]> Git Repo - linux.git/blob - drivers/scsi/scsi_lib.c
tools/bpftool: Add ringbuf map to a list of known map types
[linux.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 ([email protected]).
8  *                        Based upon conversations with large numbers
9  *                        of people at Linux Expo.
10  */
11
12 #include <linux/bio.h>
13 #include <linux/bitops.h>
14 #include <linux/blkdev.h>
15 #include <linux/completion.h>
16 #include <linux/kernel.h>
17 #include <linux/export.h>
18 #include <linux/init.h>
19 #include <linux/pci.h>
20 #include <linux/delay.h>
21 #include <linux/hardirq.h>
22 #include <linux/scatterlist.h>
23 #include <linux/blk-mq.h>
24 #include <linux/ratelimit.h>
25 #include <asm/unaligned.h>
26
27 #include <scsi/scsi.h>
28 #include <scsi/scsi_cmnd.h>
29 #include <scsi/scsi_dbg.h>
30 #include <scsi/scsi_device.h>
31 #include <scsi/scsi_driver.h>
32 #include <scsi/scsi_eh.h>
33 #include <scsi/scsi_host.h>
34 #include <scsi/scsi_transport.h> /* __scsi_init_queue() */
35 #include <scsi/scsi_dh.h>
36
37 #include <trace/events/scsi.h>
38
39 #include "scsi_debugfs.h"
40 #include "scsi_priv.h"
41 #include "scsi_logging.h"
42
43 /*
44  * Size of integrity metadata is usually small, 1 inline sg should
45  * cover normal cases.
46  */
47 #ifdef CONFIG_ARCH_NO_SG_CHAIN
48 #define  SCSI_INLINE_PROT_SG_CNT  0
49 #define  SCSI_INLINE_SG_CNT  0
50 #else
51 #define  SCSI_INLINE_PROT_SG_CNT  1
52 #define  SCSI_INLINE_SG_CNT  2
53 #endif
54
55 static struct kmem_cache *scsi_sdb_cache;
56 static struct kmem_cache *scsi_sense_cache;
57 static struct kmem_cache *scsi_sense_isadma_cache;
58 static DEFINE_MUTEX(scsi_sense_cache_mutex);
59
60 static void scsi_mq_uninit_cmd(struct scsi_cmnd *cmd);
61
62 static inline struct kmem_cache *
63 scsi_select_sense_cache(bool unchecked_isa_dma)
64 {
65         return unchecked_isa_dma ? scsi_sense_isadma_cache : scsi_sense_cache;
66 }
67
68 static void scsi_free_sense_buffer(bool unchecked_isa_dma,
69                                    unsigned char *sense_buffer)
70 {
71         kmem_cache_free(scsi_select_sense_cache(unchecked_isa_dma),
72                         sense_buffer);
73 }
74
75 static unsigned char *scsi_alloc_sense_buffer(bool unchecked_isa_dma,
76         gfp_t gfp_mask, int numa_node)
77 {
78         return kmem_cache_alloc_node(scsi_select_sense_cache(unchecked_isa_dma),
79                                      gfp_mask, numa_node);
80 }
81
82 int scsi_init_sense_cache(struct Scsi_Host *shost)
83 {
84         struct kmem_cache *cache;
85         int ret = 0;
86
87         mutex_lock(&scsi_sense_cache_mutex);
88         cache = scsi_select_sense_cache(shost->unchecked_isa_dma);
89         if (cache)
90                 goto exit;
91
92         if (shost->unchecked_isa_dma) {
93                 scsi_sense_isadma_cache =
94                         kmem_cache_create("scsi_sense_cache(DMA)",
95                                 SCSI_SENSE_BUFFERSIZE, 0,
96                                 SLAB_HWCACHE_ALIGN | SLAB_CACHE_DMA, NULL);
97                 if (!scsi_sense_isadma_cache)
98                         ret = -ENOMEM;
99         } else {
100                 scsi_sense_cache =
101                         kmem_cache_create_usercopy("scsi_sense_cache",
102                                 SCSI_SENSE_BUFFERSIZE, 0, SLAB_HWCACHE_ALIGN,
103                                 0, SCSI_SENSE_BUFFERSIZE, NULL);
104                 if (!scsi_sense_cache)
105                         ret = -ENOMEM;
106         }
107  exit:
108         mutex_unlock(&scsi_sense_cache_mutex);
109         return ret;
110 }
111
112 /*
113  * When to reinvoke queueing after a resource shortage. It's 3 msecs to
114  * not change behaviour from the previous unplug mechanism, experimentation
115  * may prove this needs changing.
116  */
117 #define SCSI_QUEUE_DELAY        3
118
119 static void
120 scsi_set_blocked(struct scsi_cmnd *cmd, int reason)
121 {
122         struct Scsi_Host *host = cmd->device->host;
123         struct scsi_device *device = cmd->device;
124         struct scsi_target *starget = scsi_target(device);
125
126         /*
127          * Set the appropriate busy bit for the device/host.
128          *
129          * If the host/device isn't busy, assume that something actually
130          * completed, and that we should be able to queue a command now.
131          *
132          * Note that the prior mid-layer assumption that any host could
133          * always queue at least one command is now broken.  The mid-layer
134          * will implement a user specifiable stall (see
135          * scsi_host.max_host_blocked and scsi_device.max_device_blocked)
136          * if a command is requeued with no other commands outstanding
137          * either for the device or for the host.
138          */
139         switch (reason) {
140         case SCSI_MLQUEUE_HOST_BUSY:
141                 atomic_set(&host->host_blocked, host->max_host_blocked);
142                 break;
143         case SCSI_MLQUEUE_DEVICE_BUSY:
144         case SCSI_MLQUEUE_EH_RETRY:
145                 atomic_set(&device->device_blocked,
146                            device->max_device_blocked);
147                 break;
148         case SCSI_MLQUEUE_TARGET_BUSY:
149                 atomic_set(&starget->target_blocked,
150                            starget->max_target_blocked);
151                 break;
152         }
153 }
154
155 static void scsi_mq_requeue_cmd(struct scsi_cmnd *cmd)
156 {
157         if (cmd->request->rq_flags & RQF_DONTPREP) {
158                 cmd->request->rq_flags &= ~RQF_DONTPREP;
159                 scsi_mq_uninit_cmd(cmd);
160         } else {
161                 WARN_ON_ONCE(true);
162         }
163         blk_mq_requeue_request(cmd->request, true);
164 }
165
166 /**
167  * __scsi_queue_insert - private queue insertion
168  * @cmd: The SCSI command being requeued
169  * @reason:  The reason for the requeue
170  * @unbusy: Whether the queue should be unbusied
171  *
172  * This is a private queue insertion.  The public interface
173  * scsi_queue_insert() always assumes the queue should be unbusied
174  * because it's always called before the completion.  This function is
175  * for a requeue after completion, which should only occur in this
176  * file.
177  */
178 static void __scsi_queue_insert(struct scsi_cmnd *cmd, int reason, bool unbusy)
179 {
180         struct scsi_device *device = cmd->device;
181
182         SCSI_LOG_MLQUEUE(1, scmd_printk(KERN_INFO, cmd,
183                 "Inserting command %p into mlqueue\n", cmd));
184
185         scsi_set_blocked(cmd, reason);
186
187         /*
188          * Decrement the counters, since these commands are no longer
189          * active on the host/device.
190          */
191         if (unbusy)
192                 scsi_device_unbusy(device, cmd);
193
194         /*
195          * Requeue this command.  It will go before all other commands
196          * that are already in the queue. Schedule requeue work under
197          * lock such that the kblockd_schedule_work() call happens
198          * before blk_cleanup_queue() finishes.
199          */
200         cmd->result = 0;
201
202         blk_mq_requeue_request(cmd->request, true);
203 }
204
205 /**
206  * scsi_queue_insert - Reinsert a command in the queue.
207  * @cmd:    command that we are adding to queue.
208  * @reason: why we are inserting command to queue.
209  *
210  * We do this for one of two cases. Either the host is busy and it cannot accept
211  * any more commands for the time being, or the device returned QUEUE_FULL and
212  * can accept no more commands.
213  *
214  * Context: This could be called either from an interrupt context or a normal
215  * process context.
216  */
217 void scsi_queue_insert(struct scsi_cmnd *cmd, int reason)
218 {
219         __scsi_queue_insert(cmd, reason, true);
220 }
221
222
223 /**
224  * __scsi_execute - insert request and wait for the result
225  * @sdev:       scsi device
226  * @cmd:        scsi command
227  * @data_direction: data direction
228  * @buffer:     data buffer
229  * @bufflen:    len of buffer
230  * @sense:      optional sense buffer
231  * @sshdr:      optional decoded sense header
232  * @timeout:    request timeout in seconds
233  * @retries:    number of times to retry request
234  * @flags:      flags for ->cmd_flags
235  * @rq_flags:   flags for ->rq_flags
236  * @resid:      optional residual length
237  *
238  * Returns the scsi_cmnd result field if a command was executed, or a negative
239  * Linux error code if we didn't get that far.
240  */
241 int __scsi_execute(struct scsi_device *sdev, const unsigned char *cmd,
242                  int data_direction, void *buffer, unsigned bufflen,
243                  unsigned char *sense, struct scsi_sense_hdr *sshdr,
244                  int timeout, int retries, u64 flags, req_flags_t rq_flags,
245                  int *resid)
246 {
247         struct request *req;
248         struct scsi_request *rq;
249         int ret = DRIVER_ERROR << 24;
250
251         req = blk_get_request(sdev->request_queue,
252                         data_direction == DMA_TO_DEVICE ?
253                         REQ_OP_SCSI_OUT : REQ_OP_SCSI_IN, BLK_MQ_REQ_PREEMPT);
254         if (IS_ERR(req))
255                 return ret;
256         rq = scsi_req(req);
257
258         if (bufflen &&  blk_rq_map_kern(sdev->request_queue, req,
259                                         buffer, bufflen, GFP_NOIO))
260                 goto out;
261
262         rq->cmd_len = COMMAND_SIZE(cmd[0]);
263         memcpy(rq->cmd, cmd, rq->cmd_len);
264         rq->retries = retries;
265         req->timeout = timeout;
266         req->cmd_flags |= flags;
267         req->rq_flags |= rq_flags | RQF_QUIET;
268
269         /*
270          * head injection *required* here otherwise quiesce won't work
271          */
272         blk_execute_rq(req->q, NULL, req, 1);
273
274         /*
275          * Some devices (USB mass-storage in particular) may transfer
276          * garbage data together with a residue indicating that the data
277          * is invalid.  Prevent the garbage from being misinterpreted
278          * and prevent security leaks by zeroing out the excess data.
279          */
280         if (unlikely(rq->resid_len > 0 && rq->resid_len <= bufflen))
281                 memset(buffer + (bufflen - rq->resid_len), 0, rq->resid_len);
282
283         if (resid)
284                 *resid = rq->resid_len;
285         if (sense && rq->sense_len)
286                 memcpy(sense, rq->sense, SCSI_SENSE_BUFFERSIZE);
287         if (sshdr)
288                 scsi_normalize_sense(rq->sense, rq->sense_len, sshdr);
289         ret = rq->result;
290  out:
291         blk_put_request(req);
292
293         return ret;
294 }
295 EXPORT_SYMBOL(__scsi_execute);
296
297 /**
298  * scsi_init_cmd_errh - Initialize cmd fields related to error handling.
299  * @cmd:  command that is ready to be queued.
300  *
301  * This function has the job of initializing a number of fields related to error
302  * handling. Typically this will be called once for each command, as required.
303  */
304 static void scsi_init_cmd_errh(struct scsi_cmnd *cmd)
305 {
306         scsi_set_resid(cmd, 0);
307         memset(cmd->sense_buffer, 0, SCSI_SENSE_BUFFERSIZE);
308         if (cmd->cmd_len == 0)
309                 cmd->cmd_len = scsi_command_size(cmd->cmnd);
310 }
311
312 /*
313  * Wake up the error handler if necessary. Avoid as follows that the error
314  * handler is not woken up if host in-flight requests number ==
315  * shost->host_failed: use call_rcu() in scsi_eh_scmd_add() in combination
316  * with an RCU read lock in this function to ensure that this function in
317  * its entirety either finishes before scsi_eh_scmd_add() increases the
318  * host_failed counter or that it notices the shost state change made by
319  * scsi_eh_scmd_add().
320  */
321 static void scsi_dec_host_busy(struct Scsi_Host *shost, struct scsi_cmnd *cmd)
322 {
323         unsigned long flags;
324
325         rcu_read_lock();
326         __clear_bit(SCMD_STATE_INFLIGHT, &cmd->state);
327         if (unlikely(scsi_host_in_recovery(shost))) {
328                 spin_lock_irqsave(shost->host_lock, flags);
329                 if (shost->host_failed || shost->host_eh_scheduled)
330                         scsi_eh_wakeup(shost);
331                 spin_unlock_irqrestore(shost->host_lock, flags);
332         }
333         rcu_read_unlock();
334 }
335
336 void scsi_device_unbusy(struct scsi_device *sdev, struct scsi_cmnd *cmd)
337 {
338         struct Scsi_Host *shost = sdev->host;
339         struct scsi_target *starget = scsi_target(sdev);
340
341         scsi_dec_host_busy(shost, cmd);
342
343         if (starget->can_queue > 0)
344                 atomic_dec(&starget->target_busy);
345
346         atomic_dec(&sdev->device_busy);
347 }
348
349 static void scsi_kick_queue(struct request_queue *q)
350 {
351         blk_mq_run_hw_queues(q, false);
352 }
353
354 /*
355  * Called for single_lun devices on IO completion. Clear starget_sdev_user,
356  * and call blk_run_queue for all the scsi_devices on the target -
357  * including current_sdev first.
358  *
359  * Called with *no* scsi locks held.
360  */
361 static void scsi_single_lun_run(struct scsi_device *current_sdev)
362 {
363         struct Scsi_Host *shost = current_sdev->host;
364         struct scsi_device *sdev, *tmp;
365         struct scsi_target *starget = scsi_target(current_sdev);
366         unsigned long flags;
367
368         spin_lock_irqsave(shost->host_lock, flags);
369         starget->starget_sdev_user = NULL;
370         spin_unlock_irqrestore(shost->host_lock, flags);
371
372         /*
373          * Call blk_run_queue for all LUNs on the target, starting with
374          * current_sdev. We race with others (to set starget_sdev_user),
375          * but in most cases, we will be first. Ideally, each LU on the
376          * target would get some limited time or requests on the target.
