xref: /openbmc/linux/drivers/scsi/scsi_lib.c (revision 92b19ff5)
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/mempool.h>
18 #include <linux/slab.h>
19 #include <linux/init.h>
20 #include <linux/pci.h>
21 #include <linux/delay.h>
22 #include <linux/hardirq.h>
23 #include <linux/scatterlist.h>
24 #include <linux/blk-mq.h>
25 #include <linux/ratelimit.h>
26 
27 #include <scsi/scsi.h>
28 #include <scsi/scsi_cmnd.h>
29 #include <scsi/scsi_dbg.h>
30 #include <scsi/scsi_device.h>
31 #include <scsi/scsi_driver.h>
32 #include <scsi/scsi_eh.h>
33 #include <scsi/scsi_host.h>
34 
35 #include <trace/events/scsi.h>
36 
37 #include "scsi_priv.h"
38 #include "scsi_logging.h"
39 
40 
41 #define SG_MEMPOOL_NR		ARRAY_SIZE(scsi_sg_pools)
42 #define SG_MEMPOOL_SIZE		2
43 
44 struct scsi_host_sg_pool {
45 	size_t		size;
46 	char		*name;
47 	struct kmem_cache	*slab;
48 	mempool_t	*pool;
49 };
50 
51 #define SP(x) { .size = x, "sgpool-" __stringify(x) }
52 #if (SCSI_MAX_SG_SEGMENTS < 32)
53 #error SCSI_MAX_SG_SEGMENTS is too small (must be 32 or greater)
54 #endif
55 static struct scsi_host_sg_pool scsi_sg_pools[] = {
56 	SP(8),
57 	SP(16),
58 #if (SCSI_MAX_SG_SEGMENTS > 32)
59 	SP(32),
60 #if (SCSI_MAX_SG_SEGMENTS > 64)
61 	SP(64),
62 #if (SCSI_MAX_SG_SEGMENTS > 128)
63 	SP(128),
64 #if (SCSI_MAX_SG_SEGMENTS > 256)
65 #error SCSI_MAX_SG_SEGMENTS is too large (256 MAX)
66 #endif
67 #endif
68 #endif
69 #endif
70 	SP(SCSI_MAX_SG_SEGMENTS)
71 };
72 #undef SP
73 
74 struct kmem_cache *scsi_sdb_cache;
75 
76 /*
77  * When to reinvoke queueing after a resource shortage. It's 3 msecs to
78  * not change behaviour from the previous unplug mechanism, experimentation
79  * may prove this needs changing.
80  */
81 #define SCSI_QUEUE_DELAY	3
82 
83 static void
84 scsi_set_blocked(struct scsi_cmnd *cmd, int reason)
85 {
86 	struct Scsi_Host *host = cmd->device->host;
87 	struct scsi_device *device = cmd->device;
88 	struct scsi_target *starget = scsi_target(device);
89 
90 	/*
91 	 * Set the appropriate busy bit for the device/host.
92 	 *
93 	 * If the host/device isn't busy, assume that something actually
94 	 * completed, and that we should be able to queue a command now.
95 	 *
96 	 * Note that the prior mid-layer assumption that any host could
97 	 * always queue at least one command is now broken.  The mid-layer
98 	 * will implement a user specifiable stall (see
99 	 * scsi_host.max_host_blocked and scsi_device.max_device_blocked)
100 	 * if a command is requeued with no other commands outstanding
101 	 * either for the device or for the host.
102 	 */
103 	switch (reason) {
104 	case SCSI_MLQUEUE_HOST_BUSY:
105 		atomic_set(&host->host_blocked, host->max_host_blocked);
106 		break;
107 	case SCSI_MLQUEUE_DEVICE_BUSY:
108 	case SCSI_MLQUEUE_EH_RETRY:
109 		atomic_set(&device->device_blocked,
110 			   device->max_device_blocked);
111 		break;
112 	case SCSI_MLQUEUE_TARGET_BUSY:
113 		atomic_set(&starget->target_blocked,
114 			   starget->max_target_blocked);
115 		break;
116 	}
117 }
118 
119 static void scsi_mq_requeue_cmd(struct scsi_cmnd *cmd)
120 {
121 	struct scsi_device *sdev = cmd->device;
122 	struct request_queue *q = cmd->request->q;
123 
124 	blk_mq_requeue_request(cmd->request);
125 	blk_mq_kick_requeue_list(q);
126 	put_device(&sdev->sdev_gendev);
127 }
128 
129 /**
130  * __scsi_queue_insert - private queue insertion
131  * @cmd: The SCSI command being requeued
132  * @reason:  The reason for the requeue
133  * @unbusy: Whether the queue should be unbusied
134  *
135  * This is a private queue insertion.  The public interface
136  * scsi_queue_insert() always assumes the queue should be unbusied
137  * because it's always called before the completion.  This function is
138  * for a requeue after completion, which should only occur in this
139  * file.
140  */
141 static void __scsi_queue_insert(struct scsi_cmnd *cmd, int reason, int unbusy)
142 {
143 	struct scsi_device *device = cmd->device;
144 	struct request_queue *q = device->request_queue;
145 	unsigned long flags;
146 
147 	SCSI_LOG_MLQUEUE(1, scmd_printk(KERN_INFO, cmd,
148 		"Inserting command %p into mlqueue\n", cmd));
149 
150 	scsi_set_blocked(cmd, reason);
151 
152 	/*
153 	 * Decrement the counters, since these commands are no longer
154 	 * active on the host/device.
155 	 */
156 	if (unbusy)
157 		scsi_device_unbusy(device);
158 
159 	/*
160 	 * Requeue this command.  It will go before all other commands
161 	 * that are already in the queue. Schedule requeue work under
162 	 * lock such that the kblockd_schedule_work() call happens
163 	 * before blk_cleanup_queue() finishes.
164 	 */
165 	cmd->result = 0;
166 	if (q->mq_ops) {
167 		scsi_mq_requeue_cmd(cmd);
168 		return;
169 	}
170 	spin_lock_irqsave(q->queue_lock, flags);
171 	blk_requeue_request(q, cmd->request);
172 	kblockd_schedule_work(&device->requeue_work);
173 	spin_unlock_irqrestore(q->queue_lock, flags);
174 }
175 
176 /*
177  * Function:    scsi_queue_insert()
178  *
179  * Purpose:     Insert a command in the midlevel queue.
180  *
181  * Arguments:   cmd    - command that we are adding to queue.
182  *              reason - why we are inserting command to queue.
183  *
184  * Lock status: Assumed that lock is not held upon entry.
185  *
186  * Returns:     Nothing.
187  *
188  * Notes:       We do this for one of two cases.  Either the host is busy
189  *              and it cannot accept any more commands for the time being,
190  *              or the device returned QUEUE_FULL and can accept no more
191  *              commands.
192  * Notes:       This could be called either from an interrupt context or a
193  *              normal process context.
194  */
195 void scsi_queue_insert(struct scsi_cmnd *cmd, int reason)
196 {
197 	__scsi_queue_insert(cmd, reason, 1);
198 }
199 /**
200  * scsi_execute - insert request and wait for the result
201  * @sdev:	scsi device
202  * @cmd:	scsi command
203  * @data_direction: data direction
204  * @buffer:	data buffer
205  * @bufflen:	len of buffer
206  * @sense:	optional sense buffer
207  * @timeout:	request timeout in seconds
208  * @retries:	number of times to retry request
209  * @flags:	or into request flags;
210  * @resid:	optional residual length
211  *
212  * returns the req->errors value which is the scsi_cmnd result
213  * field.
214  */
215 int scsi_execute(struct scsi_device *sdev, const unsigned char *cmd,
216 		 int data_direction, void *buffer, unsigned bufflen,
217 		 unsigned char *sense, int timeout, int retries, u64 flags,
218 		 int *resid)
219 {
220 	struct request *req;
221 	int write = (data_direction == DMA_TO_DEVICE);
222 	int ret = DRIVER_ERROR << 24;
223 
224 	req = blk_get_request(sdev->request_queue, write, __GFP_WAIT);
225 	if (IS_ERR(req))
226 		return ret;
227 	blk_rq_set_block_pc(req);
228 
229 	if (bufflen &&	blk_rq_map_kern(sdev->request_queue, req,
230 					buffer, bufflen, __GFP_WAIT))
231 		goto out;
232 
233 	req->cmd_len = COMMAND_SIZE(cmd[0]);
234 	memcpy(req->cmd, cmd, req->cmd_len);
235 	req->sense = sense;
236 	req->sense_len = 0;
237 	req->retries = retries;
238 	req->timeout = timeout;
239 	req->cmd_flags |= flags | REQ_QUIET | REQ_PREEMPT;
240 
241 	/*
242 	 * head injection *required* here otherwise quiesce won't work
243 	 */
244 	blk_execute_rq(req->q, NULL, req, 1);
245 
246 	/*
247 	 * Some devices (USB mass-storage in particular) may transfer
248 	 * garbage data together with a residue indicating that the data
249 	 * is invalid.  Prevent the garbage from being misinterpreted
250 	 * and prevent security leaks by zeroing out the excess data.
251 	 */
252 	if (unlikely(req->resid_len > 0 && req->resid_len <= bufflen))
253 		memset(buffer + (bufflen - req->resid_len), 0, req->resid_len);
254 
255 	if (resid)
256 		*resid = req->resid_len;
257 	ret = req->errors;
258  out:
259 	blk_put_request(req);
260 
261 	return ret;
262 }
263 EXPORT_SYMBOL(scsi_execute);
264 
265 int scsi_execute_req_flags(struct scsi_device *sdev, const unsigned char *cmd,
266 		     int data_direction, void *buffer, unsigned bufflen,
267 		     struct scsi_sense_hdr *sshdr, int timeout, int retries,
268 		     int *resid, u64 flags)
269 {
270 	char *sense = NULL;
271 	int result;
272 
273 	if (sshdr) {
274 		sense = kzalloc(SCSI_SENSE_BUFFERSIZE, GFP_NOIO);
275 		if (!sense)
276 			return DRIVER_ERROR << 24;
277 	}
278 	result = scsi_execute(sdev, cmd, data_direction, buffer, bufflen,
279 			      sense, timeout, retries, flags, resid);
280 	if (sshdr)
281 		scsi_normalize_sense(sense, SCSI_SENSE_BUFFERSIZE, sshdr);
282 
283 	kfree(sense);
284 	return result;
285 }
286 EXPORT_SYMBOL(scsi_execute_req_flags);
287 
288 /*
289  * Function:    scsi_init_cmd_errh()
290  *
291  * Purpose:     Initialize cmd fields related to error handling.
292  *
293  * Arguments:   cmd	- command that is ready to be queued.
294  *
295  * Notes:       This function has the job of initializing a number of
296  *              fields related to error handling.   Typically this will
297  *              be called once for each command, as required.
298  */
299 static void scsi_init_cmd_errh(struct scsi_cmnd *cmd)
300 {
301 	cmd->serial_number = 0;
302 	scsi_set_resid(cmd, 0);
303 	memset(cmd->sense_buffer, 0, SCSI_SENSE_BUFFERSIZE);
304 	if (cmd->cmd_len == 0)
305 		cmd->cmd_len = scsi_command_size(cmd->cmnd);
306 }
307 
308 void scsi_device_unbusy(struct scsi_device *sdev)
309 {
310 	struct Scsi_Host *shost = sdev->host;
311 	struct scsi_target *starget = scsi_target(sdev);
312 	unsigned long flags;
313 
314 	atomic_dec(&shost->host_busy);
315 	if (starget->can_queue > 0)
316 		atomic_dec(&starget->target_busy);
317 
318 	if (unlikely(scsi_host_in_recovery(shost) &&
319 		     (shost->host_failed || shost->host_eh_scheduled))) {
320 		spin_lock_irqsave(shost->host_lock, flags);
321 		scsi_eh_wakeup(shost);
322 		spin_unlock_irqrestore(shost->host_lock, flags);
323 	}
324 
325 	atomic_dec(&sdev->device_busy);
326 }
327 
328 static void scsi_kick_queue(struct request_queue *q)
329 {
330 	if (q->mq_ops)
331 		blk_mq_start_hw_queues(q);
332 	else
333 		blk_run_queue(q);
334 }
335 
336 /*
337  * Called for single_lun devices on IO completion. Clear starget_sdev_user,
338  * and call blk_run_queue for all the scsi_devices on the target -
339  * including current_sdev first.
340  *
341  * Called with *no* scsi locks held.
342  */
343 static void scsi_single_lun_run(struct scsi_device *current_sdev)
344 {
345 	struct Scsi_Host *shost = current_sdev->host;
346 	struct scsi_device *sdev, *tmp;
347 	struct scsi_target *starget = scsi_target(current_sdev);
348 	unsigned long flags;
349 
350 	spin_lock_irqsave(shost->host_lock, flags);
351 	starget->starget_sdev_user = NULL;
352 	spin_unlock_irqrestore(shost->host_lock, flags);
353 
354 	/*
355 	 * Call blk_run_queue for all LUNs on the target, starting with
356 	 * current_sdev. We race with others (to set starget_sdev_user),
357 	 * but in most cases, we will be first. Ideally, each LU on the
358 	 * target would get some limited time or requests on the target.
359 	 */
360 	scsi_kick_queue(current_sdev->request_queue);
361 
362 	spin_lock_irqsave(shost->host_lock, flags);
363 	if (starget->starget_sdev_user)
364 		goto out;
365 	list_for_each_entry_safe(sdev, tmp, &starget->devices,
366 			same_target_siblings) {
367 		if (sdev == current_sdev)
368 			continue;
369 		if (scsi_device_get(sdev))
370 			continue;
371 
372 		spin_unlock_irqrestore(shost->host_lock, flags);
373 		scsi_kick_queue(sdev->request_queue);
374 		spin_lock_irqsave(shost->host_lock, flags);
375 
376 		scsi_device_put(sdev);
377 	}
378  out:
379 	spin_unlock_irqrestore(shost->host_lock, flags);
380 }
381 
382 static inline bool scsi_device_is_busy(struct scsi_device *sdev)
383 {
384 	if (atomic_read(&sdev->device_busy) >= sdev->queue_depth)
385 		return true;
386 	if (atomic_read(&sdev->device_blocked) > 0)
387 		return true;
388 	return false;
389 }
390 
391 static inline bool scsi_target_is_busy(struct scsi_target *starget)
392 {
393 	if (starget->can_queue > 0) {
394 		if (atomic_read(&starget->target_busy) >= starget->can_queue)
395 			return true;
396 		if (atomic_read(&starget->target_blocked) > 0)
397 			return true;
398 	}
399 	return false;
400 }
401 
402 static inline bool scsi_host_is_busy(struct Scsi_Host *shost)
403 {
404 	if (shost->can_queue > 0 &&
405 	    atomic_read(&shost->host_busy) >= shost->can_queue)
406 		return true;
407 	if (atomic_read(&shost->host_blocked) > 0)
408 		return true;
409 	if (shost->host_self_blocked)
410 		return true;
411 	return false;
412 }
413 
414 static void scsi_starved_list_run(struct Scsi_Host *shost)
415 {
416 	LIST_HEAD(starved_list);
417 	struct scsi_device *sdev;
418 	unsigned long flags;
419 
420 	spin_lock_irqsave(shost->host_lock, flags);
421 	list_splice_init(&shost->starved_list, &starved_list);
422 
423 	while (!list_empty(&starved_list)) {
424 		struct request_queue *slq;
425 
426 		/*
427 		 * As long as shost is accepting commands and we have
428 		 * starved queues, call blk_run_queue. scsi_request_fn
429 		 * drops the queue_lock and can add us back to the
430 		 * starved_list.
