GNU Linux-libre 5.4.241-gnu1
[releases.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);
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  * Decrement the host_busy counter and wake up the error handler if necessary.
325  * Avoid as follows that the error handler is not woken up if shost->host_busy
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 its
328  * 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)
333 {
334         unsigned long flags;
335
336         rcu_read_lock();
337         atomic_dec(&shost->host_busy);
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)
348 {
349         struct Scsi_Host *shost = sdev->host;
350         struct scsi_target *starget = scsi_target(sdev);
351
352         scsi_dec_host_busy(shost);
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 (shost->can_queue > 0 &&
434             atomic_read(&shost->host_busy) >= shost->can_queue)
435                 return true;
436         if (atomic_read(&shost->host_blocked) > 0)
437                 return true;
438         if (shost->host_self_blocked)
439                 return true;
440         return false;
441 }
442
443 static void scsi_starved_list_run(struct Scsi_Host *shost)
444 {
445         LIST_HEAD(starved_list);
446         struct scsi_device *sdev;
447         unsigned long flags;
448
449         spin_lock_irqsave(shost->host_lock, flags);
450         list_splice_init(&shost->starved_list, &starved_list);
451
452         while (!list_empty(&starved_list)) {
453                 struct request_queue *slq;
454
455                 /*
456                  * As long as shost is accepting commands and we have
457                  * starved queues, call blk_run_queue. scsi_request_fn
458                  * drops the queue_lock and can add us back to the
459                  * starved_list.
460                  *
461                  * host_lock protects the starved_list and starved_entry.
462                  * scsi_request_fn must get the host_lock before checking
463                  * or modifying starved_list or starved_entry.
464                  */
465                 if (scsi_host_is_busy(shost))
466                         break;
467
468                 sdev = list_entry(starved_list.next,
469                                   struct scsi_device, starved_entry);
470                 list_del_init(&sdev->starved_entry);
471                 if (scsi_target_is_busy(scsi_target(sdev))) {
472                         list_move_tail(&sdev->starved_entry,
473                                        &shost->starved_list);
474                         continue;
475                 }
476
477                 /*
478                  * Once we drop the host lock, a racing scsi_remove_device()
479                  * call may remove the sdev from the starved list and destroy
480                  * it and the queue.  Mitigate by taking a reference to the
481                  * queue and never touching the sdev again after we drop the
482                  * host lock.  Note: if __scsi_remove_device() invokes
483                  * blk_cleanup_queue() before the queue is run from this
484                  * function then blk_run_queue() will return immediately since
485                  * blk_cleanup_queue() marks the queue with QUEUE_FLAG_DYING.
486                  */
487                 slq = sdev->request_queue;
488                 if (!blk_get_queue(slq))
489                         continue;
490                 spin_unlock_irqrestore(shost->host_lock, flags);
491
492                 scsi_kick_queue(slq);
493                 blk_put_queue(slq);
494
495                 spin_lock_irqsave(shost->host_lock, flags);
496         }
497         /* put any unprocessed entries back */
498         list_splice(&starved_list, &shost->starved_list);
499         spin_unlock_irqrestore(shost->host_lock, flags);
500 }
501
502 /*
503  * Function:   scsi_run_queue()
504  *
505  * Purpose:    Select a proper request queue to serve next
506  *
507  * Arguments:  q       - last request's queue
508  *
509  * Returns:     Nothing
510  *
511  * Notes:      The previous command was completely finished, start
512  *             a new one if possible.
513  */
514 static void scsi_run_queue(struct request_queue *q)
515 {
516         struct scsi_device *sdev = q->queuedata;
517
518         if (scsi_target(sdev)->single_lun)
519                 scsi_single_lun_run(sdev);
520         if (!list_empty(&sdev->host->starved_list))
521                 scsi_starved_list_run(sdev->host);
522
523         blk_mq_run_hw_queues(q, false);
524 }
525
526 void scsi_requeue_run_queue(struct work_struct *work)
527 {
528         struct scsi_device *sdev;
529         struct request_queue *q;
530
531         sdev = container_of(work, struct scsi_device, requeue_work);
532         q = sdev->request_queue;
533         scsi_run_queue(q);
534 }
535
536 void scsi_run_host_queues(struct Scsi_Host *shost)
537 {
538         struct scsi_device *sdev;
539
540         shost_for_each_device(sdev, shost)
541                 scsi_run_queue(sdev->request_queue);
542 }
543
544 static void scsi_uninit_cmd(struct scsi_cmnd *cmd)
545 {
546         if (!blk_rq_is_passthrough(cmd->request)) {
547                 struct scsi_driver *drv = scsi_cmd_to_driver(cmd);
548
549                 if (drv->uninit_command)
550                         drv->uninit_command(cmd);
551         }
552 }
553
554 static void scsi_free_sgtables(struct scsi_cmnd *cmd)
555 {
556         if (cmd->sdb.table.nents)
557                 sg_free_table_chained(&cmd->sdb.table,
558                                 SCSI_INLINE_SG_CNT);
559         if (scsi_prot_sg_count(cmd))
560                 sg_free_table_chained(&cmd->prot_sdb->table,
561                                 SCSI_INLINE_PROT_SG_CNT);
562 }
563
564 static void scsi_mq_uninit_cmd(struct scsi_cmnd *cmd)
565 {
566         scsi_free_sgtables(cmd);
567         scsi_uninit_cmd(cmd);
568         scsi_del_cmd_from_list(cmd);
569 }
570
571 static void scsi_run_queue_async(struct scsi_device *sdev)
572 {
573         if (scsi_target(sdev)->single_lun ||
574             !list_empty(&sdev->host->starved_list))
575                 kblockd_schedule_work(&sdev->requeue_work);
576         else
577                 blk_mq_run_hw_queues(sdev->request_queue, true);
578 }
579
580 /* Returns false when no more bytes to process, true if there are more */
581 static bool scsi_end_request(struct request *req, blk_status_t error,
582                 unsigned int bytes)
583 {
584         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
585         struct scsi_device *sdev = cmd->device;
586         struct request_queue *q = sdev->request_queue;
587
588         if (blk_update_request(req, error, bytes))
589                 return true;
590
591         if (blk_queue_add_random(q))
592                 add_disk_randomness(req->rq_disk);
593
594         if (!blk_rq_is_scsi(req)) {
595                 WARN_ON_ONCE(!(cmd->flags & SCMD_INITIALIZED));
596                 cmd->flags &= ~SCMD_INITIALIZED;
597         }
598
599         /*
600          * Calling rcu_barrier() is not necessary here because the
601          * SCSI error handler guarantees that the function called by
602          * call_rcu() has been called before scsi_end_request() is
603          * called.
604          */
605         destroy_rcu_head(&cmd->rcu);
606
607         /*
608          * In the MQ case the command gets freed by __blk_mq_end_request,
609          * so we have to do all cleanup that depends on it earlier.
610          *
611          * We also can't kick the queues from irq context, so we
612          * will have to defer it to a workqueue.
613          */
614         scsi_mq_uninit_cmd(cmd);
615
616         /*
617          * queue is still alive, so grab the ref for preventing it
618          * from being cleaned up during running queue.
619          */
620         percpu_ref_get(&q->q_usage_counter);
621
622         __blk_mq_end_request(req, error);
623
624         scsi_run_queue_async(sdev);
625
626         percpu_ref_put(&q->q_usage_counter);
627         return false;
628 }
629
630 /**
631  * scsi_result_to_blk_status - translate a SCSI result code into blk_status_t
632  * @cmd:        SCSI command
633  * @result:     scsi error code
634  *
635  * Translate a SCSI result code into a blk_status_t value. May reset the host
636  * byte of @cmd->result.
637  */
638 static blk_status_t scsi_result_to_blk_status(struct scsi_cmnd *cmd, int result)
639 {
640         switch (host_byte(result)) {
641         case DID_OK:
642                 /*
643                  * Also check the other bytes than the status byte in result
644                  * to handle the case when a SCSI LLD sets result to
645                  * DRIVER_SENSE << 24 without setting SAM_STAT_CHECK_CONDITION.
646                  */
647                 if (scsi_status_is_good(result) && (result & ~0xff) == 0)
648                         return BLK_STS_OK;
649                 return BLK_STS_IOERR;
650         case DID_TRANSPORT_FAILFAST:
651                 return BLK_STS_TRANSPORT;
652         case DID_TARGET_FAILURE:
653                 set_host_byte(cmd, DID_OK);
654                 return BLK_STS_TARGET;
655         case DID_NEXUS_FAILURE:
656                 set_host_byte(cmd, DID_OK);
657                 return BLK_STS_NEXUS;
658         case DID_ALLOC_FAILURE:
659                 set_host_byte(cmd, DID_OK);
660                 return BLK_STS_NOSPC;
661         case DID_MEDIUM_ERROR:
662                 set_host_byte(cmd, DID_OK);
663                 return BLK_STS_MEDIUM;
664         default:
665                 return BLK_STS_IOERR;
666         }
667 }
668
669 /* Helper for scsi_io_completion() when "reprep" action required. */
670 static void scsi_io_completion_reprep(struct scsi_cmnd *cmd,
671                                       struct request_queue *q)
672 {
673         /* A new command will be prepared and issued. */
674         scsi_mq_requeue_cmd(cmd);
675 }
676
677 /* Helper for scsi_io_completion() when special action required. */
678 static void scsi_io_completion_action(struct scsi_cmnd *cmd, int result)
679 {
680         struct request_queue *q = cmd->device->request_queue;
681         struct request *req = cmd->request;
682         int level = 0;
683         enum {ACTION_FAIL, ACTION_REPREP, ACTION_RETRY,
684               ACTION_DELAYED_RETRY} action;
685         unsigned long wait_for = (cmd->allowed + 1) * req->timeout;
686         struct scsi_sense_hdr sshdr;
687         bool sense_valid;
688         bool sense_current = true;      /* false implies "deferred sense" */
689         blk_status_t blk_stat;
690
691         sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
692         if (sense_valid)
693                 sense_current = !scsi_sense_is_deferred(&sshdr);
694
695         blk_stat = scsi_result_to_blk_status(cmd, result);
696
697         if (host_byte(result) == DID_RESET) {
698                 /* Third party bus reset or reset for error recovery
699                  * reasons.  Just retry the command and see what
700                  * happens.
701                  */
702                 action = ACTION_RETRY;
703         } else if (sense_valid && sense_current) {
704                 switch (sshdr.sense_key) {
705                 case UNIT_ATTENTION:
706                         if (cmd->device->removable) {
707                                 /* Detected disc change.  Set a bit
708                                  * and quietly refuse further access.
709                                  */
710                                 cmd->device->changed = 1;
711                                 action = ACTION_FAIL;
712                         } else {
713                                 /* Must have been a power glitch, or a
714                                  * bus reset.  Could not have been a
715                                  * media change, so we just retry the
716                                  * command and see what happens.
717                                  */
718                                 action = ACTION_RETRY;
719                         }
720                         break;
721                 case ILLEGAL_REQUEST:
722                         /* If we had an ILLEGAL REQUEST returned, then
723                          * we may have performed an unsupported
724                          * command.  The only thing this should be
725                          * would be a ten byte read where only a six
726                          * byte read was supported.  Also, on a system
727                          * where READ CAPACITY failed, we may have
728                          * read past the end of the disk.
729                          */
730                         if ((cmd->device->use_10_for_rw &&
731                             sshdr.asc == 0x20 && sshdr.ascq == 0x00) &&
732                             (cmd->cmnd[0] == READ_10 ||
733                              cmd->cmnd[0] == WRITE_10)) {
734                                 /* This will issue a new 6-byte command. */
735                                 cmd->device->use_10_for_rw = 0;
736                                 action = ACTION_REPREP;
737                         } else if (sshdr.asc == 0x10) /* DIX */ {
738                                 action = ACTION_FAIL;
739                                 blk_stat = BLK_STS_PROTECTION;
740                         /* INVALID COMMAND OPCODE or INVALID FIELD IN CDB */
741                         } else if (sshdr.asc == 0x20 || sshdr.asc == 0x24) {
742                                 action = ACTION_FAIL;
743                                 blk_stat = BLK_STS_TARGET;
744                         } else
745                                 action = ACTION_FAIL;
746                         break;
747                 case ABORTED_COMMAND:
748                         action = ACTION_FAIL;
749                         if (sshdr.asc == 0x10) /* DIF */
750                                 blk_stat = BLK_STS_PROTECTION;
751                         break;
752                 case NOT_READY:
753                         /* If the device is in the process of becoming
754                          * ready, or has a temporary blockage, retry.
