scsi: qla2xxx: remove double assignment in qla2x00_update_fcport
[linux-2.6-microblaze.git] / drivers / scsi / scsi_lib.c
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (C) 1999 Eric Youngdale
4  * Copyright (C) 2014 Christoph Hellwig
5  *
6  *  SCSI queueing library.
7  *      Initial versions: Eric Youngdale (eric@andante.org).
8  *                        Based upon conversations with large numbers
9  *                        of people at Linux Expo.
10  */
11
12 #include <linux/bio.h>
13 #include <linux/bitops.h>
14 #include <linux/blkdev.h>
15 #include <linux/completion.h>
16 #include <linux/kernel.h>
17 #include <linux/export.h>
18 #include <linux/init.h>
19 #include <linux/pci.h>
20 #include <linux/delay.h>
21 #include <linux/hardirq.h>
22 #include <linux/scatterlist.h>
23 #include <linux/blk-mq.h>
24 #include <linux/ratelimit.h>
25 #include <asm/unaligned.h>
26
27 #include <scsi/scsi.h>
28 #include <scsi/scsi_cmnd.h>
29 #include <scsi/scsi_dbg.h>
30 #include <scsi/scsi_device.h>
31 #include <scsi/scsi_driver.h>
32 #include <scsi/scsi_eh.h>
33 #include <scsi/scsi_host.h>
34 #include <scsi/scsi_transport.h> /* __scsi_init_queue() */
35 #include <scsi/scsi_dh.h>
36
37 #include <trace/events/scsi.h>
38
39 #include "scsi_debugfs.h"
40 #include "scsi_priv.h"
41 #include "scsi_logging.h"
42
43 static struct kmem_cache *scsi_sdb_cache;
44 static struct kmem_cache *scsi_sense_cache;
45 static struct kmem_cache *scsi_sense_isadma_cache;
46 static DEFINE_MUTEX(scsi_sense_cache_mutex);
47
48 static void scsi_mq_uninit_cmd(struct scsi_cmnd *cmd);
49
50 static inline struct kmem_cache *
51 scsi_select_sense_cache(bool unchecked_isa_dma)
52 {
53         return unchecked_isa_dma ? scsi_sense_isadma_cache : scsi_sense_cache;
54 }
55
56 static void scsi_free_sense_buffer(bool unchecked_isa_dma,
57                                    unsigned char *sense_buffer)
58 {
59         kmem_cache_free(scsi_select_sense_cache(unchecked_isa_dma),
60                         sense_buffer);
61 }
62
63 static unsigned char *scsi_alloc_sense_buffer(bool unchecked_isa_dma,
64         gfp_t gfp_mask, int numa_node)
65 {
66         return kmem_cache_alloc_node(scsi_select_sense_cache(unchecked_isa_dma),
67                                      gfp_mask, numa_node);
68 }
69
70 int scsi_init_sense_cache(struct Scsi_Host *shost)
71 {
72         struct kmem_cache *cache;
73         int ret = 0;
74
75         cache = scsi_select_sense_cache(shost->unchecked_isa_dma);
76         if (cache)
77                 return 0;
78
79         mutex_lock(&scsi_sense_cache_mutex);
80         if (shost->unchecked_isa_dma) {
81                 scsi_sense_isadma_cache =
82                         kmem_cache_create("scsi_sense_cache(DMA)",
83                                 SCSI_SENSE_BUFFERSIZE, 0,
84                                 SLAB_HWCACHE_ALIGN | SLAB_CACHE_DMA, NULL);
85                 if (!scsi_sense_isadma_cache)
86                         ret = -ENOMEM;
87         } else {
88                 scsi_sense_cache =
89                         kmem_cache_create_usercopy("scsi_sense_cache",
90                                 SCSI_SENSE_BUFFERSIZE, 0, SLAB_HWCACHE_ALIGN,
91                                 0, SCSI_SENSE_BUFFERSIZE, NULL);
92                 if (!scsi_sense_cache)
93                         ret = -ENOMEM;
94         }
95
96         mutex_unlock(&scsi_sense_cache_mutex);
97         return ret;
98 }
99
100 /*
101  * When to reinvoke queueing after a resource shortage. It's 3 msecs to
102  * not change behaviour from the previous unplug mechanism, experimentation
103  * may prove this needs changing.
104  */
105 #define SCSI_QUEUE_DELAY        3
106
107 static void
108 scsi_set_blocked(struct scsi_cmnd *cmd, int reason)
109 {
110         struct Scsi_Host *host = cmd->device->host;
111         struct scsi_device *device = cmd->device;
112         struct scsi_target *starget = scsi_target(device);
113
114         /*
115          * Set the appropriate busy bit for the device/host.
116          *
117          * If the host/device isn't busy, assume that something actually
118          * completed, and that we should be able to queue a command now.
119          *
120          * Note that the prior mid-layer assumption that any host could
121          * always queue at least one command is now broken.  The mid-layer
122          * will implement a user specifiable stall (see
123          * scsi_host.max_host_blocked and scsi_device.max_device_blocked)
124          * if a command is requeued with no other commands outstanding
125          * either for the device or for the host.
126          */
127         switch (reason) {
128         case SCSI_MLQUEUE_HOST_BUSY:
129                 atomic_set(&host->host_blocked, host->max_host_blocked);
130                 break;
131         case SCSI_MLQUEUE_DEVICE_BUSY:
132         case SCSI_MLQUEUE_EH_RETRY:
133                 atomic_set(&device->device_blocked,
134                            device->max_device_blocked);
135                 break;
136         case SCSI_MLQUEUE_TARGET_BUSY:
137                 atomic_set(&starget->target_blocked,
138                            starget->max_target_blocked);
139                 break;
140         }
141 }
142
143 static void scsi_mq_requeue_cmd(struct scsi_cmnd *cmd)
144 {
145         if (cmd->request->rq_flags & RQF_DONTPREP) {
146                 cmd->request->rq_flags &= ~RQF_DONTPREP;
147                 scsi_mq_uninit_cmd(cmd);
148         } else {
149                 WARN_ON_ONCE(true);
150         }
151         blk_mq_requeue_request(cmd->request, true);
152 }
153
154 /**
155  * __scsi_queue_insert - private queue insertion
156  * @cmd: The SCSI command being requeued
157  * @reason:  The reason for the requeue
158  * @unbusy: Whether the queue should be unbusied
159  *
160  * This is a private queue insertion.  The public interface
161  * scsi_queue_insert() always assumes the queue should be unbusied
162  * because it's always called before the completion.  This function is
163  * for a requeue after completion, which should only occur in this
164  * file.
165  */
166 static void __scsi_queue_insert(struct scsi_cmnd *cmd, int reason, bool unbusy)
167 {
168         struct scsi_device *device = cmd->device;
169
170         SCSI_LOG_MLQUEUE(1, scmd_printk(KERN_INFO, cmd,
171                 "Inserting command %p into mlqueue\n", cmd));
172
173         scsi_set_blocked(cmd, reason);
174
175         /*
176          * Decrement the counters, since these commands are no longer
177          * active on the host/device.
178          */
179         if (unbusy)
180                 scsi_device_unbusy(device);
181
182         /*
183          * Requeue this command.  It will go before all other commands
184          * that are already in the queue. Schedule requeue work under
185          * lock such that the kblockd_schedule_work() call happens
186          * before blk_cleanup_queue() finishes.
187          */
188         cmd->result = 0;
189
190         blk_mq_requeue_request(cmd->request, true);
191 }
192
193 /*
194  * Function:    scsi_queue_insert()
195  *
196  * Purpose:     Insert a command in the midlevel queue.
197  *
198  * Arguments:   cmd    - command that we are adding to queue.
199  *              reason - why we are inserting command to queue.
200  *
201  * Lock status: Assumed that lock is not held upon entry.
202  *
203  * Returns:     Nothing.
204  *
205  * Notes:       We do this for one of two cases.  Either the host is busy
206  *              and it cannot accept any more commands for the time being,
207  *              or the device returned QUEUE_FULL and can accept no more
208  *              commands.
209  * Notes:       This could be called either from an interrupt context or a
210  *              normal process context.
211  */
212 void scsi_queue_insert(struct scsi_cmnd *cmd, int reason)
213 {
214         __scsi_queue_insert(cmd, reason, true);
215 }
216
217
218 /**
219  * __scsi_execute - insert request and wait for the result
220  * @sdev:       scsi device
221  * @cmd:        scsi command
222  * @data_direction: data direction
223  * @buffer:     data buffer
224  * @bufflen:    len of buffer
225  * @sense:      optional sense buffer
226  * @sshdr:      optional decoded sense header
227  * @timeout:    request timeout in seconds
228  * @retries:    number of times to retry request
229  * @flags:      flags for ->cmd_flags
230  * @rq_flags:   flags for ->rq_flags
231  * @resid:      optional residual length
232  *
233  * Returns the scsi_cmnd result field if a command was executed, or a negative
234  * Linux error code if we didn't get that far.
235  */
236 int __scsi_execute(struct scsi_device *sdev, const unsigned char *cmd,
237                  int data_direction, void *buffer, unsigned bufflen,
238                  unsigned char *sense, struct scsi_sense_hdr *sshdr,
239                  int timeout, int retries, u64 flags, req_flags_t rq_flags,
240                  int *resid)
241 {
242         struct request *req;
243         struct scsi_request *rq;
244         int ret = DRIVER_ERROR << 24;
245
246         req = blk_get_request(sdev->request_queue,
247                         data_direction == DMA_TO_DEVICE ?
248                         REQ_OP_SCSI_OUT : REQ_OP_SCSI_IN, BLK_MQ_REQ_PREEMPT);
249         if (IS_ERR(req))
250                 return ret;
251         rq = scsi_req(req);
252
253         if (bufflen &&  blk_rq_map_kern(sdev->request_queue, req,
254                                         buffer, bufflen, GFP_NOIO))
255                 goto out;
256
257         rq->cmd_len = COMMAND_SIZE(cmd[0]);
258         memcpy(rq->cmd, cmd, rq->cmd_len);
259         rq->retries = retries;
260         req->timeout = timeout;
261         req->cmd_flags |= flags;
262         req->rq_flags |= rq_flags | RQF_QUIET;
263
264         /*
265          * head injection *required* here otherwise quiesce won't work
266          */
267         blk_execute_rq(req->q, NULL, req, 1);
268
269         /*
270          * Some devices (USB mass-storage in particular) may transfer
271          * garbage data together with a residue indicating that the data
272          * is invalid.  Prevent the garbage from being misinterpreted
273          * and prevent security leaks by zeroing out the excess data.
274          */
275         if (unlikely(rq->resid_len > 0 && rq->resid_len <= bufflen))
276                 memset(buffer + (bufflen - rq->resid_len), 0, rq->resid_len);
277
278         if (resid)
279                 *resid = rq->resid_len;
280         if (sense && rq->sense_len)
281                 memcpy(sense, rq->sense, SCSI_SENSE_BUFFERSIZE);
282         if (sshdr)
283                 scsi_normalize_sense(rq->sense, rq->sense_len, sshdr);
284         ret = rq->result;
285  out:
286         blk_put_request(req);
287
288         return ret;
289 }
290 EXPORT_SYMBOL(__scsi_execute);
291
292 /*
293  * Function:    scsi_init_cmd_errh()
294  *
295  * Purpose:     Initialize cmd fields related to error handling.
296  *
297  * Arguments:   cmd     - command that is ready to be queued.
298  *
299  * Notes:       This function has the job of initializing a number of
300  *              fields related to error handling.   Typically this will
301  *              be called once for each command, as required.
302  */
303 static void scsi_init_cmd_errh(struct scsi_cmnd *cmd)
304 {
305         scsi_set_resid(cmd, 0);
306         memset(cmd->sense_buffer, 0, SCSI_SENSE_BUFFERSIZE);
307         if (cmd->cmd_len == 0)
308                 cmd->cmd_len = scsi_command_size(cmd->cmnd);
309 }
310
311 /*
312  * Decrement the host_busy counter and wake up the error handler if necessary.
313  * Avoid as follows that the error handler is not woken up if shost->host_busy
314  * == shost->host_failed: use call_rcu() in scsi_eh_scmd_add() in combination
315  * with an RCU read lock in this function to ensure that this function in its
316  * entirety either finishes before scsi_eh_scmd_add() increases the
317  * host_failed counter or that it notices the shost state change made by
318  * scsi_eh_scmd_add().
319  */
320 static void scsi_dec_host_busy(struct Scsi_Host *shost)
321 {
322         unsigned long flags;
323
324         rcu_read_lock();
325         atomic_dec(&shost->host_busy);
326         if (unlikely(scsi_host_in_recovery(shost))) {
327                 spin_lock_irqsave(shost->host_lock, flags);
328                 if (shost->host_failed || shost->host_eh_scheduled)
329                         scsi_eh_wakeup(shost);
330                 spin_unlock_irqrestore(shost->host_lock, flags);
331         }
332         rcu_read_unlock();
333 }
334
335 void scsi_device_unbusy(struct scsi_device *sdev)
336 {
337         struct Scsi_Host *shost = sdev->host;
338         struct scsi_target *starget = scsi_target(sdev);
339
340         scsi_dec_host_busy(shost);
341
342         if (starget->can_queue > 0)
343                 atomic_dec(&starget->target_busy);
344
345         atomic_dec(&sdev->device_busy);
346 }
347
348 static void scsi_kick_queue(struct request_queue *q)
349 {
350         blk_mq_run_hw_queues(q, false);
351 }
352
353 /*
354  * Called for single_lun devices on IO completion. Clear starget_sdev_user,
355  * and call blk_run_queue for all the scsi_devices on the target -
356  * including current_sdev first.
357  *
358  * Called with *no* scsi locks held.
359  */
360 static void scsi_single_lun_run(struct scsi_device *current_sdev)
361 {
362         struct Scsi_Host *shost = current_sdev->host;
363         struct scsi_device *sdev, *tmp;
364         struct scsi_target *starget = scsi_target(current_sdev);
365         unsigned long flags;
366
367         spin_lock_irqsave(shost->host_lock, flags);
368         starget->starget_sdev_user = NULL;
369         spin_unlock_irqrestore(shost->host_lock, flags);
370
371         /*
372          * Call blk_run_queue for all LUNs on the target, starting with
373          * current_sdev. We race with others (to set starget_sdev_user),
374          * but in most cases, we will be first. Ideally, each LU on the
375          * target would get some limited time or requests on the target.
