scsi_lib.c revision 6bd522f6a226f435508433d24e0de4619e016a9d
1/*
2 *  scsi_lib.c Copyright (C) 1999 Eric Youngdale
3 *
4 *  SCSI queueing library.
5 *      Initial versions: Eric Youngdale (eric@andante.org).
6 *                        Based upon conversations with large numbers
7 *                        of people at Linux Expo.
8 */
9
10#include <linux/bio.h>
11#include <linux/bitops.h>
12#include <linux/blkdev.h>
13#include <linux/completion.h>
14#include <linux/kernel.h>
15#include <linux/mempool.h>
16#include <linux/slab.h>
17#include <linux/init.h>
18#include <linux/pci.h>
19#include <linux/delay.h>
20#include <linux/hardirq.h>
21#include <linux/scatterlist.h>
22
23#include <scsi/scsi.h>
24#include <scsi/scsi_cmnd.h>
25#include <scsi/scsi_dbg.h>
26#include <scsi/scsi_device.h>
27#include <scsi/scsi_driver.h>
28#include <scsi/scsi_eh.h>
29#include <scsi/scsi_host.h>
30
31#include "scsi_priv.h"
32#include "scsi_logging.h"
33
34
35#define SG_MEMPOOL_NR		ARRAY_SIZE(scsi_sg_pools)
36#define SG_MEMPOOL_SIZE		2
37
38struct scsi_host_sg_pool {
39	size_t		size;
40	char		*name;
41	struct kmem_cache	*slab;
42	mempool_t	*pool;
43};
44
45#define SP(x) { x, "sgpool-" __stringify(x) }
46#if (SCSI_MAX_SG_SEGMENTS < 32)
47#error SCSI_MAX_SG_SEGMENTS is too small (must be 32 or greater)
48#endif
49static struct scsi_host_sg_pool scsi_sg_pools[] = {
50	SP(8),
51	SP(16),
52#if (SCSI_MAX_SG_SEGMENTS > 32)
53	SP(32),
54#if (SCSI_MAX_SG_SEGMENTS > 64)
55	SP(64),
56#if (SCSI_MAX_SG_SEGMENTS > 128)
57	SP(128),
58#if (SCSI_MAX_SG_SEGMENTS > 256)
59#error SCSI_MAX_SG_SEGMENTS is too large (256 MAX)
60#endif
61#endif
62#endif
63#endif
64	SP(SCSI_MAX_SG_SEGMENTS)
65};
66#undef SP
67
68struct kmem_cache *scsi_sdb_cache;
69
70static void scsi_run_queue(struct request_queue *q);
71
72/*
73 * Function:	scsi_unprep_request()
74 *
75 * Purpose:	Remove all preparation done for a request, including its
76 *		associated scsi_cmnd, so that it can be requeued.
77 *
78 * Arguments:	req	- request to unprepare
79 *
80 * Lock status:	Assumed that no locks are held upon entry.
81 *
82 * Returns:	Nothing.
83 */
84static void scsi_unprep_request(struct request *req)
85{
86	struct scsi_cmnd *cmd = req->special;
87
88	req->cmd_flags &= ~REQ_DONTPREP;
89	req->special = NULL;
90
91	scsi_put_command(cmd);
92}
93
94/*
95 * Function:    scsi_queue_insert()
96 *
97 * Purpose:     Insert a command in the midlevel queue.
98 *
99 * Arguments:   cmd    - command that we are adding to queue.
100 *              reason - why we are inserting command to queue.
101 *
102 * Lock status: Assumed that lock is not held upon entry.
103 *
104 * Returns:     Nothing.
105 *
106 * Notes:       We do this for one of two cases.  Either the host is busy
107 *              and it cannot accept any more commands for the time being,
108 *              or the device returned QUEUE_FULL and can accept no more
109 *              commands.
110 * Notes:       This could be called either from an interrupt context or a
111 *              normal process context.
112 */
113int scsi_queue_insert(struct scsi_cmnd *cmd, int reason)
114{
115	struct Scsi_Host *host = cmd->device->host;
116	struct scsi_device *device = cmd->device;
117	struct request_queue *q = device->request_queue;
118	unsigned long flags;
119
120	SCSI_LOG_MLQUEUE(1,
121		 printk("Inserting command %p into mlqueue\n", cmd));
122
123	/*
124	 * Set the appropriate busy bit for the device/host.
125	 *
126	 * If the host/device isn't busy, assume that something actually
127	 * completed, and that we should be able to queue a command now.
128	 *
129	 * Note that the prior mid-layer assumption that any host could
130	 * always queue at least one command is now broken.  The mid-layer
131	 * will implement a user specifiable stall (see
132	 * scsi_host.max_host_blocked and scsi_device.max_device_blocked)
133	 * if a command is requeued with no other commands outstanding
134	 * either for the device or for the host.
135	 */
136	if (reason == SCSI_MLQUEUE_HOST_BUSY)
137		host->host_blocked = host->max_host_blocked;
138	else if (reason == SCSI_MLQUEUE_DEVICE_BUSY)
139		device->device_blocked = device->max_device_blocked;
140
141	/*
142	 * Decrement the counters, since these commands are no longer
143	 * active on the host/device.
144	 */
145	scsi_device_unbusy(device);
146
147	/*
148	 * Requeue this command.  It will go before all other commands
149	 * that are already in the queue.
150	 *
151	 * NOTE: there is magic here about the way the queue is plugged if
152	 * we have no outstanding commands.
153	 *
154	 * Although we *don't* plug the queue, we call the request
155	 * function.  The SCSI request function detects the blocked condition
156	 * and plugs the queue appropriately.
157         */
158	spin_lock_irqsave(q->queue_lock, flags);
159	blk_requeue_request(q, cmd->request);
160	spin_unlock_irqrestore(q->queue_lock, flags);
161
162	scsi_run_queue(q);
163
164	return 0;
165}
166
167/**
168 * scsi_execute - insert request and wait for the result
169 * @sdev:	scsi device
170 * @cmd:	scsi command
171 * @data_direction: data direction
172 * @buffer:	data buffer
173 * @bufflen:	len of buffer
174 * @sense:	optional sense buffer
175 * @timeout:	request timeout in seconds
176 * @retries:	number of times to retry request
177 * @flags:	or into request flags;
178 *
179 * returns the req->errors value which is the scsi_cmnd result
180 * field.