755                          */
756                         if (sshdr.asc == 0x04) {
757                                 switch (sshdr.ascq) {
758                                 case 0x01: /* becoming ready */
759                                 case 0x04: /* format in progress */
760                                 case 0x05: /* rebuild in progress */
761                                 case 0x06: /* recalculation in progress */
762                                 case 0x07: /* operation in progress */
763                                 case 0x08: /* Long write in progress */
764                                 case 0x09: /* self test in progress */
765                                 case 0x11: /* notify (enable spinup) required */
766                                 case 0x14: /* space allocation in progress */
767                                 case 0x1a: /* start stop unit in progress */
768                                 case 0x1b: /* sanitize in progress */
769                                 case 0x1d: /* configuration in progress */
770                                 case 0x24: /* depopulation in progress */
771                                         action = ACTION_DELAYED_RETRY;
772                                         break;
773                                 default:
774                                         action = ACTION_FAIL;
775                                         break;
776                                 }
777                         } else
778                                 action = ACTION_FAIL;
779                         break;
780                 case VOLUME_OVERFLOW:
781                         /* See SSC3rXX or current. */
782                         action = ACTION_FAIL;
783                         break;
784                 default:
785                         action = ACTION_FAIL;
786                         break;
787                 }
788         } else
789                 action = ACTION_FAIL;
790
791         if (action != ACTION_FAIL &&
792             time_before(cmd->jiffies_at_alloc + wait_for, jiffies))
793                 action = ACTION_FAIL;
794
795         switch (action) {
796         case ACTION_FAIL:
797                 /* Give up and fail the remainder of the request */
798                 if (!(req->rq_flags & RQF_QUIET)) {
799                         static DEFINE_RATELIMIT_STATE(_rs,
800                                         DEFAULT_RATELIMIT_INTERVAL,
801                                         DEFAULT_RATELIMIT_BURST);
802
803                         if (unlikely(scsi_logging_level))
804                                 level =
805                                      SCSI_LOG_LEVEL(SCSI_LOG_MLCOMPLETE_SHIFT,
806                                                     SCSI_LOG_MLCOMPLETE_BITS);
807
808                         /*
809                          * if logging is enabled the failure will be printed
810                          * in scsi_log_completion(), so avoid duplicate messages
811                          */
812                         if (!level && __ratelimit(&_rs)) {
813                                 scsi_print_result(cmd, NULL, FAILED);
814                                 if (driver_byte(result) == DRIVER_SENSE)
815                                         scsi_print_sense(cmd);
816                                 scsi_print_command(cmd);
817                         }
818                 }
819                 if (!scsi_end_request(req, blk_stat, blk_rq_err_bytes(req)))
820                         return;
821                 /*FALLTHRU*/
822         case ACTION_REPREP:
823                 scsi_io_completion_reprep(cmd, q);
824                 break;
825         case ACTION_RETRY:
826                 /* Retry the same command immediately */
827                 __scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY, false);
828                 break;
829         case ACTION_DELAYED_RETRY:
830                 /* Retry the same command after a delay */
831                 __scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY, false);
832                 break;
833         }
834 }
835
836 /*
837  * Helper for scsi_io_completion() when cmd->result is non-zero. Returns a
838  * new result that may suppress further error checking. Also modifies
839  * *blk_statp in some cases.
840  */
841 static int scsi_io_completion_nz_result(struct scsi_cmnd *cmd, int result,
842                                         blk_status_t *blk_statp)
843 {
844         bool sense_valid;
845         bool sense_current = true;      /* false implies "deferred sense" */
846         struct request *req = cmd->request;
847         struct scsi_sense_hdr sshdr;
848
849         sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
850         if (sense_valid)
851                 sense_current = !scsi_sense_is_deferred(&sshdr);
852
853         if (blk_rq_is_passthrough(req)) {
854                 if (sense_valid) {
855                         /*
856                          * SG_IO wants current and deferred errors
857                          */
858                         scsi_req(req)->sense_len =
859                                 min(8 + cmd->sense_buffer[7],
860                                     SCSI_SENSE_BUFFERSIZE);
861                 }
862                 if (sense_current)
863                         *blk_statp = scsi_result_to_blk_status(cmd, result);
864         } else if (blk_rq_bytes(req) == 0 && sense_current) {
865                 /*
866                  * Flush commands do not transfers any data, and thus cannot use
867                  * good_bytes != blk_rq_bytes(req) as the signal for an error.
868                  * This sets *blk_statp explicitly for the problem case.
869                  */
870                 *blk_statp = scsi_result_to_blk_status(cmd, result);
871         }
872         /*
873          * Recovered errors need reporting, but they're always treated as
874          * success, so fiddle the result code here.  For passthrough requests
875          * we already took a copy of the original into sreq->result which
876          * is what gets returned to the user
877          */
878         if (sense_valid && (sshdr.sense_key == RECOVERED_ERROR)) {
879                 bool do_print = true;
880                 /*
881                  * if ATA PASS-THROUGH INFORMATION AVAILABLE [0x0, 0x1d]
882                  * skip print since caller wants ATA registers. Only occurs
883                  * on SCSI ATA PASS_THROUGH commands when CK_COND=1
884                  */
885                 if ((sshdr.asc == 0x0) && (sshdr.ascq == 0x1d))
886                         do_print = false;
887                 else if (req->rq_flags & RQF_QUIET)
888                         do_print = false;
889                 if (do_print)
890                         scsi_print_sense(cmd);
891                 result = 0;
892                 /* for passthrough, *blk_statp may be set */
893                 *blk_statp = BLK_STS_OK;
894         }
895         /*
896          * Another corner case: the SCSI status byte is non-zero but 'good'.
897          * Example: PRE-FETCH command returns SAM_STAT_CONDITION_MET when
898          * it is able to fit nominated LBs in its cache (and SAM_STAT_GOOD
899          * if it can't fit). Treat SAM_STAT_CONDITION_MET and the related
900          * intermediate statuses (both obsolete in SAM-4) as good.
901          */
902         if (status_byte(result) && scsi_status_is_good(result)) {
903                 result = 0;
904                 *blk_statp = BLK_STS_OK;
905         }
906         return result;
907 }
908
909 /*
910  * Function:    scsi_io_completion()
911  *
912  * Purpose:     Completion processing for block device I/O requests.
913  *
914  * Arguments:   cmd   - command that is finished.
915  *
916  * Lock status: Assumed that no lock is held upon entry.
917  *
918  * Returns:     Nothing
919  *
920  * Notes:       We will finish off the specified number of sectors.  If we
921  *              are done, the command block will be released and the queue
922  *              function will be goosed.  If we are not done then we have to
923  *              figure out what to do next:
924  *
925  *              a) We can call scsi_requeue_command().  The request
926  *                 will be unprepared and put back on the queue.  Then
927  *                 a new command will be created for it.  This should
928  *                 be used if we made forward progress, or if we want
929  *                 to switch from READ(10) to READ(6) for example.
930  *
931  *              b) We can call __scsi_queue_insert().  The request will
932  *                 be put back on the queue and retried using the same
933  *                 command as before, possibly after a delay.
934  *
935  *              c) We can call scsi_end_request() with blk_stat other than
936  *                 BLK_STS_OK, to fail the remainder of the request.
937  */
938 void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes)
939 {
940         int result = cmd->result;
941         struct request_queue *q = cmd->device->request_queue;
942         struct request *req = cmd->request;
943         blk_status_t blk_stat = BLK_STS_OK;
944
945         if (unlikely(result))   /* a nz result may or may not be an error */
946                 result = scsi_io_completion_nz_result(cmd, result, &blk_stat);
947
948         if (unlikely(blk_rq_is_passthrough(req))) {
949                 /*
950                  * scsi_result_to_blk_status may have reset the host_byte
951                  */
952                 scsi_req(req)->result = cmd->result;
953         }
954
955         /*
956          * Next deal with any sectors which we were able to correctly
957          * handle.
958          */
959         SCSI_LOG_HLCOMPLETE(1, scmd_printk(KERN_INFO, cmd,
960                 "%u sectors total, %d bytes done.\n",
961                 blk_rq_sectors(req), good_bytes));
962
963         /*
964          * Next deal with any sectors which we were able to correctly
965          * handle. Failed, zero length commands always need to drop down
966          * to retry code. Fast path should return in this block.
967          */
968         if (likely(blk_rq_bytes(req) > 0 || blk_stat == BLK_STS_OK)) {
969                 if (likely(!scsi_end_request(req, blk_stat, good_bytes)))
970                         return; /* no bytes remaining */
971         }
972
973         /* Kill remainder if no retries. */
974         if (unlikely(blk_stat && scsi_noretry_cmd(cmd))) {
975                 if (scsi_end_request(req, blk_stat, blk_rq_bytes(req)))
976                         WARN_ONCE(true,
977                             "Bytes remaining after failed, no-retry command");
978                 return;
979         }
980
981         /*
982          * If there had been no error, but we have leftover bytes in the
983          * requeues just queue the command up again.
984          */
985         if (likely(result == 0))
986                 scsi_io_completion_reprep(cmd, q);
987         else
988                 scsi_io_completion_action(cmd, result);
989 }
990
991 static blk_status_t scsi_init_sgtable(struct request *req,
992                 struct scsi_data_buffer *sdb)
993 {
994         int count;
995
996         /*
997          * If sg table allocation fails, requeue request later.
998          */
999         if (unlikely(sg_alloc_table_chained(&sdb->table,
1000                         blk_rq_nr_phys_segments(req), sdb->table.sgl,
1001                         SCSI_INLINE_SG_CNT)))
1002                 return BLK_STS_RESOURCE;
1003
1004         /* 
1005          * Next, walk the list, and fill in the addresses and sizes of
1006          * each segment.
1007          */
1008         count = blk_rq_map_sg(req->q, req, sdb->table.sgl);
1009         BUG_ON(count > sdb->table.nents);
1010         sdb->table.nents = count;
1011         sdb->length = blk_rq_payload_bytes(req);
1012         return BLK_STS_OK;
1013 }
1014
1015 /*
1016  * Function:    scsi_init_io()
1017  *
1018  * Purpose:     SCSI I/O initialize function.
1019  *
1020  * Arguments:   cmd   - Command descriptor we wish to initialize
1021  *
1022  * Returns:     BLK_STS_OK on success
1023  *              BLK_STS_RESOURCE if the failure is retryable
1024  *              BLK_STS_IOERR if the failure is fatal
1025  */
1026 blk_status_t scsi_init_io(struct scsi_cmnd *cmd)
1027 {
1028         struct request *rq = cmd->request;
1029         blk_status_t ret;
1030
1031         if (WARN_ON_ONCE(!blk_rq_nr_phys_segments(rq)))
1032                 return BLK_STS_IOERR;
1033
1034         ret = scsi_init_sgtable(rq, &cmd->sdb);
1035         if (ret)
1036                 return ret;
1037
1038         if (blk_integrity_rq(rq)) {
1039                 struct scsi_data_buffer *prot_sdb = cmd->prot_sdb;
1040                 int ivecs, count;
1041
1042                 if (WARN_ON_ONCE(!prot_sdb)) {
1043                         /*
1044                          * This can happen if someone (e.g. multipath)
1045                          * queues a command to a device on an adapter
1046                          * that does not support DIX.
1047                          */
1048                         ret = BLK_STS_IOERR;
1049                         goto out_free_sgtables;
1050                 }
1051
1052                 ivecs = blk_rq_count_integrity_sg(rq->q, rq->bio);
1053
1054                 if (sg_alloc_table_chained(&prot_sdb->table, ivecs,
1055                                 prot_sdb->table.sgl,
1056                                 SCSI_INLINE_PROT_SG_CNT)) {
1057                         ret = BLK_STS_RESOURCE;
1058                         goto out_free_sgtables;
1059                 }
1060
1061                 count = blk_rq_map_integrity_sg(rq->q, rq->bio,
1062                                                 prot_sdb->table.sgl);
1063                 BUG_ON(count > ivecs);
1064                 BUG_ON(count > queue_max_integrity_segments(rq->q));
1065
1066                 cmd->prot_sdb = prot_sdb;
1067                 cmd->prot_sdb->table.nents = count;
1068         }
1069
1070         return BLK_STS_OK;
1071 out_free_sgtables:
1072         scsi_free_sgtables(cmd);
1073         return ret;
1074 }
1075 EXPORT_SYMBOL(scsi_init_io);
1076
1077 /**
1078  * scsi_initialize_rq - initialize struct scsi_cmnd partially
1079  * @rq: Request associated with the SCSI command to be initialized.
1080  *
1081  * This function initializes the members of struct scsi_cmnd that must be
1082  * initialized before request processing starts and that won't be
1083  * reinitialized if a SCSI command is requeued.
1084  *
1085  * Called from inside blk_get_request() for pass-through requests and from
1086  * inside scsi_init_command() for filesystem requests.
