Merge tag 'scsi-misc' of git://git.kernel.org/pub/scm/linux/kernel/git/jejb/scsi
[sfrench/cifs-2.6.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         struct scsi_device *sdev = cmd->device;
145
146         if (cmd->request->rq_flags & RQF_DONTPREP) {
147                 cmd->request->rq_flags &= ~RQF_DONTPREP;
148                 scsi_mq_uninit_cmd(cmd);
149         } else {
150                 WARN_ON_ONCE(true);
151         }
152         blk_mq_requeue_request(cmd->request, true);
153         put_device(&sdev->sdev_gendev);
154 }
155
156 /**
157  * __scsi_queue_insert - private queue insertion
158  * @cmd: The SCSI command being requeued
159  * @reason:  The reason for the requeue
160  * @unbusy: Whether the queue should be unbusied
161  *
162  * This is a private queue insertion.  The public interface
163  * scsi_queue_insert() always assumes the queue should be unbusied
164  * because it's always called before the completion.  This function is
165  * for a requeue after completion, which should only occur in this
166  * file.
167  */
168 static void __scsi_queue_insert(struct scsi_cmnd *cmd, int reason, bool unbusy)
169 {
170         struct scsi_device *device = cmd->device;
171
172         SCSI_LOG_MLQUEUE(1, scmd_printk(KERN_INFO, cmd,
173                 "Inserting command %p into mlqueue\n", cmd));
174
175         scsi_set_blocked(cmd, reason);
176
177         /*
178          * Decrement the counters, since these commands are no longer
179          * active on the host/device.
180          */
181         if (unbusy)
182                 scsi_device_unbusy(device);
183
184         /*
185          * Requeue this command.  It will go before all other commands
186          * that are already in the queue. Schedule requeue work under
187          * lock such that the kblockd_schedule_work() call happens
188          * before blk_cleanup_queue() finishes.
189          */
190         cmd->result = 0;
191
192         /*
193          * Before a SCSI command is dispatched,
194          * get_device(&sdev->sdev_gendev) is called and the host,
195          * target and device busy counters are increased. Since
196          * requeuing a request causes these actions to be repeated and
197          * since scsi_device_unbusy() has already been called,
198          * put_device(&device->sdev_gendev) must still be called. Call
199          * put_device() after blk_mq_requeue_request() to avoid that
200          * removal of the SCSI device can start before requeueing has
201          * happened.
202          */
203         blk_mq_requeue_request(cmd->request, true);
204         put_device(&device->sdev_gendev);
205 }
206
207 /*
208  * Function:    scsi_queue_insert()
209  *
210  * Purpose:     Insert a command in the midlevel queue.
211  *
212  * Arguments:   cmd    - command that we are adding to queue.
213  *              reason - why we are inserting command to queue.
214  *
215  * Lock status: Assumed that lock is not held upon entry.
216  *
217  * Returns:     Nothing.
218  *
219  * Notes:       We do this for one of two cases.  Either the host is busy
220  *              and it cannot accept any more commands for the time being,
221  *              or the device returned QUEUE_FULL and can accept no more
222  *              commands.
223  * Notes:       This could be called either from an interrupt context or a
224  *              normal process context.
225  */
226 void scsi_queue_insert(struct scsi_cmnd *cmd, int reason)
227 {
228         __scsi_queue_insert(cmd, reason, true);
229 }
230
231
232 /**
233  * __scsi_execute - insert request and wait for the result
234  * @sdev:       scsi device
235  * @cmd:        scsi command
236  * @data_direction: data direction
237  * @buffer:     data buffer
238  * @bufflen:    len of buffer
239  * @sense:      optional sense buffer
240  * @sshdr:      optional decoded sense header
241  * @timeout:    request timeout in seconds
242  * @retries:    number of times to retry request
243  * @flags:      flags for ->cmd_flags
244  * @rq_flags:   flags for ->rq_flags
245  * @resid:      optional residual length
246  *
247  * Returns the scsi_cmnd result field if a command was executed, or a negative
248  * Linux error code if we didn't get that far.
249  */
250 int __scsi_execute(struct scsi_device *sdev, const unsigned char *cmd,
251                  int data_direction, void *buffer, unsigned bufflen,
252                  unsigned char *sense, struct scsi_sense_hdr *sshdr,
253                  int timeout, int retries, u64 flags, req_flags_t rq_flags,
254                  int *resid)
255 {
256         struct request *req;
257         struct scsi_request *rq;
258         int ret = DRIVER_ERROR << 24;
259
260         req = blk_get_request(sdev->request_queue,
261                         data_direction == DMA_TO_DEVICE ?
262                         REQ_OP_SCSI_OUT : REQ_OP_SCSI_IN, BLK_MQ_REQ_PREEMPT);
263         if (IS_ERR(req))
264                 return ret;
265         rq = scsi_req(req);
266
267         if (bufflen &&  blk_rq_map_kern(sdev->request_queue, req,
268                                         buffer, bufflen, GFP_NOIO))
269                 goto out;
270
271         rq->cmd_len = COMMAND_SIZE(cmd[0]);
272         memcpy(rq->cmd, cmd, rq->cmd_len);
273         rq->retries = retries;
274         req->timeout = timeout;
275         req->cmd_flags |= flags;
276         req->rq_flags |= rq_flags | RQF_QUIET;
277
278         /*
279          * head injection *required* here otherwise quiesce won't work
280          */
281         blk_execute_rq(req->q, NULL, req, 1);
282
283         /*
284          * Some devices (USB mass-storage in particular) may transfer
285          * garbage data together with a residue indicating that the data
286          * is invalid.  Prevent the garbage from being misinterpreted
287          * and prevent security leaks by zeroing out the excess data.
288          */
289         if (unlikely(rq->resid_len > 0 && rq->resid_len <= bufflen))
290                 memset(buffer + (bufflen - rq->resid_len), 0, rq->resid_len);
291
292         if (resid)
293                 *resid = rq->resid_len;
294         if (sense && rq->sense_len)
295                 memcpy(sense, rq->sense, SCSI_SENSE_BUFFERSIZE);
296         if (sshdr)
297                 scsi_normalize_sense(rq->sense, rq->sense_len, sshdr);
298         ret = rq->result;
299  out:
300         blk_put_request(req);
301
302         return ret;
303 }
304 EXPORT_SYMBOL(__scsi_execute);
305
306 /*
307  * Function:    scsi_init_cmd_errh()
308  *
309  * Purpose:     Initialize cmd fields related to error handling.
310  *
311  * Arguments:   cmd     - command that is ready to be queued.
312  *
313  * Notes:       This function has the job of initializing a number of
314  *              fields related to error handling.   Typically this will
315  *              be called once for each command, as required.
316  */
317 static void scsi_init_cmd_errh(struct scsi_cmnd *cmd)
318 {
319         scsi_set_resid(cmd, 0);
320         memset(cmd->sense_buffer, 0, SCSI_SENSE_BUFFERSIZE);
321         if (cmd->cmd_len == 0)
322                 cmd->cmd_len = scsi_command_size(cmd->cmnd);
323 }
324
325 /*
326  * Decrement the host_busy counter and wake up the error handler if necessary.
327  * Avoid as follows that the error handler is not woken up if shost->host_busy
328  * == shost->host_failed: use call_rcu() in scsi_eh_scmd_add() in combination
329  * with an RCU read lock in this function to ensure that this function in its
330  * entirety either finishes before scsi_eh_scmd_add() increases the
331  * host_failed counter or that it notices the shost state change made by
332  * scsi_eh_scmd_add().
333  */
334 static void scsi_dec_host_busy(struct Scsi_Host *shost)
335 {
336         unsigned long flags;
337
338         rcu_read_lock();
339         atomic_dec(&shost->host_busy);
340         if (unlikely(scsi_host_in_recovery(shost))) {
341                 spin_lock_irqsave(shost->host_lock, flags);
342                 if (shost->host_failed || shost->host_eh_scheduled)
343                         scsi_eh_wakeup(shost);
344                 spin_unlock_irqrestore(shost->host_lock, flags);
345         }
346         rcu_read_unlock();
347 }
348
349 void scsi_device_unbusy(struct scsi_device *sdev)
350 {
351         struct Scsi_Host *shost = sdev->host;
352         struct scsi_target *starget = scsi_target(sdev);
353
354         scsi_dec_host_busy(shost);
355
356         if (starget->can_queue > 0)
357                 atomic_dec(&starget->target_busy);
358
359         atomic_dec(&sdev->device_busy);
360 }
361
362 static void scsi_kick_queue(struct request_queue *q)
363 {
364         blk_mq_run_hw_queues(q, false);
365 }
366
367 /*
368  * Called for single_lun devices on IO completion. Clear starget_sdev_user,
369  * and call blk_run_queue for all the scsi_devices on the target -
370  * including current_sdev first.
371  *
372  * Called with *no* scsi locks held.
373  */
374 static void scsi_single_lun_run(struct scsi_device *current_sdev)
375 {
376         struct Scsi_Host *shost = current_sdev->host;
377         struct scsi_device *sdev, *tmp;
378         struct scsi_target *starget = scsi_target(current_sdev);
379         unsigned long flags;
380
381         spin_lock_irqsave(shost->host_lock, flags);
382         starget->starget_sdev_user = NULL;
383         spin_unlock_irqrestore(shost->host_lock, flags);
384
385         /*
386          * Call blk_run_queue for all LUNs on the target, starting with
387          * current_sdev. We race with others (to set starget_sdev_user),
388          * but in most cases, we will be first. Ideally, each LU on the
389          * target would get some limited time or requests on the target.
390          */
391         scsi_kick_queue(current_sdev->request_queue);
392
393         spin_lock_irqsave(shost->host_lock, flags);
394         if (starget->starget_sdev_user)
395                 goto out;
396         list_for_each_entry_safe(sdev, tmp, &starget->devices,
397                         same_target_siblings) {
398                 if (sdev == current_sdev)
399                         continue;
400                 if (scsi_device_get(sdev))
401                         continue;
402
403                 spin_unlock_irqrestore(shost->host_lock, flags);
404                 scsi_kick_queue(sdev->request_queue);
405                 spin_lock_irqsave(shost->host_lock, flags);
406         
407                 scsi_device_put(sdev);
408         }
409  out:
410         spin_unlock_irqrestore(shost->host_lock, flags);
411 }
412
413 static inline bool scsi_device_is_busy(struct scsi_device *sdev)
414 {
415         if (atomic_read(&sdev->device_busy) >= sdev->queue_depth)
416                 return true;
417         if (atomic_read(&sdev->device_blocked) > 0)
418                 return true;
419         return false;
420 }
421
422 static inline bool scsi_target_is_busy(struct scsi_target *starget)
423 {
424         if (starget->can_queue > 0) {
425                 if (atomic_read(&starget->target_busy) >= starget->can_queue)
426                         return true;
427                 if (atomic_read(&starget->target_blocked) > 0)
428                         return true;
429         }
430         return false;
431 }
432
433 static inline bool scsi_host_is_busy(struct Scsi_Host *shost)
434 {
435         if (shost->can_queue > 0 &&
436             atomic_read(&shost->host_busy) >= shost->can_queue)
437                 return true;
438         if (atomic_read(&shost->host_blocked) > 0)
439                 return true;
440         if (shost->host_self_blocked)
441                 return true;
442         return false;
443 }
444
445 static void scsi_starved_list_run(struct Scsi_Host *shost)
446 {
447         LIST_HEAD(starved_list);
448         struct scsi_device *sdev;
449         unsigned long flags;
450
451         spin_lock_irqsave(shost->host_lock, flags);
452         list_splice_init(&shost->starved_list, &starved_list);
453
454         while (!list_empty(&starved_list)) {
455                 struct request_queue *slq;
456
457                 /*
458                  * As long as shost is accepting commands and we have
459                  * starved queues, call blk_run_queue. scsi_request_fn
460                  * drops the queue_lock and can add us back to the
461                  * starved_list.
462                  *
463                  * host_lock protects the starved_list and starved_entry.
464                  * scsi_request_fn must get the host_lock before checking
465                  * or modifying starved_list or starved_entry.
466                  */
467                 if (scsi_host_is_busy(shost))
468                         break;
469
470                 sdev = list_entry(starved_list.next,
471                                   struct scsi_device, starved_entry);
472                 list_del_init(&sdev->starved_entry);
473                 if (scsi_target_is_busy(scsi_target(sdev))) {
474                         list_move_tail(&sdev->starved_entry,
475                                        &shost->starved_list);
476                         continue;
477                 }
478
479                 /*
480                  * Once we drop the host lock, a racing scsi_remove_device()
481                  * call may remove the sdev from the starved list and destroy
482                  * it and the queue.  Mitigate by taking a reference to the
483                  * queue and never touching the sdev again after we drop the
484                  * host lock.  Note: if __scsi_remove_device() invokes
485                  * blk_cleanup_queue() before the queue is run from this
486                  * function then blk_run_queue() will return immediately since
487                  * blk_cleanup_queue() marks the queue with QUEUE_FLAG_DYING.
488                  */
489                 slq = sdev->request_queue;
490                 if (!blk_get_queue(slq))
491                         continue;
492                 spin_unlock_irqrestore(shost->host_lock, flags);
493
494                 scsi_kick_queue(slq);
495                 blk_put_queue(slq);
496
497                 spin_lock_irqsave(shost->host_lock, flags);
498         }
499         /* put any unprocessed entries back */
500         list_splice(&starved_list, &shost->starved_list);
501         spin_unlock_irqrestore(shost->host_lock, flags);
502 }
503
504 /*
505  * Function:   scsi_run_queue()
506  *
507  * Purpose:    Select a proper request queue to serve next
508  *
509  * Arguments:  q       - last request's queue
510  *
511  * Returns:     Nothing
512  *
513  * Notes:      The previous command was completely finished, start
514  *             a new one if possible.
