cfg80211: handle failed skb allocation
[cascardo/linux.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_dh.h>
34
35 #include <trace/events/scsi.h>
36
37 #include "scsi_priv.h"
38 #include "scsi_logging.h"
39
40
41 struct kmem_cache *scsi_sdb_cache;
42
43 /*
44  * When to reinvoke queueing after a resource shortage. It's 3 msecs to
45  * not change behaviour from the previous unplug mechanism, experimentation
46  * may prove this needs changing.
47  */
48 #define SCSI_QUEUE_DELAY        3
49
50 static void
51 scsi_set_blocked(struct scsi_cmnd *cmd, int reason)
52 {
53         struct Scsi_Host *host = cmd->device->host;
54         struct scsi_device *device = cmd->device;
55         struct scsi_target *starget = scsi_target(device);
56
57         /*
58          * Set the appropriate busy bit for the device/host.
59          *
60          * If the host/device isn't busy, assume that something actually
61          * completed, and that we should be able to queue a command now.
62          *
63          * Note that the prior mid-layer assumption that any host could
64          * always queue at least one command is now broken.  The mid-layer
65          * will implement a user specifiable stall (see
66          * scsi_host.max_host_blocked and scsi_device.max_device_blocked)
67          * if a command is requeued with no other commands outstanding
68          * either for the device or for the host.
69          */
70         switch (reason) {
71         case SCSI_MLQUEUE_HOST_BUSY:
72                 atomic_set(&host->host_blocked, host->max_host_blocked);
73                 break;
74         case SCSI_MLQUEUE_DEVICE_BUSY:
75         case SCSI_MLQUEUE_EH_RETRY:
76                 atomic_set(&device->device_blocked,
77                            device->max_device_blocked);
78                 break;
79         case SCSI_MLQUEUE_TARGET_BUSY:
80                 atomic_set(&starget->target_blocked,
81                            starget->max_target_blocked);
82                 break;
83         }
84 }
85
86 static void scsi_mq_requeue_cmd(struct scsi_cmnd *cmd)
87 {
88         struct scsi_device *sdev = cmd->device;
89         struct request_queue *q = cmd->request->q;
90
91         blk_mq_requeue_request(cmd->request);
92         blk_mq_kick_requeue_list(q);
93         put_device(&sdev->sdev_gendev);
94 }
95
96 /**
97  * __scsi_queue_insert - private queue insertion
98  * @cmd: The SCSI command being requeued
99  * @reason:  The reason for the requeue
100  * @unbusy: Whether the queue should be unbusied
101  *
102  * This is a private queue insertion.  The public interface
103  * scsi_queue_insert() always assumes the queue should be unbusied
104  * because it's always called before the completion.  This function is
105  * for a requeue after completion, which should only occur in this
106  * file.
107  */
108 static void __scsi_queue_insert(struct scsi_cmnd *cmd, int reason, int unbusy)
109 {
110         struct scsi_device *device = cmd->device;
111         struct request_queue *q = device->request_queue;
112         unsigned long flags;
113
114         SCSI_LOG_MLQUEUE(1, scmd_printk(KERN_INFO, cmd,
115                 "Inserting command %p into mlqueue\n", cmd));
116
117         scsi_set_blocked(cmd, reason);
118
119         /*
120          * Decrement the counters, since these commands are no longer
121          * active on the host/device.
122          */
123         if (unbusy)
124                 scsi_device_unbusy(device);
125
126         /*
127          * Requeue this command.  It will go before all other commands
128          * that are already in the queue. Schedule requeue work under
129          * lock such that the kblockd_schedule_work() call happens
130          * before blk_cleanup_queue() finishes.
131          */
132         cmd->result = 0;
133         if (q->mq_ops) {
134                 scsi_mq_requeue_cmd(cmd);
135                 return;
136         }
137         spin_lock_irqsave(q->queue_lock, flags);
138         blk_requeue_request(q, cmd->request);
139         kblockd_schedule_work(&device->requeue_work);
140         spin_unlock_irqrestore(q->queue_lock, flags);
141 }
142
143 /*
144  * Function:    scsi_queue_insert()
145  *
146  * Purpose:     Insert a command in the midlevel queue.
147  *
148  * Arguments:   cmd    - command that we are adding to queue.
149  *              reason - why we are inserting command to queue.
150  *
151  * Lock status: Assumed that lock is not held upon entry.
152  *
153  * Returns:     Nothing.
154  *
155  * Notes:       We do this for one of two cases.  Either the host is busy
156  *              and it cannot accept any more commands for the time being,
157  *              or the device returned QUEUE_FULL and can accept no more
158  *              commands.
159  * Notes:       This could be called either from an interrupt context or a
160  *              normal process context.
161  */
162 void scsi_queue_insert(struct scsi_cmnd *cmd, int reason)
163 {
164         __scsi_queue_insert(cmd, reason, 1);
165 }
166 /**
167  * scsi_execute - insert request and wait for the result
168  * @sdev:       scsi device
169  * @cmd:        scsi command
170  * @data_direction: data direction
171  * @buffer:     data buffer
172  * @bufflen:    len of buffer
173  * @sense:      optional sense buffer
174  * @timeout:    request timeout in seconds
175  * @retries:    number of times to retry request
176  * @flags:      or into request flags;
177  * @resid:      optional residual length
178  *
179  * returns the req->errors value which is the scsi_cmnd result
180  * field.
181  */
182 int scsi_execute(struct scsi_device *sdev, const unsigned char *cmd,
183                  int data_direction, void *buffer, unsigned bufflen,
184                  unsigned char *sense, int timeout, int retries, u64 flags,
185                  int *resid)
186 {
187         struct request *req;
188         int write = (data_direction == DMA_TO_DEVICE);
189         int ret = DRIVER_ERROR << 24;
190
191         req = blk_get_request(sdev->request_queue, write, __GFP_RECLAIM);
192         if (IS_ERR(req))
193                 return ret;
194         blk_rq_set_block_pc(req);
195
196         if (bufflen &&  blk_rq_map_kern(sdev->request_queue, req,
197                                         buffer, bufflen, __GFP_RECLAIM))
198                 goto out;
199
200         req->cmd_len = COMMAND_SIZE(cmd[0]);
201         memcpy(req->cmd, cmd, req->cmd_len);
202         req->sense = sense;
203         req->sense_len = 0;
204         req->retries = retries;
205         req->timeout = timeout;
206         req->cmd_flags |= flags | REQ_QUIET | REQ_PREEMPT;
207
208         /*
209          * head injection *required* here otherwise quiesce won't work
210          */
211         blk_execute_rq(req->q, NULL, req, 1);
212
213         /*
214          * Some devices (USB mass-storage in particular) may transfer
215          * garbage data together with a residue indicating that the data
216          * is invalid.  Prevent the garbage from being misinterpreted
217          * and prevent security leaks by zeroing out the excess data.
218          */
219         if (unlikely(req->resid_len > 0 && req->resid_len <= bufflen))
220                 memset(buffer + (bufflen - req->resid_len), 0, req->resid_len);
221
222         if (resid)
223                 *resid = req->resid_len;
224         ret = req->errors;
225  out:
226         blk_put_request(req);
227
228         return ret;
229 }
230 EXPORT_SYMBOL(scsi_execute);
231
232 int scsi_execute_req_flags(struct scsi_device *sdev, const unsigned char *cmd,
233                      int data_direction, void *buffer, unsigned bufflen,
234                      struct scsi_sense_hdr *sshdr, int timeout, int retries,
235                      int *resid, u64 flags)
236 {
237         char *sense = NULL;
238         int result;
239         
240         if (sshdr) {
241                 sense = kzalloc(SCSI_SENSE_BUFFERSIZE, GFP_NOIO);
242                 if (!sense)
243                         return DRIVER_ERROR << 24;
244         }
245         result = scsi_execute(sdev, cmd, data_direction, buffer, bufflen,
246                               sense, timeout, retries, flags, resid);
247         if (sshdr)
248                 scsi_normalize_sense(sense, SCSI_SENSE_BUFFERSIZE, sshdr);
249
250         kfree(sense);
251         return result;
252 }
253 EXPORT_SYMBOL(scsi_execute_req_flags);
254
255 /*
256  * Function:    scsi_init_cmd_errh()
257  *
258  * Purpose:     Initialize cmd fields related to error handling.
259  *
260  * Arguments:   cmd     - command that is ready to be queued.
261  *
262  * Notes:       This function has the job of initializing a number of
263  *              fields related to error handling.   Typically this will
264  *              be called once for each command, as required.
265  */
266 static void scsi_init_cmd_errh(struct scsi_cmnd *cmd)
267 {
268         cmd->serial_number = 0;
269         scsi_set_resid(cmd, 0);
270         memset(cmd->sense_buffer, 0, SCSI_SENSE_BUFFERSIZE);
271         if (cmd->cmd_len == 0)
272                 cmd->cmd_len = scsi_command_size(cmd->cmnd);
273 }
274
275 void scsi_device_unbusy(struct scsi_device *sdev)
276 {
277         struct Scsi_Host *shost = sdev->host;
278         struct scsi_target *starget = scsi_target(sdev);
279         unsigned long flags;
280
281         atomic_dec(&shost->host_busy);
282         if (starget->can_queue > 0)
283                 atomic_dec(&starget->target_busy);
284
285         if (unlikely(scsi_host_in_recovery(shost) &&
286                      (shost->host_failed || shost->host_eh_scheduled))) {
287                 spin_lock_irqsave(shost->host_lock, flags);
288                 scsi_eh_wakeup(shost);
289                 spin_unlock_irqrestore(shost->host_lock, flags);
290         }
291
292         atomic_dec(&sdev->device_busy);
293 }
294
295 static void scsi_kick_queue(struct request_queue *q)
296 {
297         if (q->mq_ops)
298                 blk_mq_start_hw_queues(q);
299         else
300                 blk_run_queue(q);
301 }
302
303 /*
304  * Called for single_lun devices on IO completion. Clear starget_sdev_user,
305  * and call blk_run_queue for all the scsi_devices on the target -
306  * including current_sdev first.
307  *
308  * Called with *no* scsi locks held.
309  */
310 static void scsi_single_lun_run(struct scsi_device *current_sdev)
311 {
312         struct Scsi_Host *shost = current_sdev->host;
313         struct scsi_device *sdev, *tmp;
314         struct scsi_target *starget = scsi_target(current_sdev);
315         unsigned long flags;
316
317         spin_lock_irqsave(shost->host_lock, flags);
318         starget->starget_sdev_user = NULL;
319         spin_unlock_irqrestore(shost->host_lock, flags);
320
321         /*
322          * Call blk_run_queue for all LUNs on the target, starting with
323          * current_sdev. We race with others (to set starget_sdev_user),
324          * but in most cases, we will be first. Ideally, each LU on the
325          * target would get some limited time or requests on the target.
326          */
327         scsi_kick_queue(current_sdev->request_queue);
328
329         spin_lock_irqsave(shost->host_lock, flags);
330         if (starget->starget_sdev_user)
331                 goto out;
332         list_for_each_entry_safe(sdev, tmp, &starget->devices,
333                         same_target_siblings) {
334                 if (sdev == current_sdev)
335                         continue;
336                 if (scsi_device_get(sdev))
337                         continue;
338
339                 spin_unlock_irqrestore(shost->host_lock, flags);
340                 scsi_kick_queue(sdev->request_queue);
341                 spin_lock_irqsave(shost->host_lock, flags);
342         
343                 scsi_device_put(sdev);
344         }
345  out:
346         spin_unlock_irqrestore(shost->host_lock, flags);
347 }
348
349 static inline bool scsi_device_is_busy(struct scsi_device *sdev)
350 {
351         if (atomic_read(&sdev->device_busy) >= sdev->queue_depth)
352                 return true;
353         if (atomic_read(&sdev->device_blocked) > 0)
354                 return true;
355         return false;
356 }
357
358 static inline bool scsi_target_is_busy(struct scsi_target *starget)
359 {
360         if (starget->can_queue > 0) {
361                 if (atomic_read(&starget->target_busy) >= starget->can_queue)
362                         return true;
363                 if (atomic_read(&starget->target_blocked) > 0)
364                         return true;
365         }
366         return false;
367 }
368
369 static inline bool scsi_host_is_busy(struct Scsi_Host *shost)
370 {
371         if (shost->can_queue > 0 &&
372             atomic_read(&shost->host_busy) >= shost->can_queue)
373                 return true;
374         if (atomic_read(&shost->host_blocked) > 0)
375                 return true;
376         if (shost->host_self_blocked)
377                 return true;
378         return false;
379 }
380
381 static void scsi_starved_list_run(struct Scsi_Host *shost)
382 {
383         LIST_HEAD(starved_list);
384         struct scsi_device *sdev;
385         unsigned long flags;
386
387         spin_lock_irqsave(shost->host_lock, flags);
388         list_splice_init(&shost->starved_list, &starved_list);
389
390         while (!list_empty(&starved_list)) {
391                 struct request_queue *slq;
392
393                 /*
394                  * As long as shost is accepting commands and we have
395                  * starved queues, call blk_run_queue. scsi_request_fn
396                  * drops the queue_lock and can add us back to the
397                  * starved_list.
398                  *
399                  * host_lock protects the starved_list and starved_entry.
400                  * scsi_request_fn must get the host_lock before checking
401                  * or modifying starved_list or starved_entry.
402                  */
403                 if (scsi_host_is_busy(shost))
404                         break;
405
406                 sdev = list_entry(starved_list.next,
407                                   struct scsi_device, starved_entry);
408                 list_del_init(&sdev->starved_entry);
409                 if (scsi_target_is_busy(scsi_target(sdev))) {
410                         list_move_tail(&sdev->starved_entry,
411                                        &shost->starved_list);
412                         continue;
413                 }
414
415                 /*
416                  * Once we drop the host lock, a racing scsi_remove_device()
417                  * call may remove the sdev from the starved list and destroy
418                  * it and the queue.  Mitigate by taking a reference to the
419                  * queue and never touching the sdev again after we drop the
420                  * host lock.  Note: if __scsi_remove_device() invokes
421                  * blk_cleanup_queue() before the queue is run from this
422                  * function then blk_run_queue() will return immediately since
423                  * blk_cleanup_queue() marks the queue with QUEUE_FLAG_DYING.
424                  */
425                 slq = sdev->request_queue;
426                 if (!blk_get_queue(slq))
427                         continue;
428                 spin_unlock_irqrestore(shost->host_lock, flags);
429
430                 scsi_kick_queue(slq);
431                 blk_put_queue(slq);
432
433                 spin_lock_irqsave(shost->host_lock, flags);
434         }
435         /* put any unprocessed entries back */
436         list_splice(&starved_list, &shost->starved_list);
437         spin_unlock_irqrestore(shost->host_lock, flags);
438 }
439
440 /*
441  * Function:   scsi_run_queue()
442  *
443  * Purpose:    Select a proper request queue to serve next
444  *
445  * Arguments:  q       - last request's queue
446  *
447  * Returns:     Nothing
448  *
449  * Notes:      The previous command was completely finished, start
450  *             a new one if possible.