377          */
378         scsi_kick_queue(current_sdev->request_queue);
379
380         spin_lock_irqsave(shost->host_lock, flags);
381         if (starget->starget_sdev_user)
382                 goto out;
383         list_for_each_entry_safe(sdev, tmp, &starget->devices,
384                         same_target_siblings) {
385                 if (sdev == current_sdev)
386                         continue;
387                 if (scsi_device_get(sdev))
388                         continue;
389
390                 spin_unlock_irqrestore(shost->host_lock, flags);
391                 scsi_kick_queue(sdev->request_queue);
392                 spin_lock_irqsave(shost->host_lock, flags);
393         
394                 scsi_device_put(sdev);
395         }
396  out:
397         spin_unlock_irqrestore(shost->host_lock, flags);
398 }
399
400 static inline bool scsi_device_is_busy(struct scsi_device *sdev)
401 {
402         if (atomic_read(&sdev->device_busy) >= sdev->queue_depth)
403                 return true;
404         if (atomic_read(&sdev->device_blocked) > 0)
405                 return true;
406         return false;
407 }
408
409 static inline bool scsi_target_is_busy(struct scsi_target *starget)
410 {
411         if (starget->can_queue > 0) {
412                 if (atomic_read(&starget->target_busy) >= starget->can_queue)
413                         return true;
414                 if (atomic_read(&starget->target_blocked) > 0)
415                         return true;
416         }
417         return false;
418 }
419
420 static inline bool scsi_host_is_busy(struct Scsi_Host *shost)
421 {
422         if (atomic_read(&shost->host_blocked) > 0)
423                 return true;
424         if (shost->host_self_blocked)
425                 return true;
426         return false;
427 }
428
429 static void scsi_starved_list_run(struct Scsi_Host *shost)
430 {
431         LIST_HEAD(starved_list);
432         struct scsi_device *sdev;
433         unsigned long flags;
434
435         spin_lock_irqsave(shost->host_lock, flags);
436         list_splice_init(&shost->starved_list, &starved_list);
437
438         while (!list_empty(&starved_list)) {
439                 struct request_queue *slq;
440
441                 /*
442                  * As long as shost is accepting commands and we have
443                  * starved queues, call blk_run_queue. scsi_request_fn
444                  * drops the queue_lock and can add us back to the
445                  * starved_list.
446                  *
447                  * host_lock protects the starved_list and starved_entry.
448                  * scsi_request_fn must get the host_lock before checking
449                  * or modifying starved_list or starved_entry.
450                  */
451                 if (scsi_host_is_busy(shost))
452                         break;
453
454                 sdev = list_entry(starved_list.next,
455                                   struct scsi_device, starved_entry);
456                 list_del_init(&sdev->starved_entry);
457                 if (scsi_target_is_busy(scsi_target(sdev))) {
458                         list_move_tail(&sdev->starved_entry,
459                                        &shost->starved_list);
460                         continue;
461                 }
462
463                 /*
464                  * Once we drop the host lock, a racing scsi_remove_device()
465                  * call may remove the sdev from the starved list and destroy
466                  * it and the queue.  Mitigate by taking a reference to the
467                  * queue and never touching the sdev again after we drop the
468                  * host lock.  Note: if __scsi_remove_device() invokes
469                  * blk_cleanup_queue() before the queue is run from this
470                  * function then blk_run_queue() will return immediately since
471                  * blk_cleanup_queue() marks the queue with QUEUE_FLAG_DYING.
472                  */
473                 slq = sdev->request_queue;
474                 if (!blk_get_queue(slq))
475                         continue;
476                 spin_unlock_irqrestore(shost->host_lock, flags);
477
478                 scsi_kick_queue(slq);
479                 blk_put_queue(slq);
480
481                 spin_lock_irqsave(shost->host_lock, flags);
482         }
483         /* put any unprocessed entries back */
484         list_splice(&starved_list, &shost->starved_list);
485         spin_unlock_irqrestore(shost->host_lock, flags);
486 }
487
488 /**
489  * scsi_run_queue - Select a proper request queue to serve next.
490  * @q:  last request's queue
491  *
492  * The previous command was completely finished, start a new one if possible.
493  */
494 static void scsi_run_queue(struct request_queue *q)
495 {
496         struct scsi_device *sdev = q->queuedata;
497
498         if (scsi_target(sdev)->single_lun)
499                 scsi_single_lun_run(sdev);
500         if (!list_empty(&sdev->host->starved_list))
501                 scsi_starved_list_run(sdev->host);
502
503         blk_mq_run_hw_queues(q, false);
504 }
505
506 void scsi_requeue_run_queue(struct work_struct *work)
507 {
508         struct scsi_device *sdev;
509         struct request_queue *q;
510
511         sdev = container_of(work, struct scsi_device, requeue_work);
512         q = sdev->request_queue;
513         scsi_run_queue(q);
514 }
515
516 void scsi_run_host_queues(struct Scsi_Host *shost)
517 {
518         struct scsi_device *sdev;
519
520         shost_for_each_device(sdev, shost)
521                 scsi_run_queue(sdev->request_queue);
522 }
523
524 static void scsi_uninit_cmd(struct scsi_cmnd *cmd)
525 {
526         if (!blk_rq_is_passthrough(cmd->request)) {
527                 struct scsi_driver *drv = scsi_cmd_to_driver(cmd);
528
529                 if (drv->uninit_command)
530                         drv->uninit_command(cmd);
531         }
532 }
533
534 static void scsi_free_sgtables(struct scsi_cmnd *cmd)
535 {
536         if (cmd->sdb.table.nents)
537                 sg_free_table_chained(&cmd->sdb.table,
538                                 SCSI_INLINE_SG_CNT);
539         if (scsi_prot_sg_count(cmd))
540                 sg_free_table_chained(&cmd->prot_sdb->table,
541                                 SCSI_INLINE_PROT_SG_CNT);
542 }
543
544 static void scsi_mq_uninit_cmd(struct scsi_cmnd *cmd)
545 {
546         scsi_free_sgtables(cmd);
547         scsi_uninit_cmd(cmd);
548 }
549
550 /* Returns false when no more bytes to process, true if there are more */
551 static bool scsi_end_request(struct request *req, blk_status_t error,
552                 unsigned int bytes)
553 {
554         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
555         struct scsi_device *sdev = cmd->device;
556         struct request_queue *q = sdev->request_queue;
557
558         if (blk_update_request(req, error, bytes))
559                 return true;
560
561         if (blk_queue_add_random(q))
562                 add_disk_randomness(req->rq_disk);
563
564         if (!blk_rq_is_scsi(req)) {
565                 WARN_ON_ONCE(!(cmd->flags & SCMD_INITIALIZED));
566                 cmd->flags &= ~SCMD_INITIALIZED;
567         }
568
569         /*
570          * Calling rcu_barrier() is not necessary here because the
571          * SCSI error handler guarantees that the function called by
572          * call_rcu() has been called before scsi_end_request() is
573          * called.
574          */
575         destroy_rcu_head(&cmd->rcu);
576
577         /*
578          * In the MQ case the command gets freed by __blk_mq_end_request,
579          * so we have to do all cleanup that depends on it earlier.
580          *
581          * We also can't kick the queues from irq context, so we
582          * will have to defer it to a workqueue.
583          */
584         scsi_mq_uninit_cmd(cmd);
585
586         /*
587          * queue is still alive, so grab the ref for preventing it
588          * from being cleaned up during running queue.
589          */
590         percpu_ref_get(&q->q_usage_counter);
591
592         __blk_mq_end_request(req, error);
593
594         if (scsi_target(sdev)->single_lun ||
595             !list_empty(&sdev->host->starved_list))
596                 kblockd_schedule_work(&sdev->requeue_work);
597         else
598                 blk_mq_run_hw_queues(q, true);
599
600         percpu_ref_put(&q->q_usage_counter);
601         return false;
602 }
603
604 /**
605  * scsi_result_to_blk_status - translate a SCSI result code into blk_status_t
606  * @cmd:        SCSI command
607  * @result:     scsi error code
608  *
609  * Translate a SCSI result code into a blk_status_t value. May reset the host
610  * byte of @cmd->result.
611  */
612 static blk_status_t scsi_result_to_blk_status(struct scsi_cmnd *cmd, int result)
613 {
614         switch (host_byte(result)) {
615         case DID_OK:
616                 /*
617                  * Also check the other bytes than the status byte in result
618                  * to handle the case when a SCSI LLD sets result to
619                  * DRIVER_SENSE << 24 without setting SAM_STAT_CHECK_CONDITION.
620                  */
621                 if (scsi_status_is_good(result) && (result & ~0xff) == 0)
622                         return BLK_STS_OK;
623                 return BLK_STS_IOERR;
624         case DID_TRANSPORT_FAILFAST:
625                 return BLK_STS_TRANSPORT;
626         case DID_TARGET_FAILURE:
627                 set_host_byte(cmd, DID_OK);
628                 return BLK_STS_TARGET;
629         case DID_NEXUS_FAILURE:
630                 set_host_byte(cmd, DID_OK);
631                 return BLK_STS_NEXUS;
632         case DID_ALLOC_FAILURE:
633                 set_host_byte(cmd, DID_OK);
634                 return BLK_STS_NOSPC;
635         case DID_MEDIUM_ERROR:
636                 set_host_byte(cmd, DID_OK);
637                 return BLK_STS_MEDIUM;
638         default:
639                 return BLK_STS_IOERR;
640         }
641 }
642
643 /* Helper for scsi_io_completion() when "reprep" action required. */
644 static void scsi_io_completion_reprep(struct scsi_cmnd *cmd,
645                                       struct request_queue *q)
646 {
647         /* A new command will be prepared and issued. */
648         scsi_mq_requeue_cmd(cmd);
649 }
650
651 /* Helper for scsi_io_completion() when special action required. */
652 static void scsi_io_completion_action(struct scsi_cmnd *cmd, int result)
653 {
654         struct request_queue *q = cmd->device->request_queue;
655         struct request *req = cmd->request;
656         int level = 0;
657         enum {ACTION_FAIL, ACTION_REPREP, ACTION_RETRY,
658               ACTION_DELAYED_RETRY} action;
659         unsigned long wait_for = (cmd->allowed + 1) * req->timeout;
660         struct scsi_sense_hdr sshdr;
661         bool sense_valid;
662         bool sense_current = true;      /* false implies "deferred sense" */
663         blk_status_t blk_stat;
664
665         sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
666         if (sense_valid)
667                 sense_current = !scsi_sense_is_deferred(&sshdr);
668
669         blk_stat = scsi_result_to_blk_status(cmd, result);
670
671         if (host_byte(result) == DID_RESET) {
672                 /* Third party bus reset or reset for error recovery
673                  * reasons.  Just retry the command and see what
674                  * happens.
675                  */
676                 action = ACTION_RETRY;
677         } else if (sense_valid && sense_current) {
678                 switch (sshdr.sense_key) {
679                 case UNIT_ATTENTION:
680                         if (cmd->device->removable) {
681                                 /* Detected disc change.  Set a bit
682                                  * and quietly refuse further access.
683                                  */
684                                 cmd->device->changed = 1;
685                                 action = ACTION_FAIL;
686                         } else {
687                                 /* Must have been a power glitch, or a
688                                  * bus reset.  Could not have been a
689                                  * media change, so we just retry the
690                                  * command and see what happens.
691                                  */
692                                 action = ACTION_RETRY;
693                         }
694                         break;
695                 case ILLEGAL_REQUEST:
696                         /* If we had an ILLEGAL REQUEST returned, then
697                          * we may have performed an unsupported
698                          * command.  The only thing this should be
699                          * would be a ten byte read where only a six
700                          * byte read was supported.  Also, on a system
701                          * where READ CAPACITY failed, we may have
702                          * read past the end of the disk.
703                          */
704                         if ((cmd->device->use_10_for_rw &&
705                             sshdr.asc == 0x20 && sshdr.ascq == 0x00) &&
706                             (cmd->cmnd[0] == READ_10 ||
707                              cmd->cmnd[0] == WRITE_10)) {
708                                 /* This will issue a new 6-byte command. */
709                                 cmd->device->use_10_for_rw = 0;
710                                 action = ACTION_REPREP;
711                         } else if (sshdr.asc == 0x10) /* DIX */ {
712                                 action = ACTION_FAIL;
713                                 blk_stat = BLK_STS_PROTECTION;
714                         /* INVALID COMMAND OPCODE or INVALID FIELD IN CDB */
715                         } else if (sshdr.asc == 0x20 || sshdr.asc == 0x24) {
716                                 action = ACTION_FAIL;
717                                 blk_stat = BLK_STS_TARGET;
718                         } else
719                                 action = ACTION_FAIL;
720                         break;
721                 case ABORTED_COMMAND:
722                         action = ACTION_FAIL;
723                         if (sshdr.asc == 0x10) /* DIF */
724                                 blk_stat = BLK_STS_PROTECTION;
725                         break;
726                 case NOT_READY:
727                         /* If the device is in the process of becoming
728                          * ready, or has a temporary blockage, retry.
729                          */
730                         if (sshdr.asc == 0x04) {
731                                 switch (sshdr.ascq) {
732                                 case 0x01: /* becoming ready */
733                                 case 0x04: /* format in progress */
734                                 case 0x05: /* rebuild in progress */
735                                 case 0x06: /* recalculation in progress */
736                                 case 0x07: /* operation in progress */
737                                 case 0x08: /* Long write in progress */
738                                 case 0x09: /* self test in progress */
739                                 case 0x14: /* space allocation in progress */
740                                 case 0x1a: /* start stop unit in progress */
741                                 case 0x1b: /* sanitize in progress */
742                                 case 0x1d: /* configuration in progress */
743                                 case 0x24: /* depopulation in progress */
744                                         action = ACTION_DELAYED_RETRY;
745                                         break;
746                                 default:
747                                         action = ACTION_FAIL;
748                                         break;
749                                 }
750                         } else
751                                 action = ACTION_FAIL;
752                         break;
753                 case VOLUME_OVERFLOW:
754                         /* See SSC3rXX or current. */
755                         action = ACTION_FAIL;
756                         break;
757                 default:
758                         action = ACTION_FAIL;
759                         break;
760                 }
761         } else
762                 action = ACTION_FAIL;
763
764         if (action != ACTION_FAIL &&
765             time_before(cmd->jiffies_at_alloc + wait_for, jiffies))
766                 action = ACTION_FAIL;
767
768         switch (action) {
769         case ACTION_FAIL:
770                 /* Give up and fail the remainder of the request */
771                 if (!(req->rq_flags & RQF_QUIET)) {
772                         static DEFINE_RATELIMIT_STATE(_rs,
773                                         DEFAULT_RATELIMIT_INTERVAL,
774                                         DEFAULT_RATELIMIT_BURST);
775
776                         if (unlikely(scsi_logging_level))
777                                 level =
778                                      SCSI_LOG_LEVEL(SCSI_LOG_MLCOMPLETE_SHIFT,
779                                                     SCSI_LOG_MLCOMPLETE_BITS);
780
781                         /*
782                          * if logging is enabled the failure will be printed
783                          * in scsi_log_completion(), so avoid duplicate messages
784                          */
785                         if (!level && __ratelimit(&_rs)) {
786                                 scsi_print_result(cmd, NULL, FAILED);
787                                 if (driver_byte(result) == DRIVER_SENSE)
788                                         scsi_print_sense(cmd);
789                                 scsi_print_command(cmd);
790                         }
791                 }
792                 if (!scsi_end_request(req, blk_stat, blk_rq_err_bytes(req)))
793                         return;
794                 /*FALLTHRU*/
795         case ACTION_REPREP:
796                 scsi_io_completion_reprep(cmd, q);
797                 break;
798         case ACTION_RETRY:
799                 /* Retry the same command immediately */
800                 __scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY, false);
801                 break;
802         case ACTION_DELAYED_RETRY:
803                 /* Retry the same command after a delay */
804                 __scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY, false);
805                 break;
806         }
807 }
808
809 /*
810  * Helper for scsi_io_completion() when cmd->result is non-zero. Returns a
811  * new result that may suppress further error checking. Also modifies
812  * *blk_statp in some cases.
