_device kobjects
2424  *      as uevents.
2425  */
2426 void scsi_evt_thread(struct work_struct *work)
2427 {
2428         struct scsi_device *sdev;
2429         enum scsi_device_event evt_type;
2430         LIST_HEAD(event_list);
2431
2432         sdev = container_of(work, struct scsi_device, event_work);
2433
2434         for (evt_type = SDEV_EVT_FIRST; evt_type <= SDEV_EVT_LAST; evt_type++)
2435                 if (test_and_clear_bit(evt_type, sdev->pending_events))
2436                         sdev_evt_send_simple(sdev, evt_type, GFP_KERNEL);
2437
2438         while (1) {
2439                 struct scsi_event *evt;
2440                 struct list_head *this, *tmp;
2441                 unsigned long flags;
2442
2443                 spin_lock_irqsave(&sdev->list_lock, flags);
2444                 list_splice_init(&sdev->event_list, &event_list);
2445                 spin_unlock_irqrestore(&sdev->list_lock, flags);
2446
2447                 if (list_empty(&event_list))
2448                         break;
2449
2450                 list_for_each_safe(this, tmp, &event_list) {
2451                         evt = list_entry(this, struct scsi_event, node);
2452                         list_del(&evt->node);
2453                         scsi_evt_emit(sdev, evt);
2454                         kfree(evt);
2455                 }
2456         }
2457 }
2458
2459 /**
2460  *      sdev_evt_send - send asserted event to uevent thread
2461  *      @sdev: scsi_device event occurred on
2462  *      @evt: event to send
2463  *
2464  *      Assert scsi device event asynchronously.
2465  */
2466 void sdev_evt_send(struct scsi_device *sdev, struct scsi_event *evt)
2467 {
2468         unsigned long flags;
2469
2470 #if 0
2471         /* FIXME: currently this check eliminates all media change events
2472          * for polled devices.  Need to update to discriminate between AN
2473          * and polled events */
2474         if (!test_bit(evt->evt_type, sdev->supported_events)) {
2475                 kfree(evt);
2476                 return;
2477         }
2478 #endif
2479
2480         spin_lock_irqsave(&sdev->list_lock, flags);
2481         list_add_tail(&evt->node, &sdev->event_list);
2482         schedule_work(&sdev->event_work);
2483         spin_unlock_irqrestore(&sdev->list_lock, flags);
2484 }
2485 EXPORT_SYMBOL_GPL(sdev_evt_send);
2486
2487 /**
2488  *      sdev_evt_alloc - allocate a new scsi event
2489  *      @evt_type: type of event to allocate
2490  *      @gfpflags: GFP flags for allocation
2491  *
2492  *      Allocates and returns a new scsi_event.
2493  */
2494 struct scsi_event *sdev_evt_alloc(enum scsi_device_event evt_type,
2495                                   gfp_t gfpflags)
2496 {
2497         struct scsi_event *evt = kzalloc(sizeof(struct scsi_event), gfpflags);
2498         if (!evt)
2499                 return NULL;
2500
2501         evt->evt_type = evt_type;
2502         INIT_LIST_HEAD(&evt->node);
2503
2504         /* evt_type-specific initialization, if any */
2505         switch (evt_type) {
2506         case SDEV_EVT_MEDIA_CHANGE:
2507         case SDEV_EVT_INQUIRY_CHANGE_REPORTED:
2508         case SDEV_EVT_CAPACITY_CHANGE_REPORTED:
2509         case SDEV_EVT_SOFT_THRESHOLD_REACHED_REPORTED:
2510         case SDEV_EVT_MODE_PARAMETER_CHANGE_REPORTED:
2511         case SDEV_EVT_LUN_CHANGE_REPORTED:
2512         case SDEV_EVT_ALUA_STATE_CHANGE_REPORTED:
2513         case SDEV_EVT_POWER_ON_RESET_OCCURRED:
2514         default:
2515                 /* do nothing */
2516                 break;
2517         }
2518
2519         return evt;
2520 }
2521 EXPORT_SYMBOL_GPL(sdev_evt_alloc);
2522
2523 /**
2524  *      sdev_evt_send_simple - send asserted event to uevent thread
2525  *      @sdev: scsi_device event occurred on
2526  *      @evt_type: type of event to send
2527  *      @gfpflags: GFP flags for allocation
2528  *
2529  *      Assert scsi device event asynchronously, given an event type.