515  */
516 static void scsi_run_queue(struct request_queue *q)
517 {
518         struct scsi_device *sdev = q->queuedata;
519
520         if (scsi_target(sdev)->single_lun)
521                 scsi_single_lun_run(sdev);
522         if (!list_empty(&sdev->host->starved_list))
523                 scsi_starved_list_run(sdev->host);
524
525         blk_mq_run_hw_queues(q, false);
526 }
527
528 void scsi_requeue_run_queue(struct work_struct *work)
529 {
530         struct scsi_device *sdev;
531         struct request_queue *q;
532
533         sdev = container_of(work, struct scsi_device, requeue_work);
534         q = sdev->request_queue;
535         scsi_run_queue(q);
536 }
537
538 void scsi_run_host_queues(struct Scsi_Host *shost)
539 {
540         struct scsi_device *sdev;
541
542         shost_for_each_device(sdev, shost)
543                 scsi_run_queue(sdev->request_queue);
544 }
545
546 static void scsi_uninit_cmd(struct scsi_cmnd *cmd)
547 {
548         if (!blk_rq_is_passthrough(cmd->request)) {
549                 struct scsi_driver *drv = scsi_cmd_to_driver(cmd);
550
551                 if (drv->uninit_command)
552                         drv->uninit_command(cmd);
553         }
554 }
555
556 static void scsi_mq_free_sgtables(struct scsi_cmnd *cmd)
557 {
558         if (cmd->sdb.table.nents)
559                 sg_free_table_chained(&cmd->sdb.table, true);
560         if (scsi_prot_sg_count(cmd))
561                 sg_free_table_chained(&cmd->prot_sdb->table, true);
562 }
563
564 static void scsi_mq_uninit_cmd(struct scsi_cmnd *cmd)
565 {
566         scsi_mq_free_sgtables(cmd);
567         scsi_uninit_cmd(cmd);
568         scsi_del_cmd_from_list(cmd);
569 }
570
571 /* Returns false when no more bytes to process, true if there are more */
572 static bool scsi_end_request(struct request *req, blk_status_t error,
573                 unsigned int bytes)
574 {
575         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
576         struct scsi_device *sdev = cmd->device;
577         struct request_queue *q = sdev->request_queue;
578
579         if (blk_update_request(req, error, bytes))
580                 return true;
581
582         if (blk_queue_add_random(q))
583                 add_disk_randomness(req->rq_disk);
584
585         if (!blk_rq_is_scsi(req)) {
586                 WARN_ON_ONCE(!(cmd->flags & SCMD_INITIALIZED));
587                 cmd->flags &= ~SCMD_INITIALIZED;
588                 destroy_rcu_head(&cmd->rcu);
589         }
590
591         /*
592          * In the MQ case the command gets freed by __blk_mq_end_request,
593          * so we have to do all cleanup that depends on it earlier.
594          *
595          * We also can't kick the queues from irq context, so we
596          * will have to defer it to a workqueue.
597          */
598         scsi_mq_uninit_cmd(cmd);
599
600         /*
601          * queue is still alive, so grab the ref for preventing it
602          * from being cleaned up during running queue.
603          */
604         percpu_ref_get(&q->q_usage_counter);
605
606         __blk_mq_end_request(req, error);
607
608         if (scsi_target(sdev)->single_lun ||
609             !list_empty(&sdev->host->starved_list))
610                 kblockd_schedule_work(&sdev->requeue_work);
611         else
612                 blk_mq_run_hw_queues(q, true);
613
614         percpu_ref_put(&q->q_usage_counter);
615         put_device(&sdev->sdev_gendev);
616         return false;
617 }
618
619 /**
620  * scsi_result_to_blk_status - translate a SCSI result code into blk_status_t
621  * @cmd:        SCSI command
622  * @result:     scsi error code
623  *
624  * Translate a SCSI result code into a blk_status_t value. May reset the host
625  * byte of @cmd->result.
626  */
627 static blk_status_t scsi_result_to_blk_status(struct scsi_cmnd *cmd, int result)
628 {
629         switch (host_byte(result)) {
630         case DID_OK:
631                 /*
632                  * Also check the other bytes than the status byte in result
633                  * to handle the case when a SCSI LLD sets result to
634                  * DRIVER_SENSE << 24 without setting SAM_STAT_CHECK_CONDITION.
635                  */
636                 if (scsi_status_is_good(result) && (result & ~0xff) == 0)
637                         return BLK_STS_OK;
638                 return BLK_STS_IOERR;
639         case DID_TRANSPORT_FAILFAST:
640                 return BLK_STS_TRANSPORT;
641         case DID_TARGET_FAILURE:
642                 set_host_byte(cmd, DID_OK);
643                 return BLK_STS_TARGET;
644         case DID_NEXUS_FAILURE:
645                 set_host_byte(cmd, DID_OK);
646                 return BLK_STS_NEXUS;
647         case DID_ALLOC_FAILURE:
648                 set_host_byte(cmd, DID_OK);
649                 return BLK_STS_NOSPC;
650         case DID_MEDIUM_ERROR:
651                 set_host_byte(cmd, DID_OK);
652                 return BLK_STS_MEDIUM;
653         default:
654                 return BLK_STS_IOERR;
655         }
656 }
657
658 /* Helper for scsi_io_completion() when "reprep" action required. */
659 static void scsi_io_completion_reprep(struct scsi_cmnd *cmd,
660                                       struct request_queue *q)
661 {
662         /* A new command will be prepared and issued. */
663         scsi_mq_requeue_cmd(cmd);
664 }
665
666 /* Helper for scsi_io_completion() when special action required. */
667 static void scsi_io_completion_action(struct scsi_cmnd *cmd, int result)
668 {
669         struct request_queue *q = cmd->device->request_queue;
670         struct request *req = cmd->request;
671         int level = 0;
672         enum {ACTION_FAIL, ACTION_REPREP, ACTION_RETRY,
673               ACTION_DELAYED_RETRY} action;
674         unsigned long wait_for = (cmd->allowed + 1) * req->timeout;
675         struct scsi_sense_hdr sshdr;
676         bool sense_valid;
677         bool sense_current = true;      /* false implies "deferred sense" */
678         blk_status_t blk_stat;
679
680         sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
681         if (sense_valid)
682                 sense_current = !scsi_sense_is_deferred(&sshdr);
683
684         blk_stat = scsi_result_to_blk_status(cmd, result);
685
686         if (host_byte(result) == DID_RESET) {
687                 /* Third party bus reset or reset for error recovery
688                  * reasons.  Just retry the command and see what
689                  * happens.
690                  */
691                 action = ACTION_RETRY;
692         } else if (sense_valid && sense_current) {
693                 switch (sshdr.sense_key) {
694                 case UNIT_ATTENTION:
695                         if (cmd->device->removable) {
696                                 /* Detected disc change.  Set a bit
697                                  * and quietly refuse further access.
698                                  */
699                                 cmd->device->changed = 1;
700                                 action = ACTION_FAIL;
701                         } else {
702                                 /* Must have been a power glitch, or a
703                                  * bus reset.  Could not have been a
704                                  * media change, so we just retry the
705                                  * command and see what happens.
706                                  */
707                                 action = ACTION_RETRY;
708                         }
709                         break;
710                 case ILLEGAL_REQUEST:
711                         /* If we had an ILLEGAL REQUEST returned, then
712                          * we may have performed an unsupported
713                          * command.  The only thing this should be
714                          * would be a ten byte read where only a six
715                          * byte read was supported.  Also, on a system
716                          * where READ CAPACITY failed, we may have
717                          * read past the end of the disk.
718                          */
719                         if ((cmd->device->use_10_for_rw &&
720                             sshdr.asc == 0x20 && sshdr.ascq == 0x00) &&
721                             (cmd->cmnd[0] == READ_10 ||
722                              cmd->cmnd[0] == WRITE_10)) {
723                                 /* This will issue a new 6-byte command. */
724                                 cmd->device->use_10_for_rw = 0;
725                                 action = ACTION_REPREP;
726                         } else if (sshdr.asc == 0x10) /* DIX */ {
727                                 action = ACTION_FAIL;
728                                 blk_stat = BLK_STS_PROTECTION;
729                         /* INVALID COMMAND OPCODE or INVALID FIELD IN CDB */
730                         } else if (sshdr.asc == 0x20 || sshdr.asc == 0x24) {
731                                 action = ACTION_FAIL;
732                                 blk_stat = BLK_STS_TARGET;
733                         } else
734                                 action = ACTION_FAIL;
735                         break;
736                 case ABORTED_COMMAND:
737                         action = ACTION_FAIL;
738                         if (sshdr.asc == 0x10) /* DIF */
739                                 blk_stat = BLK_STS_PROTECTION;
740                         break;
741                 case NOT_READY:
742                         /* If the device is in the process of becoming
743                          * ready, or has a temporary blockage, retry.
744                          */
745                         if (sshdr.asc == 0x04) {
746                                 switch (sshdr.ascq) {
747                                 case 0x01: /* becoming ready */
748                                 case 0x04: /* format in progress */
749                                 case 0x05: /* rebuild in progress */
750                                 case 0x06: /* recalculation in progress */
751                                 case 0x07: /* operation in progress */
752                                 case 0x08: /* Long write in progress */
753                                 case 0x09: /* self test in progress */
754                                 case 0x14: /* space allocation in progress */
755                                 case 0x1a: /* start stop unit in progress */
756                                 case 0x1b: /* sanitize in progress */
757                                 case 0x1d: /* configuration in progress */
758                                 case 0x24: /* depopulation in progress */
759                                         action = ACTION_DELAYED_RETRY;
760                                         break;
761                                 default:
762                                         action = ACTION_FAIL;
763                                         break;
764                                 }
765                         } else
766                                 action = ACTION_FAIL;
767                         break;
768                 case VOLUME_OVERFLOW:
769                         /* See SSC3rXX or current. */
770                         action = ACTION_FAIL;
771                         break;
772                 default:
773                         action = ACTION_FAIL;
774                         break;
775                 }
776         } else
777                 action = ACTION_FAIL;
778
779         if (action != ACTION_FAIL &&
780             time_before(cmd->jiffies_at_alloc + wait_for, jiffies))
781                 action = ACTION_FAIL;
782
783         switch (action) {
784         case ACTION_FAIL:
785                 /* Give up and fail the remainder of the request */
786                 if (!(req->rq_flags & RQF_QUIET)) {
787                         static DEFINE_RATELIMIT_STATE(_rs,
788                                         DEFAULT_RATELIMIT_INTERVAL,
789                                         DEFAULT_RATELIMIT_BURST);
790
791                         if (unlikely(scsi_logging_level))
792                                 level =
793                                      SCSI_LOG_LEVEL(SCSI_LOG_MLCOMPLETE_SHIFT,
794                                                     SCSI_LOG_MLCOMPLETE_BITS);
795
796                         /*
797                          * if logging is enabled the failure will be printed
798                          * in scsi_log_completion(), so avoid duplicate messages
799                          */
800                         if (!level && __ratelimit(&_rs)) {
801                                 scsi_print_result(cmd, NULL, FAILED);
802                                 if (driver_byte(result) == DRIVER_SENSE)
803                                         scsi_print_sense(cmd);
804                                 scsi_print_command(cmd);
805                         }
806                 }
807                 if (!scsi_end_request(req, blk_stat, blk_rq_err_bytes(req)))
808                         return;
809                 /*FALLTHRU*/
810         case ACTION_REPREP:
811                 scsi_io_completion_reprep(cmd, q);
812                 break;
813         case ACTION_RETRY:
814                 /* Retry the same command immediately */
815                 __scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY, false);
816                 break;
817         case ACTION_DELAYED_RETRY:
818                 /* Retry the same command after a delay */
819                 __scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY, false);
820                 break;
821         }
822 }
823
824 /*
825  * Helper for scsi_io_completion() when cmd->result is non-zero. Returns a
826  * new result that may suppress further error checking. Also modifies
827  * *blk_statp in some cases.
828  */
829 static int scsi_io_completion_nz_result(struct scsi_cmnd *cmd, int result,
830                                         blk_status_t *blk_statp)
831 {
832         bool sense_valid;
833         bool sense_current = true;      /* false implies "deferred sense" */
834         struct request *req = cmd->request;
835         struct scsi_sense_hdr sshdr;
836
837         sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
838         if (sense_valid)
839                 sense_current = !scsi_sense_is_deferred(&sshdr);
840
841         if (blk_rq_is_passthrough(req)) {
842                 if (sense_valid) {
843                         /*
844                          * SG_IO wants current and deferred errors
845                          */
846                         scsi_req(req)->sense_len =
847                                 min(8 + cmd->sense_buffer[7],
848                                     SCSI_SENSE_BUFFERSIZE);
849                 }
850                 if (sense_current)
851                         *blk_statp = scsi_result_to_blk_status(cmd, result);
852         } else if (blk_rq_bytes(req) == 0 && sense_current) {
853                 /*
854                  * Flush commands do not transfers any data, and thus cannot use
855                  * good_bytes != blk_rq_bytes(req) as the signal for an error.
856                  * This sets *blk_statp explicitly for the problem case.
857                  */
858                 *blk_statp = scsi_result_to_blk_status(cmd, result);
859         }
860         /*
861          * Recovered errors need reporting, but they're always treated as
862          * success, so fiddle the result code here.  For passthrough requests
863          * we already took a copy of the original into sreq->result which
864          * is what gets returned to the user
865          */
866         if (sense_valid && (sshdr.sense_key == RECOVERED_ERROR)) {
867                 bool do_print = true;
868                 /*
869                  * if ATA PASS-THROUGH INFORMATION AVAILABLE [0x0, 0x1d]
870                  * skip print since caller wants ATA registers. Only occurs
871                  * on SCSI ATA PASS_THROUGH commands when CK_COND=1
872                  */
873                 if ((sshdr.asc == 0x0) && (sshdr.ascq == 0x1d))
874                         do_print = false;
875                 else if (req->rq_flags & RQF_QUIET)
876                         do_print = false;
877                 if (do_print)
878                         scsi_print_sense(cmd);
879                 result = 0;
880                 /* for passthrough, *blk_statp may be set */
881                 *blk_statp = BLK_STS_OK;
882         }
883         /*
884          * Another corner case: the SCSI status byte is non-zero but 'good'.