813  */
814 static int scsi_io_completion_nz_result(struct scsi_cmnd *cmd, int result,
815                                         blk_status_t *blk_statp)
816 {
817         bool sense_valid;
818         bool sense_current = true;      /* false implies "deferred sense" */
819         struct request *req = cmd->request;
820         struct scsi_sense_hdr sshdr;
821
822         sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
823         if (sense_valid)
824                 sense_current = !scsi_sense_is_deferred(&sshdr);
825
826         if (blk_rq_is_passthrough(req)) {
827                 if (sense_valid) {
828                         /*
829                          * SG_IO wants current and deferred errors
830                          */
831                         scsi_req(req)->sense_len =
832                                 min(8 + cmd->sense_buffer[7],
833                                     SCSI_SENSE_BUFFERSIZE);
834                 }
835                 if (sense_current)
836                         *blk_statp = scsi_result_to_blk_status(cmd, result);
837         } else if (blk_rq_bytes(req) == 0 && sense_current) {
838                 /*
839                  * Flush commands do not transfers any data, and thus cannot use
840                  * good_bytes != blk_rq_bytes(req) as the signal for an error.
841                  * This sets *blk_statp explicitly for the problem case.
842                  */
843                 *blk_statp = scsi_result_to_blk_status(cmd, result);
844         }
845         /*
846          * Recovered errors need reporting, but they're always treated as
847          * success, so fiddle the result code here.  For passthrough requests
848          * we already took a copy of the original into sreq->result which
849          * is what gets returned to the user
850          */
851         if (sense_valid && (sshdr.sense_key == RECOVERED_ERROR)) {
852                 bool do_print = true;
853                 /*
854                  * if ATA PASS-THROUGH INFORMATION AVAILABLE [0x0, 0x1d]
855                  * skip print since caller wants ATA registers. Only occurs
856                  * on SCSI ATA PASS_THROUGH commands when CK_COND=1
857                  */
858                 if ((sshdr.asc == 0x0) && (sshdr.ascq == 0x1d))
859                         do_print = false;
860                 else if (req->rq_flags & RQF_QUIET)
861                         do_print = false;
862                 if (do_print)
863                         scsi_print_sense(cmd);
864                 result = 0;
865                 /* for passthrough, *blk_statp may be set */
866                 *blk_statp = BLK_STS_OK;
867         }
868         /*
869          * Another corner case: the SCSI status byte is non-zero but 'good'.
870          * Example: PRE-FETCH command returns SAM_STAT_CONDITION_MET when
871          * it is able to fit nominated LBs in its cache (and SAM_STAT_GOOD
872          * if it can't fit). Treat SAM_STAT_CONDITION_MET and the related
873          * intermediate statuses (both obsolete in SAM-4) as good.
874          */
875         if (status_byte(result) && scsi_status_is_good(result)) {
876                 result = 0;
877                 *blk_statp = BLK_STS_OK;
878         }
879         return result;
880 }
881
882 /**
883  * scsi_io_completion - Completion processing for SCSI commands.
884  * @cmd:        command that is finished.
885  * @good_bytes: number of processed bytes.
886  *
887  * We will finish off the specified number of sectors. If we are done, the
888  * command block will be released and the queue function will be goosed. If we
889  * are not done then we have to figure out what to do next:
890  *
891  *   a) We can call scsi_io_completion_reprep().  The request will be
892  *      unprepared and put back on the queue.  Then a new command will
893  *      be created for it.  This should be used if we made forward
894  *      progress, or if we want to switch from READ(10) to READ(6) for
895  *      example.
896  *
897  *   b) We can call scsi_io_completion_action().  The request will be
898  *      put back on the queue and retried using the same command as
899  *      before, possibly after a delay.
900  *
901  *   c) We can call scsi_end_request() with blk_stat other than
902  *      BLK_STS_OK, to fail the remainder of the request.
903  */
904 void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes)
905 {
906         int result = cmd->result;
907         struct request_queue *q = cmd->device->request_queue;
908         struct request *req = cmd->request;
909         blk_status_t blk_stat = BLK_STS_OK;
910
911         if (unlikely(result))   /* a nz result may or may not be an error */
912                 result = scsi_io_completion_nz_result(cmd, result, &blk_stat);
913
914         if (unlikely(blk_rq_is_passthrough(req))) {
915                 /*
916                  * scsi_result_to_blk_status may have reset the host_byte
917                  */
918                 scsi_req(req)->result = cmd->result;
919         }
920
921         /*
922          * Next deal with any sectors which we were able to correctly
923          * handle.
924          */
925         SCSI_LOG_HLCOMPLETE(1, scmd_printk(KERN_INFO, cmd,
926                 "%u sectors total, %d bytes done.\n",
927                 blk_rq_sectors(req), good_bytes));
928
929         /*
930          * Failed, zero length commands always need to drop down
931          * to retry code. Fast path should return in this block.
932          */
933         if (likely(blk_rq_bytes(req) > 0 || blk_stat == BLK_STS_OK)) {
934                 if (likely(!scsi_end_request(req, blk_stat, good_bytes)))
935                         return; /* no bytes remaining */
936         }
937
938         /* Kill remainder if no retries. */
939         if (unlikely(blk_stat && scsi_noretry_cmd(cmd))) {
940                 if (scsi_end_request(req, blk_stat, blk_rq_bytes(req)))
941                         WARN_ONCE(true,
942                             "Bytes remaining after failed, no-retry command");
943                 return;
944         }
945
946         /*
947          * If there had been no error, but we have leftover bytes in the
948          * requeues just queue the command up again.
949          */
950         if (likely(result == 0))
951                 scsi_io_completion_reprep(cmd, q);
952         else
953                 scsi_io_completion_action(cmd, result);
954 }
955
956 static inline bool scsi_cmd_needs_dma_drain(struct scsi_device *sdev,
957                 struct request *rq)
958 {
959         return sdev->dma_drain_len && blk_rq_is_passthrough(rq) &&
960                !op_is_write(req_op(rq)) &&
961                sdev->host->hostt->dma_need_drain(rq);
962 }
963
964 /**
965  * scsi_init_io - SCSI I/O initialization function.
966  * @cmd:  command descriptor we wish to initialize
967  *
968  * Returns:
969  * * BLK_STS_OK       - on success
970  * * BLK_STS_RESOURCE - if the failure is retryable
971  * * BLK_STS_IOERR    - if the failure is fatal
972  */
973 blk_status_t scsi_init_io(struct scsi_cmnd *cmd)
974 {
975         struct scsi_device *sdev = cmd->device;
976         struct request *rq = cmd->request;
977         unsigned short nr_segs = blk_rq_nr_phys_segments(rq);
978         struct scatterlist *last_sg = NULL;
979         blk_status_t ret;
980         bool need_drain = scsi_cmd_needs_dma_drain(sdev, rq);
981         int count;
982
983         if (WARN_ON_ONCE(!nr_segs))
984                 return BLK_STS_IOERR;
985
986         /*
987          * Make sure there is space for the drain.  The driver must adjust
988          * max_hw_segments to be prepared for this.
989          */
990         if (need_drain)
991                 nr_segs++;
992
993         /*
994          * If sg table allocation fails, requeue request later.
995          */
996         if (unlikely(sg_alloc_table_chained(&cmd->sdb.table, nr_segs,
997                         cmd->sdb.table.sgl, SCSI_INLINE_SG_CNT)))
998                 return BLK_STS_RESOURCE;
999
1000         /*
1001          * Next, walk the list, and fill in the addresses and sizes of
1002          * each segment.
1003          */
1004         count = __blk_rq_map_sg(rq->q, rq, cmd->sdb.table.sgl, &last_sg);
1005
1006         if (blk_rq_bytes(rq) & rq->q->dma_pad_mask) {
1007                 unsigned int pad_len =
1008                         (rq->q->dma_pad_mask & ~blk_rq_bytes(rq)) + 1;
1009
1010                 last_sg->length += pad_len;
1011                 cmd->extra_len += pad_len;
1012         }
1013
1014         if (need_drain) {
1015                 sg_unmark_end(last_sg);
1016                 last_sg = sg_next(last_sg);
1017                 sg_set_buf(last_sg, sdev->dma_drain_buf, sdev->dma_drain_len);
1018                 sg_mark_end(last_sg);
1019
1020                 cmd->extra_len += sdev->dma_drain_len;
1021                 count++;
1022         }
1023
1024         BUG_ON(count > cmd->sdb.table.nents);
1025         cmd->sdb.table.nents = count;
1026         cmd->sdb.length = blk_rq_payload_bytes(rq);
1027
1028         if (blk_integrity_rq(rq)) {
1029                 struct scsi_data_buffer *prot_sdb = cmd->prot_sdb;
1030                 int ivecs;
1031
1032                 if (WARN_ON_ONCE(!prot_sdb)) {
1033                         /*
1034                          * This can happen if someone (e.g. multipath)
1035                          * queues a command to a device on an adapter
1036                          * that does not support DIX.
1037                          */
1038                         ret = BLK_STS_IOERR;
1039                         goto out_free_sgtables;
1040                 }
1041
1042                 ivecs = blk_rq_count_integrity_sg(rq->q, rq->bio);
1043
1044                 if (sg_alloc_table_chained(&prot_sdb->table, ivecs,
1045                                 prot_sdb->table.sgl,
1046                                 SCSI_INLINE_PROT_SG_CNT)) {
1047                         ret = BLK_STS_RESOURCE;
1048                         goto out_free_sgtables;
1049                 }
1050
1051                 count = blk_rq_map_integrity_sg(rq->q, rq->bio,
1052                                                 prot_sdb->table.sgl);
1053                 BUG_ON(count > ivecs);
1054                 BUG_ON(count > queue_max_integrity_segments(rq->q));
1055
1056                 cmd->prot_sdb = prot_sdb;
1057                 cmd->prot_sdb->table.nents = count;
1058         }
1059
1060         return BLK_STS_OK;
1061 out_free_sgtables:
1062         scsi_free_sgtables(cmd);
1063         return ret;
1064 }
1065 EXPORT_SYMBOL(scsi_init_io);
1066
1067 /**
1068  * scsi_initialize_rq - initialize struct scsi_cmnd partially
1069  * @rq: Request associated with the SCSI command to be initialized.
1070  *
1071  * This function initializes the members of struct scsi_cmnd that must be
1072  * initialized before request processing starts and that won't be
1073  * reinitialized if a SCSI command is requeued.
1074  *
1075  * Called from inside blk_get_request() for pass-through requests and from
1076  * inside scsi_init_command() for filesystem requests.
1077  */
1078 static void scsi_initialize_rq(struct request *rq)
1079 {
1080         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1081
1082         scsi_req_init(&cmd->req);
1083         init_rcu_head(&cmd->rcu);
1084         cmd->jiffies_at_alloc = jiffies;
1085         cmd->retries = 0;
1086 }
1087
1088 /*
1089  * Only called when the request isn't completed by SCSI, and not freed by
1090  * SCSI
1091  */
1092 static void scsi_cleanup_rq(struct request *rq)
1093 {
1094         if (rq->rq_flags & RQF_DONTPREP) {
1095                 scsi_mq_uninit_cmd(blk_mq_rq_to_pdu(rq));
1096                 rq->rq_flags &= ~RQF_DONTPREP;
1097         }
1098 }
1099
1100 /* Called before a request is prepared. See also scsi_mq_prep_fn(). */
1101 void scsi_init_command(struct scsi_device *dev, struct scsi_cmnd *cmd)
1102 {
1103         void *buf = cmd->sense_buffer;
1104         void *prot = cmd->prot_sdb;
1105         struct request *rq = blk_mq_rq_from_pdu(cmd);
1106         unsigned int flags = cmd->flags & SCMD_PRESERVED_FLAGS;
1107         unsigned long jiffies_at_alloc;
1108         int retries, to_clear;
1109         bool in_flight;
1110
1111         if (!blk_rq_is_scsi(rq) && !(flags & SCMD_INITIALIZED)) {
1112                 flags |= SCMD_INITIALIZED;
1113                 scsi_initialize_rq(rq);
1114         }
1115
1116         jiffies_at_alloc = cmd->jiffies_at_alloc;
1117         retries = cmd->retries;
1118         in_flight = test_bit(SCMD_STATE_INFLIGHT, &cmd->state);
1119         /*
1120          * Zero out the cmd, except for the embedded scsi_request. Only clear
1121          * the driver-private command data if the LLD does not supply a
1122          * function to initialize that data.
1123          */
1124         to_clear = sizeof(*cmd) - sizeof(cmd->req);
1125         if (!dev->host->hostt->init_cmd_priv)
1126                 to_clear += dev->host->hostt->cmd_size;
1127         memset((char *)cmd + sizeof(cmd->req), 0, to_clear);
1128
1129         cmd->device = dev;
1130         cmd->sense_buffer = buf;
1131         cmd->prot_sdb = prot;
1132         cmd->flags = flags;
1133         INIT_DELAYED_WORK(&cmd->abort_work, scmd_eh_abort_handler);
1134         cmd->jiffies_at_alloc = jiffies_at_alloc;
1135         cmd->retries = retries;
1136         if (in_flight)
1137                 __set_bit(SCMD_STATE_INFLIGHT, &cmd->state);
1138
1139 }
1140
1141 static blk_status_t scsi_setup_scsi_cmnd(struct scsi_device *sdev,
1142                 struct request *req)
1143 {
1144         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1145
1146         /*
1147          * Passthrough requests may transfer data, in which case they must
1148          * a bio attached to them.  Or they might contain a SCSI command
1149          * that does not transfer data, in which case they may optionally
1150          * submit a request without an attached bio.
1151          */
1152         if (req->bio) {
1153                 blk_status_t ret = scsi_init_io(cmd);
1154                 if (unlikely(ret != BLK_STS_OK))
1155                         return ret;
1156         } else {
1157                 BUG_ON(blk_rq_bytes(req));
1158
1159                 memset(&cmd->sdb, 0, sizeof(cmd->sdb));
1160         }
1161
1162         cmd->cmd_len = scsi_req(req)->cmd_len;
1163         cmd->cmnd = scsi_req(req)->cmd;
1164         cmd->transfersize = blk_rq_bytes(req);
1165         cmd->allowed = scsi_req(req)->retries;
1166         return BLK_STS_OK;
1167 }
1168
1169 /*
1170  * Setup a normal block command.  These are simple request from filesystems
1171  * that still need to be translated to SCSI CDBs from the ULD.