451  */
452 static void scsi_run_queue(struct request_queue *q)
453 {
454         struct scsi_device *sdev = q->queuedata;
455
456         if (scsi_target(sdev)->single_lun)
457                 scsi_single_lun_run(sdev);
458         if (!list_empty(&sdev->host->starved_list))
459                 scsi_starved_list_run(sdev->host);
460
461         if (q->mq_ops)
462                 blk_mq_start_stopped_hw_queues(q, false);
463         else
464                 blk_run_queue(q);
465 }
466
467 void scsi_requeue_run_queue(struct work_struct *work)
468 {
469         struct scsi_device *sdev;
470         struct request_queue *q;
471
472         sdev = container_of(work, struct scsi_device, requeue_work);
473         q = sdev->request_queue;
474         scsi_run_queue(q);
475 }
476
477 /*
478  * Function:    scsi_requeue_command()
479  *
480  * Purpose:     Handle post-processing of completed commands.
481  *
482  * Arguments:   q       - queue to operate on
483  *              cmd     - command that may need to be requeued.
484  *
485  * Returns:     Nothing
486  *
487  * Notes:       After command completion, there may be blocks left
488  *              over which weren't finished by the previous command
489  *              this can be for a number of reasons - the main one is
490  *              I/O errors in the middle of the request, in which case
491  *              we need to request the blocks that come after the bad
492  *              sector.
493  * Notes:       Upon return, cmd is a stale pointer.
494  */
495 static void scsi_requeue_command(struct request_queue *q, struct scsi_cmnd *cmd)
496 {
497         struct scsi_device *sdev = cmd->device;
498         struct request *req = cmd->request;
499         unsigned long flags;
500
501         spin_lock_irqsave(q->queue_lock, flags);
502         blk_unprep_request(req);
503         req->special = NULL;
504         scsi_put_command(cmd);
505         blk_requeue_request(q, req);
506         spin_unlock_irqrestore(q->queue_lock, flags);
507
508         scsi_run_queue(q);
509
510         put_device(&sdev->sdev_gendev);
511 }
512
513 void scsi_run_host_queues(struct Scsi_Host *shost)
514 {
515         struct scsi_device *sdev;
516
517         shost_for_each_device(sdev, shost)
518                 scsi_run_queue(sdev->request_queue);
519 }
520
521 static void scsi_uninit_cmd(struct scsi_cmnd *cmd)
522 {
523         if (cmd->request->cmd_type == REQ_TYPE_FS) {
524                 struct scsi_driver *drv = scsi_cmd_to_driver(cmd);
525
526                 if (drv->uninit_command)
527                         drv->uninit_command(cmd);
528         }
529 }
530
531 static void scsi_mq_free_sgtables(struct scsi_cmnd *cmd)
532 {
533         struct scsi_data_buffer *sdb;
534
535         if (cmd->sdb.table.nents)
536                 sg_free_table_chained(&cmd->sdb.table, true);
537         if (cmd->request->next_rq) {
538                 sdb = cmd->request->next_rq->special;
539                 if (sdb)
540                         sg_free_table_chained(&sdb->table, true);
541         }
542         if (scsi_prot_sg_count(cmd))
543                 sg_free_table_chained(&cmd->prot_sdb->table, true);
544 }
545
546 static void scsi_mq_uninit_cmd(struct scsi_cmnd *cmd)
547 {
548         struct scsi_device *sdev = cmd->device;
549         struct Scsi_Host *shost = sdev->host;
550         unsigned long flags;
551
552         scsi_mq_free_sgtables(cmd);
553         scsi_uninit_cmd(cmd);
554
555         if (shost->use_cmd_list) {
556                 BUG_ON(list_empty(&cmd->list));
557                 spin_lock_irqsave(&sdev->list_lock, flags);
558                 list_del_init(&cmd->list);
559                 spin_unlock_irqrestore(&sdev->list_lock, flags);
560         }
561 }
562
563 /*
564  * Function:    scsi_release_buffers()
565  *
566  * Purpose:     Free resources allocate for a scsi_command.
567  *
568  * Arguments:   cmd     - command that we are bailing.
569  *
570  * Lock status: Assumed that no lock is held upon entry.
571  *
572  * Returns:     Nothing
573  *
574  * Notes:       In the event that an upper level driver rejects a
575  *              command, we must release resources allocated during
576  *              the __init_io() function.  Primarily this would involve
577  *              the scatter-gather table.
578  */
579 static void scsi_release_buffers(struct scsi_cmnd *cmd)
580 {
581         if (cmd->sdb.table.nents)
582                 sg_free_table_chained(&cmd->sdb.table, false);
583
584         memset(&cmd->sdb, 0, sizeof(cmd->sdb));
585
586         if (scsi_prot_sg_count(cmd))
587                 sg_free_table_chained(&cmd->prot_sdb->table, false);
588 }
589
590 static void scsi_release_bidi_buffers(struct scsi_cmnd *cmd)
591 {
592         struct scsi_data_buffer *bidi_sdb = cmd->request->next_rq->special;
593
594         sg_free_table_chained(&bidi_sdb->table, false);
595         kmem_cache_free(scsi_sdb_cache, bidi_sdb);
596         cmd->request->next_rq->special = NULL;
597 }
598
599 static bool scsi_end_request(struct request *req, int error,
600                 unsigned int bytes, unsigned int bidi_bytes)
601 {
602         struct scsi_cmnd *cmd = req->special;
603         struct scsi_device *sdev = cmd->device;
604         struct request_queue *q = sdev->request_queue;
605
606         if (blk_update_request(req, error, bytes))
607                 return true;
608
609         /* Bidi request must be completed as a whole */
610         if (unlikely(bidi_bytes) &&
611             blk_update_request(req->next_rq, error, bidi_bytes))
612                 return true;
613
614         if (blk_queue_add_random(q))
615                 add_disk_randomness(req->rq_disk);
616
617         if (req->mq_ctx) {
618                 /*
619                  * In the MQ case the command gets freed by __blk_mq_end_request,
620                  * so we have to do all cleanup that depends on it earlier.
621                  *
622                  * We also can't kick the queues from irq context, so we
623                  * will have to defer it to a workqueue.
624                  */
625                 scsi_mq_uninit_cmd(cmd);
626
627                 __blk_mq_end_request(req, error);
628
629                 if (scsi_target(sdev)->single_lun ||
630                     !list_empty(&sdev->host->starved_list))
631                         kblockd_schedule_work(&sdev->requeue_work);
632                 else
633                         blk_mq_start_stopped_hw_queues(q, true);
634         } else {
635                 unsigned long flags;
636
637                 if (bidi_bytes)
638                         scsi_release_bidi_buffers(cmd);
639
640                 spin_lock_irqsave(q->queue_lock, flags);
641                 blk_finish_request(req, error);
642                 spin_unlock_irqrestore(q->queue_lock, flags);
643
644                 scsi_release_buffers(cmd);
645
646                 scsi_put_command(cmd);
647                 scsi_run_queue(q);
648         }
649
650         put_device(&sdev->sdev_gendev);
651         return false;
652 }
653
654 /**
655  * __scsi_error_from_host_byte - translate SCSI error code into errno
656  * @cmd:        SCSI command (unused)
657  * @result:     scsi error code
658  *
659  * Translate SCSI error code into standard UNIX errno.
660  * Return values:
661  * -ENOLINK     temporary transport failure
662  * -EREMOTEIO   permanent target failure, do not retry
663  * -EBADE       permanent nexus failure, retry on other path
664  * -ENOSPC      No write space available
665  * -ENODATA     Medium error
666  * -EIO         unspecified I/O error
667  */
668 static int __scsi_error_from_host_byte(struct scsi_cmnd *cmd, int result)
669 {
670         int error = 0;
671
672         switch(host_byte(result)) {
673         case DID_TRANSPORT_FAILFAST:
674                 error = -ENOLINK;
675                 break;
676         case DID_TARGET_FAILURE:
677                 set_host_byte(cmd, DID_OK);
678                 error = -EREMOTEIO;
679                 break;
680         case DID_NEXUS_FAILURE:
681                 set_host_byte(cmd, DID_OK);
682                 error = -EBADE;
683                 break;
684         case DID_ALLOC_FAILURE:
685                 set_host_byte(cmd, DID_OK);
686                 error = -ENOSPC;
687                 break;
688         case DID_MEDIUM_ERROR:
689                 set_host_byte(cmd, DID_OK);
690                 error = -ENODATA;
691                 break;
692         default:
693                 error = -EIO;
694                 break;
695         }
696
697         return error;
698 }
699
700 /*
701  * Function:    scsi_io_completion()
702  *
703  * Purpose:     Completion processing for block device I/O requests.
704  *
705  * Arguments:   cmd   - command that is finished.
706  *
707  * Lock status: Assumed that no lock is held upon entry.
708  *
709  * Returns:     Nothing
710  *
711  * Notes:       We will finish off the specified number of sectors.  If we
712  *              are done, the command block will be released and the queue
713  *              function will be goosed.  If we are not done then we have to
714  *              figure out what to do next:
715  *
716  *              a) We can call scsi_requeue_command().  The request
717  *                 will be unprepared and put back on the queue.  Then
718  *                 a new command will be created for it.  This should
719  *                 be used if we made forward progress, or if we want
720  *                 to switch from READ(10) to READ(6) for example.
721  *
722  *              b) We can call __scsi_queue_insert().  The request will
723  *                 be put back on the queue and retried using the same
724  *                 command as before, possibly after a delay.
725  *
726  *              c) We can call scsi_end_request() with -EIO to fail
727  *                 the remainder of the request.
728  */
729 void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes)
730 {
731         int result = cmd->result;
732         struct request_queue *q = cmd->device->request_queue;
733         struct request *req = cmd->request;
734         int error = 0;
735         struct scsi_sense_hdr sshdr;
736         bool sense_valid = false;
737         int sense_deferred = 0, level = 0;
738         enum {ACTION_FAIL, ACTION_REPREP, ACTION_RETRY,
739               ACTION_DELAYED_RETRY} action;
740         unsigned long wait_for = (cmd->allowed + 1) * req->timeout;
741
742         if (result) {
743                 sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
744                 if (sense_valid)
745                         sense_deferred = scsi_sense_is_deferred(&sshdr);
746         }
747
748         if (req->cmd_type == REQ_TYPE_BLOCK_PC) { /* SG_IO ioctl from block level */
749                 if (result) {
750                         if (sense_valid && req->sense) {
751                                 /*
752                                  * SG_IO wants current and deferred errors
753                                  */
754                                 int len = 8 + cmd->sense_buffer[7];
755
756                                 if (len > SCSI_SENSE_BUFFERSIZE)
757                                         len = SCSI_SENSE_BUFFERSIZE;
758                                 memcpy(req->sense, cmd->sense_buffer,  len);
759                                 req->sense_len = len;
760                         }
761                         if (!sense_deferred)
762                                 error = __scsi_error_from_host_byte(cmd, result);
763                 }
764                 /*
765                  * __scsi_error_from_host_byte may have reset the host_byte
766                  */
767                 req->errors = cmd->result;
768
769                 req->resid_len = scsi_get_resid(cmd);
770
771                 if (scsi_bidi_cmnd(cmd)) {
772                         /*
773                          * Bidi commands Must be complete as a whole,
774                          * both sides at once.
775                          */
776                         req->next_rq->resid_len = scsi_in(cmd)->resid;
777                         if (scsi_end_request(req, 0, blk_rq_bytes(req),
778                                         blk_rq_bytes(req->next_rq)))
779                                 BUG();
780                         return;
781                 }
782         } else if (blk_rq_bytes(req) == 0 && result && !sense_deferred) {
783                 /*
784                  * Certain non BLOCK_PC requests are commands that don't
785                  * actually transfer anything (FLUSH), so cannot use
786                  * good_bytes != blk_rq_bytes(req) as the signal for an error.
787                  * This sets the error explicitly for the problem case.
788                  */
789                 error = __scsi_error_from_host_byte(cmd, result);
790         }
791
792         /* no bidi support for !REQ_TYPE_BLOCK_PC yet */
793         BUG_ON(blk_bidi_rq(req));
794
795         /*
796          * Next deal with any sectors which we were able to correctly
797          * handle.
798          */
799         SCSI_LOG_HLCOMPLETE(1, scmd_printk(KERN_INFO, cmd,
800                 "%u sectors total, %d bytes done.\n",
801                 blk_rq_sectors(req), good_bytes));
802
803         /*
804          * Recovered errors need reporting, but they're always treated
805          * as success, so fiddle the result code here.  For BLOCK_PC
806          * we already took a copy of the original into rq->errors which
807          * is what gets returned to the user
808          */
809         if (sense_valid && (sshdr.sense_key == RECOVERED_ERROR)) {
810                 /* if ATA PASS-THROUGH INFORMATION AVAILABLE skip
811                  * print since caller wants ATA registers. Only occurs on
812                  * SCSI ATA PASS_THROUGH commands when CK_COND=1
813                  */
814                 if ((sshdr.asc == 0x0) && (sshdr.ascq == 0x1d))
815                         ;
816                 else if (!(req->cmd_flags & REQ_QUIET))
817                         scsi_print_sense(cmd);
818                 result = 0;
819                 /* BLOCK_PC may have set error */
820                 error = 0;
821         }
822
823         /*
824          * If we finished all bytes in the request we are done now.
825          */
826         if (!scsi_end_request(req, error, good_bytes, 0))
827                 return;
828
829         /*
830          * Kill remainder if no retrys.
831          */
832         if (error && scsi_noretry_cmd(cmd)) {
833                 if (scsi_end_request(req, error, blk_rq_bytes(req), 0))
834                         BUG();
835                 return;
836         }
837
838         /*
839          * If there had been no error, but we have leftover bytes in the
840          * requeues just queue the command up again.
841          */
842         if (result == 0)
843                 goto requeue;
844
845         error = __scsi_error_from_host_byte(cmd, result);
846
847         if (host_byte(result) == DID_RESET) {
848                 /* Third party bus reset or reset for error recovery
849                  * reasons.  Just retry the command and see what
850                  * happens.
851                  */
852                 action = ACTION_RETRY;
853         } else if (sense_valid && !sense_deferred) {
854                 switch (sshdr.sense_key) {
855                 case UNIT_ATTENTION:
856                         if (cmd->device->removable) {
857                                 /* Detected disc change.  Set a bit
858                                  * and quietly refuse further access.
