scsi: stop setting up request->special
[linux-2.6-microblaze.git] / drivers / scsi / scsi_lib.c
1 /*
2  * Copyright (C) 1999 Eric Youngdale
3  * Copyright (C) 2014 Christoph Hellwig
4  *
5  *  SCSI queueing library.
6  *      Initial versions: Eric Youngdale (eric@andante.org).
7  *                        Based upon conversations with large numbers
8  *                        of people at Linux Expo.
9  */
10
11 #include <linux/bio.h>
12 #include <linux/bitops.h>
13 #include <linux/blkdev.h>
14 #include <linux/completion.h>
15 #include <linux/kernel.h>
16 #include <linux/export.h>
17 #include <linux/init.h>
18 #include <linux/pci.h>
19 #include <linux/delay.h>
20 #include <linux/hardirq.h>
21 #include <linux/scatterlist.h>
22 #include <linux/blk-mq.h>
23 #include <linux/ratelimit.h>
24 #include <asm/unaligned.h>
25
26 #include <scsi/scsi.h>
27 #include <scsi/scsi_cmnd.h>
28 #include <scsi/scsi_dbg.h>
29 #include <scsi/scsi_device.h>
30 #include <scsi/scsi_driver.h>
31 #include <scsi/scsi_eh.h>
32 #include <scsi/scsi_host.h>
33 #include <scsi/scsi_transport.h> /* __scsi_init_queue() */
34 #include <scsi/scsi_dh.h>
35
36 #include <trace/events/scsi.h>
37
38 #include "scsi_debugfs.h"
39 #include "scsi_priv.h"
40 #include "scsi_logging.h"
41
42 static struct kmem_cache *scsi_sdb_cache;
43 static struct kmem_cache *scsi_sense_cache;
44 static struct kmem_cache *scsi_sense_isadma_cache;
45 static DEFINE_MUTEX(scsi_sense_cache_mutex);
46
47 static void scsi_mq_uninit_cmd(struct scsi_cmnd *cmd);
48
49 static inline struct kmem_cache *
50 scsi_select_sense_cache(bool unchecked_isa_dma)
51 {
52         return unchecked_isa_dma ? scsi_sense_isadma_cache : scsi_sense_cache;
53 }
54
55 static void scsi_free_sense_buffer(bool unchecked_isa_dma,
56                                    unsigned char *sense_buffer)
57 {
58         kmem_cache_free(scsi_select_sense_cache(unchecked_isa_dma),
59                         sense_buffer);
60 }
61
62 static unsigned char *scsi_alloc_sense_buffer(bool unchecked_isa_dma,
63         gfp_t gfp_mask, int numa_node)
64 {
65         return kmem_cache_alloc_node(scsi_select_sense_cache(unchecked_isa_dma),
66                                      gfp_mask, numa_node);
67 }
68
69 int scsi_init_sense_cache(struct Scsi_Host *shost)
70 {
71         struct kmem_cache *cache;
72         int ret = 0;
73
74         cache = scsi_select_sense_cache(shost->unchecked_isa_dma);
75         if (cache)
76                 return 0;
77
78         mutex_lock(&scsi_sense_cache_mutex);
79         if (shost->unchecked_isa_dma) {
80                 scsi_sense_isadma_cache =
81                         kmem_cache_create("scsi_sense_cache(DMA)",
82                                 SCSI_SENSE_BUFFERSIZE, 0,
83                                 SLAB_HWCACHE_ALIGN | SLAB_CACHE_DMA, NULL);
84                 if (!scsi_sense_isadma_cache)
85                         ret = -ENOMEM;
86         } else {
87                 scsi_sense_cache =
88                         kmem_cache_create_usercopy("scsi_sense_cache",
89                                 SCSI_SENSE_BUFFERSIZE, 0, SLAB_HWCACHE_ALIGN,
90                                 0, SCSI_SENSE_BUFFERSIZE, NULL);
91                 if (!scsi_sense_cache)
92                         ret = -ENOMEM;
93         }
94
95         mutex_unlock(&scsi_sense_cache_mutex);
96         return ret;
97 }
98
99 /*
100  * When to reinvoke queueing after a resource shortage. It's 3 msecs to
101  * not change behaviour from the previous unplug mechanism, experimentation
102  * may prove this needs changing.
103  */
104 #define SCSI_QUEUE_DELAY        3
105
106 static void
107 scsi_set_blocked(struct scsi_cmnd *cmd, int reason)
108 {
109         struct Scsi_Host *host = cmd->device->host;
110         struct scsi_device *device = cmd->device;
111         struct scsi_target *starget = scsi_target(device);
112
113         /*
114          * Set the appropriate busy bit for the device/host.
115          *
116          * If the host/device isn't busy, assume that something actually
117          * completed, and that we should be able to queue a command now.
118          *
119          * Note that the prior mid-layer assumption that any host could
120          * always queue at least one command is now broken.  The mid-layer
121          * will implement a user specifiable stall (see
122          * scsi_host.max_host_blocked and scsi_device.max_device_blocked)
123          * if a command is requeued with no other commands outstanding
124          * either for the device or for the host.
125          */
126         switch (reason) {
127         case SCSI_MLQUEUE_HOST_BUSY:
128                 atomic_set(&host->host_blocked, host->max_host_blocked);
129                 break;
130         case SCSI_MLQUEUE_DEVICE_BUSY:
131         case SCSI_MLQUEUE_EH_RETRY:
132                 atomic_set(&device->device_blocked,
133                            device->max_device_blocked);
134                 break;
135         case SCSI_MLQUEUE_TARGET_BUSY:
136                 atomic_set(&starget->target_blocked,
137                            starget->max_target_blocked);
138                 break;
139         }
140 }
141
142 static void scsi_mq_requeue_cmd(struct scsi_cmnd *cmd)
143 {
144         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         cmd->serial_number = 0;
320         scsi_set_resid(cmd, 0);
321         memset(cmd->sense_buffer, 0, SCSI_SENSE_BUFFERSIZE);
322         if (cmd->cmd_len == 0)
323                 cmd->cmd_len = scsi_command_size(cmd->cmnd);
324 }
325
326 /*
327  * Decrement the host_busy counter and wake up the error handler if necessary.
328  * Avoid as follows that the error handler is not woken up if shost->host_busy
329  * == shost->host_failed: use call_rcu() in scsi_eh_scmd_add() in combination
330  * with an RCU read lock in this function to ensure that this function in its
331  * entirety either finishes before scsi_eh_scmd_add() increases the
332  * host_failed counter or that it notices the shost state change made by
333  * scsi_eh_scmd_add().
334  */
335 static void scsi_dec_host_busy(struct Scsi_Host *shost)
336 {
337         unsigned long flags;
338
339         rcu_read_lock();
340         atomic_dec(&shost->host_busy);
341         if (unlikely(scsi_host_in_recovery(shost))) {
342                 spin_lock_irqsave(shost->host_lock, flags);
343                 if (shost->host_failed || shost->host_eh_scheduled)
344                         scsi_eh_wakeup(shost);
345                 spin_unlock_irqrestore(shost->host_lock, flags);
346         }
347         rcu_read_unlock();
348 }
349
350 void scsi_device_unbusy(struct scsi_device *sdev)
351 {
352         struct Scsi_Host *shost = sdev->host;
353         struct scsi_target *starget = scsi_target(sdev);
354
355         scsi_dec_host_busy(shost);
356
357         if (starget->can_queue > 0)
358                 atomic_dec(&starget->target_busy);
359
360         atomic_dec(&sdev->device_busy);
361 }
362
363 static void scsi_kick_queue(struct request_queue *q)
364 {
365         blk_mq_run_hw_queues(q, false);
366 }
367
368 /*
369  * Called for single_lun devices on IO completion. Clear starget_sdev_user,
370  * and call blk_run_queue for all the scsi_devices on the target -
371  * including current_sdev first.
372  *
373  * Called with *no* scsi locks held.
374  */
375 static void scsi_single_lun_run(struct scsi_device *current_sdev)
376 {
377         struct Scsi_Host *shost = current_sdev->host;
378         struct scsi_device *sdev, *tmp;
379         struct scsi_target *starget = scsi_target(current_sdev);
380         unsigned long flags;
381
382         spin_lock_irqsave(shost->host_lock, flags);
383         starget->starget_sdev_user = NULL;
384         spin_unlock_irqrestore(shost->host_lock, flags);
385
386         /*
387          * Call blk_run_queue for all LUNs on the target, starting with
388          * current_sdev. We race with others (to set starget_sdev_user),
389          * but in most cases, we will be first. Ideally, each LU on the
390          * target would get some limited time or requests on the target.
391          */
392         scsi_kick_queue(current_sdev->request_queue);
393
394         spin_lock_irqsave(shost->host_lock, flags);
395         if (starget->starget_sdev_user)
396                 goto out;
397         list_for_each_entry_safe(sdev, tmp, &starget->devices,
398                         same_target_siblings) {
399                 if (sdev == current_sdev)
400                         continue;
401                 if (scsi_device_get(sdev))
402                         continue;
403
404                 spin_unlock_irqrestore(shost->host_lock, flags);
405                 scsi_kick_queue(sdev->request_queue);
406                 spin_lock_irqsave(shost->host_lock, flags);
407         
408                 scsi_device_put(sdev);
409         }
410  out:
411         spin_unlock_irqrestore(shost->host_lock, flags);
412 }
413
414 static inline bool scsi_device_is_busy(struct scsi_device *sdev)
415 {
416         if (atomic_read(&sdev->device_busy) >= sdev->queue_depth)
417                 return true;
418         if (atomic_read(&sdev->device_blocked) > 0)
419                 return true;
420         return false;
421 }
422
423 static inline bool scsi_target_is_busy(struct scsi_target *starget)
424 {
425         if (starget->can_queue > 0) {
426                 if (atomic_read(&starget->target_busy) >= starget->can_queue)
427                         return true;
428                 if (atomic_read(&starget->target_blocked) > 0)
429                         return true;
430         }
431         return false;
432 }
433
434 static inline bool scsi_host_is_busy(struct Scsi_Host *shost)
435 {
436         if (shost->can_queue > 0 &&
437             atomic_read(&shost->host_busy) >= shost->can_queue)
438                 return true;
439         if (atomic_read(&shost->host_blocked) > 0)
440                 return true;
441         if (shost->host_self_blocked)
442                 return true;
443         return false;
444 }
445
446 static void scsi_starved_list_run(struct Scsi_Host *shost)
447 {
448         LIST_HEAD(starved_list);
449         struct scsi_device *sdev;
450         unsigned long flags;
451
452         spin_lock_irqsave(shost->host_lock, flags);
453         list_splice_init(&shost->starved_list, &starved_list);
454
455         while (!list_empty(&starved_list)) {
456                 struct request_queue *slq;
457
458                 /*
459                  * As long as shost is accepting commands and we have
460                  * starved queues, call blk_run_queue. scsi_request_fn
461                  * drops the queue_lock and can add us back to the
462                  * starved_list.
463                  *
464                  * host_lock protects the starved_list and starved_entry.
465                  * scsi_request_fn must get the host_lock before checking
466                  * or modifying starved_list or starved_entry.
467                  */
468                 if (scsi_host_is_busy(shost))
469                         break;
470
471                 sdev = list_entry(starved_list.next,
472                                   struct scsi_device, starved_entry);
473                 list_del_init(&sdev->starved_entry);
474                 if (scsi_target_is_busy(scsi_target(sdev))) {
475                         list_move_tail(&sdev->starved_entry,
476                                        &shost->starved_list);
477                         continue;
478                 }
479
480                 /*
481                  * Once we drop the host lock, a racing scsi_remove_device()
482                  * call may remove the sdev from the starved list and destroy
483                  * it and the queue.  Mitigate by taking a reference to the
484                  * queue and never touching the sdev again after we drop the
485                  * host lock.  Note: if __scsi_remove_device() invokes
486                  * blk_cleanup_queue() before the queue is run from this
487                  * function then blk_run_queue() will return immediately since
488                  * blk_cleanup_queue() marks the queue with QUEUE_FLAG_DYING.