1087  */
1088 static void scsi_initialize_rq(struct request *rq)
1089 {
1090         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1091
1092         scsi_req_init(&cmd->req);
1093         init_rcu_head(&cmd->rcu);
1094         cmd->jiffies_at_alloc = jiffies;
1095         cmd->retries = 0;
1096 }
1097
1098 /*
1099  * Only called when the request isn't completed by SCSI, and not freed by
1100  * SCSI
1101  */
1102 static void scsi_cleanup_rq(struct request *rq)
1103 {
1104         if (rq->rq_flags & RQF_DONTPREP) {
1105                 scsi_mq_uninit_cmd(blk_mq_rq_to_pdu(rq));
1106                 rq->rq_flags &= ~RQF_DONTPREP;
1107         }
1108 }
1109
1110 /* Add a command to the list used by the aacraid and dpt_i2o drivers */
1111 void scsi_add_cmd_to_list(struct scsi_cmnd *cmd)
1112 {
1113         struct scsi_device *sdev = cmd->device;
1114         struct Scsi_Host *shost = sdev->host;
1115         unsigned long flags;
1116
1117         if (shost->use_cmd_list) {
1118                 spin_lock_irqsave(&sdev->list_lock, flags);
1119                 list_add_tail(&cmd->list, &sdev->cmd_list);
1120                 spin_unlock_irqrestore(&sdev->list_lock, flags);
1121         }
1122 }
1123
1124 /* Remove a command from the list used by the aacraid and dpt_i2o drivers */
1125 void scsi_del_cmd_from_list(struct scsi_cmnd *cmd)
1126 {
1127         struct scsi_device *sdev = cmd->device;
1128         struct Scsi_Host *shost = sdev->host;
1129         unsigned long flags;
1130
1131         if (shost->use_cmd_list) {
1132                 spin_lock_irqsave(&sdev->list_lock, flags);
1133                 BUG_ON(list_empty(&cmd->list));
1134                 list_del_init(&cmd->list);
1135                 spin_unlock_irqrestore(&sdev->list_lock, flags);
1136         }
1137 }
1138
1139 /* Called after a request has been started. */
1140 void scsi_init_command(struct scsi_device *dev, struct scsi_cmnd *cmd)
1141 {
1142         void *buf = cmd->sense_buffer;
1143         void *prot = cmd->prot_sdb;
1144         struct request *rq = blk_mq_rq_from_pdu(cmd);
1145         unsigned int flags = cmd->flags & SCMD_PRESERVED_FLAGS;
1146         unsigned long jiffies_at_alloc;
1147         int retries;
1148
1149         if (!blk_rq_is_scsi(rq) && !(flags & SCMD_INITIALIZED)) {
1150                 flags |= SCMD_INITIALIZED;
1151                 scsi_initialize_rq(rq);
1152         }
1153
1154         jiffies_at_alloc = cmd->jiffies_at_alloc;
1155         retries = cmd->retries;
1156         /* zero out the cmd, except for the embedded scsi_request */
1157         memset((char *)cmd + sizeof(cmd->req), 0,
1158                 sizeof(*cmd) - sizeof(cmd->req) + dev->host->hostt->cmd_size);
1159
1160         cmd->device = dev;
1161         cmd->sense_buffer = buf;
1162         cmd->prot_sdb = prot;
1163         cmd->flags = flags;
1164         INIT_DELAYED_WORK(&cmd->abort_work, scmd_eh_abort_handler);
1165         cmd->jiffies_at_alloc = jiffies_at_alloc;
1166         cmd->retries = retries;
1167
1168         scsi_add_cmd_to_list(cmd);
1169 }
1170
1171 static blk_status_t scsi_setup_scsi_cmnd(struct scsi_device *sdev,
1172                 struct request *req)
1173 {
1174         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1175
1176         /*
1177          * Passthrough requests may transfer data, in which case they must
1178          * a bio attached to them.  Or they might contain a SCSI command
1179          * that does not transfer data, in which case they may optionally
1180          * submit a request without an attached bio.
1181          */
1182         if (req->bio) {
1183                 blk_status_t ret = scsi_init_io(cmd);
1184                 if (unlikely(ret != BLK_STS_OK))
1185                         return ret;
1186         } else {
1187                 BUG_ON(blk_rq_bytes(req));
1188
1189                 memset(&cmd->sdb, 0, sizeof(cmd->sdb));
1190         }
1191
1192         cmd->cmd_len = scsi_req(req)->cmd_len;
1193         cmd->cmnd = scsi_req(req)->cmd;
1194         cmd->transfersize = blk_rq_bytes(req);
1195         cmd->allowed = scsi_req(req)->retries;
1196         return BLK_STS_OK;
1197 }
1198
1199 /*
1200  * Setup a normal block command.  These are simple request from filesystems
1201  * that still need to be translated to SCSI CDBs from the ULD.
1202  */
1203 static blk_status_t scsi_setup_fs_cmnd(struct scsi_device *sdev,
1204                 struct request *req)
1205 {
1206         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1207
1208         if (unlikely(sdev->handler && sdev->handler->prep_fn)) {
1209                 blk_status_t ret = sdev->handler->prep_fn(sdev, req);
1210                 if (ret != BLK_STS_OK)
1211                         return ret;
1212         }
1213
1214         cmd->cmnd = scsi_req(req)->cmd = scsi_req(req)->__cmd;
1215         memset(cmd->cmnd, 0, BLK_MAX_CDB);
1216         return scsi_cmd_to_driver(cmd)->init_command(cmd);
1217 }
1218
1219 static blk_status_t scsi_setup_cmnd(struct scsi_device *sdev,
1220                 struct request *req)
1221 {
1222         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1223         blk_status_t ret;
1224
1225         if (!blk_rq_bytes(req))
1226                 cmd->sc_data_direction = DMA_NONE;
1227         else if (rq_data_dir(req) == WRITE)
1228                 cmd->sc_data_direction = DMA_TO_DEVICE;
1229         else
1230                 cmd->sc_data_direction = DMA_FROM_DEVICE;
1231
1232         if (blk_rq_is_scsi(req))
1233                 ret = scsi_setup_scsi_cmnd(sdev, req);
1234         else
1235                 ret = scsi_setup_fs_cmnd(sdev, req);
1236
1237         if (ret != BLK_STS_OK)
1238                 scsi_free_sgtables(cmd);
1239
1240         return ret;
1241 }
1242
1243 static blk_status_t
1244 scsi_prep_state_check(struct scsi_device *sdev, struct request *req)
1245 {
1246         switch (sdev->sdev_state) {
1247         case SDEV_OFFLINE:
1248         case SDEV_TRANSPORT_OFFLINE:
1249                 /*
1250                  * If the device is offline we refuse to process any
1251                  * commands.  The device must be brought online
1252                  * before trying any recovery commands.
1253                  */
1254                 sdev_printk(KERN_ERR, sdev,
1255                             "rejecting I/O to offline device\n");
1256                 return BLK_STS_IOERR;
1257         case SDEV_DEL:
1258                 /*
1259                  * If the device is fully deleted, we refuse to
1260                  * process any commands as well.
1261                  */
1262                 sdev_printk(KERN_ERR, sdev,
1263                             "rejecting I/O to dead device\n");
1264                 return BLK_STS_IOERR;
1265         case SDEV_BLOCK:
1266         case SDEV_CREATED_BLOCK:
1267                 return BLK_STS_RESOURCE;
1268         case SDEV_QUIESCE:
1269                 /*
1270                  * If the devices is blocked we defer normal commands.
1271                  */
1272                 if (req && !(req->rq_flags & RQF_PREEMPT))
1273                         return BLK_STS_RESOURCE;
1274                 return BLK_STS_OK;
1275         default:
1276                 /*
1277                  * For any other not fully online state we only allow
1278                  * special commands.  In particular any user initiated
1279                  * command is not allowed.
1280                  */
1281                 if (req && !(req->rq_flags & RQF_PREEMPT))
1282                         return BLK_STS_IOERR;
1283                 return BLK_STS_OK;
1284         }
1285 }
1286
1287 /*
1288  * scsi_dev_queue_ready: if we can send requests to sdev, return 1 else
1289  * return 0.
1290  *
1291  * Called with the queue_lock held.
1292  */
1293 static inline int scsi_dev_queue_ready(struct request_queue *q,
1294                                   struct scsi_device *sdev)
1295 {
1296         unsigned int busy;
1297
1298         busy = atomic_inc_return(&sdev->device_busy) - 1;
1299         if (atomic_read(&sdev->device_blocked)) {
1300                 if (busy)
1301                         goto out_dec;
1302
1303                 /*
1304                  * unblock after device_blocked iterates to zero
1305                  */
1306                 if (atomic_dec_return(&sdev->device_blocked) > 0)
1307                         goto out_dec;
1308                 SCSI_LOG_MLQUEUE(3, sdev_printk(KERN_INFO, sdev,
1309                                    "unblocking device at zero depth\n"));
1310         }
1311
1312         if (busy >= sdev->queue_depth)
1313                 goto out_dec;
1314
1315         return 1;
1316 out_dec:
1317         atomic_dec(&sdev->device_busy);
1318         return 0;
1319 }
1320
1321 /*
1322  * scsi_target_queue_ready: checks if there we can send commands to target
1323  * @sdev: scsi device on starget to check.
1324  */
1325 static inline int scsi_target_queue_ready(struct Scsi_Host *shost,
1326                                            struct scsi_device *sdev)
1327 {
1328         struct scsi_target *starget = scsi_target(sdev);
1329         unsigned int busy;
1330
1331         if (starget->single_lun) {
1332                 spin_lock_irq(shost->host_lock);
1333                 if (starget->starget_sdev_user &&
1334                     starget->starget_sdev_user != sdev) {
1335                         spin_unlock_irq(shost->host_lock);
1336                         return 0;
1337                 }
1338                 starget->starget_sdev_user = sdev;
1339                 spin_unlock_irq(shost->host_lock);
1340         }
1341
1342         if (starget->can_queue <= 0)
1343                 return 1;
1344
1345         busy = atomic_inc_return(&starget->target_busy) - 1;
1346         if (atomic_read(&starget->target_blocked) > 0) {
1347                 if (busy)
1348                         goto starved;
1349
1350                 /*
1351                  * unblock after target_blocked iterates to zero
1352                  */
1353                 if (atomic_dec_return(&starget->target_blocked) > 0)
1354                         goto out_dec;
1355
1356                 SCSI_LOG_MLQUEUE(3, starget_printk(KERN_INFO, starget,
1357                                  "unblocking target at zero depth\n"));
1358         }
1359
1360         if (busy >= starget->can_queue)
1361                 goto starved;
1362
1363         return 1;
1364
1365 starved:
1366         spin_lock_irq(shost->host_lock);
1367         list_move_tail(&sdev->starved_entry, &shost->starved_list);
1368         spin_unlock_irq(shost->host_lock);
1369 out_dec:
1370         if (starget->can_queue > 0)
1371                 atomic_dec(&starget->target_busy);
1372         return 0;
1373 }
1374
1375 /*
1376  * scsi_host_queue_ready: if we can send requests to shost, return 1 else
1377  * return 0. We must end up running the queue again whenever 0 is
1378  * returned, else IO can hang.
1379  */
1380 static inline int scsi_host_queue_ready(struct request_queue *q,
1381                                    struct Scsi_Host *shost,
1382                                    struct scsi_device *sdev)
1383 {
1384         unsigned int busy;
1385
1386         if (scsi_host_in_recovery(shost))
1387                 return 0;
1388
1389         busy = atomic_inc_return(&shost->host_busy) - 1;
1390         if (atomic_read(&shost->host_blocked) > 0) {
1391                 if (busy)
1392                         goto starved;
1393
1394                 /*
1395                  * unblock after host_blocked iterates to zero
1396                  */
1397                 if (atomic_dec_return(&shost->host_blocked) > 0)
1398                         goto out_dec;
1399
1400                 SCSI_LOG_MLQUEUE(3,
1401                         shost_printk(KERN_INFO, shost,
1402                                      "unblocking host at zero depth\n"));
1403         }
1404
1405         if (shost->can_queue > 0 && busy >= shost->can_queue)
1406                 goto starved;
1407         if (shost->host_self_blocked)
1408                 goto starved;
1409
1410         /* We're OK to process the command, so we can't be starved */
1411         if (!list_empty(&sdev->starved_entry)) {
1412                 spin_lock_irq(shost->host_lock);
1413                 if (!list_empty(&sdev->starved_entry))
1414                         list_del_init(&sdev->starved_entry);
1415                 spin_unlock_irq(shost->host_lock);
1416         }
1417
1418         return 1;
1419
1420 starved:
1421         spin_lock_irq(shost->host_lock);
1422         if (list_empty(&sdev->starved_entry))
1423                 list_add_tail(&sdev->starved_entry, &shost->starved_list);
1424         spin_unlock_irq(shost->host_lock);
1425 out_dec:
1426         scsi_dec_host_busy(shost);
1427         return 0;
1428 }
1429
1430 /*
1431  * Busy state exporting function for request stacking drivers.
1432  *
1433  * For efficiency, no lock is taken to check the busy state of
1434  * shost/starget/sdev, since the returned value is not guaranteed and
1435  * may be changed after request stacking drivers call the function,
1436  * regardless of taking lock or not.
1437  *
1438  * When scsi can't dispatch I/Os anymore and needs to kill I/Os scsi
1439  * needs to return 'not busy'. Otherwise, request stacking drivers
1440  * may hold requests forever.
1441  */
1442 static bool scsi_mq_lld_busy(struct request_queue *q)
1443 {
1444         struct scsi_device *sdev = q->queuedata;
1445         struct Scsi_Host *shost;
1446
1447         if (blk_queue_dying(q))
1448                 return false;
1449
1450         shost = sdev->host;
1451
1452         /*
1453          * Ignore host/starget busy state.