2530  */
2531 void sdev_evt_send_simple(struct scsi_device *sdev,
2532                           enum scsi_device_event evt_type, gfp_t gfpflags)
2533 {
2534         struct scsi_event *evt = sdev_evt_alloc(evt_type, gfpflags);
2535         if (!evt) {
2536                 sdev_printk(KERN_ERR, sdev, "event %d eaten due to OOM\n",
2537                             evt_type);
2538                 return;
2539         }
2540
2541         sdev_evt_send(sdev, evt);
2542 }
2543 EXPORT_SYMBOL_GPL(sdev_evt_send_simple);
2544
2545 /**
2546  *      scsi_device_quiesce - Block user issued commands.
2547  *      @sdev:  scsi device to quiesce.
2548  *
2549  *      This works by trying to transition to the SDEV_QUIESCE state
2550  *      (which must be a legal transition).  When the device is in this
2551  *      state, only special requests will be accepted, all others will
2552  *      be deferred.  Since special requests may also be requeued requests,
2553  *      a successful return doesn't guarantee the device will be 
2554  *      totally quiescent.
2555  *
2556  *      Must be called with user context, may sleep.
2557  *
2558  *      Returns zero if unsuccessful or an error if not.
2559  */
2560 int
2561 scsi_device_quiesce(struct scsi_device *sdev)
2562 {
2563         struct request_queue *q = sdev->request_queue;
2564         int err;
2565
2566         /*
2567          * It is allowed to call scsi_device_quiesce() multiple times from
2568          * the same context but concurrent scsi_device_quiesce() calls are
2569          * not allowed.
2570          */
2571         WARN_ON_ONCE(sdev->quiesced_by && sdev->quiesced_by != current);
2572
2573         if (sdev->quiesced_by == current)
2574                 return 0;
2575
2576         blk_set_pm_only(q);
2577
2578         blk_mq_freeze_queue(q);
2579         /*
2580          * Ensure that the effect of blk_set_pm_only() will be visible
2581          * for percpu_ref_tryget() callers that occur after the queue
2582          * unfreeze even if the queue was already frozen before this function
2583          * was called. See also https://lwn.net/Articles/573497/.
2584          */
2585         synchronize_rcu();
2586         blk_mq_unfreeze_queue(q);
2587
2588         mutex_lock(&sdev->state_mutex);
2589         err = scsi_device_set_state(sdev, SDEV_QUIESCE);
2590         if (err == 0)
2591                 sdev->quiesced_by = current;
2592         else
2593                 blk_clear_pm_only(q);
2594         mutex_unlock(&sdev->state_mutex);
2595
2596         return err;
2597 }
2598 EXPORT_SYMBOL(scsi_device_quiesce);
2599
2600 /**
2601  *      scsi_device_resume - Restart user issued commands to a quiesced device.
2602  *      @sdev:  scsi device to resume.
2603  *
2604  *      Moves the device from quiesced back to running and restarts the
2605  *      queues.
2606  *
2607  *      Must be called with user context, may sleep.