431 		 *
432 		 * host_lock protects the starved_list and starved_entry.
433 		 * scsi_request_fn must get the host_lock before checking
434 		 * or modifying starved_list or starved_entry.
435 		 */
436 		if (scsi_host_is_busy(shost))
437 			break;
438 
439 		sdev = list_entry(starved_list.next,
440 				  struct scsi_device, starved_entry);
441 		list_del_init(&sdev->starved_entry);
442 		if (scsi_target_is_busy(scsi_target(sdev))) {
443 			list_move_tail(&sdev->starved_entry,
444 				       &shost->starved_list);
445 			continue;
446 		}
447 
448 		/*
449 		 * Once we drop the host lock, a racing scsi_remove_device()
450 		 * call may remove the sdev from the starved list and destroy
451 		 * it and the queue.  Mitigate by taking a reference to the
452 		 * queue and never touching the sdev again after we drop the
453 		 * host lock.  Note: if __scsi_remove_device() invokes
454 		 * blk_cleanup_queue() before the queue is run from this
455 		 * function then blk_run_queue() will return immediately since
456 		 * blk_cleanup_queue() marks the queue with QUEUE_FLAG_DYING.
457 		 */
458 		slq = sdev->request_queue;
459 		if (!blk_get_queue(slq))
460 			continue;
461 		spin_unlock_irqrestore(shost->host_lock, flags);
462 
463 		scsi_kick_queue(slq);
464 		blk_put_queue(slq);
465 
466 		spin_lock_irqsave(shost->host_lock, flags);
467 	}
468 	/* put any unprocessed entries back */
469 	list_splice(&starved_list, &shost->starved_list);
470 	spin_unlock_irqrestore(shost->host_lock, flags);
471 }
472 
473 /*
474  * Function:   scsi_run_queue()
475  *
476  * Purpose:    Select a proper request queue to serve next
477  *
478  * Arguments:  q       - last request's queue
479  *
480  * Returns:     Nothing
481  *
482  * Notes:      The previous command was completely finished, start
483  *             a new one if possible.
484  */
485 static void scsi_run_queue(struct request_queue *q)
486 {
487 	struct scsi_device *sdev = q->queuedata;
488 
489 	if (scsi_target(sdev)->single_lun)
490 		scsi_single_lun_run(sdev);
491 	if (!list_empty(&sdev->host->starved_list))
492 		scsi_starved_list_run(sdev->host);
493 
494 	if (q->mq_ops)
495 		blk_mq_start_stopped_hw_queues(q, false);
496 	else
497 		blk_run_queue(q);
498 }
499 
500 void scsi_requeue_run_queue(struct work_struct *work)
501 {
502 	struct scsi_device *sdev;
503 	struct request_queue *q;
504 
505 	sdev = container_of(work, struct scsi_device, requeue_work);
506 	q = sdev->request_queue;
507 	scsi_run_queue(q);
508 }
509 
510 /*
511  * Function:	scsi_requeue_command()
512  *
513  * Purpose:	Handle post-processing of completed commands.
514  *
515  * Arguments:	q	- queue to operate on
516  *		cmd	- command that may need to be requeued.
517  *
518  * Returns:	Nothing
519  *
520  * Notes:	After command completion, there may be blocks left
521  *		over which weren't finished by the previous command
522  *		this can be for a number of reasons - the main one is
523  *		I/O errors in the middle of the request, in which case
524  *		we need to request the blocks that come after the bad
525  *		sector.
526  * Notes:	Upon return, cmd is a stale pointer.
527  */
528 static void scsi_requeue_command(struct request_queue *q, struct scsi_cmnd *cmd)
529 {
530 	struct scsi_device *sdev = cmd->device;
531 	struct request *req = cmd->request;
532 	unsigned long flags;
533 
534 	spin_lock_irqsave(q->queue_lock, flags);
535 	blk_unprep_request(req);
536 	req->special = NULL;
537 	scsi_put_command(cmd);
538 	blk_requeue_request(q, req);
539 	spin_unlock_irqrestore(q->queue_lock, flags);
540 
541 	scsi_run_queue(q);
542 
543 	put_device(&sdev->sdev_gendev);
544 }
545 
546 void scsi_run_host_queues(struct Scsi_Host *shost)
547 {
548 	struct scsi_device *sdev;
549 
550 	shost_for_each_device(sdev, shost)
551 		scsi_run_queue(sdev->request_queue);
552 }
553 
554 static inline unsigned int scsi_sgtable_index(unsigned short nents)
555 {
556 	unsigned int index;
557 
558 	BUG_ON(nents > SCSI_MAX_SG_SEGMENTS);
559 
560 	if (nents <= 8)
561 		index = 0;
562 	else
563 		index = get_count_order(nents) - 3;
564 
565 	return index;
566 }
567 
568 static void scsi_sg_free(struct scatterlist *sgl, unsigned int nents)
569 {
570 	struct scsi_host_sg_pool *sgp;
571 
572 	sgp = scsi_sg_pools + scsi_sgtable_index(nents);
573 	mempool_free(sgl, sgp->pool);
574 }
575 
576 static struct scatterlist *scsi_sg_alloc(unsigned int nents, gfp_t gfp_mask)
577 {
578 	struct scsi_host_sg_pool *sgp;
579 
580 	sgp = scsi_sg_pools + scsi_sgtable_index(nents);
581 	return mempool_alloc(sgp->pool, gfp_mask);
582 }
583 
584 static void scsi_free_sgtable(struct scsi_data_buffer *sdb, bool mq)
585 {
586 	if (mq && sdb->table.nents <= SCSI_MAX_SG_SEGMENTS)
587 		return;
588 	__sg_free_table(&sdb->table, SCSI_MAX_SG_SEGMENTS, mq, scsi_sg_free);
589 }
590 
591 static int scsi_alloc_sgtable(struct scsi_data_buffer *sdb, int nents, bool mq)
592 {
593 	struct scatterlist *first_chunk = NULL;
594 	int ret;
595 
596 	BUG_ON(!nents);
597 
598 	if (mq) {
599 		if (nents <= SCSI_MAX_SG_SEGMENTS) {
600 			sdb->table.nents = nents;
601 			sg_init_table(sdb->table.sgl, sdb->table.nents);
602 			return 0;
603 		}
604 		first_chunk = sdb->table.sgl;
605 	}
606 
607 	ret = __sg_alloc_table(&sdb->table, nents, SCSI_MAX_SG_SEGMENTS,
608 			       first_chunk, GFP_ATOMIC, scsi_sg_alloc);
609 	if (unlikely(ret))
610 		scsi_free_sgtable(sdb, mq);
611 	return ret;
612 }
613 
614 static void scsi_uninit_cmd(struct scsi_cmnd *cmd)
615 {
616 	if (cmd->request->cmd_type == REQ_TYPE_FS) {
617 		struct scsi_driver *drv = scsi_cmd_to_driver(cmd);
618 
619 		if (drv->uninit_command)
620 			drv->uninit_command(cmd);
621 	}
622 }
623 
624 static void scsi_mq_free_sgtables(struct scsi_cmnd *cmd)
625 {
626 	if (cmd->sdb.table.nents)
627 		scsi_free_sgtable(&cmd->sdb, true);
628 	if (cmd->request->next_rq && cmd->request->next_rq->special)
629 		scsi_free_sgtable(cmd->request->next_rq->special, true);
630 	if (scsi_prot_sg_count(cmd))
631 		scsi_free_sgtable(cmd->prot_sdb, true);
632 }
633 
634 static void scsi_mq_uninit_cmd(struct scsi_cmnd *cmd)
635 {
636 	struct scsi_device *sdev = cmd->device;
637 	struct Scsi_Host *shost = sdev->host;
638 	unsigned long flags;
639 
640 	scsi_mq_free_sgtables(cmd);
641 	scsi_uninit_cmd(cmd);
642 
643 	if (shost->use_cmd_list) {
644 		BUG_ON(list_empty(&cmd->list));
645 		spin_lock_irqsave(&sdev->list_lock, flags);
646 		list_del_init(&cmd->list);
647 		spin_unlock_irqrestore(&sdev->list_lock, flags);
648 	}
649 }
650 
651 /*
652  * Function:    scsi_release_buffers()
653  *
654  * Purpose:     Free resources allocate for a scsi_command.
655  *
656  * Arguments:   cmd	- command that we are bailing.
657  *
658  * Lock status: Assumed that no lock is held upon entry.
659  *
660  * Returns:     Nothing
661  *
662  * Notes:       In the event that an upper level driver rejects a
663  *		command, we must release resources allocated during
664  *		the __init_io() function.  Primarily this would involve
665  *		the scatter-gather table.
666  */
667 static void scsi_release_buffers(struct scsi_cmnd *cmd)
668 {
669 	if (cmd->sdb.table.nents)
670 		scsi_free_sgtable(&cmd->sdb, false);
671 
672 	memset(&cmd->sdb, 0, sizeof(cmd->sdb));
673 
674 	if (scsi_prot_sg_count(cmd))
675 		scsi_free_sgtable(cmd->prot_sdb, false);
676 }
677 
678 static void scsi_release_bidi_buffers(struct scsi_cmnd *cmd)
679 {
680 	struct scsi_data_buffer *bidi_sdb = cmd->request->next_rq->special;
681 
682 	scsi_free_sgtable(bidi_sdb, false);
683 	kmem_cache_free(scsi_sdb_cache, bidi_sdb);
684 	cmd->request->next_rq->special = NULL;
685 }
686 
687 static bool scsi_end_request(struct request *req, int error,
688 		unsigned int bytes, unsigned int bidi_bytes)
689 {
690 	struct scsi_cmnd *cmd = req->special;
691 	struct scsi_device *sdev = cmd->device;
692 	struct request_queue *q = sdev->request_queue;
693 
694 	if (blk_update_request(req, error, bytes))
695 		return true;
696 
697 	/* Bidi request must be completed as a whole */
698 	if (unlikely(bidi_bytes) &&
699 	    blk_update_request(req->next_rq, error, bidi_bytes))
700 		return true;
701 
702 	if (blk_queue_add_random(q))
703 		add_disk_randomness(req->rq_disk);
704 
705 	if (req->mq_ctx) {
706 		/*
707 		 * In the MQ case the command gets freed by __blk_mq_end_request,
708 		 * so we have to do all cleanup that depends on it earlier.
709 		 *
710 		 * We also can't kick the queues from irq context, so we
711 		 * will have to defer it to a workqueue.
712 		 */
713 		scsi_mq_uninit_cmd(cmd);
714 
715 		__blk_mq_end_request(req, error);
716 
717 		if (scsi_target(sdev)->single_lun ||
718 		    !list_empty(&sdev->host->starved_list))
719 			kblockd_schedule_work(&sdev->requeue_work);
720 		else
721 			blk_mq_start_stopped_hw_queues(q, true);
722 	} else {
723 		unsigned long flags;
724 
725 		if (bidi_bytes)
726 			scsi_release_bidi_buffers(cmd);
727 
728 		spin_lock_irqsave(q->queue_lock, flags);
729 		blk_finish_request(req, error);
730 		spin_unlock_irqrestore(q->queue_lock, flags);
731 
732 		scsi_release_buffers(cmd);
733 
734 		scsi_put_command(cmd);
735 		scsi_run_queue(q);
736 	}
737 
738 	put_device(&sdev->sdev_gendev);
739 	return false;
740 }
741 
742 /**
743  * __scsi_error_from_host_byte - translate SCSI error code into errno
744  * @cmd:	SCSI command (unused)
745  * @result:	scsi error code
746  *
747  * Translate SCSI error code into standard UNIX errno.
748  * Return values:
749  * -ENOLINK	temporary transport failure
750  * -EREMOTEIO	permanent target failure, do not retry
751  * -EBADE	permanent nexus failure, retry on other path
752  * -ENOSPC	No write space available
753  * -ENODATA	Medium error
754  * -EIO		unspecified I/O error
755  */
756 static int __scsi_error_from_host_byte(struct scsi_cmnd *cmd, int result)
757 {
758 	int error = 0;
759 
760 	switch(host_byte(result)) {
761 	case DID_TRANSPORT_FAILFAST:
762 		error = -ENOLINK;
763 		break;
764 	case DID_TARGET_FAILURE:
765 		set_host_byte(cmd, DID_OK);
766 		error = -EREMOTEIO;
767 		break;
768 	case DID_NEXUS_FAILURE:
769 		set_host_byte(cmd, DID_OK);
770 		error = -EBADE;
771 		break;
772 	case DID_ALLOC_FAILURE:
773 		set_host_byte(cmd, DID_OK);
774 		error = -ENOSPC;
775 		break;
776 	case DID_MEDIUM_ERROR:
777 		set_host_byte(cmd, DID_OK);
778 		error = -ENODATA;
779 		break;
780 	default:
781 		error = -EIO;
782 		break;
783 	}
784 
785 	return error;
786 }
787 
788 /*
789  * Function:    scsi_io_completion()
790  *
791  * Purpose:     Completion processing for block device I/O requests.
792  *
793  * Arguments:   cmd   - command that is finished.
794  *
795  * Lock status: Assumed that no lock is held upon entry.
796  *
797  * Returns:     Nothing
798  *
799  * Notes:       We will finish off the specified number of sectors.  If we
800  *		are done, the command block will be released and the queue
801  *		function will be goosed.  If we are not done then we have to
802  *		figure out what to do next:
803  *
804  *		a) We can call scsi_requeue_command().  The request
805  *		   will be unprepared and put back on the queue.  Then
806  *		   a new command will be created for it.  This should
807  *		   be used if we made forward progress, or if we want
808  *		   to switch from READ(10) to READ(6) for example.
809  *
810  *		b) We can call __scsi_queue_insert().  The request will
811  *		   be put back on the queue and retried using the same
812  *		   command as before, possibly after a delay.
813  *
814  *		c) We can call scsi_end_request() with -EIO to fail
815  *		   the remainder of the request.