1172  */
1173 static blk_status_t scsi_setup_fs_cmnd(struct scsi_device *sdev,
1174                 struct request *req)
1175 {
1176         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1177
1178         if (unlikely(sdev->handler && sdev->handler->prep_fn)) {
1179                 blk_status_t ret = sdev->handler->prep_fn(sdev, req);
1180                 if (ret != BLK_STS_OK)
1181                         return ret;
1182         }
1183
1184         cmd->cmnd = scsi_req(req)->cmd = scsi_req(req)->__cmd;
1185         memset(cmd->cmnd, 0, BLK_MAX_CDB);
1186         return scsi_cmd_to_driver(cmd)->init_command(cmd);
1187 }
1188
1189 static blk_status_t scsi_setup_cmnd(struct scsi_device *sdev,
1190                 struct request *req)
1191 {
1192         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1193         blk_status_t ret;
1194
1195         if (!blk_rq_bytes(req))
1196                 cmd->sc_data_direction = DMA_NONE;
1197         else if (rq_data_dir(req) == WRITE)
1198                 cmd->sc_data_direction = DMA_TO_DEVICE;
1199         else
1200                 cmd->sc_data_direction = DMA_FROM_DEVICE;
1201
1202         if (blk_rq_is_scsi(req))
1203                 ret = scsi_setup_scsi_cmnd(sdev, req);
1204         else
1205                 ret = scsi_setup_fs_cmnd(sdev, req);
1206
1207         if (ret != BLK_STS_OK)
1208                 scsi_free_sgtables(cmd);
1209
1210         return ret;
1211 }
1212
1213 static blk_status_t
1214 scsi_prep_state_check(struct scsi_device *sdev, struct request *req)
1215 {
1216         switch (sdev->sdev_state) {
1217         case SDEV_OFFLINE:
1218         case SDEV_TRANSPORT_OFFLINE:
1219                 /*
1220                  * If the device is offline we refuse to process any
1221                  * commands.  The device must be brought online
1222                  * before trying any recovery commands.
1223                  */
1224                 if (!sdev->offline_already) {
1225                         sdev->offline_already = true;
1226                         sdev_printk(KERN_ERR, sdev,
1227                                     "rejecting I/O to offline device\n");
1228                 }
1229                 return BLK_STS_IOERR;
1230         case SDEV_DEL:
1231                 /*
1232                  * If the device is fully deleted, we refuse to
1233                  * process any commands as well.
1234                  */
1235                 sdev_printk(KERN_ERR, sdev,
1236                             "rejecting I/O to dead device\n");
1237                 return BLK_STS_IOERR;
1238         case SDEV_BLOCK:
1239         case SDEV_CREATED_BLOCK:
1240                 return BLK_STS_RESOURCE;
1241         case SDEV_QUIESCE:
1242                 /*
1243                  * If the devices is blocked we defer normal commands.
1244                  */
1245                 if (req && !(req->rq_flags & RQF_PREEMPT))
1246                         return BLK_STS_RESOURCE;
1247                 return BLK_STS_OK;
1248         default:
1249                 /*
1250                  * For any other not fully online state we only allow
1251                  * special commands.  In particular any user initiated
1252                  * command is not allowed.
1253                  */
1254                 if (req && !(req->rq_flags & RQF_PREEMPT))
1255                         return BLK_STS_IOERR;
1256                 return BLK_STS_OK;
1257         }
1258 }
1259
1260 /*
1261  * scsi_dev_queue_ready: if we can send requests to sdev, return 1 else
1262  * return 0.
1263  *
1264  * Called with the queue_lock held.
1265  */
1266 static inline int scsi_dev_queue_ready(struct request_queue *q,
1267                                   struct scsi_device *sdev)
1268 {
1269         unsigned int busy;
1270
1271         busy = atomic_inc_return(&sdev->device_busy) - 1;
1272         if (atomic_read(&sdev->device_blocked)) {
1273                 if (busy)
1274                         goto out_dec;
1275
1276                 /*
1277                  * unblock after device_blocked iterates to zero
1278                  */
1279                 if (atomic_dec_return(&sdev->device_blocked) > 0)
1280                         goto out_dec;
1281                 SCSI_LOG_MLQUEUE(3, sdev_printk(KERN_INFO, sdev,
1282                                    "unblocking device at zero depth\n"));
1283         }
1284
1285         if (busy >= sdev->queue_depth)
1286                 goto out_dec;
1287
1288         return 1;
1289 out_dec:
1290         atomic_dec(&sdev->device_busy);
1291         return 0;
1292 }
1293
1294 /*
1295  * scsi_target_queue_ready: checks if there we can send commands to target
1296  * @sdev: scsi device on starget to check.
1297  */
1298 static inline int scsi_target_queue_ready(struct Scsi_Host *shost,
1299                                            struct scsi_device *sdev)
1300 {
1301         struct scsi_target *starget = scsi_target(sdev);
1302         unsigned int busy;
1303
1304         if (starget->single_lun) {
1305                 spin_lock_irq(shost->host_lock);
1306                 if (starget->starget_sdev_user &&
1307                     starget->starget_sdev_user != sdev) {
1308                         spin_unlock_irq(shost->host_lock);
1309                         return 0;
1310                 }
1311                 starget->starget_sdev_user = sdev;
1312                 spin_unlock_irq(shost->host_lock);
1313         }
1314
1315         if (starget->can_queue <= 0)
1316                 return 1;
1317
1318         busy = atomic_inc_return(&starget->target_busy) - 1;
1319         if (atomic_read(&starget->target_blocked) > 0) {
1320                 if (busy)
1321                         goto starved;
1322
1323                 /*
1324                  * unblock after target_blocked iterates to zero
1325                  */
1326                 if (atomic_dec_return(&starget->target_blocked) > 0)
1327                         goto out_dec;
1328
1329                 SCSI_LOG_MLQUEUE(3, starget_printk(KERN_INFO, starget,
1330                                  "unblocking target at zero depth\n"));
1331         }
1332
1333         if (busy >= starget->can_queue)
1334                 goto starved;
1335
1336         return 1;
1337
1338 starved:
1339         spin_lock_irq(shost->host_lock);
1340         list_move_tail(&sdev->starved_entry, &shost->starved_list);
1341         spin_unlock_irq(shost->host_lock);
1342 out_dec:
1343         if (starget->can_queue > 0)
1344                 atomic_dec(&starget->target_busy);
1345         return 0;
1346 }
1347
1348 /*
1349  * scsi_host_queue_ready: if we can send requests to shost, return 1 else
1350  * return 0. We must end up running the queue again whenever 0 is
1351  * returned, else IO can hang.
1352  */
1353 static inline int scsi_host_queue_ready(struct request_queue *q,
1354                                    struct Scsi_Host *shost,
1355                                    struct scsi_device *sdev,
1356                                    struct scsi_cmnd *cmd)
1357 {
1358         if (scsi_host_in_recovery(shost))
1359                 return 0;
1360
1361         if (atomic_read(&shost->host_blocked) > 0) {
1362                 if (scsi_host_busy(shost) > 0)
1363                         goto starved;
1364
1365                 /*
1366                  * unblock after host_blocked iterates to zero
1367                  */
1368                 if (atomic_dec_return(&shost->host_blocked) > 0)
1369                         goto out_dec;
1370
1371                 SCSI_LOG_MLQUEUE(3,
1372                         shost_printk(KERN_INFO, shost,
1373                                      "unblocking host at zero depth\n"));
1374         }
1375
1376         if (shost->host_self_blocked)
1377                 goto starved;
1378
1379         /* We're OK to process the command, so we can't be starved */
1380         if (!list_empty(&sdev->starved_entry)) {
1381                 spin_lock_irq(shost->host_lock);
1382                 if (!list_empty(&sdev->starved_entry))
1383                         list_del_init(&sdev->starved_entry);
1384                 spin_unlock_irq(shost->host_lock);
1385         }
1386
1387         __set_bit(SCMD_STATE_INFLIGHT, &cmd->state);
1388
1389         return 1;
1390
1391 starved:
1392         spin_lock_irq(shost->host_lock);
1393         if (list_empty(&sdev->starved_entry))
1394                 list_add_tail(&sdev->starved_entry, &shost->starved_list);
1395         spin_unlock_irq(shost->host_lock);
1396 out_dec:
1397         scsi_dec_host_busy(shost, cmd);
1398         return 0;
1399 }
1400
1401 /*
1402  * Busy state exporting function for request stacking drivers.
1403  *
1404  * For efficiency, no lock is taken to check the busy state of
1405  * shost/starget/sdev, since the returned value is not guaranteed and
1406  * may be changed after request stacking drivers call the function,
1407  * regardless of taking lock or not.
1408  *
1409  * When scsi can't dispatch I/Os anymore and needs to kill I/Os scsi
1410  * needs to return 'not busy'. Otherwise, request stacking drivers
1411  * may hold requests forever.
1412  */
1413 static bool scsi_mq_lld_busy(struct request_queue *q)
1414 {
1415         struct scsi_device *sdev = q->queuedata;
1416         struct Scsi_Host *shost;
1417
1418         if (blk_queue_dying(q))
1419                 return false;
1420
1421         shost = sdev->host;
1422
1423         /*
1424          * Ignore host/starget busy state.
1425          * Since block layer does not have a concept of fairness across
1426          * multiple queues, congestion of host/starget needs to be handled
1427          * in SCSI layer.
1428          */
1429         if (scsi_host_in_recovery(shost) || scsi_device_is_busy(sdev))
1430                 return true;
1431
1432         return false;
1433 }
1434
1435 static void scsi_softirq_done(struct request *rq)
1436 {
1437         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1438         unsigned long wait_for = (cmd->allowed + 1) * rq->timeout;
1439         int disposition;
1440
1441         INIT_LIST_HEAD(&cmd->eh_entry);
1442
1443         atomic_inc(&cmd->device->iodone_cnt);
1444         if (cmd->result)
1445                 atomic_inc(&cmd->device->ioerr_cnt);
1446
1447         disposition = scsi_decide_disposition(cmd);
1448         if (disposition != SUCCESS &&
1449             time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) {
1450                 scmd_printk(KERN_ERR, cmd,
1451                             "timing out command, waited %lus\n",
1452                             wait_for/HZ);
1453                 disposition = SUCCESS;
1454         }
1455
1456         scsi_log_completion(cmd, disposition);
1457
1458         switch (disposition) {
1459                 case SUCCESS:
1460                         scsi_finish_command(cmd);
1461                         break;
1462                 case NEEDS_RETRY:
1463                         scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY);
1464                         break;
1465                 case ADD_TO_MLQUEUE:
1466                         scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY);
1467                         break;
1468                 default:
1469                         scsi_eh_scmd_add(cmd);
1470                         break;
1471         }
1472 }
1473
1474 /**
1475  * scsi_dispatch_command - Dispatch a command to the low-level driver.
1476  * @cmd: command block we are dispatching.
1477  *
1478  * Return: nonzero return request was rejected and device's queue needs to be
1479  * plugged.
1480  */
1481 static int scsi_dispatch_cmd(struct scsi_cmnd *cmd)
1482 {
1483         struct Scsi_Host *host = cmd->device->host;
1484         int rtn = 0;
1485
1486         atomic_inc(&cmd->device->iorequest_cnt);
1487
1488         /* check if the device is still usable */
1489         if (unlikely(cmd->device->sdev_state == SDEV_DEL)) {
1490                 /* in SDEV_DEL we error all commands. DID_NO_CONNECT
1491                  * returns an immediate error upwards, and signals
1492                  * that the device is no longer present */
1493                 cmd->result = DID_NO_CONNECT << 16;
1494                 goto done;
1495         }
1496
1497         /* Check to see if the scsi lld made this device blocked. */
1498         if (unlikely(scsi_device_blocked(cmd->device))) {
1499                 /*
1500                  * in blocked state, the command is just put back on
1501                  * the device queue.  The suspend state has already
1502                  * blocked the queue so future requests should not
1503                  * occur until the device transitions out of the
1504                  * suspend state.
1505                  */
1506                 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1507                         "queuecommand : device blocked\n"));
1508                 return SCSI_MLQUEUE_DEVICE_BUSY;
1509         }
1510
1511         /* Store the LUN value in cmnd, if needed. */
1512         if (cmd->device->lun_in_cdb)
1513                 cmd->cmnd[1] = (cmd->cmnd[1] & 0x1f) |
1514                                (cmd->device->lun << 5 & 0xe0);
1515
1516         scsi_log_send(cmd);
1517
1518         /*
1519          * Before we queue this command, check if the command
1520          * length exceeds what the host adapter can handle.
1521          */
1522         if (cmd->cmd_len > cmd->device->host->max_cmd_len) {
1523                 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1524                                "queuecommand : command too long. "
1525                                "cdb_size=%d host->max_cmd_len=%d\n",
1526                                cmd->cmd_len, cmd->device->host->max_cmd_len));
1527                 cmd->result = (DID_ABORT << 16);
1528                 goto done;
1529         }
1530
1531         if (unlikely(host->shost_state == SHOST_DEL)) {
1532                 cmd->result = (DID_NO_CONNECT << 16);
1533                 goto done;
1534
1535         }
1536
1537         trace_scsi_dispatch_cmd_start(cmd);
1538         rtn = host->hostt->queuecommand(host, cmd);
1539         if (rtn) {
1540                 trace_scsi_dispatch_cmd_error(cmd, rtn);
1541                 if (rtn != SCSI_MLQUEUE_DEVICE_BUSY &&
1542                     rtn != SCSI_MLQUEUE_TARGET_BUSY)
1543                         rtn = SCSI_MLQUEUE_HOST_BUSY;
1544
1545                 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1546                         "queuecommand : request rejected\n"));
1547         }
1548
1549         return rtn;
1550  done:
1551         cmd->scsi_done(cmd);
1552         return 0;
1553 }
1554
1555 /* Size in bytes of the sg-list stored in the scsi-mq command-private data. */
1556 static unsigned int scsi_mq_inline_sgl_size(struct Scsi_Host *shost)
1557 {
1558         return min_t(unsigned int, shost->sg_tablesize, SCSI_INLINE_SG_CNT) *
1559                 sizeof(struct scatterlist);
1560 }
1561
1562 static blk_status_t scsi_mq_prep_fn(struct request *req)
1563 {
1564         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1565         struct scsi_device *sdev = req->q->queuedata;
1566         struct Scsi_Host *shost = sdev->host;
1567         struct scatterlist *sg;
1568
1569         scsi_init_command(sdev, cmd);
1570
1571         cmd->request = req;
1572         cmd->tag = req->tag;
1573         cmd->prot_op = SCSI_PROT_NORMAL;
1574
1575         sg = (void *)cmd + sizeof(struct scsi_cmnd) + shost->hostt->cmd_size;
1576         cmd->sdb.table.sgl = sg;
1577
1578         if (scsi_host_get_prot(shost)) {
1579                 memset(cmd->prot_sdb, 0, sizeof(struct scsi_data_buffer));
1580
1581                 cmd->prot_sdb->table.sgl =
1582                         (struct scatterlist *)(cmd->prot_sdb + 1);
1583         }
1584
1585         blk_mq_start_request(req);
1586
1587         return scsi_setup_cmnd(sdev, req);
1588 }
1589
1590 static void scsi_mq_done(struct scsi_cmnd *cmd)
1591 {
1592         if (unlikely(test_and_set_bit(SCMD_STATE_COMPLETE, &cmd->state)))
1593                 return;
1594         trace_scsi_dispatch_cmd_done(cmd);
1595
1596         /*
1597          * If the block layer didn't complete the request due to a timeout
1598          * injection, scsi must clear its internal completed state so that the
1599          * timeout handler will see it needs to escalate its own error
1600          * recovery.