376          */
377         scsi_kick_queue(current_sdev->request_queue);
378
379         spin_lock_irqsave(shost->host_lock, flags);
380         if (starget->starget_sdev_user)
381                 goto out;
382         list_for_each_entry_safe(sdev, tmp, &starget->devices,
383                         same_target_siblings) {
384                 if (sdev == current_sdev)
385                         continue;
386                 if (scsi_device_get(sdev))
387                         continue;
388
389                 spin_unlock_irqrestore(shost->host_lock, flags);
390                 scsi_kick_queue(sdev->request_queue);
391                 spin_lock_irqsave(shost->host_lock, flags);
392         
393                 scsi_device_put(sdev);
394         }
395  out:
396         spin_unlock_irqrestore(shost->host_lock, flags);
397 }
398
399 static inline bool scsi_device_is_busy(struct scsi_device *sdev)
400 {
401         if (atomic_read(&sdev->device_busy) >= sdev->queue_depth)
402                 return true;
403         if (atomic_read(&sdev->device_blocked) > 0)
404                 return true;
405         return false;
406 }
407
408 static inline bool scsi_target_is_busy(struct scsi_target *starget)
409 {
410         if (starget->can_queue > 0) {
411                 if (atomic_read(&starget->target_busy) >= starget->can_queue)
412                         return true;
413                 if (atomic_read(&starget->target_blocked) > 0)
414                         return true;
415         }
416         return false;
417 }
418
419 static inline bool scsi_host_is_busy(struct Scsi_Host *shost)
420 {
421         if (shost->can_queue > 0 &&
422             atomic_read(&shost->host_busy) >= shost->can_queue)
423                 return true;
424         if (atomic_read(&shost->host_blocked) > 0)
425                 return true;
426         if (shost->host_self_blocked)
427                 return true;
428         return false;
429 }
430
431 static void scsi_starved_list_run(struct Scsi_Host *shost)
432 {
433         LIST_HEAD(starved_list);
434         struct scsi_device *sdev;
435         unsigned long flags;
436
437         spin_lock_irqsave(shost->host_lock, flags);
438         list_splice_init(&shost->starved_list, &starved_list);
439
440         while (!list_empty(&starved_list)) {
441                 struct request_queue *slq;
442
443                 /*
444                  * As long as shost is accepting commands and we have
445                  * starved queues, call blk_run_queue. scsi_request_fn
446                  * drops the queue_lock and can add us back to the
447                  * starved_list.
448                  *
449                  * host_lock protects the starved_list and starved_entry.
450                  * scsi_request_fn must get the host_lock before checking
451                  * or modifying starved_list or starved_entry.
452                  */
453                 if (scsi_host_is_busy(shost))
454                         break;
455
456                 sdev = list_entry(starved_list.next,
457                                   struct scsi_device, starved_entry);
458                 list_del_init(&sdev->starved_entry);
459                 if (scsi_target_is_busy(scsi_target(sdev))) {
460                         list_move_tail(&sdev->starved_entry,
461                                        &shost->starved_list);
462                         continue;
463                 }
464
465                 /*
466                  * Once we drop the host lock, a racing scsi_remove_device()
467                  * call may remove the sdev from the starved list and destroy
468                  * it and the queue.  Mitigate by taking a reference to the
469                  * queue and never touching the sdev again after we drop the
470                  * host lock.  Note: if __scsi_remove_device() invokes
471                  * blk_cleanup_queue() before the queue is run from this
472                  * function then blk_run_queue() will return immediately since
473                  * blk_cleanup_queue() marks the queue with QUEUE_FLAG_DYING.
474                  */
475                 slq = sdev->request_queue;
476                 if (!blk_get_queue(slq))
477                         continue;
478                 spin_unlock_irqrestore(shost->host_lock, flags);
479
480                 scsi_kick_queue(slq);
481                 blk_put_queue(slq);
482
483                 spin_lock_irqsave(shost->host_lock, flags);
484         }
485         /* put any unprocessed entries back */
486         list_splice(&starved_list, &shost->starved_list);
487         spin_unlock_irqrestore(shost->host_lock, flags);
488 }
489
490 /*
491  * Function:   scsi_run_queue()
492  *
493  * Purpose:    Select a proper request queue to serve next
494  *
495  * Arguments:  q       - last request's queue
496  *
497  * Returns:     Nothing
498  *
499  * Notes:      The previous command was completely finished, start
500  *             a new one if possible.
501  */
502 static void scsi_run_queue(struct request_queue *q)
503 {
504         struct scsi_device *sdev = q->queuedata;
505
506         if (scsi_target(sdev)->single_lun)
507                 scsi_single_lun_run(sdev);
508         if (!list_empty(&sdev->host->starved_list))
509                 scsi_starved_list_run(sdev->host);
510
511         blk_mq_run_hw_queues(q, false);
512 }
513
514 void scsi_requeue_run_queue(struct work_struct *work)
515 {
516         struct scsi_device *sdev;
517         struct request_queue *q;
518
519         sdev = container_of(work, struct scsi_device, requeue_work);
520         q = sdev->request_queue;
521         scsi_run_queue(q);
522 }
523
524 void scsi_run_host_queues(struct Scsi_Host *shost)
525 {
526         struct scsi_device *sdev;
527
528         shost_for_each_device(sdev, shost)
529                 scsi_run_queue(sdev->request_queue);
530 }
531
532 static void scsi_uninit_cmd(struct scsi_cmnd *cmd)
533 {
534         if (!blk_rq_is_passthrough(cmd->request)) {
535                 struct scsi_driver *drv = scsi_cmd_to_driver(cmd);
536
537                 if (drv->uninit_command)
538                         drv->uninit_command(cmd);
539         }
540 }
541
542 static void scsi_mq_free_sgtables(struct scsi_cmnd *cmd)
543 {
544         if (cmd->sdb.table.nents)
545                 sg_free_table_chained(&cmd->sdb.table, true);
546         if (scsi_prot_sg_count(cmd))
547                 sg_free_table_chained(&cmd->prot_sdb->table, true);
548 }
549
550 static void scsi_mq_uninit_cmd(struct scsi_cmnd *cmd)
551 {
552         scsi_mq_free_sgtables(cmd);
553         scsi_uninit_cmd(cmd);
554         scsi_del_cmd_from_list(cmd);
555 }
556
557 /* Returns false when no more bytes to process, true if there are more */
558 static bool scsi_end_request(struct request *req, blk_status_t error,
559                 unsigned int bytes)
560 {
561         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
562         struct scsi_device *sdev = cmd->device;
563         struct request_queue *q = sdev->request_queue;
564
565         if (blk_update_request(req, error, bytes))
566                 return true;
567
568         if (blk_queue_add_random(q))
569                 add_disk_randomness(req->rq_disk);
570
571         if (!blk_rq_is_scsi(req)) {
572                 WARN_ON_ONCE(!(cmd->flags & SCMD_INITIALIZED));
573                 cmd->flags &= ~SCMD_INITIALIZED;
574         }
575
576         /*
577          * Calling rcu_barrier() is not necessary here because the
578          * SCSI error handler guarantees that the function called by
579          * call_rcu() has been called before scsi_end_request() is
580          * called.
581          */
582         destroy_rcu_head(&cmd->rcu);
583
584         /*
585          * In the MQ case the command gets freed by __blk_mq_end_request,
586          * so we have to do all cleanup that depends on it earlier.
587          *
588          * We also can't kick the queues from irq context, so we
589          * will have to defer it to a workqueue.
590          */
591         scsi_mq_uninit_cmd(cmd);
592
593         /*
594          * queue is still alive, so grab the ref for preventing it
595          * from being cleaned up during running queue.
596          */
597         percpu_ref_get(&q->q_usage_counter);
598
599         __blk_mq_end_request(req, error);
600
601         if (scsi_target(sdev)->single_lun ||
602             !list_empty(&sdev->host->starved_list))
603                 kblockd_schedule_work(&sdev->requeue_work);
604         else
605                 blk_mq_run_hw_queues(q, true);
606
607         percpu_ref_put(&q->q_usage_counter);
608         return false;
609 }
610
611 /**
612  * scsi_result_to_blk_status - translate a SCSI result code into blk_status_t
613  * @cmd:        SCSI command
614  * @result:     scsi error code
615  *
616  * Translate a SCSI result code into a blk_status_t value. May reset the host
617  * byte of @cmd->result.
618  */
619 static blk_status_t scsi_result_to_blk_status(struct scsi_cmnd *cmd, int result)
620 {
621         switch (host_byte(result)) {
622         case DID_OK:
623                 /*
624                  * Also check the other bytes than the status byte in result
625                  * to handle the case when a SCSI LLD sets result to
626                  * DRIVER_SENSE << 24 without setting SAM_STAT_CHECK_CONDITION.
627                  */
628                 if (scsi_status_is_good(result) && (result & ~0xff) == 0)
629                         return BLK_STS_OK;
630                 return BLK_STS_IOERR;
631         case DID_TRANSPORT_FAILFAST:
632                 return BLK_STS_TRANSPORT;
633         case DID_TARGET_FAILURE:
634                 set_host_byte(cmd, DID_OK);
635                 return BLK_STS_TARGET;
636         case DID_NEXUS_FAILURE:
637                 set_host_byte(cmd, DID_OK);
638                 return BLK_STS_NEXUS;
639         case DID_ALLOC_FAILURE:
640                 set_host_byte(cmd, DID_OK);
641                 return BLK_STS_NOSPC;
642         case DID_MEDIUM_ERROR:
643                 set_host_byte(cmd, DID_OK);
644                 return BLK_STS_MEDIUM;
645         default:
646                 return BLK_STS_IOERR;
647         }
648 }
649
650 /* Helper for scsi_io_completion() when "reprep" action required. */
651 static void scsi_io_completion_reprep(struct scsi_cmnd *cmd,
652                                       struct request_queue *q)
653 {
654         /* A new command will be prepared and issued. */
655         scsi_mq_requeue_cmd(cmd);
656 }
657
658 /* Helper for scsi_io_completion() when special action required. */
659 static void scsi_io_completion_action(struct scsi_cmnd *cmd, int result)
660 {
661         struct request_queue *q = cmd->device->request_queue;
662         struct request *req = cmd->request;
663         int level = 0;
664         enum {ACTION_FAIL, ACTION_REPREP, ACTION_RETRY,
665               ACTION_DELAYED_RETRY} action;
666         unsigned long wait_for = (cmd->allowed + 1) * req->timeout;
667         struct scsi_sense_hdr sshdr;
668         bool sense_valid;
669         bool sense_current = true;      /* false implies "deferred sense" */
670         blk_status_t blk_stat;
671
672         sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
673         if (sense_valid)
674                 sense_current = !scsi_sense_is_deferred(&sshdr);
675
676         blk_stat = scsi_result_to_blk_status(cmd, result);
677
678         if (host_byte(result) == DID_RESET) {
679                 /* Third party bus reset or reset for error recovery
680                  * reasons.  Just retry the command and see what
681                  * happens.
682                  */
683                 action = ACTION_RETRY;
684         } else if (sense_valid && sense_current) {
685                 switch (sshdr.sense_key) {
686                 case UNIT_ATTENTION:
687                         if (cmd->device->removable) {
688                                 /* Detected disc change.  Set a bit
689                                  * and quietly refuse further access.
690                                  */
691                                 cmd->device->changed = 1;
692                                 action = ACTION_FAIL;
693                         } else {
694                                 /* Must have been a power glitch, or a
695                                  * bus reset.  Could not have been a
696                                  * media change, so we just retry the
697                                  * command and see what happens.
698                                  */
699                                 action = ACTION_RETRY;
700                         }
701                         break;
702                 case ILLEGAL_REQUEST:
703                         /* If we had an ILLEGAL REQUEST returned, then
704                          * we may have performed an unsupported
705                          * command.  The only thing this should be
706                          * would be a ten byte read where only a six
707                          * byte read was supported.  Also, on a system
708                          * where READ CAPACITY failed, we may have
709                          * read past the end of the disk.
710                          */
711                         if ((cmd->device->use_10_for_rw &&
712                             sshdr.asc == 0x20 && sshdr.ascq == 0x00) &&
713                             (cmd->cmnd[0] == READ_10 ||
714                              cmd->cmnd[0] == WRITE_10)) {
715                                 /* This will issue a new 6-byte command. */
716                                 cmd->device->use_10_for_rw = 0;
717                                 action = ACTION_REPREP;
718                         } else if (sshdr.asc == 0x10) /* DIX */ {
719                                 action = ACTION_FAIL;
720                                 blk_stat = BLK_STS_PROTECTION;
721                         /* INVALID COMMAND OPCODE or INVALID FIELD IN CDB */
722                         } else if (sshdr.asc == 0x20 || sshdr.asc == 0x24) {
723                                 action = ACTION_FAIL;
724                                 blk_stat = BLK_STS_TARGET;
725                         } else
726                                 action = ACTION_FAIL;
727                         break;
728                 case ABORTED_COMMAND:
729                         action = ACTION_FAIL;
730                         if (sshdr.asc == 0x10) /* DIF */
731                                 blk_stat = BLK_STS_PROTECTION;
732                         break;
733                 case NOT_READY:
734                         /* If the device is in the process of becoming
735                          * ready, or has a temporary blockage, retry.