2608  */
2609 void scsi_device_resume(struct scsi_device *sdev)
2610 {
2611         /* check if the device state was mutated prior to resume, and if
2612          * so assume the state is being managed elsewhere (for example
2613          * device deleted during suspend)
2614          */
2615         mutex_lock(&sdev->state_mutex);
2616         if (sdev->quiesced_by) {
2617                 sdev->quiesced_by = NULL;
2618                 blk_clear_pm_only(sdev->request_queue);
2619         }
2620         if (sdev->sdev_state == SDEV_QUIESCE)
2621                 scsi_device_set_state(sdev, SDEV_RUNNING);
2622         mutex_unlock(&sdev->state_mutex);
2623 }
2624 EXPORT_SYMBOL(scsi_device_resume);
2625
2626 static void
2627 device_quiesce_fn(struct scsi_device *sdev, void *data)
2628 {
2629         scsi_device_quiesce(sdev);
2630 }
2631
2632 void
2633 scsi_target_quiesce(struct scsi_target *starget)
2634 {
2635         starget_for_each_device(starget, NULL, device_quiesce_fn);
2636 }
2637 EXPORT_SYMBOL(scsi_target_quiesce);
2638
2639 static void
2640 device_resume_fn(struct scsi_device *sdev, void *data)
2641 {
2642         scsi_device_resume(sdev);
2643 }
2644
2645 void
2646 scsi_target_resume(struct scsi_target *starget)
2647 {
2648         starget_for_each_device(starget, NULL, device_resume_fn);
2649 }
2650 EXPORT_SYMBOL(scsi_target_resume);
2651
2652 /**
2653  * scsi_internal_device_block_nowait - try to transition to the SDEV_BLOCK state
2654  * @sdev: device to block
2655  *
2656  * Pause SCSI command processing on the specified device. Does not sleep.
2657  *
2658  * Returns zero if successful or a negative error code upon failure.
2659  *
2660  * Notes:
2661  * This routine transitions the device to the SDEV_BLOCK state (which must be
2662  * a legal transition). When the device is in this state, command processing
2663  * is paused until the device leaves the SDEV_BLOCK state. See also
2664  * scsi_internal_device_unblock_nowait().
2665  */
2666 int scsi_internal_device_block_nowait(struct scsi_device *sdev)
2667 {
2668         struct request_queue *q = sdev->request_queue;
2669         int err = 0;
2670
2671         err = scsi_device_set_state(sdev, SDEV_BLOCK);
2672         if (err) {
2673                 err = scsi_device_set_state(sdev, SDEV_CREATED_BLOCK);
2674
2675                 if (err)
2676                         return err;
2677         }
2678
2679         /* 
2680          * The device has transitioned to SDEV_BLOCK.  Stop the
2681          * block layer from calling the midlayer with this device's
2682          * request queue. 
2683          */
2684         blk_mq_quiesce_queue_nowait(q);
2685         return 0;
2686 }
2687 EXPORT_SYMBOL_GPL(scsi_internal_device_block_nowait);
2688
2689 /**
2690  * scsi_internal_device_block - try to transition to the SDEV_BLOCK state
2691  * @sdev: device to block
2692  *
2693  * Pause SCSI command processing on the specified device and wait until all
2694  * ongoing scsi_request_fn() / scsi_queue_rq() calls have finished. May sleep.
2695  *
2696  * Returns zero if successful or a negative error code upon failure.
2697  *
2698  * Note:
2699  * This routine transitions the device to the SDEV_BLOCK state (which must be
2700  * a legal transition). When the device is in this state, command processing
2701  * is paused until the device leaves the SDEV_BLOCK state. See also
2702  * scsi_internal_device_unblock().
2703  */
2704 static int scsi_internal_device_block(struct scsi_device *sdev)
2705 {
2706         struct request_queue *q = sdev->request_queue;
2707         int err;
2708
2709         mutex_lock(&sdev->state_mutex);
2710         err = scsi_internal_device_block_nowait(sdev);
2711         if (err == 0)
2712                 blk_mq_quiesce_queue(q);
2713         mutex_unlock(&sdev->state_mutex);
2714
2715         return err;
2716 }
2717  
2718 void scsi_start_queue(struct scsi_device *sdev)
2719 {
2720         struct request_queue *q = sdev->request_queue;
2721
2722         blk_mq_unquiesce_queue(q);
2723 }
2724
2725 /**
2726  * scsi_internal_device_unblock_nowait - resume a device after a block request
2727  * @sdev:       device to resume
2728  * @new_state:  state to set the device to after unblocking
2729  *
2730  * Restart the device queue for a previously suspended SCSI device. Does not
2731  * sleep.