181 */
182int scsi_execute(struct scsi_device *sdev, const unsigned char *cmd,
183		 int data_direction, void *buffer, unsigned bufflen,
184		 unsigned char *sense, int timeout, int retries, int flags)
185{
186	struct request *req;
187	int write = (data_direction == DMA_TO_DEVICE);
188	int ret = DRIVER_ERROR << 24;
189
190	req = blk_get_request(sdev->request_queue, write, __GFP_WAIT);
191
192	if (bufflen &&	blk_rq_map_kern(sdev->request_queue, req,
193					buffer, bufflen, __GFP_WAIT))
194		goto out;
195
196	req->cmd_len = COMMAND_SIZE(cmd[0]);
197	memcpy(req->cmd, cmd, req->cmd_len);
198	req->sense = sense;
199	req->sense_len = 0;
200	req->retries = retries;
201	req->timeout = timeout;
202	req->cmd_type = REQ_TYPE_BLOCK_PC;
203	req->cmd_flags |= flags | REQ_QUIET | REQ_PREEMPT;
204
205	/*
206	 * head injection *required* here otherwise quiesce won't work
207	 */
208	blk_execute_rq(req->q, NULL, req, 1);
209
210	/*
211	 * Some devices (USB mass-storage in particular) may transfer
212	 * garbage data together with a residue indicating that the data
213	 * is invalid.  Prevent the garbage from being misinterpreted
214	 * and prevent security leaks by zeroing out the excess data.
215	 */
216	if (unlikely(req->data_len > 0 && req->data_len <= bufflen))
217		memset(buffer + (bufflen - req->data_len), 0, req->data_len);
218
219	ret = req->errors;
220 out:
221	blk_put_request(req);
222
223	return ret;
224}
225EXPORT_SYMBOL(scsi_execute);
226
227
228int scsi_execute_req(struct scsi_device *sdev, const unsigned char *cmd,
229		     int data_direction, void *buffer, unsigned bufflen,
230		     struct scsi_sense_hdr *sshdr, int timeout, int retries)
231{
232	char *sense = NULL;
233	int result;
234
235	if (sshdr) {
236		sense = kzalloc(SCSI_SENSE_BUFFERSIZE, GFP_NOIO);
237		if (!sense)
238			return DRIVER_ERROR << 24;
239	}
240	result = scsi_execute(sdev, cmd, data_direction, buffer, bufflen,
241			      sense, timeout, retries, 0);
242	if (sshdr)
243		scsi_normalize_sense(sense, SCSI_SENSE_BUFFERSIZE, sshdr);
244
245	kfree(sense);
246	return result;
247}
248EXPORT_SYMBOL(scsi_execute_req);
249
250struct scsi_io_context {
251	void *data;
252	void (*done)(void *data, char *sense, int result, int resid);
253	char sense[SCSI_SENSE_BUFFERSIZE];
254};
255
256static struct kmem_cache *scsi_io_context_cache;
257
258static void scsi_end_async(struct request *req, int uptodate)
259{
260	struct scsi_io_context *sioc = req->end_io_data;
261
262	if (sioc->done)
263		sioc->done(sioc->data, sioc->sense, req->errors, req->data_len);
264
265	kmem_cache_free(scsi_io_context_cache, sioc);
266	__blk_put_request(req->q, req);
267}
268
269static int scsi_merge_bio(struct request *rq, struct bio *bio)
270{
271	struct request_queue *q = rq->q;
272
273	bio->bi_flags &= ~(1 << BIO_SEG_VALID);
274	if (rq_data_dir(rq) == WRITE)
275		bio->bi_rw |= (1 << BIO_RW);
276	blk_queue_bounce(q, &bio);
277
278	return blk_rq_append_bio(q, rq, bio);
279}
280
281static void scsi_bi_endio(struct bio *bio, int error)
282{
283	bio_put(bio);
284}
285
286/**
287 * scsi_req_map_sg - map a scatterlist into a request
288 * @rq:		request to fill
289 * @sgl:	scatterlist
290 * @nsegs:	number of elements
291 * @bufflen:	len of buffer
292 * @gfp:	memory allocation flags
293 *
294 * scsi_req_map_sg maps a scatterlist into a request so that the
295 * request can be sent to the block layer. We do not trust the scatterlist
296 * sent to use, as some ULDs use that struct to only organize the pages.
297 */
298static int scsi_req_map_sg(struct request *rq, struct scatterlist *sgl,
299			   int nsegs, unsigned bufflen, gfp_t gfp)
300{
301	struct request_queue *q = rq->q;
302	int nr_pages = (bufflen + sgl[0].offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
303	unsigned int data_len = bufflen, len, bytes, off;
304	struct scatterlist *sg;
305	struct page *page;
306	struct bio *bio = NULL;
307	int i, err, nr_vecs = 0;
308
309	for_each_sg(sgl, sg, nsegs, i) {
310		page = sg_page(sg);
311		off = sg->offset;
312		len = sg->length;
313
314		while (len > 0 && data_len > 0) {
315			/*
316			 * sg sends a scatterlist that is larger than
317			 * the data_len it wants transferred for certain
318			 * IO sizes
319			 */
320			bytes = min_t(unsigned int, len, PAGE_SIZE - off);
321			bytes = min(bytes, data_len);
322
323			if (!bio) {
324				nr_vecs = min_t(int, BIO_MAX_PAGES, nr_pages);
325				nr_pages -= nr_vecs;
326
327				bio = bio_alloc(gfp, nr_vecs);
328				if (!bio) {
329					err = -ENOMEM;
330					goto free_bios;
331				}
332				bio->bi_end_io = scsi_bi_endio;
333			}
334
335			if (bio_add_pc_page(q, bio, page, bytes, off) !=
336			    bytes) {
337				bio_put(bio);
338				err = -EINVAL;
339				goto free_bios;
340			}
341
342			if (bio->bi_vcnt >= nr_vecs) {
343				err = scsi_merge_bio(rq, bio);
344				if (err) {
345					bio_endio(bio, 0);
346					goto free_bios;
347				}
348				bio = NULL;
349			}
350
351			page++;
352			len -= bytes;
353			data_len -=bytes;
354			off = 0;
355		}
356	}
357
358	rq->buffer = rq->data = NULL;
359	rq->data_len = bufflen;
360	return 0;
361
362free_bios:
363	while ((bio = rq->bio) != NULL) {
364		rq->bio = bio->bi_next;
365		/*
366		 * call endio instead of bio_put incase it was bounced
367		 */
368		bio_endio(bio, 0);
369	}
370
371	return err;
372}
373
374/**
375 * scsi_execute_async - insert request
376 * @sdev:	scsi device
377 * @cmd:	scsi command
378 * @cmd_len:	length of scsi cdb
379 * @data_direction: DMA_TO_DEVICE, DMA_FROM_DEVICE, or DMA_NONE
380 * @buffer:	data buffer (this can be a kernel buffer or scatterlist)
381 * @bufflen:	len of buffer
382 * @use_sg:	if buffer is a scatterlist this is the number of elements
383 * @timeout:	request timeout in seconds
384 * @retries:	number of times to retry request
385 * @privdata:	data passed to done()
386 * @done:	callback function when done
387 * @gfp:	memory allocation flags
388 */
389int scsi_execute_async(struct scsi_device *sdev, const unsigned char *cmd,
390		       int cmd_len, int data_direction, void *buffer, unsigned bufflen,
391		       int use_sg, int timeout, int retries, void *privdata,
392		       void (*done)(void *, char *, int, int), gfp_t gfp)
393{
394	struct request *req;
395	struct scsi_io_context *sioc;
396	int err = 0;
397	int write = (data_direction == DMA_TO_DEVICE);
398
399	sioc = kmem_cache_zalloc(scsi_io_context_cache, gfp);
400	if (!sioc)
401		return DRIVER_ERROR << 24;
402
403	req = blk_get_request(sdev->request_queue, write, gfp);
404	if (!req)
405		goto free_sense;
406	req->cmd_type = REQ_TYPE_BLOCK_PC;
407	req->cmd_flags |= REQ_QUIET;
408
409	if (use_sg)
410		err = scsi_req_map_sg(req, buffer, use_sg, bufflen, gfp);
411	else if (bufflen)
412		err = blk_rq_map_kern(req->q, req, buffer, bufflen, gfp);
413
414	if (err)
415		goto free_req;
416
417	req->cmd_len = cmd_len;
418	memset(req->cmd, 0, BLK_MAX_CDB); /* ATAPI hates garbage after CDB */
419	memcpy(req->cmd, cmd, req->cmd_len);
420	req->sense = sioc->sense;
421	req->sense_len = 0;
422	req->timeout = timeout;
423	req->retries = retries;
424	req->end_io_data = sioc;
425
426	sioc->data = privdata;
427	sioc->done = done;
428
429	blk_execute_rq_nowait(req->q, NULL, req, 1, scsi_end_async);
430	return 0;
431
432free_req:
433	blk_put_request(req);
434free_sense:
435	kmem_cache_free(scsi_io_context_cache, sioc);
436	return DRIVER_ERROR << 24;
437}
438EXPORT_SYMBOL_GPL(scsi_execute_async);
439
440/*
441 * Function:    scsi_init_cmd_errh()
442 *
443 * Purpose:     Initialize cmd fields related to error handling.