736                          */
737                         if (sshdr.asc == 0x04) {
738                                 switch (sshdr.ascq) {
739                                 case 0x01: /* becoming ready */
740                                 case 0x04: /* format in progress */
741                                 case 0x05: /* rebuild in progress */
742                                 case 0x06: /* recalculation in progress */
743                                 case 0x07: /* operation in progress */
744                                 case 0x08: /* Long write in progress */
745                                 case 0x09: /* self test in progress */
746                                 case 0x14: /* space allocation in progress */
747                                 case 0x1a: /* start stop unit in progress */
748                                 case 0x1b: /* sanitize in progress */
749                                 case 0x1d: /* configuration in progress */
750                                 case 0x24: /* depopulation in progress */
751                                         action = ACTION_DELAYED_RETRY;
752                                         break;
753                                 default:
754                                         action = ACTION_FAIL;
755                                         break;
756                                 }
757                         } else
758                                 action = ACTION_FAIL;
759                         break;
760                 case VOLUME_OVERFLOW:
761                         /* See SSC3rXX or current. */
762                         action = ACTION_FAIL;
763                         break;
764                 default:
765                         action = ACTION_FAIL;
766                         break;
767                 }
768         } else
769                 action = ACTION_FAIL;
770
771         if (action != ACTION_FAIL &&
772             time_before(cmd->jiffies_at_alloc + wait_for, jiffies))
773                 action = ACTION_FAIL;
774
775         switch (action) {
776         case ACTION_FAIL:
777                 /* Give up and fail the remainder of the request */
778                 if (!(req->rq_flags & RQF_QUIET)) {
779                         static DEFINE_RATELIMIT_STATE(_rs,
780                                         DEFAULT_RATELIMIT_INTERVAL,
781                                         DEFAULT_RATELIMIT_BURST);
782
783                         if (unlikely(scsi_logging_level))
784                                 level =
785                                      SCSI_LOG_LEVEL(SCSI_LOG_MLCOMPLETE_SHIFT,
786                                                     SCSI_LOG_MLCOMPLETE_BITS);
787
788                         /*
789                          * if logging is enabled the failure will be printed
790                          * in scsi_log_completion(), so avoid duplicate messages
791                          */
792                         if (!level && __ratelimit(&_rs)) {
793                                 scsi_print_result(cmd, NULL, FAILED);
794                                 if (driver_byte(result) == DRIVER_SENSE)
795                                         scsi_print_sense(cmd);
796                                 scsi_print_command(cmd);
797                         }
798                 }
799                 if (!scsi_end_request(req, blk_stat, blk_rq_err_bytes(req)))
800                         return;
801                 /*FALLTHRU*/
802         case ACTION_REPREP:
803                 scsi_io_completion_reprep(cmd, q);
804                 break;
805         case ACTION_RETRY:
806                 /* Retry the same command immediately */
807                 __scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY, false);
808                 break;
809         case ACTION_DELAYED_RETRY:
810                 /* Retry the same command after a delay */
811                 __scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY, false);
812                 break;
813         }
814 }
815
816 /*
817  * Helper for scsi_io_completion() when cmd->result is non-zero. Returns a
818  * new result that may suppress further error checking. Also modifies
819  * *blk_statp in some cases.
820  */
821 static int scsi_io_completion_nz_result(struct scsi_cmnd *cmd, int result,
822                                         blk_status_t *blk_statp)
823 {
824         bool sense_valid;
825         bool sense_current = true;      /* false implies "deferred sense" */
826         struct request *req = cmd->request;
827         struct scsi_sense_hdr sshdr;
828
829         sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
830         if (sense_valid)
831                 sense_current = !scsi_sense_is_deferred(&sshdr);
832
833         if (blk_rq_is_passthrough(req)) {
834                 if (sense_valid) {
835                         /*
836                          * SG_IO wants current and deferred errors
837                          */
838                         scsi_req(req)->sense_len =
839                                 min(8 + cmd->sense_buffer[7],
840                                     SCSI_SENSE_BUFFERSIZE);
841                 }
842                 if (sense_current)
843                         *blk_statp = scsi_result_to_blk_status(cmd, result);
844         } else if (blk_rq_bytes(req) == 0 && sense_current) {
845                 /*
846                  * Flush commands do not transfers any data, and thus cannot use
847                  * good_bytes != blk_rq_bytes(req) as the signal for an error.
848                  * This sets *blk_statp explicitly for the problem case.
849                  */
850                 *blk_statp = scsi_result_to_blk_status(cmd, result);
851         }
852         /*
853          * Recovered errors need reporting, but they're always treated as
854          * success, so fiddle the result code here.  For passthrough requests
855          * we already took a copy of the original into sreq->result which
856          * is what gets returned to the user
857          */
858         if (sense_valid && (sshdr.sense_key == RECOVERED_ERROR)) {
859                 bool do_print = true;
860                 /*
861                  * if ATA PASS-THROUGH INFORMATION AVAILABLE [0x0, 0x1d]
862                  * skip print since caller wants ATA registers. Only occurs
863                  * on SCSI ATA PASS_THROUGH commands when CK_COND=1
864                  */
865                 if ((sshdr.asc == 0x0) && (sshdr.ascq == 0x1d))
866                         do_print = false;
867                 else if (req->rq_flags & RQF_QUIET)
868                         do_print = false;
869                 if (do_print)
870                         scsi_print_sense(cmd);
871                 result = 0;
872                 /* for passthrough, *blk_statp may be set */
873                 *blk_statp = BLK_STS_OK;
874         }
875         /*
876          * Another corner case: the SCSI status byte is non-zero but 'good'.
877          * Example: PRE-FETCH command returns SAM_STAT_CONDITION_MET when
878          * it is able to fit nominated LBs in its cache (and SAM_STAT_GOOD
879          * if it can't fit). Treat SAM_STAT_CONDITION_MET and the related
880          * intermediate statuses (both obsolete in SAM-4) as good.
881          */
882         if (status_byte(result) && scsi_status_is_good(result)) {
883                 result = 0;
884                 *blk_statp = BLK_STS_OK;
885         }
886         return result;
887 }
888
889 /*
890  * Function:    scsi_io_completion()
891  *
892  * Purpose:     Completion processing for block device I/O requests.
893  *
894  * Arguments:   cmd   - command that is finished.
895  *
896  * Lock status: Assumed that no lock is held upon entry.
897  *
898  * Returns:     Nothing
899  *
900  * Notes:       We will finish off the specified number of sectors.  If we
901  *              are done, the command block will be released and the queue
902  *              function will be goosed.  If we are not done then we have to
903  *              figure out what to do next:
904  *
905  *              a) We can call scsi_requeue_command().  The request
906  *                 will be unprepared and put back on the queue.  Then
907  *                 a new command will be created for it.  This should
908  *                 be used if we made forward progress, or if we want
909  *                 to switch from READ(10) to READ(6) for example.
910  *
911  *              b) We can call __scsi_queue_insert().  The request will
912  *                 be put back on the queue and retried using the same
913  *                 command as before, possibly after a delay.
914  *
915  *              c) We can call scsi_end_request() with blk_stat other than
916  *                 BLK_STS_OK, to fail the remainder of the request.
917  */
918 void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes)
919 {
920         int result = cmd->result;
921         struct request_queue *q = cmd->device->request_queue;
922         struct request *req = cmd->request;
923         blk_status_t blk_stat = BLK_STS_OK;
924
925         if (unlikely(result))   /* a nz result may or may not be an error */
926                 result = scsi_io_completion_nz_result(cmd, result, &blk_stat);
927
928         if (unlikely(blk_rq_is_passthrough(req))) {
929                 /*
930                  * scsi_result_to_blk_status may have reset the host_byte
931                  */
932                 scsi_req(req)->result = cmd->result;
933         }
934
935         /*
936          * Next deal with any sectors which we were able to correctly
937          * handle.
938          */
939         SCSI_LOG_HLCOMPLETE(1, scmd_printk(KERN_INFO, cmd,
940                 "%u sectors total, %d bytes done.\n",
941                 blk_rq_sectors(req), good_bytes));
942
943         /*
944          * Next deal with any sectors which we were able to correctly
945          * handle. Failed, zero length commands always need to drop down
946          * to retry code. Fast path should return in this block.
947          */
948         if (likely(blk_rq_bytes(req) > 0 || blk_stat == BLK_STS_OK)) {
949                 if (likely(!scsi_end_request(req, blk_stat, good_bytes)))
950                         return; /* no bytes remaining */
951         }
952
953         /* Kill remainder if no retries. */
954         if (unlikely(blk_stat && scsi_noretry_cmd(cmd))) {
955                 if (scsi_end_request(req, blk_stat, blk_rq_bytes(req)))
956                         WARN_ONCE(true,
957                             "Bytes remaining after failed, no-retry command");
958                 return;
959         }
960
961         /*
962          * If there had been no error, but we have leftover bytes in the
963          * requeues just queue the command up again.
964          */
965         if (likely(result == 0))
966                 scsi_io_completion_reprep(cmd, q);
967         else
968                 scsi_io_completion_action(cmd, result);
969 }
970
971 static blk_status_t scsi_init_sgtable(struct request *req,
972                 struct scsi_data_buffer *sdb)
973 {
974         int count;
975
976         /*
977          * If sg table allocation fails, requeue request later.
978          */
979         if (unlikely(sg_alloc_table_chained(&sdb->table,
980                         blk_rq_nr_phys_segments(req), sdb->table.sgl)))
981                 return BLK_STS_RESOURCE;
982
983         /* 
984          * Next, walk the list, and fill in the addresses and sizes of
985          * each segment.
986          */
987         count = blk_rq_map_sg(req->q, req, sdb->table.sgl);
988         BUG_ON(count > sdb->table.nents);
989         sdb->table.nents = count;
990         sdb->length = blk_rq_payload_bytes(req);
991         return BLK_STS_OK;
992 }
993
994 /*
995  * Function:    scsi_init_io()
996  *
997  * Purpose:     SCSI I/O initialize function.
998  *
999  * Arguments:   cmd   - Command descriptor we wish to initialize
1000  *
1001  * Returns:     BLK_STS_OK on success
1002  *              BLK_STS_RESOURCE if the failure is retryable
1003  *              BLK_STS_IOERR if the failure is fatal
1004  */
1005 blk_status_t scsi_init_io(struct scsi_cmnd *cmd)
1006 {
1007         struct request *rq = cmd->request;
1008         blk_status_t ret;
1009
1010         if (WARN_ON_ONCE(!blk_rq_nr_phys_segments(rq)))
1011                 return BLK_STS_IOERR;
1012
1013         ret = scsi_init_sgtable(rq, &cmd->sdb);
1014         if (ret)
1015                 return ret;
1016
1017         if (blk_integrity_rq(rq)) {
1018                 struct scsi_data_buffer *prot_sdb = cmd->prot_sdb;
1019                 int ivecs, count;
1020
1021                 if (WARN_ON_ONCE(!prot_sdb)) {
1022                         /*
1023                          * This can happen if someone (e.g. multipath)
1024                          * queues a command to a device on an adapter
1025                          * that does not support DIX.
1026                          */
1027                         ret = BLK_STS_IOERR;
1028                         goto out_free_sgtables;
1029                 }
1030
1031                 ivecs = blk_rq_count_integrity_sg(rq->q, rq->bio);
1032
1033                 if (sg_alloc_table_chained(&prot_sdb->table, ivecs,
1034                                 prot_sdb->table.sgl)) {
1035                         ret = BLK_STS_RESOURCE;
1036                         goto out_free_sgtables;
1037                 }
1038
1039                 count = blk_rq_map_integrity_sg(rq->q, rq->bio,
1040                                                 prot_sdb->table.sgl);
1041                 BUG_ON(count > ivecs);
1042                 BUG_ON(count > queue_max_integrity_segments(rq->q));
1043
1044                 cmd->prot_sdb = prot_sdb;
1045                 cmd->prot_sdb->table.nents = count;
1046         }
1047
1048         return BLK_STS_OK;
1049 out_free_sgtables:
1050         scsi_mq_free_sgtables(cmd);
1051         return ret;
1052 }
1053 EXPORT_SYMBOL(scsi_init_io);
1054
1055 /**
1056  * scsi_initialize_rq - initialize struct scsi_cmnd partially
1057  * @rq: Request associated with the SCSI command to be initialized.
1058  *
1059  * This function initializes the members of struct scsi_cmnd that must be
1060  * initialized before request processing starts and that won't be
1061  * reinitialized if a SCSI command is requeued.
1062  *
1063  * Called from inside blk_get_request() for pass-through requests and from
1064  * inside scsi_init_command() for filesystem requests.
1065  */
1066 static void scsi_initialize_rq(struct request *rq)
1067 {
1068         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1069
1070         scsi_req_init(&cmd->req);
1071         init_rcu_head(&cmd->rcu);
1072         cmd->jiffies_at_alloc = jiffies;
1073         cmd->retries = 0;
1074 }
1075
1076 /* Add a command to the list used by the aacraid and dpt_i2o drivers */
1077 void scsi_add_cmd_to_list(struct scsi_cmnd *cmd)
1078 {
1079         struct scsi_device *sdev = cmd->device;
1080         struct Scsi_Host *shost = sdev->host;
1081         unsigned long flags;
1082
1083         if (shost->use_cmd_list) {
1084                 spin_lock_irqsave(&sdev->list_lock, flags);
1085                 list_add_tail(&cmd->list, &sdev->cmd_list);
1086                 spin_unlock_irqrestore(&sdev->list_lock, flags);
1087         }
1088 }
1089
1090 /* Remove a command from the list used by the aacraid and dpt_i2o drivers */
1091 void scsi_del_cmd_from_list(struct scsi_cmnd *cmd)
1092 {
1093         struct scsi_device *sdev = cmd->device;
1094         struct Scsi_Host *shost = sdev->host;
1095         unsigned long flags;
1096
1097         if (shost->use_cmd_list) {
1098                 spin_lock_irqsave(&sdev->list_lock, flags);
1099                 BUG_ON(list_empty(&cmd->list));
1100                 list_del_init(&cmd->list);
1101                 spin_unlock_irqrestore(&sdev->list_lock, flags);
1102         }
1103 }
1104
1105 /* Called after a request has been started. */
1106 void scsi_init_command(struct scsi_device *dev, struct scsi_cmnd *cmd)
1107 {
1108         void *buf = cmd->sense_buffer;
1109         void *prot = cmd->prot_sdb;
1110         struct request *rq = blk_mq_rq_from_pdu(cmd);
1111         unsigned int flags = cmd->flags & SCMD_PRESERVED_FLAGS;
1112         unsigned long jiffies_at_alloc;
1113         int retries;
1114
1115         if (!blk_rq_is_scsi(rq) && !(flags & SCMD_INITIALIZED)) {
1116                 flags |= SCMD_INITIALIZED;
1117                 scsi_initialize_rq(rq);
1118         }
1119
1120         jiffies_at_alloc = cmd->jiffies_at_alloc;
1121         retries = cmd->retries;
1122         /* zero out the cmd, except for the embedded scsi_request */
1123         memset((char *)cmd + sizeof(cmd->req), 0,
1124                 sizeof(*cmd) - sizeof(cmd->req) + dev->host->hostt->cmd_size);
1125
1126         cmd->device = dev;
1127         cmd->sense_buffer = buf;
1128         cmd->prot_sdb = prot;
1129         cmd->flags = flags;
1130         INIT_DELAYED_WORK(&cmd->abort_work, scmd_eh_abort_handler);
1131         cmd->jiffies_at_alloc = jiffies_at_alloc;
1132         cmd->retries = retries;
1133
1134         scsi_add_cmd_to_list(cmd);
1135 }
1136
1137 static blk_status_t scsi_setup_scsi_cmnd(struct scsi_device *sdev,
1138                 struct request *req)
1139 {
1140         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1141
1142         /*
1143          * Passthrough requests may transfer data, in which case they must
1144          * a bio attached to them.  Or they might contain a SCSI command
1145          * that does not transfer data, in which case they may optionally
1146          * submit a request without an attached bio.