885          * Example: PRE-FETCH command returns SAM_STAT_CONDITION_MET when
886          * it is able to fit nominated LBs in its cache (and SAM_STAT_GOOD
887          * if it can't fit). Treat SAM_STAT_CONDITION_MET and the related
888          * intermediate statuses (both obsolete in SAM-4) as good.
889          */
890         if (status_byte(result) && scsi_status_is_good(result)) {
891                 result = 0;
892                 *blk_statp = BLK_STS_OK;
893         }
894         return result;
895 }
896
897 /*
898  * Function:    scsi_io_completion()
899  *
900  * Purpose:     Completion processing for block device I/O requests.
901  *
902  * Arguments:   cmd   - command that is finished.
903  *
904  * Lock status: Assumed that no lock is held upon entry.
905  *
906  * Returns:     Nothing
907  *
908  * Notes:       We will finish off the specified number of sectors.  If we
909  *              are done, the command block will be released and the queue
910  *              function will be goosed.  If we are not done then we have to
911  *              figure out what to do next:
912  *
913  *              a) We can call scsi_requeue_command().  The request
914  *                 will be unprepared and put back on the queue.  Then
915  *                 a new command will be created for it.  This should
916  *                 be used if we made forward progress, or if we want
917  *                 to switch from READ(10) to READ(6) for example.
918  *
919  *              b) We can call __scsi_queue_insert().  The request will
920  *                 be put back on the queue and retried using the same
921  *                 command as before, possibly after a delay.
922  *
923  *              c) We can call scsi_end_request() with blk_stat other than
924  *                 BLK_STS_OK, to fail the remainder of the request.
925  */
926 void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes)
927 {
928         int result = cmd->result;
929         struct request_queue *q = cmd->device->request_queue;
930         struct request *req = cmd->request;
931         blk_status_t blk_stat = BLK_STS_OK;
932
933         if (unlikely(result))   /* a nz result may or may not be an error */
934                 result = scsi_io_completion_nz_result(cmd, result, &blk_stat);
935
936         if (unlikely(blk_rq_is_passthrough(req))) {
937                 /*
938                  * scsi_result_to_blk_status may have reset the host_byte
939                  */
940                 scsi_req(req)->result = cmd->result;
941         }
942
943         /*
944          * Next deal with any sectors which we were able to correctly
945          * handle.
946          */
947         SCSI_LOG_HLCOMPLETE(1, scmd_printk(KERN_INFO, cmd,
948                 "%u sectors total, %d bytes done.\n",
949                 blk_rq_sectors(req), good_bytes));
950
951         /*
952          * Next deal with any sectors which we were able to correctly
953          * handle. Failed, zero length commands always need to drop down
954          * to retry code. Fast path should return in this block.
955          */
956         if (likely(blk_rq_bytes(req) > 0 || blk_stat == BLK_STS_OK)) {
957                 if (likely(!scsi_end_request(req, blk_stat, good_bytes)))
958                         return; /* no bytes remaining */
959         }
960
961         /* Kill remainder if no retries. */
962         if (unlikely(blk_stat && scsi_noretry_cmd(cmd))) {
963                 if (scsi_end_request(req, blk_stat, blk_rq_bytes(req)))
964                         WARN_ONCE(true,
965                             "Bytes remaining after failed, no-retry command");
966                 return;
967         }
968
969         /*
970          * If there had been no error, but we have leftover bytes in the
971          * requeues just queue the command up again.
972          */
973         if (likely(result == 0))
974                 scsi_io_completion_reprep(cmd, q);
975         else
976                 scsi_io_completion_action(cmd, result);
977 }
978
979 static blk_status_t scsi_init_sgtable(struct request *req,
980                 struct scsi_data_buffer *sdb)
981 {
982         int count;
983
984         /*
985          * If sg table allocation fails, requeue request later.
986          */
987         if (unlikely(sg_alloc_table_chained(&sdb->table,
988                         blk_rq_nr_phys_segments(req), sdb->table.sgl)))
989                 return BLK_STS_RESOURCE;
990
991         /* 
992          * Next, walk the list, and fill in the addresses and sizes of
993          * each segment.
994          */
995         count = blk_rq_map_sg(req->q, req, sdb->table.sgl);
996         BUG_ON(count > sdb->table.nents);
997         sdb->table.nents = count;
998         sdb->length = blk_rq_payload_bytes(req);
999         return BLK_STS_OK;
1000 }
1001
1002 /*
1003  * Function:    scsi_init_io()
1004  *
1005  * Purpose:     SCSI I/O initialize function.
1006  *
1007  * Arguments:   cmd   - Command descriptor we wish to initialize
1008  *
1009  * Returns:     BLK_STS_OK on success
1010  *              BLK_STS_RESOURCE if the failure is retryable
1011  *              BLK_STS_IOERR if the failure is fatal
1012  */
1013 blk_status_t scsi_init_io(struct scsi_cmnd *cmd)
1014 {
1015         struct request *rq = cmd->request;
1016         blk_status_t ret;
1017
1018         if (WARN_ON_ONCE(!blk_rq_nr_phys_segments(rq)))
1019                 return BLK_STS_IOERR;
1020
1021         ret = scsi_init_sgtable(rq, &cmd->sdb);
1022         if (ret)
1023                 return ret;
1024
1025         if (blk_integrity_rq(rq)) {
1026                 struct scsi_data_buffer *prot_sdb = cmd->prot_sdb;
1027                 int ivecs, count;
1028
1029                 if (WARN_ON_ONCE(!prot_sdb)) {
1030                         /*
1031                          * This can happen if someone (e.g. multipath)
1032                          * queues a command to a device on an adapter
1033                          * that does not support DIX.
1034                          */
1035                         ret = BLK_STS_IOERR;
1036                         goto out_free_sgtables;
1037                 }
1038
1039                 ivecs = blk_rq_count_integrity_sg(rq->q, rq->bio);
1040
1041                 if (sg_alloc_table_chained(&prot_sdb->table, ivecs,
1042                                 prot_sdb->table.sgl)) {
1043                         ret = BLK_STS_RESOURCE;
1044                         goto out_free_sgtables;
1045                 }
1046
1047                 count = blk_rq_map_integrity_sg(rq->q, rq->bio,
1048                                                 prot_sdb->table.sgl);
1049                 BUG_ON(count > ivecs);
1050                 BUG_ON(count > queue_max_integrity_segments(rq->q));
1051
1052                 cmd->prot_sdb = prot_sdb;
1053                 cmd->prot_sdb->table.nents = count;
1054         }
1055
1056         return BLK_STS_OK;
1057 out_free_sgtables:
1058         scsi_mq_free_sgtables(cmd);
1059         return ret;
1060 }
1061 EXPORT_SYMBOL(scsi_init_io);
1062
1063 /**
1064  * scsi_initialize_rq - initialize struct scsi_cmnd partially
1065  * @rq: Request associated with the SCSI command to be initialized.
1066  *
1067  * This function initializes the members of struct scsi_cmnd that must be
1068  * initialized before request processing starts and that won't be
1069  * reinitialized if a SCSI command is requeued.
1070  *
1071  * Called from inside blk_get_request() for pass-through requests and from
1072  * inside scsi_init_command() for filesystem requests.
1073  */
1074 static void scsi_initialize_rq(struct request *rq)
1075 {
1076         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1077
1078         scsi_req_init(&cmd->req);
1079         init_rcu_head(&cmd->rcu);
1080         cmd->jiffies_at_alloc = jiffies;
1081         cmd->retries = 0;
1082 }
1083
1084 /* Add a command to the list used by the aacraid and dpt_i2o drivers */
1085 void scsi_add_cmd_to_list(struct scsi_cmnd *cmd)
1086 {
1087         struct scsi_device *sdev = cmd->device;
1088         struct Scsi_Host *shost = sdev->host;
1089         unsigned long flags;
1090
1091         if (shost->use_cmd_list) {
1092                 spin_lock_irqsave(&sdev->list_lock, flags);
1093                 list_add_tail(&cmd->list, &sdev->cmd_list);
1094                 spin_unlock_irqrestore(&sdev->list_lock, flags);
1095         }
1096 }
1097
1098 /* Remove a command from the list used by the aacraid and dpt_i2o drivers */
1099 void scsi_del_cmd_from_list(struct scsi_cmnd *cmd)
1100 {
1101         struct scsi_device *sdev = cmd->device;
1102         struct Scsi_Host *shost = sdev->host;
1103         unsigned long flags;
1104
1105         if (shost->use_cmd_list) {
1106                 spin_lock_irqsave(&sdev->list_lock, flags);
1107                 BUG_ON(list_empty(&cmd->list));
1108                 list_del_init(&cmd->list);
1109                 spin_unlock_irqrestore(&sdev->list_lock, flags);
1110         }
1111 }
1112
1113 /* Called after a request has been started. */
1114 void scsi_init_command(struct scsi_device *dev, struct scsi_cmnd *cmd)
1115 {
1116         void *buf = cmd->sense_buffer;
1117         void *prot = cmd->prot_sdb;
1118         struct request *rq = blk_mq_rq_from_pdu(cmd);
1119         unsigned int flags = cmd->flags & SCMD_PRESERVED_FLAGS;
1120         unsigned long jiffies_at_alloc;
1121         int retries;
1122
1123         if (!blk_rq_is_scsi(rq) && !(flags & SCMD_INITIALIZED)) {
1124                 flags |= SCMD_INITIALIZED;
1125                 scsi_initialize_rq(rq);
1126         }
1127
1128         jiffies_at_alloc = cmd->jiffies_at_alloc;
1129         retries = cmd->retries;
1130         /* zero out the cmd, except for the embedded scsi_request */
1131         memset((char *)cmd + sizeof(cmd->req), 0,
1132                 sizeof(*cmd) - sizeof(cmd->req) + dev->host->hostt->cmd_size);
1133
1134         cmd->device = dev;
1135         cmd->sense_buffer = buf;
1136         cmd->prot_sdb = prot;
1137         cmd->flags = flags;
1138         INIT_DELAYED_WORK(&cmd->abort_work, scmd_eh_abort_handler);
1139         cmd->jiffies_at_alloc = jiffies_at_alloc;
1140         cmd->retries = retries;
1141
1142         scsi_add_cmd_to_list(cmd);
1143 }
1144
1145 static blk_status_t scsi_setup_scsi_cmnd(struct scsi_device *sdev,
1146                 struct request *req)
1147 {
1148         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1149
1150         /*
1151          * Passthrough requests may transfer data, in which case they must
1152          * a bio attached to them.  Or they might contain a SCSI command
1153          * that does not transfer data, in which case they may optionally
1154          * submit a request without an attached bio.
1155          */
1156         if (req->bio) {
1157                 blk_status_t ret = scsi_init_io(cmd);
1158                 if (unlikely(ret != BLK_STS_OK))
1159                         return ret;
1160         } else {
1161                 BUG_ON(blk_rq_bytes(req));
1162
1163                 memset(&cmd->sdb, 0, sizeof(cmd->sdb));
1164         }
1165
1166         cmd->cmd_len = scsi_req(req)->cmd_len;
1167         cmd->cmnd = scsi_req(req)->cmd;
1168         cmd->transfersize = blk_rq_bytes(req);
1169         cmd->allowed = scsi_req(req)->retries;
1170         return BLK_STS_OK;
1171 }
1172
1173 /*
1174  * Setup a normal block command.  These are simple request from filesystems
1175  * that still need to be translated to SCSI CDBs from the ULD.
1176  */
1177 static blk_status_t scsi_setup_fs_cmnd(struct scsi_device *sdev,
1178                 struct request *req)
1179 {
1180         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1181
1182         if (unlikely(sdev->handler && sdev->handler->prep_fn)) {
1183                 blk_status_t ret = sdev->handler->prep_fn(sdev, req);
1184                 if (ret != BLK_STS_OK)
1185                         return ret;
1186         }
1187
1188         cmd->cmnd = scsi_req(req)->cmd = scsi_req(req)->__cmd;
1189         memset(cmd->cmnd, 0, BLK_MAX_CDB);
1190         return scsi_cmd_to_driver(cmd)->init_command(cmd);
1191 }
1192
1193 static blk_status_t scsi_setup_cmnd(struct scsi_device *sdev,
1194                 struct request *req)
1195 {
1196         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1197
1198         if (!blk_rq_bytes(req))
1199                 cmd->sc_data_direction = DMA_NONE;
1200         else if (rq_data_dir(req) == WRITE)
1201                 cmd->sc_data_direction = DMA_TO_DEVICE;
1202         else
1203                 cmd->sc_data_direction = DMA_FROM_DEVICE;
1204
1205         if (blk_rq_is_scsi(req))
1206                 return scsi_setup_scsi_cmnd(sdev, req);
1207         else
1208                 return scsi_setup_fs_cmnd(sdev, req);
1209 }
1210
1211 static blk_status_t
1212 scsi_prep_state_check(struct scsi_device *sdev, struct request *req)
1213 {
1214         switch (sdev->sdev_state) {
1215         case SDEV_OFFLINE:
1216         case SDEV_TRANSPORT_OFFLINE:
1217                 /*
1218                  * If the device is offline we refuse to process any
1219                  * commands.  The device must be brought online
1220                  * before trying any recovery commands.
1221                  */
1222                 sdev_printk(KERN_ERR, sdev,
1223                             "rejecting I/O to offline device\n");
1224                 return BLK_STS_IOERR;
1225         case SDEV_DEL:
1226                 /*
1227                  * If the device is fully deleted, we refuse to
1228                  * process any commands as well.