2732  *
2733  * Returns zero if successful or a negative error code upon failure.
2734  *
2735  * Notes:
2736  * This routine transitions the device to the SDEV_RUNNING state or to one of
2737  * the offline states (which must be a legal transition) allowing the midlayer
2738  * to goose the queue for this device.
2739  */
2740 int scsi_internal_device_unblock_nowait(struct scsi_device *sdev,
2741                                         enum scsi_device_state new_state)
2742 {
2743         switch (new_state) {
2744         case SDEV_RUNNING:
2745         case SDEV_TRANSPORT_OFFLINE:
2746                 break;
2747         default:
2748                 return -EINVAL;
2749         }
2750
2751         /*
2752          * Try to transition the scsi device to SDEV_RUNNING or one of the
2753          * offlined states and goose the device queue if successful.
2754          */
2755         switch (sdev->sdev_state) {
2756         case SDEV_BLOCK:
2757         case SDEV_TRANSPORT_OFFLINE:
2758                 sdev->sdev_state = new_state;
2759                 break;
2760         case SDEV_CREATED_BLOCK:
2761                 if (new_state == SDEV_TRANSPORT_OFFLINE ||
2762                     new_state == SDEV_OFFLINE)
2763                         sdev->sdev_state = new_state;
2764                 else
2765                         sdev->sdev_state = SDEV_CREATED;
2766                 break;
2767         case SDEV_CANCEL:
2768         case SDEV_OFFLINE:
2769                 break;
2770         default:
2771                 return -EINVAL;
2772         }
2773         scsi_start_queue(sdev);
2774
2775         return 0;
2776 }
2777 EXPORT_SYMBOL_GPL(scsi_internal_device_unblock_nowait);
2778
2779 /**
2780  * scsi_internal_device_unblock - resume a device after a block request
2781  * @sdev:       device to resume
2782  * @new_state:  state to set the device to after unblocking
2783  *
2784  * Restart the device queue for a previously suspended SCSI device. May sleep.
2785  *
2786  * Returns zero if successful or a negative error code upon failure.
2787  *
2788  * Notes:
2789  * This routine transitions the device to the SDEV_RUNNING state or to one of
2790  * the offline states (which must be a legal transition) allowing the midlayer
2791  * to goose the queue for this device.
2792  */
2793 static int scsi_internal_device_unblock(struct scsi_device *sdev,
2794                                         enum scsi_device_state new_state)
2795 {
2796         int ret;
2797
2798         mutex_lock(&sdev->state_mutex);
2799         ret = scsi_internal_device_unblock_nowait(sdev, new_state);
2800         mutex_unlock(&sdev->state_mutex);
2801
2802         return ret;
2803 }
2804
2805 static void
2806 device_block(struct scsi_device *sdev, void *data)
2807 {
2808         int ret;
2809
2810         ret = scsi_internal_device_block(sdev);
2811
2812         WARN_ONCE(ret, "scsi_internal_device_block(%s) failed: ret = %d\n",
2813                   dev_name(&sdev->sdev_gendev), ret);
2814 }
2815
2816 static int
2817 target_block(struct device *dev, void *data)
2818 {
2819         if (scsi_is_target_device(dev))
2820                 starget_for_each_device(to_scsi_target(dev), NULL,
2821                                         device_block);
2822         return 0;
2823 }
2824
2825 void
2826 scsi_target_block(struct device *dev)
2827 {
2828         if (scsi_is_target_device(dev))
2829                 starget_for_each_device(to_scsi_target(dev), NULL,
2830                                         device_block);
2831         else
2832                 device_for_each_child(dev, NULL, target_block);
2833 }
2834 EXPORT_SYMBOL_GPL(scsi_target_block);
2835
2836 static void
2837 device_unblock(struct scsi_device *sdev, void *data)
2838 {
2839         scsi_internal_device_unblock(sdev, *(enum scsi_device_state *)data);
2840 }
2841
2842 static int
2843 target_unblock(struct device *dev, void *data)
2844 {
2845         if (scsi_is_target_device(dev))
2846                 starget_for_each_device(to_scsi_target(dev), data,
2847                                         device_unblock);
2848         return 0;
2849 }
2850
2851 void
2852 scsi_target_unblock(struct device *dev, enum scsi_device_state new_state)