1601          */
1602         if (unlikely(!blk_mq_complete_request(cmd->request)))
1603                 clear_bit(SCMD_STATE_COMPLETE, &cmd->state);
1604 }
1605
1606 static void scsi_mq_put_budget(struct blk_mq_hw_ctx *hctx)
1607 {
1608         struct request_queue *q = hctx->queue;
1609         struct scsi_device *sdev = q->queuedata;
1610
1611         atomic_dec(&sdev->device_busy);
1612 }
1613
1614 static bool scsi_mq_get_budget(struct blk_mq_hw_ctx *hctx)
1615 {
1616         struct request_queue *q = hctx->queue;
1617         struct scsi_device *sdev = q->queuedata;
1618
1619         return scsi_dev_queue_ready(q, sdev);
1620 }
1621
1622 static blk_status_t scsi_queue_rq(struct blk_mq_hw_ctx *hctx,
1623                          const struct blk_mq_queue_data *bd)
1624 {
1625         struct request *req = bd->rq;
1626         struct request_queue *q = req->q;
1627         struct scsi_device *sdev = q->queuedata;
1628         struct Scsi_Host *shost = sdev->host;
1629         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1630         blk_status_t ret;
1631         int reason;
1632
1633         /*
1634          * If the device is not in running state we will reject some or all
1635          * commands.
1636          */
1637         if (unlikely(sdev->sdev_state != SDEV_RUNNING)) {
1638                 ret = scsi_prep_state_check(sdev, req);
1639                 if (ret != BLK_STS_OK)
1640                         goto out_put_budget;
1641         }
1642
1643         ret = BLK_STS_RESOURCE;
1644         if (!scsi_target_queue_ready(shost, sdev))
1645                 goto out_put_budget;
1646         if (!scsi_host_queue_ready(q, shost, sdev, cmd))
1647                 goto out_dec_target_busy;
1648
1649         if (!(req->rq_flags & RQF_DONTPREP)) {
1650                 ret = scsi_mq_prep_fn(req);
1651                 if (ret != BLK_STS_OK)
1652                         goto out_dec_host_busy;
1653                 req->rq_flags |= RQF_DONTPREP;
1654         } else {
1655                 clear_bit(SCMD_STATE_COMPLETE, &cmd->state);
1656                 blk_mq_start_request(req);
1657         }
1658
1659         cmd->flags &= SCMD_PRESERVED_FLAGS;
1660         if (sdev->simple_tags)
1661                 cmd->flags |= SCMD_TAGGED;
1662         if (bd->last)
1663                 cmd->flags |= SCMD_LAST;
1664
1665         scsi_init_cmd_errh(cmd);
1666         cmd->scsi_done = scsi_mq_done;
1667
1668         reason = scsi_dispatch_cmd(cmd);
1669         if (reason) {
1670                 scsi_set_blocked(cmd, reason);
1671                 ret = BLK_STS_RESOURCE;
1672                 goto out_dec_host_busy;
1673         }
1674
1675         return BLK_STS_OK;
1676
1677 out_dec_host_busy:
1678         scsi_dec_host_busy(shost, cmd);
1679 out_dec_target_busy:
1680         if (scsi_target(sdev)->can_queue > 0)
1681                 atomic_dec(&scsi_target(sdev)->target_busy);
1682 out_put_budget:
1683         scsi_mq_put_budget(hctx);
1684         switch (ret) {
1685         case BLK_STS_OK:
1686                 break;
1687         case BLK_STS_RESOURCE:
1688         case BLK_STS_ZONE_RESOURCE:
1689                 if (atomic_read(&sdev->device_busy) ||
1690                     scsi_device_blocked(sdev))
1691                         ret = BLK_STS_DEV_RESOURCE;
1692                 break;
1693         default:
1694                 if (unlikely(!scsi_device_online(sdev)))
1695                         scsi_req(req)->result = DID_NO_CONNECT << 16;
1696                 else
1697                         scsi_req(req)->result = DID_ERROR << 16;
1698                 /*
1699                  * Make sure to release all allocated resources when
1700                  * we hit an error, as we will never see this command
1701                  * again.
1702                  */
1703                 if (req->rq_flags & RQF_DONTPREP)
1704                         scsi_mq_uninit_cmd(cmd);
1705                 break;
1706         }
1707         return ret;
1708 }
1709
1710 static enum blk_eh_timer_return scsi_timeout(struct request *req,
1711                 bool reserved)
1712 {
1713         if (reserved)
1714                 return BLK_EH_RESET_TIMER;
1715         return scsi_times_out(req);
1716 }
1717
1718 static int scsi_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
1719                                 unsigned int hctx_idx, unsigned int numa_node)
1720 {
1721         struct Scsi_Host *shost = set->driver_data;
1722         const bool unchecked_isa_dma = shost->unchecked_isa_dma;
1723         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1724         struct scatterlist *sg;
1725         int ret = 0;
1726
1727         if (unchecked_isa_dma)
1728                 cmd->flags |= SCMD_UNCHECKED_ISA_DMA;
1729         cmd->sense_buffer = scsi_alloc_sense_buffer(unchecked_isa_dma,
1730                                                     GFP_KERNEL, numa_node);
1731         if (!cmd->sense_buffer)
1732                 return -ENOMEM;
1733         cmd->req.sense = cmd->sense_buffer;
1734
1735         if (scsi_host_get_prot(shost)) {
1736                 sg = (void *)cmd + sizeof(struct scsi_cmnd) +
1737                         shost->hostt->cmd_size;
1738                 cmd->prot_sdb = (void *)sg + scsi_mq_inline_sgl_size(shost);
1739         }
1740
1741         if (shost->hostt->init_cmd_priv) {
1742                 ret = shost->hostt->init_cmd_priv(shost, cmd);
1743                 if (ret < 0)
1744                         scsi_free_sense_buffer(unchecked_isa_dma,
1745                                                cmd->sense_buffer);
1746         }
1747
1748         return ret;
1749 }
1750
1751 static void scsi_mq_exit_request(struct blk_mq_tag_set *set, struct request *rq,
1752                                  unsigned int hctx_idx)
1753 {
1754         struct Scsi_Host *shost = set->driver_data;
1755         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1756
1757         if (shost->hostt->exit_cmd_priv)
1758                 shost->hostt->exit_cmd_priv(shost, cmd);
1759         scsi_free_sense_buffer(cmd->flags & SCMD_UNCHECKED_ISA_DMA,
1760                                cmd->sense_buffer);
1761 }
1762
1763 static int scsi_map_queues(struct blk_mq_tag_set *set)
1764 {
1765         struct Scsi_Host *shost = container_of(set, struct Scsi_Host, tag_set);
1766
1767         if (shost->hostt->map_queues)
1768                 return shost->hostt->map_queues(shost);
1769         return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
1770 }
1771
1772 void __scsi_init_queue(struct Scsi_Host *shost, struct request_queue *q)
1773 {
1774         struct device *dev = shost->dma_dev;
1775
1776         /*
1777          * this limit is imposed by hardware restrictions
1778          */
1779         blk_queue_max_segments(q, min_t(unsigned short, shost->sg_tablesize,
1780                                         SG_MAX_SEGMENTS));
1781
1782         if (scsi_host_prot_dma(shost)) {
1783                 shost->sg_prot_tablesize =
1784                         min_not_zero(shost->sg_prot_tablesize,
1785                                      (unsigned short)SCSI_MAX_PROT_SG_SEGMENTS);
1786                 BUG_ON(shost->sg_prot_tablesize < shost->sg_tablesize);
1787                 blk_queue_max_integrity_segments(q, shost->sg_prot_tablesize);
1788         }
1789
1790         if (dev->dma_mask) {
1791                 shost->max_sectors = min_t(unsigned int, shost->max_sectors,
1792                                 dma_max_mapping_size(dev) >> SECTOR_SHIFT);
1793         }
1794         blk_queue_max_hw_sectors(q, shost->max_sectors);
1795         if (shost->unchecked_isa_dma)
1796                 blk_queue_bounce_limit(q, BLK_BOUNCE_ISA);
1797         blk_queue_segment_boundary(q, shost->dma_boundary);
1798         dma_set_seg_boundary(dev, shost->dma_boundary);
1799
1800         blk_queue_max_segment_size(q, shost->max_segment_size);
1801         blk_queue_virt_boundary(q, shost->virt_boundary_mask);
1802         dma_set_max_seg_size(dev, queue_max_segment_size(q));
1803
1804         /*
1805          * Set a reasonable default alignment:  The larger of 32-byte (dword),
1806          * which is a common minimum for HBAs, and the minimum DMA alignment,
1807          * which is set by the platform.
1808          *
1809          * Devices that require a bigger alignment can increase it later.
1810          */
1811         blk_queue_dma_alignment(q, max(4, dma_get_cache_alignment()) - 1);
1812 }
1813 EXPORT_SYMBOL_GPL(__scsi_init_queue);
1814
1815 static const struct blk_mq_ops scsi_mq_ops_no_commit = {
1816         .get_budget     = scsi_mq_get_budget,
1817         .put_budget     = scsi_mq_put_budget,
1818         .queue_rq       = scsi_queue_rq,
1819         .complete       = scsi_softirq_done,
1820         .timeout        = scsi_timeout,
1821 #ifdef CONFIG_BLK_DEBUG_FS
1822         .show_rq        = scsi_show_rq,
1823 #endif
1824         .init_request   = scsi_mq_init_request,
1825         .exit_request   = scsi_mq_exit_request,
1826         .initialize_rq_fn = scsi_initialize_rq,
1827         .cleanup_rq     = scsi_cleanup_rq,
1828         .busy           = scsi_mq_lld_busy,
1829         .map_queues     = scsi_map_queues,
1830 };
1831
1832
1833 static void scsi_commit_rqs(struct blk_mq_hw_ctx *hctx)
1834 {
1835         struct request_queue *q = hctx->queue;
1836         struct scsi_device *sdev = q->queuedata;
1837         struct Scsi_Host *shost = sdev->host;
1838
1839         shost->hostt->commit_rqs(shost, hctx->queue_num);
1840 }
1841
1842 static const struct blk_mq_ops scsi_mq_ops = {
1843         .get_budget     = scsi_mq_get_budget,
1844         .put_budget     = scsi_mq_put_budget,
1845         .queue_rq       = scsi_queue_rq,
1846         .commit_rqs     = scsi_commit_rqs,
1847         .complete       = scsi_softirq_done,
1848         .timeout        = scsi_timeout,
1849 #ifdef CONFIG_BLK_DEBUG_FS
1850         .show_rq        = scsi_show_rq,
1851 #endif
1852         .init_request   = scsi_mq_init_request,
1853         .exit_request   = scsi_mq_exit_request,
1854         .initialize_rq_fn = scsi_initialize_rq,
1855         .cleanup_rq     = scsi_cleanup_rq,
1856         .busy           = scsi_mq_lld_busy,
1857         .map_queues     = scsi_map_queues,
1858 };
1859
1860 struct request_queue *scsi_mq_alloc_queue(struct scsi_device *sdev)
1861 {
1862         sdev->request_queue = blk_mq_init_queue(&sdev->host->tag_set);
1863         if (IS_ERR(sdev->request_queue))
1864                 return NULL;
1865
1866         sdev->request_queue->queuedata = sdev;
1867         __scsi_init_queue(sdev->host, sdev->request_queue);
1868         blk_queue_flag_set(QUEUE_FLAG_SCSI_PASSTHROUGH, sdev->request_queue);
1869         return sdev->request_queue;
1870 }
1871
1872 int scsi_mq_setup_tags(struct Scsi_Host *shost)
1873 {
1874         unsigned int cmd_size, sgl_size;
1875         struct blk_mq_tag_set *tag_set = &shost->tag_set;
1876
1877         sgl_size = max_t(unsigned int, sizeof(struct scatterlist),
1878                                 scsi_mq_inline_sgl_size(shost));
1879         cmd_size = sizeof(struct scsi_cmnd) + shost->hostt->cmd_size + sgl_size;
1880         if (scsi_host_get_prot(shost))
1881                 cmd_size += sizeof(struct scsi_data_buffer) +
1882                         sizeof(struct scatterlist) * SCSI_INLINE_PROT_SG_CNT;
1883
1884         memset(tag_set, 0, sizeof(*tag_set));
1885         if (shost->hostt->commit_rqs)
1886                 tag_set->ops = &scsi_mq_ops;
1887         else
1888                 tag_set->ops = &scsi_mq_ops_no_commit;
1889         tag_set->nr_hw_queues = shost->nr_hw_queues ? : 1;
1890         tag_set->queue_depth = shost->can_queue;
1891         tag_set->cmd_size = cmd_size;
1892         tag_set->numa_node = NUMA_NO_NODE;
1893         tag_set->flags = BLK_MQ_F_SHOULD_MERGE;
1894         tag_set->flags |=
1895                 BLK_ALLOC_POLICY_TO_MQ_FLAG(shost->hostt->tag_alloc_policy);
1896         tag_set->driver_data = shost;
1897
1898         return blk_mq_alloc_tag_set(tag_set);
1899 }
1900
1901 void scsi_mq_destroy_tags(struct Scsi_Host *shost)
1902 {
1903         blk_mq_free_tag_set(&shost->tag_set);
1904 }
1905
1906 /**
1907  * scsi_device_from_queue - return sdev associated with a request_queue
1908  * @q: The request queue to return the sdev from
1909  *
1910  * Return the sdev associated with a request queue or NULL if the
1911  * request_queue does not reference a SCSI device.
1912  */
1913 struct scsi_device *scsi_device_from_queue(struct request_queue *q)
1914 {
1915         struct scsi_device *sdev = NULL;
1916
1917         if (q->mq_ops == &scsi_mq_ops_no_commit ||
1918             q->mq_ops == &scsi_mq_ops)
1919                 sdev = q->queuedata;
1920         if (!sdev || !get_device(&sdev->sdev_gendev))
1921                 sdev = NULL;
1922
1923         return sdev;
1924 }
1925 EXPORT_SYMBOL_GPL(scsi_device_from_queue);
1926
1927 /**
1928  * scsi_block_requests - Utility function used by low-level drivers to prevent
1929  * further commands from being queued to the device.
1930  * @shost:  host in question
1931  *
1932  * There is no timer nor any other means by which the requests get unblocked
1933  * other than the low-level driver calling scsi_unblock_requests().
1934  */
1935 void scsi_block_requests(struct Scsi_Host *shost)
1936 {
1937         shost->host_self_blocked = 1;
1938 }
1939 EXPORT_SYMBOL(scsi_block_requests);
1940
1941 /**
1942  * scsi_unblock_requests - Utility function used by low-level drivers to allow
1943  * further commands to be queued to the device.