1454          * Since block layer does not have a concept of fairness across
1455          * multiple queues, congestion of host/starget needs to be handled
1456          * in SCSI layer.
1457          */
1458         if (scsi_host_in_recovery(shost) || scsi_device_is_busy(sdev))
1459                 return true;
1460
1461         return false;
1462 }
1463
1464 static void scsi_softirq_done(struct request *rq)
1465 {
1466         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1467         unsigned long wait_for = (cmd->allowed + 1) * rq->timeout;
1468         int disposition;
1469
1470         INIT_LIST_HEAD(&cmd->eh_entry);
1471
1472         atomic_inc(&cmd->device->iodone_cnt);
1473         if (cmd->result)
1474                 atomic_inc(&cmd->device->ioerr_cnt);
1475
1476         disposition = scsi_decide_disposition(cmd);
1477         if (disposition != SUCCESS &&
1478             time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) {
1479                 scmd_printk(KERN_ERR, cmd,
1480                             "timing out command, waited %lus\n",
1481                             wait_for/HZ);
1482                 disposition = SUCCESS;
1483         }
1484
1485         scsi_log_completion(cmd, disposition);
1486
1487         switch (disposition) {
1488                 case SUCCESS:
1489                         scsi_finish_command(cmd);
1490                         break;
1491                 case NEEDS_RETRY:
1492                         scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY);
1493                         break;
1494                 case ADD_TO_MLQUEUE:
1495                         scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY);
1496                         break;
1497                 default:
1498                         scsi_eh_scmd_add(cmd);
1499                         break;
1500         }
1501 }
1502
1503 /**
1504  * scsi_dispatch_command - Dispatch a command to the low-level driver.
1505  * @cmd: command block we are dispatching.
1506  *
1507  * Return: nonzero return request was rejected and device's queue needs to be
1508  * plugged.
1509  */
1510 static int scsi_dispatch_cmd(struct scsi_cmnd *cmd)
1511 {
1512         struct Scsi_Host *host = cmd->device->host;
1513         int rtn = 0;
1514
1515         atomic_inc(&cmd->device->iorequest_cnt);
1516
1517         /* check if the device is still usable */
1518         if (unlikely(cmd->device->sdev_state == SDEV_DEL)) {
1519                 /* in SDEV_DEL we error all commands. DID_NO_CONNECT
1520                  * returns an immediate error upwards, and signals
1521                  * that the device is no longer present */
1522                 cmd->result = DID_NO_CONNECT << 16;
1523                 goto done;
1524         }
1525
1526         /* Check to see if the scsi lld made this device blocked. */
1527         if (unlikely(scsi_device_blocked(cmd->device))) {
1528                 /*
1529                  * in blocked state, the command is just put back on
1530                  * the device queue.  The suspend state has already
1531                  * blocked the queue so future requests should not
1532                  * occur until the device transitions out of the
1533                  * suspend state.
1534                  */
1535                 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1536                         "queuecommand : device blocked\n"));
1537                 return SCSI_MLQUEUE_DEVICE_BUSY;
1538         }
1539
1540         /* Store the LUN value in cmnd, if needed. */
1541         if (cmd->device->lun_in_cdb)
1542                 cmd->cmnd[1] = (cmd->cmnd[1] & 0x1f) |
1543                                (cmd->device->lun << 5 & 0xe0);
1544
1545         scsi_log_send(cmd);
1546
1547         /*
1548          * Before we queue this command, check if the command
1549          * length exceeds what the host adapter can handle.
1550          */
1551         if (cmd->cmd_len > cmd->device->host->max_cmd_len) {
1552                 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1553                                "queuecommand : command too long. "
1554                                "cdb_size=%d host->max_cmd_len=%d\n",
1555                                cmd->cmd_len, cmd->device->host->max_cmd_len));
1556                 cmd->result = (DID_ABORT << 16);
1557                 goto done;
1558         }
1559
1560         if (unlikely(host->shost_state == SHOST_DEL)) {
1561                 cmd->result = (DID_NO_CONNECT << 16);
1562                 goto done;
1563
1564         }
1565
1566         trace_scsi_dispatch_cmd_start(cmd);
1567         rtn = host->hostt->queuecommand(host, cmd);
1568         if (rtn) {
1569                 trace_scsi_dispatch_cmd_error(cmd, rtn);
1570                 if (rtn != SCSI_MLQUEUE_DEVICE_BUSY &&
1571                     rtn != SCSI_MLQUEUE_TARGET_BUSY)
1572                         rtn = SCSI_MLQUEUE_HOST_BUSY;
1573
1574                 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1575                         "queuecommand : request rejected\n"));
1576         }
1577
1578         return rtn;
1579  done:
1580         cmd->scsi_done(cmd);
1581         return 0;
1582 }
1583
1584 /* Size in bytes of the sg-list stored in the scsi-mq command-private data. */
1585 static unsigned int scsi_mq_inline_sgl_size(struct Scsi_Host *shost)
1586 {
1587         return min_t(unsigned int, shost->sg_tablesize, SCSI_INLINE_SG_CNT) *
1588                 sizeof(struct scatterlist);
1589 }
1590
1591 static blk_status_t scsi_mq_prep_fn(struct request *req)
1592 {
1593         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1594         struct scsi_device *sdev = req->q->queuedata;
1595         struct Scsi_Host *shost = sdev->host;
1596         struct scatterlist *sg;
1597
1598         scsi_init_command(sdev, cmd);
1599
1600         cmd->request = req;
1601         cmd->tag = req->tag;
1602         cmd->prot_op = SCSI_PROT_NORMAL;
1603
1604         sg = (void *)cmd + sizeof(struct scsi_cmnd) + shost->hostt->cmd_size;
1605         cmd->sdb.table.sgl = sg;
1606
1607         if (scsi_host_get_prot(shost)) {
1608                 memset(cmd->prot_sdb, 0, sizeof(struct scsi_data_buffer));
1609
1610                 cmd->prot_sdb->table.sgl =
1611                         (struct scatterlist *)(cmd->prot_sdb + 1);
1612         }
1613
1614         blk_mq_start_request(req);
1615
1616         return scsi_setup_cmnd(sdev, req);
1617 }
1618
1619 static void scsi_mq_done(struct scsi_cmnd *cmd)
1620 {
1621         if (unlikely(test_and_set_bit(SCMD_STATE_COMPLETE, &cmd->state)))
1622                 return;
1623         trace_scsi_dispatch_cmd_done(cmd);
1624
1625         /*
1626          * If the block layer didn't complete the request due to a timeout
1627          * injection, scsi must clear its internal completed state so that the
1628          * timeout handler will see it needs to escalate its own error
1629          * recovery.
1630          */
1631         if (unlikely(!blk_mq_complete_request(cmd->request)))
1632                 clear_bit(SCMD_STATE_COMPLETE, &cmd->state);
1633 }
1634
1635 static void scsi_mq_put_budget(struct blk_mq_hw_ctx *hctx)
1636 {
1637         struct request_queue *q = hctx->queue;
1638         struct scsi_device *sdev = q->queuedata;
1639
1640         atomic_dec(&sdev->device_busy);
1641 }
1642
1643 static bool scsi_mq_get_budget(struct blk_mq_hw_ctx *hctx)
1644 {
1645         struct request_queue *q = hctx->queue;
1646         struct scsi_device *sdev = q->queuedata;
1647
1648         return scsi_dev_queue_ready(q, sdev);
1649 }
1650
1651 static blk_status_t scsi_queue_rq(struct blk_mq_hw_ctx *hctx,
1652                          const struct blk_mq_queue_data *bd)
1653 {
1654         struct request *req = bd->rq;
1655         struct request_queue *q = req->q;
1656         struct scsi_device *sdev = q->queuedata;
1657         struct Scsi_Host *shost = sdev->host;
1658         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1659         blk_status_t ret;
1660         int reason;
1661
1662         /*
1663          * If the device is not in running state we will reject some or all
1664          * commands.
1665          */
1666         if (unlikely(sdev->sdev_state != SDEV_RUNNING)) {
1667                 ret = scsi_prep_state_check(sdev, req);
1668                 if (ret != BLK_STS_OK)
1669                         goto out_put_budget;
1670         }
1671
1672         ret = BLK_STS_RESOURCE;
1673         if (!scsi_target_queue_ready(shost, sdev))
1674                 goto out_put_budget;
1675         if (!scsi_host_queue_ready(q, shost, sdev))
1676                 goto out_dec_target_busy;
1677
1678         if (!(req->rq_flags & RQF_DONTPREP)) {
1679                 ret = scsi_mq_prep_fn(req);
1680                 if (ret != BLK_STS_OK)
1681                         goto out_dec_host_busy;
1682                 req->rq_flags |= RQF_DONTPREP;
1683         } else {
1684                 clear_bit(SCMD_STATE_COMPLETE, &cmd->state);
1685                 blk_mq_start_request(req);
1686         }
1687
1688         cmd->flags &= SCMD_PRESERVED_FLAGS;
1689         if (sdev->simple_tags)
1690                 cmd->flags |= SCMD_TAGGED;
1691         if (bd->last)
1692                 cmd->flags |= SCMD_LAST;
1693
1694         scsi_init_cmd_errh(cmd);
1695         cmd->scsi_done = scsi_mq_done;
1696
1697         reason = scsi_dispatch_cmd(cmd);
1698         if (reason) {
1699                 scsi_set_blocked(cmd, reason);
1700                 ret = BLK_STS_RESOURCE;
1701                 goto out_dec_host_busy;
1702         }
1703
1704         return BLK_STS_OK;
1705
1706 out_dec_host_busy:
1707         scsi_dec_host_busy(shost);
1708 out_dec_target_busy:
1709         if (scsi_target(sdev)->can_queue > 0)
1710                 atomic_dec(&scsi_target(sdev)->target_busy);
1711 out_put_budget:
1712         scsi_mq_put_budget(hctx);
1713         switch (ret) {
1714         case BLK_STS_OK:
1715                 break;
1716         case BLK_STS_RESOURCE:
1717                 if (scsi_device_blocked(sdev))
1718                         ret = BLK_STS_DEV_RESOURCE;
1719                 break;
1720         default:
1721                 if (unlikely(!scsi_device_online(sdev)))
1722                         scsi_req(req)->result = DID_NO_CONNECT << 16;
1723                 else
1724                         scsi_req(req)->result = DID_ERROR << 16;
1725                 /*
1726                  * Make sure to release all allocated resources when
1727                  * we hit an error, as we will never see this command
1728                  * again.
1729                  */
1730                 if (req->rq_flags & RQF_DONTPREP)
1731                         scsi_mq_uninit_cmd(cmd);
1732                 scsi_run_queue_async(sdev);
1733                 break;
1734         }
1735         return ret;
1736 }
1737
1738 static enum blk_eh_timer_return scsi_timeout(struct request *req,
1739                 bool reserved)
1740 {
1741         if (reserved)
1742                 return BLK_EH_RESET_TIMER;
1743         return scsi_times_out(req);
1744 }
1745
1746 static int scsi_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
1747                                 unsigned int hctx_idx, unsigned int numa_node)
1748 {
1749         struct Scsi_Host *shost = set->driver_data;
1750         const bool unchecked_isa_dma = shost->unchecked_isa_dma;
1751         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1752         struct scatterlist *sg;
1753
1754         if (unchecked_isa_dma)
1755                 cmd->flags |= SCMD_UNCHECKED_ISA_DMA;
1756         cmd->sense_buffer = scsi_alloc_sense_buffer(unchecked_isa_dma,
1757                                                     GFP_KERNEL, numa_node);
1758         if (!cmd->sense_buffer)
1759                 return -ENOMEM;
1760         cmd->req.sense = cmd->sense_buffer;
1761
1762         if (scsi_host_get_prot(shost)) {
1763                 sg = (void *)cmd + sizeof(struct scsi_cmnd) +
1764                         shost->hostt->cmd_size;
1765                 cmd->prot_sdb = (void *)sg + scsi_mq_inline_sgl_size(shost);
1766         }
1767
1768         return 0;
1769 }
1770
1771 static void scsi_mq_exit_request(struct blk_mq_tag_set *set, struct request *rq,
1772                                  unsigned int hctx_idx)
1773 {
1774         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1775
1776         scsi_free_sense_buffer(cmd->flags & SCMD_UNCHECKED_ISA_DMA,
1777                                cmd->sense_buffer);
1778 }
1779
1780 static int scsi_map_queues(struct blk_mq_tag_set *set)
1781 {
1782         struct Scsi_Host *shost = container_of(set, struct Scsi_Host, tag_set);
1783
1784         if (shost->hostt->map_queues)
1785                 return shost->hostt->map_queues(shost);
1786         return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
1787 }
1788
1789 void __scsi_init_queue(struct Scsi_Host *shost, struct request_queue *q)
1790 {
1791         struct device *dev = shost->dma_dev;
1792
1793         /*
1794          * this limit is imposed by hardware restrictions
1795          */
1796         blk_queue_max_segments(q, min_t(unsigned short, shost->sg_tablesize,
1797                                         SG_MAX_SEGMENTS));
1798
1799         if (scsi_host_prot_dma(shost)) {
1800                 shost->sg_prot_tablesize =
1801                         min_not_zero(shost->sg_prot_tablesize,
1802                                      (unsigned short)SCSI_MAX_PROT_SG_SEGMENTS);
1803                 BUG_ON(shost->sg_prot_tablesize < shost->sg_tablesize);
1804                 blk_queue_max_integrity_segments(q, shost->sg_prot_tablesize);
1805         }
1806
1807         if (dev->dma_mask) {
1808                 shost->max_sectors = min_t(unsigned int, shost->max_sectors,
1809                                 dma_max_mapping_size(dev) >> SECTOR_SHIFT);
1810         }
1811         blk_queue_max_hw_sectors(q, shost->max_sectors);
1812         if (shost->unchecked_isa_dma)
1813                 blk_queue_bounce_limit(q, BLK_BOUNCE_ISA);
1814         blk_queue_segment_boundary(q, shost->dma_boundary);
1815         dma_set_seg_boundary(dev, shost->dma_boundary);
1816
1817         blk_queue_max_segment_size(q, shost->max_segment_size);
1818         blk_queue_virt_boundary(q, shost->virt_boundary_mask);
1819         dma_set_max_seg_size(dev, queue_max_segment_size(q));
1820
1821         /*
1822          * Set a reasonable default alignment:  The larger of 32-byte (dword),
1823          * which is a common minimum for HBAs, and the minimum DMA alignment,
1824          * which is set by the platform.