2853 {
2854         if (scsi_is_target_device(dev))
2855                 starget_for_each_device(to_scsi_target(dev), &new_state,
2856                                         device_unblock);
2857         else
2858                 device_for_each_child(dev, &new_state, target_unblock);
2859 }
2860 EXPORT_SYMBOL_GPL(scsi_target_unblock);
2861
2862 int
2863 scsi_host_block(struct Scsi_Host *shost)
2864 {
2865         struct scsi_device *sdev;
2866         int ret = 0;
2867
2868         shost_for_each_device(sdev, shost) {
2869                 ret = scsi_internal_device_block(sdev);
2870                 if (ret)
2871                         break;
2872         }
2873         return ret;
2874 }
2875 EXPORT_SYMBOL_GPL(scsi_host_block);
2876
2877 int
2878 scsi_host_unblock(struct Scsi_Host *shost, int new_state)
2879 {
2880         struct scsi_device *sdev;
2881         int ret = 0;
2882
2883         shost_for_each_device(sdev, shost) {
2884                 ret = scsi_internal_device_unblock(sdev, new_state);
2885                 if (ret)
2886                         break;
2887         }
2888         return ret;
2889 }
2890 EXPORT_SYMBOL_GPL(scsi_host_unblock);
2891
2892 /**
2893  * scsi_kmap_atomic_sg - find and atomically map an sg-elemnt
2894  * @sgl:        scatter-gather list
2895  * @sg_count:   number of segments in sg
2896  * @offset:     offset in bytes into sg, on return offset into the mapped area
2897  * @len:        bytes to map, on return number of bytes mapped
2898  *
2899  * Returns virtual address of the start of the mapped page
2900  */
2901 void *scsi_kmap_atomic_sg(struct scatterlist *sgl, int sg_count,
2902                           size_t *offset, size_t *len)
2903 {
2904         int i;
2905         size_t sg_len = 0, len_complete = 0;
2906         struct scatterlist *sg;
2907         struct page *page;
2908
2909         WARN_ON(!irqs_disabled());
2910
2911         for_each_sg(sgl, sg, sg_count, i) {
2912                 len_complete = sg_len; /* Complete sg-entries */
2913                 sg_len += sg->length;
2914                 if (sg_len > *offset)
2915                         break;
2916         }
2917
2918         if (unlikely(i == sg_count)) {
2919                 printk(KERN_ERR "%s: Bytes in sg: %zu, requested offset %zu, "
2920                         "elements %d\n",
2921                        __func__, sg_len, *offset, sg_count);
2922                 WARN_ON(1);
2923                 return NULL;
2924         }
2925
2926         /* Offset starting from the beginning of first page in this sg-entry */
2927         *offset = *offset - len_complete + sg->offset;
2928
2929         /* Assumption: contiguous pages can be accessed as "page + i" */
2930         page = nth_page(sg_page(sg), (*offset >> PAGE_SHIFT));
2931         *offset &= ~PAGE_MASK;
2932
2933         /* Bytes in this sg-entry from *offset to the end of the page */
2934         sg_len = PAGE_SIZE - *offset;
2935         if (*len > sg_len)
2936                 *len = sg_len;
2937
2938         return kmap_atomic(page);
2939 }
2940 EXPORT_SYMBOL(scsi_kmap_atomic_sg);
2941
2942 /**
2943  * scsi_kunmap_atomic_sg - atomically unmap a virtual address, previously mapped with scsi_kmap_atomic_sg
2944  * @virt:       virtual address to be unmapped
2945  */
2946 void scsi_kunmap_atomic_sg(void *virt)
2947 {
2948         kunmap_atomic(virt);
2949 }
2950 EXPORT_SYMBOL(scsi_kunmap_atomic_sg);
2951
2952 void sdev_disable_disk_events(struct scsi_device *sdev)
2953 {
2954         atomic_inc(&sdev->disk_events_disable_depth);
2955 }
2956 EXPORT_SYMBOL(sdev_disable_disk_events);
2957
2958 void sdev_enable_disk_events(struct scsi_device *sdev)
2959 {
2960         if (WARN_ON_ONCE(atomic_read(&sdev->disk_events_disable_depth) <= 0))
2961                 return;
2962         atomic_dec(&sdev->disk_events_disable_depth);
2963 }
2964 EXPORT_SYMBOL(sdev_enable_disk_events);
2965
2966 /**
2967  * scsi_vpd_lun_id - return a unique device identification
2968  * @sdev: SCSI device
2969  * @id:   buffer for the identification
2970  * @id_len:  length of the buffer
2971  *
2972  * Copies a unique device identification into @id based
2973  * on the information in the VPD page 0x83 of the device.