859                                  */
860                                 cmd->device->changed = 1;
861                                 action = ACTION_FAIL;
862                         } else {
863                                 /* Must have been a power glitch, or a
864                                  * bus reset.  Could not have been a
865                                  * media change, so we just retry the
866                                  * command and see what happens.
867                                  */
868                                 action = ACTION_RETRY;
869                         }
870                         break;
871                 case ILLEGAL_REQUEST:
872                         /* If we had an ILLEGAL REQUEST returned, then
873                          * we may have performed an unsupported
874                          * command.  The only thing this should be
875                          * would be a ten byte read where only a six
876                          * byte read was supported.  Also, on a system
877                          * where READ CAPACITY failed, we may have
878                          * read past the end of the disk.
879                          */
880                         if ((cmd->device->use_10_for_rw &&
881                             sshdr.asc == 0x20 && sshdr.ascq == 0x00) &&
882                             (cmd->cmnd[0] == READ_10 ||
883                              cmd->cmnd[0] == WRITE_10)) {
884                                 /* This will issue a new 6-byte command. */
885                                 cmd->device->use_10_for_rw = 0;
886                                 action = ACTION_REPREP;
887                         } else if (sshdr.asc == 0x10) /* DIX */ {
888                                 action = ACTION_FAIL;
889                                 error = -EILSEQ;
890                         /* INVALID COMMAND OPCODE or INVALID FIELD IN CDB */
891                         } else if (sshdr.asc == 0x20 || sshdr.asc == 0x24) {
892                                 action = ACTION_FAIL;
893                                 error = -EREMOTEIO;
894                         } else
895                                 action = ACTION_FAIL;
896                         break;
897                 case ABORTED_COMMAND:
898                         action = ACTION_FAIL;
899                         if (sshdr.asc == 0x10) /* DIF */
900                                 error = -EILSEQ;
901                         break;
902                 case NOT_READY:
903                         /* If the device is in the process of becoming
904                          * ready, or has a temporary blockage, retry.
905                          */
906                         if (sshdr.asc == 0x04) {
907                                 switch (sshdr.ascq) {
908                                 case 0x01: /* becoming ready */
909                                 case 0x04: /* format in progress */
910                                 case 0x05: /* rebuild in progress */
911                                 case 0x06: /* recalculation in progress */
912                                 case 0x07: /* operation in progress */
913                                 case 0x08: /* Long write in progress */
914                                 case 0x09: /* self test in progress */
915                                 case 0x14: /* space allocation in progress */
916                                         action = ACTION_DELAYED_RETRY;
917                                         break;
918                                 default:
919                                         action = ACTION_FAIL;
920                                         break;
921                                 }
922                         } else
923                                 action = ACTION_FAIL;
924                         break;
925                 case VOLUME_OVERFLOW:
926                         /* See SSC3rXX or current. */
927                         action = ACTION_FAIL;
928                         break;
929                 default:
930                         action = ACTION_FAIL;
931                         break;
932                 }
933         } else
934                 action = ACTION_FAIL;
935
936         if (action != ACTION_FAIL &&
937             time_before(cmd->jiffies_at_alloc + wait_for, jiffies))
938                 action = ACTION_FAIL;
939
940         switch (action) {
941         case ACTION_FAIL:
942                 /* Give up and fail the remainder of the request */
943                 if (!(req->cmd_flags & REQ_QUIET)) {
944                         static DEFINE_RATELIMIT_STATE(_rs,
945                                         DEFAULT_RATELIMIT_INTERVAL,
946                                         DEFAULT_RATELIMIT_BURST);
947
948                         if (unlikely(scsi_logging_level))
949                                 level = SCSI_LOG_LEVEL(SCSI_LOG_MLCOMPLETE_SHIFT,
950                                                        SCSI_LOG_MLCOMPLETE_BITS);
951
952                         /*
953                          * if logging is enabled the failure will be printed
954                          * in scsi_log_completion(), so avoid duplicate messages
955                          */
956                         if (!level && __ratelimit(&_rs)) {
957                                 scsi_print_result(cmd, NULL, FAILED);
958                                 if (driver_byte(result) & DRIVER_SENSE)
959                                         scsi_print_sense(cmd);
960                                 scsi_print_command(cmd);
961                         }
962                 }
963                 if (!scsi_end_request(req, error, blk_rq_err_bytes(req), 0))
964                         return;
965                 /*FALLTHRU*/
966         case ACTION_REPREP:
967         requeue:
968                 /* Unprep the request and put it back at the head of the queue.
969                  * A new command will be prepared and issued.
970                  */
971                 if (q->mq_ops) {
972                         cmd->request->cmd_flags &= ~REQ_DONTPREP;
973                         scsi_mq_uninit_cmd(cmd);
974                         scsi_mq_requeue_cmd(cmd);
975                 } else {
976                         scsi_release_buffers(cmd);
977                         scsi_requeue_command(q, cmd);
978                 }
979                 break;
980         case ACTION_RETRY:
981                 /* Retry the same command immediately */
982                 __scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY, 0);
983                 break;
984         case ACTION_DELAYED_RETRY:
985                 /* Retry the same command after a delay */
986                 __scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY, 0);
987                 break;
988         }
989 }
990
991 static int scsi_init_sgtable(struct request *req, struct scsi_data_buffer *sdb)
992 {
993         int count;
994
995         /*
996          * If sg table allocation fails, requeue request later.
997          */
998         if (unlikely(sg_alloc_table_chained(&sdb->table, req->nr_phys_segments,
999                                         sdb->table.sgl)))
1000                 return BLKPREP_DEFER;
1001
1002         /* 
1003          * Next, walk the list, and fill in the addresses and sizes of
1004          * each segment.
1005          */
1006         count = blk_rq_map_sg(req->q, req, sdb->table.sgl);
1007         BUG_ON(count > sdb->table.nents);
1008         sdb->table.nents = count;
1009         sdb->length = blk_rq_bytes(req);
1010         return BLKPREP_OK;
1011 }
1012
1013 /*
1014  * Function:    scsi_init_io()
1015  *
1016  * Purpose:     SCSI I/O initialize function.
1017  *
1018  * Arguments:   cmd   - Command descriptor we wish to initialize
1019  *
1020  * Returns:     0 on success
1021  *              BLKPREP_DEFER if the failure is retryable
1022  *              BLKPREP_KILL if the failure is fatal
1023  */
1024 int scsi_init_io(struct scsi_cmnd *cmd)
1025 {
1026         struct scsi_device *sdev = cmd->device;
1027         struct request *rq = cmd->request;
1028         bool is_mq = (rq->mq_ctx != NULL);
1029         int error;
1030
1031         BUG_ON(!rq->nr_phys_segments);
1032
1033         error = scsi_init_sgtable(rq, &cmd->sdb);
1034         if (error)
1035                 goto err_exit;
1036
1037         if (blk_bidi_rq(rq)) {
1038                 if (!rq->q->mq_ops) {
1039                         struct scsi_data_buffer *bidi_sdb =
1040                                 kmem_cache_zalloc(scsi_sdb_cache, GFP_ATOMIC);
1041                         if (!bidi_sdb) {
1042                                 error = BLKPREP_DEFER;
1043                                 goto err_exit;
1044                         }
1045
1046                         rq->next_rq->special = bidi_sdb;
1047                 }
1048
1049                 error = scsi_init_sgtable(rq->next_rq, rq->next_rq->special);
1050                 if (error)
1051                         goto err_exit;
1052         }
1053
1054         if (blk_integrity_rq(rq)) {
1055                 struct scsi_data_buffer *prot_sdb = cmd->prot_sdb;
1056                 int ivecs, count;
1057
1058                 if (prot_sdb == NULL) {
1059                         /*
1060                          * This can happen if someone (e.g. multipath)
1061                          * queues a command to a device on an adapter
1062                          * that does not support DIX.
1063                          */
1064                         WARN_ON_ONCE(1);
1065                         error = BLKPREP_KILL;
1066                         goto err_exit;
1067                 }
1068
1069                 ivecs = blk_rq_count_integrity_sg(rq->q, rq->bio);
1070
1071                 if (sg_alloc_table_chained(&prot_sdb->table, ivecs,
1072                                 prot_sdb->table.sgl)) {
1073                         error = BLKPREP_DEFER;
1074                         goto err_exit;
1075                 }
1076
1077                 count = blk_rq_map_integrity_sg(rq->q, rq->bio,
1078                                                 prot_sdb->table.sgl);
1079                 BUG_ON(unlikely(count > ivecs));
1080                 BUG_ON(unlikely(count > queue_max_integrity_segments(rq->q)));
1081
1082                 cmd->prot_sdb = prot_sdb;
1083                 cmd->prot_sdb->table.nents = count;
1084         }
1085
1086         return BLKPREP_OK;
1087 err_exit:
1088         if (is_mq) {
1089                 scsi_mq_free_sgtables(cmd);
1090         } else {
1091                 scsi_release_buffers(cmd);
1092                 cmd->request->special = NULL;
1093                 scsi_put_command(cmd);
1094                 put_device(&sdev->sdev_gendev);
1095         }
1096         return error;
1097 }
1098 EXPORT_SYMBOL(scsi_init_io);
1099
1100 static struct scsi_cmnd *scsi_get_cmd_from_req(struct scsi_device *sdev,
1101                 struct request *req)
1102 {
1103         struct scsi_cmnd *cmd;
1104
1105         if (!req->special) {
1106                 /* Bail if we can't get a reference to the device */
1107                 if (!get_device(&sdev->sdev_gendev))
1108                         return NULL;
1109
1110                 cmd = scsi_get_command(sdev, GFP_ATOMIC);
1111                 if (unlikely(!cmd)) {
1112                         put_device(&sdev->sdev_gendev);
1113                         return NULL;
1114                 }
1115                 req->special = cmd;
1116         } else {
1117                 cmd = req->special;
1118         }
1119
1120         /* pull a tag out of the request if we have one */
1121         cmd->tag = req->tag;
1122         cmd->request = req;
1123
1124         cmd->cmnd = req->cmd;
1125         cmd->prot_op = SCSI_PROT_NORMAL;
1126
1127         return cmd;
1128 }
1129
1130 static int scsi_setup_blk_pc_cmnd(struct scsi_device *sdev, struct request *req)
1131 {
1132         struct scsi_cmnd *cmd = req->special;
1133
1134         /*
1135          * BLOCK_PC requests may transfer data, in which case they must
1136          * a bio attached to them.  Or they might contain a SCSI command
1137          * that does not transfer data, in which case they may optionally
1138          * submit a request without an attached bio.
1139          */
1140         if (req->bio) {
1141                 int ret = scsi_init_io(cmd);
1142                 if (unlikely(ret))
1143                         return ret;
1144         } else {
1145                 BUG_ON(blk_rq_bytes(req));
1146
1147                 memset(&cmd->sdb, 0, sizeof(cmd->sdb));
1148         }
1149
1150         cmd->cmd_len = req->cmd_len;
1151         cmd->transfersize = blk_rq_bytes(req);
1152         cmd->allowed = req->retries;
1153         return BLKPREP_OK;
1154 }
1155
1156 /*
1157  * Setup a REQ_TYPE_FS command.  These are simple request from filesystems
1158  * that still need to be translated to SCSI CDBs from the ULD.
1159  */
1160 static int scsi_setup_fs_cmnd(struct scsi_device *sdev, struct request *req)
1161 {
1162         struct scsi_cmnd *cmd = req->special;
1163
1164         if (unlikely(sdev->handler && sdev->handler->prep_fn)) {
1165                 int ret = sdev->handler->prep_fn(sdev, req);
1166                 if (ret != BLKPREP_OK)
1167                         return ret;
1168         }
1169
1170         memset(cmd->cmnd, 0, BLK_MAX_CDB);
1171         return scsi_cmd_to_driver(cmd)->init_command(cmd);
1172 }
1173
1174 static int scsi_setup_cmnd(struct scsi_device *sdev, struct request *req)
1175 {
1176         struct scsi_cmnd *cmd = req->special;
1177
1178         if (!blk_rq_bytes(req))
1179                 cmd->sc_data_direction = DMA_NONE;
1180         else if (rq_data_dir(req) == WRITE)
1181                 cmd->sc_data_direction = DMA_TO_DEVICE;
1182         else
1183                 cmd->sc_data_direction = DMA_FROM_DEVICE;
1184
1185         switch (req->cmd_type) {
1186         case REQ_TYPE_FS:
1187                 return scsi_setup_fs_cmnd(sdev, req);
1188         case REQ_TYPE_BLOCK_PC:
1189                 return scsi_setup_blk_pc_cmnd(sdev, req);
1190         default:
1191                 return BLKPREP_KILL;
1192         }
1193 }
1194
1195 static int
1196 scsi_prep_state_check(struct scsi_device *sdev, struct request *req)
1197 {
1198         int ret = BLKPREP_OK;
1199
1200         /*
1201          * If the device is not in running state we will reject some
1202          * or all commands.
1203          */
1204         if (unlikely(sdev->sdev_state != SDEV_RUNNING)) {
1205                 switch (sdev->sdev_state) {
1206                 case SDEV_OFFLINE:
1207                 case SDEV_TRANSPORT_OFFLINE:
1208                         /*
1209                          * If the device is offline we refuse to process any
1210                          * commands.  The device must be brought online
1211                          * before trying any recovery commands.
1212                          */
1213                         sdev_printk(KERN_ERR, sdev,
1214                                     "rejecting I/O to offline device\n");
1215                         ret = BLKPREP_KILL;
1216                         break;
1217                 case SDEV_DEL:
1218                         /*
1219                          * If the device is fully deleted, we refuse to
1220                          * process any commands as well.
1221                          */
1222                         sdev_printk(KERN_ERR, sdev,
1223                                     "rejecting I/O to dead device\n");
1224                         ret = BLKPREP_KILL;
1225                         break;
1226                 case SDEV_BLOCK:
1227                 case SDEV_CREATED_BLOCK:
1228                         ret = BLKPREP_DEFER;
1229                         break;
1230                 case SDEV_QUIESCE:
1231                         /*
1232                          * If the devices is blocked we defer normal commands.
1233                          */
1234                         if (!(req->cmd_flags & REQ_PREEMPT))
1235                                 ret = BLKPREP_DEFER;
1236                         break;
1237                 default:
1238                         /*
1239                          * For any other not fully online state we only allow
1240                          * special commands.  In particular any user initiated
1241                          * command is not allowed.