444 *
445 * Arguments:   cmd	- command that is ready to be queued.
446 *
447 * Notes:       This function has the job of initializing a number of
448 *              fields related to error handling.   Typically this will
449 *              be called once for each command, as required.
450 */
451static void scsi_init_cmd_errh(struct scsi_cmnd *cmd)
452{
453	cmd->serial_number = 0;
454	scsi_set_resid(cmd, 0);
455	memset(cmd->sense_buffer, 0, SCSI_SENSE_BUFFERSIZE);
456	if (cmd->cmd_len == 0)
457		cmd->cmd_len = scsi_command_size(cmd->cmnd);
458}
459
460void scsi_device_unbusy(struct scsi_device *sdev)
461{
462	struct Scsi_Host *shost = sdev->host;
463	unsigned long flags;
464
465	spin_lock_irqsave(shost->host_lock, flags);
466	shost->host_busy--;
467	if (unlikely(scsi_host_in_recovery(shost) &&
468		     (shost->host_failed || shost->host_eh_scheduled)))
469		scsi_eh_wakeup(shost);
470	spin_unlock(shost->host_lock);
471	spin_lock(sdev->request_queue->queue_lock);
472	sdev->device_busy--;
473	spin_unlock_irqrestore(sdev->request_queue->queue_lock, flags);
474}
475
476/*
477 * Called for single_lun devices on IO completion. Clear starget_sdev_user,
478 * and call blk_run_queue for all the scsi_devices on the target -
479 * including current_sdev first.
480 *
481 * Called with *no* scsi locks held.
482 */
483static void scsi_single_lun_run(struct scsi_device *current_sdev)
484{
485	struct Scsi_Host *shost = current_sdev->host;
486	struct scsi_device *sdev, *tmp;
487	struct scsi_target *starget = scsi_target(current_sdev);
488	unsigned long flags;
489
490	spin_lock_irqsave(shost->host_lock, flags);
491	starget->starget_sdev_user = NULL;
492	spin_unlock_irqrestore(shost->host_lock, flags);
493
494	/*
495	 * Call blk_run_queue for all LUNs on the target, starting with
496	 * current_sdev. We race with others (to set starget_sdev_user),
497	 * but in most cases, we will be first. Ideally, each LU on the
498	 * target would get some limited time or requests on the target.
499	 */
500	blk_run_queue(current_sdev->request_queue);
501
502	spin_lock_irqsave(shost->host_lock, flags);
503	if (starget->starget_sdev_user)
504		goto out;
505	list_for_each_entry_safe(sdev, tmp, &starget->devices,
506			same_target_siblings) {
507		if (sdev == current_sdev)
508			continue;
509		if (scsi_device_get(sdev))
510			continue;
511
512		spin_unlock_irqrestore(shost->host_lock, flags);
513		blk_run_queue(sdev->request_queue);
514		spin_lock_irqsave(shost->host_lock, flags);
515
516		scsi_device_put(sdev);
517	}
518 out:
519	spin_unlock_irqrestore(shost->host_lock, flags);
520}
521
522/*
523 * Function:	scsi_run_queue()
524 *
525 * Purpose:	Select a proper request queue to serve next
526 *
527 * Arguments:	q	- last request's queue
528 *
529 * Returns:     Nothing
530 *
531 * Notes:	The previous command was completely finished, start
532 *		a new one if possible.
533 */
534static void scsi_run_queue(struct request_queue *q)
535{
536	struct scsi_device *sdev = q->queuedata;
537	struct Scsi_Host *shost = sdev->host;
538	unsigned long flags;
539
540	if (scsi_target(sdev)->single_lun)
541		scsi_single_lun_run(sdev);
542
543	spin_lock_irqsave(shost->host_lock, flags);
544	while (!list_empty(&shost->starved_list) &&
545	       !shost->host_blocked && !shost->host_self_blocked &&
546		!((shost->can_queue > 0) &&
547		  (shost->host_busy >= shost->can_queue))) {
548
549		int flagset;
550
551		/*
552		 * As long as shost is accepting commands and we have
553		 * starved queues, call blk_run_queue. scsi_request_fn
554		 * drops the queue_lock and can add us back to the
555		 * starved_list.
556		 *
557		 * host_lock protects the starved_list and starved_entry.
558		 * scsi_request_fn must get the host_lock before checking
559		 * or modifying starved_list or starved_entry.
560		 */
561		sdev = list_entry(shost->starved_list.next,
562					  struct scsi_device, starved_entry);
563		list_del_init(&sdev->starved_entry);
564		spin_unlock(shost->host_lock);
565
566		spin_lock(sdev->request_queue->queue_lock);
567		flagset = test_bit(QUEUE_FLAG_REENTER, &q->queue_flags) &&
568				!test_bit(QUEUE_FLAG_REENTER,
569					&sdev->request_queue->queue_flags);
570		if (flagset)
571			queue_flag_set(QUEUE_FLAG_REENTER, sdev->request_queue);
572		__blk_run_queue(sdev->request_queue);
573		if (flagset)
574			queue_flag_clear(QUEUE_FLAG_REENTER, sdev->request_queue);
575		spin_unlock(sdev->request_queue->queue_lock);
576
577		spin_lock(shost->host_lock);
578		if (unlikely(!list_empty(&sdev->starved_entry)))
579			/*
580			 * sdev lost a race, and was put back on the
581			 * starved list. This is unlikely but without this
582			 * in theory we could loop forever.