1229                  */
1230                 sdev_printk(KERN_ERR, sdev,
1231                             "rejecting I/O to dead device\n");
1232                 return BLK_STS_IOERR;
1233         case SDEV_BLOCK:
1234         case SDEV_CREATED_BLOCK:
1235                 return BLK_STS_RESOURCE;
1236         case SDEV_QUIESCE:
1237                 /*
1238                  * If the devices is blocked we defer normal commands.
1239                  */
1240                 if (req && !(req->rq_flags & RQF_PREEMPT))
1241                         return BLK_STS_RESOURCE;
1242                 return BLK_STS_OK;
1243         default:
1244                 /*
1245                  * For any other not fully online state we only allow
1246                  * special commands.  In particular any user initiated
1247                  * command is not allowed.
1248                  */
1249                 if (req && !(req->rq_flags & RQF_PREEMPT))
1250                         return BLK_STS_IOERR;
1251                 return BLK_STS_OK;
1252         }
1253 }
1254
1255 /*
1256  * scsi_dev_queue_ready: if we can send requests to sdev, return 1 else
1257  * return 0.
1258  *
1259  * Called with the queue_lock held.
1260  */
1261 static inline int scsi_dev_queue_ready(struct request_queue *q,
1262                                   struct scsi_device *sdev)
1263 {
1264         unsigned int busy;
1265
1266         busy = atomic_inc_return(&sdev->device_busy) - 1;
1267         if (atomic_read(&sdev->device_blocked)) {
1268                 if (busy)
1269                         goto out_dec;
1270
1271                 /*
1272                  * unblock after device_blocked iterates to zero
1273                  */
1274                 if (atomic_dec_return(&sdev->device_blocked) > 0)
1275                         goto out_dec;
1276                 SCSI_LOG_MLQUEUE(3, sdev_printk(KERN_INFO, sdev,
1277                                    "unblocking device at zero depth\n"));
1278         }
1279
1280         if (busy >= sdev->queue_depth)
1281                 goto out_dec;
1282
1283         return 1;
1284 out_dec:
1285         atomic_dec(&sdev->device_busy);
1286         return 0;
1287 }
1288
1289 /*
1290  * scsi_target_queue_ready: checks if there we can send commands to target
1291  * @sdev: scsi device on starget to check.
1292  */
1293 static inline int scsi_target_queue_ready(struct Scsi_Host *shost,
1294                                            struct scsi_device *sdev)
1295 {
1296         struct scsi_target *starget = scsi_target(sdev);
1297         unsigned int busy;
1298
1299         if (starget->single_lun) {
1300                 spin_lock_irq(shost->host_lock);
1301                 if (starget->starget_sdev_user &&
1302                     starget->starget_sdev_user != sdev) {
1303                         spin_unlock_irq(shost->host_lock);
1304                         return 0;
1305                 }
1306                 starget->starget_sdev_user = sdev;
1307                 spin_unlock_irq(shost->host_lock);
1308         }
1309
1310         if (starget->can_queue <= 0)
1311                 return 1;
1312
1313         busy = atomic_inc_return(&starget->target_busy) - 1;
1314         if (atomic_read(&starget->target_blocked) > 0) {
1315                 if (busy)
1316                         goto starved;
1317
1318                 /*
1319                  * unblock after target_blocked iterates to zero
1320                  */
1321                 if (atomic_dec_return(&starget->target_blocked) > 0)
1322                         goto out_dec;
1323
1324                 SCSI_LOG_MLQUEUE(3, starget_printk(KERN_INFO, starget,
1325                                  "unblocking target at zero depth\n"));
1326         }
1327
1328         if (busy >= starget->can_queue)
1329                 goto starved;
1330
1331         return 1;
1332
1333 starved:
1334         spin_lock_irq(shost->host_lock);
1335         list_move_tail(&sdev->starved_entry, &shost->starved_list);
1336         spin_unlock_irq(shost->host_lock);
1337 out_dec:
1338         if (starget->can_queue > 0)
1339                 atomic_dec(&starget->target_busy);
1340         return 0;
1341 }
1342
1343 /*
1344  * scsi_host_queue_ready: if we can send requests to shost, return 1 else
1345  * return 0. We must end up running the queue again whenever 0 is
1346  * returned, else IO can hang.
1347  */
1348 static inline int scsi_host_queue_ready(struct request_queue *q,
1349                                    struct Scsi_Host *shost,
1350                                    struct scsi_device *sdev)
1351 {
1352         unsigned int busy;
1353
1354         if (scsi_host_in_recovery(shost))
1355                 return 0;
1356
1357         busy = atomic_inc_return(&shost->host_busy) - 1;
1358         if (atomic_read(&shost->host_blocked) > 0) {
1359                 if (busy)
1360                         goto starved;
1361
1362                 /*
1363                  * unblock after host_blocked iterates to zero
1364                  */
1365                 if (atomic_dec_return(&shost->host_blocked) > 0)
1366                         goto out_dec;
1367
1368                 SCSI_LOG_MLQUEUE(3,
1369                         shost_printk(KERN_INFO, shost,
1370                                      "unblocking host at zero depth\n"));
1371         }
1372
1373         if (shost->can_queue > 0 && busy >= shost->can_queue)
1374                 goto starved;
1375         if (shost->host_self_blocked)
1376                 goto starved;
1377
1378         /* We're OK to process the command, so we can't be starved */
1379         if (!list_empty(&sdev->starved_entry)) {
1380                 spin_lock_irq(shost->host_lock);
1381                 if (!list_empty(&sdev->starved_entry))
1382                         list_del_init(&sdev->starved_entry);
1383                 spin_unlock_irq(shost->host_lock);
1384         }
1385
1386         return 1;
1387
1388 starved:
1389         spin_lock_irq(shost->host_lock);
1390         if (list_empty(&sdev->starved_entry))
1391                 list_add_tail(&sdev->starved_entry, &shost->starved_list);
1392         spin_unlock_irq(shost->host_lock);
1393 out_dec:
1394         scsi_dec_host_busy(shost);
1395         return 0;
1396 }
1397
1398 /*
1399  * Busy state exporting function for request stacking drivers.
1400  *
1401  * For efficiency, no lock is taken to check the busy state of
1402  * shost/starget/sdev, since the returned value is not guaranteed and
1403  * may be changed after request stacking drivers call the function,
1404  * regardless of taking lock or not.
1405  *
1406  * When scsi can't dispatch I/Os anymore and needs to kill I/Os scsi
1407  * needs to return 'not busy'. Otherwise, request stacking drivers
1408  * may hold requests forever.
1409  */
1410 static bool scsi_mq_lld_busy(struct request_queue *q)
1411 {
1412         struct scsi_device *sdev = q->queuedata;
1413         struct Scsi_Host *shost;
1414
1415         if (blk_queue_dying(q))
1416                 return false;
1417
1418         shost = sdev->host;
1419
1420         /*
1421          * Ignore host/starget busy state.
1422          * Since block layer does not have a concept of fairness across
1423          * multiple queues, congestion of host/starget needs to be handled
1424          * in SCSI layer.
1425          */
1426         if (scsi_host_in_recovery(shost) || scsi_device_is_busy(sdev))
1427                 return true;
1428
1429         return false;
1430 }
1431
1432 static void scsi_softirq_done(struct request *rq)
1433 {
1434         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1435         unsigned long wait_for = (cmd->allowed + 1) * rq->timeout;
1436         int disposition;
1437
1438         INIT_LIST_HEAD(&cmd->eh_entry);
1439
1440         atomic_inc(&cmd->device->iodone_cnt);
1441         if (cmd->result)
1442                 atomic_inc(&cmd->device->ioerr_cnt);
1443
1444         disposition = scsi_decide_disposition(cmd);
1445         if (disposition != SUCCESS &&
1446             time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) {
1447                 sdev_printk(KERN_ERR, cmd->device,
1448                             "timing out command, waited %lus\n",
1449                             wait_for/HZ);
1450                 disposition = SUCCESS;
1451         }
1452
1453         scsi_log_completion(cmd, disposition);
1454
1455         switch (disposition) {
1456                 case SUCCESS:
1457                         scsi_finish_command(cmd);
1458                         break;
1459                 case NEEDS_RETRY:
1460                         scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY);
1461                         break;
1462                 case ADD_TO_MLQUEUE:
1463                         scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY);
1464                         break;
1465                 default:
1466                         scsi_eh_scmd_add(cmd);
1467                         break;
1468         }
1469 }
1470
1471 /**
1472  * scsi_dispatch_command - Dispatch a command to the low-level driver.
1473  * @cmd: command block we are dispatching.
1474  *
1475  * Return: nonzero return request was rejected and device's queue needs to be
1476  * plugged.
1477  */
1478 static int scsi_dispatch_cmd(struct scsi_cmnd *cmd)
1479 {
1480         struct Scsi_Host *host = cmd->device->host;
1481         int rtn = 0;
1482
1483         atomic_inc(&cmd->device->iorequest_cnt);
1484
1485         /* check if the device is still usable */
1486         if (unlikely(cmd->device->sdev_state == SDEV_DEL)) {
1487                 /* in SDEV_DEL we error all commands. DID_NO_CONNECT
1488                  * returns an immediate error upwards, and signals
1489                  * that the device is no longer present */
1490                 cmd->result = DID_NO_CONNECT << 16;
1491                 goto done;
1492         }
1493
1494         /* Check to see if the scsi lld made this device blocked. */
1495         if (unlikely(scsi_device_blocked(cmd->device))) {
1496                 /*
1497                  * in blocked state, the command is just put back on
1498                  * the device queue.  The suspend state has already
1499                  * blocked the queue so future requests should not
1500                  * occur until the device transitions out of the
1501                  * suspend state.
1502                  */
1503                 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1504                         "queuecommand : device blocked\n"));
1505                 return SCSI_MLQUEUE_DEVICE_BUSY;
1506         }
1507
1508         /* Store the LUN value in cmnd, if needed. */
1509         if (cmd->device->lun_in_cdb)
1510                 cmd->cmnd[1] = (cmd->cmnd[1] & 0x1f) |
1511                                (cmd->device->lun << 5 & 0xe0);
1512
1513         scsi_log_send(cmd);
1514
1515         /*
1516          * Before we queue this command, check if the command
1517          * length exceeds what the host adapter can handle.
1518          */
1519         if (cmd->cmd_len > cmd->device->host->max_cmd_len) {
1520                 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1521                                "queuecommand : command too long. "
1522                                "cdb_size=%d host->max_cmd_len=%d\n",
1523                                cmd->cmd_len, cmd->device->host->max_cmd_len));
1524                 cmd->result = (DID_ABORT << 16);
1525                 goto done;
1526         }
1527
1528         if (unlikely(host->shost_state == SHOST_DEL)) {
1529                 cmd->result = (DID_NO_CONNECT << 16);
1530                 goto done;
1531
1532         }
1533
1534         trace_scsi_dispatch_cmd_start(cmd);
1535         rtn = host->hostt->queuecommand(host, cmd);
1536         if (rtn) {
1537                 trace_scsi_dispatch_cmd_error(cmd, rtn);
1538                 if (rtn != SCSI_MLQUEUE_DEVICE_BUSY &&
1539                     rtn != SCSI_MLQUEUE_TARGET_BUSY)
1540                         rtn = SCSI_MLQUEUE_HOST_BUSY;
1541
1542                 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1543                         "queuecommand : request rejected\n"));
1544         }
1545
1546         return rtn;
1547  done:
1548         cmd->scsi_done(cmd);
1549         return 0;
1550 }
1551
1552 /* Size in bytes of the sg-list stored in the scsi-mq command-private data. */
1553 static unsigned int scsi_mq_sgl_size(struct Scsi_Host *shost)
1554 {
1555         return min_t(unsigned int, shost->sg_tablesize, SG_CHUNK_SIZE) *
1556                 sizeof(struct scatterlist);
1557 }
1558
1559 static blk_status_t scsi_mq_prep_fn(struct request *req)
1560 {
1561         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1562         struct scsi_device *sdev = req->q->queuedata;
1563         struct Scsi_Host *shost = sdev->host;
1564         struct scatterlist *sg;
1565
1566         scsi_init_command(sdev, cmd);
1567
1568         cmd->request = req;
1569         cmd->tag = req->tag;
1570         cmd->prot_op = SCSI_PROT_NORMAL;
1571
1572         sg = (void *)cmd + sizeof(struct scsi_cmnd) + shost->hostt->cmd_size;
1573         cmd->sdb.table.sgl = sg;
1574
1575         if (scsi_host_get_prot(shost)) {
1576                 memset(cmd->prot_sdb, 0, sizeof(struct scsi_data_buffer));
1577
1578                 cmd->prot_sdb->table.sgl =
1579                         (struct scatterlist *)(cmd->prot_sdb + 1);
1580         }
1581
1582         blk_mq_start_request(req);
1583
1584         return scsi_setup_cmnd(sdev, req);
1585 }
1586
1587 static void scsi_mq_done(struct scsi_cmnd *cmd)
1588 {
1589         if (unlikely(test_and_set_bit(SCMD_STATE_COMPLETE, &cmd->state)))
1590                 return;
1591         trace_scsi_dispatch_cmd_done(cmd);
1592
1593         /*
1594          * If the block layer didn't complete the request due to a timeout
1595          * injection, scsi must clear its internal completed state so that the
1596          * timeout handler will see it needs to escalate its own error
1597          * recovery.