1944  * @shost:  host in question
1945  *
1946  * There is no timer nor any other means by which the requests get unblocked
1947  * other than the low-level driver calling scsi_unblock_requests(). This is done
1948  * as an API function so that changes to the internals of the scsi mid-layer
1949  * won't require wholesale changes to drivers that use this feature.
1950  */
1951 void scsi_unblock_requests(struct Scsi_Host *shost)
1952 {
1953         shost->host_self_blocked = 0;
1954         scsi_run_host_queues(shost);
1955 }
1956 EXPORT_SYMBOL(scsi_unblock_requests);
1957
1958 int __init scsi_init_queue(void)
1959 {
1960         scsi_sdb_cache = kmem_cache_create("scsi_data_buffer",
1961                                            sizeof(struct scsi_data_buffer),
1962                                            0, 0, NULL);
1963         if (!scsi_sdb_cache) {
1964                 printk(KERN_ERR "SCSI: can't init scsi sdb cache\n");
1965                 return -ENOMEM;
1966         }
1967
1968         return 0;
1969 }
1970
1971 void scsi_exit_queue(void)
1972 {
1973         kmem_cache_destroy(scsi_sense_cache);
1974         kmem_cache_destroy(scsi_sense_isadma_cache);
1975         kmem_cache_destroy(scsi_sdb_cache);
1976 }
1977
1978 /**
1979  *      scsi_mode_select - issue a mode select
1980  *      @sdev:  SCSI device to be queried
1981  *      @pf:    Page format bit (1 == standard, 0 == vendor specific)
1982  *      @sp:    Save page bit (0 == don't save, 1 == save)
1983  *      @modepage: mode page being requested
1984  *      @buffer: request buffer (may not be smaller than eight bytes)
1985  *      @len:   length of request buffer.
1986  *      @timeout: command timeout
1987  *      @retries: number of retries before failing
1988  *      @data: returns a structure abstracting the mode header data
1989  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
1990  *              must be SCSI_SENSE_BUFFERSIZE big.
1991  *
1992  *      Returns zero if successful; negative error number or scsi
1993  *      status on error
1994  *
1995  */
1996 int
1997 scsi_mode_select(struct scsi_device *sdev, int pf, int sp, int modepage,
1998                  unsigned char *buffer, int len, int timeout, int retries,
1999                  struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
2000 {
2001         unsigned char cmd[10];
2002         unsigned char *real_buffer;
2003         int ret;
2004
2005         memset(cmd, 0, sizeof(cmd));
2006         cmd[1] = (pf ? 0x10 : 0) | (sp ? 0x01 : 0);
2007
2008         if (sdev->use_10_for_ms) {
2009                 if (len > 65535)
2010                         return -EINVAL;
2011                 real_buffer = kmalloc(8 + len, GFP_KERNEL);
2012                 if (!real_buffer)
2013                         return -ENOMEM;
2014                 memcpy(real_buffer + 8, buffer, len);
2015                 len += 8;
2016                 real_buffer[0] = 0;
2017                 real_buffer[1] = 0;
2018                 real_buffer[2] = data->medium_type;
2019                 real_buffer[3] = data->device_specific;
2020                 real_buffer[4] = data->longlba ? 0x01 : 0;
2021                 real_buffer[5] = 0;
2022                 real_buffer[6] = data->block_descriptor_length >> 8;
2023                 real_buffer[7] = data->block_descriptor_length;
2024
2025                 cmd[0] = MODE_SELECT_10;
2026                 cmd[7] = len >> 8;
2027                 cmd[8] = len;
2028         } else {
2029                 if (len > 255 || data->block_descriptor_length > 255 ||
2030                     data->longlba)
2031                         return -EINVAL;
2032
2033                 real_buffer = kmalloc(4 + len, GFP_KERNEL);
2034                 if (!real_buffer)
2035                         return -ENOMEM;
2036                 memcpy(real_buffer + 4, buffer, len);
2037                 len += 4;
2038                 real_buffer[0] = 0;
2039                 real_buffer[1] = data->medium_type;
2040                 real_buffer[2] = data->device_specific;
2041                 real_buffer[3] = data->block_descriptor_length;
2042                 
2043
2044                 cmd[0] = MODE_SELECT;
2045                 cmd[4] = len;
2046         }
2047
2048         ret = scsi_execute_req(sdev, cmd, DMA_TO_DEVICE, real_buffer, len,
2049                                sshdr, timeout, retries, NULL);
2050         kfree(real_buffer);
2051         return ret;
2052 }
2053 EXPORT_SYMBOL_GPL(scsi_mode_select);
2054
2055 /**
2056  *      scsi_mode_sense - issue a mode sense, falling back from 10 to six bytes if necessary.
2057  *      @sdev:  SCSI device to be queried
2058  *      @dbd:   set if mode sense will allow block descriptors to be returned
2059  *      @modepage: mode page being requested
2060  *      @buffer: request buffer (may not be smaller than eight bytes)
2061  *      @len:   length of request buffer.
2062  *      @timeout: command timeout
2063  *      @retries: number of retries before failing
2064  *      @data: returns a structure abstracting the mode header data
2065  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
2066  *              must be SCSI_SENSE_BUFFERSIZE big.
2067  *
2068  *      Returns zero if unsuccessful, or the header offset (either 4
2069  *      or 8 depending on whether a six or ten byte command was
2070  *      issued) if successful.
2071  */
2072 int
2073 scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
2074                   unsigned char *buffer, int len, int timeout, int retries,
2075                   struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
2076 {
2077         unsigned char cmd[12];
2078         int use_10_for_ms;
2079         int header_length;
2080         int result, retry_count = retries;
2081         struct scsi_sense_hdr my_sshdr;
2082
2083         memset(data, 0, sizeof(*data));
2084         memset(&cmd[0], 0, 12);
2085
2086         dbd = sdev->set_dbd_for_ms ? 8 : dbd;
2087         cmd[1] = dbd & 0x18;    /* allows DBD and LLBA bits */
2088         cmd[2] = modepage;
2089
2090         /* caller might not be interested in sense, but we need it */
2091         if (!sshdr)
2092                 sshdr = &my_sshdr;
2093
2094  retry:
2095         use_10_for_ms = sdev->use_10_for_ms;
2096
2097         if (use_10_for_ms) {
2098                 if (len < 8)
2099                         len = 8;
2100
2101                 cmd[0] = MODE_SENSE_10;
2102                 cmd[8] = len;
2103                 header_length = 8;
2104         } else {
2105                 if (len < 4)
2106                         len = 4;
2107
2108                 cmd[0] = MODE_SENSE;
2109                 cmd[4] = len;
2110                 header_length = 4;
2111         }
2112
2113         memset(buffer, 0, len);
2114
2115         result = scsi_execute_req(sdev, cmd, DMA_FROM_DEVICE, buffer, len,
2116                                   sshdr, timeout, retries, NULL);
2117
2118         /* This code looks awful: what it's doing is making sure an
2119          * ILLEGAL REQUEST sense return identifies the actual command
2120          * byte as the problem.  MODE_SENSE commands can return
2121          * ILLEGAL REQUEST if the code page isn't supported */
2122
2123         if (use_10_for_ms && !scsi_status_is_good(result) &&
2124             driver_byte(result) == DRIVER_SENSE) {
2125                 if (scsi_sense_valid(sshdr)) {
2126                         if ((sshdr->sense_key == ILLEGAL_REQUEST) &&
2127                             (sshdr->asc == 0x20) && (sshdr->ascq == 0)) {
2128                                 /* 
2129                                  * Invalid command operation code
2130                                  */
2131                                 sdev->use_10_for_ms = 0;
2132                                 goto retry;
2133                         }
2134                 }
2135         }
2136
2137         if(scsi_status_is_good(result)) {
2138                 if (unlikely(buffer[0] == 0x86 && buffer[1] == 0x0b &&
2139                              (modepage == 6 || modepage == 8))) {
2140                         /* Initio breakage? */
2141                         header_length = 0;
2142                         data->length = 13;
2143                         data->medium_type = 0;
2144                         data->device_specific = 0;
2145                         data->longlba = 0;
2146                         data->block_descriptor_length = 0;
2147                 } else if(use_10_for_ms) {
2148                         data->length = buffer[0]*256 + buffer[1] + 2;
2149                         data->medium_type = buffer[2];
2150                         data->device_specific = buffer[3];
2151                         data->longlba = buffer[4] & 0x01;
2152                         data->block_descriptor_length = buffer[6]*256
2153                                 + buffer[7];
2154                 } else {
2155                         data->length = buffer[0] + 1;
2156                         data->medium_type = buffer[1];
2157                         data->device_specific = buffer[2];
2158                         data->block_descriptor_length = buffer[3];
2159                 }
2160                 data->header_length = header_length;
2161         } else if ((status_byte(result) == CHECK_CONDITION) &&
2162                    scsi_sense_valid(sshdr) &&
2163                    sshdr->sense_key == UNIT_ATTENTION && retry_count) {
2164                 retry_count--;
2165                 goto retry;
2166         }
2167
2168         return result;
2169 }
2170 EXPORT_SYMBOL(scsi_mode_sense);
2171
2172 /**
2173  *      scsi_test_unit_ready - test if unit is ready
2174  *      @sdev:  scsi device to change the state of.
2175  *      @timeout: command timeout
2176  *      @retries: number of retries before failing
2177  *      @sshdr: outpout pointer for decoded sense information.
2178  *
2179  *      Returns zero if unsuccessful or an error if TUR failed.  For
2180  *      removable media, UNIT_ATTENTION sets ->changed flag.
2181  **/
2182 int
2183 scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries,
2184                      struct scsi_sense_hdr *sshdr)
2185 {
2186         char cmd[] = {
2187                 TEST_UNIT_READY, 0, 0, 0, 0, 0,
2188         };
2189         int result;
2190
2191         /* try to eat the UNIT_ATTENTION if there are enough retries */
2192         do {
2193                 result = scsi_execute_req(sdev, cmd, DMA_NONE, NULL, 0, sshdr,
2194                                           timeout, 1, NULL);
2195                 if (sdev->removable && scsi_sense_valid(sshdr) &&
2196                     sshdr->sense_key == UNIT_ATTENTION)
2197                         sdev->changed = 1;
2198         } while (scsi_sense_valid(sshdr) &&
2199                  sshdr->sense_key == UNIT_ATTENTION && --retries);
2200
2201         return result;
2202 }
2203 EXPORT_SYMBOL(scsi_test_unit_ready);
2204
2205 /**
2206  *      scsi_device_set_state - Take the given device through the device state model.
2207  *      @sdev:  scsi device to change the state of.
2208  *      @state: state to change to.
2209  *
2210  *      Returns zero if successful or an error if the requested
2211  *      transition is illegal.
2212  */
2213 int
2214 scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
2215 {
2216         enum scsi_device_state oldstate = sdev->sdev_state;
2217
2218         if (state == oldstate)
2219                 return 0;
2220
2221         switch (state) {
2222         case SDEV_CREATED:
2223                 switch (oldstate) {
2224                 case SDEV_CREATED_BLOCK:
2225                         break;
2226                 default:
2227                         goto illegal;
2228                 }
2229                 break;
2230                         
2231         case SDEV_RUNNING:
2232                 switch (oldstate) {
2233                 case SDEV_CREATED:
2234                 case SDEV_OFFLINE:
2235                 case SDEV_TRANSPORT_OFFLINE:
2236                 case SDEV_QUIESCE:
2237                 case SDEV_BLOCK:
2238                         break;
2239                 default:
2240                         goto illegal;
2241                 }
2242                 break;
2243
2244         case SDEV_QUIESCE:
2245                 switch (oldstate) {
2246                 case SDEV_RUNNING:
2247                 case SDEV_OFFLINE:
2248                 case SDEV_TRANSPORT_OFFLINE:
2249                         break;
2250                 default:
2251                         goto illegal;
2252                 }
2253                 break;
2254
2255         case SDEV_OFFLINE:
2256         case SDEV_TRANSPORT_OFFLINE:
2257                 switch (oldstate) {
2258                 case SDEV_CREATED:
2259                 case SDEV_RUNNING:
2260                 case SDEV_QUIESCE:
2261                 case SDEV_BLOCK:
2262                         break;
2263                 default:
2264                         goto illegal;
2265                 }
2266                 break;
2267
2268         case SDEV_BLOCK:
2269                 switch (oldstate) {
2270                 case SDEV_RUNNING:
2271                 case SDEV_CREATED_BLOCK:
2272                 case SDEV_QUIESCE:
2273                 case SDEV_OFFLINE:
2274                         break;
2275                 default:
2276                         goto illegal;
2277                 }
2278                 break;
2279
2280         case SDEV_CREATED_BLOCK:
2281                 switch (oldstate) {
2282                 case SDEV_CREATED:
2283                         break;
2284                 default:
2285                         goto illegal;
2286                 }
2287                 break;
2288
2289         case SDEV_CANCEL:
2290                 switch (oldstate) {
2291                 case SDEV_CREATED:
2292                 case SDEV_RUNNING:
2293                 case SDEV_QUIESCE:
2294                 case SDEV_OFFLINE:
2295                 case SDEV_TRANSPORT_OFFLINE:
2296                         break;
2297                 default:
2298                         goto illegal;
2299                 }
2300                 break;
2301
2302         case SDEV_DEL:
2303                 switch (oldstate) {
2304                 case SDEV_CREATED:
2305                 case SDEV_RUNNING:
2306                 case SDEV_OFFLINE:
2307                 case SDEV_TRANSPORT_OFFLINE:
2308                 case SDEV_CANCEL:
2309                 case SDEV_BLOCK:
2310                 case SDEV_CREATED_BLOCK:
2311                         break;
2312                 default:
2313                         goto illegal;
2314                 }
2315                 break;
2316
2317         }
2318         sdev->offline_already = false;
2319         sdev->sdev_state = state;
2320         return 0;
2321
2322  illegal:
2323         SCSI_LOG_ERROR_RECOVERY(1,
2324                                 sdev_printk(KERN_ERR, sdev,
2325                                             "Illegal state transition %s->%s",
2326                                             scsi_device_state_name(oldstate),
2327                                             scsi_device_state_name(state))
2328                                 );
2329         return -EINVAL;
2330 }
2331 EXPORT_SYMBOL(scsi_device_set_state);
2332
2333 /**
2334  *      sdev_evt_emit - emit a single SCSI device uevent
2335  *      @sdev: associated SCSI device
2336  *      @evt: event to emit
2337  *
2338  *      Send a single uevent (scsi_event) to the associated scsi_device.