489                  */
490                 slq = sdev->request_queue;
491                 if (!blk_get_queue(slq))
492                         continue;
493                 spin_unlock_irqrestore(shost->host_lock, flags);
494
495                 scsi_kick_queue(slq);
496                 blk_put_queue(slq);
497
498                 spin_lock_irqsave(shost->host_lock, flags);
499         }
500         /* put any unprocessed entries back */
501         list_splice(&starved_list, &shost->starved_list);
502         spin_unlock_irqrestore(shost->host_lock, flags);
503 }
504
505 /*
506  * Function:   scsi_run_queue()
507  *
508  * Purpose:    Select a proper request queue to serve next
509  *
510  * Arguments:  q       - last request's queue
511  *
512  * Returns:     Nothing
513  *
514  * Notes:      The previous command was completely finished, start
515  *             a new one if possible.
516  */
517 static void scsi_run_queue(struct request_queue *q)
518 {
519         struct scsi_device *sdev = q->queuedata;
520
521         if (scsi_target(sdev)->single_lun)
522                 scsi_single_lun_run(sdev);
523         if (!list_empty(&sdev->host->starved_list))
524                 scsi_starved_list_run(sdev->host);
525
526         blk_mq_run_hw_queues(q, false);
527 }
528
529 void scsi_requeue_run_queue(struct work_struct *work)
530 {
531         struct scsi_device *sdev;
532         struct request_queue *q;
533
534         sdev = container_of(work, struct scsi_device, requeue_work);
535         q = sdev->request_queue;
536         scsi_run_queue(q);
537 }
538
539 void scsi_run_host_queues(struct Scsi_Host *shost)
540 {
541         struct scsi_device *sdev;
542
543         shost_for_each_device(sdev, shost)
544                 scsi_run_queue(sdev->request_queue);
545 }
546
547 static void scsi_uninit_cmd(struct scsi_cmnd *cmd)
548 {
549         if (!blk_rq_is_passthrough(cmd->request)) {
550                 struct scsi_driver *drv = scsi_cmd_to_driver(cmd);
551
552                 if (drv->uninit_command)
553                         drv->uninit_command(cmd);
554         }
555 }
556
557 static void scsi_mq_free_sgtables(struct scsi_cmnd *cmd)
558 {
559         if (cmd->sdb.table.nents)
560                 sg_free_table_chained(&cmd->sdb.table, true);
561         if (scsi_prot_sg_count(cmd))
562                 sg_free_table_chained(&cmd->prot_sdb->table, true);
563 }
564
565 static void scsi_mq_uninit_cmd(struct scsi_cmnd *cmd)
566 {
567         scsi_mq_free_sgtables(cmd);
568         scsi_uninit_cmd(cmd);
569         scsi_del_cmd_from_list(cmd);
570 }
571
572 /* Returns false when no more bytes to process, true if there are more */
573 static bool scsi_end_request(struct request *req, blk_status_t error,
574                 unsigned int bytes)
575 {
576         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
577         struct scsi_device *sdev = cmd->device;
578         struct request_queue *q = sdev->request_queue;
579
580         if (blk_update_request(req, error, bytes))
581                 return true;
582
583         if (blk_queue_add_random(q))
584                 add_disk_randomness(req->rq_disk);
585
586         if (!blk_rq_is_scsi(req)) {
587                 WARN_ON_ONCE(!(cmd->flags & SCMD_INITIALIZED));
588                 cmd->flags &= ~SCMD_INITIALIZED;
589                 destroy_rcu_head(&cmd->rcu);
590         }
591
592         /*
593          * In the MQ case the command gets freed by __blk_mq_end_request,
594          * so we have to do all cleanup that depends on it earlier.
595          *
596          * We also can't kick the queues from irq context, so we
597          * will have to defer it to a workqueue.
598          */
599         scsi_mq_uninit_cmd(cmd);
600
601         /*
602          * queue is still alive, so grab the ref for preventing it
603          * from being cleaned up during running queue.
604          */
605         percpu_ref_get(&q->q_usage_counter);
606
607         __blk_mq_end_request(req, error);
608
609         if (scsi_target(sdev)->single_lun ||
610             !list_empty(&sdev->host->starved_list))
611                 kblockd_schedule_work(&sdev->requeue_work);
612         else
613                 blk_mq_run_hw_queues(q, true);
614
615         percpu_ref_put(&q->q_usage_counter);
616         put_device(&sdev->sdev_gendev);
617         return false;
618 }
619
620 /**
621  * scsi_result_to_blk_status - translate a SCSI result code into blk_status_t
622  * @cmd:        SCSI command
623  * @result:     scsi error code
624  *
625  * Translate a SCSI result code into a blk_status_t value. May reset the host
626  * byte of @cmd->result.
627  */
628 static blk_status_t scsi_result_to_blk_status(struct scsi_cmnd *cmd, int result)
629 {
630         switch (host_byte(result)) {
631         case DID_OK:
632                 /*
633                  * Also check the other bytes than the status byte in result
634                  * to handle the case when a SCSI LLD sets result to
635                  * DRIVER_SENSE << 24 without setting SAM_STAT_CHECK_CONDITION.
636                  */
637                 if (scsi_status_is_good(result) && (result & ~0xff) == 0)
638                         return BLK_STS_OK;
639                 return BLK_STS_IOERR;
640         case DID_TRANSPORT_FAILFAST:
641                 return BLK_STS_TRANSPORT;
642         case DID_TARGET_FAILURE:
643                 set_host_byte(cmd, DID_OK);
644                 return BLK_STS_TARGET;
645         case DID_NEXUS_FAILURE:
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                 scsi_req(req)->resid_len = scsi_get_resid(cmd);
942         }
943
944         /*
945          * Next deal with any sectors which we were able to correctly
946          * handle.
947          */
948         SCSI_LOG_HLCOMPLETE(1, scmd_printk(KERN_INFO, cmd,
949                 "%u sectors total, %d bytes done.\n",
950                 blk_rq_sectors(req), good_bytes));
951
952         /*
953          * Next deal with any sectors which we were able to correctly
954          * handle. Failed, zero length commands always need to drop down
955          * to retry code. Fast path should return in this block.
956          */
957         if (likely(blk_rq_bytes(req) > 0 || blk_stat == BLK_STS_OK)) {
958                 if (likely(!scsi_end_request(req, blk_stat, good_bytes)))
959                         return; /* no bytes remaining */
960         }
961
962         /* Kill remainder if no retries. */
963         if (unlikely(blk_stat && scsi_noretry_cmd(cmd))) {
964                 if (scsi_end_request(req, blk_stat, blk_rq_bytes(req)))
965                         WARN_ONCE(true,
966                             "Bytes remaining after failed, no-retry command");
967                 return;
968         }
969
970         /*
971          * If there had been no error, but we have leftover bytes in the
972          * requeues just queue the command up again.
973          */
974         if (likely(result == 0))
975                 scsi_io_completion_reprep(cmd, q);
976         else
977                 scsi_io_completion_action(cmd, result);
978 }
979
980 static blk_status_t scsi_init_sgtable(struct request *req,
981                 struct scsi_data_buffer *sdb)
982 {
983         int count;
984
985         /*
986          * If sg table allocation fails, requeue request later.
987          */
988         if (unlikely(sg_alloc_table_chained(&sdb->table,
989                         blk_rq_nr_phys_segments(req), sdb->table.sgl)))
990                 return BLK_STS_RESOURCE;
991
992         /* 
993          * Next, walk the list, and fill in the addresses and sizes of
994          * each segment.
995          */
996         count = blk_rq_map_sg(req->q, req, sdb->table.sgl);
997         BUG_ON(count > sdb->table.nents);
998         sdb->table.nents = count;
999         sdb->length = blk_rq_payload_bytes(req);
1000         return BLK_STS_OK;
1001 }
1002
1003 /*
1004  * Function:    scsi_init_io()
1005  *
1006  * Purpose:     SCSI I/O initialize function.
1007  *
1008  * Arguments:   cmd   - Command descriptor we wish to initialize
1009  *
1010  * Returns:     BLK_STS_OK on success
1011  *              BLK_STS_RESOURCE if the failure is retryable
1012  *              BLK_STS_IOERR if the failure is fatal
1013  */
1014 blk_status_t scsi_init_io(struct scsi_cmnd *cmd)
1015 {
1016         struct request *rq = cmd->request;
1017         blk_status_t ret;
1018
1019         if (WARN_ON_ONCE(!blk_rq_nr_phys_segments(rq)))
1020                 return BLK_STS_IOERR;
1021
1022         ret = scsi_init_sgtable(rq, &cmd->sdb);
1023         if (ret)
1024                 return ret;
1025
1026         if (blk_integrity_rq(rq)) {
1027                 struct scsi_data_buffer *prot_sdb = cmd->prot_sdb;
1028                 int ivecs, count;
1029
1030                 if (WARN_ON_ONCE(!prot_sdb)) {
1031                         /*
1032                          * This can happen if someone (e.g. multipath)
1033                          * queues a command to a device on an adapter
1034                          * that does not support DIX.
1035                          */
1036                         ret = BLK_STS_IOERR;
1037                         goto out_free_sgtables;
1038                 }
1039
1040                 ivecs = blk_rq_count_integrity_sg(rq->q, rq->bio);
1041
1042                 if (sg_alloc_table_chained(&prot_sdb->table, ivecs,
1043                                 prot_sdb->table.sgl)) {
1044                         ret = BLK_STS_RESOURCE;
1045                         goto out_free_sgtables;
1046                 }
1047
1048                 count = blk_rq_map_integrity_sg(rq->q, rq->bio,
1049                                                 prot_sdb->table.sgl);
1050                 BUG_ON(count > ivecs);
1051                 BUG_ON(count > queue_max_integrity_segments(rq->q));
1052
1053                 cmd->prot_sdb = prot_sdb;
1054                 cmd->prot_sdb->table.nents = count;
1055         }
1056
1057         return BLK_STS_OK;
1058 out_free_sgtables:
1059         scsi_mq_free_sgtables(cmd);
1060         return ret;
1061 }
1062 EXPORT_SYMBOL(scsi_init_io);
1063
1064 /**
1065  * scsi_initialize_rq - initialize struct scsi_cmnd partially
1066  * @rq: Request associated with the SCSI command to be initialized.
1067  *
1068  * This function initializes the members of struct scsi_cmnd that must be
1069  * initialized before request processing starts and that won't be
1070  * reinitialized if a SCSI command is requeued.
1071  *
1072  * Called from inside blk_get_request() for pass-through requests and from
1073  * inside scsi_init_command() for filesystem requests.
1074  */
1075 static void scsi_initialize_rq(struct request *rq)
1076 {
1077         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1078
1079         scsi_req_init(&cmd->req);
1080         init_rcu_head(&cmd->rcu);
1081         cmd->jiffies_at_alloc = jiffies;
1082         cmd->retries = 0;
1083 }
1084
1085 /* Add a command to the list used by the aacraid and dpt_i2o drivers */
1086 void scsi_add_cmd_to_list(struct scsi_cmnd *cmd)
1087 {
1088         struct scsi_device *sdev = cmd->device;
1089         struct Scsi_Host *shost = sdev->host;
1090         unsigned long flags;
1091
1092         if (shost->use_cmd_list) {
1093                 spin_lock_irqsave(&sdev->list_lock, flags);
1094                 list_add_tail(&cmd->list, &sdev->cmd_list);
1095                 spin_unlock_irqrestore(&sdev->list_lock, flags);
1096         }
1097 }
1098
1099 /* Remove a command from the list used by the aacraid and dpt_i2o drivers */
1100 void scsi_del_cmd_from_list(struct scsi_cmnd *cmd)
1101 {
1102         struct scsi_device *sdev = cmd->device;
1103         struct Scsi_Host *shost = sdev->host;
1104         unsigned long flags;
1105
1106         if (shost->use_cmd_list) {
1107                 spin_lock_irqsave(&sdev->list_lock, flags);
1108                 BUG_ON(list_empty(&cmd->list));
1109                 list_del_init(&cmd->list);
1110                 spin_unlock_irqrestore(&sdev->list_lock, flags);
1111         }
1112 }
1113
1114 /* Called after a request has been started. */
1115 void scsi_init_command(struct scsi_device *dev, struct scsi_cmnd *cmd)
1116 {
1117         void *buf = cmd->sense_buffer;
1118         void *prot = cmd->prot_sdb;
1119         struct request *rq = blk_mq_rq_from_pdu(cmd);
1120         unsigned int flags = cmd->flags & SCMD_PRESERVED_FLAGS;
1121         unsigned long jiffies_at_alloc;
1122         int retries;
1123
1124         if (!blk_rq_is_scsi(rq) && !(flags & SCMD_INITIALIZED)) {
1125                 flags |= SCMD_INITIALIZED;
1126                 scsi_initialize_rq(rq);
1127         }
1128
1129         jiffies_at_alloc = cmd->jiffies_at_alloc;
1130         retries = cmd->retries;
1131         /* zero out the cmd, except for the embedded scsi_request */
1132         memset((char *)cmd + sizeof(cmd->req), 0,
1133                 sizeof(*cmd) - sizeof(cmd->req) + dev->host->hostt->cmd_size);
1134
1135         cmd->device = dev;
1136         cmd->sense_buffer = buf;
1137         cmd->prot_sdb = prot;
1138         cmd->flags = flags;
1139         INIT_DELAYED_WORK(&cmd->abort_work, scmd_eh_abort_handler);
1140         cmd->jiffies_at_alloc = jiffies_at_alloc;
1141         cmd->retries = retries;
1142
1143         scsi_add_cmd_to_list(cmd);
1144 }
1145
1146 static blk_status_t scsi_setup_scsi_cmnd(struct scsi_device *sdev,
1147                 struct request *req)
1148 {
1149         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1150
1151         /*
1152          * Passthrough requests may transfer data, in which case they must
1153          * a bio attached to them.  Or they might contain a SCSI command
1154          * that does not transfer data, in which case they may optionally
1155          * submit a request without an attached bio.