816  */
817 void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes)
818 {
819 	int result = cmd->result;
820 	struct request_queue *q = cmd->device->request_queue;
821 	struct request *req = cmd->request;
822 	int error = 0;
823 	struct scsi_sense_hdr sshdr;
824 	bool sense_valid = false;
825 	int sense_deferred = 0, level = 0;
826 	enum {ACTION_FAIL, ACTION_REPREP, ACTION_RETRY,
827 	      ACTION_DELAYED_RETRY} action;
828 	unsigned long wait_for = (cmd->allowed + 1) * req->timeout;
829 
830 	if (result) {
831 		sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
832 		if (sense_valid)
833 			sense_deferred = scsi_sense_is_deferred(&sshdr);
834 	}
835 
836 	if (req->cmd_type == REQ_TYPE_BLOCK_PC) { /* SG_IO ioctl from block level */
837 		if (result) {
838 			if (sense_valid && req->sense) {
839 				/*
840 				 * SG_IO wants current and deferred errors
841 				 */
842 				int len = 8 + cmd->sense_buffer[7];
843 
844 				if (len > SCSI_SENSE_BUFFERSIZE)
845 					len = SCSI_SENSE_BUFFERSIZE;
846 				memcpy(req->sense, cmd->sense_buffer,  len);
847 				req->sense_len = len;
848 			}
849 			if (!sense_deferred)
850 				error = __scsi_error_from_host_byte(cmd, result);
851 		}
852 		/*
853 		 * __scsi_error_from_host_byte may have reset the host_byte
854 		 */
855 		req->errors = cmd->result;
856 
857 		req->resid_len = scsi_get_resid(cmd);
858 
859 		if (scsi_bidi_cmnd(cmd)) {
860 			/*
861 			 * Bidi commands Must be complete as a whole,
862 			 * both sides at once.
863 			 */
864 			req->next_rq->resid_len = scsi_in(cmd)->resid;
865 			if (scsi_end_request(req, 0, blk_rq_bytes(req),
866 					blk_rq_bytes(req->next_rq)))
867 				BUG();
868 			return;
869 		}
870 	} else if (blk_rq_bytes(req) == 0 && result && !sense_deferred) {
871 		/*
872 		 * Certain non BLOCK_PC requests are commands that don't
873 		 * actually transfer anything (FLUSH), so cannot use
874 		 * good_bytes != blk_rq_bytes(req) as the signal for an error.
875 		 * This sets the error explicitly for the problem case.
876 		 */
877 		error = __scsi_error_from_host_byte(cmd, result);
878 	}
879 
880 	/* no bidi support for !REQ_TYPE_BLOCK_PC yet */
881 	BUG_ON(blk_bidi_rq(req));
882 
883 	/*
884 	 * Next deal with any sectors which we were able to correctly
885 	 * handle.
886 	 */
887 	SCSI_LOG_HLCOMPLETE(1, scmd_printk(KERN_INFO, cmd,
888 		"%u sectors total, %d bytes done.\n",
889 		blk_rq_sectors(req), good_bytes));
890 
891 	/*
892 	 * Recovered errors need reporting, but they're always treated
893 	 * as success, so fiddle the result code here.  For BLOCK_PC
894 	 * we already took a copy of the original into rq->errors which
895 	 * is what gets returned to the user
896 	 */
897 	if (sense_valid && (sshdr.sense_key == RECOVERED_ERROR)) {
898 		/* if ATA PASS-THROUGH INFORMATION AVAILABLE skip
899 		 * print since caller wants ATA registers. Only occurs on
900 		 * SCSI ATA PASS_THROUGH commands when CK_COND=1
901 		 */
902 		if ((sshdr.asc == 0x0) && (sshdr.ascq == 0x1d))
903 			;
904 		else if (!(req->cmd_flags & REQ_QUIET))
905 			scsi_print_sense(cmd);
906 		result = 0;
907 		/* BLOCK_PC may have set error */
908 		error = 0;
909 	}
910 
911 	/*
912 	 * If we finished all bytes in the request we are done now.
913 	 */
914 	if (!scsi_end_request(req, error, good_bytes, 0))
915 		return;
916 
917 	/*
918 	 * Kill remainder if no retrys.
919 	 */
920 	if (error && scsi_noretry_cmd(cmd)) {
921 		if (scsi_end_request(req, error, blk_rq_bytes(req), 0))
922 			BUG();
923 		return;
924 	}
925 
926 	/*
927 	 * If there had been no error, but we have leftover bytes in the
928 	 * requeues just queue the command up again.
929 	 */
930 	if (result == 0)
931 		goto requeue;
932 
933 	error = __scsi_error_from_host_byte(cmd, result);
934 
935 	if (host_byte(result) == DID_RESET) {
936 		/* Third party bus reset or reset for error recovery
937 		 * reasons.  Just retry the command and see what
938 		 * happens.
939 		 */
940 		action = ACTION_RETRY;
941 	} else if (sense_valid && !sense_deferred) {
942 		switch (sshdr.sense_key) {
943 		case UNIT_ATTENTION:
944 			if (cmd->device->removable) {
945 				/* Detected disc change.  Set a bit
946 				 * and quietly refuse further access.
947 				 */
948 				cmd->device->changed = 1;
949 				action = ACTION_FAIL;
950 			} else {
951 				/* Must have been a power glitch, or a
952 				 * bus reset.  Could not have been a
953 				 * media change, so we just retry the
954 				 * command and see what happens.
955 				 */
956 				action = ACTION_RETRY;
957 			}
958 			break;
959 		case ILLEGAL_REQUEST:
960 			/* If we had an ILLEGAL REQUEST returned, then
961 			 * we may have performed an unsupported
962 			 * command.  The only thing this should be
963 			 * would be a ten byte read where only a six
964 			 * byte read was supported.  Also, on a system
965 			 * where READ CAPACITY failed, we may have
966 			 * read past the end of the disk.
967 			 */
968 			if ((cmd->device->use_10_for_rw &&
969 			    sshdr.asc == 0x20 && sshdr.ascq == 0x00) &&
970 			    (cmd->cmnd[0] == READ_10 ||
971 			     cmd->cmnd[0] == WRITE_10)) {
972 				/* This will issue a new 6-byte command. */
973 				cmd->device->use_10_for_rw = 0;
974 				action = ACTION_REPREP;
975 			} else if (sshdr.asc == 0x10) /* DIX */ {
976 				action = ACTION_FAIL;
977 				error = -EILSEQ;
978 			/* INVALID COMMAND OPCODE or INVALID FIELD IN CDB */
979 			} else if (sshdr.asc == 0x20 || sshdr.asc == 0x24) {
980 				action = ACTION_FAIL;
981 				error = -EREMOTEIO;
982 			} else
983 				action = ACTION_FAIL;
984 			break;
985 		case ABORTED_COMMAND:
986 			action = ACTION_FAIL;
987 			if (sshdr.asc == 0x10) /* DIF */
988 				error = -EILSEQ;
989 			break;
990 		case NOT_READY:
991 			/* If the device is in the process of becoming
992 			 * ready, or has a temporary blockage, retry.
993 			 */
994 			if (sshdr.asc == 0x04) {
995 				switch (sshdr.ascq) {
996 				case 0x01: /* becoming ready */
997 				case 0x04: /* format in progress */
998 				case 0x05: /* rebuild in progress */
999 				case 0x06: /* recalculation in progress */
1000 				case 0x07: /* operation in progress */
1001 				case 0x08: /* Long write in progress */
1002 				case 0x09: /* self test in progress */
1003 				case 0x14: /* space allocation in progress */
1004 					action = ACTION_DELAYED_RETRY;
1005 					break;
1006 				default:
1007 					action = ACTION_FAIL;
1008 					break;
1009 				}
1010 			} else
1011 				action = ACTION_FAIL;
1012 			break;
1013 		case VOLUME_OVERFLOW:
1014 			/* See SSC3rXX or current. */
1015 			action = ACTION_FAIL;
1016 			break;
1017 		default:
1018 			action = ACTION_FAIL;
1019 			break;
1020 		}
1021 	} else
1022 		action = ACTION_FAIL;
1023 
1024 	if (action != ACTION_FAIL &&
1025 	    time_before(cmd->jiffies_at_alloc + wait_for, jiffies))
1026 		action = ACTION_FAIL;
1027 
1028 	switch (action) {
1029 	case ACTION_FAIL:
1030 		/* Give up and fail the remainder of the request */
1031 		if (!(req->cmd_flags & REQ_QUIET)) {
1032 			static DEFINE_RATELIMIT_STATE(_rs,
1033 					DEFAULT_RATELIMIT_INTERVAL,
1034 					DEFAULT_RATELIMIT_BURST);
1035 
1036 			if (unlikely(scsi_logging_level))
1037 				level = SCSI_LOG_LEVEL(SCSI_LOG_MLCOMPLETE_SHIFT,
1038 						       SCSI_LOG_MLCOMPLETE_BITS);
1039 
1040 			/*
1041 			 * if logging is enabled the failure will be printed
1042 			 * in scsi_log_completion(), so avoid duplicate messages
1043 			 */
1044 			if (!level && __ratelimit(&_rs)) {
1045 				scsi_print_result(cmd, NULL, FAILED);
1046 				if (driver_byte(result) & DRIVER_SENSE)
1047 					scsi_print_sense(cmd);
1048 				scsi_print_command(cmd);
1049 			}
1050 		}
1051 		if (!scsi_end_request(req, error, blk_rq_err_bytes(req), 0))
1052 			return;
1053 		/*FALLTHRU*/
1054 	case ACTION_REPREP:
1055 	requeue:
1056 		/* Unprep the request and put it back at the head of the queue.
1057 		 * A new command will be prepared and issued.
1058 		 */
1059 		if (q->mq_ops) {
1060 			cmd->request->cmd_flags &= ~REQ_DONTPREP;
1061 			scsi_mq_uninit_cmd(cmd);
1062 			scsi_mq_requeue_cmd(cmd);
1063 		} else {
1064 			scsi_release_buffers(cmd);
1065 			scsi_requeue_command(q, cmd);
1066 		}
1067 		break;
1068 	case ACTION_RETRY:
1069 		/* Retry the same command immediately */
1070 		__scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY, 0);
1071 		break;
1072 	case ACTION_DELAYED_RETRY:
1073 		/* Retry the same command after a delay */
1074 		__scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY, 0);
1075 		break;
1076 	}
1077 }
1078 
1079 static int scsi_init_sgtable(struct request *req, struct scsi_data_buffer *sdb)
1080 {
1081 	int count;
1082 
1083 	/*
1084 	 * If sg table allocation fails, requeue request later.
1085 	 */
1086 	if (unlikely(scsi_alloc_sgtable(sdb, req->nr_phys_segments,
1087 					req->mq_ctx != NULL)))
1088 		return BLKPREP_DEFER;
1089 
1090 	/*
1091 	 * Next, walk the list, and fill in the addresses and sizes of
1092 	 * each segment.
1093 	 */
1094 	count = blk_rq_map_sg(req->q, req, sdb->table.sgl);
1095 	BUG_ON(count > sdb->table.nents);
1096 	sdb->table.nents = count;
1097 	sdb->length = blk_rq_bytes(req);
1098 	return BLKPREP_OK;
1099 }
1100 
1101 /*
1102  * Function:    scsi_init_io()
1103  *
1104  * Purpose:     SCSI I/O initialize function.
1105  *
1106  * Arguments:   cmd   - Command descriptor we wish to initialize
1107  *
1108  * Returns:     0 on success
1109  *		BLKPREP_DEFER if the failure is retryable
1110  *		BLKPREP_KILL if the failure is fatal
1111  */
1112 int scsi_init_io(struct scsi_cmnd *cmd)
1113 {
1114 	struct scsi_device *sdev = cmd->device;
1115 	struct request *rq = cmd->request;
1116 	bool is_mq = (rq->mq_ctx != NULL);
1117 	int error;
1118 
1119 	BUG_ON(!rq->nr_phys_segments);
1120 
1121 	error = scsi_init_sgtable(rq, &cmd->sdb);
1122 	if (error)
1123 		goto err_exit;
1124 
1125 	if (blk_bidi_rq(rq)) {
1126 		if (!rq->q->mq_ops) {
1127 			struct scsi_data_buffer *bidi_sdb =
1128 				kmem_cache_zalloc(scsi_sdb_cache, GFP_ATOMIC);
1129 			if (!bidi_sdb) {
1130 				error = BLKPREP_DEFER;
1131 				goto err_exit;
1132 			}
1133 
1134 			rq->next_rq->special = bidi_sdb;
1135 		}
1136 
1137 		error = scsi_init_sgtable(rq->next_rq, rq->next_rq->special);
1138 		if (error)
1139 			goto err_exit;
1140 	}
1141 
1142 	if (blk_integrity_rq(rq)) {
1143 		struct scsi_data_buffer *prot_sdb = cmd->prot_sdb;
1144 		int ivecs, count;
1145 
1146 		if (prot_sdb == NULL) {
1147 			/*
1148 			 * This can happen if someone (e.g. multipath)
1149 			 * queues a command to a device on an adapter
1150 			 * that does not support DIX.
1151 			 */
1152 			WARN_ON_ONCE(1);
1153 			error = BLKPREP_KILL;
1154 			goto err_exit;
1155 		}
1156 
1157 		ivecs = blk_rq_count_integrity_sg(rq->q, rq->bio);
1158 
1159 		if (scsi_alloc_sgtable(prot_sdb, ivecs, is_mq)) {
1160 			error = BLKPREP_DEFER;
1161 			goto err_exit;
1162 		}
1163 
1164 		count = blk_rq_map_integrity_sg(rq->q, rq->bio,
1165 						prot_sdb->table.sgl);
1166 		BUG_ON(unlikely(count > ivecs));
1167 		BUG_ON(unlikely(count > queue_max_integrity_segments(rq->q)));
1168 
1169 		cmd->prot_sdb = prot_sdb;
1170 		cmd->prot_sdb->table.nents = count;
1171 	}
1172 
1173 	return BLKPREP_OK;
1174 err_exit:
1175 	if (is_mq) {
1176 		scsi_mq_free_sgtables(cmd);
1177 	} else {
1178 		scsi_release_buffers(cmd);
1179 		cmd->request->special = NULL;
1180 		scsi_put_command(cmd);
1181 		put_device(&sdev->sdev_gendev);
1182 	}
1183 	return error;
1184 }
1185 EXPORT_SYMBOL(scsi_init_io);
1186 
1187 static struct scsi_cmnd *scsi_get_cmd_from_req(struct scsi_device *sdev,
1188 		struct request *req)
1189 {
1190 	struct scsi_cmnd *cmd;
1191 
1192 	if (!req->special) {
1193 		/* Bail if we can't get a reference to the device */
1194 		if (!get_device(&sdev->sdev_gendev))
1195 			return NULL;
1196 
1197 		cmd = scsi_get_command(sdev, GFP_ATOMIC);
1198 		if (unlikely(!cmd)) {
1199 			put_device(&sdev->sdev_gendev);
1200 			return NULL;
1201 		}
1202 		req->special = cmd;
1203 	} else {
1204 		cmd = req->special;
1205 	}
1206 
1207 	/* pull a tag out of the request if we have one */
1208 	cmd->tag = req->tag;
1209 	cmd->request = req;
1210 
1211 	cmd->cmnd = req->cmd;
1212 	cmd->prot_op = SCSI_PROT_NORMAL;
1213 
1214 	return cmd;
1215 }
1216 
1217 static int scsi_setup_blk_pc_cmnd(struct scsi_device *sdev, struct request *req)
1218 {
1219 	struct scsi_cmnd *cmd = req->special;
1220 
1221 	/*
1222 	 * BLOCK_PC requests may transfer data, in which case they must
1223 	 * a bio attached to them.  Or they might contain a SCSI command
1224 	 * that does not transfer data, in which case they may optionally
1225 	 * submit a request without an attached bio.