1242                          */
1243                         if (!(req->cmd_flags & REQ_PREEMPT))
1244                                 ret = BLKPREP_KILL;
1245                         break;
1246                 }
1247         }
1248         return ret;
1249 }
1250
1251 static int
1252 scsi_prep_return(struct request_queue *q, struct request *req, int ret)
1253 {
1254         struct scsi_device *sdev = q->queuedata;
1255
1256         switch (ret) {
1257         case BLKPREP_KILL:
1258         case BLKPREP_INVALID:
1259                 req->errors = DID_NO_CONNECT << 16;
1260                 /* release the command and kill it */
1261                 if (req->special) {
1262                         struct scsi_cmnd *cmd = req->special;
1263                         scsi_release_buffers(cmd);
1264                         scsi_put_command(cmd);
1265                         put_device(&sdev->sdev_gendev);
1266                         req->special = NULL;
1267                 }
1268                 break;
1269         case BLKPREP_DEFER:
1270                 /*
1271                  * If we defer, the blk_peek_request() returns NULL, but the
1272                  * queue must be restarted, so we schedule a callback to happen
1273                  * shortly.
1274                  */
1275                 if (atomic_read(&sdev->device_busy) == 0)
1276                         blk_delay_queue(q, SCSI_QUEUE_DELAY);
1277                 break;
1278         default:
1279                 req->cmd_flags |= REQ_DONTPREP;
1280         }
1281
1282         return ret;
1283 }
1284
1285 static int scsi_prep_fn(struct request_queue *q, struct request *req)
1286 {
1287         struct scsi_device *sdev = q->queuedata;
1288         struct scsi_cmnd *cmd;
1289         int ret;
1290
1291         ret = scsi_prep_state_check(sdev, req);
1292         if (ret != BLKPREP_OK)
1293                 goto out;
1294
1295         cmd = scsi_get_cmd_from_req(sdev, req);
1296         if (unlikely(!cmd)) {
1297                 ret = BLKPREP_DEFER;
1298                 goto out;
1299         }
1300
1301         ret = scsi_setup_cmnd(sdev, req);
1302 out:
1303         return scsi_prep_return(q, req, ret);
1304 }
1305
1306 static void scsi_unprep_fn(struct request_queue *q, struct request *req)
1307 {
1308         scsi_uninit_cmd(req->special);
1309 }
1310
1311 /*
1312  * scsi_dev_queue_ready: if we can send requests to sdev, return 1 else
1313  * return 0.
1314  *
1315  * Called with the queue_lock held.
1316  */
1317 static inline int scsi_dev_queue_ready(struct request_queue *q,
1318                                   struct scsi_device *sdev)
1319 {
1320         unsigned int busy;
1321
1322         busy = atomic_inc_return(&sdev->device_busy) - 1;
1323         if (atomic_read(&sdev->device_blocked)) {
1324                 if (busy)
1325                         goto out_dec;
1326
1327                 /*
1328                  * unblock after device_blocked iterates to zero
1329                  */
1330                 if (atomic_dec_return(&sdev->device_blocked) > 0) {
1331                         /*
1332                          * For the MQ case we take care of this in the caller.
1333                          */
1334                         if (!q->mq_ops)
1335                                 blk_delay_queue(q, SCSI_QUEUE_DELAY);
1336                         goto out_dec;
1337                 }
1338                 SCSI_LOG_MLQUEUE(3, sdev_printk(KERN_INFO, sdev,
1339                                    "unblocking device at zero depth\n"));
1340         }
1341
1342         if (busy >= sdev->queue_depth)
1343                 goto out_dec;
1344
1345         return 1;
1346 out_dec:
1347         atomic_dec(&sdev->device_busy);
1348         return 0;
1349 }
1350
1351 /*
1352  * scsi_target_queue_ready: checks if there we can send commands to target
1353  * @sdev: scsi device on starget to check.
1354  */
1355 static inline int scsi_target_queue_ready(struct Scsi_Host *shost,
1356                                            struct scsi_device *sdev)
1357 {
1358         struct scsi_target *starget = scsi_target(sdev);
1359         unsigned int busy;
1360
1361         if (starget->single_lun) {
1362                 spin_lock_irq(shost->host_lock);
1363                 if (starget->starget_sdev_user &&
1364                     starget->starget_sdev_user != sdev) {
1365                         spin_unlock_irq(shost->host_lock);
1366                         return 0;
1367                 }
1368                 starget->starget_sdev_user = sdev;
1369                 spin_unlock_irq(shost->host_lock);
1370         }
1371
1372         if (starget->can_queue <= 0)
1373                 return 1;
1374
1375         busy = atomic_inc_return(&starget->target_busy) - 1;
1376         if (atomic_read(&starget->target_blocked) > 0) {
1377                 if (busy)
1378                         goto starved;
1379
1380                 /*
1381                  * unblock after target_blocked iterates to zero
1382                  */
1383                 if (atomic_dec_return(&starget->target_blocked) > 0)
1384                         goto out_dec;
1385
1386                 SCSI_LOG_MLQUEUE(3, starget_printk(KERN_INFO, starget,
1387                                  "unblocking target at zero depth\n"));
1388         }
1389
1390         if (busy >= starget->can_queue)
1391                 goto starved;
1392
1393         return 1;
1394
1395 starved:
1396         spin_lock_irq(shost->host_lock);
1397         list_move_tail(&sdev->starved_entry, &shost->starved_list);
1398         spin_unlock_irq(shost->host_lock);
1399 out_dec:
1400         if (starget->can_queue > 0)
1401                 atomic_dec(&starget->target_busy);
1402         return 0;
1403 }
1404
1405 /*
1406  * scsi_host_queue_ready: if we can send requests to shost, return 1 else
1407  * return 0. We must end up running the queue again whenever 0 is
1408  * returned, else IO can hang.
1409  */
1410 static inline int scsi_host_queue_ready(struct request_queue *q,
1411                                    struct Scsi_Host *shost,
1412                                    struct scsi_device *sdev)
1413 {
1414         unsigned int busy;
1415
1416         if (scsi_host_in_recovery(shost))
1417                 return 0;
1418
1419         busy = atomic_inc_return(&shost->host_busy) - 1;
1420         if (atomic_read(&shost->host_blocked) > 0) {
1421                 if (busy)
1422                         goto starved;
1423
1424                 /*
1425                  * unblock after host_blocked iterates to zero
1426                  */
1427                 if (atomic_dec_return(&shost->host_blocked) > 0)
1428                         goto out_dec;
1429
1430                 SCSI_LOG_MLQUEUE(3,
1431                         shost_printk(KERN_INFO, shost,
1432                                      "unblocking host at zero depth\n"));
1433         }
1434
1435         if (shost->can_queue > 0 && busy >= shost->can_queue)
1436                 goto starved;
1437         if (shost->host_self_blocked)
1438                 goto starved;
1439
1440         /* We're OK to process the command, so we can't be starved */
1441         if (!list_empty(&sdev->starved_entry)) {
1442                 spin_lock_irq(shost->host_lock);
1443                 if (!list_empty(&sdev->starved_entry))
1444                         list_del_init(&sdev->starved_entry);
1445                 spin_unlock_irq(shost->host_lock);
1446         }
1447
1448         return 1;
1449
1450 starved:
1451         spin_lock_irq(shost->host_lock);
1452         if (list_empty(&sdev->starved_entry))
1453                 list_add_tail(&sdev->starved_entry, &shost->starved_list);
1454         spin_unlock_irq(shost->host_lock);
1455 out_dec:
1456         atomic_dec(&shost->host_busy);
1457         return 0;
1458 }
1459
1460 /*
1461  * Busy state exporting function for request stacking drivers.
1462  *
1463  * For efficiency, no lock is taken to check the busy state of
1464  * shost/starget/sdev, since the returned value is not guaranteed and
1465  * may be changed after request stacking drivers call the function,
1466  * regardless of taking lock or not.
1467  *
1468  * When scsi can't dispatch I/Os anymore and needs to kill I/Os scsi
1469  * needs to return 'not busy'. Otherwise, request stacking drivers
1470  * may hold requests forever.
1471  */
1472 static int scsi_lld_busy(struct request_queue *q)
1473 {
1474         struct scsi_device *sdev = q->queuedata;
1475         struct Scsi_Host *shost;
1476
1477         if (blk_queue_dying(q))
1478                 return 0;
1479
1480         shost = sdev->host;
1481
1482         /*
1483          * Ignore host/starget busy state.
1484          * Since block layer does not have a concept of fairness across
1485          * multiple queues, congestion of host/starget needs to be handled
1486          * in SCSI layer.
1487          */
1488         if (scsi_host_in_recovery(shost) || scsi_device_is_busy(sdev))
1489                 return 1;
1490
1491         return 0;
1492 }
1493
1494 /*
1495  * Kill a request for a dead device
1496  */
1497 static void scsi_kill_request(struct request *req, struct request_queue *q)
1498 {
1499         struct scsi_cmnd *cmd = req->special;
1500         struct scsi_device *sdev;
1501         struct scsi_target *starget;
1502         struct Scsi_Host *shost;
1503
1504         blk_start_request(req);
1505
1506         scmd_printk(KERN_INFO, cmd, "killing request\n");
1507
1508         sdev = cmd->device;
1509         starget = scsi_target(sdev);
1510         shost = sdev->host;
1511         scsi_init_cmd_errh(cmd);
1512         cmd->result = DID_NO_CONNECT << 16;
1513         atomic_inc(&cmd->device->iorequest_cnt);
1514
1515         /*
1516          * SCSI request completion path will do scsi_device_unbusy(),
1517          * bump busy counts.  To bump the counters, we need to dance
1518          * with the locks as normal issue path does.
1519          */
1520         atomic_inc(&sdev->device_busy);
1521         atomic_inc(&shost->host_busy);
1522         if (starget->can_queue > 0)
1523                 atomic_inc(&starget->target_busy);
1524
1525         blk_complete_request(req);
1526 }
1527
1528 static void scsi_softirq_done(struct request *rq)
1529 {
1530         struct scsi_cmnd *cmd = rq->special;
1531         unsigned long wait_for = (cmd->allowed + 1) * rq->timeout;
1532         int disposition;
1533
1534         INIT_LIST_HEAD(&cmd->eh_entry);
1535
1536         atomic_inc(&cmd->device->iodone_cnt);
1537         if (cmd->result)
1538                 atomic_inc(&cmd->device->ioerr_cnt);
1539
1540         disposition = scsi_decide_disposition(cmd);
1541         if (disposition != SUCCESS &&
1542             time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) {
1543                 sdev_printk(KERN_ERR, cmd->device,
1544                             "timing out command, waited %lus\n",
1545                             wait_for/HZ);
1546                 disposition = SUCCESS;
1547         }
1548
1549         scsi_log_completion(cmd, disposition);
1550
1551         switch (disposition) {
1552                 case SUCCESS:
1553                         scsi_finish_command(cmd);
1554                         break;
1555                 case NEEDS_RETRY:
1556                         scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY);
1557                         break;
1558                 case ADD_TO_MLQUEUE:
1559                         scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY);
1560                         break;
1561                 default:
1562                         if (!scsi_eh_scmd_add(cmd, 0))
1563                                 scsi_finish_command(cmd);
1564         }
1565 }
1566
1567 /**
1568  * scsi_dispatch_command - Dispatch a command to the low-level driver.
1569  * @cmd: command block we are dispatching.
1570  *
1571  * Return: nonzero return request was rejected and device's queue needs to be
1572  * plugged.
1573  */
1574 static int scsi_dispatch_cmd(struct scsi_cmnd *cmd)
1575 {
1576         struct Scsi_Host *host = cmd->device->host;
1577         int rtn = 0;
1578
1579         atomic_inc(&cmd->device->iorequest_cnt);
1580
1581         /* check if the device is still usable */
1582         if (unlikely(cmd->device->sdev_state == SDEV_DEL)) {
1583                 /* in SDEV_DEL we error all commands. DID_NO_CONNECT
1584                  * returns an immediate error upwards, and signals
1585                  * that the device is no longer present */
1586                 cmd->result = DID_NO_CONNECT << 16;
1587                 goto done;
1588         }
1589
1590         /* Check to see if the scsi lld made this device blocked. */
1591         if (unlikely(scsi_device_blocked(cmd->device))) {
1592                 /*
1593                  * in blocked state, the command is just put back on
1594                  * the device queue.  The suspend state has already
1595                  * blocked the queue so future requests should not
1596                  * occur until the device transitions out of the
1597                  * suspend state.
1598                  */
1599                 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1600                         "queuecommand : device blocked\n"));
1601                 return SCSI_MLQUEUE_DEVICE_BUSY;
1602         }
1603
1604         /* Store the LUN value in cmnd, if needed. */
1605         if (cmd->device->lun_in_cdb)
1606                 cmd->cmnd[1] = (cmd->cmnd[1] & 0x1f) |
1607                                (cmd->device->lun << 5 & 0xe0);
1608
1609         scsi_log_send(cmd);
1610
1611         /*
1612          * Before we queue this command, check if the command
1613          * length exceeds what the host adapter can handle.
1614          */
1615         if (cmd->cmd_len > cmd->device->host->max_cmd_len) {
1616                 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1617                                "queuecommand : command too long. "
1618                                "cdb_size=%d host->max_cmd_len=%d\n",
1619                                cmd->cmd_len, cmd->device->host->max_cmd_len));
1620                 cmd->result = (DID_ABORT << 16);
1621                 goto done;
1622         }
1623
1624         if (unlikely(host->shost_state == SHOST_DEL)) {
1625                 cmd->result = (DID_NO_CONNECT << 16);
1626                 goto done;
1627
1628         }
1629
1630         trace_scsi_dispatch_cmd_start(cmd);
1631         rtn = host->hostt->queuecommand(host, cmd);
1632         if (rtn) {
1633                 trace_scsi_dispatch_cmd_error(cmd, rtn);
1634                 if (rtn != SCSI_MLQUEUE_DEVICE_BUSY &&
1635                     rtn != SCSI_MLQUEUE_TARGET_BUSY)
1636                         rtn = SCSI_MLQUEUE_HOST_BUSY;
1637
1638                 SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1639                         "queuecommand : request rejected\n"));
1640         }
1641
1642         return rtn;
1643  done:
1644         cmd->scsi_done(cmd);
1645         return 0;
1646 }
1647
1648 /**
1649  * scsi_done - Invoke completion on finished SCSI command.
1650  * @cmd: The SCSI Command for which a low-level device driver (LLDD) gives
1651  * ownership back to SCSI Core -- i.e. the LLDD has finished with it.
1652  *
1653  * Description: This function is the mid-level's (SCSI Core) interrupt routine,
1654  * which regains ownership of the SCSI command (de facto) from a LLDD, and
1655  * calls blk_complete_request() for further processing.
1656  *
1657  * This function is interrupt context safe.
1658  */
1659 static void scsi_done(struct scsi_cmnd *cmd)
1660 {
1661         trace_scsi_dispatch_cmd_done(cmd);
1662         blk_complete_request(cmd->request);
1663 }
1664
1665 /*
1666  * Function:    scsi_request_fn()
1667  *
1668  * Purpose:     Main strategy routine for SCSI.
1669  *
1670  * Arguments:   q       - Pointer to actual queue.