1156          */
1157         if (req->bio) {
1158                 blk_status_t ret = scsi_init_io(cmd);
1159                 if (unlikely(ret != BLK_STS_OK))
1160                         return ret;
1161         } else {
1162                 BUG_ON(blk_rq_bytes(req));
1163
1164                 memset(&cmd->sdb, 0, sizeof(cmd->sdb));
1165         }
1166
1167         cmd->cmd_len = scsi_req(req)->cmd_len;
1168         cmd->cmnd = scsi_req(req)->cmd;
1169         cmd->transfersize = blk_rq_bytes(req);
1170         cmd->allowed = scsi_req(req)->retries;
1171         return BLK_STS_OK;
1172 }
1173
1174 /*
1175  * Setup a normal block command.  These are simple request from filesystems
1176  * that still need to be translated to SCSI CDBs from the ULD.
1177  */
1178 static blk_status_t scsi_setup_fs_cmnd(struct scsi_device *sdev,
1179                 struct request *req)
1180 {
1181         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1182
1183         if (unlikely(sdev->handler && sdev->handler->prep_fn)) {
1184                 blk_status_t ret = sdev->handler->prep_fn(sdev, req);
1185                 if (ret != BLK_STS_OK)
1186                         return ret;
1187         }
1188
1189         cmd->cmnd = scsi_req(req)->cmd = scsi_req(req)->__cmd;
1190         memset(cmd->cmnd, 0, BLK_MAX_CDB);
1191         return scsi_cmd_to_driver(cmd)->init_command(cmd);
1192 }
1193
1194 static blk_status_t scsi_setup_cmnd(struct scsi_device *sdev,
1195                 struct request *req)
1196 {
1197         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1198
1199         if (!blk_rq_bytes(req))
1200                 cmd->sc_data_direction = DMA_NONE;
1201         else if (rq_data_dir(req) == WRITE)
1202                 cmd->sc_data_direction = DMA_TO_DEVICE;
1203         else
1204                 cmd->sc_data_direction = DMA_FROM_DEVICE;
1205
1206         if (blk_rq_is_scsi(req))
1207                 return scsi_setup_scsi_cmnd(sdev, req);
1208         else
1209                 return scsi_setup_fs_cmnd(sdev, req);
1210 }
1211
1212 static blk_status_t
1213 scsi_prep_state_check(struct scsi_device *sdev, struct request *req)
1214 {
1215         switch (sdev->sdev_state) {
1216         case SDEV_OFFLINE:
1217         case SDEV_TRANSPORT_OFFLINE:
1218                 /*
1219                  * If the device is offline we refuse to process any
1220                  * commands.  The device must be brought online
1221                  * before trying any recovery commands.
1222                  */
1223                 sdev_printk(KERN_ERR, sdev,
1224                             "rejecting I/O to offline device\n");
1225                 return BLK_STS_IOERR;
1226         case SDEV_DEL:
1227                 /*
1228                  * If the device is fully deleted, we refuse to
1229                  * process any commands as well.
1230                  */
1231                 sdev_printk(KERN_ERR, sdev,
1232                             "rejecting I/O to dead device\n");
1233                 return BLK_STS_IOERR;
1234         case SDEV_BLOCK:
1235         case SDEV_CREATED_BLOCK:
1236                 return BLK_STS_RESOURCE;
1237         case SDEV_QUIESCE:
1238                 /*
1239                  * If the devices is blocked we defer normal commands.
1240                  */
1241                 if (req && !(req->rq_flags & RQF_PREEMPT))
1242                         return BLK_STS_RESOURCE;
1243                 return BLK_STS_OK;
1244         default:
1245                 /*
1246                  * For any other not fully online state we only allow
1247                  * special commands.  In particular any user initiated
1248                  * command is not allowed.
1249                  */
1250                 if (req && !(req->rq_flags & RQF_PREEMPT))
1251                         return BLK_STS_IOERR;
1252                 return BLK_STS_OK;
1253         }
1254 }
1255
1256 /*
1257  * scsi_dev_queue_ready: if we can send requests to sdev, return 1 else
1258  * return 0.
1259  *
1260  * Called with the queue_lock held.
1261  */
1262 static inline int scsi_dev_queue_ready(struct request_queue *q,
1263                                   struct scsi_device *sdev)
1264 {
1265         unsigned int busy;
1266
1267         busy = atomic_inc_return(&sdev->device_busy) - 1;
1268         if (atomic_read(&sdev->device_blocked)) {
1269                 if (busy)
1270                         goto out_dec;
1271
1272                 /*
1273                  * unblock after device_blocked iterates to zero
1274                  */
1275                 if (atomic_dec_return(&sdev->device_blocked) > 0)
1276                         goto out_dec;
1277                 SCSI_LOG_MLQUEUE(3, sdev_printk(KERN_INFO, sdev,
1278                                    "unblocking device at zero depth\n"));
1279         }
1280
1281         if (busy >= sdev->queue_depth)
1282                 goto out_dec;
1283
1284         return 1;
1285 out_dec:
1286         atomic_dec(&sdev->device_busy);
1287         return 0;
1288 }
1289
1290 /*
1291  * scsi_target_queue_ready: checks if there we can send commands to target
1292  * @sdev: scsi device on starget to check.
1293  */
1294 static inline int scsi_target_queue_ready(struct Scsi_Host *shost,
1295                                            struct scsi_device *sdev)
1296 {
1297         struct scsi_target *starget = scsi_target(sdev);
1298         unsigned int busy;
1299
1300         if (starget->single_lun) {
1301                 spin_lock_irq(shost->host_lock);
1302                 if (starget->starget_sdev_user &&
1303                     starget->starget_sdev_user != sdev) {
1304                         spin_unlock_irq(shost->host_lock);
1305                         return 0;
1306                 }
1307                 starget->starget_sdev_user = sdev;
1308                 spin_unlock_irq(shost->host_lock);
1309         }
1310
1311         if (starget->can_queue <= 0)
1312                 return 1;
1313
1314         busy = atomic_inc_return(&starget->target_busy) - 1;
1315         if (atomic_read(&starget->target_blocked) > 0) {
1316                 if (busy)
1317                         goto starved;
1318
1319                 /*
1320                  * unblock after target_blocked iterates to zero
1321                  */
1322                 if (atomic_dec_return(&starget->target_blocked) > 0)
1323                         goto out_dec;
1324
1325                 SCSI_LOG_MLQUEUE(3, starget_printk(KERN_INFO, starget,
1326                                  "unblocking target at zero depth\n"));
1327         }
1328
1329         if (busy >= starget->can_queue)
1330                 goto starved;
1331
1332         return 1;
1333
1334 starved:
1335         spin_lock_irq(shost->host_lock);
1336         list_move_tail(&sdev->starved_entry, &shost->starved_list);
1337         spin_unlock_irq(shost->host_lock);
1338 out_dec:
1339         if (starget->can_queue > 0)
1340                 atomic_dec(&starget->target_busy);
1341         return 0;
1342 }
1343
1344 /*
1345  * scsi_host_queue_ready: if we can send requests to shost, return 1 else
1346  * return 0. We must end up running the queue again whenever 0 is
1347  * returned, else IO can hang.
1348  */
1349 static inline int scsi_host_queue_ready(struct request_queue *q,
1350                                    struct Scsi_Host *shost,
1351                                    struct scsi_device *sdev)
1352 {
1353         unsigned int busy;
1354
1355         if (scsi_host_in_recovery(shost))
1356                 return 0;
1357
1358         busy = atomic_inc_return(&shost->host_busy) - 1;
1359         if (atomic_read(&shost->host_blocked) > 0) {
1360                 if (busy)
1361                         goto starved;
1362
1363                 /*
1364                  * unblock after host_blocked iterates to zero
1365                  */
1366                 if (atomic_dec_return(&shost->host_blocked) > 0)
1367                         goto out_dec;
1368
1369                 SCSI_LOG_MLQUEUE(3,
1370                         shost_printk(KERN_INFO, shost,
1371                                      "unblocking host at zero depth\n"));
1372         }
1373
1374         if (shost->can_queue > 0 && busy >= shost->can_queue)
1375                 goto starved;
1376         if (shost->host_self_blocked)
1377                 goto starved;
1378
1379         /* We're OK to process the command, so we can't be starved */
1380         if (!list_empty(&sdev->starved_entry)) {
1381                 spin_lock_irq(shost->host_lock);
1382                 if (!list_empty(&sdev->starved_entry))
1383                         list_del_init(&sdev->starved_entry);
1384                 spin_unlock_irq(shost->host_lock);
1385         }
1386
1387         return 1;
1388
1389 starved:
1390         spin_lock_irq(shost->host_lock);
1391         if (list_empty(&sdev->starved_entry))
1392                 list_add_tail(&sdev->starved_entry, &shost->starved_list);
1393         spin_unlock_irq(shost->host_lock);
1394 out_dec:
1395         scsi_dec_host_busy(shost);
1396         return 0;
1397 }
1398
1399 /*
1400  * Busy state exporting function for request stacking drivers.
1401  *
1402  * For efficiency, no lock is taken to check the busy state of
1403  * shost/starget/sdev, since the returned value is not guaranteed and
1404  * may be changed after request stacking drivers call the function,
1405  * regardless of taking lock or not.
1406  *
1407  * When scsi can't dispatch I/Os anymore and needs to kill I/Os scsi
1408  * needs to return 'not busy'. Otherwise, request stacking drivers
1409  * may hold requests forever.
1410  */
1411 static bool scsi_mq_lld_busy(struct request_queue *q)
1412 {
1413         struct scsi_device *sdev = q->queuedata;
1414         struct Scsi_Host *shost;
1415
1416         if (blk_queue_dying(q))
1417                 return false;
1418
1419         shost = sdev->host;
1420
1421         /*
1422          * Ignore host/starget busy state.
1423          * Since block layer does not have a concept of fairness across
1424          * multiple queues, congestion of host/starget needs to be handled
1425          * in SCSI layer.
1426          */
1427         if (scsi_host_in_recovery(shost) || scsi_device_is_busy(sdev))
1428                 return true;
1429
1430         return false;
1431 }
1432
1433 static void scsi_softirq_done(struct request *rq)
1434 {
1435         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1436         unsigned long wait_for = (cmd->allowed + 1) * rq->timeout;
1437         int disposition;
1438
1439         INIT_LIST_HEAD(&cmd->eh_entry);
1440
1441         atomic_inc(&cmd->device->iodone_cnt);
1442         if (cmd->result)
1443                 atomic_inc(&cmd->device->ioerr_cnt);
1444
1445         disposition = scsi_decide_disposition(cmd);
1446         if (disposition != SUCCESS &&
1447             time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) {
1448                 sdev_printk(KERN_ERR, cmd->device,
1449                             "timing out command, waited %lus\n",
1450                             wait_for/HZ);
1451                 disposition = SUCCESS;
1452         }
1453
1454         scsi_log_completion(cmd, disposition);
1455
1456         switch (disposition) {
1457                 case SUCCESS:
1458                         scsi_finish_command(cmd);
1459                         break;
1460                 case NEEDS_RETRY:
1461                         scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY);
1462                         break;
1463                 case ADD_TO_MLQUEUE:
1464                         scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY);
1465                         break;
1466                 default:
1467                         scsi_eh_scmd_add(cmd);
1468                         break;
1469         }
1470 }
1471
1472 /**
1473  * scsi_dispatch_command - Dispatch a command to the low-level driver.