2339  */
2340 static void scsi_evt_emit(struct scsi_device *sdev, struct scsi_event *evt)
2341 {
2342         int idx = 0;
2343         char *envp[3];
2344
2345         switch (evt->evt_type) {
2346         case SDEV_EVT_MEDIA_CHANGE:
2347                 envp[idx++] = "SDEV_MEDIA_CHANGE=1";
2348                 break;
2349         case SDEV_EVT_INQUIRY_CHANGE_REPORTED:
2350                 scsi_rescan_device(&sdev->sdev_gendev);
2351                 envp[idx++] = "SDEV_UA=INQUIRY_DATA_HAS_CHANGED";
2352                 break;
2353         case SDEV_EVT_CAPACITY_CHANGE_REPORTED:
2354                 envp[idx++] = "SDEV_UA=CAPACITY_DATA_HAS_CHANGED";
2355                 break;
2356         case SDEV_EVT_SOFT_THRESHOLD_REACHED_REPORTED:
2357                envp[idx++] = "SDEV_UA=THIN_PROVISIONING_SOFT_THRESHOLD_REACHED";
2358                 break;
2359         case SDEV_EVT_MODE_PARAMETER_CHANGE_REPORTED:
2360                 envp[idx++] = "SDEV_UA=MODE_PARAMETERS_CHANGED";
2361                 break;
2362         case SDEV_EVT_LUN_CHANGE_REPORTED:
2363                 envp[idx++] = "SDEV_UA=REPORTED_LUNS_DATA_HAS_CHANGED";
2364                 break;
2365         case SDEV_EVT_ALUA_STATE_CHANGE_REPORTED:
2366                 envp[idx++] = "SDEV_UA=ASYMMETRIC_ACCESS_STATE_CHANGED";
2367                 break;
2368         case SDEV_EVT_POWER_ON_RESET_OCCURRED:
2369                 envp[idx++] = "SDEV_UA=POWER_ON_RESET_OCCURRED";
2370                 break;
2371         default:
2372                 /* do nothing */
2373                 break;
2374         }
2375
2376         envp[idx++] = NULL;
2377
2378         kobject_uevent_env(&sdev->sdev_gendev.kobj, KOBJ_CHANGE, envp);
2379 }
2380
2381 /**
2382  *      sdev_evt_thread - send a uevent for each scsi event
2383  *      @work: work struct for scsi_device
2384  *
2385  *      Dispatch queued events to their associated scsi_device kobjects
2386  *      as uevents.
2387  */
2388 void scsi_evt_thread(struct work_struct *work)
2389 {
2390         struct scsi_device *sdev;
2391         enum scsi_device_event evt_type;
2392         LIST_HEAD(event_list);
2393
2394         sdev = container_of(work, struct scsi_device, event_work);
2395
2396         for (evt_type = SDEV_EVT_FIRST; evt_type <= SDEV_EVT_LAST; evt_type++)
2397                 if (test_and_clear_bit(evt_type, sdev->pending_events))
2398                         sdev_evt_send_simple(sdev, evt_type, GFP_KERNEL);
2399
2400         while (1) {
2401                 struct scsi_event *evt;
2402                 struct list_head *this, *tmp;
2403                 unsigned long flags;
2404
2405                 spin_lock_irqsave(&sdev->list_lock, flags);
2406                 list_splice_init(&sdev->event_list, &event_list);
2407                 spin_unlock_irqrestore(&sdev->list_lock, flags);
2408
2409                 if (list_empty(&event_list))
2410                         break;
2411
2412                 list_for_each_safe(this, tmp, &event_list) {
2413                         evt = list_entry(this, struct scsi_event, node);
2414                         list_del(&evt->node);
2415                         scsi_evt_emit(sdev, evt);
2416                         kfree(evt);
2417                 }
2418         }
2419 }
2420
2421 /**
2422  *      sdev_evt_send - send asserted event to uevent thread
2423  *      @sdev: scsi_device event occurred on
2424  *      @evt: event to send
2425  *
2426  *      Assert scsi device event asynchronously.
2427  */
2428 void sdev_evt_send(struct scsi_device *sdev, struct scsi_event *evt)
2429 {
2430         unsigned long flags;
2431
2432 #if 0
2433         /* FIXME: currently this check eliminates all media change events
2434          * for polled devices.  Need to update to discriminate between AN
2435          * and polled events */
2436         if (!test_bit(evt->evt_type, sdev->supported_events)) {
2437                 kfree(evt);
2438                 return;
2439         }
2440 #endif
2441
2442         spin_lock_irqsave(&sdev->list_lock, flags);
2443         list_add_tail(&evt->node, &sdev->event_list);
2444         schedule_work(&sdev->event_work);
2445         spin_unlock_irqrestore(&sdev->list_lock, flags);
2446 }
2447 EXPORT_SYMBOL_GPL(sdev_evt_send);
2448
2449 /**
2450  *      sdev_evt_alloc - allocate a new scsi event
2451  *      @evt_type: type of event to allocate
2452  *      @gfpflags: GFP flags for allocation
2453  *
2454  *      Allocates and returns a new scsi_event.
2455  */
2456 struct scsi_event *sdev_evt_alloc(enum scsi_device_event evt_type,
2457                                   gfp_t gfpflags)
2458 {
2459         struct scsi_event *evt = kzalloc(sizeof(struct scsi_event), gfpflags);
2460         if (!evt)
2461                 return NULL;
2462
2463         evt->evt_type = evt_type;
2464         INIT_LIST_HEAD(&evt->node);
2465
2466         /* evt_type-specific initialization, if any */
2467         switch (evt_type) {
2468         case SDEV_EVT_MEDIA_CHANGE:
2469         case SDEV_EVT_INQUIRY_CHANGE_REPORTED:
2470         case SDEV_EVT_CAPACITY_CHANGE_REPORTED:
2471         case SDEV_EVT_SOFT_THRESHOLD_REACHED_REPORTED:
2472         case SDEV_EVT_MODE_PARAMETER_CHANGE_REPORTED:
2473         case SDEV_EVT_LUN_CHANGE_REPORTED:
2474         case SDEV_EVT_ALUA_STATE_CHANGE_REPORTED:
2475         case SDEV_EVT_POWER_ON_RESET_OCCURRED:
2476         default:
2477                 /* do nothing */
2478                 break;
2479         }
2480
2481         return evt;
2482 }
2483 EXPORT_SYMBOL_GPL(sdev_evt_alloc);
2484
2485 /**
2486  *      sdev_evt_send_simple - send asserted event to uevent thread
2487  *      @sdev: scsi_device event occurred on
2488  *      @evt_type: type of event to send
2489  *      @gfpflags: GFP flags for allocation
2490  *
2491  *      Assert scsi device event asynchronously, given an event type.
2492  */
2493 void sdev_evt_send_simple(struct scsi_device *sdev,
2494                           enum scsi_device_event evt_type, gfp_t gfpflags)
2495 {
2496         struct scsi_event *evt = sdev_evt_alloc(evt_type, gfpflags);
2497         if (!evt) {
2498                 sdev_printk(KERN_ERR, sdev, "event %d eaten due to OOM\n",
2499                             evt_type);
2500                 return;
2501         }
2502
2503         sdev_evt_send(sdev, evt);
2504 }
2505 EXPORT_SYMBOL_GPL(sdev_evt_send_simple);
2506
2507 /**
2508  *      scsi_device_quiesce - Block user issued commands.
2509  *      @sdev:  scsi device to quiesce.
2510  *
2511  *      This works by trying to transition to the SDEV_QUIESCE state
2512  *      (which must be a legal transition).  When the device is in this
2513  *      state, only special requests will be accepted, all others will
2514  *      be deferred.  Since special requests may also be requeued requests,
2515  *      a successful return doesn't guarantee the device will be 
2516  *      totally quiescent.
2517  *
2518  *      Must be called with user context, may sleep.
2519  *
2520  *      Returns zero if unsuccessful or an error if not.
2521  */
2522 int
2523 scsi_device_quiesce(struct scsi_device *sdev)
2524 {
2525         struct request_queue *q = sdev->request_queue;
2526         int err;
2527
2528         /*
2529          * It is allowed to call scsi_device_quiesce() multiple times from
2530          * the same context but concurrent scsi_device_quiesce() calls are
2531          * not allowed.
2532          */
2533         WARN_ON_ONCE(sdev->quiesced_by && sdev->quiesced_by != current);
2534
2535         if (sdev->quiesced_by == current)
2536                 return 0;
2537
2538         blk_set_pm_only(q);
2539
2540         blk_mq_freeze_queue(q);
2541         /*
2542          * Ensure that the effect of blk_set_pm_only() will be visible
2543          * for percpu_ref_tryget() callers that occur after the queue
2544          * unfreeze even if the queue was already frozen before this function
2545          * was called. See also https://lwn.net/Articles/573497/.
2546          */
2547         synchronize_rcu();
2548         blk_mq_unfreeze_queue(q);
2549
2550         mutex_lock(&sdev->state_mutex);
2551         err = scsi_device_set_state(sdev, SDEV_QUIESCE);
2552         if (err == 0)
2553                 sdev->quiesced_by = current;
2554         else
2555                 blk_clear_pm_only(q);
2556         mutex_unlock(&sdev->state_mutex);
2557
2558         return err;
2559 }
2560 EXPORT_SYMBOL(scsi_device_quiesce);
2561
2562 /**
2563  *      scsi_device_resume - Restart user issued commands to a quiesced device.
2564  *      @sdev:  scsi device to resume.
2565  *
2566  *      Moves the device from quiesced back to running and restarts the
2567  *      queues.
2568  *
2569  *      Must be called with user context, may sleep.
2570  */
2571 void scsi_device_resume(struct scsi_device *sdev)
2572 {
2573         /* check if the device state was mutated prior to resume, and if
2574          * so assume the state is being managed elsewhere (for example
2575          * device deleted during suspend)
2576          */
2577         mutex_lock(&sdev->state_mutex);
2578         if (sdev->quiesced_by) {
2579                 sdev->quiesced_by = NULL;
2580                 blk_clear_pm_only(sdev->request_queue);
2581         }
2582         if (sdev->sdev_state == SDEV_QUIESCE)
2583                 scsi_device_set_state(sdev, SDEV_RUNNING);
2584         mutex_unlock(&sdev->state_mutex);
2585 }
2586 EXPORT_SYMBOL(scsi_device_resume);
2587
2588 static void
2589 device_quiesce_fn(struct scsi_device *sdev, void *data)
2590 {
2591         scsi_device_quiesce(sdev);
2592 }
2593
2594 void
2595 scsi_target_quiesce(struct scsi_target *starget)
2596 {
2597         starget_for_each_device(starget, NULL, device_quiesce_fn);
2598 }
2599 EXPORT_SYMBOL(scsi_target_quiesce);
2600
2601 static void
2602 device_resume_fn(struct scsi_device *sdev, void *data)
2603 {
2604         scsi_device_resume(sdev);
2605 }
2606
2607 void
2608 scsi_target_resume(struct scsi_target *starget)
2609 {
2610         starget_for_each_device(starget, NULL, device_resume_fn);
2611 }
2612 EXPORT_SYMBOL(scsi_target_resume);
2613
2614 /**
2615  * scsi_internal_device_block_nowait - try to transition to the SDEV_BLOCK state
2616  * @sdev: device to block
2617  *
2618  * Pause SCSI command processing on the specified device. Does not sleep.
2619  *
2620  * Returns zero if successful or a negative error code upon failure.
2621  *
2622  * Notes:
2623  * This routine transitions the device to the SDEV_BLOCK state (which must be
2624  * a legal transition). When the device is in this state, command processing
2625  * is paused until the device leaves the SDEV_BLOCK state. See also
2626  * scsi_internal_device_unblock_nowait().
2627  */
2628 int scsi_internal_device_block_nowait(struct scsi_device *sdev)
2629 {
2630         struct request_queue *q = sdev->request_queue;
2631         int err = 0;
2632
2633         err = scsi_device_set_state(sdev, SDEV_BLOCK);
2634         if (err) {
2635                 err = scsi_device_set_state(sdev, SDEV_CREATED_BLOCK);
2636
2637                 if (err)
2638                         return err;
2639         }
2640
2641         /* 
2642          * The device has transitioned to SDEV_BLOCK.  Stop the
2643          * block layer from calling the midlayer with this device's
2644          * request queue. 
2645          */
2646         blk_mq_quiesce_queue_nowait(q);
2647         return 0;
2648 }
2649 EXPORT_SYMBOL_GPL(scsi_internal_device_block_nowait);
2650
2651 /**
2652  * scsi_internal_device_block - try to transition to the SDEV_BLOCK state
2653  * @sdev: device to block
2654  *
2655  * Pause SCSI command processing on the specified device and wait until all
2656  * ongoing scsi_request_fn() / scsi_queue_rq() calls have finished. May sleep.
2657  *
2658  * Returns zero if successful or a negative error code upon failure.
2659  *
2660  * Note:
2661  * This routine transitions the device to the SDEV_BLOCK state (which must be
2662  * a legal transition). When the device is in this state, command processing
2663  * is paused until the device leaves the SDEV_BLOCK state. See also
2664  * scsi_internal_device_unblock().
2665  */
2666 static int scsi_internal_device_block(struct scsi_device *sdev)
2667 {
2668         struct request_queue *q = sdev->request_queue;
2669         int err;
2670
2671         mutex_lock(&sdev->state_mutex);
2672         err = scsi_internal_device_block_nowait(sdev);
2673         if (err == 0)
2674                 blk_mq_quiesce_queue(q);
2675         mutex_unlock(&sdev->state_mutex);
2676
2677         return err;
2678 }
2679  
2680 void scsi_start_queue(struct scsi_device *sdev)
2681 {
2682         struct request_queue *q = sdev->request_queue;
2683
2684         blk_mq_unquiesce_queue(q);
2685 }
2686
2687 /**
2688  * scsi_internal_device_unblock_nowait - resume a device after a block request
2689  * @sdev:       device to resume
2690  * @new_state:  state to set the device to after unblocking
2691  *
2692  * Restart the device queue for a previously suspended SCSI device. Does not
2693  * sleep.
2694  *
2695  * Returns zero if successful or a negative error code upon failure.
2696  *
2697  * Notes:
2698  * This routine transitions the device to the SDEV_RUNNING state or to one of
2699  * the offline states (which must be a legal transition) allowing the midlayer
2700  * to goose the queue for this device.
2701  */
2702 int scsi_internal_device_unblock_nowait(struct scsi_device *sdev,
2703                                         enum scsi_device_state new_state)
2704 {
2705         switch (new_state) {
2706         case SDEV_RUNNING:
2707         case SDEV_TRANSPORT_OFFLINE:
2708                 break;
2709         default:
2710                 return -EINVAL;
2711         }
2712
2713         /*
2714          * Try to transition the scsi device to SDEV_RUNNING or one of the
2715          * offlined states and goose the device queue if successful.