1147          */
1148         if (req->bio) {
1149                 blk_status_t ret = scsi_init_io(cmd);
1150                 if (unlikely(ret != BLK_STS_OK))
1151                         return ret;
1152         } else {
1153                 BUG_ON(blk_rq_bytes(req));
1154
1155                 memset(&cmd->sdb, 0, sizeof(cmd->sdb));
1156         }
1157
1158         cmd->cmd_len = scsi_req(req)->cmd_len;
1159         cmd->cmnd = scsi_req(req)->cmd;
1160         cmd->transfersize = blk_rq_bytes(req);
1161         cmd->allowed = scsi_req(req)->retries;
1162         return BLK_STS_OK;
1163 }
1164
1165 /*
1166  * Setup a normal block command.  These are simple request from filesystems
1167  * that still need to be translated to SCSI CDBs from the ULD.
1168  */
1169 static blk_status_t scsi_setup_fs_cmnd(struct scsi_device *sdev,
1170                 struct request *req)
1171 {
1172         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1173
1174         if (unlikely(sdev->handler && sdev->handler->prep_fn)) {
1175                 blk_status_t ret = sdev->handler->prep_fn(sdev, req);
1176                 if (ret != BLK_STS_OK)
1177                         return ret;
1178         }
1179
1180         cmd->cmnd = scsi_req(req)->cmd = scsi_req(req)->__cmd;
1181         memset(cmd->cmnd, 0, BLK_MAX_CDB);
1182         return scsi_cmd_to_driver(cmd)->init_command(cmd);
1183 }
1184
1185 static blk_status_t scsi_setup_cmnd(struct scsi_device *sdev,
1186                 struct request *req)
1187 {
1188         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1189
1190         if (!blk_rq_bytes(req))
1191                 cmd->sc_data_direction = DMA_NONE;
1192         else if (rq_data_dir(req) == WRITE)
1193                 cmd->sc_data_direction = DMA_TO_DEVICE;
1194         else
1195                 cmd->sc_data_direction = DMA_FROM_DEVICE;
1196
1197         if (blk_rq_is_scsi(req))
1198                 return scsi_setup_scsi_cmnd(sdev, req);
1199         else
1200                 return scsi_setup_fs_cmnd(sdev, req);
1201 }
1202
1203 static blk_status_t
1204 scsi_prep_state_check(struct scsi_device *sdev, struct request *req)
1205 {
1206         switch (sdev->sdev_state) {
1207         case SDEV_OFFLINE:
1208         case SDEV_TRANSPORT_OFFLINE:
1209                 /*
1210                  * If the device is offline we refuse to process any
1211                  * commands.  The device must be brought online
1212                  * before trying any recovery commands.
1213                  */
1214                 sdev_printk(KERN_ERR, sdev,
1215                             "rejecting I/O to offline device\n");
1216                 return BLK_STS_IOERR;
1217         case SDEV_DEL:
1218                 /*
1219                  * If the device is fully deleted, we refuse to
1220                  * process any commands as well.
1221                  */
1222                 sdev_printk(KERN_ERR, sdev,
1223                             "rejecting I/O to dead device\n");
1224                 return BLK_STS_IOERR;
1225         case SDEV_BLOCK:
1226         case SDEV_CREATED_BLOCK:
1227                 return BLK_STS_RESOURCE;
1228         case SDEV_QUIESCE:
1229                 /*
1230                  * If the devices is blocked we defer normal commands.
1231                  */
1232                 if (req && !(req->rq_flags & RQF_PREEMPT))
1233                         return BLK_STS_RESOURCE;
1234                 return BLK_STS_OK;
1235         default:
1236                 /*
1237                  * For any other not fully online state we only allow
1238                  * special commands.  In particular any user initiated
1239                  * command is not allowed.
1240                  */
1241                 if (req && !(req->rq_flags & RQF_PREEMPT))
1242                         return BLK_STS_IOERR;
1243                 return BLK_STS_OK;
1244         }
1245 }
1246
1247 /*
1248  * scsi_dev_queue_ready: if we can send requests to sdev, return 1 else
1249  * return 0.
1250  *
1251  * Called with the queue_lock held.
1252  */
1253 static inline int scsi_dev_queue_ready(struct request_queue *q,
1254                                   struct scsi_device *sdev)
1255 {
1256         unsigned int busy;
1257
1258         busy = atomic_inc_return(&sdev->device_busy) - 1;
1259         if (atomic_read(&sdev->device_blocked)) {
1260                 if (busy)
1261                         goto out_dec;
1262
1263                 /*
1264                  * unblock after device_blocked iterates to zero
1265                  */
1266                 if (atomic_dec_return(&sdev->device_blocked) > 0)
1267                         goto out_dec;
1268                 SCSI_LOG_MLQUEUE(3, sdev_printk(KERN_INFO, sdev,
1269                                    "unblocking device at zero depth\n"));
1270         }
1271
1272         if (busy >= sdev->queue_depth)
1273                 goto out_dec;
1274
1275         return 1;
1276 out_dec:
1277         atomic_dec(&sdev->device_busy);
1278         return 0;
1279 }
1280
1281 /*
1282  * scsi_target_queue_ready: checks if there we can send commands to target
1283  * @sdev: scsi device on starget to check.
1284  */
1285 static inline int scsi_target_queue_ready(struct Scsi_Host *shost,
1286                                            struct scsi_device *sdev)
1287 {
1288         struct scsi_target *starget = scsi_target(sdev);
1289         unsigned int busy;
1290
1291         if (starget->single_lun) {
1292                 spin_lock_irq(shost->host_lock);
1293                 if (starget->starget_sdev_user &&
1294                     starget->starget_sdev_user != sdev) {
1295                         spin_unlock_irq(shost->host_lock);
1296                         return 0;
1297                 }
1298                 starget->starget_sdev_user = sdev;
1299                 spin_unlock_irq(shost->host_lock);
1300         }
1301
1302         if (starget->can_queue <= 0)
1303                 return 1;
1304
1305         busy = atomic_inc_return(&starget->target_busy) - 1;
1306         if (atomic_read(&starget->target_blocked) > 0) {
1307                 if (busy)
1308                         goto starved;
1309
1310                 /*
1311                  * unblock after target_blocked iterates to zero
1312                  */
1313                 if (atomic_dec_return(&starget->target_blocked) > 0)
1314                         goto out_dec;
1315
1316                 SCSI_LOG_MLQUEUE(3, starget_printk(KERN_INFO, starget,
1317                                  "unblocking target at zero depth\n"));
1318         }
1319
1320         if (busy >= starget->can_queue)
1321                 goto starved;
1322
1323         return 1;
1324
1325 starved:
1326         spin_lock_irq(shost->host_lock);
1327         list_move_tail(&sdev->starved_entry, &shost->starved_list);
1328         spin_unlock_irq(shost->host_lock);
1329 out_dec:
1330         if (starget->can_queue > 0)
1331                 atomic_dec(&starget->target_busy);
1332         return 0;
1333 }
1334
1335 /*
1336  * scsi_host_queue_ready: if we can send requests to shost, return 1 else
1337  * return 0. We must end up running the queue again whenever 0 is
1338  * returned, else IO can hang.
1339  */
1340 static inline int scsi_host_queue_ready(struct request_queue *q,
1341                                    struct Scsi_Host *shost,
1342                                    struct scsi_device *sdev)
1343 {
1344         unsigned int busy;
1345
1346         if (scsi_host_in_recovery(shost))
1347                 return 0;
1348
1349         busy = atomic_inc_return(&shost->host_busy) - 1;
1350         if (atomic_read(&shost->host_blocked) > 0) {
1351                 if (busy)
1352                         goto starved;
1353
1354                 /*
1355                  * unblock after host_blocked iterates to zero
1356                  */
1357                 if (atomic_dec_return(&shost->host_blocked) > 0)
1358                         goto out_dec;
1359
1360                 SCSI_LOG_MLQUEUE(3,
1361                         shost_printk(KERN_INFO, shost,
1362                                      "unblocking host at zero depth\n"));
1363         }
1364
1365         if (shost->can_queue > 0 && busy >= shost->can_queue)
1366                 goto starved;
1367         if (shost->host_self_blocked)
1368                 goto starved;
1369
1370         /* We're OK to process the command, so we can't be starved */
1371         if (!list_empty(&sdev->starved_entry)) {
1372                 spin_lock_irq(shost->host_lock);
1373                 if (!list_empty(&sdev->starved_entry))
1374                         list_del_init(&sdev->starved_entry);
1375                 spin_unlock_irq(shost->host_lock);
1376         }
1377
1378         return 1;
1379
1380 starved:
1381         spin_lock_irq(shost->host_lock);
1382         if (list_empty(&sdev->starved_entry))
1383                 list_add_tail(&sdev->starved_entry, &shost->starved_list);
1384         spin_unlock_irq(shost->host_lock);
1385 out_dec:
1386         scsi_dec_host_busy(shost);
1387         return 0;
1388 }
1389
1390 /*
1391  * Busy state exporting function for request stacking drivers.
1392  *
1393  * For efficiency, no lock is taken to check the busy state of
1394  * shost/starget/sdev, since the returned value is not guaranteed and
1395  * may be changed after request stacking drivers call the function,
1396  * regardless of taking lock or not.
1397  *
1398  * When scsi can't dispatch I/Os anymore and needs to kill I/Os scsi
1399  * needs to return 'not busy'. Otherwise, request stacking drivers
1400  * may hold requests forever.
1401  */
1402 static bool scsi_mq_lld_busy(struct request_queue *q)
1403 {
1404         struct scsi_device *sdev = q->queuedata;
1405         struct Scsi_Host *shost;
1406
1407         if (blk_queue_dying(q))
1408                 return false;
1409
1410         shost = sdev->host;
1411
1412         /*
1413          * Ignore host/starget busy state.
1414          * Since block layer does not have a concept of fairness across
1415          * multiple queues, congestion of host/starget needs to be handled
1416          * in SCSI layer.
1417          */
1418         if (scsi_host_in_recovery(shost) || scsi_device_is_busy(sdev))
1419                 return true;
1420
1421         return false;
1422 }
1423
1424 static void scsi_softirq_done(struct request *rq)
1425 {
1426         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1427         unsigned long wait_for = (cmd->allowed + 1) * rq->timeout;
1428         int disposition;
1429
1430         INIT_LIST_HEAD(&cmd->eh_entry);
1431
1432         atomic_inc(&cmd->device->iodone_cnt);
1433         if (cmd->result)
1434                 atomic_inc(&cmd->device->ioerr_cnt);
1435
1436         disposition = scsi_decide_disposition(cmd);
1437         if (disposition != SUCCESS &&
1438             time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) {
1439                 sdev_printk(KERN_ERR, cmd->device,
1440                             "timing out command, waited %lus\n",
1441                             wait_for/HZ);
1442                 disposition = SUCCESS;
1443         }
1444
1445         scsi_log_completion(cmd, disposition);
1446
1447         switch (disposition) {
1448                 case SUCCESS:
1449                         scsi_finish_command(cmd);
1450                         break;
1451                 case NEEDS_RETRY:
1452                         scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY);
1453                         break;
1454                 case ADD_TO_MLQUEUE:
1455                         scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY);
1456                         break;
1457                 default:
1458                         scsi_eh_scmd_add(cmd);
1459                         break;
1460         }
1461 }
1462
1463 /**
1464  * scsi_dispatch_command - Dispatch a command to the low-level driver.
1465  * @cmd: command block we are dispatching.
1466  *
1467  * Return: nonzero return request was rejected and device's queue needs to be
1468  * plugged.
1469  */
1470 static int scsi_dispatch_cmd(struct scsi_cmnd *cmd)
1471 {
1472         struct Scsi_Host *host = cmd->device->host;
1473         int rtn = 0;
1474
1475         atomic_inc(&cmd->device->iorequest_cnt);
1476
1477         /* check if the device is still usable */
1478         if (unlikely(cmd->device->sdev_state == SDEV_DEL)) {
1479                 /* in SDEV_DEL we error all commands. DID_NO_CONNECT
1480                  * returns an immediate error upwards, and signals
1481                  * that the device is no longer present */
1482                 cmd->result = DID_NO_CONNECT << 16;
1483                 goto done;
1484         }
1485
1486         /* Check to see if the scsi lld made this device blocked. */
1487         if (unlikely(scsi_device_blocked(cmd->device))) {
1488                 /*
1489                  * in blocked state, the command is just put back on
1490                  * the device queue.  The suspend state has already
1491                  * blocked the queue so future requests should not
1492                  * occur until the device transitions out of the
1493                  * suspend state.
1494                  */
1495                 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1496                         "queuecommand : device blocked\n"));
1497                 return SCSI_MLQUEUE_DEVICE_BUSY;
1498         }
1499
1500         /* Store the LUN value in cmnd, if needed. */
1501         if (cmd->device->lun_in_cdb)
1502                 cmd->cmnd[1] = (cmd->cmnd[1] & 0x1f) |
1503                                (cmd->device->lun << 5 & 0xe0);
1504
1505         scsi_log_send(cmd);
1506
1507         /*
1508          * Before we queue this command, check if the command
1509          * length exceeds what the host adapter can handle.