1474  * @cmd: command block we are dispatching.
1475  *
1476  * Return: nonzero return request was rejected and device's queue needs to be
1477  * plugged.
1478  */
1479 static int scsi_dispatch_cmd(struct scsi_cmnd *cmd)
1480 {
1481         struct Scsi_Host *host = cmd->device->host;
1482         int rtn = 0;
1483
1484         atomic_inc(&cmd->device->iorequest_cnt);
1485
1486         /* check if the device is still usable */
1487         if (unlikely(cmd->device->sdev_state == SDEV_DEL)) {
1488                 /* in SDEV_DEL we error all commands. DID_NO_CONNECT
1489                  * returns an immediate error upwards, and signals
1490                  * that the device is no longer present */
1491                 cmd->result = DID_NO_CONNECT << 16;
1492                 goto done;
1493         }
1494
1495         /* Check to see if the scsi lld made this device blocked. */
1496         if (unlikely(scsi_device_blocked(cmd->device))) {
1497                 /*
1498                  * in blocked state, the command is just put back on
1499                  * the device queue.  The suspend state has already
1500                  * blocked the queue so future requests should not
1501                  * occur until the device transitions out of the
1502                  * suspend state.
1503                  */
1504                 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1505                         "queuecommand : device blocked\n"));
1506                 return SCSI_MLQUEUE_DEVICE_BUSY;
1507         }
1508
1509         /* Store the LUN value in cmnd, if needed. */
1510         if (cmd->device->lun_in_cdb)
1511                 cmd->cmnd[1] = (cmd->cmnd[1] & 0x1f) |
1512                                (cmd->device->lun << 5 & 0xe0);
1513
1514         scsi_log_send(cmd);
1515
1516         /*
1517          * Before we queue this command, check if the command
1518          * length exceeds what the host adapter can handle.
1519          */
1520         if (cmd->cmd_len > cmd->device->host->max_cmd_len) {
1521                 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1522                                "queuecommand : command too long. "
1523                                "cdb_size=%d host->max_cmd_len=%d\n",
1524                                cmd->cmd_len, cmd->device->host->max_cmd_len));
1525                 cmd->result = (DID_ABORT << 16);
1526                 goto done;
1527         }
1528
1529         if (unlikely(host->shost_state == SHOST_DEL)) {
1530                 cmd->result = (DID_NO_CONNECT << 16);
1531                 goto done;
1532
1533         }
1534
1535         trace_scsi_dispatch_cmd_start(cmd);
1536         rtn = host->hostt->queuecommand(host, cmd);
1537         if (rtn) {
1538                 trace_scsi_dispatch_cmd_error(cmd, rtn);
1539                 if (rtn != SCSI_MLQUEUE_DEVICE_BUSY &&
1540                     rtn != SCSI_MLQUEUE_TARGET_BUSY)
1541                         rtn = SCSI_MLQUEUE_HOST_BUSY;
1542
1543                 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1544                         "queuecommand : request rejected\n"));
1545         }
1546
1547         return rtn;
1548  done:
1549         cmd->scsi_done(cmd);
1550         return 0;
1551 }
1552
1553 /* Size in bytes of the sg-list stored in the scsi-mq command-private data. */
1554 static unsigned int scsi_mq_sgl_size(struct Scsi_Host *shost)
1555 {
1556         return min_t(unsigned int, shost->sg_tablesize, SG_CHUNK_SIZE) *
1557                 sizeof(struct scatterlist);
1558 }
1559
1560 static blk_status_t scsi_mq_prep_fn(struct request *req)
1561 {
1562         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1563         struct scsi_device *sdev = req->q->queuedata;
1564         struct Scsi_Host *shost = sdev->host;
1565         struct scatterlist *sg;
1566
1567         scsi_init_command(sdev, cmd);
1568
1569         cmd->request = req;
1570         cmd->tag = req->tag;
1571         cmd->prot_op = SCSI_PROT_NORMAL;
1572
1573         sg = (void *)cmd + sizeof(struct scsi_cmnd) + shost->hostt->cmd_size;
1574         cmd->sdb.table.sgl = sg;
1575
1576         if (scsi_host_get_prot(shost)) {
1577                 memset(cmd->prot_sdb, 0, sizeof(struct scsi_data_buffer));
1578
1579                 cmd->prot_sdb->table.sgl =
1580                         (struct scatterlist *)(cmd->prot_sdb + 1);
1581         }
1582
1583         blk_mq_start_request(req);
1584
1585         return scsi_setup_cmnd(sdev, req);
1586 }
1587
1588 static void scsi_mq_done(struct scsi_cmnd *cmd)
1589 {
1590         if (unlikely(test_and_set_bit(SCMD_STATE_COMPLETE, &cmd->state)))
1591                 return;
1592         trace_scsi_dispatch_cmd_done(cmd);
1593
1594         /*
1595          * If the block layer didn't complete the request due to a timeout
1596          * injection, scsi must clear its internal completed state so that the
1597          * timeout handler will see it needs to escalate its own error
1598          * recovery.
1599          */
1600         if (unlikely(!blk_mq_complete_request(cmd->request)))
1601                 clear_bit(SCMD_STATE_COMPLETE, &cmd->state);
1602 }
1603
1604 static void scsi_mq_put_budget(struct blk_mq_hw_ctx *hctx)
1605 {
1606         struct request_queue *q = hctx->queue;
1607         struct scsi_device *sdev = q->queuedata;
1608
1609         atomic_dec(&sdev->device_busy);
1610         put_device(&sdev->sdev_gendev);
1611 }
1612
1613 static bool scsi_mq_get_budget(struct blk_mq_hw_ctx *hctx)
1614 {
1615         struct request_queue *q = hctx->queue;
1616         struct scsi_device *sdev = q->queuedata;
1617
1618         if (!get_device(&sdev->sdev_gendev))
1619                 goto out;
1620         if (!scsi_dev_queue_ready(q, sdev))
1621                 goto out_put_device;
1622
1623         return true;
1624
1625 out_put_device:
1626         put_device(&sdev->sdev_gendev);
1627 out:
1628         if (atomic_read(&sdev->device_busy) == 0 && !scsi_device_blocked(sdev))
1629                 blk_mq_delay_run_hw_queue(hctx, SCSI_QUEUE_DELAY);
1630         return false;
1631 }
1632
1633 static blk_status_t scsi_queue_rq(struct blk_mq_hw_ctx *hctx,
1634                          const struct blk_mq_queue_data *bd)
1635 {
1636         struct request *req = bd->rq;
1637         struct request_queue *q = req->q;
1638         struct scsi_device *sdev = q->queuedata;
1639         struct Scsi_Host *shost = sdev->host;
1640         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1641         blk_status_t ret;
1642         int reason;
1643
1644         /*
1645          * If the device is not in running state we will reject some or all
1646          * commands.
1647          */
1648         if (unlikely(sdev->sdev_state != SDEV_RUNNING)) {
1649                 ret = scsi_prep_state_check(sdev, req);
1650                 if (ret != BLK_STS_OK)
1651                         goto out_put_budget;
1652         }
1653
1654         ret = BLK_STS_RESOURCE;
1655         if (!scsi_target_queue_ready(shost, sdev))
1656                 goto out_put_budget;
1657         if (!scsi_host_queue_ready(q, shost, sdev))
1658                 goto out_dec_target_busy;
1659
1660         if (!(req->rq_flags & RQF_DONTPREP)) {
1661                 ret = scsi_mq_prep_fn(req);
1662                 if (ret != BLK_STS_OK)
1663                         goto out_dec_host_busy;
1664                 req->rq_flags |= RQF_DONTPREP;
1665         } else {
1666                 clear_bit(SCMD_STATE_COMPLETE, &cmd->state);
1667                 blk_mq_start_request(req);
1668         }
1669
1670         if (sdev->simple_tags)
1671                 cmd->flags |= SCMD_TAGGED;
1672         else
1673                 cmd->flags &= ~SCMD_TAGGED;
1674
1675         scsi_init_cmd_errh(cmd);
1676         cmd->scsi_done = scsi_mq_done;
1677
1678         reason = scsi_dispatch_cmd(cmd);
1679         if (reason) {
1680                 scsi_set_blocked(cmd, reason);
1681                 ret = BLK_STS_RESOURCE;
1682                 goto out_dec_host_busy;
1683         }
1684
1685         return BLK_STS_OK;
1686
1687 out_dec_host_busy:
1688         scsi_dec_host_busy(shost);
1689 out_dec_target_busy:
1690         if (scsi_target(sdev)->can_queue > 0)
1691                 atomic_dec(&scsi_target(sdev)->target_busy);
1692 out_put_budget:
1693         scsi_mq_put_budget(hctx);
1694         switch (ret) {
1695         case BLK_STS_OK:
1696                 break;
1697         case BLK_STS_RESOURCE:
1698                 if (atomic_read(&sdev->device_busy) ||
1699                     scsi_device_blocked(sdev))
1700                         ret = BLK_STS_DEV_RESOURCE;
1701                 break;
1702         default:
1703                 /*
1704                  * Make sure to release all allocated ressources when
1705                  * we hit an error, as we will never see this command
1706                  * again.
1707                  */
1708                 if (req->rq_flags & RQF_DONTPREP)
1709                         scsi_mq_uninit_cmd(cmd);
1710                 break;
1711         }
1712         return ret;
1713 }
1714
1715 static enum blk_eh_timer_return scsi_timeout(struct request *req,
1716                 bool reserved)
1717 {
1718         if (reserved)
1719                 return BLK_EH_RESET_TIMER;
1720         return scsi_times_out(req);
1721 }
1722
1723 static int scsi_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
1724                                 unsigned int hctx_idx, unsigned int numa_node)
1725 {
1726         struct Scsi_Host *shost = set->driver_data;
1727         const bool unchecked_isa_dma = shost->unchecked_isa_dma;
1728         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1729         struct scatterlist *sg;
1730
1731         if (unchecked_isa_dma)
1732                 cmd->flags |= SCMD_UNCHECKED_ISA_DMA;
1733         cmd->sense_buffer = scsi_alloc_sense_buffer(unchecked_isa_dma,
1734                                                     GFP_KERNEL, numa_node);
1735         if (!cmd->sense_buffer)
1736                 return -ENOMEM;
1737         cmd->req.sense = cmd->sense_buffer;
1738
1739         if (scsi_host_get_prot(shost)) {
1740                 sg = (void *)cmd + sizeof(struct scsi_cmnd) +
1741                         shost->hostt->cmd_size;
1742                 cmd->prot_sdb = (void *)sg + scsi_mq_sgl_size(shost);
1743         }
1744
1745         return 0;
1746 }
1747
1748 static void scsi_mq_exit_request(struct blk_mq_tag_set *set, struct request *rq,
1749                                  unsigned int hctx_idx)
1750 {
1751         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1752
1753         scsi_free_sense_buffer(cmd->flags & SCMD_UNCHECKED_ISA_DMA,
1754                                cmd->sense_buffer);
1755 }
1756
1757 static int scsi_map_queues(struct blk_mq_tag_set *set)
1758 {
1759         struct Scsi_Host *shost = container_of(set, struct Scsi_Host, tag_set);
1760
1761         if (shost->hostt->map_queues)
1762                 return shost->hostt->map_queues(shost);
1763         return blk_mq_map_queues(&set->map[0]);
1764 }
1765
1766 void __scsi_init_queue(struct Scsi_Host *shost, struct request_queue *q)
1767 {
1768         struct device *dev = shost->dma_dev;
1769
1770         /*
1771          * this limit is imposed by hardware restrictions
1772          */
1773         blk_queue_max_segments(q, min_t(unsigned short, shost->sg_tablesize,
1774                                         SG_MAX_SEGMENTS));
1775
1776         if (scsi_host_prot_dma(shost)) {
1777                 shost->sg_prot_tablesize =
1778                         min_not_zero(shost->sg_prot_tablesize,
1779                                      (unsigned short)SCSI_MAX_PROT_SG_SEGMENTS);
1780                 BUG_ON(shost->sg_prot_tablesize < shost->sg_tablesize);
1781                 blk_queue_max_integrity_segments(q, shost->sg_prot_tablesize);
1782         }
1783
1784         blk_queue_max_hw_sectors(q, shost->max_sectors);
1785         if (shost->unchecked_isa_dma)
1786                 blk_queue_bounce_limit(q, BLK_BOUNCE_ISA);
1787         blk_queue_segment_boundary(q, shost->dma_boundary);
1788         dma_set_seg_boundary(dev, shost->dma_boundary);
1789
1790         blk_queue_max_segment_size(q,
1791                 min(shost->max_segment_size, dma_get_max_seg_size(dev)));
1792
1793         /*
1794          * Set a reasonable default alignment:  The larger of 32-byte (dword),
1795          * which is a common minimum for HBAs, and the minimum DMA alignment,
1796          * which is set by the platform.