583			 */
584			break;
585	}
586	spin_unlock_irqrestore(shost->host_lock, flags);
587
588	blk_run_queue(q);
589}
590
591/*
592 * Function:	scsi_requeue_command()
593 *
594 * Purpose:	Handle post-processing of completed commands.
595 *
596 * Arguments:	q	- queue to operate on
597 *		cmd	- command that may need to be requeued.
598 *
599 * Returns:	Nothing
600 *
601 * Notes:	After command completion, there may be blocks left
602 *		over which weren't finished by the previous command
603 *		this can be for a number of reasons - the main one is
604 *		I/O errors in the middle of the request, in which case
605 *		we need to request the blocks that come after the bad
606 *		sector.
607 * Notes:	Upon return, cmd is a stale pointer.
608 */
609static void scsi_requeue_command(struct request_queue *q, struct scsi_cmnd *cmd)
610{
611	struct request *req = cmd->request;
612	unsigned long flags;
613
614	scsi_unprep_request(req);
615	spin_lock_irqsave(q->queue_lock, flags);
616	blk_requeue_request(q, req);
617	spin_unlock_irqrestore(q->queue_lock, flags);
618
619	scsi_run_queue(q);
620}
621
622void scsi_next_command(struct scsi_cmnd *cmd)
623{
624	struct scsi_device *sdev = cmd->device;
625	struct request_queue *q = sdev->request_queue;
626
627	/* need to hold a reference on the device before we let go of the cmd */
628	get_device(&sdev->sdev_gendev);
629
630	scsi_put_command(cmd);
631	scsi_run_queue(q);
632
633	/* ok to remove device now */
634	put_device(&sdev->sdev_gendev);
635}
636
637void scsi_run_host_queues(struct Scsi_Host *shost)
638{
639	struct scsi_device *sdev;
640
641	shost_for_each_device(sdev, shost)
642		scsi_run_queue(sdev->request_queue);
643}
644
645/*
646 * Function:    scsi_end_request()
647 *
648 * Purpose:     Post-processing of completed commands (usually invoked at end
649 *		of upper level post-processing and scsi_io_completion).
650 *
651 * Arguments:   cmd	 - command that is complete.
652 *              error    - 0 if I/O indicates success, < 0 for I/O error.
653 *              bytes    - number of bytes of completed I/O
654 *		requeue  - indicates whether we should requeue leftovers.
655 *
656 * Lock status: Assumed that lock is not held upon entry.
657 *
658 * Returns:     cmd if requeue required, NULL otherwise.
659 *
660 * Notes:       This is called for block device requests in order to
661 *              mark some number of sectors as complete.
662 *
663 *		We are guaranteeing that the request queue will be goosed
664 *		at some point during this call.
665 * Notes:	If cmd was requeued, upon return it will be a stale pointer.
666 */
667static struct scsi_cmnd *scsi_end_request(struct scsi_cmnd *cmd, int error,
668					  int bytes, int requeue)
669{
670	struct request_queue *q = cmd->device->request_queue;
671	struct request *req = cmd->request;
672
673	/*
674	 * If there are blocks left over at the end, set up the command
675	 * to queue the remainder of them.
676	 */
677	if (blk_end_request(req, error, bytes)) {
678		int leftover = (req->hard_nr_sectors << 9);
679
680		if (blk_pc_request(req))
681			leftover = req->data_len;
682
683		/* kill remainder if no retrys */
684		if (error && blk_noretry_request(req))
685			blk_end_request(req, error, leftover);
686		else {
687			if (requeue) {
688				/*
689				 * Bleah.  Leftovers again.  Stick the
690				 * leftovers in the front of the
691				 * queue, and goose the queue again.
692				 */
693				scsi_requeue_command(q, cmd);
694				cmd = NULL;
695			}
696			return cmd;
697		}
698	}
699
700	/*
701	 * This will goose the queue request function at the end, so we don't
702	 * need to worry about launching another command.
703	 */
704	scsi_next_command(cmd);
705	return NULL;
706}
707
708static inline unsigned int scsi_sgtable_index(unsigned short nents)
709{
710	unsigned int index;
711
712	BUG_ON(nents > SCSI_MAX_SG_SEGMENTS);
713
714	if (nents <= 8)
715		index = 0;
716	else
717		index = get_count_order(nents) - 3;
718
719	return index;
720}
721
722static void scsi_sg_free(struct scatterlist *sgl, unsigned int nents)
723{
724	struct scsi_host_sg_pool *sgp;
725
726	sgp = scsi_sg_pools + scsi_sgtable_index(nents);
727	mempool_free(sgl, sgp->pool);
728}
729
730static struct scatterlist *scsi_sg_alloc(unsigned int nents, gfp_t gfp_mask)
731{
732	struct scsi_host_sg_pool *sgp;
733
734	sgp = scsi_sg_pools + scsi_sgtable_index(nents);
735	return mempool_alloc(sgp->pool, gfp_mask);
736}
737
738static int scsi_alloc_sgtable(struct scsi_data_buffer *sdb, int nents,
739			      gfp_t gfp_mask)
740{
741	int ret;
742
743	BUG_ON(!nents);
744
745	ret = __sg_alloc_table(&sdb->table, nents, SCSI_MAX_SG_SEGMENTS,
746			       gfp_mask, scsi_sg_alloc);
747	if (unlikely(ret))
748		__sg_free_table(&sdb->table, SCSI_MAX_SG_SEGMENTS,
749				scsi_sg_free);
750
751	return ret;
752}
753
754static void scsi_free_sgtable(struct scsi_data_buffer *sdb)
755{
756	__sg_free_table(&sdb->table, SCSI_MAX_SG_SEGMENTS, scsi_sg_free);
757}
758
759/*
760 * Function:    scsi_release_buffers()
761 *
762 * Purpose:     Completion processing for block device I/O requests.
763 *
764 * Arguments:   cmd	- command that we are bailing.
765 *
766 * Lock status: Assumed that no lock is held upon entry.
767 *
768 * Returns:     Nothing
769 *
770 * Notes:       In the event that an upper level driver rejects a
771 *		command, we must release resources allocated during
772 *		the __init_io() function.  Primarily this would involve
773 *		the scatter-gather table, and potentially any bounce
774 *		buffers.
775 */
776void scsi_release_buffers(struct scsi_cmnd *cmd)
777{
778	if (cmd->sdb.table.nents)
779		scsi_free_sgtable(&cmd->sdb);
780
781	memset(&cmd->sdb, 0, sizeof(cmd->sdb));
782
783	if (scsi_bidi_cmnd(cmd)) {
784		struct scsi_data_buffer *bidi_sdb =
785			cmd->request->next_rq->special;
786		scsi_free_sgtable(bidi_sdb);
787		kmem_cache_free(scsi_sdb_cache, bidi_sdb);
788		cmd->request->next_rq->special = NULL;
789	}
790
791	if (scsi_prot_sg_count(cmd))
792		scsi_free_sgtable(cmd->prot_sdb);
793}
794EXPORT_SYMBOL(scsi_release_buffers);
795
796/*
797 * Bidi commands Must be complete as a whole, both sides at once.