1510          */
1511         if (cmd->cmd_len > cmd->device->host->max_cmd_len) {
1512                 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1513                                "queuecommand : command too long. "
1514                                "cdb_size=%d host->max_cmd_len=%d\n",
1515                                cmd->cmd_len, cmd->device->host->max_cmd_len));
1516                 cmd->result = (DID_ABORT << 16);
1517                 goto done;
1518         }
1519
1520         if (unlikely(host->shost_state == SHOST_DEL)) {
1521                 cmd->result = (DID_NO_CONNECT << 16);
1522                 goto done;
1523
1524         }
1525
1526         trace_scsi_dispatch_cmd_start(cmd);
1527         rtn = host->hostt->queuecommand(host, cmd);
1528         if (rtn) {
1529                 trace_scsi_dispatch_cmd_error(cmd, rtn);
1530                 if (rtn != SCSI_MLQUEUE_DEVICE_BUSY &&
1531                     rtn != SCSI_MLQUEUE_TARGET_BUSY)
1532                         rtn = SCSI_MLQUEUE_HOST_BUSY;
1533
1534                 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1535                         "queuecommand : request rejected\n"));
1536         }
1537
1538         return rtn;
1539  done:
1540         cmd->scsi_done(cmd);
1541         return 0;
1542 }
1543
1544 /* Size in bytes of the sg-list stored in the scsi-mq command-private data. */
1545 static unsigned int scsi_mq_sgl_size(struct Scsi_Host *shost)
1546 {
1547         return min_t(unsigned int, shost->sg_tablesize, SG_CHUNK_SIZE) *
1548                 sizeof(struct scatterlist);
1549 }
1550
1551 static blk_status_t scsi_mq_prep_fn(struct request *req)
1552 {
1553         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1554         struct scsi_device *sdev = req->q->queuedata;
1555         struct Scsi_Host *shost = sdev->host;
1556         struct scatterlist *sg;
1557
1558         scsi_init_command(sdev, cmd);
1559
1560         cmd->request = req;
1561         cmd->tag = req->tag;
1562         cmd->prot_op = SCSI_PROT_NORMAL;
1563
1564         sg = (void *)cmd + sizeof(struct scsi_cmnd) + shost->hostt->cmd_size;
1565         cmd->sdb.table.sgl = sg;
1566
1567         if (scsi_host_get_prot(shost)) {
1568                 memset(cmd->prot_sdb, 0, sizeof(struct scsi_data_buffer));
1569
1570                 cmd->prot_sdb->table.sgl =
1571                         (struct scatterlist *)(cmd->prot_sdb + 1);
1572         }
1573
1574         blk_mq_start_request(req);
1575
1576         return scsi_setup_cmnd(sdev, req);
1577 }
1578
1579 static void scsi_mq_done(struct scsi_cmnd *cmd)
1580 {
1581         if (unlikely(test_and_set_bit(SCMD_STATE_COMPLETE, &cmd->state)))
1582                 return;
1583         trace_scsi_dispatch_cmd_done(cmd);
1584
1585         /*
1586          * If the block layer didn't complete the request due to a timeout
1587          * injection, scsi must clear its internal completed state so that the
1588          * timeout handler will see it needs to escalate its own error
1589          * recovery.
1590          */
1591         if (unlikely(!blk_mq_complete_request(cmd->request)))
1592                 clear_bit(SCMD_STATE_COMPLETE, &cmd->state);
1593 }
1594
1595 static void scsi_mq_put_budget(struct blk_mq_hw_ctx *hctx)
1596 {
1597         struct request_queue *q = hctx->queue;
1598         struct scsi_device *sdev = q->queuedata;
1599
1600         atomic_dec(&sdev->device_busy);
1601 }
1602
1603 static bool scsi_mq_get_budget(struct blk_mq_hw_ctx *hctx)
1604 {
1605         struct request_queue *q = hctx->queue;
1606         struct scsi_device *sdev = q->queuedata;
1607
1608         if (scsi_dev_queue_ready(q, sdev))
1609                 return true;
1610
1611         if (atomic_read(&sdev->device_busy) == 0 && !scsi_device_blocked(sdev))
1612                 blk_mq_delay_run_hw_queue(hctx, SCSI_QUEUE_DELAY);
1613         return false;
1614 }
1615
1616 static blk_status_t scsi_queue_rq(struct blk_mq_hw_ctx *hctx,
1617                          const struct blk_mq_queue_data *bd)
1618 {
1619         struct request *req = bd->rq;
1620         struct request_queue *q = req->q;
1621         struct scsi_device *sdev = q->queuedata;
1622         struct Scsi_Host *shost = sdev->host;
1623         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1624         blk_status_t ret;
1625         int reason;
1626
1627         /*
1628          * If the device is not in running state we will reject some or all
1629          * commands.
1630          */
1631         if (unlikely(sdev->sdev_state != SDEV_RUNNING)) {
1632                 ret = scsi_prep_state_check(sdev, req);
1633                 if (ret != BLK_STS_OK)
1634                         goto out_put_budget;
1635         }
1636
1637         ret = BLK_STS_RESOURCE;
1638         if (!scsi_target_queue_ready(shost, sdev))
1639                 goto out_put_budget;
1640         if (!scsi_host_queue_ready(q, shost, sdev))
1641                 goto out_dec_target_busy;
1642
1643         if (!(req->rq_flags & RQF_DONTPREP)) {
1644                 ret = scsi_mq_prep_fn(req);
1645                 if (ret != BLK_STS_OK)
1646                         goto out_dec_host_busy;
1647                 req->rq_flags |= RQF_DONTPREP;
1648         } else {
1649                 clear_bit(SCMD_STATE_COMPLETE, &cmd->state);
1650                 blk_mq_start_request(req);
1651         }
1652
1653         if (sdev->simple_tags)
1654                 cmd->flags |= SCMD_TAGGED;
1655         else
1656                 cmd->flags &= ~SCMD_TAGGED;
1657
1658         scsi_init_cmd_errh(cmd);
1659         cmd->scsi_done = scsi_mq_done;
1660
1661         reason = scsi_dispatch_cmd(cmd);
1662         if (reason) {
1663                 scsi_set_blocked(cmd, reason);
1664                 ret = BLK_STS_RESOURCE;
1665                 goto out_dec_host_busy;
1666         }
1667
1668         return BLK_STS_OK;
1669
1670 out_dec_host_busy:
1671         scsi_dec_host_busy(shost);
1672 out_dec_target_busy:
1673         if (scsi_target(sdev)->can_queue > 0)
1674                 atomic_dec(&scsi_target(sdev)->target_busy);
1675 out_put_budget:
1676         scsi_mq_put_budget(hctx);
1677         switch (ret) {
1678         case BLK_STS_OK:
1679                 break;
1680         case BLK_STS_RESOURCE:
1681                 if (atomic_read(&sdev->device_busy) ||
1682                     scsi_device_blocked(sdev))
1683                         ret = BLK_STS_DEV_RESOURCE;
1684                 break;
1685         default:
1686                 if (unlikely(!scsi_device_online(sdev)))
1687                         scsi_req(req)->result = DID_NO_CONNECT << 16;
1688                 else
1689                         scsi_req(req)->result = DID_ERROR << 16;
1690                 /*
1691                  * Make sure to release all allocated resources when
1692                  * we hit an error, as we will never see this command
1693                  * again.
1694                  */
1695                 if (req->rq_flags & RQF_DONTPREP)
1696                         scsi_mq_uninit_cmd(cmd);
1697                 break;
1698         }
1699         return ret;
1700 }
1701
1702 static enum blk_eh_timer_return scsi_timeout(struct request *req,
1703                 bool reserved)
1704 {
1705         if (reserved)
1706                 return BLK_EH_RESET_TIMER;
1707         return scsi_times_out(req);
1708 }
1709
1710 static int scsi_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
1711                                 unsigned int hctx_idx, unsigned int numa_node)
1712 {
1713         struct Scsi_Host *shost = set->driver_data;
1714         const bool unchecked_isa_dma = shost->unchecked_isa_dma;
1715         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1716         struct scatterlist *sg;
1717
1718         if (unchecked_isa_dma)
1719                 cmd->flags |= SCMD_UNCHECKED_ISA_DMA;
1720         cmd->sense_buffer = scsi_alloc_sense_buffer(unchecked_isa_dma,
1721                                                     GFP_KERNEL, numa_node);
1722         if (!cmd->sense_buffer)
1723                 return -ENOMEM;
1724         cmd->req.sense = cmd->sense_buffer;
1725
1726         if (scsi_host_get_prot(shost)) {
1727                 sg = (void *)cmd + sizeof(struct scsi_cmnd) +
1728                         shost->hostt->cmd_size;
1729                 cmd->prot_sdb = (void *)sg + scsi_mq_sgl_size(shost);
1730         }
1731
1732         return 0;
1733 }
1734
1735 static void scsi_mq_exit_request(struct blk_mq_tag_set *set, struct request *rq,
1736                                  unsigned int hctx_idx)
1737 {
1738         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1739
1740         scsi_free_sense_buffer(cmd->flags & SCMD_UNCHECKED_ISA_DMA,
1741                                cmd->sense_buffer);
1742 }
1743
1744 static int scsi_map_queues(struct blk_mq_tag_set *set)
1745 {
1746         struct Scsi_Host *shost = container_of(set, struct Scsi_Host, tag_set);
1747
1748         if (shost->hostt->map_queues)
1749                 return shost->hostt->map_queues(shost);
1750         return blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
1751 }
1752
1753 void __scsi_init_queue(struct Scsi_Host *shost, struct request_queue *q)
1754 {
1755         struct device *dev = shost->dma_dev;
1756
1757         /*
1758          * this limit is imposed by hardware restrictions
1759          */
1760         blk_queue_max_segments(q, min_t(unsigned short, shost->sg_tablesize,
1761                                         SG_MAX_SEGMENTS));
1762
1763         if (scsi_host_prot_dma(shost)) {
1764                 shost->sg_prot_tablesize =
1765                         min_not_zero(shost->sg_prot_tablesize,
1766                                      (unsigned short)SCSI_MAX_PROT_SG_SEGMENTS);
1767                 BUG_ON(shost->sg_prot_tablesize < shost->sg_tablesize);
1768                 blk_queue_max_integrity_segments(q, shost->sg_prot_tablesize);
1769         }
1770
1771         blk_queue_max_hw_sectors(q, shost->max_sectors);
1772         if (shost->unchecked_isa_dma)
1773                 blk_queue_bounce_limit(q, BLK_BOUNCE_ISA);
1774         blk_queue_segment_boundary(q, shost->dma_boundary);
1775         dma_set_seg_boundary(dev, shost->dma_boundary);
1776
1777         blk_queue_max_segment_size(q, shost->max_segment_size);
1778         dma_set_max_seg_size(dev, shost->max_segment_size);
1779
1780         /*
1781          * Set a reasonable default alignment:  The larger of 32-byte (dword),
1782          * which is a common minimum for HBAs, and the minimum DMA alignment,
1783          * which is set by the platform.
1784          *
1785          * Devices that require a bigger alignment can increase it later.
1786          */
1787         blk_queue_dma_alignment(q, max(4, dma_get_cache_alignment()) - 1);
1788 }
1789 EXPORT_SYMBOL_GPL(__scsi_init_queue);
1790
1791 static const struct blk_mq_ops scsi_mq_ops = {
1792         .get_budget     = scsi_mq_get_budget,
1793         .put_budget     = scsi_mq_put_budget,
1794         .queue_rq       = scsi_queue_rq,
1795         .complete       = scsi_softirq_done,
1796         .timeout        = scsi_timeout,
1797 #ifdef CONFIG_BLK_DEBUG_FS
1798         .show_rq        = scsi_show_rq,
1799 #endif
1800         .init_request   = scsi_mq_init_request,
1801         .exit_request   = scsi_mq_exit_request,
1802         .initialize_rq_fn = scsi_initialize_rq,
1803         .busy           = scsi_mq_lld_busy,
1804         .map_queues     = scsi_map_queues,
1805 };
1806
1807 struct request_queue *scsi_mq_alloc_queue(struct scsi_device *sdev)
1808 {
1809         sdev->request_queue = blk_mq_init_queue(&sdev->host->tag_set);
1810         if (IS_ERR(sdev->request_queue))
1811                 return NULL;
1812
1813         sdev->request_queue->queuedata = sdev;
1814         __scsi_init_queue(sdev->host, sdev->request_queue);
1815         blk_queue_flag_set(QUEUE_FLAG_SCSI_PASSTHROUGH, sdev->request_queue);
1816         return sdev->request_queue;
1817 }
1818
1819 int scsi_mq_setup_tags(struct Scsi_Host *shost)
1820 {
1821         unsigned int cmd_size, sgl_size;
1822
1823         sgl_size = scsi_mq_sgl_size(shost);
1824         cmd_size = sizeof(struct scsi_cmnd) + shost->hostt->cmd_size + sgl_size;
1825         if (scsi_host_get_prot(shost))
1826                 cmd_size += sizeof(struct scsi_data_buffer) + sgl_size;
1827
1828         memset(&shost->tag_set, 0, sizeof(shost->tag_set));
1829         shost->tag_set.ops = &scsi_mq_ops;
1830         shost->tag_set.nr_hw_queues = shost->nr_hw_queues ? : 1;
1831         shost->tag_set.queue_depth = shost->can_queue;
1832         shost->tag_set.cmd_size = cmd_size;
1833         shost->tag_set.numa_node = NUMA_NO_NODE;
1834         shost->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
1835         shost->tag_set.flags |=
1836                 BLK_ALLOC_POLICY_TO_MQ_FLAG(shost->hostt->tag_alloc_policy);
1837         shost->tag_set.driver_data = shost;
1838
1839         return blk_mq_alloc_tag_set(&shost->tag_set);
1840 }
1841
1842 void scsi_mq_destroy_tags(struct Scsi_Host *shost)
1843 {
1844         blk_mq_free_tag_set(&shost->tag_set);
1845 }
1846
1847 /**
1848  * scsi_device_from_queue - return sdev associated with a request_queue
1849  * @q: The request queue to return the sdev from
1850  *
1851  * Return the sdev associated with a request queue or NULL if the
1852  * request_queue does not reference a SCSI device.
1853  */
1854 struct scsi_device *scsi_device_from_queue(struct request_queue *q)
1855 {
1856         struct scsi_device *sdev = NULL;
1857
1858         if (q->mq_ops == &scsi_mq_ops)
1859                 sdev = q->queuedata;
1860         if (!sdev || !get_device(&sdev->sdev_gendev))
1861                 sdev = NULL;
1862
1863         return sdev;
1864 }
1865 EXPORT_SYMBOL_GPL(scsi_device_from_queue);
1866
1867 /*
1868  * Function:    scsi_block_requests()
1869  *
1870  * Purpose:     Utility function used by low-level drivers to prevent further
1871  *              commands from being queued to the device.
1872  *
1873  * Arguments:   shost       - Host in question
1874  *
1875  * Returns:     Nothing
1876  *
1877  * Lock status: No locks are assumed held.
1878  *
1879  * Notes:       There is no timer nor any other means by which the requests
1880  *              get unblocked other than the low-level driver calling
1881  *              scsi_unblock_requests().
1882  */
1883 void scsi_block_requests(struct Scsi_Host *shost)
1884 {
1885         shost->host_self_blocked = 1;
1886 }
1887 EXPORT_SYMBOL(scsi_block_requests);
1888
1889 /*
1890  * Function:    scsi_unblock_requests()
1891  *
1892  * Purpose:     Utility function used by low-level drivers to allow further
1893  *              commands from being queued to the device.