1797          *
1798          * Devices that require a bigger alignment can increase it later.
1799          */
1800         blk_queue_dma_alignment(q, max(4, dma_get_cache_alignment()) - 1);
1801 }
1802 EXPORT_SYMBOL_GPL(__scsi_init_queue);
1803
1804 static const struct blk_mq_ops scsi_mq_ops = {
1805         .get_budget     = scsi_mq_get_budget,
1806         .put_budget     = scsi_mq_put_budget,
1807         .queue_rq       = scsi_queue_rq,
1808         .complete       = scsi_softirq_done,
1809         .timeout        = scsi_timeout,
1810 #ifdef CONFIG_BLK_DEBUG_FS
1811         .show_rq        = scsi_show_rq,
1812 #endif
1813         .init_request   = scsi_mq_init_request,
1814         .exit_request   = scsi_mq_exit_request,
1815         .initialize_rq_fn = scsi_initialize_rq,
1816         .busy           = scsi_mq_lld_busy,
1817         .map_queues     = scsi_map_queues,
1818 };
1819
1820 struct request_queue *scsi_mq_alloc_queue(struct scsi_device *sdev)
1821 {
1822         sdev->request_queue = blk_mq_init_queue(&sdev->host->tag_set);
1823         if (IS_ERR(sdev->request_queue))
1824                 return NULL;
1825
1826         sdev->request_queue->queuedata = sdev;
1827         __scsi_init_queue(sdev->host, sdev->request_queue);
1828         blk_queue_flag_set(QUEUE_FLAG_SCSI_PASSTHROUGH, sdev->request_queue);
1829         return sdev->request_queue;
1830 }
1831
1832 int scsi_mq_setup_tags(struct Scsi_Host *shost)
1833 {
1834         unsigned int cmd_size, sgl_size;
1835
1836         sgl_size = scsi_mq_sgl_size(shost);
1837         cmd_size = sizeof(struct scsi_cmnd) + shost->hostt->cmd_size + sgl_size;
1838         if (scsi_host_get_prot(shost))
1839                 cmd_size += sizeof(struct scsi_data_buffer) + sgl_size;
1840
1841         memset(&shost->tag_set, 0, sizeof(shost->tag_set));
1842         shost->tag_set.ops = &scsi_mq_ops;
1843         shost->tag_set.nr_hw_queues = shost->nr_hw_queues ? : 1;
1844         shost->tag_set.queue_depth = shost->can_queue;
1845         shost->tag_set.cmd_size = cmd_size;
1846         shost->tag_set.numa_node = NUMA_NO_NODE;
1847         shost->tag_set.flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_SG_MERGE;
1848         shost->tag_set.flags |=
1849                 BLK_ALLOC_POLICY_TO_MQ_FLAG(shost->hostt->tag_alloc_policy);
1850         shost->tag_set.driver_data = shost;
1851
1852         return blk_mq_alloc_tag_set(&shost->tag_set);
1853 }
1854
1855 void scsi_mq_destroy_tags(struct Scsi_Host *shost)
1856 {
1857         blk_mq_free_tag_set(&shost->tag_set);
1858 }
1859
1860 /**
1861  * scsi_device_from_queue - return sdev associated with a request_queue
1862  * @q: The request queue to return the sdev from
1863  *
1864  * Return the sdev associated with a request queue or NULL if the
1865  * request_queue does not reference a SCSI device.
1866  */
1867 struct scsi_device *scsi_device_from_queue(struct request_queue *q)
1868 {
1869         struct scsi_device *sdev = NULL;
1870
1871         if (q->mq_ops == &scsi_mq_ops)
1872                 sdev = q->queuedata;
1873         if (!sdev || !get_device(&sdev->sdev_gendev))
1874                 sdev = NULL;
1875
1876         return sdev;
1877 }
1878 EXPORT_SYMBOL_GPL(scsi_device_from_queue);
1879
1880 /*
1881  * Function:    scsi_block_requests()
1882  *
1883  * Purpose:     Utility function used by low-level drivers to prevent further
1884  *              commands from being queued to the device.
1885  *
1886  * Arguments:   shost       - Host in question
1887  *
1888  * Returns:     Nothing
1889  *
1890  * Lock status: No locks are assumed held.
1891  *
1892  * Notes:       There is no timer nor any other means by which the requests
1893  *              get unblocked other than the low-level driver calling
1894  *              scsi_unblock_requests().
1895  */
1896 void scsi_block_requests(struct Scsi_Host *shost)
1897 {
1898         shost->host_self_blocked = 1;
1899 }
1900 EXPORT_SYMBOL(scsi_block_requests);
1901
1902 /*
1903  * Function:    scsi_unblock_requests()
1904  *
1905  * Purpose:     Utility function used by low-level drivers to allow further
1906  *              commands from being queued to the device.
1907  *
1908  * Arguments:   shost       - Host in question
1909  *
1910  * Returns:     Nothing
1911  *
1912  * Lock status: No locks are assumed held.
1913  *
1914  * Notes:       There is no timer nor any other means by which the requests
1915  *              get unblocked other than the low-level driver calling
1916  *              scsi_unblock_requests().
1917  *
1918  *              This is done as an API function so that changes to the
1919  *              internals of the scsi mid-layer won't require wholesale
1920  *              changes to drivers that use this feature.
1921  */
1922 void scsi_unblock_requests(struct Scsi_Host *shost)
1923 {
1924         shost->host_self_blocked = 0;
1925         scsi_run_host_queues(shost);
1926 }
1927 EXPORT_SYMBOL(scsi_unblock_requests);
1928
1929 int __init scsi_init_queue(void)
1930 {
1931         scsi_sdb_cache = kmem_cache_create("scsi_data_buffer",
1932                                            sizeof(struct scsi_data_buffer),
1933                                            0, 0, NULL);
1934         if (!scsi_sdb_cache) {
1935                 printk(KERN_ERR "SCSI: can't init scsi sdb cache\n");
1936                 return -ENOMEM;
1937         }
1938
1939         return 0;
1940 }
1941
1942 void scsi_exit_queue(void)
1943 {
1944         kmem_cache_destroy(scsi_sense_cache);
1945         kmem_cache_destroy(scsi_sense_isadma_cache);
1946         kmem_cache_destroy(scsi_sdb_cache);
1947 }
1948
1949 /**
1950  *      scsi_mode_select - issue a mode select
1951  *      @sdev:  SCSI device to be queried
1952  *      @pf:    Page format bit (1 == standard, 0 == vendor specific)
1953  *      @sp:    Save page bit (0 == don't save, 1 == save)
1954  *      @modepage: mode page being requested
1955  *      @buffer: request buffer (may not be smaller than eight bytes)
1956  *      @len:   length of request buffer.
1957  *      @timeout: command timeout
1958  *      @retries: number of retries before failing
1959  *      @data: returns a structure abstracting the mode header data
1960  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
1961  *              must be SCSI_SENSE_BUFFERSIZE big.
1962  *
1963  *      Returns zero if successful; negative error number or scsi
1964  *      status on error
1965  *
1966  */
1967 int
1968 scsi_mode_select(struct scsi_device *sdev, int pf, int sp, int modepage,
1969                  unsigned char *buffer, int len, int timeout, int retries,
1970                  struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1971 {
1972         unsigned char cmd[10];
1973         unsigned char *real_buffer;
1974         int ret;
1975
1976         memset(cmd, 0, sizeof(cmd));
1977         cmd[1] = (pf ? 0x10 : 0) | (sp ? 0x01 : 0);
1978
1979         if (sdev->use_10_for_ms) {
1980                 if (len > 65535)
1981                         return -EINVAL;
1982                 real_buffer = kmalloc(8 + len, GFP_KERNEL);
1983                 if (!real_buffer)
1984                         return -ENOMEM;
1985                 memcpy(real_buffer + 8, buffer, len);
1986                 len += 8;
1987                 real_buffer[0] = 0;
1988                 real_buffer[1] = 0;
1989                 real_buffer[2] = data->medium_type;
1990                 real_buffer[3] = data->device_specific;
1991                 real_buffer[4] = data->longlba ? 0x01 : 0;
1992                 real_buffer[5] = 0;
1993                 real_buffer[6] = data->block_descriptor_length >> 8;
1994                 real_buffer[7] = data->block_descriptor_length;
1995
1996                 cmd[0] = MODE_SELECT_10;
1997                 cmd[7] = len >> 8;
1998                 cmd[8] = len;
1999         } else {
2000                 if (len > 255 || data->block_descriptor_length > 255 ||
2001                     data->longlba)
2002                         return -EINVAL;
2003
2004                 real_buffer = kmalloc(4 + len, GFP_KERNEL);
2005                 if (!real_buffer)
2006                         return -ENOMEM;
2007                 memcpy(real_buffer + 4, buffer, len);
2008                 len += 4;
2009                 real_buffer[0] = 0;
2010                 real_buffer[1] = data->medium_type;
2011                 real_buffer[2] = data->device_specific;
2012                 real_buffer[3] = data->block_descriptor_length;
2013                 
2014
2015                 cmd[0] = MODE_SELECT;
2016                 cmd[4] = len;
2017         }
2018
2019         ret = scsi_execute_req(sdev, cmd, DMA_TO_DEVICE, real_buffer, len,
2020                                sshdr, timeout, retries, NULL);
2021         kfree(real_buffer);
2022         return ret;
2023 }
2024 EXPORT_SYMBOL_GPL(scsi_mode_select);
2025
2026 /**
2027  *      scsi_mode_sense - issue a mode sense, falling back from 10 to six bytes if necessary.
2028  *      @sdev:  SCSI device to be queried
2029  *      @dbd:   set if mode sense will allow block descriptors to be returned
2030  *      @modepage: mode page being requested
2031  *      @buffer: request buffer (may not be smaller than eight bytes)
2032  *      @len:   length of request buffer.
2033  *      @timeout: command timeout
2034  *      @retries: number of retries before failing
2035  *      @data: returns a structure abstracting the mode header data
2036  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
2037  *              must be SCSI_SENSE_BUFFERSIZE big.
2038  *
2039  *      Returns zero if unsuccessful, or the header offset (either 4
2040  *      or 8 depending on whether a six or ten byte command was
2041  *      issued) if successful.