1825          *
1826          * Devices that require a bigger alignment can increase it later.
1827          */
1828         blk_queue_dma_alignment(q, max(4, dma_get_cache_alignment()) - 1);
1829 }
1830 EXPORT_SYMBOL_GPL(__scsi_init_queue);
1831
1832 static const struct blk_mq_ops scsi_mq_ops_no_commit = {
1833         .get_budget     = scsi_mq_get_budget,
1834         .put_budget     = scsi_mq_put_budget,
1835         .queue_rq       = scsi_queue_rq,
1836         .complete       = scsi_softirq_done,
1837         .timeout        = scsi_timeout,
1838 #ifdef CONFIG_BLK_DEBUG_FS
1839         .show_rq        = scsi_show_rq,
1840 #endif
1841         .init_request   = scsi_mq_init_request,
1842         .exit_request   = scsi_mq_exit_request,
1843         .initialize_rq_fn = scsi_initialize_rq,
1844         .cleanup_rq     = scsi_cleanup_rq,
1845         .busy           = scsi_mq_lld_busy,
1846         .map_queues     = scsi_map_queues,
1847 };
1848
1849
1850 static void scsi_commit_rqs(struct blk_mq_hw_ctx *hctx)
1851 {
1852         struct request_queue *q = hctx->queue;
1853         struct scsi_device *sdev = q->queuedata;
1854         struct Scsi_Host *shost = sdev->host;
1855
1856         shost->hostt->commit_rqs(shost, hctx->queue_num);
1857 }
1858
1859 static const struct blk_mq_ops scsi_mq_ops = {
1860         .get_budget     = scsi_mq_get_budget,
1861         .put_budget     = scsi_mq_put_budget,
1862         .queue_rq       = scsi_queue_rq,
1863         .commit_rqs     = scsi_commit_rqs,
1864         .complete       = scsi_softirq_done,
1865         .timeout        = scsi_timeout,
1866 #ifdef CONFIG_BLK_DEBUG_FS
1867         .show_rq        = scsi_show_rq,
1868 #endif
1869         .init_request   = scsi_mq_init_request,
1870         .exit_request   = scsi_mq_exit_request,
1871         .initialize_rq_fn = scsi_initialize_rq,
1872         .cleanup_rq     = scsi_cleanup_rq,
1873         .busy           = scsi_mq_lld_busy,
1874         .map_queues     = scsi_map_queues,
1875 };
1876
1877 struct request_queue *scsi_mq_alloc_queue(struct scsi_device *sdev)
1878 {
1879         sdev->request_queue = blk_mq_init_queue(&sdev->host->tag_set);
1880         if (IS_ERR(sdev->request_queue))
1881                 return NULL;
1882
1883         sdev->request_queue->queuedata = sdev;
1884         __scsi_init_queue(sdev->host, sdev->request_queue);
1885         blk_queue_flag_set(QUEUE_FLAG_SCSI_PASSTHROUGH, sdev->request_queue);
1886         return sdev->request_queue;
1887 }
1888
1889 int scsi_mq_setup_tags(struct Scsi_Host *shost)
1890 {
1891         unsigned int cmd_size, sgl_size;
1892
1893         sgl_size = max_t(unsigned int, sizeof(struct scatterlist),
1894                                 scsi_mq_inline_sgl_size(shost));
1895         cmd_size = sizeof(struct scsi_cmnd) + shost->hostt->cmd_size + sgl_size;
1896         if (scsi_host_get_prot(shost))
1897                 cmd_size += sizeof(struct scsi_data_buffer) +
1898                         sizeof(struct scatterlist) * SCSI_INLINE_PROT_SG_CNT;
1899
1900         memset(&shost->tag_set, 0, sizeof(shost->tag_set));
1901         if (shost->hostt->commit_rqs)
1902                 shost->tag_set.ops = &scsi_mq_ops;
1903         else
1904                 shost->tag_set.ops = &scsi_mq_ops_no_commit;
1905         shost->tag_set.nr_hw_queues = shost->nr_hw_queues ? : 1;
1906         shost->tag_set.queue_depth = shost->can_queue;
1907         shost->tag_set.cmd_size = cmd_size;
1908         shost->tag_set.numa_node = NUMA_NO_NODE;
1909         shost->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
1910         shost->tag_set.flags |=
1911                 BLK_ALLOC_POLICY_TO_MQ_FLAG(shost->hostt->tag_alloc_policy);
1912         shost->tag_set.driver_data = shost;
1913
1914         return blk_mq_alloc_tag_set(&shost->tag_set);
1915 }
1916
1917 void scsi_mq_destroy_tags(struct Scsi_Host *shost)
1918 {
1919         blk_mq_free_tag_set(&shost->tag_set);
1920 }
1921
1922 /**
1923  * scsi_device_from_queue - return sdev associated with a request_queue
1924  * @q: The request queue to return the sdev from
1925  *
1926  * Return the sdev associated with a request queue or NULL if the
1927  * request_queue does not reference a SCSI device.
1928  */
1929 struct scsi_device *scsi_device_from_queue(struct request_queue *q)
1930 {
1931         struct scsi_device *sdev = NULL;
1932
1933         if (q->mq_ops == &scsi_mq_ops_no_commit ||
1934             q->mq_ops == &scsi_mq_ops)
1935                 sdev = q->queuedata;
1936         if (!sdev || !get_device(&sdev->sdev_gendev))
1937                 sdev = NULL;
1938
1939         return sdev;
1940 }
1941 EXPORT_SYMBOL_GPL(scsi_device_from_queue);
1942
1943 /*
1944  * Function:    scsi_block_requests()
1945  *
1946  * Purpose:     Utility function used by low-level drivers to prevent further
1947  *              commands from being queued to the device.
1948  *
1949  * Arguments:   shost       - Host in question
1950  *
1951  * Returns:     Nothing
1952  *
1953  * Lock status: No locks are assumed held.
1954  *
1955  * Notes:       There is no timer nor any other means by which the requests
1956  *              get unblocked other than the low-level driver calling
1957  *              scsi_unblock_requests().
1958  */
1959 void scsi_block_requests(struct Scsi_Host *shost)
1960 {
1961         shost->host_self_blocked = 1;
1962 }
1963 EXPORT_SYMBOL(scsi_block_requests);
1964
1965 /*
1966  * Function:    scsi_unblock_requests()
1967  *
1968  * Purpose:     Utility function used by low-level drivers to allow further
1969  *              commands from being queued to the device.
1970  *
1971  * Arguments:   shost       - Host in question
1972  *
1973  * Returns:     Nothing
1974  *
1975  * Lock status: No locks are assumed held.
1976  *
1977  * Notes:       There is no timer nor any other means by which the requests
1978  *              get unblocked other than the low-level driver calling
1979  *              scsi_unblock_requests().
1980  *
1981  *              This is done as an API function so that changes to the
1982  *              internals of the scsi mid-layer won't require wholesale
1983  *              changes to drivers that use this feature.
1984  */
1985 void scsi_unblock_requests(struct Scsi_Host *shost)
1986 {
1987         shost->host_self_blocked = 0;
1988         scsi_run_host_queues(shost);
1989 }
1990 EXPORT_SYMBOL(scsi_unblock_requests);
1991
1992 int __init scsi_init_queue(void)
1993 {
1994         scsi_sdb_cache = kmem_cache_create("scsi_data_buffer",
1995                                            sizeof(struct scsi_data_buffer),
1996                                            0, 0, NULL);
1997         if (!scsi_sdb_cache) {
1998                 printk(KERN_ERR "SCSI: can't init scsi sdb cache\n");
1999                 return -ENOMEM;
2000         }
2001
2002         return 0;
2003 }
2004
2005 void scsi_exit_queue(void)
2006 {
2007         kmem_cache_destroy(scsi_sense_cache);
2008         kmem_cache_destroy(scsi_sense_isadma_cache);
2009         kmem_cache_destroy(scsi_sdb_cache);
2010 }
2011
2012 /**
2013  *      scsi_mode_select - issue a mode select
2014  *      @sdev:  SCSI device to be queried
2015  *      @pf:    Page format bit (1 == standard, 0 == vendor specific)
2016  *      @sp:    Save page bit (0 == don't save, 1 == save)
2017  *      @modepage: mode page being requested
2018  *      @buffer: request buffer (may not be smaller than eight bytes)
2019  *      @len:   length of request buffer.
2020  *      @timeout: command timeout
2021  *      @retries: number of retries before failing
2022  *      @data: returns a structure abstracting the mode header data
2023  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
2024  *              must be SCSI_SENSE_BUFFERSIZE big.
2025  *
2026  *      Returns zero if successful; negative error number or scsi
2027  *      status on error
2028  *
2029  */
2030 int
2031 scsi_mode_select(struct scsi_device *sdev, int pf, int sp, int modepage,
2032                  unsigned char *buffer, int len, int timeout, int retries,
2033                  struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
2034 {
2035         unsigned char cmd[10];
2036         unsigned char *real_buffer;
2037         int ret;
2038
2039         memset(cmd, 0, sizeof(cmd));
2040         cmd[1] = (pf ? 0x10 : 0) | (sp ? 0x01 : 0);
2041
2042         if (sdev->use_10_for_ms) {
2043                 if (len > 65535)
2044                         return -EINVAL;
2045                 real_buffer = kmalloc(8 + len, GFP_KERNEL);
2046                 if (!real_buffer)
2047                         return -ENOMEM;
2048                 memcpy(real_buffer + 8, buffer, len);
2049                 len += 8;
2050                 real_buffer[0] = 0;
2051                 real_buffer[1] = 0;
2052                 real_buffer[2] = data->medium_type;
2053                 real_buffer[3] = data->device_specific;
2054                 real_buffer[4] = data->longlba ? 0x01 : 0;
2055                 real_buffer[5] = 0;
2056                 real_buffer[6] = data->block_descriptor_length >> 8;
2057                 real_buffer[7] = data->block_descriptor_length;
2058
2059                 cmd[0] = MODE_SELECT_10;
2060                 cmd[7] = len >> 8;
2061                 cmd[8] = len;
2062         } else {
2063                 if (len > 255 || data->block_descriptor_length > 255 ||
2064                     data->longlba)
2065                         return -EINVAL;
2066
2067                 real_buffer = kmalloc(4 + len, GFP_KERNEL);
2068                 if (!real_buffer)
2069                         return -ENOMEM;
2070                 memcpy(real_buffer + 4, buffer, len);
2071                 len += 4;
2072                 real_buffer[0] = 0;
2073                 real_buffer[1] = data->medium_type;
2074                 real_buffer[2] = data->device_specific;
2075                 real_buffer[3] = data->block_descriptor_length;
2076                 
2077
2078                 cmd[0] = MODE_SELECT;
2079                 cmd[4] = len;
2080         }
2081
2082         ret = scsi_execute_req(sdev, cmd, DMA_TO_DEVICE, real_buffer, len,
2083                                sshdr, timeout, retries, NULL);
2084         kfree(real_buffer);
2085         return ret;
2086 }
2087 EXPORT_SYMBOL_GPL(scsi_mode_select);
2088
2089 /**
2090  *      scsi_mode_sense - issue a mode sense, falling back from 10 to six bytes if necessary.
2091  *      @sdev:  SCSI device to be queried
2092  *      @dbd:   set if mode sense will allow block descriptors to be returned
2093  *      @modepage: mode page being requested
2094  *      @buffer: request buffer (may not be smaller than eight bytes)
2095  *      @len:   length of request buffer.