1894  *
1895  * Arguments:   shost       - Host in question
1896  *
1897  * Returns:     Nothing
1898  *
1899  * Lock status: No locks are assumed held.
1900  *
1901  * Notes:       There is no timer nor any other means by which the requests
1902  *              get unblocked other than the low-level driver calling
1903  *              scsi_unblock_requests().
1904  *
1905  *              This is done as an API function so that changes to the
1906  *              internals of the scsi mid-layer won't require wholesale
1907  *              changes to drivers that use this feature.
1908  */
1909 void scsi_unblock_requests(struct Scsi_Host *shost)
1910 {
1911         shost->host_self_blocked = 0;
1912         scsi_run_host_queues(shost);
1913 }
1914 EXPORT_SYMBOL(scsi_unblock_requests);
1915
1916 int __init scsi_init_queue(void)
1917 {
1918         scsi_sdb_cache = kmem_cache_create("scsi_data_buffer",
1919                                            sizeof(struct scsi_data_buffer),
1920                                            0, 0, NULL);
1921         if (!scsi_sdb_cache) {
1922                 printk(KERN_ERR "SCSI: can't init scsi sdb cache\n");
1923                 return -ENOMEM;
1924         }
1925
1926         return 0;
1927 }
1928
1929 void scsi_exit_queue(void)
1930 {
1931         kmem_cache_destroy(scsi_sense_cache);
1932         kmem_cache_destroy(scsi_sense_isadma_cache);
1933         kmem_cache_destroy(scsi_sdb_cache);
1934 }
1935
1936 /**
1937  *      scsi_mode_select - issue a mode select
1938  *      @sdev:  SCSI device to be queried
1939  *      @pf:    Page format bit (1 == standard, 0 == vendor specific)
1940  *      @sp:    Save page bit (0 == don't save, 1 == save)
1941  *      @modepage: mode page being requested
1942  *      @buffer: request buffer (may not be smaller than eight bytes)
1943  *      @len:   length of request buffer.
1944  *      @timeout: command timeout
1945  *      @retries: number of retries before failing
1946  *      @data: returns a structure abstracting the mode header data
1947  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
1948  *              must be SCSI_SENSE_BUFFERSIZE big.
1949  *
1950  *      Returns zero if successful; negative error number or scsi
1951  *      status on error
1952  *
1953  */
1954 int
1955 scsi_mode_select(struct scsi_device *sdev, int pf, int sp, int modepage,
1956                  unsigned char *buffer, int len, int timeout, int retries,
1957                  struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1958 {
1959         unsigned char cmd[10];
1960         unsigned char *real_buffer;
1961         int ret;
1962
1963         memset(cmd, 0, sizeof(cmd));
1964         cmd[1] = (pf ? 0x10 : 0) | (sp ? 0x01 : 0);
1965
1966         if (sdev->use_10_for_ms) {
1967                 if (len > 65535)
1968                         return -EINVAL;
1969                 real_buffer = kmalloc(8 + len, GFP_KERNEL);
1970                 if (!real_buffer)
1971                         return -ENOMEM;
1972                 memcpy(real_buffer + 8, buffer, len);
1973                 len += 8;
1974                 real_buffer[0] = 0;
1975                 real_buffer[1] = 0;
1976                 real_buffer[2] = data->medium_type;
1977                 real_buffer[3] = data->device_specific;
1978                 real_buffer[4] = data->longlba ? 0x01 : 0;
1979                 real_buffer[5] = 0;
1980                 real_buffer[6] = data->block_descriptor_length >> 8;
1981                 real_buffer[7] = data->block_descriptor_length;
1982
1983                 cmd[0] = MODE_SELECT_10;
1984                 cmd[7] = len >> 8;
1985                 cmd[8] = len;
1986         } else {
1987                 if (len > 255 || data->block_descriptor_length > 255 ||
1988                     data->longlba)
1989                         return -EINVAL;
1990
1991                 real_buffer = kmalloc(4 + len, GFP_KERNEL);
1992                 if (!real_buffer)
1993                         return -ENOMEM;
1994                 memcpy(real_buffer + 4, buffer, len);
1995                 len += 4;
1996                 real_buffer[0] = 0;
1997                 real_buffer[1] = data->medium_type;
1998                 real_buffer[2] = data->device_specific;
1999                 real_buffer[3] = data->block_descriptor_length;
2000                 
2001
2002                 cmd[0] = MODE_SELECT;
2003                 cmd[4] = len;
2004         }
2005
2006         ret = scsi_execute_req(sdev, cmd, DMA_TO_DEVICE, real_buffer, len,
2007                                sshdr, timeout, retries, NULL);
2008         kfree(real_buffer);
2009         return ret;
2010 }
2011 EXPORT_SYMBOL_GPL(scsi_mode_select);
2012
2013 /**
2014  *      scsi_mode_sense - issue a mode sense, falling back from 10 to six bytes if necessary.
2015  *      @sdev:  SCSI device to be queried
2016  *      @dbd:   set if mode sense will allow block descriptors to be returned
2017  *      @modepage: mode page being requested
2018  *      @buffer: request buffer (may not be smaller than eight bytes)
2019  *      @len:   length of request buffer.
2020  *      @timeout: command timeout
2021  *      @retries: number of retries before failing
2022  *      @data: returns a structure abstracting the mode header data
2023  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
2024  *              must be SCSI_SENSE_BUFFERSIZE big.
2025  *
2026  *      Returns zero if unsuccessful, or the header offset (either 4
2027  *      or 8 depending on whether a six or ten byte command was
2028  *      issued) if successful.
2029  */
2030 int
2031 scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
2032                   unsigned char *buffer, int len, int timeout, int retries,
2033                   struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
2034 {
2035         unsigned char cmd[12];
2036         int use_10_for_ms;
2037         int header_length;
2038         int result, retry_count = retries;
2039         struct scsi_sense_hdr my_sshdr;
2040
2041         memset(data, 0, sizeof(*data));
2042         memset(&cmd[0], 0, 12);
2043         cmd[1] = dbd & 0x18;    /* allows DBD and LLBA bits */
2044         cmd[2] = modepage;
2045
2046         /* caller might not be interested in sense, but we need it */
2047         if (!sshdr)
2048                 sshdr = &my_sshdr;
2049
2050  retry:
2051         use_10_for_ms = sdev->use_10_for_ms;
2052
2053         if (use_10_for_ms) {
2054                 if (len < 8)
2055                         len = 8;
2056
2057                 cmd[0] = MODE_SENSE_10;
2058                 cmd[8] = len;
2059                 header_length = 8;
2060         } else {
2061                 if (len < 4)
2062                         len = 4;
2063
2064                 cmd[0] = MODE_SENSE;
2065                 cmd[4] = len;
2066                 header_length = 4;
2067         }
2068
2069         memset(buffer, 0, len);
2070
2071         result = scsi_execute_req(sdev, cmd, DMA_FROM_DEVICE, buffer, len,
2072                                   sshdr, timeout, retries, NULL);
2073
2074         /* This code looks awful: what it's doing is making sure an
2075          * ILLEGAL REQUEST sense return identifies the actual command
2076          * byte as the problem.  MODE_SENSE commands can return
2077          * ILLEGAL REQUEST if the code page isn't supported */
2078
2079         if (use_10_for_ms && !scsi_status_is_good(result) &&
2080             driver_byte(result) == DRIVER_SENSE) {
2081                 if (scsi_sense_valid(sshdr)) {
2082                         if ((sshdr->sense_key == ILLEGAL_REQUEST) &&
2083                             (sshdr->asc == 0x20) && (sshdr->ascq == 0)) {
2084                                 /* 
2085                                  * Invalid command operation code
2086                                  */
2087                                 sdev->use_10_for_ms = 0;
2088                                 goto retry;
2089                         }
2090                 }
2091         }
2092
2093         if(scsi_status_is_good(result)) {
2094                 if (unlikely(buffer[0] == 0x86 && buffer[1] == 0x0b &&
2095                              (modepage == 6 || modepage == 8))) {
2096                         /* Initio breakage? */
2097                         header_length = 0;
2098                         data->length = 13;
2099                         data->medium_type = 0;
2100                         data->device_specific = 0;
2101                         data->longlba = 0;
2102                         data->block_descriptor_length = 0;
2103                 } else if(use_10_for_ms) {
2104                         data->length = buffer[0]*256 + buffer[1] + 2;
2105                         data->medium_type = buffer[2];
2106                         data->device_specific = buffer[3];
2107                         data->longlba = buffer[4] & 0x01;
2108                         data->block_descriptor_length = buffer[6]*256
2109                                 + buffer[7];
2110                 } else {
2111                         data->length = buffer[0] + 1;
2112                         data->medium_type = buffer[1];
2113                         data->device_specific = buffer[2];
2114                         data->block_descriptor_length = buffer[3];
2115                 }
2116                 data->header_length = header_length;
2117         } else if ((status_byte(result) == CHECK_CONDITION) &&
2118                    scsi_sense_valid(sshdr) &&
2119                    sshdr->sense_key == UNIT_ATTENTION && retry_count) {
2120                 retry_count--;
2121                 goto retry;
2122         }
2123
2124         return result;
2125 }
2126 EXPORT_SYMBOL(scsi_mode_sense);
2127
2128 /**
2129  *      scsi_test_unit_ready - test if unit is ready
2130  *      @sdev:  scsi device to change the state of.
2131  *      @timeout: command timeout
2132  *      @retries: number of retries before failing
2133  *      @sshdr: outpout pointer for decoded sense information.
2134  *
2135  *      Returns zero if unsuccessful or an error if TUR failed.  For
2136  *      removable media, UNIT_ATTENTION sets ->changed flag.
2137  **/
2138 int
2139 scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries,
2140                      struct scsi_sense_hdr *sshdr)
2141 {
2142         char cmd[] = {
2143                 TEST_UNIT_READY, 0, 0, 0, 0, 0,
2144         };
2145         int result;
2146
2147         /* try to eat the UNIT_ATTENTION if there are enough retries */
2148         do {
2149                 result = scsi_execute_req(sdev, cmd, DMA_NONE, NULL, 0, sshdr,
2150                                           timeout, 1, NULL);
2151                 if (sdev->removable && scsi_sense_valid(sshdr) &&
2152                     sshdr->sense_key == UNIT_ATTENTION)
2153                         sdev->changed = 1;
2154         } while (scsi_sense_valid(sshdr) &&
2155                  sshdr->sense_key == UNIT_ATTENTION && --retries);
2156
2157         return result;
2158 }
2159 EXPORT_SYMBOL(scsi_test_unit_ready);
2160
2161 /**
2162  *      scsi_device_set_state - Take the given device through the device state model.
2163  *      @sdev:  scsi device to change the state of.
2164  *      @state: state to change to.
2165  *
2166  *      Returns zero if successful or an error if the requested
2167  *      transition is illegal.
2168  */
2169 int
2170 scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
2171 {
2172         enum scsi_device_state oldstate = sdev->sdev_state;
2173
2174         if (state == oldstate)
2175                 return 0;
2176
2177         switch (state) {
2178         case SDEV_CREATED:
2179                 switch (oldstate) {
2180                 case SDEV_CREATED_BLOCK:
2181                         break;
2182                 default:
2183                         goto illegal;
2184                 }
2185                 break;
2186                         
2187         case SDEV_RUNNING:
2188                 switch (oldstate) {
2189                 case SDEV_CREATED:
2190                 case SDEV_OFFLINE:
2191                 case SDEV_TRANSPORT_OFFLINE:
2192                 case SDEV_QUIESCE:
2193                 case SDEV_BLOCK:
2194                         break;
2195                 default:
2196                         goto illegal;
2197                 }
2198                 break;
2199
2200         case SDEV_QUIESCE:
2201                 switch (oldstate) {
2202                 case SDEV_RUNNING:
2203                 case SDEV_OFFLINE:
2204                 case SDEV_TRANSPORT_OFFLINE:
2205                         break;
2206                 default:
2207                         goto illegal;
2208                 }
2209                 break;
2210
2211         case SDEV_OFFLINE:
2212         case SDEV_TRANSPORT_OFFLINE:
2213                 switch (oldstate) {
2214                 case SDEV_CREATED:
2215                 case SDEV_RUNNING:
2216                 case SDEV_QUIESCE:
2217                 case SDEV_BLOCK:
2218                         break;
2219                 default:
2220                         goto illegal;
2221                 }
2222                 break;
2223
2224         case SDEV_BLOCK:
2225                 switch (oldstate) {
2226                 case SDEV_RUNNING:
2227                 case SDEV_CREATED_BLOCK:
2228                 case SDEV_OFFLINE:
2229                         break;
2230                 default:
2231                         goto illegal;
2232                 }
2233                 break;
2234
2235         case SDEV_CREATED_BLOCK:
2236                 switch (oldstate) {
2237                 case SDEV_CREATED:
2238                         break;
2239                 default:
2240                         goto illegal;
2241                 }
2242                 break;
2243
2244         case SDEV_CANCEL:
2245                 switch (oldstate) {
2246                 case SDEV_CREATED:
2247                 case SDEV_RUNNING:
2248                 case SDEV_QUIESCE:
2249                 case SDEV_OFFLINE:
2250                 case SDEV_TRANSPORT_OFFLINE:
2251                         break;
2252                 default:
2253                         goto illegal;
2254                 }
2255                 break;
2256
2257         case SDEV_DEL:
2258                 switch (oldstate) {
2259                 case SDEV_CREATED:
2260                 case SDEV_RUNNING:
2261                 case SDEV_OFFLINE:
2262                 case SDEV_TRANSPORT_OFFLINE:
2263                 case SDEV_CANCEL:
2264                 case SDEV_BLOCK:
2265                 case SDEV_CREATED_BLOCK:
2266                         break;
2267                 default:
2268                         goto illegal;
2269                 }
2270                 break;
2271
2272         }
2273         sdev->sdev_state = state;
2274         return 0;
2275
2276  illegal:
2277         SCSI_LOG_ERROR_RECOVERY(1,
2278                                 sdev_printk(KERN_ERR, sdev,
2279                                             "Illegal state transition %s->%s",
2280                                             scsi_device_state_name(oldstate),
2281                                             scsi_device_state_name(state))
2282                                 );
2283         return -EINVAL;
2284 }
2285 EXPORT_SYMBOL(scsi_device_set_state);
2286
2287 /**
2288  *      sdev_evt_emit - emit a single SCSI device uevent
2289  *      @sdev: associated SCSI device
2290  *      @evt: event to emit
2291  *
2292  *      Send a single uevent (scsi_event) to the associated scsi_device.