2042  */
2043 int
2044 scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
2045                   unsigned char *buffer, int len, int timeout, int retries,
2046                   struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
2047 {
2048         unsigned char cmd[12];
2049         int use_10_for_ms;
2050         int header_length;
2051         int result, retry_count = retries;
2052         struct scsi_sense_hdr my_sshdr;
2053
2054         memset(data, 0, sizeof(*data));
2055         memset(&cmd[0], 0, 12);
2056         cmd[1] = dbd & 0x18;    /* allows DBD and LLBA bits */
2057         cmd[2] = modepage;
2058
2059         /* caller might not be interested in sense, but we need it */
2060         if (!sshdr)
2061                 sshdr = &my_sshdr;
2062
2063  retry:
2064         use_10_for_ms = sdev->use_10_for_ms;
2065
2066         if (use_10_for_ms) {
2067                 if (len < 8)
2068                         len = 8;
2069
2070                 cmd[0] = MODE_SENSE_10;
2071                 cmd[8] = len;
2072                 header_length = 8;
2073         } else {
2074                 if (len < 4)
2075                         len = 4;
2076
2077                 cmd[0] = MODE_SENSE;
2078                 cmd[4] = len;
2079                 header_length = 4;
2080         }
2081
2082         memset(buffer, 0, len);
2083
2084         result = scsi_execute_req(sdev, cmd, DMA_FROM_DEVICE, buffer, len,
2085                                   sshdr, timeout, retries, NULL);
2086
2087         /* This code looks awful: what it's doing is making sure an
2088          * ILLEGAL REQUEST sense return identifies the actual command
2089          * byte as the problem.  MODE_SENSE commands can return
2090          * ILLEGAL REQUEST if the code page isn't supported */
2091
2092         if (use_10_for_ms && !scsi_status_is_good(result) &&
2093             driver_byte(result) == DRIVER_SENSE) {
2094                 if (scsi_sense_valid(sshdr)) {
2095                         if ((sshdr->sense_key == ILLEGAL_REQUEST) &&
2096                             (sshdr->asc == 0x20) && (sshdr->ascq == 0)) {
2097                                 /* 
2098                                  * Invalid command operation code
2099                                  */
2100                                 sdev->use_10_for_ms = 0;
2101                                 goto retry;
2102                         }
2103                 }
2104         }
2105
2106         if(scsi_status_is_good(result)) {
2107                 if (unlikely(buffer[0] == 0x86 && buffer[1] == 0x0b &&
2108                              (modepage == 6 || modepage == 8))) {
2109                         /* Initio breakage? */
2110                         header_length = 0;
2111                         data->length = 13;
2112                         data->medium_type = 0;
2113                         data->device_specific = 0;
2114                         data->longlba = 0;
2115                         data->block_descriptor_length = 0;
2116                 } else if(use_10_for_ms) {
2117                         data->length = buffer[0]*256 + buffer[1] + 2;
2118                         data->medium_type = buffer[2];
2119                         data->device_specific = buffer[3];
2120                         data->longlba = buffer[4] & 0x01;
2121                         data->block_descriptor_length = buffer[6]*256
2122                                 + buffer[7];
2123                 } else {
2124                         data->length = buffer[0] + 1;
2125                         data->medium_type = buffer[1];
2126                         data->device_specific = buffer[2];
2127                         data->block_descriptor_length = buffer[3];
2128                 }
2129                 data->header_length = header_length;
2130         } else if ((status_byte(result) == CHECK_CONDITION) &&
2131                    scsi_sense_valid(sshdr) &&
2132                    sshdr->sense_key == UNIT_ATTENTION && retry_count) {
2133                 retry_count--;
2134                 goto retry;
2135         }
2136
2137         return result;
2138 }
2139 EXPORT_SYMBOL(scsi_mode_sense);
2140
2141 /**
2142  *      scsi_test_unit_ready - test if unit is ready
2143  *      @sdev:  scsi device to change the state of.
2144  *      @timeout: command timeout
2145  *      @retries: number of retries before failing
2146  *      @sshdr: outpout pointer for decoded sense information.
2147  *
2148  *      Returns zero if unsuccessful or an error if TUR failed.  For
2149  *      removable media, UNIT_ATTENTION sets ->changed flag.
2150  **/
2151 int
2152 scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries,
2153                      struct scsi_sense_hdr *sshdr)
2154 {
2155         char cmd[] = {
2156                 TEST_UNIT_READY, 0, 0, 0, 0, 0,
2157         };
2158         int result;
2159
2160         /* try to eat the UNIT_ATTENTION if there are enough retries */
2161         do {
2162                 result = scsi_execute_req(sdev, cmd, DMA_NONE, NULL, 0, sshdr,
2163                                           timeout, 1, NULL);
2164                 if (sdev->removable && scsi_sense_valid(sshdr) &&
2165                     sshdr->sense_key == UNIT_ATTENTION)
2166                         sdev->changed = 1;
2167         } while (scsi_sense_valid(sshdr) &&
2168                  sshdr->sense_key == UNIT_ATTENTION && --retries);
2169
2170         return result;
2171 }
2172 EXPORT_SYMBOL(scsi_test_unit_ready);
2173
2174 /**
2175  *      scsi_device_set_state - Take the given device through the device state model.
2176  *      @sdev:  scsi device to change the state of.
2177  *      @state: state to change to.
2178  *
2179  *      Returns zero if successful or an error if the requested
2180  *      transition is illegal.
2181  */
2182 int
2183 scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
2184 {
2185         enum scsi_device_state oldstate = sdev->sdev_state;
2186
2187         if (state == oldstate)
2188                 return 0;
2189
2190         switch (state) {
2191         case SDEV_CREATED:
2192                 switch (oldstate) {
2193                 case SDEV_CREATED_BLOCK:
2194                         break;
2195                 default:
2196                         goto illegal;
2197                 }
2198                 break;
2199                         
2200         case SDEV_RUNNING:
2201                 switch (oldstate) {
2202                 case SDEV_CREATED:
2203                 case SDEV_OFFLINE:
2204                 case SDEV_TRANSPORT_OFFLINE:
2205                 case SDEV_QUIESCE:
2206                 case SDEV_BLOCK:
2207                         break;
2208                 default:
2209                         goto illegal;
2210                 }
2211                 break;
2212
2213         case SDEV_QUIESCE:
2214                 switch (oldstate) {
2215                 case SDEV_RUNNING:
2216                 case SDEV_OFFLINE:
2217                 case SDEV_TRANSPORT_OFFLINE:
2218                         break;
2219                 default:
2220                         goto illegal;
2221                 }
2222                 break;
2223
2224         case SDEV_OFFLINE:
2225         case SDEV_TRANSPORT_OFFLINE:
2226                 switch (oldstate) {
2227                 case SDEV_CREATED:
2228                 case SDEV_RUNNING:
2229                 case SDEV_QUIESCE:
2230                 case SDEV_BLOCK:
2231                         break;
2232                 default:
2233                         goto illegal;
2234                 }
2235                 break;
2236
2237         case SDEV_BLOCK:
2238                 switch (oldstate) {
2239                 case SDEV_RUNNING:
2240                 case SDEV_CREATED_BLOCK:
2241                 case SDEV_OFFLINE:
2242                         break;
2243                 default:
2244                         goto illegal;
2245                 }
2246                 break;
2247
2248         case SDEV_CREATED_BLOCK:
2249                 switch (oldstate) {
2250                 case SDEV_CREATED:
2251                         break;
2252                 default:
2253                         goto illegal;
2254                 }
2255                 break;
2256
2257         case SDEV_CANCEL:
2258                 switch (oldstate) {
2259                 case SDEV_CREATED:
2260                 case SDEV_RUNNING:
2261                 case SDEV_QUIESCE:
2262                 case SDEV_OFFLINE:
2263                 case SDEV_TRANSPORT_OFFLINE:
2264                         break;
2265                 default:
2266                         goto illegal;
2267                 }
2268                 break;
2269
2270         case SDEV_DEL:
2271                 switch (oldstate) {
2272                 case SDEV_CREATED:
2273                 case SDEV_RUNNING:
2274                 case SDEV_OFFLINE:
2275                 case SDEV_TRANSPORT_OFFLINE:
2276                 case SDEV_CANCEL:
2277                 case SDEV_BLOCK:
2278                 case SDEV_CREATED_BLOCK:
2279                         break;
2280                 default:
2281                         goto illegal;
2282                 }
2283                 break;
2284
2285         }
2286         sdev->sdev_state = state;
2287         return 0;
2288
2289  illegal:
2290         SCSI_LOG_ERROR_RECOVERY(1,
2291                                 sdev_printk(KERN_ERR, sdev,
2292                                             "Illegal state transition %s->%s",
2293                                             scsi_device_state_name(oldstate),
2294                                             scsi_device_state_name(state))
2295                                 );
2296         return -EINVAL;
2297 }
2298 EXPORT_SYMBOL(scsi_device_set_state);
2299
2300 /**
2301  *      sdev_evt_emit - emit a single SCSI device uevent
2302  *      @sdev: associated SCSI device
2303  *      @evt: event to emit
2304  *
2305  *      Send a single uevent (scsi_event) to the associated scsi_device.
2306  */
2307 static void scsi_evt_emit(struct scsi_device *sdev, struct scsi_event *evt)
2308 {
2309         int idx = 0;
2310         char *envp[3];
2311
2312         switch (evt->evt_type) {
2313         case SDEV_EVT_MEDIA_CHANGE:
2314                 envp[idx++] = "SDEV_MEDIA_CHANGE=1";
2315                 break;
2316         case SDEV_EVT_INQUIRY_CHANGE_REPORTED:
2317                 scsi_rescan_device(&sdev->sdev_gendev);
2318                 envp[idx++] = "SDEV_UA=INQUIRY_DATA_HAS_CHANGED";
2319                 break;
2320         case SDEV_EVT_CAPACITY_CHANGE_REPORTED:
2321                 envp[idx++] = "SDEV_UA=CAPACITY_DATA_HAS_CHANGED";
2322                 break;
2323         case SDEV_EVT_SOFT_THRESHOLD_REACHED_REPORTED:
2324                envp[idx++] = "SDEV_UA=THIN_PROVISIONING_SOFT_THRESHOLD_REACHED";
2325                 break;
2326         case SDEV_EVT_MODE_PARAMETER_CHANGE_REPORTED:
2327                 envp[idx++] = "SDEV_UA=MODE_PARAMETERS_CHANGED";
2328                 break;
2329         case SDEV_EVT_LUN_CHANGE_REPORTED:
2330                 envp[idx++] = "SDEV_UA=REPORTED_LUNS_DATA_HAS_CHANGED";
2331                 break;
2332         case SDEV_EVT_ALUA_STATE_CHANGE_REPORTED:
2333                 envp[idx++] = "SDEV_UA=ASYMMETRIC_ACCESS_STATE_CHANGED";
2334                 break;
2335         case SDEV_EVT_POWER_ON_RESET_OCCURRED:
2336                 envp[idx++] = "SDEV_UA=POWER_ON_RESET_OCCURRED";
2337                 break;
2338         default:
2339                 /* do nothing */
2340                 break;
2341         }
2342
2343         envp[idx++] = NULL;
2344
2345         kobject_uevent_env(&sdev->sdev_gendev.kobj, KOBJ_CHANGE, envp);
2346 }
2347
2348 /**
2349  *      sdev_evt_thread - send a uevent for each scsi event
2350  *      @work: work struct for scsi_device
2351  *
2352  *      Dispatch queued events to their associated scsi_device kobjects
2353  *      as uevents.
2354  */
2355 void scsi_evt_thread(struct work_struct *work)
2356 {
2357         struct scsi_device *sdev;
2358         enum scsi_device_event evt_type;
2359         LIST_HEAD(event_list);
2360
2361         sdev = container_of(work, struct scsi_device, event_work);
2362
2363         for (evt_type = SDEV_EVT_FIRST; evt_type <= SDEV_EVT_LAST; evt_type++)
2364                 if (test_and_clear_bit(evt_type, sdev->pending_events))
2365                         sdev_evt_send_simple(sdev, evt_type, GFP_KERNEL);
2366
2367         while (1) {
2368                 struct scsi_event *evt;
2369                 struct list_head *this, *tmp;
2370                 unsigned long flags;
2371
2372                 spin_lock_irqsave(&sdev->list_lock, flags);
2373                 list_splice_init(&sdev->event_list, &event_list);
2374                 spin_unlock_irqrestore(&sdev->list_lock, flags);
2375
2376                 if (list_empty(&event_list))
2377                         break;
2378
2379                 list_for_each_safe(this, tmp, &event_list) {
2380                         evt = list_entry(this, struct scsi_event, node);
2381                         list_del(&evt->node);
2382                         scsi_evt_emit(sdev, evt);
2383                         kfree(evt);
2384                 }
2385         }
2386 }
2387
2388 /**
2389  *      sdev_evt_send - send asserted event to uevent thread
2390  *      @sdev: scsi_device event occurred on
2391  *      @evt: event to send
2392  *
2393  *      Assert scsi device event asynchronously.