1671  *
1672  * Returns:     Nothing
1673  *
1674  * Lock status: IO request lock assumed to be held when called.
1675  */
1676 static void scsi_request_fn(struct request_queue *q)
1677         __releases(q->queue_lock)
1678         __acquires(q->queue_lock)
1679 {
1680         struct scsi_device *sdev = q->queuedata;
1681         struct Scsi_Host *shost;
1682         struct scsi_cmnd *cmd;
1683         struct request *req;
1684
1685         /*
1686          * To start with, we keep looping until the queue is empty, or until
1687          * the host is no longer able to accept any more requests.
1688          */
1689         shost = sdev->host;
1690         for (;;) {
1691                 int rtn;
1692                 /*
1693                  * get next queueable request.  We do this early to make sure
1694                  * that the request is fully prepared even if we cannot
1695                  * accept it.
1696                  */
1697                 req = blk_peek_request(q);
1698                 if (!req)
1699                         break;
1700
1701                 if (unlikely(!scsi_device_online(sdev))) {
1702                         sdev_printk(KERN_ERR, sdev,
1703                                     "rejecting I/O to offline device\n");
1704                         scsi_kill_request(req, q);
1705                         continue;
1706                 }
1707
1708                 if (!scsi_dev_queue_ready(q, sdev))
1709                         break;
1710
1711                 /*
1712                  * Remove the request from the request list.
1713                  */
1714                 if (!(blk_queue_tagged(q) && !blk_queue_start_tag(q, req)))
1715                         blk_start_request(req);
1716
1717                 spin_unlock_irq(q->queue_lock);
1718                 cmd = req->special;
1719                 if (unlikely(cmd == NULL)) {
1720                         printk(KERN_CRIT "impossible request in %s.\n"
1721                                          "please mail a stack trace to "
1722                                          "linux-scsi@vger.kernel.org\n",
1723                                          __func__);
1724                         blk_dump_rq_flags(req, "foo");
1725                         BUG();
1726                 }
1727
1728                 /*
1729                  * We hit this when the driver is using a host wide
1730                  * tag map. For device level tag maps the queue_depth check
1731                  * in the device ready fn would prevent us from trying
1732                  * to allocate a tag. Since the map is a shared host resource
1733                  * we add the dev to the starved list so it eventually gets
1734                  * a run when a tag is freed.
1735                  */
1736                 if (blk_queue_tagged(q) && !(req->cmd_flags & REQ_QUEUED)) {
1737                         spin_lock_irq(shost->host_lock);
1738                         if (list_empty(&sdev->starved_entry))
1739                                 list_add_tail(&sdev->starved_entry,
1740                                               &shost->starved_list);
1741                         spin_unlock_irq(shost->host_lock);
1742                         goto not_ready;
1743                 }
1744
1745                 if (!scsi_target_queue_ready(shost, sdev))
1746                         goto not_ready;
1747
1748                 if (!scsi_host_queue_ready(q, shost, sdev))
1749                         goto host_not_ready;
1750         
1751                 if (sdev->simple_tags)
1752                         cmd->flags |= SCMD_TAGGED;
1753                 else
1754                         cmd->flags &= ~SCMD_TAGGED;
1755
1756                 /*
1757                  * Finally, initialize any error handling parameters, and set up
1758                  * the timers for timeouts.
1759                  */
1760                 scsi_init_cmd_errh(cmd);
1761
1762                 /*
1763                  * Dispatch the command to the low-level driver.
1764                  */
1765                 cmd->scsi_done = scsi_done;
1766                 rtn = scsi_dispatch_cmd(cmd);
1767                 if (rtn) {
1768                         scsi_queue_insert(cmd, rtn);
1769                         spin_lock_irq(q->queue_lock);
1770                         goto out_delay;
1771                 }
1772                 spin_lock_irq(q->queue_lock);
1773         }
1774
1775         return;
1776
1777  host_not_ready:
1778         if (scsi_target(sdev)->can_queue > 0)
1779                 atomic_dec(&scsi_target(sdev)->target_busy);
1780  not_ready:
1781         /*
1782          * lock q, handle tag, requeue req, and decrement device_busy. We
1783          * must return with queue_lock held.
1784          *
1785          * Decrementing device_busy without checking it is OK, as all such
1786          * cases (host limits or settings) should run the queue at some
1787          * later time.
1788          */
1789         spin_lock_irq(q->queue_lock);
1790         blk_requeue_request(q, req);
1791         atomic_dec(&sdev->device_busy);
1792 out_delay:
1793         if (!atomic_read(&sdev->device_busy) && !scsi_device_blocked(sdev))
1794                 blk_delay_queue(q, SCSI_QUEUE_DELAY);
1795 }
1796
1797 static inline int prep_to_mq(int ret)
1798 {
1799         switch (ret) {
1800         case BLKPREP_OK:
1801                 return 0;
1802         case BLKPREP_DEFER:
1803                 return BLK_MQ_RQ_QUEUE_BUSY;
1804         default:
1805                 return BLK_MQ_RQ_QUEUE_ERROR;
1806         }
1807 }
1808
1809 static int scsi_mq_prep_fn(struct request *req)
1810 {
1811         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1812         struct scsi_device *sdev = req->q->queuedata;
1813         struct Scsi_Host *shost = sdev->host;
1814         unsigned char *sense_buf = cmd->sense_buffer;
1815         struct scatterlist *sg;
1816
1817         memset(cmd, 0, sizeof(struct scsi_cmnd));
1818
1819         req->special = cmd;
1820
1821         cmd->request = req;
1822         cmd->device = sdev;
1823         cmd->sense_buffer = sense_buf;
1824
1825         cmd->tag = req->tag;
1826
1827         cmd->cmnd = req->cmd;
1828         cmd->prot_op = SCSI_PROT_NORMAL;
1829
1830         INIT_LIST_HEAD(&cmd->list);
1831         INIT_DELAYED_WORK(&cmd->abort_work, scmd_eh_abort_handler);
1832         cmd->jiffies_at_alloc = jiffies;
1833
1834         if (shost->use_cmd_list) {
1835                 spin_lock_irq(&sdev->list_lock);
1836                 list_add_tail(&cmd->list, &sdev->cmd_list);
1837                 spin_unlock_irq(&sdev->list_lock);
1838         }
1839
1840         sg = (void *)cmd + sizeof(struct scsi_cmnd) + shost->hostt->cmd_size;
1841         cmd->sdb.table.sgl = sg;
1842
1843         if (scsi_host_get_prot(shost)) {
1844                 cmd->prot_sdb = (void *)sg +
1845                         min_t(unsigned int,
1846                               shost->sg_tablesize, SG_CHUNK_SIZE) *
1847                         sizeof(struct scatterlist);
1848                 memset(cmd->prot_sdb, 0, sizeof(struct scsi_data_buffer));
1849
1850                 cmd->prot_sdb->table.sgl =
1851                         (struct scatterlist *)(cmd->prot_sdb + 1);
1852         }
1853
1854         if (blk_bidi_rq(req)) {
1855                 struct request *next_rq = req->next_rq;
1856                 struct scsi_data_buffer *bidi_sdb = blk_mq_rq_to_pdu(next_rq);
1857
1858                 memset(bidi_sdb, 0, sizeof(struct scsi_data_buffer));
1859                 bidi_sdb->table.sgl =
1860                         (struct scatterlist *)(bidi_sdb + 1);
1861
1862                 next_rq->special = bidi_sdb;
1863         }
1864
1865         blk_mq_start_request(req);
1866
1867         return scsi_setup_cmnd(sdev, req);
1868 }
1869
1870 static void scsi_mq_done(struct scsi_cmnd *cmd)
1871 {
1872         trace_scsi_dispatch_cmd_done(cmd);
1873         blk_mq_complete_request(cmd->request, cmd->request->errors);
1874 }
1875
1876 static int scsi_queue_rq(struct blk_mq_hw_ctx *hctx,
1877                          const struct blk_mq_queue_data *bd)
1878 {
1879         struct request *req = bd->rq;
1880         struct request_queue *q = req->q;
1881         struct scsi_device *sdev = q->queuedata;
1882         struct Scsi_Host *shost = sdev->host;
1883         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1884         int ret;
1885         int reason;
1886
1887         ret = prep_to_mq(scsi_prep_state_check(sdev, req));
1888         if (ret)
1889                 goto out;
1890
1891         ret = BLK_MQ_RQ_QUEUE_BUSY;
1892         if (!get_device(&sdev->sdev_gendev))
1893                 goto out;
1894
1895         if (!scsi_dev_queue_ready(q, sdev))
1896                 goto out_put_device;
1897         if (!scsi_target_queue_ready(shost, sdev))
1898                 goto out_dec_device_busy;
1899         if (!scsi_host_queue_ready(q, shost, sdev))
1900                 goto out_dec_target_busy;
1901
1902
1903         if (!(req->cmd_flags & REQ_DONTPREP)) {
1904                 ret = prep_to_mq(scsi_mq_prep_fn(req));
1905                 if (ret)
1906                         goto out_dec_host_busy;
1907                 req->cmd_flags |= REQ_DONTPREP;
1908         } else {
1909                 blk_mq_start_request(req);
1910         }
1911
1912         if (sdev->simple_tags)
1913                 cmd->flags |= SCMD_TAGGED;
1914         else
1915                 cmd->flags &= ~SCMD_TAGGED;
1916
1917         scsi_init_cmd_errh(cmd);
1918         cmd->scsi_done = scsi_mq_done;
1919
1920         reason = scsi_dispatch_cmd(cmd);
1921         if (reason) {
1922                 scsi_set_blocked(cmd, reason);
1923                 ret = BLK_MQ_RQ_QUEUE_BUSY;
1924                 goto out_dec_host_busy;
1925         }
1926
1927         return BLK_MQ_RQ_QUEUE_OK;
1928
1929 out_dec_host_busy:
1930         atomic_dec(&shost->host_busy);
1931 out_dec_target_busy:
1932         if (scsi_target(sdev)->can_queue > 0)
1933                 atomic_dec(&scsi_target(sdev)->target_busy);
1934 out_dec_device_busy:
1935         atomic_dec(&sdev->device_busy);
1936 out_put_device:
1937         put_device(&sdev->sdev_gendev);
1938 out:
1939         switch (ret) {
1940         case BLK_MQ_RQ_QUEUE_BUSY:
1941                 blk_mq_stop_hw_queue(hctx);
1942                 if (atomic_read(&sdev->device_busy) == 0 &&
1943                     !scsi_device_blocked(sdev))
1944                         blk_mq_delay_queue(hctx, SCSI_QUEUE_DELAY);
1945                 break;
1946         case BLK_MQ_RQ_QUEUE_ERROR:
1947                 /*
1948                  * Make sure to release all allocated ressources when
1949                  * we hit an error, as we will never see this command
1950                  * again.
1951                  */
1952                 if (req->cmd_flags & REQ_DONTPREP)
1953                         scsi_mq_uninit_cmd(cmd);
1954                 break;
1955         default:
1956                 break;
1957         }
1958         return ret;
1959 }
1960
1961 static enum blk_eh_timer_return scsi_timeout(struct request *req,
1962                 bool reserved)
1963 {
1964         if (reserved)
1965                 return BLK_EH_RESET_TIMER;
1966         return scsi_times_out(req);
1967 }
1968
1969 static int scsi_init_request(void *data, struct request *rq,
1970                 unsigned int hctx_idx, unsigned int request_idx,
1971                 unsigned int numa_node)
1972 {
1973         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1974
1975         cmd->sense_buffer = kzalloc_node(SCSI_SENSE_BUFFERSIZE, GFP_KERNEL,
1976                         numa_node);
1977         if (!cmd->sense_buffer)
1978                 return -ENOMEM;
1979         return 0;
1980 }
1981
1982 static void scsi_exit_request(void *data, struct request *rq,
1983                 unsigned int hctx_idx, unsigned int request_idx)
1984 {
1985         struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
1986
1987         kfree(cmd->sense_buffer);
1988 }
1989
1990 static u64 scsi_calculate_bounce_limit(struct Scsi_Host *shost)
1991 {
1992         struct device *host_dev;
1993         u64 bounce_limit = 0xffffffff;
1994
1995         if (shost->unchecked_isa_dma)
1996                 return BLK_BOUNCE_ISA;
1997         /*
1998          * Platforms with virtual-DMA translation
1999          * hardware have no practical limit.
2000          */
2001         if (!PCI_DMA_BUS_IS_PHYS)
2002                 return BLK_BOUNCE_ANY;
2003
2004         host_dev = scsi_get_device(shost);
2005         if (host_dev && host_dev->dma_mask)
2006                 bounce_limit = (u64)dma_max_pfn(host_dev) << PAGE_SHIFT;
2007
2008         return bounce_limit;
2009 }
2010
2011 static void __scsi_init_queue(struct Scsi_Host *shost, struct request_queue *q)
2012 {
2013         struct device *dev = shost->dma_dev;
2014
2015         /*
2016          * this limit is imposed by hardware restrictions
2017          */
2018         blk_queue_max_segments(q, min_t(unsigned short, shost->sg_tablesize,
2019                                         SG_MAX_SEGMENTS));
2020
2021         if (scsi_host_prot_dma(shost)) {
2022                 shost->sg_prot_tablesize =
2023                         min_not_zero(shost->sg_prot_tablesize,
2024                                      (unsigned short)SCSI_MAX_PROT_SG_SEGMENTS);
2025                 BUG_ON(shost->sg_prot_tablesize < shost->sg_tablesize);
2026                 blk_queue_max_integrity_segments(q, shost->sg_prot_tablesize);
2027         }
2028
2029         blk_queue_max_hw_sectors(q, shost->max_sectors);
2030         blk_queue_bounce_limit(q, scsi_calculate_bounce_limit(shost));
2031         blk_queue_segment_boundary(q, shost->dma_boundary);
2032         dma_set_seg_boundary(dev, shost->dma_boundary);
2033
2034         blk_queue_max_segment_size(q, dma_get_max_seg_size(dev));
2035
2036         if (!shost->use_clustering)
2037                 q->limits.cluster = 0;
2038
2039         /*
2040          * set a reasonable default alignment on word boundaries: the
2041          * host and device may alter it using
2042          * blk_queue_update_dma_alignment() later.