2974  * The string will be formatted as a SCSI name string.
2975  *
2976  * Returns the length of the identification or error on failure.
2977  * If the identifier is longer than the supplied buffer the actual
2978  * identifier length is returned and the buffer is not zero-padded.
2979  */
2980 int scsi_vpd_lun_id(struct scsi_device *sdev, char *id, size_t id_len)
2981 {
2982         u8 cur_id_type = 0xff;
2983         u8 cur_id_size = 0;
2984         const unsigned char *d, *cur_id_str;
2985         const struct scsi_vpd *vpd_pg83;
2986         int id_size = -EINVAL;
2987
2988         rcu_read_lock();
2989         vpd_pg83 = rcu_dereference(sdev->vpd_pg83);
2990         if (!vpd_pg83) {
2991                 rcu_read_unlock();
2992                 return -ENXIO;
2993         }
2994
2995         /*
2996          * Look for the correct descriptor.
2997          * Order of preference for lun descriptor:
2998          * - SCSI name string
2999          * - NAA IEEE Registered Extended
3000          * - EUI-64 based 16-byte
3001          * - EUI-64 based 12-byte
3002          * - NAA IEEE Registered
3003          * - NAA IEEE Extended
3004          * - T10 Vendor ID
3005          * as longer descriptors reduce the likelyhood
3006          * of identification clashes.
3007          */
3008
3009         /* The id string must be at least 20 bytes + terminating NULL byte */
3010         if (id_len < 21) {
3011                 rcu_read_unlock();
3012                 return -EINVAL;
3013         }
3014
3015         memset(id, 0, id_len);
3016         d = vpd_pg83->data + 4;
3017         while (d < vpd_pg83->data + vpd_pg83->len) {
3018                 /* Skip designators not referring to the LUN */
3019                 if ((d[1] & 0x30) != 0x00)
3020                         goto next_desig;
3021
3022                 switch (d[1] & 0xf) {
3023                 case 0x1:
3024                         /* T10 Vendor ID */
3025                         if (cur_id_size > d[3])
3026                                 break;
3027                         /* Prefer anything */
3028                         if (cur_id_type > 0x01 && cur_id_type != 0xff)
3029                                 break;
3030                         cur_id_size = d[3];
3031                         if (cur_id_size + 4 > id_len)
3032                                 cur_id_size = id_len - 4;
3033                         cur_id_str = d + 4;
3034                         cur_id_type = d[1] & 0xf;
3035                         id_size = snprintf(id, id_len, "t10.%*pE",
3036                                            cur_id_size, cur_id_str);
3037                         break;
3038                 case 0x2:
3039                         /* EUI-64 */
3040                         if (cur_id_size > d[3])
3041                                 break;
3042                         /* Prefer NAA IEEE Registered Extended */
3043                         if (cur_id_type == 0x3 &&
3044                             cur_id_size == d[3])
3045                                 break;
3046                         cur_id_size = d[3];
3047                         cur_id_str = d + 4;
3048                         cur_id_type = d[1] & 0xf;
3049                         switch (cur_id_size) {
3050                         case 8:
3051                                 id_size = snprintf(id, id_len,
3052                                                    "eui.%8phN",
3053                                                    cur_id_str);
3054                                 break;
3055                         case 12:
3056                                 id_size = snprintf(id, id_len,
3057                                                    "eui.%12phN",
3058                                                    cur_id_str);
3059                                 break;
3060                         case 16:
3061                                 id_size = snprintf(id, id_len,
3062                                                    "eui.%16phN",
3063                                                    cur_id_str);
3064                                 break;