1226 	 */
1227 	if (req->bio) {
1228 		int ret = scsi_init_io(cmd);
1229 		if (unlikely(ret))
1230 			return ret;
1231 	} else {
1232 		BUG_ON(blk_rq_bytes(req));
1233 
1234 		memset(&cmd->sdb, 0, sizeof(cmd->sdb));
1235 	}
1236 
1237 	cmd->cmd_len = req->cmd_len;
1238 	cmd->transfersize = blk_rq_bytes(req);
1239 	cmd->allowed = req->retries;
1240 	return BLKPREP_OK;
1241 }
1242 
1243 /*
1244  * Setup a REQ_TYPE_FS command.  These are simple request from filesystems
1245  * that still need to be translated to SCSI CDBs from the ULD.
1246  */
1247 static int scsi_setup_fs_cmnd(struct scsi_device *sdev, struct request *req)
1248 {
1249 	struct scsi_cmnd *cmd = req->special;
1250 
1251 	if (unlikely(sdev->scsi_dh_data && sdev->scsi_dh_data->scsi_dh
1252 			 && sdev->scsi_dh_data->scsi_dh->prep_fn)) {
1253 		int ret = sdev->scsi_dh_data->scsi_dh->prep_fn(sdev, req);
1254 		if (ret != BLKPREP_OK)
1255 			return ret;
1256 	}
1257 
1258 	memset(cmd->cmnd, 0, BLK_MAX_CDB);
1259 	return scsi_cmd_to_driver(cmd)->init_command(cmd);
1260 }
1261 
1262 static int scsi_setup_cmnd(struct scsi_device *sdev, struct request *req)
1263 {
1264 	struct scsi_cmnd *cmd = req->special;
1265 
1266 	if (!blk_rq_bytes(req))
1267 		cmd->sc_data_direction = DMA_NONE;
1268 	else if (rq_data_dir(req) == WRITE)
1269 		cmd->sc_data_direction = DMA_TO_DEVICE;
1270 	else
1271 		cmd->sc_data_direction = DMA_FROM_DEVICE;
1272 
1273 	switch (req->cmd_type) {
1274 	case REQ_TYPE_FS:
1275 		return scsi_setup_fs_cmnd(sdev, req);
1276 	case REQ_TYPE_BLOCK_PC:
1277 		return scsi_setup_blk_pc_cmnd(sdev, req);
1278 	default:
1279 		return BLKPREP_KILL;
1280 	}
1281 }
1282 
1283 static int
1284 scsi_prep_state_check(struct scsi_device *sdev, struct request *req)
1285 {
1286 	int ret = BLKPREP_OK;
1287 
1288 	/*
1289 	 * If the device is not in running state we will reject some
1290 	 * or all commands.
1291 	 */
1292 	if (unlikely(sdev->sdev_state != SDEV_RUNNING)) {
1293 		switch (sdev->sdev_state) {
1294 		case SDEV_OFFLINE:
1295 		case SDEV_TRANSPORT_OFFLINE:
1296 			/*
1297 			 * If the device is offline we refuse to process any
1298 			 * commands.  The device must be brought online
1299 			 * before trying any recovery commands.
1300 			 */
1301 			sdev_printk(KERN_ERR, sdev,
1302 				    "rejecting I/O to offline device\n");
1303 			ret = BLKPREP_KILL;
1304 			break;
1305 		case SDEV_DEL:
1306 			/*
1307 			 * If the device is fully deleted, we refuse to
1308 			 * process any commands as well.
1309 			 */
1310 			sdev_printk(KERN_ERR, sdev,
1311 				    "rejecting I/O to dead device\n");
1312 			ret = BLKPREP_KILL;
1313 			break;
1314 		case SDEV_BLOCK:
1315 		case SDEV_CREATED_BLOCK:
1316 			ret = BLKPREP_DEFER;
1317 			break;
1318 		case SDEV_QUIESCE:
1319 			/*
1320 			 * If the devices is blocked we defer normal commands.
1321 			 */
1322 			if (!(req->cmd_flags & REQ_PREEMPT))
1323 				ret = BLKPREP_DEFER;
1324 			break;
1325 		default:
1326 			/*
1327 			 * For any other not fully online state we only allow
1328 			 * special commands.  In particular any user initiated
1329 			 * command is not allowed.
1330 			 */
1331 			if (!(req->cmd_flags & REQ_PREEMPT))
1332 				ret = BLKPREP_KILL;
1333 			break;
1334 		}
1335 	}
1336 	return ret;
1337 }
1338 
1339 static int
1340 scsi_prep_return(struct request_queue *q, struct request *req, int ret)
1341 {
1342 	struct scsi_device *sdev = q->queuedata;
1343 
1344 	switch (ret) {
1345 	case BLKPREP_KILL:
1346 		req->errors = DID_NO_CONNECT << 16;
1347 		/* release the command and kill it */
1348 		if (req->special) {
1349 			struct scsi_cmnd *cmd = req->special;
1350 			scsi_release_buffers(cmd);
1351 			scsi_put_command(cmd);
1352 			put_device(&sdev->sdev_gendev);
1353 			req->special = NULL;
1354 		}
1355 		break;
1356 	case BLKPREP_DEFER:
1357 		/*
1358 		 * If we defer, the blk_peek_request() returns NULL, but the
1359 		 * queue must be restarted, so we schedule a callback to happen
1360 		 * shortly.
1361 		 */
1362 		if (atomic_read(&sdev->device_busy) == 0)
1363 			blk_delay_queue(q, SCSI_QUEUE_DELAY);
1364 		break;
1365 	default:
1366 		req->cmd_flags |= REQ_DONTPREP;
1367 	}
1368 
1369 	return ret;
1370 }
1371 
1372 static int scsi_prep_fn(struct request_queue *q, struct request *req)
1373 {
1374 	struct scsi_device *sdev = q->queuedata;
1375 	struct scsi_cmnd *cmd;
1376 	int ret;
1377 
1378 	ret = scsi_prep_state_check(sdev, req);
1379 	if (ret != BLKPREP_OK)
1380 		goto out;
1381 
1382 	cmd = scsi_get_cmd_from_req(sdev, req);
1383 	if (unlikely(!cmd)) {
1384 		ret = BLKPREP_DEFER;
1385 		goto out;
1386 	}
1387 
1388 	ret = scsi_setup_cmnd(sdev, req);
1389 out:
1390 	return scsi_prep_return(q, req, ret);
1391 }
1392 
1393 static void scsi_unprep_fn(struct request_queue *q, struct request *req)
1394 {
1395 	scsi_uninit_cmd(req->special);
1396 }
1397 
1398 /*
1399  * scsi_dev_queue_ready: if we can send requests to sdev, return 1 else
1400  * return 0.
1401  *
1402  * Called with the queue_lock held.
1403  */
1404 static inline int scsi_dev_queue_ready(struct request_queue *q,
1405 				  struct scsi_device *sdev)
1406 {
1407 	unsigned int busy;
1408 
1409 	busy = atomic_inc_return(&sdev->device_busy) - 1;
1410 	if (atomic_read(&sdev->device_blocked)) {
1411 		if (busy)
1412 			goto out_dec;
1413 
1414 		/*
1415 		 * unblock after device_blocked iterates to zero
1416 		 */
1417 		if (atomic_dec_return(&sdev->device_blocked) > 0) {
1418 			/*
1419 			 * For the MQ case we take care of this in the caller.
1420 			 */
1421 			if (!q->mq_ops)
1422 				blk_delay_queue(q, SCSI_QUEUE_DELAY);
1423 			goto out_dec;
1424 		}
1425 		SCSI_LOG_MLQUEUE(3, sdev_printk(KERN_INFO, sdev,
1426 				   "unblocking device at zero depth\n"));
1427 	}
1428 
1429 	if (busy >= sdev->queue_depth)
1430 		goto out_dec;
1431 
1432 	return 1;
1433 out_dec:
1434 	atomic_dec(&sdev->device_busy);
1435 	return 0;
1436 }
1437 
1438 /*
1439  * scsi_target_queue_ready: checks if there we can send commands to target
1440  * @sdev: scsi device on starget to check.
1441  */
1442 static inline int scsi_target_queue_ready(struct Scsi_Host *shost,
1443 					   struct scsi_device *sdev)
1444 {
1445 	struct scsi_target *starget = scsi_target(sdev);
1446 	unsigned int busy;
1447 
1448 	if (starget->single_lun) {
1449 		spin_lock_irq(shost->host_lock);
1450 		if (starget->starget_sdev_user &&
1451 		    starget->starget_sdev_user != sdev) {
1452 			spin_unlock_irq(shost->host_lock);
1453 			return 0;
1454 		}
1455 		starget->starget_sdev_user = sdev;
1456 		spin_unlock_irq(shost->host_lock);
1457 	}
1458 
1459 	if (starget->can_queue <= 0)
1460 		return 1;
1461 
1462 	busy = atomic_inc_return(&starget->target_busy) - 1;
1463 	if (atomic_read(&starget->target_blocked) > 0) {
1464 		if (busy)
1465 			goto starved;
1466 
1467 		/*
1468 		 * unblock after target_blocked iterates to zero
1469 		 */
1470 		if (atomic_dec_return(&starget->target_blocked) > 0)
1471 			goto out_dec;
1472 
1473 		SCSI_LOG_MLQUEUE(3, starget_printk(KERN_INFO, starget,
1474 				 "unblocking target at zero depth\n"));
1475 	}
1476 
1477 	if (busy >= starget->can_queue)
1478 		goto starved;
1479 
1480 	return 1;
1481 
1482 starved:
1483 	spin_lock_irq(shost->host_lock);
1484 	list_move_tail(&sdev->starved_entry, &shost->starved_list);
1485 	spin_unlock_irq(shost->host_lock);
1486 out_dec:
1487 	if (starget->can_queue > 0)
1488 		atomic_dec(&starget->target_busy);
1489 	return 0;
1490 }
1491 
1492 /*
1493  * scsi_host_queue_ready: if we can send requests to shost, return 1 else
1494  * return 0. We must end up running the queue again whenever 0 is
1495  * returned, else IO can hang.
1496  */
1497 static inline int scsi_host_queue_ready(struct request_queue *q,
1498 				   struct Scsi_Host *shost,
1499 				   struct scsi_device *sdev)
1500 {
1501 	unsigned int busy;
1502 
1503 	if (scsi_host_in_recovery(shost))
1504 		return 0;
1505 
1506 	busy = atomic_inc_return(&shost->host_busy) - 1;
1507 	if (atomic_read(&shost->host_blocked) > 0) {
1508 		if (busy)
1509 			goto starved;
1510 
1511 		/*
1512 		 * unblock after host_blocked iterates to zero
1513 		 */
1514 		if (atomic_dec_return(&shost->host_blocked) > 0)
1515 			goto out_dec;
1516 
1517 		SCSI_LOG_MLQUEUE(3,
1518 			shost_printk(KERN_INFO, shost,
1519 				     "unblocking host at zero depth\n"));
1520 	}
1521 
1522 	if (shost->can_queue > 0 && busy >= shost->can_queue)
1523 		goto starved;
1524 	if (shost->host_self_blocked)
1525 		goto starved;
1526 
1527 	/* We're OK to process the command, so we can't be starved */
1528 	if (!list_empty(&sdev->starved_entry)) {
1529 		spin_lock_irq(shost->host_lock);
1530 		if (!list_empty(&sdev->starved_entry))
1531 			list_del_init(&sdev->starved_entry);
1532 		spin_unlock_irq(shost->host_lock);
1533 	}
1534 
1535 	return 1;
1536 
1537 starved:
1538 	spin_lock_irq(shost->host_lock);
1539 	if (list_empty(&sdev->starved_entry))
1540 		list_add_tail(&sdev->starved_entry, &shost->starved_list);
1541 	spin_unlock_irq(shost->host_lock);
1542 out_dec:
1543 	atomic_dec(&shost->host_busy);
1544 	return 0;
1545 }
1546 
1547 /*
1548  * Busy state exporting function for request stacking drivers.
1549  *
1550  * For efficiency, no lock is taken to check the busy state of
1551  * shost/starget/sdev, since the returned value is not guaranteed and
1552  * may be changed after request stacking drivers call the function,
1553  * regardless of taking lock or not.
1554  *
1555  * When scsi can't dispatch I/Os anymore and needs to kill I/Os scsi
1556  * needs to return 'not busy'. Otherwise, request stacking drivers
1557  * may hold requests forever.
1558  */
1559 static int scsi_lld_busy(struct request_queue *q)
1560 {
1561 	struct scsi_device *sdev = q->queuedata;
1562 	struct Scsi_Host *shost;
1563 
1564 	if (blk_queue_dying(q))
1565 		return 0;
1566 
1567 	shost = sdev->host;
1568 
1569 	/*
1570 	 * Ignore host/starget busy state.
1571 	 * Since block layer does not have a concept of fairness across
1572 	 * multiple queues, congestion of host/starget needs to be handled
1573 	 * in SCSI layer.
1574 	 */
1575 	if (scsi_host_in_recovery(shost) || scsi_device_is_busy(sdev))
1576 		return 1;
1577 
1578 	return 0;
1579 }
1580 
1581 /*
1582  * Kill a request for a dead device
1583  */
1584 static void scsi_kill_request(struct request *req, struct request_queue *q)
1585 {
1586 	struct scsi_cmnd *cmd = req->special;
1587 	struct scsi_device *sdev;
1588 	struct scsi_target *starget;
1589 	struct Scsi_Host *shost;
1590 
1591 	blk_start_request(req);
1592 
1593 	scmd_printk(KERN_INFO, cmd, "killing request\n");
1594 
1595 	sdev = cmd->device;
1596 	starget = scsi_target(sdev);
1597 	shost = sdev->host;
1598 	scsi_init_cmd_errh(cmd);
1599 	cmd->result = DID_NO_CONNECT << 16;
1600 	atomic_inc(&cmd->device->iorequest_cnt);
1601 
1602 	/*
1603 	 * SCSI request completion path will do scsi_device_unbusy(),
1604 	 * bump busy counts.  To bump the counters, we need to dance
1605 	 * with the locks as normal issue path does.