2096  *      @timeout: command timeout
2097  *      @retries: number of retries before failing
2098  *      @data: returns a structure abstracting the mode header data
2099  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
2100  *              must be SCSI_SENSE_BUFFERSIZE big.
2101  *
2102  *      Returns zero if unsuccessful, or the header offset (either 4
2103  *      or 8 depending on whether a six or ten byte command was
2104  *      issued) if successful.
2105  */
2106 int
2107 scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
2108                   unsigned char *buffer, int len, int timeout, int retries,
2109                   struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
2110 {
2111         unsigned char cmd[12];
2112         int use_10_for_ms;
2113         int header_length;
2114         int result, retry_count = retries;
2115         struct scsi_sense_hdr my_sshdr;
2116
2117         memset(data, 0, sizeof(*data));
2118         memset(&cmd[0], 0, 12);
2119         cmd[1] = dbd & 0x18;    /* allows DBD and LLBA bits */
2120         cmd[2] = modepage;
2121
2122         /* caller might not be interested in sense, but we need it */
2123         if (!sshdr)
2124                 sshdr = &my_sshdr;
2125
2126  retry:
2127         use_10_for_ms = sdev->use_10_for_ms;
2128
2129         if (use_10_for_ms) {
2130                 if (len < 8)
2131                         len = 8;
2132
2133                 cmd[0] = MODE_SENSE_10;
2134                 cmd[8] = len;
2135                 header_length = 8;
2136         } else {
2137                 if (len < 4)
2138                         len = 4;
2139
2140                 cmd[0] = MODE_SENSE;
2141                 cmd[4] = len;
2142                 header_length = 4;
2143         }
2144
2145         memset(buffer, 0, len);
2146
2147         result = scsi_execute_req(sdev, cmd, DMA_FROM_DEVICE, buffer, len,
2148                                   sshdr, timeout, retries, NULL);
2149
2150         /* This code looks awful: what it's doing is making sure an
2151          * ILLEGAL REQUEST sense return identifies the actual command
2152          * byte as the problem.  MODE_SENSE commands can return
2153          * ILLEGAL REQUEST if the code page isn't supported */
2154
2155         if (use_10_for_ms && !scsi_status_is_good(result) &&
2156             driver_byte(result) == DRIVER_SENSE) {
2157                 if (scsi_sense_valid(sshdr)) {
2158                         if ((sshdr->sense_key == ILLEGAL_REQUEST) &&
2159                             (sshdr->asc == 0x20) && (sshdr->ascq == 0)) {
2160                                 /* 
2161                                  * Invalid command operation code
2162                                  */
2163                                 sdev->use_10_for_ms = 0;
2164                                 goto retry;
2165                         }
2166                 }
2167         }
2168
2169         if(scsi_status_is_good(result)) {
2170                 if (unlikely(buffer[0] == 0x86 && buffer[1] == 0x0b &&
2171                              (modepage == 6 || modepage == 8))) {
2172                         /* Initio breakage? */
2173                         header_length = 0;
2174                         data->length = 13;
2175                         data->medium_type = 0;
2176                         data->device_specific = 0;
2177                         data->longlba = 0;
2178                         data->block_descriptor_length = 0;
2179                 } else if(use_10_for_ms) {
2180                         data->length = buffer[0]*256 + buffer[1] + 2;
2181                         data->medium_type = buffer[2];
2182                         data->device_specific = buffer[3];
2183                         data->longlba = buffer[4] & 0x01;
2184                         data->block_descriptor_length = buffer[6]*256
2185                                 + buffer[7];
2186                 } else {
2187                         data->length = buffer[0] + 1;
2188                         data->medium_type = buffer[1];
2189                         data->device_specific = buffer[2];
2190                         data->block_descriptor_length = buffer[3];
2191                 }
2192                 data->header_length = header_length;
2193         } else if ((status_byte(result) == CHECK_CONDITION) &&
2194                    scsi_sense_valid(sshdr) &&
2195                    sshdr->sense_key == UNIT_ATTENTION && retry_count) {
2196                 retry_count--;
2197                 goto retry;
2198         }
2199
2200         return result;
2201 }
2202 EXPORT_SYMBOL(scsi_mode_sense);
2203
2204 /**
2205  *      scsi_test_unit_ready - test if unit is ready
2206  *      @sdev:  scsi device to change the state of.
2207  *      @timeout: command timeout
2208  *      @retries: number of retries before failing
2209  *      @sshdr: outpout pointer for decoded sense information.
2210  *
2211  *      Returns zero if unsuccessful or an error if TUR failed.  For
2212  *      removable media, UNIT_ATTENTION sets ->changed flag.
2213  **/
2214 int
2215 scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries,
2216                      struct scsi_sense_hdr *sshdr)
2217 {
2218         char cmd[] = {
2219                 TEST_UNIT_READY, 0, 0, 0, 0, 0,
2220         };
2221         int result;
2222
2223         /* try to eat the UNIT_ATTENTION if there are enough retries */
2224         do {
2225                 result = scsi_execute_req(sdev, cmd, DMA_NONE, NULL, 0, sshdr,
2226                                           timeout, 1, NULL);
2227                 if (sdev->removable && scsi_sense_valid(sshdr) &&
2228                     sshdr->sense_key == UNIT_ATTENTION)
2229                         sdev->changed = 1;
2230         } while (scsi_sense_valid(sshdr) &&
2231                  sshdr->sense_key == UNIT_ATTENTION && --retries);
2232
2233         return result;
2234 }
2235 EXPORT_SYMBOL(scsi_test_unit_ready);
2236
2237 /**
2238  *      scsi_device_set_state - Take the given device through the device state model.
2239  *      @sdev:  scsi device to change the state of.
2240  *      @state: state to change to.
2241  *
2242  *      Returns zero if successful or an error if the requested
2243  *      transition is illegal.
2244  */
2245 int
2246 scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
2247 {
2248         enum scsi_device_state oldstate = sdev->sdev_state;
2249
2250         if (state == oldstate)
2251                 return 0;
2252
2253         switch (state) {
2254         case SDEV_CREATED:
2255                 switch (oldstate) {
2256                 case SDEV_CREATED_BLOCK:
2257                         break;
2258                 default:
2259                         goto illegal;
2260                 }
2261                 break;
2262                         
2263         case SDEV_RUNNING:
2264                 switch (oldstate) {
2265                 case SDEV_CREATED:
2266                 case SDEV_OFFLINE:
2267                 case SDEV_TRANSPORT_OFFLINE:
2268                 case SDEV_QUIESCE:
2269                 case SDEV_BLOCK:
2270                         break;
2271                 default:
2272                         goto illegal;
2273                 }
2274                 break;
2275
2276         case SDEV_QUIESCE:
2277                 switch (oldstate) {
2278                 case SDEV_RUNNING:
2279                 case SDEV_OFFLINE:
2280                 case SDEV_TRANSPORT_OFFLINE:
2281                         break;
2282                 default:
2283                         goto illegal;
2284                 }
2285                 break;
2286
2287         case SDEV_OFFLINE:
2288         case SDEV_TRANSPORT_OFFLINE:
2289                 switch (oldstate) {
2290                 case SDEV_CREATED:
2291                 case SDEV_RUNNING:
2292                 case SDEV_QUIESCE:
2293                 case SDEV_BLOCK:
2294                         break;
2295                 default:
2296                         goto illegal;
2297                 }
2298                 break;
2299
2300         case SDEV_BLOCK:
2301                 switch (oldstate) {
2302                 case SDEV_RUNNING:
2303                 case SDEV_CREATED_BLOCK:
2304                 case SDEV_OFFLINE:
2305                         break;
2306                 default:
2307                         goto illegal;
2308                 }
2309                 break;
2310
2311         case SDEV_CREATED_BLOCK:
2312                 switch (oldstate) {
2313                 case SDEV_CREATED:
2314                         break;
2315                 default:
2316                         goto illegal;
2317                 }
2318                 break;
2319
2320         case SDEV_CANCEL:
2321                 switch (oldstate) {
2322                 case SDEV_CREATED:
2323                 case SDEV_RUNNING:
2324                 case SDEV_QUIESCE:
2325                 case SDEV_OFFLINE:
2326                 case SDEV_TRANSPORT_OFFLINE:
2327                         break;
2328                 default:
2329                         goto illegal;
2330                 }
2331                 break;
2332
2333         case SDEV_DEL:
2334                 switch (oldstate) {
2335                 case SDEV_CREATED:
2336                 case SDEV_RUNNING:
2337                 case SDEV_OFFLINE:
2338                 case SDEV_TRANSPORT_OFFLINE:
2339                 case SDEV_CANCEL:
2340                 case SDEV_BLOCK:
2341                 case SDEV_CREATED_BLOCK:
2342                         break;
2343                 default:
2344                         goto illegal;
2345                 }
2346                 break;
2347
2348         }
2349         sdev->sdev_state = state;
2350         return 0;
2351
2352  illegal:
2353         SCSI_LOG_ERROR_RECOVERY(1,
2354                                 sdev_printk(KERN_ERR, sdev,
2355                                             "Illegal state transition %s->%s",
2356                                             scsi_device_state_name(oldstate),
2357                                             scsi_device_state_name(state))
2358                                 );
2359         return -EINVAL;
2360 }
2361 EXPORT_SYMBOL(scsi_device_set_state);
2362
2363 /**
2364  *      sdev_evt_emit - emit a single SCSI device uevent
2365  *      @sdev: associated SCSI device
2366  *      @evt: event to emit
2367  *
2368  *      Send a single uevent (scsi_event) to the associated scsi_device.
2369  */
2370 static void scsi_evt_emit(struct scsi_device *sdev, struct scsi_event *evt)
2371 {
2372         int idx = 0;
2373         char *envp[3];
2374
2375         switch (evt->evt_type) {
2376         case SDEV_EVT_MEDIA_CHANGE:
2377                 envp[idx++] = "SDEV_MEDIA_CHANGE=1";
2378                 break;
2379         case SDEV_EVT_INQUIRY_CHANGE_REPORTED:
2380                 scsi_rescan_device(&sdev->sdev_gendev);
2381                 envp[idx++] = "SDEV_UA=INQUIRY_DATA_HAS_CHANGED";
2382                 break;
2383         case SDEV_EVT_CAPACITY_CHANGE_REPORTED:
2384                 envp[idx++] = "SDEV_UA=CAPACITY_DATA_HAS_CHANGED";
2385                 break;
2386         case SDEV_EVT_SOFT_THRESHOLD_REACHED_REPORTED:
2387                envp[idx++] = "SDEV_UA=THIN_PROVISIONING_SOFT_THRESHOLD_REACHED";
2388                 break;
2389         case SDEV_EVT_MODE_PARAMETER_CHANGE_REPORTED:
2390                 envp[idx++] = "SDEV_UA=MODE_PARAMETERS_CHANGED";
2391                 break;
2392         case SDEV_EVT_LUN_CHANGE_REPORTED:
2393                 envp[idx++] = "SDEV_UA=REPORTED_LUNS_DATA_HAS_CHANGED";
2394                 break;
2395         case SDEV_EVT_ALUA_STATE_CHANGE_REPORTED:
2396                 envp[idx++] = "SDEV_UA=ASYMMETRIC_ACCESS_STATE_CHANGED";
2397                 break;
2398         case SDEV_EVT_POWER_ON_RESET_OCCURRED:
2399                 envp[idx++] = "SDEV_UA=POWER_ON_RESET_OCCURRED";
2400                 break;
2401         default:
2402                 /* do nothing */
2403                 break;
2404         }
2405
2406         envp[idx++] = NULL;
2407
2408         kobject_uevent_env(&sdev->sdev_gendev.kobj, KOBJ_CHANGE, envp);
2409 }
2410
2411 /**
2412  *      sdev_evt_thread - send a uevent for each scsi event
2413  *      @work: work struct for scsi_device
2414  *
2415  *      Dispatch queued events to their associated scsi_device kobjects
2416  *      as uevents.
2417  */
2418 void scsi_evt_thread(struct work_struct *work)
2419 {
2420         struct scsi_device *sdev;
2421         enum scsi_device_event evt_type;
2422         LIST_HEAD(event_list);
2423
2424         sdev = container_of(work, struct scsi_device, event_work);
2425
2426         for (evt_type = SDEV_EVT_FIRST; evt_type <= SDEV_EVT_LAST; evt_type++)
2427                 if (test_and_clear_bit(evt_type, sdev->pending_events))
2428                         sdev_evt_send_simple(sdev, evt_type, GFP_KERNEL);
2429
2430         while (1) {
2431                 struct scsi_event *evt;
2432                 struct list_head *this, *tmp;
2433                 unsigned long flags;
2434
2435                 spin_lock_irqsave(&sdev->list_lock, flags);
2436                 list_splice_init(&sdev->event_list, &event_list);
2437                 spin_unlock_irqrestore(&sdev->list_lock, flags);
2438
2439                 if (list_empty(&event_list))
2440                         break;
2441
2442                 list_for_each_safe(this, tmp, &event_list) {
2443                         evt = list_entry(this, struct scsi_event, node);
2444                         list_del(&evt->node);
2445                         scsi_evt_emit(sdev, evt);
2446                         kfree(evt);
2447                 }
2448         }
2449 }
2450
2451 /**
2452  *      sdev_evt_send - send asserted event to uevent thread
2453  *      @sdev: scsi_device event occurred on
2454  *      @evt: event to send
2455  *
2456  *      Assert scsi device event asynchronously.