1598          */
1599         if (unlikely(!blk_mq_complete_request(cmd->request)))
1600                 clear_bit(SCMD_STATE_COMPLETE, &cmd->state);
1601 }
1602
1603 static void scsi_mq_put_budget(struct blk_mq_hw_ctx *hctx)
1604 {
1605         struct request_queue *q = hctx->queue;
1606         struct scsi_device *sdev = q->queuedata;
1607
1608         atomic_dec(&sdev->device_busy);
1609         put_device(&sdev->sdev_gendev);
1610 }
1611
1612 static bool scsi_mq_get_budget(struct blk_mq_hw_ctx *hctx)
1613 {
1614         struct request_queue *q = hctx->queue;
1615         struct scsi_device *sdev = q->queuedata;
1616
1617         if (!get_device(&sdev->sdev_gendev))
1618                 goto out;
1619         if (!scsi_dev_queue_ready(q, sdev))
1620                 goto out_put_device;
1621
1622         return true;
1623
1624 out_put_device:
1625         put_device(&sdev->sdev_gendev);
1626 out:
1627         if (atomic_read(&sdev->device_busy) == 0 && !scsi_device_blocked(sdev))
1628                 blk_mq_delay_run_hw_queue(hctx, SCSI_QUEUE_DELAY);
1629         return false;
1630 }
1631
1632 static blk_status_t scsi_queue_rq(struct blk_mq_hw_ctx *hctx,
1633                          const struct blk_mq_queue_data *bd)
1634 {
1635         struct request *req = bd->rq;
1636         struct request_queue *q = req->q;
1637         struct scsi_device *sdev = q->queuedata;
1638         struct Scsi_Host *shost = sdev->host;
1639         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1640         blk_status_t ret;
1641         int reason;
1642
1643         /*
1644          * If the device is not in running state we will reject some or all
1645          * commands.
1646          */
1647         if (unlikely(sdev->sdev_state != SDEV_RUNNING)) {
1648                 ret = scsi_prep_state_check(sdev, req);
1649                 if (ret != BLK_STS_OK)
1650                         goto out_put_budget;
1651         }
1652
1653         ret = BLK_STS_RESOURCE;
1654         if (!scsi_target_queue_ready(shost, sdev))
1655                 goto out_put_budget;
1656         if (!scsi_host_queue_ready(q, shost, sdev))
1657                 goto out_dec_target_busy;
1658
1659         if (!(req->rq_flags & RQF_DONTPREP)) {
1660                 ret = scsi_mq_prep_fn(req);
1661                 if (ret != BLK_STS_OK)
1662                         goto out_dec_host_busy;
1663                 req->rq_flags |= RQF_DONTPREP;
1664         } else {
1665                 clear_bit(SCMD_STATE_COMPLETE, &cmd->state);
1666                 blk_mq_start_request(req);
1667         }
1668
1669         if (sdev->simple_tags)
1670                 cmd->flags |= SCMD_TAGGED;
1671         else
1672                 cmd->flags &= ~SCMD_TAGGED;
1673
1674         scsi_init_cmd_errh(cmd);
1675         cmd->scsi_done = scsi_mq_done;
1676
1677         reason = scsi_dispatch_cmd(cmd);
1678         if (reason) {
1679                 scsi_set_blocked(cmd, reason);
1680                 ret = BLK_STS_RESOURCE;
1681                 goto out_dec_host_busy;
1682         }
1683
1684         return BLK_STS_OK;
1685
1686 out_dec_host_busy:
1687         scsi_dec_host_busy(shost);
1688 out_dec_target_busy:
1689         if (scsi_target(sdev)->can_queue > 0)
1690                 atomic_dec(&scsi_target(sdev)->target_busy);
1691 out_put_budget:
1692         scsi_mq_put_budget(hctx);
1693         switch (ret) {
1694         case BLK_STS_OK:
1695                 break;
1696         case BLK_STS_RESOURCE:
1697                 if (atomic_read(&sdev->device_busy) ||
1698                     scsi_device_blocked(sdev))
1699                         ret = BLK_STS_DEV_RESOURCE;
1700                 break;
1701         default:
1702                 /*
1703                  * Make sure to release all allocated ressources when
1704                  * we hit an error, as we will never see this command
1705                  * again.
1706                  */
1707                 if (req->rq_flags & RQF_DONTPREP)
1708                         scsi_mq_uninit_cmd(cmd);
1709                 break;
1710         }
1711         return ret;
1712 }
1713
1714 static enum blk_eh_timer_return scsi_timeout(struct request *req,
1715                 bool reserved)
1716 {
1717         if (reserved)
1718                 return BLK_EH_RESET_TIMER;
1719         return scsi_times_out(req);
1720 }
1721
1722 static int scsi_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
1723                                 unsigned int hctx_idx, unsigned int numa_node)
1724 {
1725         struct Scsi_Host *shost = set->driver_data;
1726         const bool unchecked_isa_dma = shost->unchecked_isa_dma;
1727         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1728         struct scatterlist *sg;
1729
1730         if (unchecked_isa_dma)
1731                 cmd->flags |= SCMD_UNCHECKED_ISA_DMA;
1732         cmd->sense_buffer = scsi_alloc_sense_buffer(unchecked_isa_dma,
1733                                                     GFP_KERNEL, numa_node);
1734         if (!cmd->sense_buffer)
1735                 return -ENOMEM;
1736         cmd->req.sense = cmd->sense_buffer;
1737
1738         if (scsi_host_get_prot(shost)) {
1739                 sg = (void *)cmd + sizeof(struct scsi_cmnd) +
1740                         shost->hostt->cmd_size;
1741                 cmd->prot_sdb = (void *)sg + scsi_mq_sgl_size(shost);
1742         }
1743
1744         return 0;
1745 }
1746
1747 static void scsi_mq_exit_request(struct blk_mq_tag_set *set, struct request *rq,
1748                                  unsigned int hctx_idx)
1749 {
1750         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1751
1752         scsi_free_sense_buffer(cmd->flags & SCMD_UNCHECKED_ISA_DMA,
1753                                cmd->sense_buffer);
1754 }
1755
1756 static int scsi_map_queues(struct blk_mq_tag_set *set)
1757 {
1758         struct Scsi_Host *shost = container_of(set, struct Scsi_Host, tag_set);
1759
1760         if (shost->hostt->map_queues)
1761                 return shost->hostt->map_queues(shost);
1762         return blk_mq_map_queues(&set->map[0]);
1763 }
1764
1765 void __scsi_init_queue(struct Scsi_Host *shost, struct request_queue *q)
1766 {
1767         struct device *dev = shost->dma_dev;
1768
1769         /*
1770          * this limit is imposed by hardware restrictions
1771          */
1772         blk_queue_max_segments(q, min_t(unsigned short, shost->sg_tablesize,
1773                                         SG_MAX_SEGMENTS));
1774
1775         if (scsi_host_prot_dma(shost)) {
1776                 shost->sg_prot_tablesize =
1777                         min_not_zero(shost->sg_prot_tablesize,
1778                                      (unsigned short)SCSI_MAX_PROT_SG_SEGMENTS);
1779                 BUG_ON(shost->sg_prot_tablesize < shost->sg_tablesize);
1780                 blk_queue_max_integrity_segments(q, shost->sg_prot_tablesize);
1781         }
1782
1783         blk_queue_max_hw_sectors(q, shost->max_sectors);
1784         if (shost->unchecked_isa_dma)
1785                 blk_queue_bounce_limit(q, BLK_BOUNCE_ISA);
1786         blk_queue_segment_boundary(q, shost->dma_boundary);
1787         dma_set_seg_boundary(dev, shost->dma_boundary);
1788
1789         blk_queue_max_segment_size(q, shost->max_segment_size);
1790         dma_set_max_seg_size(dev, shost->max_segment_size);
1791
1792         /*
1793          * Set a reasonable default alignment:  The larger of 32-byte (dword),
1794          * which is a common minimum for HBAs, and the minimum DMA alignment,
1795          * which is set by the platform.
1796          *
1797          * Devices that require a bigger alignment can increase it later.
1798          */
1799         blk_queue_dma_alignment(q, max(4, dma_get_cache_alignment()) - 1);
1800 }
1801 EXPORT_SYMBOL_GPL(__scsi_init_queue);
1802
1803 static const struct blk_mq_ops scsi_mq_ops = {
1804         .get_budget     = scsi_mq_get_budget,
1805         .put_budget     = scsi_mq_put_budget,
1806         .queue_rq       = scsi_queue_rq,
1807         .complete       = scsi_softirq_done,
1808         .timeout        = scsi_timeout,
1809 #ifdef CONFIG_BLK_DEBUG_FS
1810         .show_rq        = scsi_show_rq,
1811 #endif
1812         .init_request   = scsi_mq_init_request,
1813         .exit_request   = scsi_mq_exit_request,
1814         .initialize_rq_fn = scsi_initialize_rq,
1815         .busy           = scsi_mq_lld_busy,
1816         .map_queues     = scsi_map_queues,
1817 };
1818
1819 struct request_queue *scsi_mq_alloc_queue(struct scsi_device *sdev)
1820 {
1821         sdev->request_queue = blk_mq_init_queue(&sdev->host->tag_set);
1822         if (IS_ERR(sdev->request_queue))
1823                 return NULL;
1824
1825         sdev->request_queue->queuedata = sdev;
1826         __scsi_init_queue(sdev->host, sdev->request_queue);
1827         blk_queue_flag_set(QUEUE_FLAG_SCSI_PASSTHROUGH, sdev->request_queue);
1828         return sdev->request_queue;
1829 }
1830
1831 int scsi_mq_setup_tags(struct Scsi_Host *shost)
1832 {
1833         unsigned int cmd_size, sgl_size;
1834
1835         sgl_size = scsi_mq_sgl_size(shost);
1836         cmd_size = sizeof(struct scsi_cmnd) + shost->hostt->cmd_size + sgl_size;
1837         if (scsi_host_get_prot(shost))
1838                 cmd_size += sizeof(struct scsi_data_buffer) + sgl_size;
1839
1840         memset(&shost->tag_set, 0, sizeof(shost->tag_set));
1841         shost->tag_set.ops = &scsi_mq_ops;
1842         shost->tag_set.nr_hw_queues = shost->nr_hw_queues ? : 1;
1843         shost->tag_set.queue_depth = shost->can_queue;
1844         shost->tag_set.cmd_size = cmd_size;
1845         shost->tag_set.numa_node = NUMA_NO_NODE;
1846         shost->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
1847         shost->tag_set.flags |=
1848                 BLK_ALLOC_POLICY_TO_MQ_FLAG(shost->hostt->tag_alloc_policy);
1849         shost->tag_set.driver_data = shost;
1850
1851         return blk_mq_alloc_tag_set(&shost->tag_set);
1852 }
1853
1854 void scsi_mq_destroy_tags(struct Scsi_Host *shost)
1855 {
1856         blk_mq_free_tag_set(&shost->tag_set);
1857 }
1858
1859 /**
1860  * scsi_device_from_queue - return sdev associated with a request_queue
1861  * @q: The request queue to return the sdev from
1862  *
1863  * Return the sdev associated with a request queue or NULL if the
1864  * request_queue does not reference a SCSI device.
1865  */
1866 struct scsi_device *scsi_device_from_queue(struct request_queue *q)
1867 {
1868         struct scsi_device *sdev = NULL;
1869
1870         if (q->mq_ops == &scsi_mq_ops)
1871                 sdev = q->queuedata;
1872         if (!sdev || !get_device(&sdev->sdev_gendev))
1873                 sdev = NULL;
1874
1875         return sdev;
1876 }
1877 EXPORT_SYMBOL_GPL(scsi_device_from_queue);
1878
1879 /*
1880  * Function:    scsi_block_requests()
1881  *
1882  * Purpose:     Utility function used by low-level drivers to prevent further
1883  *              commands from being queued to the device.
1884  *
1885  * Arguments:   shost       - Host in question
1886  *
1887  * Returns:     Nothing
1888  *
1889  * Lock status: No locks are assumed held.
1890  *
1891  * Notes:       There is no timer nor any other means by which the requests
1892  *              get unblocked other than the low-level driver calling
1893  *              scsi_unblock_requests().
1894  */
1895 void scsi_block_requests(struct Scsi_Host *shost)
1896 {
1897         shost->host_self_blocked = 1;
1898 }
1899 EXPORT_SYMBOL(scsi_block_requests);
1900
1901 /*
1902  * Function:    scsi_unblock_requests()
1903  *
1904  * Purpose:     Utility function used by low-level drivers to allow further
1905  *              commands from being queued to the device.
1906  *
1907  * Arguments:   shost       - Host in question
1908  *
1909  * Returns:     Nothing
1910  *
1911  * Lock status: No locks are assumed held.
1912  *
1913  * Notes:       There is no timer nor any other means by which the requests
1914  *              get unblocked other than the low-level driver calling
1915  *              scsi_unblock_requests().
1916  *
1917  *              This is done as an API function so that changes to the
1918  *              internals of the scsi mid-layer won't require wholesale
1919  *              changes to drivers that use this feature.
1920  */
1921 void scsi_unblock_requests(struct Scsi_Host *shost)
1922 {
1923         shost->host_self_blocked = 0;
1924         scsi_run_host_queues(shost);
1925 }
1926 EXPORT_SYMBOL(scsi_unblock_requests);
1927
1928 int __init scsi_init_queue(void)
1929 {
1930         scsi_sdb_cache = kmem_cache_create("scsi_data_buffer",
1931                                            sizeof(struct scsi_data_buffer),
1932                                            0, 0, NULL);
1933         if (!scsi_sdb_cache) {
1934                 printk(KERN_ERR "SCSI: can't init scsi sdb cache\n");
1935                 return -ENOMEM;
1936         }
1937
1938         return 0;
1939 }
1940
1941 void scsi_exit_queue(void)
1942 {
1943         kmem_cache_destroy(scsi_sense_cache);
1944         kmem_cache_destroy(scsi_sense_isadma_cache);
1945         kmem_cache_destroy(scsi_sdb_cache);
1946 }
1947
1948 /**
1949  *      scsi_mode_select - issue a mode select
1950  *      @sdev:  SCSI device to be queried
1951  *      @pf:    Page format bit (1 == standard, 0 == vendor specific)
1952  *      @sp:    Save page bit (0 == don't save, 1 == save)
1953  *      @modepage: mode page being requested
1954  *      @buffer: request buffer (may not be smaller than eight bytes)
1955  *      @len:   length of request buffer.