1606 	 */
1607 	atomic_inc(&sdev->device_busy);
1608 	atomic_inc(&shost->host_busy);
1609 	if (starget->can_queue > 0)
1610 		atomic_inc(&starget->target_busy);
1611 
1612 	blk_complete_request(req);
1613 }
1614 
1615 static void scsi_softirq_done(struct request *rq)
1616 {
1617 	struct scsi_cmnd *cmd = rq->special;
1618 	unsigned long wait_for = (cmd->allowed + 1) * rq->timeout;
1619 	int disposition;
1620 
1621 	INIT_LIST_HEAD(&cmd->eh_entry);
1622 
1623 	atomic_inc(&cmd->device->iodone_cnt);
1624 	if (cmd->result)
1625 		atomic_inc(&cmd->device->ioerr_cnt);
1626 
1627 	disposition = scsi_decide_disposition(cmd);
1628 	if (disposition != SUCCESS &&
1629 	    time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) {
1630 		sdev_printk(KERN_ERR, cmd->device,
1631 			    "timing out command, waited %lus\n",
1632 			    wait_for/HZ);
1633 		disposition = SUCCESS;
1634 	}
1635 
1636 	scsi_log_completion(cmd, disposition);
1637 
1638 	switch (disposition) {
1639 		case SUCCESS:
1640 			scsi_finish_command(cmd);
1641 			break;
1642 		case NEEDS_RETRY:
1643 			scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY);
1644 			break;
1645 		case ADD_TO_MLQUEUE:
1646 			scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY);
1647 			break;
1648 		default:
1649 			if (!scsi_eh_scmd_add(cmd, 0))
1650 				scsi_finish_command(cmd);
1651 	}
1652 }
1653 
1654 /**
1655  * scsi_dispatch_command - Dispatch a command to the low-level driver.
1656  * @cmd: command block we are dispatching.
1657  *
1658  * Return: nonzero return request was rejected and device's queue needs to be
1659  * plugged.
1660  */
1661 static int scsi_dispatch_cmd(struct scsi_cmnd *cmd)
1662 {
1663 	struct Scsi_Host *host = cmd->device->host;
1664 	int rtn = 0;
1665 
1666 	atomic_inc(&cmd->device->iorequest_cnt);
1667 
1668 	/* check if the device is still usable */
1669 	if (unlikely(cmd->device->sdev_state == SDEV_DEL)) {
1670 		/* in SDEV_DEL we error all commands. DID_NO_CONNECT
1671 		 * returns an immediate error upwards, and signals
1672 		 * that the device is no longer present */
1673 		cmd->result = DID_NO_CONNECT << 16;
1674 		goto done;
1675 	}
1676 
1677 	/* Check to see if the scsi lld made this device blocked. */
1678 	if (unlikely(scsi_device_blocked(cmd->device))) {
1679 		/*
1680 		 * in blocked state, the command is just put back on
1681 		 * the device queue.  The suspend state has already
1682 		 * blocked the queue so future requests should not
1683 		 * occur until the device transitions out of the
1684 		 * suspend state.
1685 		 */
1686 		SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1687 			"queuecommand : device blocked\n"));
1688 		return SCSI_MLQUEUE_DEVICE_BUSY;
1689 	}
1690 
1691 	/* Store the LUN value in cmnd, if needed. */
1692 	if (cmd->device->lun_in_cdb)
1693 		cmd->cmnd[1] = (cmd->cmnd[1] & 0x1f) |
1694 			       (cmd->device->lun << 5 & 0xe0);
1695 
1696 	scsi_log_send(cmd);
1697 
1698 	/*
1699 	 * Before we queue this command, check if the command
1700 	 * length exceeds what the host adapter can handle.
1701 	 */
1702 	if (cmd->cmd_len > cmd->device->host->max_cmd_len) {
1703 		SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1704 			       "queuecommand : command too long. "
1705 			       "cdb_size=%d host->max_cmd_len=%d\n",
1706 			       cmd->cmd_len, cmd->device->host->max_cmd_len));
1707 		cmd->result = (DID_ABORT << 16);
1708 		goto done;
1709 	}
1710 
1711 	if (unlikely(host->shost_state == SHOST_DEL)) {
1712 		cmd->result = (DID_NO_CONNECT << 16);
1713 		goto done;
1714 
1715 	}
1716 
1717 	trace_scsi_dispatch_cmd_start(cmd);
1718 	rtn = host->hostt->queuecommand(host, cmd);
1719 	if (rtn) {
1720 		trace_scsi_dispatch_cmd_error(cmd, rtn);
1721 		if (rtn != SCSI_MLQUEUE_DEVICE_BUSY &&
1722 		    rtn != SCSI_MLQUEUE_TARGET_BUSY)
1723 			rtn = SCSI_MLQUEUE_HOST_BUSY;
1724 
1725 		SCSI_LOG_MLQUEUE(3, scmd_printk(KERN_INFO, cmd,
1726 			"queuecommand : request rejected\n"));
1727 	}
1728 
1729 	return rtn;
1730  done:
1731 	cmd->scsi_done(cmd);
1732 	return 0;
1733 }
1734 
1735 /**
1736  * scsi_done - Invoke completion on finished SCSI command.
1737  * @cmd: The SCSI Command for which a low-level device driver (LLDD) gives
1738  * ownership back to SCSI Core -- i.e. the LLDD has finished with it.
1739  *
1740  * Description: This function is the mid-level's (SCSI Core) interrupt routine,
1741  * which regains ownership of the SCSI command (de facto) from a LLDD, and
1742  * calls blk_complete_request() for further processing.
1743  *
1744  * This function is interrupt context safe.
1745  */
1746 static void scsi_done(struct scsi_cmnd *cmd)
1747 {
1748 	trace_scsi_dispatch_cmd_done(cmd);
1749 	blk_complete_request(cmd->request);
1750 }
1751 
1752 /*
1753  * Function:    scsi_request_fn()
1754  *
1755  * Purpose:     Main strategy routine for SCSI.
1756  *
1757  * Arguments:   q       - Pointer to actual queue.
1758  *
1759  * Returns:     Nothing
1760  *
1761  * Lock status: IO request lock assumed to be held when called.
1762  */
1763 static void scsi_request_fn(struct request_queue *q)
1764 	__releases(q->queue_lock)
1765 	__acquires(q->queue_lock)
1766 {
1767 	struct scsi_device *sdev = q->queuedata;
1768 	struct Scsi_Host *shost;
1769 	struct scsi_cmnd *cmd;
1770 	struct request *req;
1771 
1772 	/*
1773 	 * To start with, we keep looping until the queue is empty, or until
1774 	 * the host is no longer able to accept any more requests.
1775 	 */
1776 	shost = sdev->host;
1777 	for (;;) {
1778 		int rtn;
1779 		/*
1780 		 * get next queueable request.  We do this early to make sure
1781 		 * that the request is fully prepared even if we cannot
1782 		 * accept it.
1783 		 */
1784 		req = blk_peek_request(q);
1785 		if (!req)
1786 			break;
1787 
1788 		if (unlikely(!scsi_device_online(sdev))) {
1789 			sdev_printk(KERN_ERR, sdev,
1790 				    "rejecting I/O to offline device\n");
1791 			scsi_kill_request(req, q);
1792 			continue;
1793 		}
1794 
1795 		if (!scsi_dev_queue_ready(q, sdev))
1796 			break;
1797 
1798 		/*
1799 		 * Remove the request from the request list.
1800 		 */
1801 		if (!(blk_queue_tagged(q) && !blk_queue_start_tag(q, req)))
1802 			blk_start_request(req);
1803 
1804 		spin_unlock_irq(q->queue_lock);
1805 		cmd = req->special;
1806 		if (unlikely(cmd == NULL)) {
1807 			printk(KERN_CRIT "impossible request in %s.\n"
1808 					 "please mail a stack trace to "
1809 					 "linux-scsi@vger.kernel.org\n",
1810 					 __func__);
1811 			blk_dump_rq_flags(req, "foo");
1812 			BUG();
1813 		}
1814 
1815 		/*
1816 		 * We hit this when the driver is using a host wide
1817 		 * tag map. For device level tag maps the queue_depth check
1818 		 * in the device ready fn would prevent us from trying
1819 		 * to allocate a tag. Since the map is a shared host resource
1820 		 * we add the dev to the starved list so it eventually gets
1821 		 * a run when a tag is freed.
1822 		 */
1823 		if (blk_queue_tagged(q) && !(req->cmd_flags & REQ_QUEUED)) {
1824 			spin_lock_irq(shost->host_lock);
1825 			if (list_empty(&sdev->starved_entry))
1826 				list_add_tail(&sdev->starved_entry,
1827 					      &shost->starved_list);
1828 			spin_unlock_irq(shost->host_lock);
1829 			goto not_ready;
1830 		}
1831 
1832 		if (!scsi_target_queue_ready(shost, sdev))
1833 			goto not_ready;
1834 
1835 		if (!scsi_host_queue_ready(q, shost, sdev))
1836 			goto host_not_ready;
1837 
1838 		if (sdev->simple_tags)
1839 			cmd->flags |= SCMD_TAGGED;
1840 		else
1841 			cmd->flags &= ~SCMD_TAGGED;
1842 
1843 		/*
1844 		 * Finally, initialize any error handling parameters, and set up
1845 		 * the timers for timeouts.
1846 		 */
1847 		scsi_init_cmd_errh(cmd);
1848 
1849 		/*
1850 		 * Dispatch the command to the low-level driver.
1851 		 */
1852 		cmd->scsi_done = scsi_done;
1853 		rtn = scsi_dispatch_cmd(cmd);
1854 		if (rtn) {
1855 			scsi_queue_insert(cmd, rtn);
1856 			spin_lock_irq(q->queue_lock);
1857 			goto out_delay;
1858 		}
1859 		spin_lock_irq(q->queue_lock);
1860 	}
1861 
1862 	return;
1863 
1864  host_not_ready:
1865 	if (scsi_target(sdev)->can_queue > 0)
1866 		atomic_dec(&scsi_target(sdev)->target_busy);
1867  not_ready:
1868 	/*
1869 	 * lock q, handle tag, requeue req, and decrement device_busy. We
1870 	 * must return with queue_lock held.
1871 	 *
1872 	 * Decrementing device_busy without checking it is OK, as all such
1873 	 * cases (host limits or settings) should run the queue at some
1874 	 * later time.
1875 	 */
1876 	spin_lock_irq(q->queue_lock);
1877 	blk_requeue_request(q, req);
1878 	atomic_dec(&sdev->device_busy);
1879 out_delay:
1880 	if (!atomic_read(&sdev->device_busy) && !scsi_device_blocked(sdev))
1881 		blk_delay_queue(q, SCSI_QUEUE_DELAY);
1882 }
1883 
1884 static inline int prep_to_mq(int ret)
1885 {
1886 	switch (ret) {
1887 	case BLKPREP_OK:
1888 		return 0;
1889 	case BLKPREP_DEFER:
1890 		return BLK_MQ_RQ_QUEUE_BUSY;
1891 	default:
1892 		return BLK_MQ_RQ_QUEUE_ERROR;
1893 	}
1894 }
1895 
1896 static int scsi_mq_prep_fn(struct request *req)
1897 {
1898 	struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1899 	struct scsi_device *sdev = req->q->queuedata;
1900 	struct Scsi_Host *shost = sdev->host;
1901 	unsigned char *sense_buf = cmd->sense_buffer;
1902 	struct scatterlist *sg;
1903 
1904 	memset(cmd, 0, sizeof(struct scsi_cmnd));
1905 
1906 	req->special = cmd;
1907 
1908 	cmd->request = req;
1909 	cmd->device = sdev;
1910 	cmd->sense_buffer = sense_buf;
1911 
1912 	cmd->tag = req->tag;
1913 
1914 	cmd->cmnd = req->cmd;
1915 	cmd->prot_op = SCSI_PROT_NORMAL;
1916 
1917 	INIT_LIST_HEAD(&cmd->list);
1918 	INIT_DELAYED_WORK(&cmd->abort_work, scmd_eh_abort_handler);
1919 	cmd->jiffies_at_alloc = jiffies;
1920 
1921 	if (shost->use_cmd_list) {
1922 		spin_lock_irq(&sdev->list_lock);
1923 		list_add_tail(&cmd->list, &sdev->cmd_list);
1924 		spin_unlock_irq(&sdev->list_lock);
1925 	}
1926 
1927 	sg = (void *)cmd + sizeof(struct scsi_cmnd) + shost->hostt->cmd_size;
1928 	cmd->sdb.table.sgl = sg;
1929 
1930 	if (scsi_host_get_prot(shost)) {
1931 		cmd->prot_sdb = (void *)sg +
1932 			min_t(unsigned int,
1933 			      shost->sg_tablesize, SCSI_MAX_SG_SEGMENTS) *
1934 			sizeof(struct scatterlist);
1935 		memset(cmd->prot_sdb, 0, sizeof(struct scsi_data_buffer));
1936 
1937 		cmd->prot_sdb->table.sgl =
1938 			(struct scatterlist *)(cmd->prot_sdb + 1);
1939 	}
1940 
1941 	if (blk_bidi_rq(req)) {
1942 		struct request *next_rq = req->next_rq;
1943 		struct scsi_data_buffer *bidi_sdb = blk_mq_rq_to_pdu(next_rq);
1944 
1945 		memset(bidi_sdb, 0, sizeof(struct scsi_data_buffer));
1946 		bidi_sdb->table.sgl =
1947 			(struct scatterlist *)(bidi_sdb + 1);
1948 
1949 		next_rq->special = bidi_sdb;
1950 	}
1951 
1952 	blk_mq_start_request(req);
1953 
1954 	return scsi_setup_cmnd(sdev, req);
1955 }
1956 
1957 static void scsi_mq_done(struct scsi_cmnd *cmd)
1958 {
1959 	trace_scsi_dispatch_cmd_done(cmd);
1960 	blk_mq_complete_request(cmd->request);
1961 }
1962 
1963 static int scsi_queue_rq(struct blk_mq_hw_ctx *hctx,
1964 			 const struct blk_mq_queue_data *bd)
1965 {
1966 	struct request *req = bd->rq;
1967 	struct request_queue *q = req->q;
1968 	struct scsi_device *sdev = q->queuedata;
1969 	struct Scsi_Host *shost = sdev->host;
1970 	struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(req);
1971 	int ret;
1972 	int reason;
1973 
1974 	ret = prep_to_mq(scsi_prep_state_check(sdev, req));
1975 	if (ret)
1976 		goto out;
1977 
1978 	ret = BLK_MQ_RQ_QUEUE_BUSY;
1979 	if (!get_device(&sdev->sdev_gendev))
1980 		goto out;
1981 
1982 	if (!scsi_dev_queue_ready(q, sdev))
1983 		goto out_put_device;
1984 	if (!scsi_target_queue_ready(shost, sdev))
1985 		goto out_dec_device_busy;
1986 	if (!scsi_host_queue_ready(q, shost, sdev))
1987 		goto out_dec_target_busy;
1988 
1989 
1990 	if (!(req->cmd_flags & REQ_DONTPREP)) {
1991 		ret = prep_to_mq(scsi_mq_prep_fn(req));
1992 		if (ret)
1993 			goto out_dec_host_busy;
1994 		req->cmd_flags |= REQ_DONTPREP;
1995 	} else {
1996 		blk_mq_start_request(req);
1997 	}
1998 
1999 	if (sdev->simple_tags)
2000 		cmd->flags |= SCMD_TAGGED;
2001 	else
2002 		cmd->flags &= ~SCMD_TAGGED;
2003 
2004 	scsi_init_cmd_errh(cmd);
2005 	cmd->scsi_done = scsi_mq_done;
2006 
2007 	reason = scsi_dispatch_cmd(cmd);
2008 	if (reason) {
2009 		scsi_set_blocked(cmd, reason);
2010 		ret = BLK_MQ_RQ_QUEUE_BUSY;
2011 		goto out_dec_host_busy;
2012 	}
2013 
2014 	return BLK_MQ_RQ_QUEUE_OK;
2015 
2016 out_dec_host_busy:
2017 	atomic_dec(&shost->host_busy);
2018 out_dec_target_busy:
2019 	if (scsi_target(sdev)->can_queue > 0)
2020 		atomic_dec(&scsi_target(sdev)->target_busy);
2021 out_dec_device_busy:
2022 	atomic_dec(&sdev->device_busy);
2023 out_put_device:
2024 	put_device(&sdev->sdev_gendev);
2025 out:
2026 	switch (ret) {
2027 	case BLK_MQ_RQ_QUEUE_BUSY:
2028 		blk_mq_stop_hw_queue(hctx);
2029 		if (atomic_read(&sdev->device_busy) == 0 &&
2030 		    !scsi_device_blocked(sdev))
2031 			blk_mq_delay_queue(hctx, SCSI_QUEUE_DELAY);
2032 		break;
2033 	case BLK_MQ_RQ_QUEUE_ERROR:
2034 		/*
2035 		 * Make sure to release all allocated ressources when
2036 		 * we hit an error, as we will never see this command
2037 		 * again.