2716          */
2717         switch (sdev->sdev_state) {
2718         case SDEV_BLOCK:
2719         case SDEV_TRANSPORT_OFFLINE:
2720                 sdev->sdev_state = new_state;
2721                 break;
2722         case SDEV_CREATED_BLOCK:
2723                 if (new_state == SDEV_TRANSPORT_OFFLINE ||
2724                     new_state == SDEV_OFFLINE)
2725                         sdev->sdev_state = new_state;
2726                 else
2727                         sdev->sdev_state = SDEV_CREATED;
2728                 break;
2729         case SDEV_CANCEL:
2730         case SDEV_OFFLINE:
2731                 break;
2732         default:
2733                 return -EINVAL;
2734         }
2735         scsi_start_queue(sdev);
2736
2737         return 0;
2738 }
2739 EXPORT_SYMBOL_GPL(scsi_internal_device_unblock_nowait);
2740
2741 /**
2742  * scsi_internal_device_unblock - resume a device after a block request
2743  * @sdev:       device to resume
2744  * @new_state:  state to set the device to after unblocking
2745  *
2746  * Restart the device queue for a previously suspended SCSI device. May sleep.
2747  *
2748  * Returns zero if successful or a negative error code upon failure.
2749  *
2750  * Notes:
2751  * This routine transitions the device to the SDEV_RUNNING state or to one of
2752  * the offline states (which must be a legal transition) allowing the midlayer
2753  * to goose the queue for this device.
2754  */
2755 static int scsi_internal_device_unblock(struct scsi_device *sdev,
2756                                         enum scsi_device_state new_state)
2757 {
2758         int ret;
2759
2760         mutex_lock(&sdev->state_mutex);
2761         ret = scsi_internal_device_unblock_nowait(sdev, new_state);
2762         mutex_unlock(&sdev->state_mutex);
2763
2764         return ret;
2765 }
2766
2767 static void
2768 device_block(struct scsi_device *sdev, void *data)
2769 {
2770         int ret;
2771
2772         ret = scsi_internal_device_block(sdev);
2773
2774         WARN_ONCE(ret, "scsi_internal_device_block(%s) failed: ret = %d\n",
2775                   dev_name(&sdev->sdev_gendev), ret);
2776 }
2777
2778 static int
2779 target_block(struct device *dev, void *data)
2780 {
2781         if (scsi_is_target_device(dev))
2782                 starget_for_each_device(to_scsi_target(dev), NULL,
2783                                         device_block);
2784         return 0;
2785 }
2786
2787 void
2788 scsi_target_block(struct device *dev)
2789 {
2790         if (scsi_is_target_device(dev))
2791                 starget_for_each_device(to_scsi_target(dev), NULL,
2792                                         device_block);
2793         else
2794                 device_for_each_child(dev, NULL, target_block);
2795 }
2796 EXPORT_SYMBOL_GPL(scsi_target_block);
2797
2798 static void
2799 device_unblock(struct scsi_device *sdev, void *data)
2800 {
2801         scsi_internal_device_unblock(sdev, *(enum scsi_device_state *)data);
2802 }
2803
2804 static int
2805 target_unblock(struct device *dev, void *data)
2806 {
2807         if (scsi_is_target_device(dev))
2808                 starget_for_each_device(to_scsi_target(dev), data,
2809                                         device_unblock);
2810         return 0;
2811 }
2812
2813 void
2814 scsi_target_unblock(struct device *dev, enum scsi_device_state new_state)
2815 {
2816         if (scsi_is_target_device(dev))
2817                 starget_for_each_device(to_scsi_target(dev), &new_state,
2818                                         device_unblock);
2819         else
2820                 device_for_each_child(dev, &new_state, target_unblock);
2821 }
2822 EXPORT_SYMBOL_GPL(scsi_target_unblock);
2823
2824 int
2825 scsi_host_block(struct Scsi_Host *shost)
2826 {
2827         struct scsi_device *sdev;
2828         int ret = 0;
2829
2830         /*
2831          * Call scsi_internal_device_block_nowait so we can avoid
2832          * calling synchronize_rcu() for each LUN.
2833          */
2834         shost_for_each_device(sdev, shost) {
2835                 mutex_lock(&sdev->state_mutex);
2836                 ret = scsi_internal_device_block_nowait(sdev);
2837                 mutex_unlock(&sdev->state_mutex);
2838                 if (ret)
2839                         break;
2840         }
2841
2842         /*
2843          * SCSI never enables blk-mq's BLK_MQ_F_BLOCKING flag so
2844          * calling synchronize_rcu() once is enough.
2845          */
2846         WARN_ON_ONCE(shost->tag_set.flags & BLK_MQ_F_BLOCKING);
2847
2848         if (!ret)
2849                 synchronize_rcu();
2850
2851         return ret;
2852 }
2853 EXPORT_SYMBOL_GPL(scsi_host_block);
2854
2855 int
2856 scsi_host_unblock(struct Scsi_Host *shost, int new_state)
2857 {
2858         struct scsi_device *sdev;
2859         int ret = 0;
2860
2861         shost_for_each_device(sdev, shost) {
2862                 ret = scsi_internal_device_unblock(sdev, new_state);
2863                 if (ret) {
2864                         scsi_device_put(sdev);
2865                         break;
2866                 }
2867         }
2868         return ret;
2869 }
2870 EXPORT_SYMBOL_GPL(scsi_host_unblock);
2871
2872 /**
2873  * scsi_kmap_atomic_sg - find and atomically map an sg-elemnt
2874  * @sgl:        scatter-gather list
2875  * @sg_count:   number of segments in sg
2876  * @offset:     offset in bytes into sg, on return offset into the mapped area
2877  * @len:        bytes to map, on return number of bytes mapped
2878  *
2879  * Returns virtual address of the start of the mapped page
2880  */
2881 void *scsi_kmap_atomic_sg(struct scatterlist *sgl, int sg_count,
2882                           size_t *offset, size_t *len)
2883 {
2884         int i;
2885         size_t sg_len = 0, len_complete = 0;
2886         struct scatterlist *sg;
2887         struct page *page;
2888
2889         WARN_ON(!irqs_disabled());
2890
2891         for_each_sg(sgl, sg, sg_count, i) {
2892                 len_complete = sg_len; /* Complete sg-entries */
2893                 sg_len += sg->length;
2894                 if (sg_len > *offset)
2895                         break;
2896         }
2897
2898         if (unlikely(i == sg_count)) {
2899                 printk(KERN_ERR "%s: Bytes in sg: %zu, requested offset %zu, "
2900                         "elements %d\n",
2901                        __func__, sg_len, *offset, sg_count);
2902                 WARN_ON(1);
2903                 return NULL;
2904         }
2905
2906         /* Offset starting from the beginning of first page in this sg-entry */
2907         *offset = *offset - len_complete + sg->offset;
2908
2909         /* Assumption: contiguous pages can be accessed as "page + i" */
2910         page = nth_page(sg_page(sg), (*offset >> PAGE_SHIFT));
2911         *offset &= ~PAGE_MASK;
2912
2913         /* Bytes in this sg-entry from *offset to the end of the page */
2914         sg_len = PAGE_SIZE - *offset;
2915         if (*len > sg_len)
2916                 *len = sg_len;
2917
2918         return kmap_atomic(page);
2919 }
2920 EXPORT_SYMBOL(scsi_kmap_atomic_sg);
2921
2922 /**
2923  * scsi_kunmap_atomic_sg - atomically unmap a virtual address, previously mapped with scsi_kmap_atomic_sg
2924  * @virt:       virtual address to be unmapped
2925  */
2926 void scsi_kunmap_atomic_sg(void *virt)
2927 {
2928         kunmap_atomic(virt);
2929 }
2930 EXPORT_SYMBOL(scsi_kunmap_atomic_sg);
2931
2932 void sdev_disable_disk_events(struct scsi_device *sdev)
2933 {
2934         atomic_inc(&sdev->disk_events_disable_depth);
2935 }
2936 EXPORT_SYMBOL(sdev_disable_disk_events);
2937
2938 void sdev_enable_disk_events(struct scsi_device *sdev)
2939 {
2940         if (WARN_ON_ONCE(atomic_read(&sdev->disk_events_disable_depth) <= 0))
2941                 return;
2942         atomic_dec(&sdev->disk_events_disable_depth);
2943 }
2944 EXPORT_SYMBOL(sdev_enable_disk_events);
2945
2946 /**
2947  * scsi_vpd_lun_id - return a unique device identification
2948  * @sdev: SCSI device
2949  * @id:   buffer for the identification
2950  * @id_len:  length of the buffer
2951  *
2952  * Copies a unique device identification into @id based
2953  * on the information in the VPD page 0x83 of the device.
2954  * The string will be formatted as a SCSI name string.
2955  *
2956  * Returns the length of the identification or error on failure.
2957  * If the identifier is longer than the supplied buffer the actual
2958  * identifier length is returned and the buffer is not zero-padded.
2959  */
2960 int scsi_vpd_lun_id(struct scsi_device *sdev, char *id, size_t id_len)
2961 {
2962         u8 cur_id_type = 0xff;
2963         u8 cur_id_size = 0;
2964         const unsigned char *d, *cur_id_str;
2965         const struct scsi_vpd *vpd_pg83;
2966         int id_size = -EINVAL;
2967
2968         rcu_read_lock();
2969         vpd_pg83 = rcu_dereference(sdev->vpd_pg83);
2970         if (!vpd_pg83) {
2971                 rcu_read_unlock();
2972                 return -ENXIO;
2973         }
2974
2975         /*
2976          * Look for the correct descriptor.
2977          * Order of preference for lun descriptor:
2978          * - SCSI name string
2979          * - NAA IEEE Registered Extended
2980          * - EUI-64 based 16-byte
2981          * - EUI-64 based 12-byte
2982          * - NAA IEEE Registered
2983          * - NAA IEEE Extended
2984          * - T10 Vendor ID
2985          * as longer descriptors reduce the likelyhood
2986          * of identification clashes.
2987          */
2988
2989         /* The id string must be at least 20 bytes + terminating NULL byte */
2990         if (id_len < 21) {
2991                 rcu_read_unlock();
2992                 return -EINVAL;
2993         }
2994
2995         memset(id, 0, id_len);
2996         d = vpd_pg83->data + 4;
2997         while (d < vpd_pg83->data + vpd_pg83->len) {
2998                 /* Skip designators not referring to the LUN */
2999                 if ((d[1] & 0x30) != 0x00)
3000                         goto next_desig;
3001
3002                 switch (d[1] & 0xf) {
3003                 case 0x1:
3004                         /* T10 Vendor ID */
3005                         if (cur_id_size > d[3])
3006                                 break;
3007                         /* Prefer anything */
3008                         if (cur_id_type > 0x01 && cur_id_type != 0xff)
3009                                 break;
3010                         cur_id_size = d[3];
3011                         if (cur_id_size + 4 > id_len)
3012                                 cur_id_size = id_len - 4;
3013                         cur_id_str = d + 4;
3014                         cur_id_type = d[1] & 0xf;
3015                         id_size = snprintf(id, id_len, "t10.%*pE",
3016                                            cur_id_size, cur_id_str);
3017                         break;
3018                 case 0x2:
3019                         /* EUI-64 */
3020                         if (cur_id_size > d[3])
3021                                 break;
3022                         /* Prefer NAA IEEE Registered Extended */
3023                         if (cur_id_type == 0x3 &&
3024                             cur_id_size == d[3])
3025                                 break;
3026                         cur_id_size = d[3];
3027                         cur_id_str = d + 4;
3028                         cur_id_type = d[1] & 0xf;
3029                         switch (cur_id_size) {
3030                         case 8:
3031                                 id_size = snprintf(id, id_len,
3032                                                    "eui.%8phN",
3033                                                    cur_id_str);
3034                                 break;
3035                         case 12:
3036                                 id_size = snprintf(id, id_len,
3037                                                    "eui.%12phN",
3038                                                    cur_id_str);
3039                                 break;
3040                         case 16:
3041                                 id_size = snprintf(id, id_len,
3042                                                    "eui.%16phN",
3043                                                    cur_id_str);
3044                                 break;
3045                         default:
3046                                 cur_id_size = 0;
3047                                 break;
3048                         }
3049                         break;
3050                 case 0x3:
3051                         /* NAA */
3052                         if (cur_id_size > d[3])
3053                                 break;
3054                         cur_id_size = d[3];
3055                         cur_id_str = d + 4;
3056                         cur_id_type = d[1] & 0xf;
3057                         switch (cur_id_size) {
3058                         case 8:
3059                                 id_size = snprintf(id, id_len,
3060                                                    "naa.%8phN",
3061                                                    cur_id_str);
3062                                 break;
3063                         case 16:
3064                                 id_size = snprintf(id, id_len,
3065                                                    "naa.%16phN",
3066                                                    cur_id_str);
3067                                 break;
3068                         default:
3069                                 cur_id_size = 0;
3070                                 break;
3071                         }
3072                         break;
3073                 case 0x8:
3074                         /* SCSI name string */
3075                         if (cur_id_size + 4 > d[3])
3076                                 break;
3077                         /* Prefer others for truncated descriptor */
3078                         if (cur_id_size && d[3] > id_len)
3079                                 break;
3080                         cur_id_size = id_size = d[3];
3081                         cur_id_str = d + 4;
3082                         cur_id_type = d[1] & 0xf;
3083                         if (cur_id_size >= id_len)
3084                                 cur_id_size = id_len - 1;
3085                         memcpy(id, cur_id_str, cur_id_size);
3086                         /* Decrease priority for truncated descriptor */
3087                         if (cur_id_size != id_size)
3088                                 cur_id_size = 6;
3089                         break;
3090                 default:
3091                         break;
3092                 }
3093 next_desig:
3094                 d += d[3] + 4;
3095         }
3096         rcu_read_unlock();
3097
3098         return id_size;
3099 }
3100 EXPORT_SYMBOL(scsi_vpd_lun_id);
3101
3102 /*
3103  * scsi_vpd_tpg_id - return a target port group identifier
3104  * @sdev: SCSI device
3105  *
3106  * Returns the Target Port Group identifier from the information
3107  * froom VPD page 0x83 of the device.
3108  *
3109  * Returns the identifier or error on failure.
3110  */
3111 int scsi_vpd_tpg_id(struct scsi_device *sdev, int *rel_id)
3112 {
3113         const unsigned char *d;
3114         const struct scsi_vpd *vpd_pg83;
3115         int group_id = -EAGAIN, rel_port = -1;
3116
3117         rcu_read_lock();
3118         vpd_pg83 = rcu_dereference(sdev->vpd_pg83);
3119         if (!vpd_pg83) {
3120                 rcu_read_unlock();
3121                 return -ENXIO;
3122         }
3123
3124         d = vpd_pg83->data + 4;
3125         while (d < vpd_pg83->data + vpd_pg83->len) {
3126                 switch (d[1] & 0xf) {
3127                 case 0x4:
3128                         /* Relative target port */
3129                         rel_port = get_unaligned_be16(&d[6]);
3130                         break;
3131                 case 0x5:
3132                         /* Target port group */
3133                         group_id = get_unaligned_be16(&d[6]);
3134                         break;
3135                 default:
3136                         break;
3137                 }
3138                 d += d[3] + 4;
3139         }
3140         rcu_read_unlock();
3141
3142         if (group_id >= 0 && rel_id && rel_port != -1)
3143                 *rel_id = rel_port;
3144
3145         return group_id;
3146 }
3147 EXPORT_SYMBOL(scsi_vpd_tpg_id);
This page took 0.211498 seconds and 4 git commands to generate.