2457  */
2458 void sdev_evt_send(struct scsi_device *sdev, struct scsi_event *evt)
2459 {
2460         unsigned long flags;
2461
2462 #if 0
2463         /* FIXME: currently this check eliminates all media change events
2464          * for polled devices.  Need to update to discriminate between AN
2465          * and polled events */
2466         if (!test_bit(evt->evt_type, sdev->supported_events)) {
2467                 kfree(evt);
2468                 return;
2469         }
2470 #endif
2471
2472         spin_lock_irqsave(&sdev->list_lock, flags);
2473         list_add_tail(&evt->node, &sdev->event_list);
2474         schedule_work(&sdev->event_work);
2475         spin_unlock_irqrestore(&sdev->list_lock, flags);
2476 }
2477 EXPORT_SYMBOL_GPL(sdev_evt_send);
2478
2479 /**
2480  *      sdev_evt_alloc - allocate a new scsi event
2481  *      @evt_type: type of event to allocate
2482  *      @gfpflags: GFP flags for allocation
2483  *
2484  *      Allocates and returns a new scsi_event.
2485  */
2486 struct scsi_event *sdev_evt_alloc(enum scsi_device_event evt_type,
2487                                   gfp_t gfpflags)
2488 {
2489         struct scsi_event *evt = kzalloc(sizeof(struct scsi_event), gfpflags);
2490         if (!evt)
2491                 return NULL;
2492
2493         evt->evt_type = evt_type;
2494         INIT_LIST_HEAD(&evt->node);
2495
2496         /* evt_type-specific initialization, if any */
2497         switch (evt_type) {
2498         case SDEV_EVT_MEDIA_CHANGE:
2499         case SDEV_EVT_INQUIRY_CHANGE_REPORTED:
2500         case SDEV_EVT_CAPACITY_CHANGE_REPORTED:
2501         case SDEV_EVT_SOFT_THRESHOLD_REACHED_REPORTED:
2502         case SDEV_EVT_MODE_PARAMETER_CHANGE_REPORTED:
2503         case SDEV_EVT_LUN_CHANGE_REPORTED:
2504         case SDEV_EVT_ALUA_STATE_CHANGE_REPORTED:
2505         case SDEV_EVT_POWER_ON_RESET_OCCURRED:
2506         default:
2507                 /* do nothing */
2508                 break;
2509         }
2510
2511         return evt;
2512 }
2513 EXPORT_SYMBOL_GPL(sdev_evt_alloc);
2514
2515 /**
2516  *      sdev_evt_send_simple - send asserted event to uevent thread
2517  *      @sdev: scsi_device event occurred on
2518  *      @evt_type: type of event to send
2519  *      @gfpflags: GFP flags for allocation
2520  *
2521  *      Assert scsi device event asynchronously, given an event type.
2522  */
2523 void sdev_evt_send_simple(struct scsi_device *sdev,
2524                           enum scsi_device_event evt_type, gfp_t gfpflags)
2525 {
2526         struct scsi_event *evt = sdev_evt_alloc(evt_type, gfpflags);
2527         if (!evt) {
2528                 sdev_printk(KERN_ERR, sdev, "event %d eaten due to OOM\n",
2529                             evt_type);
2530                 return;
2531         }
2532
2533         sdev_evt_send(sdev, evt);
2534 }
2535 EXPORT_SYMBOL_GPL(sdev_evt_send_simple);
2536
2537 /**
2538  *      scsi_device_quiesce - Block user issued commands.
2539  *      @sdev:  scsi device to quiesce.
2540  *
2541  *      This works by trying to transition to the SDEV_QUIESCE state
2542  *      (which must be a legal transition).  When the device is in this
2543  *      state, only special requests will be accepted, all others will
2544  *      be deferred.  Since special requests may also be requeued requests,
2545  *      a successful return doesn't guarantee the device will be 
2546  *      totally quiescent.
2547  *
2548  *      Must be called with user context, may sleep.
2549  *
2550  *      Returns zero if unsuccessful or an error if not.
2551  */
2552 int
2553 scsi_device_quiesce(struct scsi_device *sdev)
2554 {
2555         struct request_queue *q = sdev->request_queue;
2556         int err;
2557
2558         /*
2559          * It is allowed to call scsi_device_quiesce() multiple times from
2560          * the same context but concurrent scsi_device_quiesce() calls are
2561          * not allowed.
2562          */
2563         WARN_ON_ONCE(sdev->quiesced_by && sdev->quiesced_by != current);
2564
2565         if (sdev->quiesced_by == current)
2566                 return 0;
2567
2568         blk_set_pm_only(q);
2569
2570         blk_mq_freeze_queue(q);
2571         /*
2572          * Ensure that the effect of blk_set_pm_only() will be visible
2573          * for percpu_ref_tryget() callers that occur after the queue
2574          * unfreeze even if the queue was already frozen before this function
2575          * was called. See also https://lwn.net/Articles/573497/.
2576          */
2577         synchronize_rcu();
2578         blk_mq_unfreeze_queue(q);
2579
2580         mutex_lock(&sdev->state_mutex);
2581         err = scsi_device_set_state(sdev, SDEV_QUIESCE);
2582         if (err == 0)
2583                 sdev->quiesced_by = current;
2584         else
2585                 blk_clear_pm_only(q);
2586         mutex_unlock(&sdev->state_mutex);
2587
2588         return err;
2589 }
2590 EXPORT_SYMBOL(scsi_device_quiesce);
2591
2592 /**
2593  *      scsi_device_resume - Restart user issued commands to a quiesced device.
2594  *      @sdev:  scsi device to resume.
2595  *
2596  *      Moves the device from quiesced back to running and restarts the
2597  *      queues.
2598  *
2599  *      Must be called with user context, may sleep.
2600  */
2601 void scsi_device_resume(struct scsi_device *sdev)
2602 {
2603         /* check if the device state was mutated prior to resume, and if
2604          * so assume the state is being managed elsewhere (for example
2605          * device deleted during suspend)
2606          */
2607         mutex_lock(&sdev->state_mutex);
2608         if (sdev->quiesced_by) {
2609                 sdev->quiesced_by = NULL;
2610                 blk_clear_pm_only(sdev->request_queue);
2611         }
2612         if (sdev->sdev_state == SDEV_QUIESCE)
2613                 scsi_device_set_state(sdev, SDEV_RUNNING);
2614         mutex_unlock(&sdev->state_mutex);
2615 }
2616 EXPORT_SYMBOL(scsi_device_resume);
2617
2618 static void
2619 device_quiesce_fn(struct scsi_device *sdev, void *data)
2620 {
2621         scsi_device_quiesce(sdev);
2622 }
2623
2624 void
2625 scsi_target_quiesce(struct scsi_target *starget)
2626 {
2627         starget_for_each_device(starget, NULL, device_quiesce_fn);
2628 }
2629 EXPORT_SYMBOL(scsi_target_quiesce);
2630
2631 static void
2632 device_resume_fn(struct scsi_device *sdev, void *data)
2633 {
2634         scsi_device_resume(sdev);
2635 }
2636
2637 void
2638 scsi_target_resume(struct scsi_target *starget)
2639 {
2640         starget_for_each_device(starget, NULL, device_resume_fn);
2641 }
2642 EXPORT_SYMBOL(scsi_target_resume);
2643
2644 /**
2645  * scsi_internal_device_block_nowait - try to transition to the SDEV_BLOCK state
2646  * @sdev: device to block
2647  *
2648  * Pause SCSI command processing on the specified device. Does not sleep.
2649  *
2650  * Returns zero if successful or a negative error code upon failure.
2651  *
2652  * Notes:
2653  * This routine transitions the device to the SDEV_BLOCK state (which must be
2654  * a legal transition). When the device is in this state, command processing
2655  * is paused until the device leaves the SDEV_BLOCK state. See also
2656  * scsi_internal_device_unblock_nowait().
2657  */
2658 int scsi_internal_device_block_nowait(struct scsi_device *sdev)
2659 {
2660         struct request_queue *q = sdev->request_queue;
2661         int err = 0;
2662
2663         err = scsi_device_set_state(sdev, SDEV_BLOCK);
2664         if (err) {
2665                 err = scsi_device_set_state(sdev, SDEV_CREATED_BLOCK);
2666
2667                 if (err)
2668                         return err;
2669         }
2670
2671         /* 
2672          * The device has transitioned to SDEV_BLOCK.  Stop the
2673          * block layer from calling the midlayer with this device's
2674          * request queue. 
2675          */
2676         blk_mq_quiesce_queue_nowait(q);
2677         return 0;
2678 }
2679 EXPORT_SYMBOL_GPL(scsi_internal_device_block_nowait);
2680
2681 /**
2682  * scsi_internal_device_block - try to transition to the SDEV_BLOCK state
2683  * @sdev: device to block
2684  *
2685  * Pause SCSI command processing on the specified device and wait until all
2686  * ongoing scsi_request_fn() / scsi_queue_rq() calls have finished. May sleep.
2687  *
2688  * Returns zero if successful or a negative error code upon failure.
2689  *
2690  * Note:
2691  * This routine transitions the device to the SDEV_BLOCK state (which must be
2692  * a legal transition). When the device is in this state, command processing
2693  * is paused until the device leaves the SDEV_BLOCK state. See also
2694  * scsi_internal_device_unblock().
2695  */
2696 static int scsi_internal_device_block(struct scsi_device *sdev)
2697 {
2698         struct request_queue *q = sdev->request_queue;
2699         int err;
2700
2701         mutex_lock(&sdev->state_mutex);
2702         err = scsi_internal_device_block_nowait(sdev);
2703         if (err == 0)
2704                 blk_mq_quiesce_queue(q);
2705         mutex_unlock(&sdev->state_mutex);
2706
2707         return err;
2708 }
2709  
2710 void scsi_start_queue(struct scsi_device *sdev)
2711 {
2712         struct request_queue *q = sdev->request_queue;
2713
2714         blk_mq_unquiesce_queue(q);
2715 }
2716
2717 /**
2718  * scsi_internal_device_unblock_nowait - resume a device after a block request
2719  * @sdev:       device to resume
2720  * @new_state:  state to set the device to after unblocking
2721  *
2722  * Restart the device queue for a previously suspended SCSI device. Does not
2723  * sleep.
2724  *
2725  * Returns zero if successful or a negative error code upon failure.
2726  *
2727  * Notes:
2728  * This routine transitions the device to the SDEV_RUNNING state or to one of
2729  * the offline states (which must be a legal transition) allowing the midlayer
2730  * to goose the queue for this device.
2731  */
2732 int scsi_internal_device_unblock_nowait(struct scsi_device *sdev,
2733                                         enum scsi_device_state new_state)
2734 {
2735         switch (new_state) {
2736         case SDEV_RUNNING:
2737         case SDEV_TRANSPORT_OFFLINE:
2738                 break;
2739         default:
2740                 return -EINVAL;
2741         }
2742
2743         /*
2744          * Try to transition the scsi device to SDEV_RUNNING or one of the
2745          * offlined states and goose the device queue if successful.
2746          */
2747         switch (sdev->sdev_state) {
2748         case SDEV_BLOCK:
2749         case SDEV_TRANSPORT_OFFLINE:
2750                 sdev->sdev_state = new_state;
2751                 break;
2752         case SDEV_CREATED_BLOCK:
2753                 if (new_state == SDEV_TRANSPORT_OFFLINE ||
2754                     new_state == SDEV_OFFLINE)
2755                         sdev->sdev_state = new_state;
2756                 else
2757                         sdev->sdev_state = SDEV_CREATED;
2758                 break;
2759         case SDEV_CANCEL:
2760         case SDEV_OFFLINE:
2761                 break;
2762         default:
2763                 return -EINVAL;
2764         }
2765         scsi_start_queue(sdev);
2766
2767         return 0;
2768 }
2769 EXPORT_SYMBOL_GPL(scsi_internal_device_unblock_nowait);
2770
2771 /**
2772  * scsi_internal_device_unblock - resume a device after a block request
2773  * @sdev:       device to resume
2774  * @new_state:  state to set the device to after unblocking
2775  *
2776  * Restart the device queue for a previously suspended SCSI device. May sleep.
2777  *
2778  * Returns zero if successful or a negative error code upon failure.
2779  *
2780  * Notes:
2781  * This routine transitions the device to the SDEV_RUNNING state or to one of
2782  * the offline states (which must be a legal transition) allowing the midlayer
2783  * to goose the queue for this device.