798 * If part of the bytes were written and lld returned
799 * scsi_in()->resid and/or scsi_out()->resid this information will be left
800 * in req->data_len and req->next_rq->data_len. The upper-layer driver can
801 * decide what to do with this information.
802 */
803static void scsi_end_bidi_request(struct scsi_cmnd *cmd)
804{
805	struct request *req = cmd->request;
806	unsigned int dlen = req->data_len;
807	unsigned int next_dlen = req->next_rq->data_len;
808
809	req->data_len = scsi_out(cmd)->resid;
810	req->next_rq->data_len = scsi_in(cmd)->resid;
811
812	/* The req and req->next_rq have not been completed */
813	BUG_ON(blk_end_bidi_request(req, 0, dlen, next_dlen));
814
815	scsi_release_buffers(cmd);
816
817	/*
818	 * This will goose the queue request function at the end, so we don't
819	 * need to worry about launching another command.
820	 */
821	scsi_next_command(cmd);
822}
823
824/*
825 * Function:    scsi_io_completion()
826 *
827 * Purpose:     Completion processing for block device I/O requests.
828 *
829 * Arguments:   cmd   - command that is finished.
830 *
831 * Lock status: Assumed that no lock is held upon entry.
832 *
833 * Returns:     Nothing
834 *
835 * Notes:       This function is matched in terms of capabilities to
836 *              the function that created the scatter-gather list.
837 *              In other words, if there are no bounce buffers
838 *              (the normal case for most drivers), we don't need
839 *              the logic to deal with cleaning up afterwards.
840 *
841 *		We must do one of several things here:
842 *
843 *		a) Call scsi_end_request.  This will finish off the
844 *		   specified number of sectors.  If we are done, the
845 *		   command block will be released, and the queue
846 *		   function will be goosed.  If we are not done, then
847 *		   scsi_end_request will directly goose the queue.
848 *
849 *		b) We can just use scsi_requeue_command() here.  This would
850 *		   be used if we just wanted to retry, for example.
851 */
852void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes)
853{
854	int result = cmd->result;
855	int this_count = scsi_bufflen(cmd);
856	struct request_queue *q = cmd->device->request_queue;
857	struct request *req = cmd->request;
858	int error = 0;
859	struct scsi_sense_hdr sshdr;
860	int sense_valid = 0;
861	int sense_deferred = 0;
862
863	if (result) {
864		sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
865		if (sense_valid)
866			sense_deferred = scsi_sense_is_deferred(&sshdr);
867	}
868
869	if (blk_pc_request(req)) { /* SG_IO ioctl from block level */
870		req->errors = result;
871		if (result) {
872			if (sense_valid && req->sense) {
873				/*
874				 * SG_IO wants current and deferred errors
875				 */
876				int len = 8 + cmd->sense_buffer[7];
877
878				if (len > SCSI_SENSE_BUFFERSIZE)
879					len = SCSI_SENSE_BUFFERSIZE;
880				memcpy(req->sense, cmd->sense_buffer,  len);
881				req->sense_len = len;
882			}
883			if (!sense_deferred)
884				error = -EIO;
885		}
886		if (scsi_bidi_cmnd(cmd)) {
887			/* will also release_buffers */
888			scsi_end_bidi_request(cmd);
889			return;
890		}
891		req->data_len = scsi_get_resid(cmd);
892	}
893
894	BUG_ON(blk_bidi_rq(req)); /* bidi not support for !blk_pc_request yet */
895	scsi_release_buffers(cmd);
896
897	/*
898	 * Next deal with any sectors which we were able to correctly
899	 * handle.
900	 */
901	SCSI_LOG_HLCOMPLETE(1, printk("%ld sectors total, "
902				      "%d bytes done.\n",
903				      req->nr_sectors, good_bytes));
904
905	/* A number of bytes were successfully read.  If there
906	 * are leftovers and there is some kind of error
907	 * (result != 0), retry the rest.
908	 */
909	if (scsi_end_request(cmd, error, good_bytes, result == 0) == NULL)
910		return;
911
912	/* good_bytes = 0, or (inclusive) there were leftovers and
913	 * result = 0, so scsi_end_request couldn't retry.
914	 */
915	if (sense_valid && !sense_deferred) {
916		switch (sshdr.sense_key) {
917		case UNIT_ATTENTION:
918			if (cmd->device->removable) {
919				/* Detected disc change.  Set a bit
920				 * and quietly refuse further access.
921				 */
922				cmd->device->changed = 1;
923				scsi_end_request(cmd, -EIO, this_count, 1);
924				return;
925			} else {
926				/* Must have been a power glitch, or a
927				 * bus reset.  Could not have been a
928				 * media change, so we just retry the
929				 * request and see what happens.
930				 */
931				scsi_requeue_command(q, cmd);
932				return;
933			}
934			break;
935		case ILLEGAL_REQUEST:
936			/* If we had an ILLEGAL REQUEST returned, then
937			 * we may have performed an unsupported
938			 * command.  The only thing this should be
939			 * would be a ten byte read where only a six
940			 * byte read was supported.  Also, on a system
941			 * where READ CAPACITY failed, we may have
942			 * read past the end of the disk.
943			 */
944			if ((cmd->device->use_10_for_rw &&
945			    sshdr.asc == 0x20 && sshdr.ascq == 0x00) &&
946			    (cmd->cmnd[0] == READ_10 ||
947			     cmd->cmnd[0] == WRITE_10)) {
948				cmd->device->use_10_for_rw = 0;
949				/* This will cause a retry with a
950				 * 6-byte command.
951				 */
952				scsi_requeue_command(q, cmd);
953			} else if (sshdr.asc == 0x10) /* DIX */
954				scsi_end_request(cmd, -EIO, this_count, 0);
955			else
956				scsi_end_request(cmd, -EIO, this_count, 1);
957			return;
958		case ABORTED_COMMAND:
959			if (sshdr.asc == 0x10) { /* DIF */
960				scsi_end_request(cmd, -EIO, this_count, 0);
961				return;
962			}
963			break;
964		case NOT_READY:
965			/* If the device is in the process of becoming
966			 * ready, or has a temporary blockage, retry.