1956  *      @timeout: command timeout
1957  *      @retries: number of retries before failing
1958  *      @data: returns a structure abstracting the mode header data
1959  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
1960  *              must be SCSI_SENSE_BUFFERSIZE big.
1961  *
1962  *      Returns zero if successful; negative error number or scsi
1963  *      status on error
1964  *
1965  */
1966 int
1967 scsi_mode_select(struct scsi_device *sdev, int pf, int sp, int modepage,
1968                  unsigned char *buffer, int len, int timeout, int retries,
1969                  struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1970 {
1971         unsigned char cmd[10];
1972         unsigned char *real_buffer;
1973         int ret;
1974
1975         memset(cmd, 0, sizeof(cmd));
1976         cmd[1] = (pf ? 0x10 : 0) | (sp ? 0x01 : 0);
1977
1978         if (sdev->use_10_for_ms) {
1979                 if (len > 65535)
1980                         return -EINVAL;
1981                 real_buffer = kmalloc(8 + len, GFP_KERNEL);
1982                 if (!real_buffer)
1983                         return -ENOMEM;
1984                 memcpy(real_buffer + 8, buffer, len);
1985                 len += 8;
1986                 real_buffer[0] = 0;
1987                 real_buffer[1] = 0;
1988                 real_buffer[2] = data->medium_type;
1989                 real_buffer[3] = data->device_specific;
1990                 real_buffer[4] = data->longlba ? 0x01 : 0;
1991                 real_buffer[5] = 0;
1992                 real_buffer[6] = data->block_descriptor_length >> 8;
1993                 real_buffer[7] = data->block_descriptor_length;
1994
1995                 cmd[0] = MODE_SELECT_10;
1996                 cmd[7] = len >> 8;
1997                 cmd[8] = len;
1998         } else {
1999                 if (len > 255 || data->block_descriptor_length > 255 ||
2000                     data->longlba)
2001                         return -EINVAL;
2002
2003                 real_buffer = kmalloc(4 + len, GFP_KERNEL);
2004                 if (!real_buffer)
2005                         return -ENOMEM;
2006                 memcpy(real_buffer + 4, buffer, len);
2007                 len += 4;
2008                 real_buffer[0] = 0;
2009                 real_buffer[1] = data->medium_type;
2010                 real_buffer[2] = data->device_specific;
2011                 real_buffer[3] = data->block_descriptor_length;
2012                 
2013
2014                 cmd[0] = MODE_SELECT;
2015                 cmd[4] = len;
2016         }
2017
2018         ret = scsi_execute_req(sdev, cmd, DMA_TO_DEVICE, real_buffer, len,
2019                                sshdr, timeout, retries, NULL);
2020         kfree(real_buffer);
2021         return ret;
2022 }
2023 EXPORT_SYMBOL_GPL(scsi_mode_select);
2024
2025 /**
2026  *      scsi_mode_sense - issue a mode sense, falling back from 10 to six bytes if necessary.
2027  *      @sdev:  SCSI device to be queried
2028  *      @dbd:   set if mode sense will allow block descriptors to be returned
2029  *      @modepage: mode page being requested
2030  *      @buffer: request buffer (may not be smaller than eight bytes)
2031  *      @len:   length of request buffer.
2032  *      @timeout: command timeout
2033  *      @retries: number of retries before failing
2034  *      @data: returns a structure abstracting the mode header data
2035  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
2036  *              must be SCSI_SENSE_BUFFERSIZE big.
2037  *
2038  *      Returns zero if unsuccessful, or the header offset (either 4
2039  *      or 8 depending on whether a six or ten byte command was
2040  *      issued) if successful.
2041  */
2042 int
2043 scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
2044                   unsigned char *buffer, int len, int timeout, int retries,
2045                   struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
2046 {
2047         unsigned char cmd[12];
2048         int use_10_for_ms;
2049         int header_length;
2050         int result, retry_count = retries;
2051         struct scsi_sense_hdr my_sshdr;
2052
2053         memset(data, 0, sizeof(*data));
2054         memset(&cmd[0], 0, 12);
2055         cmd[1] = dbd & 0x18;    /* allows DBD and LLBA bits */
2056         cmd[2] = modepage;
2057
2058         /* caller might not be interested in sense, but we need it */
2059         if (!sshdr)
2060                 sshdr = &my_sshdr;
2061
2062  retry:
2063         use_10_for_ms = sdev->use_10_for_ms;
2064
2065         if (use_10_for_ms) {
2066                 if (len < 8)
2067                         len = 8;
2068
2069                 cmd[0] = MODE_SENSE_10;
2070                 cmd[8] = len;
2071                 header_length = 8;
2072         } else {
2073                 if (len < 4)
2074                         len = 4;
2075
2076                 cmd[0] = MODE_SENSE;
2077                 cmd[4] = len;
2078                 header_length = 4;
2079         }
2080
2081         memset(buffer, 0, len);
2082
2083         result = scsi_execute_req(sdev, cmd, DMA_FROM_DEVICE, buffer, len,
2084                                   sshdr, timeout, retries, NULL);
2085
2086         /* This code looks awful: what it's doing is making sure an
2087          * ILLEGAL REQUEST sense return identifies the actual command
2088          * byte as the problem.  MODE_SENSE commands can return
2089          * ILLEGAL REQUEST if the code page isn't supported */
2090
2091         if (use_10_for_ms && !scsi_status_is_good(result) &&
2092             driver_byte(result) == DRIVER_SENSE) {
2093                 if (scsi_sense_valid(sshdr)) {
2094                         if ((sshdr->sense_key == ILLEGAL_REQUEST) &&
2095                             (sshdr->asc == 0x20) && (sshdr->ascq == 0)) {
2096                                 /* 
2097                                  * Invalid command operation code
2098                                  */
2099                                 sdev->use_10_for_ms = 0;
2100                                 goto retry;
2101                         }
2102                 }
2103         }
2104
2105         if(scsi_status_is_good(result)) {
2106                 if (unlikely(buffer[0] == 0x86 && buffer[1] == 0x0b &&
2107                              (modepage == 6 || modepage == 8))) {
2108                         /* Initio breakage? */
2109                         header_length = 0;
2110                         data->length = 13;
2111                         data->medium_type = 0;
2112                         data->device_specific = 0;
2113                         data->longlba = 0;
2114                         data->block_descriptor_length = 0;
2115                 } else if(use_10_for_ms) {
2116                         data->length = buffer[0]*256 + buffer[1] + 2;
2117                         data->medium_type = buffer[2];
2118                         data->device_specific = buffer[3];
2119                         data->longlba = buffer[4] & 0x01;
2120                         data->block_descriptor_length = buffer[6]*256
2121                                 + buffer[7];
2122                 } else {
2123                         data->length = buffer[0] + 1;
2124                         data->medium_type = buffer[1];
2125                         data->device_specific = buffer[2];
2126                         data->block_descriptor_length = buffer[3];
2127                 }
2128                 data->header_length = header_length;
2129         } else if ((status_byte(result) == CHECK_CONDITION) &&
2130                    scsi_sense_valid(sshdr) &&
2131                    sshdr->sense_key == UNIT_ATTENTION && retry_count) {
2132                 retry_count--;
2133                 goto retry;
2134         }
2135
2136         return result;
2137 }
2138 EXPORT_SYMBOL(scsi_mode_sense);
2139
2140 /**
2141  *      scsi_test_unit_ready - test if unit is ready
2142  *      @sdev:  scsi device to change the state of.
2143  *      @timeout: command timeout
2144  *      @retries: number of retries before failing
2145  *      @sshdr: outpout pointer for decoded sense information.
2146  *
2147  *      Returns zero if unsuccessful or an error if TUR failed.  For
2148  *      removable media, UNIT_ATTENTION sets ->changed flag.
2149  **/
2150 int
2151 scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries,
2152                      struct scsi_sense_hdr *sshdr)
2153 {
2154         char cmd[] = {
2155                 TEST_UNIT_READY, 0, 0, 0, 0, 0,
2156         };
2157         int result;
2158
2159         /* try to eat the UNIT_ATTENTION if there are enough retries */
2160         do {
2161                 result = scsi_execute_req(sdev, cmd, DMA_NONE, NULL, 0, sshdr,
2162                                           timeout, 1, NULL);
2163                 if (sdev->removable && scsi_sense_valid(sshdr) &&
2164                     sshdr->sense_key == UNIT_ATTENTION)
2165                         sdev->changed = 1;
2166         } while (scsi_sense_valid(sshdr) &&
2167                  sshdr->sense_key == UNIT_ATTENTION && --retries);
2168
2169         return result;
2170 }
2171 EXPORT_SYMBOL(scsi_test_unit_ready);
2172
2173 /**
2174  *      scsi_device_set_state - Take the given device through the device state model.
2175  *      @sdev:  scsi device to change the state of.
2176  *      @state: state to change to.
2177  *
2178  *      Returns zero if successful or an error if the requested
2179  *      transition is illegal.
2180  */
2181 int
2182 scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
2183 {
2184         enum scsi_device_state oldstate = sdev->sdev_state;
2185
2186         if (state == oldstate)
2187                 return 0;
2188
2189         switch (state) {
2190         case SDEV_CREATED:
2191                 switch (oldstate) {
2192                 case SDEV_CREATED_BLOCK:
2193                         break;
2194                 default:
2195                         goto illegal;
2196                 }
2197                 break;
2198                         
2199         case SDEV_RUNNING:
2200                 switch (oldstate) {
2201                 case SDEV_CREATED:
2202                 case SDEV_OFFLINE:
2203                 case SDEV_TRANSPORT_OFFLINE:
2204                 case SDEV_QUIESCE:
2205                 case SDEV_BLOCK:
2206                         break;
2207                 default:
2208                         goto illegal;
2209                 }
2210                 break;
2211
2212         case SDEV_QUIESCE:
2213                 switch (oldstate) {
2214                 case SDEV_RUNNING:
2215                 case SDEV_OFFLINE:
2216                 case SDEV_TRANSPORT_OFFLINE:
2217                         break;
2218                 default:
2219                         goto illegal;
2220                 }
2221                 break;
2222
2223         case SDEV_OFFLINE:
2224         case SDEV_TRANSPORT_OFFLINE:
2225                 switch (oldstate) {
2226                 case SDEV_CREATED:
2227                 case SDEV_RUNNING:
2228                 case SDEV_QUIESCE:
2229                 case SDEV_BLOCK:
2230                         break;
2231                 default:
2232                         goto illegal;
2233                 }
2234                 break;
2235
2236         case SDEV_BLOCK:
2237                 switch (oldstate) {
2238                 case SDEV_RUNNING:
2239                 case SDEV_CREATED_BLOCK:
2240                 case SDEV_OFFLINE:
2241                         break;
2242                 default:
2243                         goto illegal;
2244                 }
2245                 break;
2246
2247         case SDEV_CREATED_BLOCK:
2248                 switch (oldstate) {
2249                 case SDEV_CREATED:
2250                         break;
2251                 default:
2252                         goto illegal;
2253                 }
2254                 break;
2255
2256         case SDEV_CANCEL:
2257                 switch (oldstate) {
2258                 case SDEV_CREATED:
2259                 case SDEV_RUNNING:
2260                 case SDEV_QUIESCE:
2261                 case SDEV_OFFLINE:
2262                 case SDEV_TRANSPORT_OFFLINE:
2263                         break;
2264                 default:
2265                         goto illegal;
2266                 }
2267                 break;
2268
2269         case SDEV_DEL:
2270                 switch (oldstate) {
2271                 case SDEV_CREATED:
2272                 case SDEV_RUNNING:
2273                 case SDEV_OFFLINE:
2274                 case SDEV_TRANSPORT_OFFLINE:
2275                 case SDEV_CANCEL:
2276                 case SDEV_BLOCK:
2277                 case SDEV_CREATED_BLOCK:
2278                         break;
2279                 default:
2280                         goto illegal;
2281                 }
2282                 break;
2283
2284         }
2285         sdev->sdev_state = state;
2286         return 0;
2287
2288  illegal:
2289         SCSI_LOG_ERROR_RECOVERY(1,
2290                                 sdev_printk(KERN_ERR, sdev,
2291                                             "Illegal state transition %s->%s",
2292                                             scsi_device_state_name(oldstate),
2293                                             scsi_device_state_name(state))
2294                                 );
2295         return -EINVAL;
2296 }
2297 EXPORT_SYMBOL(scsi_device_set_state);
2298
2299 /**
2300  *      sdev_evt_emit - emit a single SCSI device uevent
2301  *      @sdev: associated SCSI device
2302  *      @evt: event to emit
2303  *
2304  *      Send a single uevent (scsi_event) to the associated scsi_device.