2043          */
2044         blk_queue_dma_alignment(q, 0x03);
2045 }
2046
2047 struct request_queue *__scsi_alloc_queue(struct Scsi_Host *shost,
2048                                          request_fn_proc *request_fn)
2049 {
2050         struct request_queue *q;
2051
2052         q = blk_init_queue(request_fn, NULL);
2053         if (!q)
2054                 return NULL;
2055         __scsi_init_queue(shost, q);
2056         return q;
2057 }
2058 EXPORT_SYMBOL(__scsi_alloc_queue);
2059
2060 struct request_queue *scsi_alloc_queue(struct scsi_device *sdev)
2061 {
2062         struct request_queue *q;
2063
2064         q = __scsi_alloc_queue(sdev->host, scsi_request_fn);
2065         if (!q)
2066                 return NULL;
2067
2068         blk_queue_prep_rq(q, scsi_prep_fn);
2069         blk_queue_unprep_rq(q, scsi_unprep_fn);
2070         blk_queue_softirq_done(q, scsi_softirq_done);
2071         blk_queue_rq_timed_out(q, scsi_times_out);
2072         blk_queue_lld_busy(q, scsi_lld_busy);
2073         return q;
2074 }
2075
2076 static struct blk_mq_ops scsi_mq_ops = {
2077         .map_queue      = blk_mq_map_queue,
2078         .queue_rq       = scsi_queue_rq,
2079         .complete       = scsi_softirq_done,
2080         .timeout        = scsi_timeout,
2081         .init_request   = scsi_init_request,
2082         .exit_request   = scsi_exit_request,
2083 };
2084
2085 struct request_queue *scsi_mq_alloc_queue(struct scsi_device *sdev)
2086 {
2087         sdev->request_queue = blk_mq_init_queue(&sdev->host->tag_set);
2088         if (IS_ERR(sdev->request_queue))
2089                 return NULL;
2090
2091         sdev->request_queue->queuedata = sdev;
2092         __scsi_init_queue(sdev->host, sdev->request_queue);
2093         return sdev->request_queue;
2094 }
2095
2096 int scsi_mq_setup_tags(struct Scsi_Host *shost)
2097 {
2098         unsigned int cmd_size, sgl_size, tbl_size;
2099
2100         tbl_size = shost->sg_tablesize;
2101         if (tbl_size > SG_CHUNK_SIZE)
2102                 tbl_size = SG_CHUNK_SIZE;
2103         sgl_size = tbl_size * sizeof(struct scatterlist);
2104         cmd_size = sizeof(struct scsi_cmnd) + shost->hostt->cmd_size + sgl_size;
2105         if (scsi_host_get_prot(shost))
2106                 cmd_size += sizeof(struct scsi_data_buffer) + sgl_size;
2107
2108         memset(&shost->tag_set, 0, sizeof(shost->tag_set));
2109         shost->tag_set.ops = &scsi_mq_ops;
2110         shost->tag_set.nr_hw_queues = shost->nr_hw_queues ? : 1;
2111         shost->tag_set.queue_depth = shost->can_queue;
2112         shost->tag_set.cmd_size = cmd_size;
2113         shost->tag_set.numa_node = NUMA_NO_NODE;
2114         shost->tag_set.flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_SG_MERGE;
2115         shost->tag_set.flags |=
2116                 BLK_ALLOC_POLICY_TO_MQ_FLAG(shost->hostt->tag_alloc_policy);
2117         shost->tag_set.driver_data = shost;
2118
2119         return blk_mq_alloc_tag_set(&shost->tag_set);
2120 }
2121
2122 void scsi_mq_destroy_tags(struct Scsi_Host *shost)
2123 {
2124         blk_mq_free_tag_set(&shost->tag_set);
2125 }
2126
2127 /*
2128  * Function:    scsi_block_requests()
2129  *
2130  * Purpose:     Utility function used by low-level drivers to prevent further
2131  *              commands from being queued to the device.
2132  *
2133  * Arguments:   shost       - Host in question
2134  *
2135  * Returns:     Nothing
2136  *
2137  * Lock status: No locks are assumed held.
2138  *
2139  * Notes:       There is no timer nor any other means by which the requests
2140  *              get unblocked other than the low-level driver calling
2141  *              scsi_unblock_requests().
2142  */
2143 void scsi_block_requests(struct Scsi_Host *shost)
2144 {
2145         shost->host_self_blocked = 1;
2146 }
2147 EXPORT_SYMBOL(scsi_block_requests);
2148
2149 /*
2150  * Function:    scsi_unblock_requests()
2151  *
2152  * Purpose:     Utility function used by low-level drivers to allow further
2153  *              commands from being queued to the device.
2154  *
2155  * Arguments:   shost       - Host in question
2156  *
2157  * Returns:     Nothing
2158  *
2159  * Lock status: No locks are assumed held.
2160  *
2161  * Notes:       There is no timer nor any other means by which the requests
2162  *              get unblocked other than the low-level driver calling
2163  *              scsi_unblock_requests().
2164  *
2165  *              This is done as an API function so that changes to the
2166  *              internals of the scsi mid-layer won't require wholesale
2167  *              changes to drivers that use this feature.
2168  */
2169 void scsi_unblock_requests(struct Scsi_Host *shost)
2170 {
2171         shost->host_self_blocked = 0;
2172         scsi_run_host_queues(shost);
2173 }
2174 EXPORT_SYMBOL(scsi_unblock_requests);
2175
2176 int __init scsi_init_queue(void)
2177 {
2178         scsi_sdb_cache = kmem_cache_create("scsi_data_buffer",
2179                                            sizeof(struct scsi_data_buffer),
2180                                            0, 0, NULL);
2181         if (!scsi_sdb_cache) {
2182                 printk(KERN_ERR "SCSI: can't init scsi sdb cache\n");
2183                 return -ENOMEM;
2184         }
2185
2186         return 0;
2187 }
2188
2189 void scsi_exit_queue(void)
2190 {
2191         kmem_cache_destroy(scsi_sdb_cache);
2192 }
2193
2194 /**
2195  *      scsi_mode_select - issue a mode select
2196  *      @sdev:  SCSI device to be queried
2197  *      @pf:    Page format bit (1 == standard, 0 == vendor specific)
2198  *      @sp:    Save page bit (0 == don't save, 1 == save)
2199  *      @modepage: mode page being requested
2200  *      @buffer: request buffer (may not be smaller than eight bytes)
2201  *      @len:   length of request buffer.
2202  *      @timeout: command timeout
2203  *      @retries: number of retries before failing
2204  *      @data: returns a structure abstracting the mode header data
2205  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
2206  *              must be SCSI_SENSE_BUFFERSIZE big.
2207  *
2208  *      Returns zero if successful; negative error number or scsi
2209  *      status on error
2210  *
2211  */
2212 int
2213 scsi_mode_select(struct scsi_device *sdev, int pf, int sp, int modepage,
2214                  unsigned char *buffer, int len, int timeout, int retries,
2215                  struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
2216 {
2217         unsigned char cmd[10];
2218         unsigned char *real_buffer;
2219         int ret;
2220
2221         memset(cmd, 0, sizeof(cmd));
2222         cmd[1] = (pf ? 0x10 : 0) | (sp ? 0x01 : 0);
2223
2224         if (sdev->use_10_for_ms) {
2225                 if (len > 65535)
2226                         return -EINVAL;
2227                 real_buffer = kmalloc(8 + len, GFP_KERNEL);
2228                 if (!real_buffer)
2229                         return -ENOMEM;
2230                 memcpy(real_buffer + 8, buffer, len);
2231                 len += 8;
2232                 real_buffer[0] = 0;
2233                 real_buffer[1] = 0;
2234                 real_buffer[2] = data->medium_type;
2235                 real_buffer[3] = data->device_specific;
2236                 real_buffer[4] = data->longlba ? 0x01 : 0;
2237                 real_buffer[5] = 0;
2238                 real_buffer[6] = data->block_descriptor_length >> 8;
2239                 real_buffer[7] = data->block_descriptor_length;
2240
2241                 cmd[0] = MODE_SELECT_10;
2242                 cmd[7] = len >> 8;
2243                 cmd[8] = len;
2244         } else {
2245                 if (len > 255 || data->block_descriptor_length > 255 ||
2246                     data->longlba)
2247                         return -EINVAL;
2248
2249                 real_buffer = kmalloc(4 + len, GFP_KERNEL);
2250                 if (!real_buffer)
2251                         return -ENOMEM;
2252                 memcpy(real_buffer + 4, buffer, len);
2253                 len += 4;
2254                 real_buffer[0] = 0;
2255                 real_buffer[1] = data->medium_type;
2256                 real_buffer[2] = data->device_specific;
2257                 real_buffer[3] = data->block_descriptor_length;
2258                 
2259
2260                 cmd[0] = MODE_SELECT;
2261                 cmd[4] = len;
2262         }
2263
2264         ret = scsi_execute_req(sdev, cmd, DMA_TO_DEVICE, real_buffer, len,
2265                                sshdr, timeout, retries, NULL);
2266         kfree(real_buffer);
2267         return ret;
2268 }
2269 EXPORT_SYMBOL_GPL(scsi_mode_select);
2270
2271 /**
2272  *      scsi_mode_sense - issue a mode sense, falling back from 10 to six bytes if necessary.
2273  *      @sdev:  SCSI device to be queried
2274  *      @dbd:   set if mode sense will allow block descriptors to be returned
2275  *      @modepage: mode page being requested
2276  *      @buffer: request buffer (may not be smaller than eight bytes)
2277  *      @len:   length of request buffer.
2278  *      @timeout: command timeout
2279  *      @retries: number of retries before failing
2280  *      @data: returns a structure abstracting the mode header data
2281  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
2282  *              must be SCSI_SENSE_BUFFERSIZE big.
2283  *
2284  *      Returns zero if unsuccessful, or the header offset (either 4
2285  *      or 8 depending on whether a six or ten byte command was
2286  *      issued) if successful.
2287  */
2288 int
2289 scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
2290                   unsigned char *buffer, int len, int timeout, int retries,
2291                   struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
2292 {
2293         unsigned char cmd[12];
2294         int use_10_for_ms;
2295         int header_length;
2296         int result, retry_count = retries;
2297         struct scsi_sense_hdr my_sshdr;
2298
2299         memset(data, 0, sizeof(*data));
2300         memset(&cmd[0], 0, 12);
2301         cmd[1] = dbd & 0x18;    /* allows DBD and LLBA bits */
2302         cmd[2] = modepage;
2303
2304         /* caller might not be interested in sense, but we need it */
2305         if (!sshdr)
2306                 sshdr = &my_sshdr;
2307
2308  retry:
2309         use_10_for_ms = sdev->use_10_for_ms;
2310
2311         if (use_10_for_ms) {
2312                 if (len < 8)
2313                         len = 8;
2314
2315                 cmd[0] = MODE_SENSE_10;
2316                 cmd[8] = len;
2317                 header_length = 8;
2318         } else {
2319                 if (len < 4)
2320                         len = 4;
2321
2322                 cmd[0] = MODE_SENSE;
2323                 cmd[4] = len;
2324                 header_length = 4;
2325         }
2326
2327         memset(buffer, 0, len);
2328
2329         result = scsi_execute_req(sdev, cmd, DMA_FROM_DEVICE, buffer, len,
2330                                   sshdr, timeout, retries, NULL);
2331
2332         /* This code looks awful: what it's doing is making sure an
2333          * ILLEGAL REQUEST sense return identifies the actual command
2334          * byte as the problem.  MODE_SENSE commands can return
2335          * ILLEGAL REQUEST if the code page isn't supported */
2336
2337         if (use_10_for_ms && !scsi_status_is_good(result) &&
2338             (driver_byte(result) & DRIVER_SENSE)) {
2339                 if (scsi_sense_valid(sshdr)) {
2340                         if ((sshdr->sense_key == ILLEGAL_REQUEST) &&
2341                             (sshdr->asc == 0x20) && (sshdr->ascq == 0)) {
2342                                 /* 
2343                                  * Invalid command operation code
2344                                  */
2345                                 sdev->use_10_for_ms = 0;
2346                                 goto retry;
2347                         }
2348                 }
2349         }
2350
2351         if(scsi_status_is_good(result)) {
2352                 if (unlikely(buffer[0] == 0x86 && buffer[1] == 0x0b &&
2353                              (modepage == 6 || modepage == 8))) {
2354                         /* Initio breakage? */
2355                         header_length = 0;
2356                         data->length = 13;
2357                         data->medium_type = 0;
2358                         data->device_specific = 0;
2359                         data->longlba = 0;
2360                         data->block_descriptor_length = 0;
2361                 } else if(use_10_for_ms) {
2362                         data->length = buffer[0]*256 + buffer[1] + 2;
2363                         data->medium_type = buffer[2];
2364                         data->device_specific = buffer[3];
2365                         data->longlba = buffer[4] & 0x01;
2366                         data->block_descriptor_length = buffer[6]*256
2367                                 + buffer[7];
2368                 } else {
2369                         data->length = buffer[0] + 1;
2370                         data->medium_type = buffer[1];
2371                         data->device_specific = buffer[2];
2372                         data->block_descriptor_length = buffer[3];
2373                 }
2374                 data->header_length = header_length;
2375         } else if ((status_byte(result) == CHECK_CONDITION) &&
2376                    scsi_sense_valid(sshdr) &&
2377                    sshdr->sense_key == UNIT_ATTENTION && retry_count) {
2378                 retry_count--;
2379                 goto retry;
2380         }
2381
2382         return result;
2383 }
2384 EXPORT_SYMBOL(scsi_mode_sense);
2385
2386 /**
2387  *      scsi_test_unit_ready - test if unit is ready
2388  *      @sdev:  scsi device to change the state of.
2389  *      @timeout: command timeout
2390  *      @retries: number of retries before failing
2391  *      @sshdr_external: Optional pointer to struct scsi_sense_hdr for
2392  *              returning sense. Make sure that this is cleared before passing
2393  *              in.
2394  *
2395  *      Returns zero if unsuccessful or an error if TUR failed.  For
2396  *      removable media, UNIT_ATTENTION sets ->changed flag.
2397  **/
2398 int
2399 scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries,
2400                      struct scsi_sense_hdr *sshdr_external)
2401 {
2402         char cmd[] = {
2403                 TEST_UNIT_READY, 0, 0, 0, 0, 0,
2404         };
2405         struct scsi_sense_hdr *sshdr;
2406         int result;
2407
2408         if (!sshdr_external)
2409                 sshdr = kzalloc(sizeof(*sshdr), GFP_KERNEL);
2410         else
2411                 sshdr = sshdr_external;
2412
2413         /* try to eat the UNIT_ATTENTION if there are enough retries */
2414         do {
2415                 result = scsi_execute_req(sdev, cmd, DMA_NONE, NULL, 0, sshdr,
2416                                           timeout, retries, NULL);
2417                 if (sdev->removable && scsi_sense_valid(sshdr) &&
2418                     sshdr->sense_key == UNIT_ATTENTION)
2419                         sdev->changed = 1;
2420         } while (scsi_sense_valid(sshdr) &&
2421                  sshdr->sense_key == UNIT_ATTENTION && --retries);
2422
2423         if (!sshdr_external)
2424                 kfree(sshdr);
2425         return result;
2426 }
2427 EXPORT_SYMBOL(scsi_test_unit_ready);
2428
2429 /**
2430  *      scsi_device_set_state - Take the given device through the device state model.