967			 */
968			if (sshdr.asc == 0x04) {
969				switch (sshdr.ascq) {
970				case 0x01: /* becoming ready */
971				case 0x04: /* format in progress */
972				case 0x05: /* rebuild in progress */
973				case 0x06: /* recalculation in progress */
974				case 0x07: /* operation in progress */
975				case 0x08: /* Long write in progress */
976				case 0x09: /* self test in progress */
977					scsi_requeue_command(q, cmd);
978					return;
979				default:
980					break;
981				}
982			}
983			if (!(req->cmd_flags & REQ_QUIET))
984				scsi_cmd_print_sense_hdr(cmd,
985							 "Device not ready",
986							 &sshdr);
987
988			scsi_end_request(cmd, -EIO, this_count, 1);
989			return;
990		case VOLUME_OVERFLOW:
991			if (!(req->cmd_flags & REQ_QUIET)) {
992				scmd_printk(KERN_INFO, cmd,
993					    "Volume overflow, CDB: ");
994				__scsi_print_command(cmd->cmnd);
995				scsi_print_sense("", cmd);
996			}
997			/* See SSC3rXX or current. */
998			scsi_end_request(cmd, -EIO, this_count, 1);
999			return;
1000		default:
1001			break;
1002		}
1003	}
1004	if (host_byte(result) == DID_RESET) {
1005		/* Third party bus reset or reset for error recovery
1006		 * reasons.  Just retry the request and see what
1007		 * happens.
1008		 */
1009		scsi_requeue_command(q, cmd);
1010		return;
1011	}
1012	if (result) {
1013		if (!(req->cmd_flags & REQ_QUIET)) {
1014			scsi_print_result(cmd);
1015			if (driver_byte(result) & DRIVER_SENSE)
1016				scsi_print_sense("", cmd);
1017		}
1018	}
1019	scsi_end_request(cmd, -EIO, this_count, !result);
1020}
1021
1022static int scsi_init_sgtable(struct request *req, struct scsi_data_buffer *sdb,
1023			     gfp_t gfp_mask)
1024{
1025	int count;
1026
1027	/*
1028	 * If sg table allocation fails, requeue request later.
1029	 */
1030	if (unlikely(scsi_alloc_sgtable(sdb, req->nr_phys_segments,
1031					gfp_mask))) {
1032		return BLKPREP_DEFER;
1033	}
1034
1035	req->buffer = NULL;
1036
1037	/*
1038	 * Next, walk the list, and fill in the addresses and sizes of
1039	 * each segment.
1040	 */
1041	count = blk_rq_map_sg(req->q, req, sdb->table.sgl);
1042	BUG_ON(count > sdb->table.nents);
1043	sdb->table.nents = count;
1044	if (blk_pc_request(req))
1045		sdb->length = req->data_len;
1046	else
1047		sdb->length = req->nr_sectors << 9;
1048	return BLKPREP_OK;
1049}
1050
1051/*
1052 * Function:    scsi_init_io()
1053 *
1054 * Purpose:     SCSI I/O initialize function.
1055 *
1056 * Arguments:   cmd   - Command descriptor we wish to initialize
1057 *
1058 * Returns:     0 on success
1059 *		BLKPREP_DEFER if the failure is retryable
1060 *		BLKPREP_KILL if the failure is fatal
1061 */
1062int scsi_init_io(struct scsi_cmnd *cmd, gfp_t gfp_mask)
1063{
1064	int error = scsi_init_sgtable(cmd->request, &cmd->sdb, gfp_mask);
1065	if (error)
1066		goto err_exit;
1067
1068	if (blk_bidi_rq(cmd->request)) {
1069		struct scsi_data_buffer *bidi_sdb = kmem_cache_zalloc(
1070			scsi_sdb_cache, GFP_ATOMIC);
1071		if (!bidi_sdb) {
1072			error = BLKPREP_DEFER;
1073			goto err_exit;
1074		}
1075
1076		cmd->request->next_rq->special = bidi_sdb;
1077		error = scsi_init_sgtable(cmd->request->next_rq, bidi_sdb,
1078								    GFP_ATOMIC);
1079		if (error)
1080			goto err_exit;
1081	}
1082
1083	if (blk_integrity_rq(cmd->request)) {
1084		struct scsi_data_buffer *prot_sdb = cmd->prot_sdb;
1085		int ivecs, count;
1086
1087		BUG_ON(prot_sdb == NULL);
1088		ivecs = blk_rq_count_integrity_sg(cmd->request);
1089
1090		if (scsi_alloc_sgtable(prot_sdb, ivecs, gfp_mask)) {
1091			error = BLKPREP_DEFER;
1092			goto err_exit;
1093		}
1094
1095		count = blk_rq_map_integrity_sg(cmd->request,
1096						prot_sdb->table.sgl);
1097		BUG_ON(unlikely(count > ivecs));
1098
1099		cmd->prot_sdb = prot_sdb;
1100		cmd->prot_sdb->table.nents = count;
1101	}
1102
1103	return BLKPREP_OK ;
1104
1105err_exit:
1106	scsi_release_buffers(cmd);
1107	if (error == BLKPREP_KILL)
1108		scsi_put_command(cmd);
1109	else /* BLKPREP_DEFER */
1110		scsi_unprep_request(cmd->request);
1111
1112	return error;
1113}
1114EXPORT_SYMBOL(scsi_init_io);
1115
1116static struct scsi_cmnd *scsi_get_cmd_from_req(struct scsi_device *sdev,
1117		struct request *req)
1118{
1119	struct scsi_cmnd *cmd;
1120
1121	if (!req->special) {
1122		cmd = scsi_get_command(sdev, GFP_ATOMIC);
1123		if (unlikely(!cmd))
1124			return NULL;
1125		req->special = cmd;
1126	} else {
1127		cmd = req->special;
1128	}
1129
1130	/* pull a tag out of the request if we have one */
1131	cmd->tag = req->tag;
1132	cmd->request = req;
1133
1134	cmd->cmnd = req->cmd;
1135
1136	return cmd;
1137}
1138
1139int scsi_setup_blk_pc_cmnd(struct scsi_device *sdev, struct request *req)
1140{
1141	struct scsi_cmnd *cmd;
1142	int ret = scsi_prep_state_check(sdev, req);
1143
1144	if (ret != BLKPREP_OK)
1145		return ret;
1146
1147	cmd = scsi_get_cmd_from_req(sdev, req);
1148	if (unlikely(!cmd))
1149		return BLKPREP_DEFER;
1150
1151	/*
1152	 * BLOCK_PC requests may transfer data, in which case they must
1153	 * a bio attached to them.  Or they might contain a SCSI command
1154	 * that does not transfer data, in which case they may optionally
1155	 * submit a request without an attached bio.
1156	 */
1157	if (req->bio) {
1158		int ret;
1159
1160		BUG_ON(!req->nr_phys_segments);
1161
1162		ret = scsi_init_io(cmd, GFP_ATOMIC);
1163		if (unlikely(ret))
1164			return ret;
1165	} else {
1166		BUG_ON(req->data_len);
1167		BUG_ON(req->data);
1168
1169		memset(&cmd->sdb, 0, sizeof(cmd->sdb));
1170		req->buffer = NULL;
1171	}
1172
1173	cmd->cmd_len = req->cmd_len;
1174	if (!req->data_len)
1175		cmd->sc_data_direction = DMA_NONE;
1176	else if (rq_data_dir(req) == WRITE)
1177		cmd->sc_data_direction = DMA_TO_DEVICE;
1178	else
1179		cmd->sc_data_direction = DMA_FROM_DEVICE;
1180
1181	cmd->transfersize = req->data_len;
1182	cmd->allowed = req->retries;
1183	cmd->timeout_per_command = req->timeout;
1184	return BLKPREP_OK;
1185}
1186EXPORT_SYMBOL(scsi_setup_blk_pc_cmnd);
1187
1188/*
1189 * Setup a REQ_TYPE_FS command.  These are simple read/write request
1190 * from filesystems that still need to be translated to SCSI CDBs from
1191 * the ULD.