2784  */
2785 static int scsi_internal_device_unblock(struct scsi_device *sdev,
2786                                         enum scsi_device_state new_state)
2787 {
2788         int ret;
2789
2790         mutex_lock(&sdev->state_mutex);
2791         ret = scsi_internal_device_unblock_nowait(sdev, new_state);
2792         mutex_unlock(&sdev->state_mutex);
2793
2794         return ret;
2795 }
2796
2797 static void
2798 device_block(struct scsi_device *sdev, void *data)
2799 {
2800         int ret;
2801
2802         ret = scsi_internal_device_block(sdev);
2803
2804         WARN_ONCE(ret, "scsi_internal_device_block(%s) failed: ret = %d\n",
2805                   dev_name(&sdev->sdev_gendev), ret);
2806 }
2807
2808 static int
2809 target_block(struct device *dev, void *data)
2810 {
2811         if (scsi_is_target_device(dev))
2812                 starget_for_each_device(to_scsi_target(dev), NULL,
2813                                         device_block);
2814         return 0;
2815 }
2816
2817 void
2818 scsi_target_block(struct device *dev)
2819 {
2820         if (scsi_is_target_device(dev))
2821                 starget_for_each_device(to_scsi_target(dev), NULL,
2822                                         device_block);
2823         else
2824                 device_for_each_child(dev, NULL, target_block);
2825 }
2826 EXPORT_SYMBOL_GPL(scsi_target_block);
2827
2828 static void
2829 device_unblock(struct scsi_device *sdev, void *data)
2830 {
2831         scsi_internal_device_unblock(sdev, *(enum scsi_device_state *)data);
2832 }
2833
2834 static int
2835 target_unblock(struct device *dev, void *data)
2836 {
2837         if (scsi_is_target_device(dev))
2838                 starget_for_each_device(to_scsi_target(dev), data,
2839                                         device_unblock);
2840         return 0;
2841 }
2842
2843 void
2844 scsi_target_unblock(struct device *dev, enum scsi_device_state new_state)
2845 {
2846         if (scsi_is_target_device(dev))
2847                 starget_for_each_device(to_scsi_target(dev), &new_state,
2848                                         device_unblock);
2849         else
2850                 device_for_each_child(dev, &new_state, target_unblock);
2851 }
2852 EXPORT_SYMBOL_GPL(scsi_target_unblock);
2853
2854 /**
2855  * scsi_kmap_atomic_sg - find and atomically map an sg-elemnt
2856  * @sgl:        scatter-gather list
2857  * @sg_count:   number of segments in sg
2858  * @offset:     offset in bytes into sg, on return offset into the mapped area
2859  * @len:        bytes to map, on return number of bytes mapped
2860  *
2861  * Returns virtual address of the start of the mapped page
2862  */
2863 void *scsi_kmap_atomic_sg(struct scatterlist *sgl, int sg_count,
2864                           size_t *offset, size_t *len)
2865 {
2866         int i;
2867         size_t sg_len = 0, len_complete = 0;
2868         struct scatterlist *sg;
2869         struct page *page;
2870
2871         WARN_ON(!irqs_disabled());
2872
2873         for_each_sg(sgl, sg, sg_count, i) {
2874                 len_complete = sg_len; /* Complete sg-entries */
2875                 sg_len += sg->length;
2876                 if (sg_len > *offset)
2877                         break;
2878         }
2879
2880         if (unlikely(i == sg_count)) {
2881                 printk(KERN_ERR "%s: Bytes in sg: %zu, requested offset %zu, "
2882                         "elements %d\n",
2883                        __func__, sg_len, *offset, sg_count);
2884                 WARN_ON(1);
2885                 return NULL;
2886         }
2887
2888         /* Offset starting from the beginning of first page in this sg-entry */
2889         *offset = *offset - len_complete + sg->offset;
2890
2891         /* Assumption: contiguous pages can be accessed as "page + i" */
2892         page = nth_page(sg_page(sg), (*offset >> PAGE_SHIFT));
2893         *offset &= ~PAGE_MASK;
2894
2895         /* Bytes in this sg-entry from *offset to the end of the page */
2896         sg_len = PAGE_SIZE - *offset;
2897         if (*len > sg_len)
2898                 *len = sg_len;
2899
2900         return kmap_atomic(page);
2901 }
2902 EXPORT_SYMBOL(scsi_kmap_atomic_sg);
2903
2904 /**
2905  * scsi_kunmap_atomic_sg - atomically unmap a virtual address, previously mapped with scsi_kmap_atomic_sg
2906  * @virt:       virtual address to be unmapped
2907  */
2908 void scsi_kunmap_atomic_sg(void *virt)
2909 {
2910         kunmap_atomic(virt);
2911 }
2912 EXPORT_SYMBOL(scsi_kunmap_atomic_sg);
2913
2914 void sdev_disable_disk_events(struct scsi_device *sdev)
2915 {
2916         atomic_inc(&sdev->disk_events_disable_depth);
2917 }
2918 EXPORT_SYMBOL(sdev_disable_disk_events);
2919
2920 void sdev_enable_disk_events(struct scsi_device *sdev)
2921 {
2922         if (WARN_ON_ONCE(atomic_read(&sdev->disk_events_disable_depth) <= 0))
2923                 return;
2924         atomic_dec(&sdev->disk_events_disable_depth);
2925 }
2926 EXPORT_SYMBOL(sdev_enable_disk_events);
2927
2928 static unsigned char designator_prio(const unsigned char *d)
2929 {
2930         if (d[1] & 0x30)
2931                 /* not associated with LUN */
2932                 return 0;
2933
2934         if (d[3] == 0)
2935                 /* invalid length */
2936                 return 0;
2937
2938         /*
2939          * Order of preference for lun descriptor:
2940          * - SCSI name string
2941          * - NAA IEEE Registered Extended
2942          * - EUI-64 based 16-byte
2943          * - EUI-64 based 12-byte
2944          * - NAA IEEE Registered
2945          * - NAA IEEE Extended
2946          * - EUI-64 based 8-byte
2947          * - SCSI name string (truncated)
2948          * - T10 Vendor ID
2949          * as longer descriptors reduce the likelyhood
2950          * of identification clashes.
2951          */
2952
2953         switch (d[1] & 0xf) {
2954         case 8:
2955                 /* SCSI name string, variable-length UTF-8 */
2956                 return 9;
2957         case 3:
2958                 switch (d[4] >> 4) {
2959                 case 6:
2960                         /* NAA registered extended */
2961                         return 8;
2962                 case 5:
2963                         /* NAA registered */
2964                         return 5;
2965                 case 4:
2966                         /* NAA extended */
2967                         return 4;
2968                 case 3:
2969                         /* NAA locally assigned */
2970                         return 1;
2971                 default:
2972                         break;
2973                 }
2974                 break;
2975         case 2:
2976                 switch (d[3]) {
2977                 case 16:
2978                         /* EUI64-based, 16 byte */
2979                         return 7;
2980                 case 12:
2981                         /* EUI64-based, 12 byte */
2982                         return 6;
2983                 case 8:
2984                         /* EUI64-based, 8 byte */
2985                         return 3;
2986                 default:
2987                         break;
2988                 }
2989                 break;
2990         case 1:
2991                 /* T10 vendor ID */
2992                 return 1;
2993         default:
2994                 break;
2995         }
2996
2997         return 0;
2998 }
2999
3000 /**
3001  * scsi_vpd_lun_id - return a unique device identification
3002  * @sdev: SCSI device
3003  * @id:   buffer for the identification
3004  * @id_len:  length of the buffer
3005  *
3006  * Copies a unique device identification into @id based
3007  * on the information in the VPD page 0x83 of the device.
3008  * The string will be formatted as a SCSI name string.
3009  *
3010  * Returns the length of the identification or error on failure.
3011  * If the identifier is longer than the supplied buffer the actual
3012  * identifier length is returned and the buffer is not zero-padded.
3013  */
3014 int scsi_vpd_lun_id(struct scsi_device *sdev, char *id, size_t id_len)
3015 {
3016         u8 cur_id_prio = 0;
3017         u8 cur_id_size = 0;
3018         const unsigned char *d, *cur_id_str;
3019         const struct scsi_vpd *vpd_pg83;
3020         int id_size = -EINVAL;
3021
3022         rcu_read_lock();
3023         vpd_pg83 = rcu_dereference(sdev->vpd_pg83);
3024         if (!vpd_pg83) {
3025                 rcu_read_unlock();
3026                 return -ENXIO;
3027         }
3028
3029         /* The id string must be at least 20 bytes + terminating NULL byte */
3030         if (id_len < 21) {
3031                 rcu_read_unlock();
3032                 return -EINVAL;
3033         }
3034
3035         memset(id, 0, id_len);
3036         d = vpd_pg83->data + 4;
3037         while (d < vpd_pg83->data + vpd_pg83->len) {
3038                 u8 prio = designator_prio(d);
3039
3040                 if (prio == 0 || cur_id_prio > prio)
3041                         goto next_desig;
3042
3043                 switch (d[1] & 0xf) {
3044                 case 0x1:
3045                         /* T10 Vendor ID */
3046                         if (cur_id_size > d[3])
3047                                 break;
3048                         cur_id_prio = prio;
3049                         cur_id_size = d[3];
3050                         if (cur_id_size + 4 > id_len)
3051                                 cur_id_size = id_len - 4;
3052                         cur_id_str = d + 4;
3053                         id_size = snprintf(id, id_len, "t10.%*pE",
3054                                            cur_id_size, cur_id_str);
3055                         break;
3056                 case 0x2:
3057                         /* EUI-64 */
3058                         cur_id_prio = prio;
3059                         cur_id_size = d[3];
3060                         cur_id_str = d + 4;
3061                         switch (cur_id_size) {
3062                         case 8:
3063                                 id_size = snprintf(id, id_len,
3064                                                    "eui.%8phN",
3065                                                    cur_id_str);
3066                                 break;
3067                         case 12:
3068                                 id_size = snprintf(id, id_len,
3069                                                    "eui.%12phN",
3070                                                    cur_id_str);
3071                                 break;
3072                         case 16:
3073                                 id_size = snprintf(id, id_len,
3074                                                    "eui.%16phN",
3075                                                    cur_id_str);
3076                                 break;
3077                         default:
3078                                 break;
3079                         }
3080                         break;
3081                 case 0x3:
3082                         /* NAA */
3083                         cur_id_prio = prio;
3084                         cur_id_size = d[3];
3085                         cur_id_str = d + 4;
3086                         switch (cur_id_size) {
3087                         case 8:
3088                                 id_size = snprintf(id, id_len,
3089                                                    "naa.%8phN",
3090                                                    cur_id_str);
3091                                 break;
3092                         case 16:
3093                                 id_size = snprintf(id, id_len,
3094                                                    "naa.%16phN",
3095                                                    cur_id_str);
3096                                 break;
3097                         default:
3098                                 break;
3099                         }
3100                         break;
3101                 case 0x8:
3102                         /* SCSI name string */
3103                         if (cur_id_size > d[3])
3104                                 break;
3105                         /* Prefer others for truncated descriptor */
3106                         if (d[3] > id_len) {
3107                                 prio = 2;
3108                                 if (cur_id_prio > prio)
3109                                         break;
3110                         }
3111                         cur_id_prio = prio;
3112                         cur_id_size = id_size = d[3];
3113                         cur_id_str = d + 4;
3114                         if (cur_id_size >= id_len)
3115                                 cur_id_size = id_len - 1;
3116                         memcpy(id, cur_id_str, cur_id_size);
3117                         break;
3118                 default:
3119                         break;
3120                 }
3121 next_desig:
3122                 d += d[3] + 4;
3123         }
3124         rcu_read_unlock();
3125
3126         return id_size;
3127 }
3128 EXPORT_SYMBOL(scsi_vpd_lun_id);
3129
3130 /*
3131  * scsi_vpd_tpg_id - return a target port group identifier
3132  * @sdev: SCSI device
3133  *
3134  * Returns the Target Port Group identifier from the information
3135  * froom VPD page 0x83 of the device.
3136  *
3137  * Returns the identifier or error on failure.
3138  */
3139 int scsi_vpd_tpg_id(struct scsi_device *sdev, int *rel_id)
3140 {
3141         const unsigned char *d;
3142         const struct scsi_vpd *vpd_pg83;
3143         int group_id = -EAGAIN, rel_port = -1;
3144
3145         rcu_read_lock();
3146         vpd_pg83 = rcu_dereference(sdev->vpd_pg83);
3147         if (!vpd_pg83) {
3148                 rcu_read_unlock();
3149                 return -ENXIO;
3150         }
3151
3152         d = vpd_pg83->data + 4;
3153         while (d < vpd_pg83->data + vpd_pg83->len) {
3154                 switch (d[1] & 0xf) {
3155                 case 0x4:
3156                         /* Relative target port */
3157                         rel_port = get_unaligned_be16(&d[6]);
3158                         break;
3159                 case 0x5:
3160                         /* Target port group */
3161                         group_id = get_unaligned_be16(&d[6]);
3162                         break;
3163                 default:
3164                         break;
3165                 }
3166                 d += d[3] + 4;
3167         }
3168         rcu_read_unlock();
3169
3170         if (group_id >= 0 && rel_id && rel_port != -1)
3171                 *rel_id = rel_port;
3172
3173         return group_id;
3174 }
3175 EXPORT_SYMBOL(scsi_vpd_tpg_id);