2293  */
2294 static void scsi_evt_emit(struct scsi_device *sdev, struct scsi_event *evt)
2295 {
2296         int idx = 0;
2297         char *envp[3];
2298
2299         switch (evt->evt_type) {
2300         case SDEV_EVT_MEDIA_CHANGE:
2301                 envp[idx++] = "SDEV_MEDIA_CHANGE=1";
2302                 break;
2303         case SDEV_EVT_INQUIRY_CHANGE_REPORTED:
2304                 scsi_rescan_device(&sdev->sdev_gendev);
2305                 envp[idx++] = "SDEV_UA=INQUIRY_DATA_HAS_CHANGED";
2306                 break;
2307         case SDEV_EVT_CAPACITY_CHANGE_REPORTED:
2308                 envp[idx++] = "SDEV_UA=CAPACITY_DATA_HAS_CHANGED";
2309                 break;
2310         case SDEV_EVT_SOFT_THRESHOLD_REACHED_REPORTED:
2311                envp[idx++] = "SDEV_UA=THIN_PROVISIONING_SOFT_THRESHOLD_REACHED";
2312                 break;
2313         case SDEV_EVT_MODE_PARAMETER_CHANGE_REPORTED:
2314                 envp[idx++] = "SDEV_UA=MODE_PARAMETERS_CHANGED";
2315                 break;
2316         case SDEV_EVT_LUN_CHANGE_REPORTED:
2317                 envp[idx++] = "SDEV_UA=REPORTED_LUNS_DATA_HAS_CHANGED";
2318                 break;
2319         case SDEV_EVT_ALUA_STATE_CHANGE_REPORTED:
2320                 envp[idx++] = "SDEV_UA=ASYMMETRIC_ACCESS_STATE_CHANGED";
2321                 break;
2322         case SDEV_EVT_POWER_ON_RESET_OCCURRED:
2323                 envp[idx++] = "SDEV_UA=POWER_ON_RESET_OCCURRED";
2324                 break;
2325         default:
2326                 /* do nothing */
2327                 break;
2328         }
2329
2330         envp[idx++] = NULL;
2331
2332         kobject_uevent_env(&sdev->sdev_gendev.kobj, KOBJ_CHANGE, envp);
2333 }
2334
2335 /**
2336  *      sdev_evt_thread - send a uevent for each scsi event
2337  *      @work: work struct for scsi_device
2338  *
2339  *      Dispatch queued events to their associated scsi_device kobjects
2340  *      as uevents.
2341  */
2342 void scsi_evt_thread(struct work_struct *work)
2343 {
2344         struct scsi_device *sdev;
2345         enum scsi_device_event evt_type;
2346         LIST_HEAD(event_list);
2347
2348         sdev = container_of(work, struct scsi_device, event_work);
2349
2350         for (evt_type = SDEV_EVT_FIRST; evt_type <= SDEV_EVT_LAST; evt_type++)
2351                 if (test_and_clear_bit(evt_type, sdev->pending_events))
2352                         sdev_evt_send_simple(sdev, evt_type, GFP_KERNEL);
2353
2354         while (1) {
2355                 struct scsi_event *evt;
2356                 struct list_head *this, *tmp;
2357                 unsigned long flags;
2358
2359                 spin_lock_irqsave(&sdev->list_lock, flags);
2360                 list_splice_init(&sdev->event_list, &event_list);
2361                 spin_unlock_irqrestore(&sdev->list_lock, flags);
2362
2363                 if (list_empty(&event_list))
2364                         break;
2365
2366                 list_for_each_safe(this, tmp, &event_list) {
2367                         evt = list_entry(this, struct scsi_event, node);
2368                         list_del(&evt->node);
2369                         scsi_evt_emit(sdev, evt);
2370                         kfree(evt);
2371                 }
2372         }
2373 }
2374
2375 /**
2376  *      sdev_evt_send - send asserted event to uevent thread
2377  *      @sdev: scsi_device event occurred on
2378  *      @evt: event to send
2379  *
2380  *      Assert scsi device event asynchronously.
2381  */
2382 void sdev_evt_send(struct scsi_device *sdev, struct scsi_event *evt)
2383 {
2384         unsigned long flags;
2385
2386 #if 0
2387         /* FIXME: currently this check eliminates all media change events
2388          * for polled devices.  Need to update to discriminate between AN
2389          * and polled events */
2390         if (!test_bit(evt->evt_type, sdev->supported_events)) {
2391                 kfree(evt);
2392                 return;
2393         }
2394 #endif
2395
2396         spin_lock_irqsave(&sdev->list_lock, flags);
2397         list_add_tail(&evt->node, &sdev->event_list);
2398         schedule_work(&sdev->event_work);
2399         spin_unlock_irqrestore(&sdev->list_lock, flags);
2400 }
2401 EXPORT_SYMBOL_GPL(sdev_evt_send);
2402
2403 /**
2404  *      sdev_evt_alloc - allocate a new scsi event
2405  *      @evt_type: type of event to allocate
2406  *      @gfpflags: GFP flags for allocation
2407  *
2408  *      Allocates and returns a new scsi_event.
2409  */
2410 struct scsi_event *sdev_evt_alloc(enum scsi_device_event evt_type,
2411                                   gfp_t gfpflags)
2412 {
2413         struct scsi_event *evt = kzalloc(sizeof(struct scsi_event), gfpflags);
2414         if (!evt)
2415                 return NULL;
2416
2417         evt->evt_type = evt_type;
2418         INIT_LIST_HEAD(&evt->node);
2419
2420         /* evt_type-specific initialization, if any */
2421         switch (evt_type) {
2422         case SDEV_EVT_MEDIA_CHANGE:
2423         case SDEV_EVT_INQUIRY_CHANGE_REPORTED:
2424         case SDEV_EVT_CAPACITY_CHANGE_REPORTED:
2425         case SDEV_EVT_SOFT_THRESHOLD_REACHED_REPORTED:
2426         case SDEV_EVT_MODE_PARAMETER_CHANGE_REPORTED:
2427         case SDEV_EVT_LUN_CHANGE_REPORTED:
2428         case SDEV_EVT_ALUA_STATE_CHANGE_REPORTED:
2429         case SDEV_EVT_POWER_ON_RESET_OCCURRED:
2430         default:
2431                 /* do nothing */
2432                 break;
2433         }
2434
2435         return evt;
2436 }
2437 EXPORT_SYMBOL_GPL(sdev_evt_alloc);
2438
2439 /**
2440  *      sdev_evt_send_simple - send asserted event to uevent thread
2441  *      @sdev: scsi_device event occurred on
2442  *      @evt_type: type of event to send
2443  *      @gfpflags: GFP flags for allocation
2444  *
2445  *      Assert scsi device event asynchronously, given an event type.
2446  */
2447 void sdev_evt_send_simple(struct scsi_device *sdev,
2448                           enum scsi_device_event evt_type, gfp_t gfpflags)
2449 {
2450         struct scsi_event *evt = sdev_evt_alloc(evt_type, gfpflags);
2451         if (!evt) {
2452                 sdev_printk(KERN_ERR, sdev, "event %d eaten due to OOM\n",
2453                             evt_type);
2454                 return;
2455         }
2456
2457         sdev_evt_send(sdev, evt);
2458 }
2459 EXPORT_SYMBOL_GPL(sdev_evt_send_simple);
2460
2461 /**
2462  *      scsi_device_quiesce - Block user issued commands.
2463  *      @sdev:  scsi device to quiesce.
2464  *
2465  *      This works by trying to transition to the SDEV_QUIESCE state
2466  *      (which must be a legal transition).  When the device is in this
2467  *      state, only special requests will be accepted, all others will
2468  *      be deferred.  Since special requests may also be requeued requests,
2469  *      a successful return doesn't guarantee the device will be 
2470  *      totally quiescent.
2471  *
2472  *      Must be called with user context, may sleep.
2473  *
2474  *      Returns zero if unsuccessful or an error if not.
2475  */
2476 int
2477 scsi_device_quiesce(struct scsi_device *sdev)
2478 {
2479         struct request_queue *q = sdev->request_queue;
2480         int err;
2481
2482         /*
2483          * It is allowed to call scsi_device_quiesce() multiple times from
2484          * the same context but concurrent scsi_device_quiesce() calls are
2485          * not allowed.
2486          */
2487         WARN_ON_ONCE(sdev->quiesced_by && sdev->quiesced_by != current);
2488
2489         if (sdev->quiesced_by == current)
2490                 return 0;
2491
2492         blk_set_pm_only(q);
2493
2494         blk_mq_freeze_queue(q);
2495         /*
2496          * Ensure that the effect of blk_set_pm_only() will be visible
2497          * for percpu_ref_tryget() callers that occur after the queue
2498          * unfreeze even if the queue was already frozen before this function
2499          * was called. See also https://lwn.net/Articles/573497/.
2500          */
2501         synchronize_rcu();
2502         blk_mq_unfreeze_queue(q);
2503
2504         mutex_lock(&sdev->state_mutex);
2505         err = scsi_device_set_state(sdev, SDEV_QUIESCE);
2506         if (err == 0)
2507                 sdev->quiesced_by = current;
2508         else
2509                 blk_clear_pm_only(q);
2510         mutex_unlock(&sdev->state_mutex);
2511
2512         return err;
2513 }
2514 EXPORT_SYMBOL(scsi_device_quiesce);
2515
2516 /**
2517  *      scsi_device_resume - Restart user issued commands to a quiesced device.
2518  *      @sdev:  scsi device to resume.
2519  *
2520  *      Moves the device from quiesced back to running and restarts the
2521  *      queues.
2522  *
2523  *      Must be called with user context, may sleep.
2524  */
2525 void scsi_device_resume(struct scsi_device *sdev)
2526 {
2527         /* check if the device state was mutated prior to resume, and if
2528          * so assume the state is being managed elsewhere (for example
2529          * device deleted during suspend)
2530          */
2531         mutex_lock(&sdev->state_mutex);
2532         if (sdev->quiesced_by) {
2533                 sdev->quiesced_by = NULL;
2534                 blk_clear_pm_only(sdev->request_queue);
2535         }
2536         if (sdev->sdev_state == SDEV_QUIESCE)
2537                 scsi_device_set_state(sdev, SDEV_RUNNING);
2538         mutex_unlock(&sdev->state_mutex);
2539 }
2540 EXPORT_SYMBOL(scsi_device_resume);
2541
2542 static void
2543 device_quiesce_fn(struct scsi_device *sdev, void *data)
2544 {
2545         scsi_device_quiesce(sdev);
2546 }
2547
2548 void
2549 scsi_target_quiesce(struct scsi_target *starget)
2550 {
2551         starget_for_each_device(starget, NULL, device_quiesce_fn);
2552 }
2553 EXPORT_SYMBOL(scsi_target_quiesce);
2554
2555 static void
2556 device_resume_fn(struct scsi_device *sdev, void *data)
2557 {
2558         scsi_device_resume(sdev);
2559 }
2560
2561 void
2562 scsi_target_resume(struct scsi_target *starget)
2563 {
2564         starget_for_each_device(starget, NULL, device_resume_fn);
2565 }
2566 EXPORT_SYMBOL(scsi_target_resume);
2567
2568 /**
2569  * scsi_internal_device_block_nowait - try to transition to the SDEV_BLOCK state
2570  * @sdev: device to block
2571  *
2572  * Pause SCSI command processing on the specified device. Does not sleep.
2573  *
2574  * Returns zero if successful or a negative error code upon failure.
2575  *
2576  * Notes:
2577  * This routine transitions the device to the SDEV_BLOCK state (which must be
2578  * a legal transition). When the device is in this state, command processing
2579  * is paused until the device leaves the SDEV_BLOCK state. See also
2580  * scsi_internal_device_unblock_nowait().
2581  */
2582 int scsi_internal_device_block_nowait(struct scsi_device *sdev)
2583 {
2584         struct request_queue *q = sdev->request_queue;
2585         int err = 0;
2586
2587         err = scsi_device_set_state(sdev, SDEV_BLOCK);
2588         if (err) {
2589                 err = scsi_device_set_state(sdev, SDEV_CREATED_BLOCK);
2590
2591                 if (err)
2592                         return err;
2593         }
2594
2595         /* 
2596          * The device has transitioned to SDEV_BLOCK.  Stop the
2597          * block layer from calling the midlayer with this device's
2598          * request queue. 
2599          */
2600         blk_mq_quiesce_queue_nowait(q);
2601         return 0;
2602 }
2603 EXPORT_SYMBOL_GPL(scsi_internal_device_block_nowait);
2604
2605 /**
2606  * scsi_internal_device_block - try to transition to the SDEV_BLOCK state
2607  * @sdev: device to block
2608  *
2609  * Pause SCSI command processing on the specified device and wait until all
2610  * ongoing scsi_request_fn() / scsi_queue_rq() calls have finished. May sleep.
2611  *
2612  * Returns zero if successful or a negative error code upon failure.
2613  *
2614  * Note:
2615  * This routine transitions the device to the SDEV_BLOCK state (which must be
2616  * a legal transition). When the device is in this state, command processing
2617  * is paused until the device leaves the SDEV_BLOCK state. See also
2618  * scsi_internal_device_unblock().
2619  *
2620  * To do: avoid that scsi_send_eh_cmnd() calls queuecommand() after
2621  * scsi_internal_device_block() has blocked a SCSI device and also
2622  * remove the rport mutex lock and unlock calls from srp_queuecommand().
2623  */
2624 static int scsi_internal_device_block(struct scsi_device *sdev)
2625 {
2626         struct request_queue *q = sdev->request_queue;
2627         int err;
2628
2629         mutex_lock(&sdev->state_mutex);
2630         err = scsi_internal_device_block_nowait(sdev);
2631         if (err == 0)
2632                 blk_mq_quiesce_queue(q);
2633         mutex_unlock(&sdev->state_mutex);
2634
2635         return err;
2636 }
2637  
2638 void scsi_start_queue(struct scsi_device *sdev)
2639 {
2640         struct request_queue *q = sdev->request_queue;
2641
2642         blk_mq_unquiesce_queue(q);
2643 }
2644
2645 /**
2646  * scsi_internal_device_unblock_nowait - resume a device after a block request
2647  * @sdev:       device to resume
2648  * @new_state:  state to set the device to after unblocking
2649  *
2650  * Restart the device queue for a previously suspended SCSI device. Does not
2651  * sleep.