2394  */
2395 void sdev_evt_send(struct scsi_device *sdev, struct scsi_event *evt)
2396 {
2397         unsigned long flags;
2398
2399 #if 0
2400         /* FIXME: currently this check eliminates all media change events
2401          * for polled devices.  Need to update to discriminate between AN
2402          * and polled events */
2403         if (!test_bit(evt->evt_type, sdev->supported_events)) {
2404                 kfree(evt);
2405                 return;
2406         }
2407 #endif
2408
2409         spin_lock_irqsave(&sdev->list_lock, flags);
2410         list_add_tail(&evt->node, &sdev->event_list);
2411         schedule_work(&sdev->event_work);
2412         spin_unlock_irqrestore(&sdev->list_lock, flags);
2413 }
2414 EXPORT_SYMBOL_GPL(sdev_evt_send);
2415
2416 /**
2417  *      sdev_evt_alloc - allocate a new scsi event
2418  *      @evt_type: type of event to allocate
2419  *      @gfpflags: GFP flags for allocation
2420  *
2421  *      Allocates and returns a new scsi_event.
2422  */
2423 struct scsi_event *sdev_evt_alloc(enum scsi_device_event evt_type,
2424                                   gfp_t gfpflags)
2425 {
2426         struct scsi_event *evt = kzalloc(sizeof(struct scsi_event), gfpflags);
2427         if (!evt)
2428                 return NULL;
2429
2430         evt->evt_type = evt_type;
2431         INIT_LIST_HEAD(&evt->node);
2432
2433         /* evt_type-specific initialization, if any */
2434         switch (evt_type) {
2435         case SDEV_EVT_MEDIA_CHANGE:
2436         case SDEV_EVT_INQUIRY_CHANGE_REPORTED:
2437         case SDEV_EVT_CAPACITY_CHANGE_REPORTED:
2438         case SDEV_EVT_SOFT_THRESHOLD_REACHED_REPORTED:
2439         case SDEV_EVT_MODE_PARAMETER_CHANGE_REPORTED:
2440         case SDEV_EVT_LUN_CHANGE_REPORTED:
2441         case SDEV_EVT_ALUA_STATE_CHANGE_REPORTED:
2442         case SDEV_EVT_POWER_ON_RESET_OCCURRED:
2443         default:
2444                 /* do nothing */
2445                 break;
2446         }
2447
2448         return evt;
2449 }
2450 EXPORT_SYMBOL_GPL(sdev_evt_alloc);
2451
2452 /**
2453  *      sdev_evt_send_simple - send asserted event to uevent thread
2454  *      @sdev: scsi_device event occurred on
2455  *      @evt_type: type of event to send
2456  *      @gfpflags: GFP flags for allocation
2457  *
2458  *      Assert scsi device event asynchronously, given an event type.
2459  */
2460 void sdev_evt_send_simple(struct scsi_device *sdev,
2461                           enum scsi_device_event evt_type, gfp_t gfpflags)
2462 {
2463         struct scsi_event *evt = sdev_evt_alloc(evt_type, gfpflags);
2464         if (!evt) {
2465                 sdev_printk(KERN_ERR, sdev, "event %d eaten due to OOM\n",
2466                             evt_type);
2467                 return;
2468         }
2469
2470         sdev_evt_send(sdev, evt);
2471 }
2472 EXPORT_SYMBOL_GPL(sdev_evt_send_simple);
2473
2474 /**
2475  *      scsi_device_quiesce - Block user issued commands.
2476  *      @sdev:  scsi device to quiesce.
2477  *
2478  *      This works by trying to transition to the SDEV_QUIESCE state
2479  *      (which must be a legal transition).  When the device is in this
2480  *      state, only special requests will be accepted, all others will
2481  *      be deferred.  Since special requests may also be requeued requests,
2482  *      a successful return doesn't guarantee the device will be 
2483  *      totally quiescent.
2484  *
2485  *      Must be called with user context, may sleep.
2486  *
2487  *      Returns zero if unsuccessful or an error if not.
2488  */
2489 int
2490 scsi_device_quiesce(struct scsi_device *sdev)
2491 {
2492         struct request_queue *q = sdev->request_queue;
2493         int err;
2494
2495         /*
2496          * It is allowed to call scsi_device_quiesce() multiple times from
2497          * the same context but concurrent scsi_device_quiesce() calls are
2498          * not allowed.
2499          */
2500         WARN_ON_ONCE(sdev->quiesced_by && sdev->quiesced_by != current);
2501
2502         if (sdev->quiesced_by == current)
2503                 return 0;
2504
2505         blk_set_pm_only(q);
2506
2507         blk_mq_freeze_queue(q);
2508         /*
2509          * Ensure that the effect of blk_set_pm_only() will be visible
2510          * for percpu_ref_tryget() callers that occur after the queue
2511          * unfreeze even if the queue was already frozen before this function
2512          * was called. See also https://lwn.net/Articles/573497/.
2513          */
2514         synchronize_rcu();
2515         blk_mq_unfreeze_queue(q);
2516
2517         mutex_lock(&sdev->state_mutex);
2518         err = scsi_device_set_state(sdev, SDEV_QUIESCE);
2519         if (err == 0)
2520                 sdev->quiesced_by = current;
2521         else
2522                 blk_clear_pm_only(q);
2523         mutex_unlock(&sdev->state_mutex);
2524
2525         return err;
2526 }
2527 EXPORT_SYMBOL(scsi_device_quiesce);
2528
2529 /**
2530  *      scsi_device_resume - Restart user issued commands to a quiesced device.
2531  *      @sdev:  scsi device to resume.
2532  *
2533  *      Moves the device from quiesced back to running and restarts the
2534  *      queues.
2535  *
2536  *      Must be called with user context, may sleep.
2537  */
2538 void scsi_device_resume(struct scsi_device *sdev)
2539 {
2540         /* check if the device state was mutated prior to resume, and if
2541          * so assume the state is being managed elsewhere (for example
2542          * device deleted during suspend)
2543          */
2544         mutex_lock(&sdev->state_mutex);
2545         WARN_ON_ONCE(!sdev->quiesced_by);
2546         sdev->quiesced_by = NULL;
2547         blk_clear_pm_only(sdev->request_queue);
2548         if (sdev->sdev_state == SDEV_QUIESCE)
2549                 scsi_device_set_state(sdev, SDEV_RUNNING);
2550         mutex_unlock(&sdev->state_mutex);
2551 }
2552 EXPORT_SYMBOL(scsi_device_resume);
2553
2554 static void
2555 device_quiesce_fn(struct scsi_device *sdev, void *data)
2556 {
2557         scsi_device_quiesce(sdev);
2558 }
2559
2560 void
2561 scsi_target_quiesce(struct scsi_target *starget)
2562 {
2563         starget_for_each_device(starget, NULL, device_quiesce_fn);
2564 }
2565 EXPORT_SYMBOL(scsi_target_quiesce);
2566
2567 static void
2568 device_resume_fn(struct scsi_device *sdev, void *data)
2569 {
2570         scsi_device_resume(sdev);
2571 }
2572
2573 void
2574 scsi_target_resume(struct scsi_target *starget)
2575 {
2576         starget_for_each_device(starget, NULL, device_resume_fn);
2577 }
2578 EXPORT_SYMBOL(scsi_target_resume);
2579
2580 /**
2581  * scsi_internal_device_block_nowait - try to transition to the SDEV_BLOCK state
2582  * @sdev: device to block
2583  *
2584  * Pause SCSI command processing on the specified device. Does not sleep.
2585  *
2586  * Returns zero if successful or a negative error code upon failure.
2587  *
2588  * Notes:
2589  * This routine transitions the device to the SDEV_BLOCK state (which must be
2590  * a legal transition). When the device is in this state, command processing
2591  * is paused until the device leaves the SDEV_BLOCK state. See also
2592  * scsi_internal_device_unblock_nowait().
2593  */
2594 int scsi_internal_device_block_nowait(struct scsi_device *sdev)
2595 {
2596         struct request_queue *q = sdev->request_queue;
2597         int err = 0;
2598
2599         err = scsi_device_set_state(sdev, SDEV_BLOCK);
2600         if (err) {
2601                 err = scsi_device_set_state(sdev, SDEV_CREATED_BLOCK);
2602
2603                 if (err)
2604                         return err;
2605         }
2606
2607         /* 
2608          * The device has transitioned to SDEV_BLOCK.  Stop the
2609          * block layer from calling the midlayer with this device's
2610          * request queue. 
2611          */
2612         blk_mq_quiesce_queue_nowait(q);
2613         return 0;
2614 }
2615 EXPORT_SYMBOL_GPL(scsi_internal_device_block_nowait);
2616
2617 /**
2618  * scsi_internal_device_block - try to transition to the SDEV_BLOCK state
2619  * @sdev: device to block
2620  *
2621  * Pause SCSI command processing on the specified device and wait until all
2622  * ongoing scsi_request_fn() / scsi_queue_rq() calls have finished. May sleep.
2623  *
2624  * Returns zero if successful or a negative error code upon failure.
2625  *
2626  * Note:
2627  * This routine transitions the device to the SDEV_BLOCK state (which must be
2628  * a legal transition). When the device is in this state, command processing
2629  * is paused until the device leaves the SDEV_BLOCK state. See also
2630  * scsi_internal_device_unblock().
2631  *
2632  * To do: avoid that scsi_send_eh_cmnd() calls queuecommand() after
2633  * scsi_internal_device_block() has blocked a SCSI device and also
2634  * remove the rport mutex lock and unlock calls from srp_queuecommand().
2635  */
2636 static int scsi_internal_device_block(struct scsi_device *sdev)
2637 {
2638         struct request_queue *q = sdev->request_queue;
2639         int err;
2640
2641         mutex_lock(&sdev->state_mutex);
2642         err = scsi_internal_device_block_nowait(sdev);
2643         if (err == 0)
2644                 blk_mq_quiesce_queue(q);
2645         mutex_unlock(&sdev->state_mutex);
2646
2647         return err;
2648 }
2649  
2650 void scsi_start_queue(struct scsi_device *sdev)
2651 {
2652         struct request_queue *q = sdev->request_queue;
2653
2654         blk_mq_unquiesce_queue(q);
2655 }
2656
2657 /**
2658  * scsi_internal_device_unblock_nowait - resume a device after a block request
2659  * @sdev:       device to resume
2660  * @new_state:  state to set the device to after unblocking
2661  *
2662  * Restart the device queue for a previously suspended SCSI device. Does not
2663  * sleep.
2664  *
2665  * Returns zero if successful or a negative error code upon failure.
2666  *
2667  * Notes:
2668  * This routine transitions the device to the SDEV_RUNNING state or to one of
2669  * the offline states (which must be a legal transition) allowing the midlayer
2670  * to goose the queue for this device.
2671  */
2672 int scsi_internal_device_unblock_nowait(struct scsi_device *sdev,
2673                                         enum scsi_device_state new_state)
2674 {
2675         /*
2676          * Try to transition the scsi device to SDEV_RUNNING or one of the
2677          * offlined states and goose the device queue if successful.
2678          */
2679         switch (sdev->sdev_state) {
2680         case SDEV_BLOCK:
2681         case SDEV_TRANSPORT_OFFLINE:
2682                 sdev->sdev_state = new_state;
2683                 break;
2684         case SDEV_CREATED_BLOCK:
2685                 if (new_state == SDEV_TRANSPORT_OFFLINE ||
2686                     new_state == SDEV_OFFLINE)
2687                         sdev->sdev_state = new_state;
2688                 else
2689                         sdev->sdev_state = SDEV_CREATED;
2690                 break;
2691         case SDEV_CANCEL:
2692         case SDEV_OFFLINE:
2693                 break;
2694         default:
2695                 return -EINVAL;
2696         }
2697         scsi_start_queue(sdev);
2698
2699         return 0;
2700 }
2701 EXPORT_SYMBOL_GPL(scsi_internal_device_unblock_nowait);
2702
2703 /**
2704  * scsi_internal_device_unblock - resume a device after a block request
2705  * @sdev:       device to resume
2706  * @new_state:  state to set the device to after unblocking
2707  *
2708  * Restart the device queue for a previously suspended SCSI device. May sleep.
2709  *
2710  * Returns zero if successful or a negative error code upon failure.
2711  *
2712  * Notes:
2713  * This routine transitions the device to the SDEV_RUNNING state or to one of
2714  * the offline states (which must be a legal transition) allowing the midlayer
2715  * to goose the queue for this device.