2305  */
2306 static void scsi_evt_emit(struct scsi_device *sdev, struct scsi_event *evt)
2307 {
2308         int idx = 0;
2309         char *envp[3];
2310
2311         switch (evt->evt_type) {
2312         case SDEV_EVT_MEDIA_CHANGE:
2313                 envp[idx++] = "SDEV_MEDIA_CHANGE=1";
2314                 break;
2315         case SDEV_EVT_INQUIRY_CHANGE_REPORTED:
2316                 scsi_rescan_device(&sdev->sdev_gendev);
2317                 envp[idx++] = "SDEV_UA=INQUIRY_DATA_HAS_CHANGED";
2318                 break;
2319         case SDEV_EVT_CAPACITY_CHANGE_REPORTED:
2320                 envp[idx++] = "SDEV_UA=CAPACITY_DATA_HAS_CHANGED";
2321                 break;
2322         case SDEV_EVT_SOFT_THRESHOLD_REACHED_REPORTED:
2323                envp[idx++] = "SDEV_UA=THIN_PROVISIONING_SOFT_THRESHOLD_REACHED";
2324                 break;
2325         case SDEV_EVT_MODE_PARAMETER_CHANGE_REPORTED:
2326                 envp[idx++] = "SDEV_UA=MODE_PARAMETERS_CHANGED";
2327                 break;
2328         case SDEV_EVT_LUN_CHANGE_REPORTED:
2329                 envp[idx++] = "SDEV_UA=REPORTED_LUNS_DATA_HAS_CHANGED";
2330                 break;
2331         case SDEV_EVT_ALUA_STATE_CHANGE_REPORTED:
2332                 envp[idx++] = "SDEV_UA=ASYMMETRIC_ACCESS_STATE_CHANGED";
2333                 break;
2334         case SDEV_EVT_POWER_ON_RESET_OCCURRED:
2335                 envp[idx++] = "SDEV_UA=POWER_ON_RESET_OCCURRED";
2336                 break;
2337         default:
2338                 /* do nothing */
2339                 break;
2340         }
2341
2342         envp[idx++] = NULL;
2343
2344         kobject_uevent_env(&sdev->sdev_gendev.kobj, KOBJ_CHANGE, envp);
2345 }
2346
2347 /**
2348  *      sdev_evt_thread - send a uevent for each scsi event
2349  *      @work: work struct for scsi_device
2350  *
2351  *      Dispatch queued events to their associated scsi_device kobjects
2352  *      as uevents.
2353  */
2354 void scsi_evt_thread(struct work_struct *work)
2355 {
2356         struct scsi_device *sdev;
2357         enum scsi_device_event evt_type;
2358         LIST_HEAD(event_list);
2359
2360         sdev = container_of(work, struct scsi_device, event_work);
2361
2362         for (evt_type = SDEV_EVT_FIRST; evt_type <= SDEV_EVT_LAST; evt_type++)
2363                 if (test_and_clear_bit(evt_type, sdev->pending_events))
2364                         sdev_evt_send_simple(sdev, evt_type, GFP_KERNEL);
2365
2366         while (1) {
2367                 struct scsi_event *evt;
2368                 struct list_head *this, *tmp;
2369                 unsigned long flags;
2370
2371                 spin_lock_irqsave(&sdev->list_lock, flags);
2372                 list_splice_init(&sdev->event_list, &event_list);
2373                 spin_unlock_irqrestore(&sdev->list_lock, flags);
2374
2375                 if (list_empty(&event_list))
2376                         break;
2377
2378                 list_for_each_safe(this, tmp, &event_list) {
2379                         evt = list_entry(this, struct scsi_event, node);
2380                         list_del(&evt->node);
2381                         scsi_evt_emit(sdev, evt);
2382                         kfree(evt);
2383                 }
2384         }
2385 }
2386
2387 /**
2388  *      sdev_evt_send - send asserted event to uevent thread
2389  *      @sdev: scsi_device event occurred on
2390  *      @evt: event to send
2391  *
2392  *      Assert scsi device event asynchronously.
2393  */
2394 void sdev_evt_send(struct scsi_device *sdev, struct scsi_event *evt)
2395 {
2396         unsigned long flags;
2397
2398 #if 0
2399         /* FIXME: currently this check eliminates all media change events
2400          * for polled devices.  Need to update to discriminate between AN
2401          * and polled events */
2402         if (!test_bit(evt->evt_type, sdev->supported_events)) {
2403                 kfree(evt);
2404                 return;
2405         }
2406 #endif
2407
2408         spin_lock_irqsave(&sdev->list_lock, flags);
2409         list_add_tail(&evt->node, &sdev->event_list);
2410         schedule_work(&sdev->event_work);
2411         spin_unlock_irqrestore(&sdev->list_lock, flags);
2412 }
2413 EXPORT_SYMBOL_GPL(sdev_evt_send);
2414
2415 /**
2416  *      sdev_evt_alloc - allocate a new scsi event
2417  *      @evt_type: type of event to allocate
2418  *      @gfpflags: GFP flags for allocation
2419  *
2420  *      Allocates and returns a new scsi_event.
2421  */
2422 struct scsi_event *sdev_evt_alloc(enum scsi_device_event evt_type,
2423                                   gfp_t gfpflags)
2424 {
2425         struct scsi_event *evt = kzalloc(sizeof(struct scsi_event), gfpflags);
2426         if (!evt)
2427                 return NULL;
2428
2429         evt->evt_type = evt_type;
2430         INIT_LIST_HEAD(&evt->node);
2431
2432         /* evt_type-specific initialization, if any */
2433         switch (evt_type) {
2434         case SDEV_EVT_MEDIA_CHANGE:
2435         case SDEV_EVT_INQUIRY_CHANGE_REPORTED:
2436         case SDEV_EVT_CAPACITY_CHANGE_REPORTED:
2437         case SDEV_EVT_SOFT_THRESHOLD_REACHED_REPORTED:
2438         case SDEV_EVT_MODE_PARAMETER_CHANGE_REPORTED:
2439         case SDEV_EVT_LUN_CHANGE_REPORTED:
2440         case SDEV_EVT_ALUA_STATE_CHANGE_REPORTED:
2441         case SDEV_EVT_POWER_ON_RESET_OCCURRED:
2442         default:
2443                 /* do nothing */
2444                 break;
2445         }
2446
2447         return evt;
2448 }
2449 EXPORT_SYMBOL_GPL(sdev_evt_alloc);
2450
2451 /**
2452  *      sdev_evt_send_simple - send asserted event to uevent thread
2453  *      @sdev: scsi_device event occurred on
2454  *      @evt_type: type of event to send
2455  *      @gfpflags: GFP flags for allocation
2456  *
2457  *      Assert scsi device event asynchronously, given an event type.
2458  */
2459 void sdev_evt_send_simple(struct scsi_device *sdev,
2460                           enum scsi_device_event evt_type, gfp_t gfpflags)
2461 {
2462         struct scsi_event *evt = sdev_evt_alloc(evt_type, gfpflags);
2463         if (!evt) {
2464                 sdev_printk(KERN_ERR, sdev, "event %d eaten due to OOM\n",
2465                             evt_type);
2466                 return;
2467         }
2468
2469         sdev_evt_send(sdev, evt);
2470 }
2471 EXPORT_SYMBOL_GPL(sdev_evt_send_simple);
2472
2473 /**
2474  *      scsi_device_quiesce - Block user issued commands.
2475  *      @sdev:  scsi device to quiesce.
2476  *
2477  *      This works by trying to transition to the SDEV_QUIESCE state
2478  *      (which must be a legal transition).  When the device is in this
2479  *      state, only special requests will be accepted, all others will
2480  *      be deferred.  Since special requests may also be requeued requests,
2481  *      a successful return doesn't guarantee the device will be 
2482  *      totally quiescent.
2483  *
2484  *      Must be called with user context, may sleep.
2485  *
2486  *      Returns zero if unsuccessful or an error if not.
2487  */
2488 int
2489 scsi_device_quiesce(struct scsi_device *sdev)
2490 {
2491         struct request_queue *q = sdev->request_queue;
2492         int err;
2493
2494         /*
2495          * It is allowed to call scsi_device_quiesce() multiple times from
2496          * the same context but concurrent scsi_device_quiesce() calls are
2497          * not allowed.
2498          */
2499         WARN_ON_ONCE(sdev->quiesced_by && sdev->quiesced_by != current);
2500
2501         if (sdev->quiesced_by == current)
2502                 return 0;
2503
2504         blk_set_pm_only(q);
2505
2506         blk_mq_freeze_queue(q);
2507         /*
2508          * Ensure that the effect of blk_set_pm_only() will be visible
2509          * for percpu_ref_tryget() callers that occur after the queue
2510          * unfreeze even if the queue was already frozen before this function
2511          * was called. See also https://lwn.net/Articles/573497/.
2512          */
2513         synchronize_rcu();
2514         blk_mq_unfreeze_queue(q);
2515
2516         mutex_lock(&sdev->state_mutex);
2517         err = scsi_device_set_state(sdev, SDEV_QUIESCE);
2518         if (err == 0)
2519                 sdev->quiesced_by = current;
2520         else
2521                 blk_clear_pm_only(q);
2522         mutex_unlock(&sdev->state_mutex);
2523
2524         return err;
2525 }
2526 EXPORT_SYMBOL(scsi_device_quiesce);
2527
2528 /**
2529  *      scsi_device_resume - Restart user issued commands to a quiesced device.
2530  *      @sdev:  scsi device to resume.
2531  *
2532  *      Moves the device from quiesced back to running and restarts the
2533  *      queues.
2534  *
2535  *      Must be called with user context, may sleep.
2536  */
2537 void scsi_device_resume(struct scsi_device *sdev)
2538 {
2539         /* check if the device state was mutated prior to resume, and if
2540          * so assume the state is being managed elsewhere (for example
2541          * device deleted during suspend)
2542          */
2543         mutex_lock(&sdev->state_mutex);
2544         sdev->quiesced_by = NULL;
2545         blk_clear_pm_only(sdev->request_queue);
2546         if (sdev->sdev_state == SDEV_QUIESCE)
2547                 scsi_device_set_state(sdev, SDEV_RUNNING);
2548         mutex_unlock(&sdev->state_mutex);
2549 }
2550 EXPORT_SYMBOL(scsi_device_resume);
2551
2552 static void
2553 device_quiesce_fn(struct scsi_device *sdev, void *data)
2554 {
2555         scsi_device_quiesce(sdev);
2556 }
2557
2558 void
2559 scsi_target_quiesce(struct scsi_target *starget)
2560 {
2561         starget_for_each_device(starget, NULL, device_quiesce_fn);
2562 }
2563 EXPORT_SYMBOL(scsi_target_quiesce);
2564
2565 static void
2566 device_resume_fn(struct scsi_device *sdev, void *data)
2567 {
2568         scsi_device_resume(sdev);
2569 }
2570
2571 void
2572 scsi_target_resume(struct scsi_target *starget)
2573 {
2574         starget_for_each_device(starget, NULL, device_resume_fn);
2575 }
2576 EXPORT_SYMBOL(scsi_target_resume);
2577
2578 /**
2579  * scsi_internal_device_block_nowait - try to transition to the SDEV_BLOCK state
2580  * @sdev: device to block
2581  *
2582  * Pause SCSI command processing on the specified device. Does not sleep.
2583  *
2584  * Returns zero if successful or a negative error code upon failure.
2585  *
2586  * Notes:
2587  * This routine transitions the device to the SDEV_BLOCK state (which must be
2588  * a legal transition). When the device is in this state, command processing
2589  * is paused until the device leaves the SDEV_BLOCK state. See also
2590  * scsi_internal_device_unblock_nowait().
2591  */
2592 int scsi_internal_device_block_nowait(struct scsi_device *sdev)
2593 {
2594         struct request_queue *q = sdev->request_queue;
2595         int err = 0;
2596
2597         err = scsi_device_set_state(sdev, SDEV_BLOCK);
2598         if (err) {
2599                 err = scsi_device_set_state(sdev, SDEV_CREATED_BLOCK);
2600
2601                 if (err)
2602                         return err;
2603         }
2604
2605         /* 
2606          * The device has transitioned to SDEV_BLOCK.  Stop the
2607          * block layer from calling the midlayer with this device's
2608          * request queue. 
2609          */
2610         blk_mq_quiesce_queue_nowait(q);
2611         return 0;
2612 }
2613 EXPORT_SYMBOL_GPL(scsi_internal_device_block_nowait);
2614
2615 /**
2616  * scsi_internal_device_block - try to transition to the SDEV_BLOCK state
2617  * @sdev: device to block
2618  *
2619  * Pause SCSI command processing on the specified device and wait until all
2620  * ongoing scsi_request_fn() / scsi_queue_rq() calls have finished. May sleep.
2621  *
2622  * Returns zero if successful or a negative error code upon failure.
2623  *
2624  * Note:
2625  * This routine transitions the device to the SDEV_BLOCK state (which must be
2626  * a legal transition). When the device is in this state, command processing
2627  * is paused until the device leaves the SDEV_BLOCK state. See also
2628  * scsi_internal_device_unblock().
2629  *
2630  * To do: avoid that scsi_send_eh_cmnd() calls queuecommand() after
2631  * scsi_internal_device_block() has blocked a SCSI device and also
2632  * remove the rport mutex lock and unlock calls from srp_queuecommand().
2633  */
2634 static int scsi_internal_device_block(struct scsi_device *sdev)
2635 {
2636         struct request_queue *q = sdev->request_queue;
2637         int err;
2638
2639         mutex_lock(&sdev->state_mutex);
2640         err = scsi_internal_device_block_nowait(sdev);
2641         if (err == 0)
2642                 blk_mq_quiesce_queue(q);
2643         mutex_unlock(&sdev->state_mutex);
2644
2645         return err;
2646 }
2647  
2648 void scsi_start_queue(struct scsi_device *sdev)
2649 {
2650         struct request_queue *q = sdev->request_queue;
2651
2652         blk_mq_unquiesce_queue(q);
2653 }
2654
2655 /**
2656  * scsi_internal_device_unblock_nowait - resume a device after a block request
2657  * @sdev:       device to resume
2658  * @new_state:  state to set the device to after unblocking
2659  *
2660  * Restart the device queue for a previously suspended SCSI device. Does not
2661  * sleep.
2662  *
2663  * Returns zero if successful or a negative error code upon failure.
2664  *
2665  * Notes:
2666  * This routine transitions the device to the SDEV_RUNNING state or to one of
2667  * the offline states (which must be a legal transition) allowing the midlayer
2668  * to goose the queue for this device.
2669  */
2670 int scsi_internal_device_unblock_nowait(struct scsi_device *sdev,
2671                                         enum scsi_device_state new_state)
2672 {
2673         /*
2674          * Try to transition the scsi device to SDEV_RUNNING or one of the
2675          * offlined states and goose the device queue if successful.