3065                         default:
3066                                 cur_id_size = 0;
3067                                 break;
3068                         }
3069                         break;
3070                 case 0x3:
3071                         /* NAA */
3072                         if (cur_id_size > d[3])
3073                                 break;
3074                         cur_id_size = d[3];
3075                         cur_id_str = d + 4;
3076                         cur_id_type = d[1] & 0xf;
3077                         switch (cur_id_size) {
3078                         case 8:
3079                                 id_size = snprintf(id, id_len,
3080                                                    "naa.%8phN",
3081                                                    cur_id_str);
3082                                 break;
3083                         case 16:
3084                                 id_size = snprintf(id, id_len,
3085                                                    "naa.%16phN",
3086                                                    cur_id_str);
3087                                 break;
3088                         default:
3089                                 cur_id_size = 0;
3090                                 break;
3091                         }
3092                         break;
3093                 case 0x8:
3094                         /* SCSI name string */
3095                         if (cur_id_size + 4 > d[3])
3096                                 break;
3097                         /* Prefer others for truncated descriptor */
3098                         if (cur_id_size && d[3] > id_len)
3099                                 break;
3100                         cur_id_size = id_size = d[3];
3101                         cur_id_str = d + 4;
3102                         cur_id_type = d[1] & 0xf;
3103                         if (cur_id_size >= id_len)
3104                                 cur_id_size = id_len - 1;
3105                         memcpy(id, cur_id_str, cur_id_size);
3106                         /* Decrease priority for truncated descriptor */
3107                         if (cur_id_size != id_size)
3108                                 cur_id_size = 6;
3109                         break;
3110                 default:
3111                         break;
3112                 }
3113 next_desig:
3114                 d += d[3] + 4;
3115         }
3116         rcu_read_unlock();
3117
3118         return id_size;
3119 }
3120 EXPORT_SYMBOL(scsi_vpd_lun_id);
3121
3122 /*
3123  * scsi_vpd_tpg_id - return a target port group identifier
3124  * @sdev: SCSI device
3125  *
3126  * Returns the Target Port Group identifier from the information
3127  * froom VPD page 0x83 of the device.
3128  *
3129  * Returns the identifier or error on failure.
3130  */
3131 int scsi_vpd_tpg_id(struct scsi_device *sdev, int *rel_id)
3132 {
3133         const unsigned char *d;
3134         const struct scsi_vpd *vpd_pg83;
3135         int group_id = -EAGAIN, rel_port = -1;
3136
3137         rcu_read_lock();
3138         vpd_pg83 = rcu_dereference(sdev->vpd_pg83);
3139         if (!vpd_pg83) {
3140                 rcu_read_unlock();
3141                 return -ENXIO;
3142         }
3143
3144         d = vpd_pg83->data + 4;
3145         while (d < vpd_pg83->data + vpd_pg83->len) {
3146                 switch (d[1] & 0xf) {
3147                 case 0x4:
3148                         /* Relative target port */
3149                         rel_port = get_unaligned_be16(&d[6]);
3150                         break;
3151                 case 0x5:
3152                         /* Target port group */
3153                         group_id = get_unaligned_be16(&d[6]);
3154                         break;
3155                 default:
3156                         break;
3157                 }
3158                 d += d[3] + 4;
3159         }
3160         rcu_read_unlock();
3161
3162         if (group_id >= 0 && rel_id && rel_port != -1)
3163                 *rel_id = rel_port;
3164
3165         return group_id;
3166 }
3167 EXPORT_SYMBOL(scsi_vpd_tpg_id);