2652  *
2653  * Returns zero if successful or a negative error code upon failure.
2654  *
2655  * Notes:
2656  * This routine transitions the device to the SDEV_RUNNING state or to one of
2657  * the offline states (which must be a legal transition) allowing the midlayer
2658  * to goose the queue for this device.
2659  */
2660 int scsi_internal_device_unblock_nowait(struct scsi_device *sdev,
2661                                         enum scsi_device_state new_state)
2662 {
2663         /*
2664          * Try to transition the scsi device to SDEV_RUNNING or one of the
2665          * offlined states and goose the device queue if successful.
2666          */
2667         switch (sdev->sdev_state) {
2668         case SDEV_BLOCK:
2669         case SDEV_TRANSPORT_OFFLINE:
2670                 sdev->sdev_state = new_state;
2671                 break;
2672         case SDEV_CREATED_BLOCK:
2673                 if (new_state == SDEV_TRANSPORT_OFFLINE ||
2674                     new_state == SDEV_OFFLINE)
2675                         sdev->sdev_state = new_state;
2676                 else
2677                         sdev->sdev_state = SDEV_CREATED;
2678                 break;
2679         case SDEV_CANCEL:
2680         case SDEV_OFFLINE:
2681                 break;
2682         default:
2683                 return -EINVAL;
2684         }
2685         scsi_start_queue(sdev);
2686
2687         return 0;
2688 }
2689 EXPORT_SYMBOL_GPL(scsi_internal_device_unblock_nowait);
2690
2691 /**
2692  * scsi_internal_device_unblock - resume a device after a block request
2693  * @sdev:       device to resume
2694  * @new_state:  state to set the device to after unblocking
2695  *
2696  * Restart the device queue for a previously suspended SCSI device. May sleep.
2697  *
2698  * Returns zero if successful or a negative error code upon failure.
2699  *
2700  * Notes:
2701  * This routine transitions the device to the SDEV_RUNNING state or to one of
2702  * the offline states (which must be a legal transition) allowing the midlayer
2703  * to goose the queue for this device.
2704  */
2705 static int scsi_internal_device_unblock(struct scsi_device *sdev,
2706                                         enum scsi_device_state new_state)
2707 {
2708         int ret;
2709
2710         mutex_lock(&sdev->state_mutex);
2711         ret = scsi_internal_device_unblock_nowait(sdev, new_state);
2712         mutex_unlock(&sdev->state_mutex);
2713
2714         return ret;
2715 }
2716
2717 static void
2718 device_block(struct scsi_device *sdev, void *data)
2719 {
2720         scsi_internal_device_block(sdev);
2721 }
2722
2723 static int
2724 target_block(struct device *dev, void *data)
2725 {
2726         if (scsi_is_target_device(dev))
2727                 starget_for_each_device(to_scsi_target(dev), NULL,
2728                                         device_block);
2729         return 0;
2730 }
2731
2732 void
2733 scsi_target_block(struct device *dev)
2734 {
2735         if (scsi_is_target_device(dev))
2736                 starget_for_each_device(to_scsi_target(dev), NULL,
2737                                         device_block);
2738         else
2739                 device_for_each_child(dev, NULL, target_block);
2740 }
2741 EXPORT_SYMBOL_GPL(scsi_target_block);
2742
2743 static void
2744 device_unblock(struct scsi_device *sdev, void *data)
2745 {
2746         scsi_internal_device_unblock(sdev, *(enum scsi_device_state *)data);
2747 }
2748
2749 static int
2750 target_unblock(struct device *dev, void *data)
2751 {
2752         if (scsi_is_target_device(dev))
2753                 starget_for_each_device(to_scsi_target(dev), data,
2754                                         device_unblock);
2755         return 0;
2756 }
2757
2758 void
2759 scsi_target_unblock(struct device *dev, enum scsi_device_state new_state)
2760 {
2761         if (scsi_is_target_device(dev))
2762                 starget_for_each_device(to_scsi_target(dev), &new_state,
2763                                         device_unblock);
2764         else
2765                 device_for_each_child(dev, &new_state, target_unblock);
2766 }
2767 EXPORT_SYMBOL_GPL(scsi_target_unblock);
2768
2769 /**
2770  * scsi_kmap_atomic_sg - find and atomically map an sg-elemnt
2771  * @sgl:        scatter-gather list
2772  * @sg_count:   number of segments in sg
2773  * @offset:     offset in bytes into sg, on return offset into the mapped area
2774  * @len:        bytes to map, on return number of bytes mapped
2775  *
2776  * Returns virtual address of the start of the mapped page
2777  */
2778 void *scsi_kmap_atomic_sg(struct scatterlist *sgl, int sg_count,
2779                           size_t *offset, size_t *len)
2780 {
2781         int i;
2782         size_t sg_len = 0, len_complete = 0;
2783         struct scatterlist *sg;
2784         struct page *page;
2785
2786         WARN_ON(!irqs_disabled());
2787
2788         for_each_sg(sgl, sg, sg_count, i) {
2789                 len_complete = sg_len; /* Complete sg-entries */
2790                 sg_len += sg->length;
2791                 if (sg_len > *offset)
2792                         break;
2793         }
2794
2795         if (unlikely(i == sg_count)) {
2796                 printk(KERN_ERR "%s: Bytes in sg: %zu, requested offset %zu, "
2797                         "elements %d\n",
2798                        __func__, sg_len, *offset, sg_count);
2799                 WARN_ON(1);
2800                 return NULL;
2801         }
2802
2803         /* Offset starting from the beginning of first page in this sg-entry */
2804         *offset = *offset - len_complete + sg->offset;
2805
2806         /* Assumption: contiguous pages can be accessed as "page + i" */
2807         page = nth_page(sg_page(sg), (*offset >> PAGE_SHIFT));
2808         *offset &= ~PAGE_MASK;
2809
2810         /* Bytes in this sg-entry from *offset to the end of the page */
2811         sg_len = PAGE_SIZE - *offset;
2812         if (*len > sg_len)
2813                 *len = sg_len;
2814
2815         return kmap_atomic(page);
2816 }
2817 EXPORT_SYMBOL(scsi_kmap_atomic_sg);
2818
2819 /**
2820  * scsi_kunmap_atomic_sg - atomically unmap a virtual address, previously mapped with scsi_kmap_atomic_sg
2821  * @virt:       virtual address to be unmapped
2822  */
2823 void scsi_kunmap_atomic_sg(void *virt)
2824 {
2825         kunmap_atomic(virt);
2826 }
2827 EXPORT_SYMBOL(scsi_kunmap_atomic_sg);
2828
2829 void sdev_disable_disk_events(struct scsi_device *sdev)
2830 {
2831         atomic_inc(&sdev->disk_events_disable_depth);
2832 }
2833 EXPORT_SYMBOL(sdev_disable_disk_events);
2834
2835 void sdev_enable_disk_events(struct scsi_device *sdev)
2836 {
2837         if (WARN_ON_ONCE(atomic_read(&sdev->disk_events_disable_depth) <= 0))
2838                 return;
2839         atomic_dec(&sdev->disk_events_disable_depth);
2840 }
2841 EXPORT_SYMBOL(sdev_enable_disk_events);
2842
2843 /**
2844  * scsi_vpd_lun_id - return a unique device identification
2845  * @sdev: SCSI device
2846  * @id:   buffer for the identification
2847  * @id_len:  length of the buffer
2848  *
2849  * Copies a unique device identification into @id based
2850  * on the information in the VPD page 0x83 of the device.
2851  * The string will be formatted as a SCSI name string.
2852  *
2853  * Returns the length of the identification or error on failure.
2854  * If the identifier is longer than the supplied buffer the actual
2855  * identifier length is returned and the buffer is not zero-padded.
2856  */
2857 int scsi_vpd_lun_id(struct scsi_device *sdev, char *id, size_t id_len)
2858 {
2859         u8 cur_id_type = 0xff;
2860         u8 cur_id_size = 0;
2861         const unsigned char *d, *cur_id_str;
2862         const struct scsi_vpd *vpd_pg83;
2863         int id_size = -EINVAL;
2864
2865         rcu_read_lock();
2866         vpd_pg83 = rcu_dereference(sdev->vpd_pg83);
2867         if (!vpd_pg83) {
2868                 rcu_read_unlock();
2869                 return -ENXIO;
2870         }
2871
2872         /*
2873          * Look for the correct descriptor.
2874          * Order of preference for lun descriptor:
2875          * - SCSI name string
2876          * - NAA IEEE Registered Extended
2877          * - EUI-64 based 16-byte
2878          * - EUI-64 based 12-byte
2879          * - NAA IEEE Registered
2880          * - NAA IEEE Extended
2881          * - T10 Vendor ID
2882          * as longer descriptors reduce the likelyhood
2883          * of identification clashes.
2884          */
2885
2886         /* The id string must be at least 20 bytes + terminating NULL byte */
2887         if (id_len < 21) {
2888                 rcu_read_unlock();
2889                 return -EINVAL;
2890         }
2891
2892         memset(id, 0, id_len);
2893         d = vpd_pg83->data + 4;
2894         while (d < vpd_pg83->data + vpd_pg83->len) {
2895                 /* Skip designators not referring to the LUN */
2896                 if ((d[1] & 0x30) != 0x00)
2897                         goto next_desig;
2898
2899                 switch (d[1] & 0xf) {
2900                 case 0x1:
2901                         /* T10 Vendor ID */
2902                         if (cur_id_size > d[3])
2903                                 break;
2904                         /* Prefer anything */
2905                         if (cur_id_type > 0x01 && cur_id_type != 0xff)
2906                                 break;
2907                         cur_id_size = d[3];
2908                         if (cur_id_size + 4 > id_len)
2909                                 cur_id_size = id_len - 4;
2910                         cur_id_str = d + 4;
2911                         cur_id_type = d[1] & 0xf;
2912                         id_size = snprintf(id, id_len, "t10.%*pE",
2913                                            cur_id_size, cur_id_str);
2914                         break;
2915                 case 0x2:
2916                         /* EUI-64 */
2917                         if (cur_id_size > d[3])
2918                                 break;
2919                         /* Prefer NAA IEEE Registered Extended */
2920                         if (cur_id_type == 0x3 &&
2921                             cur_id_size == d[3])
2922                                 break;
2923                         cur_id_size = d[3];
2924                         cur_id_str = d + 4;
2925                         cur_id_type = d[1] & 0xf;
2926                         switch (cur_id_size) {
2927                         case 8:
2928                                 id_size = snprintf(id, id_len,
2929                                                    "eui.%8phN",
2930                                                    cur_id_str);
2931                                 break;
2932                         case 12:
2933                                 id_size = snprintf(id, id_len,
2934                                                    "eui.%12phN",
2935                                                    cur_id_str);
2936                                 break;
2937                         case 16:
2938                                 id_size = snprintf(id, id_len,
2939                                                    "eui.%16phN",
2940                                                    cur_id_str);
2941                                 break;
2942                         default:
2943                                 cur_id_size = 0;
2944                                 break;
2945                         }
2946                         break;
2947                 case 0x3:
2948                         /* NAA */
2949                         if (cur_id_size > d[3])
2950                                 break;
2951                         cur_id_size = d[3];
2952                         cur_id_str = d + 4;
2953                         cur_id_type = d[1] & 0xf;
2954                         switch (cur_id_size) {
2955                         case 8:
2956                                 id_size = snprintf(id, id_len,
2957                                                    "naa.%8phN",
2958                                                    cur_id_str);
2959                                 break;
2960                         case 16:
2961                                 id_size = snprintf(id, id_len,
2962                                                    "naa.%16phN",
2963                                                    cur_id_str);
2964                                 break;
2965                         default:
2966                                 cur_id_size = 0;
2967                                 break;
2968                         }
2969                         break;
2970                 case 0x8:
2971                         /* SCSI name string */
2972                         if (cur_id_size + 4 > d[3])
2973                                 break;
2974                         /* Prefer others for truncated descriptor */
2975                         if (cur_id_size && d[3] > id_len)
2976                                 break;
2977                         cur_id_size = id_size = d[3];
2978                         cur_id_str = d + 4;
2979                         cur_id_type = d[1] & 0xf;
2980                         if (cur_id_size >= id_len)
2981                                 cur_id_size = id_len - 1;
2982                         memcpy(id, cur_id_str, cur_id_size);
2983                         /* Decrease priority for truncated descriptor */
2984                         if (cur_id_size != id_size)
2985                                 cur_id_size = 6;
2986                         break;
2987                 default:
2988                         break;
2989                 }
2990 next_desig:
2991                 d += d[3] + 4;
2992         }
2993         rcu_read_unlock();
2994
2995         return id_size;
2996 }
2997 EXPORT_SYMBOL(scsi_vpd_lun_id);
2998
2999 /*
3000  * scsi_vpd_tpg_id - return a target port group identifier
3001  * @sdev: SCSI device
3002  *
3003  * Returns the Target Port Group identifier from the information
3004  * froom VPD page 0x83 of the device.
3005  *
3006  * Returns the identifier or error on failure.
3007  */
3008 int scsi_vpd_tpg_id(struct scsi_device *sdev, int *rel_id)
3009 {
3010         const unsigned char *d;
3011         const struct scsi_vpd *vpd_pg83;
3012         int group_id = -EAGAIN, rel_port = -1;
3013
3014         rcu_read_lock();
3015         vpd_pg83 = rcu_dereference(sdev->vpd_pg83);
3016         if (!vpd_pg83) {
3017                 rcu_read_unlock();
3018                 return -ENXIO;
3019         }
3020
3021         d = vpd_pg83->data + 4;
3022         while (d < vpd_pg83->data + vpd_pg83->len) {
3023                 switch (d[1] & 0xf) {
3024                 case 0x4:
3025                         /* Relative target port */
3026                         rel_port = get_unaligned_be16(&d[6]);
3027                         break;
3028                 case 0x5:
3029                         /* Target port group */
3030                         group_id = get_unaligned_be16(&d[6]);
3031                         break;
3032                 default:
3033                         break;
3034                 }
3035                 d += d[3] + 4;
3036         }
3037         rcu_read_unlock();
3038
3039         if (group_id >= 0 && rel_id && rel_port != -1)
3040                 *rel_id = rel_port;
3041
3042         return group_id;
3043 }
3044 EXPORT_SYMBOL(scsi_vpd_tpg_id);