2716  */
2717 static int scsi_internal_device_unblock(struct scsi_device *sdev,
2718                                         enum scsi_device_state new_state)
2719 {
2720         int ret;
2721
2722         mutex_lock(&sdev->state_mutex);
2723         ret = scsi_internal_device_unblock_nowait(sdev, new_state);
2724         mutex_unlock(&sdev->state_mutex);
2725
2726         return ret;
2727 }
2728
2729 static void
2730 device_block(struct scsi_device *sdev, void *data)
2731 {
2732         scsi_internal_device_block(sdev);
2733 }
2734
2735 static int
2736 target_block(struct device *dev, void *data)
2737 {
2738         if (scsi_is_target_device(dev))
2739                 starget_for_each_device(to_scsi_target(dev), NULL,
2740                                         device_block);
2741         return 0;
2742 }
2743
2744 void
2745 scsi_target_block(struct device *dev)
2746 {
2747         if (scsi_is_target_device(dev))
2748                 starget_for_each_device(to_scsi_target(dev), NULL,
2749                                         device_block);
2750         else
2751                 device_for_each_child(dev, NULL, target_block);
2752 }
2753 EXPORT_SYMBOL_GPL(scsi_target_block);
2754
2755 static void
2756 device_unblock(struct scsi_device *sdev, void *data)
2757 {
2758         scsi_internal_device_unblock(sdev, *(enum scsi_device_state *)data);
2759 }
2760
2761 static int
2762 target_unblock(struct device *dev, void *data)
2763 {
2764         if (scsi_is_target_device(dev))
2765                 starget_for_each_device(to_scsi_target(dev), data,
2766                                         device_unblock);
2767         return 0;
2768 }
2769
2770 void
2771 scsi_target_unblock(struct device *dev, enum scsi_device_state new_state)
2772 {
2773         if (scsi_is_target_device(dev))
2774                 starget_for_each_device(to_scsi_target(dev), &new_state,
2775                                         device_unblock);
2776         else
2777                 device_for_each_child(dev, &new_state, target_unblock);
2778 }
2779 EXPORT_SYMBOL_GPL(scsi_target_unblock);
2780
2781 /**
2782  * scsi_kmap_atomic_sg - find and atomically map an sg-elemnt
2783  * @sgl:        scatter-gather list
2784  * @sg_count:   number of segments in sg
2785  * @offset:     offset in bytes into sg, on return offset into the mapped area
2786  * @len:        bytes to map, on return number of bytes mapped
2787  *
2788  * Returns virtual address of the start of the mapped page
2789  */
2790 void *scsi_kmap_atomic_sg(struct scatterlist *sgl, int sg_count,
2791                           size_t *offset, size_t *len)
2792 {
2793         int i;
2794         size_t sg_len = 0, len_complete = 0;
2795         struct scatterlist *sg;
2796         struct page *page;
2797
2798         WARN_ON(!irqs_disabled());
2799
2800         for_each_sg(sgl, sg, sg_count, i) {
2801                 len_complete = sg_len; /* Complete sg-entries */
2802                 sg_len += sg->length;
2803                 if (sg_len > *offset)
2804                         break;
2805         }
2806
2807         if (unlikely(i == sg_count)) {
2808                 printk(KERN_ERR "%s: Bytes in sg: %zu, requested offset %zu, "
2809                         "elements %d\n",
2810                        __func__, sg_len, *offset, sg_count);
2811                 WARN_ON(1);
2812                 return NULL;
2813         }
2814
2815         /* Offset starting from the beginning of first page in this sg-entry */
2816         *offset = *offset - len_complete + sg->offset;
2817
2818         /* Assumption: contiguous pages can be accessed as "page + i" */
2819         page = nth_page(sg_page(sg), (*offset >> PAGE_SHIFT));
2820         *offset &= ~PAGE_MASK;
2821
2822         /* Bytes in this sg-entry from *offset to the end of the page */
2823         sg_len = PAGE_SIZE - *offset;
2824         if (*len > sg_len)
2825                 *len = sg_len;
2826
2827         return kmap_atomic(page);
2828 }
2829 EXPORT_SYMBOL(scsi_kmap_atomic_sg);
2830
2831 /**
2832  * scsi_kunmap_atomic_sg - atomically unmap a virtual address, previously mapped with scsi_kmap_atomic_sg
2833  * @virt:       virtual address to be unmapped
2834  */
2835 void scsi_kunmap_atomic_sg(void *virt)
2836 {
2837         kunmap_atomic(virt);
2838 }
2839 EXPORT_SYMBOL(scsi_kunmap_atomic_sg);
2840
2841 void sdev_disable_disk_events(struct scsi_device *sdev)
2842 {
2843         atomic_inc(&sdev->disk_events_disable_depth);
2844 }
2845 EXPORT_SYMBOL(sdev_disable_disk_events);
2846
2847 void sdev_enable_disk_events(struct scsi_device *sdev)
2848 {
2849         if (WARN_ON_ONCE(atomic_read(&sdev->disk_events_disable_depth) <= 0))
2850                 return;
2851         atomic_dec(&sdev->disk_events_disable_depth);
2852 }
2853 EXPORT_SYMBOL(sdev_enable_disk_events);
2854
2855 /**
2856  * scsi_vpd_lun_id - return a unique device identification
2857  * @sdev: SCSI device
2858  * @id:   buffer for the identification
2859  * @id_len:  length of the buffer
2860  *
2861  * Copies a unique device identification into @id based
2862  * on the information in the VPD page 0x83 of the device.
2863  * The string will be formatted as a SCSI name string.
2864  *
2865  * Returns the length of the identification or error on failure.
2866  * If the identifier is longer than the supplied buffer the actual
2867  * identifier length is returned and the buffer is not zero-padded.
2868  */
2869 int scsi_vpd_lun_id(struct scsi_device *sdev, char *id, size_t id_len)
2870 {
2871         u8 cur_id_type = 0xff;
2872         u8 cur_id_size = 0;
2873         const unsigned char *d, *cur_id_str;
2874         const struct scsi_vpd *vpd_pg83;
2875         int id_size = -EINVAL;
2876
2877         rcu_read_lock();
2878         vpd_pg83 = rcu_dereference(sdev->vpd_pg83);
2879         if (!vpd_pg83) {
2880                 rcu_read_unlock();
2881                 return -ENXIO;
2882         }
2883
2884         /*
2885          * Look for the correct descriptor.
2886          * Order of preference for lun descriptor:
2887          * - SCSI name string
2888          * - NAA IEEE Registered Extended
2889          * - EUI-64 based 16-byte
2890          * - EUI-64 based 12-byte
2891          * - NAA IEEE Registered
2892          * - NAA IEEE Extended
2893          * - T10 Vendor ID
2894          * as longer descriptors reduce the likelyhood
2895          * of identification clashes.
2896          */
2897
2898         /* The id string must be at least 20 bytes + terminating NULL byte */
2899         if (id_len < 21) {
2900                 rcu_read_unlock();
2901                 return -EINVAL;
2902         }
2903
2904         memset(id, 0, id_len);
2905         d = vpd_pg83->data + 4;
2906         while (d < vpd_pg83->data + vpd_pg83->len) {
2907                 /* Skip designators not referring to the LUN */
2908                 if ((d[1] & 0x30) != 0x00)
2909                         goto next_desig;
2910
2911                 switch (d[1] & 0xf) {
2912                 case 0x1:
2913                         /* T10 Vendor ID */
2914                         if (cur_id_size > d[3])
2915                                 break;
2916                         /* Prefer anything */
2917                         if (cur_id_type > 0x01 && cur_id_type != 0xff)
2918                                 break;
2919                         cur_id_size = d[3];
2920                         if (cur_id_size + 4 > id_len)
2921                                 cur_id_size = id_len - 4;
2922                         cur_id_str = d + 4;
2923                         cur_id_type = d[1] & 0xf;
2924                         id_size = snprintf(id, id_len, "t10.%*pE",
2925                                            cur_id_size, cur_id_str);
2926                         break;
2927                 case 0x2:
2928                         /* EUI-64 */
2929                         if (cur_id_size > d[3])
2930                                 break;
2931                         /* Prefer NAA IEEE Registered Extended */
2932                         if (cur_id_type == 0x3 &&
2933                             cur_id_size == d[3])
2934                                 break;
2935                         cur_id_size = d[3];
2936                         cur_id_str = d + 4;
2937                         cur_id_type = d[1] & 0xf;
2938                         switch (cur_id_size) {
2939                         case 8:
2940                                 id_size = snprintf(id, id_len,
2941                                                    "eui.%8phN",
2942                                                    cur_id_str);
2943                                 break;
2944                         case 12:
2945                                 id_size = snprintf(id, id_len,
2946                                                    "eui.%12phN",
2947                                                    cur_id_str);
2948                                 break;
2949                         case 16:
2950                                 id_size = snprintf(id, id_len,
2951                                                    "eui.%16phN",
2952                                                    cur_id_str);
2953                                 break;
2954                         default:
2955                                 cur_id_size = 0;
2956                                 break;
2957                         }
2958                         break;
2959                 case 0x3:
2960                         /* NAA */
2961                         if (cur_id_size > d[3])
2962                                 break;
2963                         cur_id_size = d[3];
2964                         cur_id_str = d + 4;
2965                         cur_id_type = d[1] & 0xf;
2966                         switch (cur_id_size) {
2967                         case 8:
2968                                 id_size = snprintf(id, id_len,
2969                                                    "naa.%8phN",
2970                                                    cur_id_str);
2971                                 break;
2972                         case 16:
2973                                 id_size = snprintf(id, id_len,
2974                                                    "naa.%16phN",
2975                                                    cur_id_str);
2976                                 break;
2977                         default:
2978                                 cur_id_size = 0;
2979                                 break;
2980                         }
2981                         break;
2982                 case 0x8:
2983                         /* SCSI name string */
2984                         if (cur_id_size + 4 > d[3])
2985                                 break;
2986                         /* Prefer others for truncated descriptor */
2987                         if (cur_id_size && d[3] > id_len)
2988                                 break;
2989                         cur_id_size = id_size = d[3];
2990                         cur_id_str = d + 4;
2991                         cur_id_type = d[1] & 0xf;
2992                         if (cur_id_size >= id_len)
2993                                 cur_id_size = id_len - 1;
2994                         memcpy(id, cur_id_str, cur_id_size);
2995                         /* Decrease priority for truncated descriptor */
2996                         if (cur_id_size != id_size)
2997                                 cur_id_size = 6;
2998                         break;
2999                 default:
3000                         break;
3001                 }
3002 next_desig:
3003                 d += d[3] + 4;
3004         }
3005         rcu_read_unlock();
3006
3007         return id_size;
3008 }
3009 EXPORT_SYMBOL(scsi_vpd_lun_id);
3010
3011 /*
3012  * scsi_vpd_tpg_id - return a target port group identifier
3013  * @sdev: SCSI device
3014  *
3015  * Returns the Target Port Group identifier from the information
3016  * froom VPD page 0x83 of the device.
3017  *
3018  * Returns the identifier or error on failure.
3019  */
3020 int scsi_vpd_tpg_id(struct scsi_device *sdev, int *rel_id)
3021 {
3022         const unsigned char *d;
3023         const struct scsi_vpd *vpd_pg83;
3024         int group_id = -EAGAIN, rel_port = -1;
3025
3026         rcu_read_lock();
3027         vpd_pg83 = rcu_dereference(sdev->vpd_pg83);
3028         if (!vpd_pg83) {
3029                 rcu_read_unlock();
3030                 return -ENXIO;
3031         }
3032
3033         d = vpd_pg83->data + 4;
3034         while (d < vpd_pg83->data + vpd_pg83->len) {
3035                 switch (d[1] & 0xf) {
3036                 case 0x4:
3037                         /* Relative target port */
3038                         rel_port = get_unaligned_be16(&d[6]);
3039                         break;
3040                 case 0x5:
3041                         /* Target port group */
3042                         group_id = get_unaligned_be16(&d[6]);
3043                         break;
3044                 default:
3045                         break;
3046                 }
3047                 d += d[3] + 4;
3048         }
3049         rcu_read_unlock();
3050
3051         if (group_id >= 0 && rel_id && rel_port != -1)
3052                 *rel_id = rel_port;
3053
3054         return group_id;
3055 }
3056 EXPORT_SYMBOL(scsi_vpd_tpg_id);