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