1192 */
1193int scsi_setup_fs_cmnd(struct scsi_device *sdev, struct request *req)
1194{
1195	struct scsi_cmnd *cmd;
1196	int ret = scsi_prep_state_check(sdev, req);
1197
1198	if (ret != BLKPREP_OK)
1199		return ret;
1200
1201	if (unlikely(sdev->scsi_dh_data && sdev->scsi_dh_data->scsi_dh
1202			 && sdev->scsi_dh_data->scsi_dh->prep_fn)) {
1203		ret = sdev->scsi_dh_data->scsi_dh->prep_fn(sdev, req);
1204		if (ret != BLKPREP_OK)
1205			return ret;
1206	}
1207
1208	/*
1209	 * Filesystem requests must transfer data.
1210	 */
1211	BUG_ON(!req->nr_phys_segments);
1212
1213	cmd = scsi_get_cmd_from_req(sdev, req);
1214	if (unlikely(!cmd))
1215		return BLKPREP_DEFER;
1216
1217	memset(cmd->cmnd, 0, BLK_MAX_CDB);
1218	return scsi_init_io(cmd, GFP_ATOMIC);
1219}
1220EXPORT_SYMBOL(scsi_setup_fs_cmnd);
1221
1222int scsi_prep_state_check(struct scsi_device *sdev, struct request *req)
1223{
1224	int ret = BLKPREP_OK;
1225
1226	/*
1227	 * If the device is not in running state we will reject some
1228	 * or all commands.
1229	 */
1230	if (unlikely(sdev->sdev_state != SDEV_RUNNING)) {
1231		switch (sdev->sdev_state) {
1232		case SDEV_OFFLINE:
1233			/*
1234			 * If the device is offline we refuse to process any
1235			 * commands.  The device must be brought online
1236			 * before trying any recovery commands.
1237			 */
1238			sdev_printk(KERN_ERR, sdev,
1239				    "rejecting I/O to offline device\n");
1240			ret = BLKPREP_KILL;
1241			break;
1242		case SDEV_DEL:
1243			/*
1244			 * If the device is fully deleted, we refuse to
1245			 * process any commands as well.
1246			 */
1247			sdev_printk(KERN_ERR, sdev,
1248				    "rejecting I/O to dead device\n");
1249			ret = BLKPREP_KILL;
1250			break;
1251		case SDEV_QUIESCE:
1252		case SDEV_BLOCK:
1253			/*
1254			 * If the devices is blocked we defer normal commands.
1255			 */
1256			if (!(req->cmd_flags & REQ_PREEMPT))
1257				ret = BLKPREP_DEFER;
1258			break;
1259		default:
1260			/*
1261			 * For any other not fully online state we only allow
1262			 * special commands.  In particular any user initiated
1263			 * command is not allowed.
1264			 */
1265			if (!(req->cmd_flags & REQ_PREEMPT))
1266				ret = BLKPREP_KILL;
1267			break;
1268		}
1269	}
1270	return ret;
1271}
1272EXPORT_SYMBOL(scsi_prep_state_check);
1273
1274int scsi_prep_return(struct request_queue *q, struct request *req, int ret)
1275{
1276	struct scsi_device *sdev = q->queuedata;
1277
1278	switch (ret) {
1279	case BLKPREP_KILL:
1280		req->errors = DID_NO_CONNECT << 16;
1281		/* release the command and kill it */
1282		if (req->special) {
1283			struct scsi_cmnd *cmd = req->special;
1284			scsi_release_buffers(cmd);
1285			scsi_put_command(cmd);
1286			req->special = NULL;
1287		}
1288		break;
1289	case BLKPREP_DEFER:
1290		/*
1291		 * If we defer, the elv_next_request() returns NULL, but the
1292		 * queue must be restarted, so we plug here if no returning
1293		 * command will automatically do that.
1294		 */
1295		if (sdev->device_busy == 0)
1296			blk_plug_device(q);
1297		break;
1298	default:
1299		req->cmd_flags |= REQ_DONTPREP;
1300	}
1301
1302	return ret;
1303}
1304EXPORT_SYMBOL(scsi_prep_return);
1305
1306int scsi_prep_fn(struct request_queue *q, struct request *req)
1307{
1308	struct scsi_device *sdev = q->queuedata;
1309	int ret = BLKPREP_KILL;
1310
1311	if (req->cmd_type == REQ_TYPE_BLOCK_PC)
1312		ret = scsi_setup_blk_pc_cmnd(sdev, req);
1313	return scsi_prep_return(q, req, ret);
1314}
1315
1316/*
1317 * scsi_dev_queue_ready: if we can send requests to sdev, return 1 else
1318 * return 0.
1319 *
1320 * Called with the queue_lock held.
1321 */
1322static inline int scsi_dev_queue_ready(struct request_queue *q,
1323				  struct scsi_device *sdev)
1324{
1325	if (sdev->device_busy >= sdev->queue_depth)
1326		return 0;
1327	if (sdev->device_busy == 0 && sdev->device_blocked) {
1328		/*
1329		 * unblock after device_blocked iterates to zero
1330		 */
1331		if (--sdev->device_blocked == 0) {
1332			SCSI_LOG_MLQUEUE(3,
1333				   sdev_printk(KERN_INFO, sdev,
1334				   "unblocking device at zero depth\n"));
1335		} else {
1336			blk_plug_device(q);
1337			return 0;
1338		}
1339	}
1340	if (sdev->device_blocked)
1341		return 0;
1342
1343	return 1;
1344}
1345
1346/*
1347 * scsi_host_queue_ready: if we can send requests to shost, return 1 else
1348 * return 0. We must end up running the queue again whenever 0 is
1349 * returned, else IO can hang.
1350 *
1351 * Called with host_lock held.