2676          */
2677         switch (sdev->sdev_state) {
2678         case SDEV_BLOCK:
2679         case SDEV_TRANSPORT_OFFLINE:
2680                 sdev->sdev_state = new_state;
2681                 break;
2682         case SDEV_CREATED_BLOCK:
2683                 if (new_state == SDEV_TRANSPORT_OFFLINE ||
2684                     new_state == SDEV_OFFLINE)
2685                         sdev->sdev_state = new_state;
2686                 else
2687                         sdev->sdev_state = SDEV_CREATED;
2688                 break;
2689         case SDEV_CANCEL:
2690         case SDEV_OFFLINE:
2691                 break;
2692         default:
2693                 return -EINVAL;
2694         }
2695         scsi_start_queue(sdev);
2696
2697         return 0;
2698 }
2699 EXPORT_SYMBOL_GPL(scsi_internal_device_unblock_nowait);
2700
2701 /**
2702  * scsi_internal_device_unblock - resume a device after a block request
2703  * @sdev:       device to resume
2704  * @new_state:  state to set the device to after unblocking
2705  *
2706  * Restart the device queue for a previously suspended SCSI device. May sleep.
2707  *
2708  * Returns zero if successful or a negative error code upon failure.
2709  *
2710  * Notes:
2711  * This routine transitions the device to the SDEV_RUNNING state or to one of
2712  * the offline states (which must be a legal transition) allowing the midlayer
2713  * to goose the queue for this device.
2714  */
2715 static int scsi_internal_device_unblock(struct scsi_device *sdev,
2716                                         enum scsi_device_state new_state)
2717 {
2718         int ret;
2719
2720         mutex_lock(&sdev->state_mutex);
2721         ret = scsi_internal_device_unblock_nowait(sdev, new_state);
2722         mutex_unlock(&sdev->state_mutex);
2723
2724         return ret;
2725 }
2726
2727 static void
2728 device_block(struct scsi_device *sdev, void *data)
2729 {
2730         scsi_internal_device_block(sdev);
2731 }
2732
2733 static int
2734 target_block(struct device *dev, void *data)
2735 {
2736         if (scsi_is_target_device(dev))
2737                 starget_for_each_device(to_scsi_target(dev), NULL,
2738                                         device_block);
2739         return 0;
2740 }
2741
2742 void
2743 scsi_target_block(struct device *dev)
2744 {
2745         if (scsi_is_target_device(dev))
2746                 starget_for_each_device(to_scsi_target(dev), NULL,
2747                                         device_block);
2748         else
2749                 device_for_each_child(dev, NULL, target_block);
2750 }
2751 EXPORT_SYMBOL_GPL(scsi_target_block);
2752
2753 static void
2754 device_unblock(struct scsi_device *sdev, void *data)
2755 {
2756         scsi_internal_device_unblock(sdev, *(enum scsi_device_state *)data);
2757 }
2758
2759 static int
2760 target_unblock(struct device *dev, void *data)
2761 {
2762         if (scsi_is_target_device(dev))
2763                 starget_for_each_device(to_scsi_target(dev), data,
2764                                         device_unblock);
2765         return 0;
2766 }
2767
2768 void
2769 scsi_target_unblock(struct device *dev, enum scsi_device_state new_state)
2770 {
2771         if (scsi_is_target_device(dev))
2772                 starget_for_each_device(to_scsi_target(dev), &new_state,
2773                                         device_unblock);
2774         else
2775                 device_for_each_child(dev, &new_state, target_unblock);
2776 }
2777 EXPORT_SYMBOL_GPL(scsi_target_unblock);
2778
2779 /**
2780  * scsi_kmap_atomic_sg - find and atomically map an sg-elemnt
2781  * @sgl:        scatter-gather list
2782  * @sg_count:   number of segments in sg
2783  * @offset:     offset in bytes into sg, on return offset into the mapped area
2784  * @len:        bytes to map, on return number of bytes mapped
2785  *
2786  * Returns virtual address of the start of the mapped page
2787  */
2788 void *scsi_kmap_atomic_sg(struct scatterlist *sgl, int sg_count,
2789                           size_t *offset, size_t *len)
2790 {
2791         int i;
2792         size_t sg_len = 0, len_complete = 0;
2793         struct scatterlist *sg;
2794         struct page *page;
2795
2796         WARN_ON(!irqs_disabled());
2797
2798         for_each_sg(sgl, sg, sg_count, i) {
2799                 len_complete = sg_len; /* Complete sg-entries */
2800                 sg_len += sg->length;
2801                 if (sg_len > *offset)
2802                         break;
2803         }
2804
2805         if (unlikely(i == sg_count)) {
2806                 printk(KERN_ERR "%s: Bytes in sg: %zu, requested offset %zu, "
2807                         "elements %d\n",
2808                        __func__, sg_len, *offset, sg_count);
2809                 WARN_ON(1);
2810                 return NULL;
2811         }
2812
2813         /* Offset starting from the beginning of first page in this sg-entry */
2814         *offset = *offset - len_complete + sg->offset;
2815
2816         /* Assumption: contiguous pages can be accessed as "page + i" */
2817         page = nth_page(sg_page(sg), (*offset >> PAGE_SHIFT));
2818         *offset &= ~PAGE_MASK;
2819
2820         /* Bytes in this sg-entry from *offset to the end of the page */
2821         sg_len = PAGE_SIZE - *offset;
2822         if (*len > sg_len)
2823                 *len = sg_len;
2824
2825         return kmap_atomic(page);
2826 }
2827 EXPORT_SYMBOL(scsi_kmap_atomic_sg);
2828
2829 /**
2830  * scsi_kunmap_atomic_sg - atomically unmap a virtual address, previously mapped with scsi_kmap_atomic_sg
2831  * @virt:       virtual address to be unmapped
2832  */
2833 void scsi_kunmap_atomic_sg(void *virt)
2834 {
2835         kunmap_atomic(virt);
2836 }
2837 EXPORT_SYMBOL(scsi_kunmap_atomic_sg);
2838
2839 void sdev_disable_disk_events(struct scsi_device *sdev)
2840 {
2841         atomic_inc(&sdev->disk_events_disable_depth);
2842 }
2843 EXPORT_SYMBOL(sdev_disable_disk_events);
2844
2845 void sdev_enable_disk_events(struct scsi_device *sdev)
2846 {
2847         if (WARN_ON_ONCE(atomic_read(&sdev->disk_events_disable_depth) <= 0))
2848                 return;
2849         atomic_dec(&sdev->disk_events_disable_depth);
2850 }
2851 EXPORT_SYMBOL(sdev_enable_disk_events);
2852
2853 /**
2854  * scsi_vpd_lun_id - return a unique device identification
2855  * @sdev: SCSI device
2856  * @id:   buffer for the identification
2857  * @id_len:  length of the buffer
2858  *
2859  * Copies a unique device identification into @id based
2860  * on the information in the VPD page 0x83 of the device.
2861  * The string will be formatted as a SCSI name string.
2862  *
2863  * Returns the length of the identification or error on failure.
2864  * If the identifier is longer than the supplied buffer the actual
2865  * identifier length is returned and the buffer is not zero-padded.
2866  */
2867 int scsi_vpd_lun_id(struct scsi_device *sdev, char *id, size_t id_len)
2868 {
2869         u8 cur_id_type = 0xff;
2870         u8 cur_id_size = 0;
2871         const unsigned char *d, *cur_id_str;
2872         const struct scsi_vpd *vpd_pg83;
2873         int id_size = -EINVAL;
2874
2875         rcu_read_lock();
2876         vpd_pg83 = rcu_dereference(sdev->vpd_pg83);
2877         if (!vpd_pg83) {
2878                 rcu_read_unlock();
2879                 return -ENXIO;
2880         }
2881
2882         /*
2883          * Look for the correct descriptor.
2884          * Order of preference for lun descriptor:
2885          * - SCSI name string
2886          * - NAA IEEE Registered Extended
2887          * - EUI-64 based 16-byte
2888          * - EUI-64 based 12-byte
2889          * - NAA IEEE Registered
2890          * - NAA IEEE Extended
2891          * - T10 Vendor ID
2892          * as longer descriptors reduce the likelyhood
2893          * of identification clashes.
2894          */
2895
2896         /* The id string must be at least 20 bytes + terminating NULL byte */
2897         if (id_len < 21) {
2898                 rcu_read_unlock();
2899                 return -EINVAL;
2900         }
2901
2902         memset(id, 0, id_len);
2903         d = vpd_pg83->data + 4;
2904         while (d < vpd_pg83->data + vpd_pg83->len) {
2905                 /* Skip designators not referring to the LUN */
2906                 if ((d[1] & 0x30) != 0x00)
2907                         goto next_desig;
2908
2909                 switch (d[1] & 0xf) {
2910                 case 0x1:
2911                         /* T10 Vendor ID */
2912                         if (cur_id_size > d[3])
2913                                 break;
2914                         /* Prefer anything */
2915                         if (cur_id_type > 0x01 && cur_id_type != 0xff)
2916                                 break;
2917                         cur_id_size = d[3];
2918                         if (cur_id_size + 4 > id_len)
2919                                 cur_id_size = id_len - 4;
2920                         cur_id_str = d + 4;
2921                         cur_id_type = d[1] & 0xf;
2922                         id_size = snprintf(id, id_len, "t10.%*pE",
2923                                            cur_id_size, cur_id_str);
2924                         break;
2925                 case 0x2:
2926                         /* EUI-64 */
2927                         if (cur_id_size > d[3])
2928                                 break;
2929                         /* Prefer NAA IEEE Registered Extended */
2930                         if (cur_id_type == 0x3 &&
2931                             cur_id_size == d[3])
2932                                 break;
2933                         cur_id_size = d[3];
2934                         cur_id_str = d + 4;
2935                         cur_id_type = d[1] & 0xf;
2936                         switch (cur_id_size) {
2937                         case 8:
2938                                 id_size = snprintf(id, id_len,
2939                                                    "eui.%8phN",
2940                                                    cur_id_str);
2941                                 break;
2942                         case 12:
2943                                 id_size = snprintf(id, id_len,
2944                                                    "eui.%12phN",
2945                                                    cur_id_str);
2946                                 break;
2947                         case 16:
2948                                 id_size = snprintf(id, id_len,
2949                                                    "eui.%16phN",
2950                                                    cur_id_str);
2951                                 break;
2952                         default:
2953                                 cur_id_size = 0;
2954                                 break;
2955                         }
2956                         break;
2957                 case 0x3:
2958                         /* NAA */
2959                         if (cur_id_size > d[3])
2960                                 break;
2961                         cur_id_size = d[3];
2962                         cur_id_str = d + 4;
2963                         cur_id_type = d[1] & 0xf;
2964                         switch (cur_id_size) {
2965                         case 8:
2966                                 id_size = snprintf(id, id_len,
2967                                                    "naa.%8phN",
2968                                                    cur_id_str);
2969                                 break;
2970                         case 16:
2971                                 id_size = snprintf(id, id_len,
2972                                                    "naa.%16phN",
2973                                                    cur_id_str);
2974                                 break;
2975                         default:
2976                                 cur_id_size = 0;
2977                                 break;
2978                         }
2979                         break;
2980                 case 0x8:
2981                         /* SCSI name string */
2982                         if (cur_id_size + 4 > d[3])
2983                                 break;
2984                         /* Prefer others for truncated descriptor */
2985                         if (cur_id_size && d[3] > id_len)
2986                                 break;
2987                         cur_id_size = id_size = d[3];
2988                         cur_id_str = d + 4;
2989                         cur_id_type = d[1] & 0xf;
2990                         if (cur_id_size >= id_len)
2991                                 cur_id_size = id_len - 1;
2992                         memcpy(id, cur_id_str, cur_id_size);
2993                         /* Decrease priority for truncated descriptor */
2994                         if (cur_id_size != id_size)
2995                                 cur_id_size = 6;
2996                         break;
2997                 default:
2998                         break;
2999                 }
3000 next_desig:
3001                 d += d[3] + 4;
3002         }
3003         rcu_read_unlock();
3004
3005         return id_size;
3006 }
3007 EXPORT_SYMBOL(scsi_vpd_lun_id);
3008
3009 /*
3010  * scsi_vpd_tpg_id - return a target port group identifier
3011  * @sdev: SCSI device
3012  *
3013  * Returns the Target Port Group identifier from the information
3014  * froom VPD page 0x83 of the device.
3015  *
3016  * Returns the identifier or error on failure.
3017  */
3018 int scsi_vpd_tpg_id(struct scsi_device *sdev, int *rel_id)
3019 {
3020         const unsigned char *d;
3021         const struct scsi_vpd *vpd_pg83;
3022         int group_id = -EAGAIN, rel_port = -1;
3023
3024         rcu_read_lock();
3025         vpd_pg83 = rcu_dereference(sdev->vpd_pg83);
3026         if (!vpd_pg83) {
3027                 rcu_read_unlock();
3028                 return -ENXIO;
3029         }
3030
3031         d = vpd_pg83->data + 4;
3032         while (d < vpd_pg83->data + vpd_pg83->len) {
3033                 switch (d[1] & 0xf) {
3034                 case 0x4:
3035                         /* Relative target port */
3036                         rel_port = get_unaligned_be16(&d[6]);
3037                         break;
3038                 case 0x5:
3039                         /* Target port group */
3040                         group_id = get_unaligned_be16(&d[6]);
3041                         break;
3042                 default:
3043                         break;
3044                 }
3045                 d += d[3] + 4;
3046         }
3047         rcu_read_unlock();
3048
3049         if (group_id >= 0 && rel_id && rel_port != -1)
3050                 *rel_id = rel_port;
3051
3052         return group_id;
3053 }
3054 EXPORT_SYMBOL(scsi_vpd_tpg_id);