2038 		 */
2039 		if (req->cmd_flags & REQ_DONTPREP)
2040 			scsi_mq_uninit_cmd(cmd);
2041 		break;
2042 	default:
2043 		break;
2044 	}
2045 	return ret;
2046 }
2047 
2048 static enum blk_eh_timer_return scsi_timeout(struct request *req,
2049 		bool reserved)
2050 {
2051 	if (reserved)
2052 		return BLK_EH_RESET_TIMER;
2053 	return scsi_times_out(req);
2054 }
2055 
2056 static int scsi_init_request(void *data, struct request *rq,
2057 		unsigned int hctx_idx, unsigned int request_idx,
2058 		unsigned int numa_node)
2059 {
2060 	struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
2061 
2062 	cmd->sense_buffer = kzalloc_node(SCSI_SENSE_BUFFERSIZE, GFP_KERNEL,
2063 			numa_node);
2064 	if (!cmd->sense_buffer)
2065 		return -ENOMEM;
2066 	return 0;
2067 }
2068 
2069 static void scsi_exit_request(void *data, struct request *rq,
2070 		unsigned int hctx_idx, unsigned int request_idx)
2071 {
2072 	struct scsi_cmnd *cmd = blk_mq_rq_to_pdu(rq);
2073 
2074 	kfree(cmd->sense_buffer);
2075 }
2076 
2077 static u64 scsi_calculate_bounce_limit(struct Scsi_Host *shost)
2078 {
2079 	struct device *host_dev;
2080 	u64 bounce_limit = 0xffffffff;
2081 
2082 	if (shost->unchecked_isa_dma)
2083 		return BLK_BOUNCE_ISA;
2084 	/*
2085 	 * Platforms with virtual-DMA translation
2086 	 * hardware have no practical limit.
2087 	 */
2088 	if (!PCI_DMA_BUS_IS_PHYS)
2089 		return BLK_BOUNCE_ANY;
2090 
2091 	host_dev = scsi_get_device(shost);
2092 	if (host_dev && host_dev->dma_mask)
2093 		bounce_limit = (u64)dma_max_pfn(host_dev) << PAGE_SHIFT;
2094 
2095 	return bounce_limit;
2096 }
2097 
2098 static void __scsi_init_queue(struct Scsi_Host *shost, struct request_queue *q)
2099 {
2100 	struct device *dev = shost->dma_dev;
2101 
2102 	/*
2103 	 * this limit is imposed by hardware restrictions
2104 	 */
2105 	blk_queue_max_segments(q, min_t(unsigned short, shost->sg_tablesize,
2106 					SCSI_MAX_SG_CHAIN_SEGMENTS));
2107 
2108 	if (scsi_host_prot_dma(shost)) {
2109 		shost->sg_prot_tablesize =
2110 			min_not_zero(shost->sg_prot_tablesize,
2111 				     (unsigned short)SCSI_MAX_PROT_SG_SEGMENTS);
2112 		BUG_ON(shost->sg_prot_tablesize < shost->sg_tablesize);
2113 		blk_queue_max_integrity_segments(q, shost->sg_prot_tablesize);
2114 	}
2115 
2116 	blk_queue_max_hw_sectors(q, shost->max_sectors);
2117 	blk_queue_bounce_limit(q, scsi_calculate_bounce_limit(shost));
2118 	blk_queue_segment_boundary(q, shost->dma_boundary);
2119 	dma_set_seg_boundary(dev, shost->dma_boundary);
2120 
2121 	blk_queue_max_segment_size(q, dma_get_max_seg_size(dev));
2122 
2123 	if (!shost->use_clustering)
2124 		q->limits.cluster = 0;
2125 
2126 	/*
2127 	 * set a reasonable default alignment on word boundaries: the
2128 	 * host and device may alter it using
2129 	 * blk_queue_update_dma_alignment() later.
2130 	 */
2131 	blk_queue_dma_alignment(q, 0x03);
2132 }
2133 
2134 struct request_queue *__scsi_alloc_queue(struct Scsi_Host *shost,
2135 					 request_fn_proc *request_fn)
2136 {
2137 	struct request_queue *q;
2138 
2139 	q = blk_init_queue(request_fn, NULL);
2140 	if (!q)
2141 		return NULL;
2142 	__scsi_init_queue(shost, q);
2143 	return q;
2144 }
2145 EXPORT_SYMBOL(__scsi_alloc_queue);
2146 
2147 struct request_queue *scsi_alloc_queue(struct scsi_device *sdev)
2148 {
2149 	struct request_queue *q;
2150 
2151 	q = __scsi_alloc_queue(sdev->host, scsi_request_fn);
2152 	if (!q)
2153 		return NULL;
2154 
2155 	blk_queue_prep_rq(q, scsi_prep_fn);
2156 	blk_queue_unprep_rq(q, scsi_unprep_fn);
2157 	blk_queue_softirq_done(q, scsi_softirq_done);
2158 	blk_queue_rq_timed_out(q, scsi_times_out);
2159 	blk_queue_lld_busy(q, scsi_lld_busy);
2160 	return q;
2161 }
2162 
2163 static struct blk_mq_ops scsi_mq_ops = {
2164 	.map_queue	= blk_mq_map_queue,
2165 	.queue_rq	= scsi_queue_rq,
2166 	.complete	= scsi_softirq_done,
2167 	.timeout	= scsi_timeout,
2168 	.init_request	= scsi_init_request,
2169 	.exit_request	= scsi_exit_request,
2170 };
2171 
2172 struct request_queue *scsi_mq_alloc_queue(struct scsi_device *sdev)
2173 {
2174 	sdev->request_queue = blk_mq_init_queue(&sdev->host->tag_set);
2175 	if (IS_ERR(sdev->request_queue))
2176 		return NULL;
2177 
2178 	sdev->request_queue->queuedata = sdev;
2179 	__scsi_init_queue(sdev->host, sdev->request_queue);
2180 	return sdev->request_queue;
2181 }
2182 
2183 int scsi_mq_setup_tags(struct Scsi_Host *shost)
2184 {
2185 	unsigned int cmd_size, sgl_size, tbl_size;
2186 
2187 	tbl_size = shost->sg_tablesize;
2188 	if (tbl_size > SCSI_MAX_SG_SEGMENTS)
2189 		tbl_size = SCSI_MAX_SG_SEGMENTS;
2190 	sgl_size = tbl_size * sizeof(struct scatterlist);
2191 	cmd_size = sizeof(struct scsi_cmnd) + shost->hostt->cmd_size + sgl_size;
2192 	if (scsi_host_get_prot(shost))
2193 		cmd_size += sizeof(struct scsi_data_buffer) + sgl_size;
2194 
2195 	memset(&shost->tag_set, 0, sizeof(shost->tag_set));
2196 	shost->tag_set.ops = &scsi_mq_ops;
2197 	shost->tag_set.nr_hw_queues = shost->nr_hw_queues ? : 1;
2198 	shost->tag_set.queue_depth = shost->can_queue;
2199 	shost->tag_set.cmd_size = cmd_size;
2200 	shost->tag_set.numa_node = NUMA_NO_NODE;
2201 	shost->tag_set.flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_SG_MERGE;
2202 	shost->tag_set.flags |=
2203 		BLK_ALLOC_POLICY_TO_MQ_FLAG(shost->hostt->tag_alloc_policy);
2204 	shost->tag_set.driver_data = shost;
2205 
2206 	return blk_mq_alloc_tag_set(&shost->tag_set);
2207 }
2208 
2209 void scsi_mq_destroy_tags(struct Scsi_Host *shost)
2210 {
2211 	blk_mq_free_tag_set(&shost->tag_set);
2212 }
2213 
2214 /*
2215  * Function:    scsi_block_requests()
2216  *
2217  * Purpose:     Utility function used by low-level drivers to prevent further
2218  *		commands from being queued to the device.
2219  *
2220  * Arguments:   shost       - Host in question
2221  *
2222  * Returns:     Nothing
2223  *
2224  * Lock status: No locks are assumed held.
2225  *
2226  * Notes:       There is no timer nor any other means by which the requests
2227  *		get unblocked other than the low-level driver calling
2228  *		scsi_unblock_requests().
2229  */
2230 void scsi_block_requests(struct Scsi_Host *shost)
2231 {
2232 	shost->host_self_blocked = 1;
2233 }
2234 EXPORT_SYMBOL(scsi_block_requests);
2235 
2236 /*
2237  * Function:    scsi_unblock_requests()
2238  *
2239  * Purpose:     Utility function used by low-level drivers to allow further
2240  *		commands from being queued to the device.
2241  *
2242  * Arguments:   shost       - Host in question
2243  *
2244  * Returns:     Nothing
2245  *
2246  * Lock status: No locks are assumed held.
2247  *
2248  * Notes:       There is no timer nor any other means by which the requests
2249  *		get unblocked other than the low-level driver calling
2250  *		scsi_unblock_requests().
2251  *
2252  *		This is done as an API function so that changes to the
2253  *		internals of the scsi mid-layer won't require wholesale
2254  *		changes to drivers that use this feature.
2255  */
2256 void scsi_unblock_requests(struct Scsi_Host *shost)
2257 {
2258 	shost->host_self_blocked = 0;
2259 	scsi_run_host_queues(shost);
2260 }
2261 EXPORT_SYMBOL(scsi_unblock_requests);
2262 
2263 int __init scsi_init_queue(void)
2264 {
2265 	int i;
2266 
2267 	scsi_sdb_cache = kmem_cache_create("scsi_data_buffer",
2268 					   sizeof(struct scsi_data_buffer),
2269 					   0, 0, NULL);
2270 	if (!scsi_sdb_cache) {
2271 		printk(KERN_ERR "SCSI: can't init scsi sdb cache\n");
2272 		return -ENOMEM;
2273 	}
2274 
2275 	for (i = 0; i < SG_MEMPOOL_NR; i++) {
2276 		struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
2277 		int size = sgp->size * sizeof(struct scatterlist);
2278 
2279 		sgp->slab = kmem_cache_create(sgp->name, size, 0,
2280 				SLAB_HWCACHE_ALIGN, NULL);
2281 		if (!sgp->slab) {
2282 			printk(KERN_ERR "SCSI: can't init sg slab %s\n",
2283 					sgp->name);
2284 			goto cleanup_sdb;
2285 		}
2286 
2287 		sgp->pool = mempool_create_slab_pool(SG_MEMPOOL_SIZE,
2288 						     sgp->slab);
2289 		if (!sgp->pool) {
2290 			printk(KERN_ERR "SCSI: can't init sg mempool %s\n",
2291 					sgp->name);
2292 			goto cleanup_sdb;
2293 		}
2294 	}
2295 
2296 	return 0;
2297 
2298 cleanup_sdb:
2299 	for (i = 0; i < SG_MEMPOOL_NR; i++) {
2300 		struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
2301 		if (sgp->pool)
2302 			mempool_destroy(sgp->pool);
2303 		if (sgp->slab)
2304 			kmem_cache_destroy(sgp->slab);
2305 	}
2306 	kmem_cache_destroy(scsi_sdb_cache);
2307 
2308 	return -ENOMEM;
2309 }
2310 
2311 void scsi_exit_queue(void)
2312 {
2313 	int i;
2314 
2315 	kmem_cache_destroy(scsi_sdb_cache);
2316 
2317 	for (i = 0; i < SG_MEMPOOL_NR; i++) {
2318 		struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
2319 		mempool_destroy(sgp->pool);
2320 		kmem_cache_destroy(sgp->slab);
2321 	}
2322 }
2323 
2324 /**
2325  *	scsi_mode_select - issue a mode select
2326  *	@sdev:	SCSI device to be queried
2327  *	@pf:	Page format bit (1 == standard, 0 == vendor specific)
2328  *	@sp:	Save page bit (0 == don't save, 1 == save)
2329  *	@modepage: mode page being requested
2330  *	@buffer: request buffer (may not be smaller than eight bytes)
2331  *	@len:	length of request buffer.
2332  *	@timeout: command timeout
2333  *	@retries: number of retries before failing
2334  *	@data: returns a structure abstracting the mode header data
2335  *	@sshdr: place to put sense data (or NULL if no sense to be collected).
2336  *		must be SCSI_SENSE_BUFFERSIZE big.