2431  *      @sdev:  scsi device to change the state of.
2432  *      @state: state to change to.
2433  *
2434  *      Returns zero if unsuccessful or an error if the requested 
2435  *      transition is illegal.
2436  */
2437 int
2438 scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
2439 {
2440         enum scsi_device_state oldstate = sdev->sdev_state;
2441
2442         if (state == oldstate)
2443                 return 0;
2444
2445         switch (state) {
2446         case SDEV_CREATED:
2447                 switch (oldstate) {
2448                 case SDEV_CREATED_BLOCK:
2449                         break;
2450                 default:
2451                         goto illegal;
2452                 }
2453                 break;
2454                         
2455         case SDEV_RUNNING:
2456                 switch (oldstate) {
2457                 case SDEV_CREATED:
2458                 case SDEV_OFFLINE:
2459                 case SDEV_TRANSPORT_OFFLINE:
2460                 case SDEV_QUIESCE:
2461                 case SDEV_BLOCK:
2462                         break;
2463                 default:
2464                         goto illegal;
2465                 }
2466                 break;
2467
2468         case SDEV_QUIESCE:
2469                 switch (oldstate) {
2470                 case SDEV_RUNNING:
2471                 case SDEV_OFFLINE:
2472                 case SDEV_TRANSPORT_OFFLINE:
2473                         break;
2474                 default:
2475                         goto illegal;
2476                 }
2477                 break;
2478
2479         case SDEV_OFFLINE:
2480         case SDEV_TRANSPORT_OFFLINE:
2481                 switch (oldstate) {
2482                 case SDEV_CREATED:
2483                 case SDEV_RUNNING:
2484                 case SDEV_QUIESCE:
2485                 case SDEV_BLOCK:
2486                         break;
2487                 default:
2488                         goto illegal;
2489                 }
2490                 break;
2491
2492         case SDEV_BLOCK:
2493                 switch (oldstate) {
2494                 case SDEV_RUNNING:
2495                 case SDEV_CREATED_BLOCK:
2496                         break;
2497                 default:
2498                         goto illegal;
2499                 }
2500                 break;
2501
2502         case SDEV_CREATED_BLOCK:
2503                 switch (oldstate) {
2504                 case SDEV_CREATED:
2505                         break;
2506                 default:
2507                         goto illegal;
2508                 }
2509                 break;
2510
2511         case SDEV_CANCEL:
2512                 switch (oldstate) {
2513                 case SDEV_CREATED:
2514                 case SDEV_RUNNING:
2515                 case SDEV_QUIESCE:
2516                 case SDEV_OFFLINE:
2517                 case SDEV_TRANSPORT_OFFLINE:
2518                 case SDEV_BLOCK:
2519                         break;
2520                 default:
2521                         goto illegal;
2522                 }
2523                 break;
2524
2525         case SDEV_DEL:
2526                 switch (oldstate) {
2527                 case SDEV_CREATED:
2528                 case SDEV_RUNNING:
2529                 case SDEV_OFFLINE:
2530                 case SDEV_TRANSPORT_OFFLINE:
2531                 case SDEV_CANCEL:
2532                 case SDEV_CREATED_BLOCK:
2533                         break;
2534                 default:
2535                         goto illegal;
2536                 }
2537                 break;
2538
2539         }
2540         sdev->sdev_state = state;
2541         return 0;
2542
2543  illegal:
2544         SCSI_LOG_ERROR_RECOVERY(1,
2545                                 sdev_printk(KERN_ERR, sdev,
2546                                             "Illegal state transition %s->%s",
2547                                             scsi_device_state_name(oldstate),
2548                                             scsi_device_state_name(state))
2549                                 );
2550         return -EINVAL;
2551 }
2552 EXPORT_SYMBOL(scsi_device_set_state);
2553
2554 /**
2555  *      sdev_evt_emit - emit a single SCSI device uevent
2556  *      @sdev: associated SCSI device
2557  *      @evt: event to emit
2558  *
2559  *      Send a single uevent (scsi_event) to the associated scsi_device.
2560  */
2561 static void scsi_evt_emit(struct scsi_device *sdev, struct scsi_event *evt)
2562 {
2563         int idx = 0;
2564         char *envp[3];
2565
2566         switch (evt->evt_type) {
2567         case SDEV_EVT_MEDIA_CHANGE:
2568                 envp[idx++] = "SDEV_MEDIA_CHANGE=1";
2569                 break;
2570         case SDEV_EVT_INQUIRY_CHANGE_REPORTED:
2571                 scsi_rescan_device(&sdev->sdev_gendev);
2572                 envp[idx++] = "SDEV_UA=INQUIRY_DATA_HAS_CHANGED";
2573                 break;
2574         case SDEV_EVT_CAPACITY_CHANGE_REPORTED:
2575                 envp[idx++] = "SDEV_UA=CAPACITY_DATA_HAS_CHANGED";
2576                 break;
2577         case SDEV_EVT_SOFT_THRESHOLD_REACHED_REPORTED:
2578                envp[idx++] = "SDEV_UA=THIN_PROVISIONING_SOFT_THRESHOLD_REACHED";
2579                 break;
2580         case SDEV_EVT_MODE_PARAMETER_CHANGE_REPORTED:
2581                 envp[idx++] = "SDEV_UA=MODE_PARAMETERS_CHANGED";
2582                 break;
2583         case SDEV_EVT_LUN_CHANGE_REPORTED:
2584                 envp[idx++] = "SDEV_UA=REPORTED_LUNS_DATA_HAS_CHANGED";
2585                 break;
2586         case SDEV_EVT_ALUA_STATE_CHANGE_REPORTED:
2587                 envp[idx++] = "SDEV_UA=ASYMMETRIC_ACCESS_STATE_CHANGED";
2588                 break;
2589         default:
2590                 /* do nothing */
2591                 break;
2592         }
2593
2594         envp[idx++] = NULL;
2595
2596         kobject_uevent_env(&sdev->sdev_gendev.kobj, KOBJ_CHANGE, envp);
2597 }
2598
2599 /**
2600  *      sdev_evt_thread - send a uevent for each scsi event
2601  *      @work: work struct for scsi_device
2602  *
2603  *      Dispatch queued events to their associated scsi_device kobjects
2604  *      as uevents.
2605  */
2606 void scsi_evt_thread(struct work_struct *work)
2607 {
2608         struct scsi_device *sdev;
2609         enum scsi_device_event evt_type;
2610         LIST_HEAD(event_list);
2611
2612         sdev = container_of(work, struct scsi_device, event_work);
2613
2614         for (evt_type = SDEV_EVT_FIRST; evt_type <= SDEV_EVT_LAST; evt_type++)
2615                 if (test_and_clear_bit(evt_type, sdev->pending_events))
2616                         sdev_evt_send_simple(sdev, evt_type, GFP_KERNEL);
2617
2618         while (1) {
2619                 struct scsi_event *evt;
2620                 struct list_head *this, *tmp;
2621                 unsigned long flags;
2622
2623                 spin_lock_irqsave(&sdev->list_lock, flags);
2624                 list_splice_init(&sdev->event_list, &event_list);
2625                 spin_unlock_irqrestore(&sdev->list_lock, flags);
2626
2627                 if (list_empty(&event_list))
2628                         break;
2629
2630                 list_for_each_safe(this, tmp, &event_list) {
2631                         evt = list_entry(this, struct scsi_event, node);
2632                         list_del(&evt->node);
2633                         scsi_evt_emit(sdev, evt);
2634                         kfree(evt);
2635                 }
2636         }
2637 }
2638
2639 /**
2640  *      sdev_evt_send - send asserted event to uevent thread
2641  *      @sdev: scsi_device event occurred on
2642  *      @evt: event to send
2643  *
2644  *      Assert scsi device event asynchronously.
2645  */
2646 void sdev_evt_send(struct scsi_device *sdev, struct scsi_event *evt)
2647 {
2648         unsigned long flags;
2649
2650 #if 0
2651         /* FIXME: currently this check eliminates all media change events
2652          * for polled devices.  Need to update to discriminate between AN
2653          * and polled events */
2654         if (!test_bit(evt->evt_type, sdev->supported_events)) {
2655                 kfree(evt);
2656                 return;
2657         }
2658 #endif
2659
2660         spin_lock_irqsave(&sdev->list_lock, flags);
2661         list_add_tail(&evt->node, &sdev->event_list);
2662         schedule_work(&sdev->event_work);
2663         spin_unlock_irqrestore(&sdev->list_lock, flags);
2664 }
2665 EXPORT_SYMBOL_GPL(sdev_evt_send);
2666
2667 /**
2668  *      sdev_evt_alloc - allocate a new scsi event
2669  *      @evt_type: type of event to allocate
2670  *      @gfpflags: GFP flags for allocation
2671  *
2672  *      Allocates and returns a new scsi_event.
2673  */
2674 struct scsi_event *sdev_evt_alloc(enum scsi_device_event evt_type,
2675                                   gfp_t gfpflags)
2676 {
2677         struct scsi_event *evt = kzalloc(sizeof(struct scsi_event), gfpflags);
2678         if (!evt)
2679                 return NULL;
2680
2681         evt->evt_type = evt_type;
2682         INIT_LIST_HEAD(&evt->node);
2683
2684         /* evt_type-specific initialization, if any */
2685         switch (evt_type) {
2686         case SDEV_EVT_MEDIA_CHANGE:
2687         case SDEV_EVT_INQUIRY_CHANGE_REPORTED:
2688         case SDEV_EVT_CAPACITY_CHANGE_REPORTED:
2689         case SDEV_EVT_SOFT_THRESHOLD_REACHED_REPORTED:
2690         case SDEV_EVT_MODE_PARAMETER_CHANGE_REPORTED:
2691         case SDEV_EVT_LUN_CHANGE_REPORTED:
2692         case SDEV_EVT_ALUA_STATE_CHANGE_REPORTED:
2693         default:
2694                 /* do nothing */
2695                 break;
2696         }
2697
2698         return evt;
2699 }
2700 EXPORT_SYMBOL_GPL(sdev_evt_alloc);
2701
2702 /**
2703  *      sdev_evt_send_simple - send asserted event to uevent thread
2704  *      @sdev: scsi_device event occurred on
2705  *      @evt_type: type of event to send
2706  *      @gfpflags: GFP flags for allocation
2707  *
2708  *      Assert scsi device event asynchronously, given an event type.
2709  */
2710 void sdev_evt_send_simple(struct scsi_device *sdev,
2711                           enum scsi_device_event evt_type, gfp_t gfpflags)
2712 {
2713         struct scsi_event *evt = sdev_evt_alloc(evt_type, gfpflags);
2714         if (!evt) {
2715                 sdev_printk(KERN_ERR, sdev, "event %d eaten due to OOM\n",
2716                             evt_type);
2717                 return;
2718         }
2719
2720         sdev_evt_send(sdev, evt);
2721 }
2722 EXPORT_SYMBOL_GPL(sdev_evt_send_simple);
2723
2724 /**
2725  *      scsi_device_quiesce - Block user issued commands.
2726  *      @sdev:  scsi device to quiesce.
2727  *
2728  *      This works by trying to transition to the SDEV_QUIESCE state
2729  *      (which must be a legal transition).  When the device is in this
2730  *      state, only special requests will be accepted, all others will
2731  *      be deferred.  Since special requests may also be requeued requests,
2732  *      a successful return doesn't guarantee the device will be 
2733  *      totally quiescent.
2734  *
2735  *      Must be called with user context, may sleep.
2736  *
2737  *      Returns zero if unsuccessful or an error if not.
2738  */
2739 int
2740 scsi_device_quiesce(struct scsi_device *sdev)
2741 {
2742         int err = scsi_device_set_state(sdev, SDEV_QUIESCE);
2743         if (err)
2744                 return err;
2745
2746         scsi_run_queue(sdev->request_queue);
2747         while (atomic_read(&sdev->device_busy)) {
2748                 msleep_interruptible(200);
2749                 scsi_run_queue(sdev->request_queue);
2750         }
2751         return 0;
2752 }
2753 EXPORT_SYMBOL(scsi_device_quiesce);
2754
2755 /**
2756  *      scsi_device_resume - Restart user issued commands to a quiesced device.
2757  *      @sdev:  scsi device to resume.
2758  *
2759  *      Moves the device from quiesced back to running and restarts the
2760  *      queues.
2761  *
2762  *      Must be called with user context, may sleep.
2763  */
2764 void scsi_device_resume(struct scsi_device *sdev)
2765 {
2766         /* check if the device state was mutated prior to resume, and if
2767          * so assume the state is being managed elsewhere (for example
2768          * device deleted during suspend)
2769          */
2770         if (sdev->sdev_state != SDEV_QUIESCE ||
2771             scsi_device_set_state(sdev, SDEV_RUNNING))
2772                 return;
2773         scsi_run_queue(sdev->request_queue);
2774 }
2775 EXPORT_SYMBOL(scsi_device_resume);
2776
2777 static void
2778 device_quiesce_fn(struct scsi_device *sdev, void *data)
2779 {
2780         scsi_device_quiesce(sdev);
2781 }
2782
2783 void
2784 scsi_target_quiesce(struct scsi_target *starget)
2785 {
2786         starget_for_each_device(starget, NULL, device_quiesce_fn);
2787 }
2788 EXPORT_SYMBOL(scsi_target_quiesce);
2789
2790 static void
2791 device_resume_fn(struct scsi_device *sdev, void *data)
2792 {
2793         scsi_device_resume(sdev);
2794 }
2795
2796 void
2797 scsi_target_resume(struct scsi_target *starget)
2798 {
2799         starget_for_each_device(starget, NULL, device_resume_fn);
2800 }
2801 EXPORT_SYMBOL(scsi_target_resume);
2802
2803 /**
2804  * scsi_internal_device_block - internal function to put a device temporarily into the SDEV_BLOCK state
2805  * @sdev:       device to block
2806  *
2807  * Block request made by scsi lld's to temporarily stop all
2808  * scsi commands on the specified device.  Called from interrupt
2809  * or normal process context.
2810  *
2811  * Returns zero if successful or error if not
2812  *
2813  * Notes:       
2814  *      This routine transitions the device to the SDEV_BLOCK state
2815  *      (which must be a legal transition).  When the device is in this
2816  *      state, all commands are deferred until the scsi lld reenables
2817  *      the device with scsi_device_unblock or device_block_tmo fires.
2818  */
2819 int
2820 scsi_internal_device_block(struct scsi_device *sdev)
2821 {
2822         struct request_queue *q = sdev->request_queue;
2823         unsigned long flags;
2824         int err = 0;
2825
2826         err = scsi_device_set_state(sdev, SDEV_BLOCK);
2827         if (err) {
2828                 err = scsi_device_set_state(sdev, SDEV_CREATED_BLOCK);
2829
2830                 if (err)
2831                         return err;
2832         }
2833
2834         /* 
2835          * The device has transitioned to SDEV_BLOCK.  Stop the
2836          * block layer from calling the midlayer with this device's
2837          * request queue. 