2337  *
2338  *	Returns zero if successful; negative error number or scsi
2339  *	status on error
2340  *
2341  */
2342 int
2343 scsi_mode_select(struct scsi_device *sdev, int pf, int sp, int modepage,
2344 		 unsigned char *buffer, int len, int timeout, int retries,
2345 		 struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
2346 {
2347 	unsigned char cmd[10];
2348 	unsigned char *real_buffer;
2349 	int ret;
2350 
2351 	memset(cmd, 0, sizeof(cmd));
2352 	cmd[1] = (pf ? 0x10 : 0) | (sp ? 0x01 : 0);
2353 
2354 	if (sdev->use_10_for_ms) {
2355 		if (len > 65535)
2356 			return -EINVAL;
2357 		real_buffer = kmalloc(8 + len, GFP_KERNEL);
2358 		if (!real_buffer)
2359 			return -ENOMEM;
2360 		memcpy(real_buffer + 8, buffer, len);
2361 		len += 8;
2362 		real_buffer[0] = 0;
2363 		real_buffer[1] = 0;
2364 		real_buffer[2] = data->medium_type;
2365 		real_buffer[3] = data->device_specific;
2366 		real_buffer[4] = data->longlba ? 0x01 : 0;
2367 		real_buffer[5] = 0;
2368 		real_buffer[6] = data->block_descriptor_length >> 8;
2369 		real_buffer[7] = data->block_descriptor_length;
2370 
2371 		cmd[0] = MODE_SELECT_10;
2372 		cmd[7] = len >> 8;
2373 		cmd[8] = len;
2374 	} else {
2375 		if (len > 255 || data->block_descriptor_length > 255 ||
2376 		    data->longlba)
2377 			return -EINVAL;
2378 
2379 		real_buffer = kmalloc(4 + len, GFP_KERNEL);
2380 		if (!real_buffer)
2381 			return -ENOMEM;
2382 		memcpy(real_buffer + 4, buffer, len);
2383 		len += 4;
2384 		real_buffer[0] = 0;
2385 		real_buffer[1] = data->medium_type;
2386 		real_buffer[2] = data->device_specific;
2387 		real_buffer[3] = data->block_descriptor_length;
2388 
2389 
2390 		cmd[0] = MODE_SELECT;
2391 		cmd[4] = len;
2392 	}
2393 
2394 	ret = scsi_execute_req(sdev, cmd, DMA_TO_DEVICE, real_buffer, len,
2395 			       sshdr, timeout, retries, NULL);
2396 	kfree(real_buffer);
2397 	return ret;
2398 }
2399 EXPORT_SYMBOL_GPL(scsi_mode_select);
2400 
2401 /**
2402  *	scsi_mode_sense - issue a mode sense, falling back from 10 to six bytes if necessary.
2403  *	@sdev:	SCSI device to be queried
2404  *	@dbd:	set if mode sense will allow block descriptors to be returned
2405  *	@modepage: mode page being requested
2406  *	@buffer: request buffer (may not be smaller than eight bytes)
2407  *	@len:	length of request buffer.
2408  *	@timeout: command timeout
2409  *	@retries: number of retries before failing
2410  *	@data: returns a structure abstracting the mode header data
2411  *	@sshdr: place to put sense data (or NULL if no sense to be collected).
2412  *		must be SCSI_SENSE_BUFFERSIZE big.
2413  *
2414  *	Returns zero if unsuccessful, or the header offset (either 4
2415  *	or 8 depending on whether a six or ten byte command was
2416  *	issued) if successful.
2417  */
2418 int
2419 scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
2420 		  unsigned char *buffer, int len, int timeout, int retries,
2421 		  struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
2422 {
2423 	unsigned char cmd[12];
2424 	int use_10_for_ms;
2425 	int header_length;
2426 	int result;
2427 	struct scsi_sense_hdr my_sshdr;
2428 
2429 	memset(data, 0, sizeof(*data));
2430 	memset(&cmd[0], 0, 12);
2431 	cmd[1] = dbd & 0x18;	/* allows DBD and LLBA bits */
2432 	cmd[2] = modepage;
2433 
2434 	/* caller might not be interested in sense, but we need it */
2435 	if (!sshdr)
2436 		sshdr = &my_sshdr;
2437 
2438  retry:
2439 	use_10_for_ms = sdev->use_10_for_ms;
2440 
2441 	if (use_10_for_ms) {
2442 		if (len < 8)
2443 			len = 8;
2444 
2445 		cmd[0] = MODE_SENSE_10;
2446 		cmd[8] = len;
2447 		header_length = 8;
2448 	} else {
2449 		if (len < 4)
2450 			len = 4;
2451 
2452 		cmd[0] = MODE_SENSE;
2453 		cmd[4] = len;
2454 		header_length = 4;
2455 	}
2456 
2457 	memset(buffer, 0, len);
2458 
2459 	result = scsi_execute_req(sdev, cmd, DMA_FROM_DEVICE, buffer, len,
2460 				  sshdr, timeout, retries, NULL);
2461 
2462 	/* This code looks awful: what it's doing is making sure an
2463 	 * ILLEGAL REQUEST sense return identifies the actual command
2464 	 * byte as the problem.  MODE_SENSE commands can return
2465 	 * ILLEGAL REQUEST if the code page isn't supported */
2466 
2467 	if (use_10_for_ms && !scsi_status_is_good(result) &&
2468 	    (driver_byte(result) & DRIVER_SENSE)) {
2469 		if (scsi_sense_valid(sshdr)) {
2470 			if ((sshdr->sense_key == ILLEGAL_REQUEST) &&
2471 			    (sshdr->asc == 0x20) && (sshdr->ascq == 0)) {
2472 				/*
2473 				 * Invalid command operation code
2474 				 */
2475 				sdev->use_10_for_ms = 0;
2476 				goto retry;
2477 			}
2478 		}
2479 	}
2480 
2481 	if(scsi_status_is_good(result)) {
2482 		if (unlikely(buffer[0] == 0x86 && buffer[1] == 0x0b &&
2483 			     (modepage == 6 || modepage == 8))) {
2484 			/* Initio breakage? */
2485 			header_length = 0;
2486 			data->length = 13;
2487 			data->medium_type = 0;
2488 			data->device_specific = 0;
2489 			data->longlba = 0;
2490 			data->block_descriptor_length = 0;
2491 		} else if(use_10_for_ms) {
2492 			data->length = buffer[0]*256 + buffer[1] + 2;
2493 			data->medium_type = buffer[2];
2494 			data->device_specific = buffer[3];
2495 			data->longlba = buffer[4] & 0x01;
2496 			data->block_descriptor_length = buffer[6]*256
2497 				+ buffer[7];
2498 		} else {
2499 			data->length = buffer[0] + 1;
2500 			data->medium_type = buffer[1];
2501 			data->device_specific = buffer[2];
2502 			data->block_descriptor_length = buffer[3];
2503 		}
2504 		data->header_length = header_length;
2505 	}
2506 
2507 	return result;
2508 }
2509 EXPORT_SYMBOL(scsi_mode_sense);
2510 
2511 /**
2512  *	scsi_test_unit_ready - test if unit is ready
2513  *	@sdev:	scsi device to change the state of.
2514  *	@timeout: command timeout
2515  *	@retries: number of retries before failing
2516  *	@sshdr_external: Optional pointer to struct scsi_sense_hdr for
2517  *		returning sense. Make sure that this is cleared before passing
2518  *		in.
2519  *
2520  *	Returns zero if unsuccessful or an error if TUR failed.  For
2521  *	removable media, UNIT_ATTENTION sets ->changed flag.
2522  **/
2523 int
2524 scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries,
2525 		     struct scsi_sense_hdr *sshdr_external)
2526 {
2527 	char cmd[] = {
2528 		TEST_UNIT_READY, 0, 0, 0, 0, 0,
2529 	};
2530 	struct scsi_sense_hdr *sshdr;
2531 	int result;
2532 
2533 	if (!sshdr_external)
2534 		sshdr = kzalloc(sizeof(*sshdr), GFP_KERNEL);
2535 	else
2536 		sshdr = sshdr_external;
2537 
2538 	/* try to eat the UNIT_ATTENTION if there are enough retries */
2539 	do {
2540 		result = scsi_execute_req(sdev, cmd, DMA_NONE, NULL, 0, sshdr,
2541 					  timeout, retries, NULL);
2542 		if (sdev->removable && scsi_sense_valid(sshdr) &&
2543 		    sshdr->sense_key == UNIT_ATTENTION)
2544 			sdev->changed = 1;
2545 	} while (scsi_sense_valid(sshdr) &&
2546 		 sshdr->sense_key == UNIT_ATTENTION && --retries);
2547 
2548 	if (!sshdr_external)
2549 		kfree(sshdr);
2550 	return result;
2551 }
2552 EXPORT_SYMBOL(scsi_test_unit_ready);
2553 
2554 /**
2555  *	scsi_device_set_state - Take the given device through the device state model.
2556  *	@sdev:	scsi device to change the state of.
2557  *	@state:	state to change to.
2558  *
2559  *	Returns zero if unsuccessful or an error if the requested
2560  *	transition is illegal.
2561  */
2562 int
2563 scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
2564 {
2565 	enum scsi_device_state oldstate = sdev->sdev_state;
2566 
2567 	if (state == oldstate)
2568 		return 0;
2569 
2570 	switch (state) {
2571 	case SDEV_CREATED:
2572 		switch (oldstate) {
2573 		case SDEV_CREATED_BLOCK:
2574 			break;
2575 		default:
2576 			goto illegal;
2577 		}
2578 		break;
2579 
2580 	case SDEV_RUNNING:
2581 		switch (oldstate) {
2582 		case SDEV_CREATED:
2583 		case SDEV_OFFLINE:
2584 		case SDEV_TRANSPORT_OFFLINE:
2585 		case SDEV_QUIESCE:
2586 		case SDEV_BLOCK:
2587 			break;
2588 		default:
2589 			goto illegal;
2590 		}
2591 		break;
2592 
2593 	case SDEV_QUIESCE:
2594 		switch (oldstate) {
2595 		case SDEV_RUNNING:
2596 		case SDEV_OFFLINE:
2597 		case SDEV_TRANSPORT_OFFLINE:
2598 			break;
2599 		default:
2600 			goto illegal;
2601 		}
2602 		break;
2603 
2604 	case SDEV_OFFLINE:
2605 	case SDEV_TRANSPORT_OFFLINE:
2606 		switch (oldstate) {
2607 		case SDEV_CREATED:
2608 		case SDEV_RUNNING:
2609 		case SDEV_QUIESCE:
2610 		case SDEV_BLOCK:
2611 			break;
2612 		default:
2613 			goto illegal;
2614 		}
2615 		break;
2616 
2617 	case SDEV_BLOCK:
2618 		switch (oldstate) {
2619 		case SDEV_RUNNING:
2620 		case SDEV_CREATED_BLOCK:
2621 			break;
2622 		default:
2623 			goto illegal;
2624 		}
2625 		break;
2626 
2627 	case SDEV_CREATED_BLOCK:
2628 		switch (oldstate) {
2629 		case SDEV_CREATED:
2630 			break;
2631 		default:
2632 			goto illegal;
2633 		}
2634 		break;
2635 
2636 	case SDEV_CANCEL:
2637 		switch (oldstate) {
2638 		case SDEV_CREATED:
2639 		case SDEV_RUNNING:
2640 		case SDEV_QUIESCE:
2641 		case SDEV_OFFLINE:
2642 		case SDEV_TRANSPORT_OFFLINE:
2643 		case SDEV_BLOCK:
2644 			break;
2645 		default:
2646 			goto illegal;
2647 		}
2648 		break;
2649 
2650 	case SDEV_DEL:
2651 		switch (oldstate) {
2652 		case SDEV_CREATED:
2653 		case SDEV_RUNNING:
2654 		case SDEV_OFFLINE:
2655 		case SDEV_TRANSPORT_OFFLINE:
2656 		case SDEV_CANCEL:
2657 		case SDEV_CREATED_BLOCK:
2658 			break;
2659 		default:
2660 			goto illegal;
2661 		}
2662 		break;
2663 
2664 	}
2665 	sdev->sdev_state = state;
2666 	return 0;
2667 
2668  illegal:
2669 	SCSI_LOG_ERROR_RECOVERY(1,
2670 				sdev_printk(KERN_ERR, sdev,
2671 					    "Illegal state transition %s->%s",
2672 					    scsi_device_state_name(oldstate),
2673 					    scsi_device_state_name(state))
2674 				);
2675 	return -EINVAL;
2676 }
2677 EXPORT_SYMBOL(scsi_device_set_state);
2678 
2679 /**
2680  * 	sdev_evt_emit - emit a single SCSI device uevent
2681  *	@sdev: associated SCSI device
2682  *	@evt: event to emit
2683  *
2684  *	Send a single uevent (scsi_event) to the associated scsi_device.
2685  */
2686 static void scsi_evt_emit(struct scsi_device *sdev, struct scsi_event *evt)
2687 {
2688 	int idx = 0;
2689 	char *envp[3];
2690 
2691 	switch (evt->evt_type) {
2692 	case SDEV_EVT_MEDIA_CHANGE:
2693 		envp[idx++] = "SDEV_MEDIA_CHANGE=1";
2694 		break;
2695 	case SDEV_EVT_INQUIRY_CHANGE_REPORTED:
2696 		envp[idx++] = "SDEV_UA=INQUIRY_DATA_HAS_CHANGED";
2697 		break;
2698 	case SDEV_EVT_CAPACITY_CHANGE_REPORTED:
2699 		envp[idx++] = "SDEV_UA=CAPACITY_DATA_HAS_CHANGED";
2700 		break;
2701 	case SDEV_EVT_SOFT_THRESHOLD_REACHED_REPORTED:
2702 	       envp[idx++] = "SDEV_UA=THIN_PROVISIONING_SOFT_THRESHOLD_REACHED";
2703 		break;
2704 	case SDEV_EVT_MODE_PARAMETER_CHANGE_REPORTED:
2705 		envp[idx++] = "SDEV_UA=MODE_PARAMETERS_CHANGED";
2706 		break;
2707 	case SDEV_EVT_LUN_CHANGE_REPORTED:
2708 		envp[idx++] = "SDEV_UA=REPORTED_LUNS_DATA_HAS_CHANGED";
2709 		break;
2710 	default:
2711 		/* do nothing */
2712 		break;
2713 	}
2714 
2715 	envp[idx++] = NULL;
2716 
2717 	kobject_uevent_env(&sdev->sdev_gendev.kobj, KOBJ_CHANGE, envp);
2718 }
2719 
2720 /**
2721  * 	sdev_evt_thread - send a uevent for each scsi event
2722  *	@work: work struct for scsi_device
2723  *
2724  *	Dispatch queued events to their associated scsi_device kobjects
2725  *	as uevents.
2726  */
2727 void scsi_evt_thread(struct work_struct *work)
2728 {
2729 	struct scsi_device *sdev;
2730 	enum scsi_device_event evt_type;
2731 	LIST_HEAD(event_list);
2732 
2733 	sdev = container_of(work, struct scsi_device, event_work);
2734 
2735 	for (evt_type = SDEV_EVT_FIRST; evt_type <= SDEV_EVT_LAST; evt_type++)
2736 		if (test_and_clear_bit(evt_type, sdev->pending_events))
2737 			sdev_evt_send_simple(sdev, evt_type, GFP_KERNEL);
2738 
2739 	while (1) {
2740 		struct scsi_event *evt;
2741 		struct list_head *this, *tmp;
2742 		unsigned long flags;
2743 
2744 		spin_lock_irqsave(&sdev->list_lock, flags);
2745 		list_splice_init(&sdev->event_list, &event_list);
2746 		spin_unlock_irqrestore(&sdev->list_lock, flags);
2747 
2748 		if (list_empty(&event_list))
2749 			break;
2750 
2751 		list_for_each_safe(this, tmp, &event_list) {
2752 			evt = list_entry(this, struct scsi_event, node);
2753 			list_del(&evt->node);
2754 			scsi_evt_emit(sdev, evt);
2755 			kfree(evt);
2756 		}
2757 	}
2758 }
2759 
2760 /**
2761  * 	sdev_evt_send - send asserted event to uevent thread
2762  *	@sdev: scsi_device event occurred on
2763  *	@evt: event to send
2764  *
2765  *	Assert scsi device event asynchronously.
2766  */
2767 void sdev_evt_send(struct scsi_device *sdev, struct scsi_event *evt)
2768 {
2769 	unsigned long flags;
2770 
2771 #if 0
2772 	/* FIXME: currently this check eliminates all media change events
2773 	 * for polled devices.  Need to update to discriminate between AN
2774 	 * and polled events */
2775 	if (!test_bit(evt->evt_type, sdev->supported_events)) {
2776 		kfree(evt);
2777 		return;
2778 	}
2779 #endif
2780 
2781 	spin_lock_irqsave(&sdev->list_lock, flags);
2782 	list_add_tail(&evt->node, &sdev->event_list);
2783 	schedule_work(&sdev->event_work);
2784 	spin_unlock_irqrestore(&sdev->list_lock, flags);
2785 }
2786 EXPORT_SYMBOL_GPL(sdev_evt_send);
2787 
2788 /**
2789  * 	sdev_evt_alloc - allocate a new scsi event
2790  *	@evt_type: type of event to allocate
2791  *	@gfpflags: GFP flags for allocation
2792  *
2793  *	Allocates and returns a new scsi_event.
2794  */
2795 struct scsi_event *sdev_evt_alloc(enum scsi_device_event evt_type,
2796 				  gfp_t gfpflags)
2797 {
2798 	struct scsi_event *evt = kzalloc(sizeof(struct scsi_event), gfpflags);
2799 	if (!evt)
2800 		return NULL;
2801 
2802 	evt->evt_type = evt_type;
2803 	INIT_LIST_HEAD(&evt->node);
2804 
2805 	/* evt_type-specific initialization, if any */
2806 	switch (evt_type) {
2807 	case SDEV_EVT_MEDIA_CHANGE:
2808 	case SDEV_EVT_INQUIRY_CHANGE_REPORTED:
2809 	case SDEV_EVT_CAPACITY_CHANGE_REPORTED:
2810 	case SDEV_EVT_SOFT_THRESHOLD_REACHED_REPORTED:
2811 	case SDEV_EVT_MODE_PARAMETER_CHANGE_REPORTED:
2812 	case SDEV_EVT_LUN_CHANGE_REPORTED:
2813 	default:
2814 		/* do nothing */
2815 		break;
2816 	}
2817 
2818 	return evt;
2819 }
2820 EXPORT_SYMBOL_GPL(sdev_evt_alloc);
2821 
2822 /**
2823  * 	sdev_evt_send_simple - send asserted event to uevent thread
2824  *	@sdev: scsi_device event occurred on
2825  *	@evt_type: type of event to send
2826  *	@gfpflags: GFP flags for allocation
2827  *
2828  *	Assert scsi device event asynchronously, given an event type.
2829  */
2830 void sdev_evt_send_simple(struct scsi_device *sdev,
2831 			  enum scsi_device_event evt_type, gfp_t gfpflags)
2832 {
2833 	struct scsi_event *evt = sdev_evt_alloc(evt_type, gfpflags);
2834 	if (!evt) {
2835 		sdev_printk(KERN_ERR, sdev, "event %d eaten due to OOM\n",
2836 			    evt_type);
2837 		return;
2838 	}
2839 
2840 	sdev_evt_send(sdev, evt);
2841 }
2842 EXPORT_SYMBOL_GPL(sdev_evt_send_simple);
2843 
2844 /**
2845  *	scsi_device_quiesce - Block user issued commands.
2846  *	@sdev:	scsi device to quiesce.
2847  *
2848  *	This works by trying to transition to the SDEV_QUIESCE state
2849  *	(which must be a legal transition).  When the device is in this
2850  *	state, only special requests will be accepted, all others will
2851  *	be deferred.  Since special requests may also be requeued requests,
2852  *	a successful return doesn't guarantee the device will be
2853  *	totally quiescent.
2854  *
2855  *	Must be called with user context, may sleep.
2856  *
2857  *	Returns zero if unsuccessful or an error if not.
2858  */
2859 int
2860 scsi_device_quiesce(struct scsi_device *sdev)
2861 {
2862 	int err = scsi_device_set_state(sdev, SDEV_QUIESCE);
2863 	if (err)
2864 		return err;
2865 
2866 	scsi_run_queue(sdev->request_queue);
2867 	while (atomic_read(&sdev->device_busy)) {
2868 		msleep_interruptible(200);
2869 		scsi_run_queue(sdev->request_queue);
2870 	}
2871 	return 0;
2872 }
2873 EXPORT_SYMBOL(scsi_device_quiesce);
2874 
2875 /**
2876  *	scsi_device_resume - Restart user issued commands to a quiesced device.
2877  *	@sdev:	scsi device to resume.
2878  *
2879  *	Moves the device from quiesced back to running and restarts the
2880  *	queues.
2881  *
2882  *	Must be called with user context, may sleep.
2883  */
2884 void scsi_device_resume(struct scsi_device *sdev)
2885 {
2886 	/* check if the device state was mutated prior to resume, and if
2887 	 * so assume the state is being managed elsewhere (for example
2888 	 * device deleted during suspend)
2889 	 */
2890 	if (sdev->sdev_state != SDEV_QUIESCE ||
2891 	    scsi_device_set_state(sdev, SDEV_RUNNING))
2892 		return;
2893 	scsi_run_queue(sdev->request_queue);
2894 }
2895 EXPORT_SYMBOL(scsi_device_resume);
2896 
2897 static void
2898 device_quiesce_fn(struct scsi_device *sdev, void *data)
2899 {
2900 	scsi_device_quiesce(sdev);
2901 }
2902 
2903 void
2904 scsi_target_quiesce(struct scsi_target *starget)
2905 {
2906 	starget_for_each_device(starget, NULL, device_quiesce_fn);
2907 }
2908 EXPORT_SYMBOL(scsi_target_quiesce);
2909 
2910 static void
2911 device_resume_fn(struct scsi_device *sdev, void *data)
2912 {
2913 	scsi_device_resume(sdev);
2914 }
2915 
2916 void
2917 scsi_target_resume(struct scsi_target *starget)
2918 {
2919 	starget_for_each_device(starget, NULL, device_resume_fn);
2920 }
2921 EXPORT_SYMBOL(scsi_target_resume);
2922 
2923 /**
2924  * scsi_internal_device_block - internal function to put a device temporarily into the SDEV_BLOCK state
2925  * @sdev:	device to block
2926  *
2927  * Block request made by scsi lld's to temporarily stop all
2928  * scsi commands on the specified device.  Called from interrupt
2929  * or normal process context.
2930  *
2931  * Returns zero if successful or error if not
2932  *
2933  * Notes:
2934  *	This routine transitions the device to the SDEV_BLOCK state
2935  *	(which must be a legal transition).  When the device is in this
2936  *	state, all commands are deferred until the scsi lld reenables
2937  *	the device with scsi_device_unblock or device_block_tmo fires.
2938  */
2939 int
2940 scsi_internal_device_block(struct scsi_device *sdev)
2941 {
2942 	struct request_queue *q = sdev->request_queue;
2943 	unsigned long flags;
2944 	int err = 0;
2945 
2946 	err = scsi_device_set_state(sdev, SDEV_BLOCK);
2947 	if (err) {
2948 		err = scsi_device_set_state(sdev, SDEV_CREATED_BLOCK);
2949 
2950 		if (err)
2951 			return err;
2952 	}
2953 
2954 	/*
2955 	 * The device has transitioned to SDEV_BLOCK.  Stop the
2956 	 * block layer from calling the midlayer with this device's
2957 	 * request queue.
2958 	 */
2959 	if (q->mq_ops) {
2960 		blk_mq_stop_hw_queues(q);
2961 	} else {
2962 		spin_lock_irqsave(q->queue_lock, flags);
2963 		blk_stop_queue(q);
2964 		spin_unlock_irqrestore(q->queue_lock, flags);
2965 	}
2966 
2967 	return 0;
2968 }
2969 EXPORT_SYMBOL_GPL(scsi_internal_device_block);
2970 
2971 /**
2972  * scsi_internal_device_unblock - resume a device after a block request
2973  * @sdev:	device to resume
2974  * @new_state:	state to set devices to after unblocking
2975  *
2976  * Called by scsi lld's or the midlayer to restart the device queue
2977  * for the previously suspended scsi device.  Called from interrupt or
2978  * normal process context.
2979  *
2980  * Returns zero if successful or error if not.
2981  *
2982  * Notes:
2983  *	This routine transitions the device to the SDEV_RUNNING state
2984  *	or to one of the offline states (which must be a legal transition)
2985  *	allowing the midlayer to goose the queue for this device.
2986  */
2987 int
2988 scsi_internal_device_unblock(struct scsi_device *sdev,
2989 			     enum scsi_device_state new_state)
2990 {
2991 	struct request_queue *q = sdev->request_queue;
2992 	unsigned long flags;
2993 
2994 	/*
2995 	 * Try to transition the scsi device to SDEV_RUNNING or one of the
2996 	 * offlined states and goose the device queue if successful.
2997 	 */
2998 	if ((sdev->sdev_state == SDEV_BLOCK) ||
2999 	    (sdev->sdev_state == SDEV_TRANSPORT_OFFLINE))
3000 		sdev->sdev_state = new_state;
3001 	else if (sdev->sdev_state == SDEV_CREATED_BLOCK) {
3002 		if (new_state == SDEV_TRANSPORT_OFFLINE ||
3003 		    new_state == SDEV_OFFLINE)
3004 			sdev->sdev_state = new_state;
3005 		else
3006 			sdev->sdev_state = SDEV_CREATED;
3007 	} else if (sdev->sdev_state != SDEV_CANCEL &&
3008 		 sdev->sdev_state != SDEV_OFFLINE)
3009 		return -EINVAL;
3010 
3011 	if (q->mq_ops) {
3012 		blk_mq_start_stopped_hw_queues(q, false);
3013 	} else {
3014 		spin_lock_irqsave(q->queue_lock, flags);
3015 		blk_start_queue(q);
3016 		spin_unlock_irqrestore(q->queue_lock, flags);
3017 	}
3018 
3019 	return 0;
3020 }
3021 EXPORT_SYMBOL_GPL(scsi_internal_device_unblock);
3022 
3023 static void
3024 device_block(struct scsi_device *sdev, void *data)
3025 {
3026 	scsi_internal_device_block(sdev);
3027 }
3028 
3029 static int
3030 target_block(struct device *dev, void *data)
3031 {
3032 	if (scsi_is_target_device(dev))
3033 		starget_for_each_device(to_scsi_target(dev), NULL,
3034 					device_block);
3035 	return 0;
3036 }
3037 
3038 void
3039 scsi_target_block(struct device *dev)
3040 {
3041 	if (scsi_is_target_device(dev))
3042 		starget_for_each_device(to_scsi_target(dev), NULL,
3043 					device_block);
3044 	else
3045 		device_for_each_child(dev, NULL, target_block);
3046 }
3047 EXPORT_SYMBOL_GPL(scsi_target_block);
3048 
3049 static void
3050 device_unblock(struct scsi_device *sdev, void *data)
3051 {
3052 	scsi_internal_device_unblock(sdev, *(enum scsi_device_state *)data);
3053 }
3054 
3055 static int
3056 target_unblock(struct device *dev, void *data)
3057 {
3058 	if (scsi_is_target_device(dev))
3059 		starget_for_each_device(to_scsi_target(dev), data,
3060 					device_unblock);
3061 	return 0;
3062 }
3063 
3064 void
3065 scsi_target_unblock(struct device *dev, enum scsi_device_state new_state)
3066 {
3067 	if (scsi_is_target_device(dev))
3068 		starget_for_each_device(to_scsi_target(dev), &new_state,
3069 					device_unblock);
3070 	else
3071 		device_for_each_child(dev, &new_state, target_unblock);
3072 }
3073 EXPORT_SYMBOL_GPL(scsi_target_unblock);
3074 
3075 /**
3076  * scsi_kmap_atomic_sg - find and atomically map an sg-elemnt
3077  * @sgl:	scatter-gather list
3078  * @sg_count:	number of segments in sg
3079  * @offset:	offset in bytes into sg, on return offset into the mapped area
3080  * @len:	bytes to map, on return number of bytes mapped
3081  *
3082  * Returns virtual address of the start of the mapped page
3083  */
3084 void *scsi_kmap_atomic_sg(struct scatterlist *sgl, int sg_count,
3085 			  size_t *offset, size_t *len)
3086 {
3087 	int i;
3088 	size_t sg_len = 0, len_complete = 0;
3089 	struct scatterlist *sg;
3090 	struct page *page;
3091 
3092 	WARN_ON(!irqs_disabled());
3093 
3094 	for_each_sg(sgl, sg, sg_count, i) {
3095 		len_complete = sg_len; /* Complete sg-entries */
3096 		sg_len += sg->length;
3097 		if (sg_len > *offset)
3098 			break;
3099 	}
3100 
3101 	if (unlikely(i == sg_count)) {
3102 		printk(KERN_ERR "%s: Bytes in sg: %zu, requested offset %zu, "
3103 			"elements %d\n",
3104 		       __func__, sg_len, *offset, sg_count);
3105 		WARN_ON(1);
3106 		return NULL;
3107 	}
3108 
3109 	/* Offset starting from the beginning of first page in this sg-entry */
3110 	*offset = *offset - len_complete + sg->offset;
3111 
3112 	/* Assumption: contiguous pages can be accessed as "page + i" */
3113 	page = nth_page(sg_page(sg), (*offset >> PAGE_SHIFT));
3114 	*offset &= ~PAGE_MASK;
3115 
3116 	/* Bytes in this sg-entry from *offset to the end of the page */
3117 	sg_len = PAGE_SIZE - *offset;
3118 	if (*len > sg_len)
3119 		*len = sg_len;
3120 
3121 	return kmap_atomic(page);
3122 }
3123 EXPORT_SYMBOL(scsi_kmap_atomic_sg);
3124 
3125 /**
3126  * scsi_kunmap_atomic_sg - atomically unmap a virtual address, previously mapped with scsi_kmap_atomic_sg
3127  * @virt:	virtual address to be unmapped
3128  */
3129 void scsi_kunmap_atomic_sg(void *virt)
3130 {
3131 	kunmap_atomic(virt);
3132 }
3133 EXPORT_SYMBOL(scsi_kunmap_atomic_sg);
3134 
3135 void sdev_disable_disk_events(struct scsi_device *sdev)
3136 {
3137 	atomic_inc(&sdev->disk_events_disable_depth);
3138 }
3139 EXPORT_SYMBOL(sdev_disable_disk_events);
3140 
3141 void sdev_enable_disk_events(struct scsi_device *sdev)
3142 {
3143 	if (WARN_ON_ONCE(atomic_read(&sdev->disk_events_disable_depth) <= 0))
3144 		return;
3145 	atomic_dec(&sdev->disk_events_disable_depth);
3146 }
3147 EXPORT_SYMBOL(sdev_enable_disk_events);
3148