2838          */
2839         if (q->mq_ops) {
2840                 blk_mq_stop_hw_queues(q);
2841         } else {
2842                 spin_lock_irqsave(q->queue_lock, flags);
2843                 blk_stop_queue(q);
2844                 spin_unlock_irqrestore(q->queue_lock, flags);
2845         }
2846
2847         return 0;
2848 }
2849 EXPORT_SYMBOL_GPL(scsi_internal_device_block);
2850  
2851 /**
2852  * scsi_internal_device_unblock - resume a device after a block request
2853  * @sdev:       device to resume
2854  * @new_state:  state to set devices to after unblocking
2855  *
2856  * Called by scsi lld's or the midlayer to restart the device queue
2857  * for the previously suspended scsi device.  Called from interrupt or
2858  * normal process context.
2859  *
2860  * Returns zero if successful or error if not.
2861  *
2862  * Notes:       
2863  *      This routine transitions the device to the SDEV_RUNNING state
2864  *      or to one of the offline states (which must be a legal transition)
2865  *      allowing the midlayer to goose the queue for this device.
2866  */
2867 int
2868 scsi_internal_device_unblock(struct scsi_device *sdev,
2869                              enum scsi_device_state new_state)
2870 {
2871         struct request_queue *q = sdev->request_queue; 
2872         unsigned long flags;
2873
2874         /*
2875          * Try to transition the scsi device to SDEV_RUNNING or one of the
2876          * offlined states and goose the device queue if successful.
2877          */
2878         if ((sdev->sdev_state == SDEV_BLOCK) ||
2879             (sdev->sdev_state == SDEV_TRANSPORT_OFFLINE))
2880                 sdev->sdev_state = new_state;
2881         else if (sdev->sdev_state == SDEV_CREATED_BLOCK) {
2882                 if (new_state == SDEV_TRANSPORT_OFFLINE ||
2883                     new_state == SDEV_OFFLINE)
2884                         sdev->sdev_state = new_state;
2885                 else
2886                         sdev->sdev_state = SDEV_CREATED;
2887         } else if (sdev->sdev_state != SDEV_CANCEL &&
2888                  sdev->sdev_state != SDEV_OFFLINE)
2889                 return -EINVAL;
2890
2891         if (q->mq_ops) {
2892                 blk_mq_start_stopped_hw_queues(q, false);
2893         } else {
2894                 spin_lock_irqsave(q->queue_lock, flags);
2895                 blk_start_queue(q);
2896                 spin_unlock_irqrestore(q->queue_lock, flags);
2897         }
2898
2899         return 0;
2900 }
2901 EXPORT_SYMBOL_GPL(scsi_internal_device_unblock);
2902
2903 static void
2904 device_block(struct scsi_device *sdev, void *data)
2905 {
2906         scsi_internal_device_block(sdev);
2907 }
2908
2909 static int
2910 target_block(struct device *dev, void *data)
2911 {
2912         if (scsi_is_target_device(dev))
2913                 starget_for_each_device(to_scsi_target(dev), NULL,
2914                                         device_block);
2915         return 0;
2916 }
2917
2918 void
2919 scsi_target_block(struct device *dev)
2920 {
2921         if (scsi_is_target_device(dev))
2922                 starget_for_each_device(to_scsi_target(dev), NULL,
2923                                         device_block);
2924         else
2925                 device_for_each_child(dev, NULL, target_block);
2926 }
2927 EXPORT_SYMBOL_GPL(scsi_target_block);
2928
2929 static void
2930 device_unblock(struct scsi_device *sdev, void *data)
2931 {
2932         scsi_internal_device_unblock(sdev, *(enum scsi_device_state *)data);
2933 }
2934
2935 static int
2936 target_unblock(struct device *dev, void *data)
2937 {
2938         if (scsi_is_target_device(dev))
2939                 starget_for_each_device(to_scsi_target(dev), data,
2940                                         device_unblock);
2941         return 0;
2942 }
2943
2944 void
2945 scsi_target_unblock(struct device *dev, enum scsi_device_state new_state)
2946 {
2947         if (scsi_is_target_device(dev))
2948                 starget_for_each_device(to_scsi_target(dev), &new_state,
2949                                         device_unblock);
2950         else
2951                 device_for_each_child(dev, &new_state, target_unblock);
2952 }
2953 EXPORT_SYMBOL_GPL(scsi_target_unblock);
2954
2955 /**
2956  * scsi_kmap_atomic_sg - find and atomically map an sg-elemnt
2957  * @sgl:        scatter-gather list
2958  * @sg_count:   number of segments in sg
2959  * @offset:     offset in bytes into sg, on return offset into the mapped area
2960  * @len:        bytes to map, on return number of bytes mapped
2961  *
2962  * Returns virtual address of the start of the mapped page
2963  */
2964 void *scsi_kmap_atomic_sg(struct scatterlist *sgl, int sg_count,
2965                           size_t *offset, size_t *len)
2966 {
2967         int i;
2968         size_t sg_len = 0, len_complete = 0;
2969         struct scatterlist *sg;
2970         struct page *page;
2971
2972         WARN_ON(!irqs_disabled());
2973
2974         for_each_sg(sgl, sg, sg_count, i) {
2975                 len_complete = sg_len; /* Complete sg-entries */
2976                 sg_len += sg->length;
2977                 if (sg_len > *offset)
2978                         break;
2979         }
2980
2981         if (unlikely(i == sg_count)) {
2982                 printk(KERN_ERR "%s: Bytes in sg: %zu, requested offset %zu, "
2983                         "elements %d\n",
2984                        __func__, sg_len, *offset, sg_count);
2985                 WARN_ON(1);
2986                 return NULL;
2987         }
2988
2989         /* Offset starting from the beginning of first page in this sg-entry */
2990         *offset = *offset - len_complete + sg->offset;
2991
2992         /* Assumption: contiguous pages can be accessed as "page + i" */
2993         page = nth_page(sg_page(sg), (*offset >> PAGE_SHIFT));
2994         *offset &= ~PAGE_MASK;
2995
2996         /* Bytes in this sg-entry from *offset to the end of the page */
2997         sg_len = PAGE_SIZE - *offset;
2998         if (*len > sg_len)
2999                 *len = sg_len;
3000
3001         return kmap_atomic(page);
3002 }
3003 EXPORT_SYMBOL(scsi_kmap_atomic_sg);
3004
3005 /**
3006  * scsi_kunmap_atomic_sg - atomically unmap a virtual address, previously mapped with scsi_kmap_atomic_sg
3007  * @virt:       virtual address to be unmapped
3008  */
3009 void scsi_kunmap_atomic_sg(void *virt)
3010 {
3011         kunmap_atomic(virt);
3012 }
3013 EXPORT_SYMBOL(scsi_kunmap_atomic_sg);
3014
3015 void sdev_disable_disk_events(struct scsi_device *sdev)
3016 {
3017         atomic_inc(&sdev->disk_events_disable_depth);
3018 }
3019 EXPORT_SYMBOL(sdev_disable_disk_events);
3020
3021 void sdev_enable_disk_events(struct scsi_device *sdev)
3022 {
3023         if (WARN_ON_ONCE(atomic_read(&sdev->disk_events_disable_depth) <= 0))
3024                 return;
3025         atomic_dec(&sdev->disk_events_disable_depth);
3026 }
3027 EXPORT_SYMBOL(sdev_enable_disk_events);
3028
3029 /**
3030  * scsi_vpd_lun_id - return a unique device identification
3031  * @sdev: SCSI device
3032  * @id:   buffer for the identification
3033  * @id_len:  length of the buffer
3034  *
3035  * Copies a unique device identification into @id based
3036  * on the information in the VPD page 0x83 of the device.
3037  * The string will be formatted as a SCSI name string.
3038  *
3039  * Returns the length of the identification or error on failure.
3040  * If the identifier is longer than the supplied buffer the actual
3041  * identifier length is returned and the buffer is not zero-padded.
3042  */
3043 int scsi_vpd_lun_id(struct scsi_device *sdev, char *id, size_t id_len)
3044 {
3045         u8 cur_id_type = 0xff;
3046         u8 cur_id_size = 0;
3047         unsigned char *d, *cur_id_str;
3048         unsigned char __rcu *vpd_pg83;
3049         int id_size = -EINVAL;
3050
3051         rcu_read_lock();
3052         vpd_pg83 = rcu_dereference(sdev->vpd_pg83);
3053         if (!vpd_pg83) {
3054                 rcu_read_unlock();
3055                 return -ENXIO;
3056         }
3057
3058         /*
3059          * Look for the correct descriptor.
3060          * Order of preference for lun descriptor:
3061          * - SCSI name string
3062          * - NAA IEEE Registered Extended
3063          * - EUI-64 based 16-byte
3064          * - EUI-64 based 12-byte
3065          * - NAA IEEE Registered
3066          * - NAA IEEE Extended
3067          * - T10 Vendor ID
3068          * as longer descriptors reduce the likelyhood
3069          * of identification clashes.
3070          */
3071
3072         /* The id string must be at least 20 bytes + terminating NULL byte */
3073         if (id_len < 21) {
3074                 rcu_read_unlock();
3075                 return -EINVAL;
3076         }
3077
3078         memset(id, 0, id_len);
3079         d = vpd_pg83 + 4;
3080         while (d < vpd_pg83 + sdev->vpd_pg83_len) {
3081                 /* Skip designators not referring to the LUN */
3082                 if ((d[1] & 0x30) != 0x00)
3083                         goto next_desig;
3084
3085                 switch (d[1] & 0xf) {
3086                 case 0x1:
3087                         /* T10 Vendor ID */
3088                         if (cur_id_size > d[3])
3089                                 break;
3090                         /* Prefer anything */
3091                         if (cur_id_type > 0x01 && cur_id_type != 0xff)
3092                                 break;
3093                         cur_id_size = d[3];
3094                         if (cur_id_size + 4 > id_len)
3095                                 cur_id_size = id_len - 4;
3096                         cur_id_str = d + 4;
3097                         cur_id_type = d[1] & 0xf;
3098                         id_size = snprintf(id, id_len, "t10.%*pE",
3099                                            cur_id_size, cur_id_str);
3100                         break;
3101                 case 0x2:
3102                         /* EUI-64 */
3103                         if (cur_id_size > d[3])
3104                                 break;
3105                         /* Prefer NAA IEEE Registered Extended */
3106                         if (cur_id_type == 0x3 &&
3107                             cur_id_size == d[3])
3108                                 break;
3109                         cur_id_size = d[3];
3110                         cur_id_str = d + 4;
3111                         cur_id_type = d[1] & 0xf;
3112                         switch (cur_id_size) {
3113                         case 8:
3114                                 id_size = snprintf(id, id_len,
3115                                                    "eui.%8phN",
3116                                                    cur_id_str);
3117                                 break;
3118                         case 12:
3119                                 id_size = snprintf(id, id_len,
3120                                                    "eui.%12phN",
3121                                                    cur_id_str);
3122                                 break;
3123                         case 16:
3124                                 id_size = snprintf(id, id_len,
3125                                                    "eui.%16phN",
3126                                                    cur_id_str);
3127                                 break;
3128                         default:
3129                                 cur_id_size = 0;
3130                                 break;
3131                         }
3132                         break;
3133                 case 0x3:
3134                         /* NAA */
3135                         if (cur_id_size > d[3])
3136                                 break;
3137                         cur_id_size = d[3];
3138                         cur_id_str = d + 4;
3139                         cur_id_type = d[1] & 0xf;
3140                         switch (cur_id_size) {
3141                         case 8:
3142                                 id_size = snprintf(id, id_len,
3143                                                    "naa.%8phN",
3144                                                    cur_id_str);
3145                                 break;
3146                         case 16:
3147                                 id_size = snprintf(id, id_len,
3148                                                    "naa.%16phN",
3149                                                    cur_id_str);
3150                                 break;
3151                         default:
3152                                 cur_id_size = 0;
3153                                 break;
3154                         }
3155                         break;
3156                 case 0x8:
3157                         /* SCSI name string */
3158                         if (cur_id_size + 4 > d[3])
3159                                 break;
3160                         /* Prefer others for truncated descriptor */
3161                         if (cur_id_size && d[3] > id_len)
3162                                 break;
3163                         cur_id_size = id_size = d[3];
3164                         cur_id_str = d + 4;
3165                         cur_id_type = d[1] & 0xf;
3166                         if (cur_id_size >= id_len)
3167                                 cur_id_size = id_len - 1;
3168                         memcpy(id, cur_id_str, cur_id_size);
3169                         /* Decrease priority for truncated descriptor */
3170                         if (cur_id_size != id_size)
3171                                 cur_id_size = 6;
3172                         break;
3173                 default:
3174                         break;
3175                 }
3176 next_desig:
3177                 d += d[3] + 4;
3178         }
3179         rcu_read_unlock();
3180
3181         return id_size;
3182 }
3183 EXPORT_SYMBOL(scsi_vpd_lun_id);
3184
3185 /*
3186  * scsi_vpd_tpg_id - return a target port group identifier
3187  * @sdev: SCSI device
3188  *
3189  * Returns the Target Port Group identifier from the information
3190  * froom VPD page 0x83 of the device.
3191  *
3192  * Returns the identifier or error on failure.
3193  */
3194 int scsi_vpd_tpg_id(struct scsi_device *sdev, int *rel_id)
3195 {
3196         unsigned char *d;
3197         unsigned char __rcu *vpd_pg83;
3198         int group_id = -EAGAIN, rel_port = -1;
3199
3200         rcu_read_lock();
3201         vpd_pg83 = rcu_dereference(sdev->vpd_pg83);
3202         if (!vpd_pg83) {
3203                 rcu_read_unlock();
3204                 return -ENXIO;
3205         }
3206
3207         d = sdev->vpd_pg83 + 4;
3208         while (d < sdev->vpd_pg83 + sdev->vpd_pg83_len) {
3209                 switch (d[1] & 0xf) {
3210                 case 0x4:
3211                         /* Relative target port */
3212                         rel_port = get_unaligned_be16(&d[6]);
3213                         break;
3214                 case 0x5:
3215                         /* Target port group */
3216                         group_id = get_unaligned_be16(&d[6]);
3217                         break;
3218                 default:
3219                         break;
3220                 }
3221                 d += d[3] + 4;
3222         }
3223         rcu_read_unlock();
3224
3225         if (group_id >= 0 && rel_id && rel_port != -1)
3226                 *rel_id = rel_port;
3227
3228         return group_id;
3229 }
3230 EXPORT_SYMBOL(scsi_vpd_tpg_id);