1352 */
1353static inline int scsi_host_queue_ready(struct request_queue *q,
1354				   struct Scsi_Host *shost,
1355				   struct scsi_device *sdev)
1356{
1357	if (scsi_host_in_recovery(shost))
1358		return 0;
1359	if (shost->host_busy == 0 && shost->host_blocked) {
1360		/*
1361		 * unblock after host_blocked iterates to zero
1362		 */
1363		if (--shost->host_blocked == 0) {
1364			SCSI_LOG_MLQUEUE(3,
1365				printk("scsi%d unblocking host at zero depth\n",
1366					shost->host_no));
1367		} else {
1368			return 0;
1369		}
1370	}
1371	if ((shost->can_queue > 0 && shost->host_busy >= shost->can_queue) ||
1372	    shost->host_blocked || shost->host_self_blocked) {
1373		if (list_empty(&sdev->starved_entry))
1374			list_add_tail(&sdev->starved_entry, &shost->starved_list);
1375		return 0;
1376	}
1377
1378	/* We're OK to process the command, so we can't be starved */
1379	if (!list_empty(&sdev->starved_entry))
1380		list_del_init(&sdev->starved_entry);
1381
1382	return 1;
1383}
1384
1385/*
1386 * Kill a request for a dead device
1387 */
1388static void scsi_kill_request(struct request *req, struct request_queue *q)
1389{
1390	struct scsi_cmnd *cmd = req->special;
1391	struct scsi_device *sdev = cmd->device;
1392	struct Scsi_Host *shost = sdev->host;
1393
1394	blkdev_dequeue_request(req);
1395
1396	if (unlikely(cmd == NULL)) {
1397		printk(KERN_CRIT "impossible request in %s.\n",
1398				 __FUNCTION__);
1399		BUG();
1400	}
1401
1402	scsi_init_cmd_errh(cmd);
1403	cmd->result = DID_NO_CONNECT << 16;
1404	atomic_inc(&cmd->device->iorequest_cnt);
1405
1406	/*
1407	 * SCSI request completion path will do scsi_device_unbusy(),
1408	 * bump busy counts.  To bump the counters, we need to dance
1409	 * with the locks as normal issue path does.
1410	 */
1411	sdev->device_busy++;
1412	spin_unlock(sdev->request_queue->queue_lock);
1413	spin_lock(shost->host_lock);
1414	shost->host_busy++;
1415	spin_unlock(shost->host_lock);
1416	spin_lock(sdev->request_queue->queue_lock);
1417
1418	__scsi_done(cmd);
1419}
1420
1421static void scsi_softirq_done(struct request *rq)
1422{
1423	struct scsi_cmnd *cmd = rq->completion_data;
1424	unsigned long wait_for = (cmd->allowed + 1) * cmd->timeout_per_command;
1425	int disposition;
1426
1427	INIT_LIST_HEAD(&cmd->eh_entry);
1428
1429	disposition = scsi_decide_disposition(cmd);
1430	if (disposition != SUCCESS &&
1431	    time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) {
1432		sdev_printk(KERN_ERR, cmd->device,
1433			    "timing out command, waited %lus\n",
1434			    wait_for/HZ);
1435		disposition = SUCCESS;
1436	}
1437
1438	scsi_log_completion(cmd, disposition);
1439
1440	switch (disposition) {
1441		case SUCCESS:
1442			scsi_finish_command(cmd);
1443			break;
1444		case NEEDS_RETRY:
1445			scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY);
1446			break;
1447		case ADD_TO_MLQUEUE:
1448			scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY);
1449			break;
1450		default:
1451			if (!scsi_eh_scmd_add(cmd, 0))
1452				scsi_finish_command(cmd);
1453	}
1454}
1455
1456/*
1457 * Function:    scsi_request_fn()
1458 *
1459 * Purpose:     Main strategy routine for SCSI.
1460 *
1461 * Arguments:   q       - Pointer to actual queue.
1462 *
1463 * Returns:     Nothing
1464 *
1465 * Lock status: IO request lock assumed to be held when called.
1466 */
1467static void scsi_request_fn(struct request_queue *q)
1468{
1469	struct scsi_device *sdev = q->queuedata;
1470	struct Scsi_Host *shost;
1471	struct scsi_cmnd *cmd;
1472	struct request *req;
1473
1474	if (!sdev) {
1475		printk("scsi: killing requests for dead queue\n");
1476		while ((req = elv_next_request(q)) != NULL)
1477			scsi_kill_request(req, q);
1478		return;
1479	}
1480
1481	if(!get_device(&sdev->sdev_gendev))
1482		/* We must be tearing the block queue down already */
1483		return;
1484
1485	/*
1486	 * To start with, we keep looping until the queue is empty, or until
1487	 * the host is no longer able to accept any more requests.
1488	 */
1489	shost = sdev->host;
1490	while (!blk_queue_plugged(q)) {
1491		int rtn;
1492		/*
1493		 * get next queueable request.  We do this early to make sure
1494		 * that the request is fully prepared even if we cannot
1495		 * accept it.
1496		 */
1497		req = elv_next_request(q);
1498		if (!req || !scsi_dev_queue_ready(q, sdev))
1499			break;
1500
1501		if (unlikely(!scsi_device_online(sdev))) {
1502			sdev_printk(KERN_ERR, sdev,
1503				    "rejecting I/O to offline device\n");
1504			scsi_kill_request(req, q);
1505			continue;
1506		}
1507
1508
1509		/*
1510		 * Remove the request from the request list.
1511		 */
1512		if (!(blk_queue_tagged(q) && !blk_queue_start_tag(q, req)))
1513			blkdev_dequeue_request(req);
1514		sdev->device_busy++;
1515
1516		spin_unlock(q->queue_lock);
1517		cmd = req->special;
1518		if (unlikely(cmd == NULL)) {
1519			printk(KERN_CRIT "impossible request in %s.\n"
1520					 "please mail a stack trace to "
1521					 "linux-scsi@vger.kernel.org\n",
1522					 __FUNCTION__);
1523			blk_dump_rq_flags(req, "foo");
1524			BUG();
1525		}
1526		spin_lock(shost->host_lock);
1527
1528		/*
1529		 * We hit this when the driver is using a host wide
1530		 * tag map. For device level tag maps the queue_depth check
1531		 * in the device ready fn would prevent us from trying
1532		 * to allocate a tag. Since the map is a shared host resource
1533		 * we add the dev to the starved list so it eventually gets
1534		 * a run when a tag is freed.
1535		 */
1536		if (blk_queue_tagged(q) && !blk_rq_tagged(req)) {
1537			if (list_empty(&sdev->starved_entry))
1538				list_add_tail(&sdev->starved_entry,
1539					      &shost->starved_list);
1540			goto not_ready;
1541		}
1542
1543		if (!scsi_host_queue_ready(q, shost, sdev))
1544			goto not_ready;
1545		if (scsi_target(sdev)->single_lun) {
1546			if (scsi_target(sdev)->starget_sdev_user &&
1547			    scsi_target(sdev)->starget_sdev_user != sdev)
1548				goto not_ready;
1549			scsi_target(sdev)->starget_sdev_user = sdev;
1550		}
1551		shost->host_busy++;
1552
1553		/*
1554		 * XXX(hch): This is rather suboptimal, scsi_dispatch_cmd will
1555		 *		take the lock again.
1556		 */
1557		spin_unlock_irq(shost->host_lock);
1558
1559		/*
1560		 * Finally, initialize any error handling parameters, and set up
1561		 * the timers for timeouts.
1562		 */
1563		scsi_init_cmd_errh(cmd);
1564
1565		/*
1566		 * Dispatch the command to the low-level driver.
1567		 */
1568		rtn = scsi_dispatch_cmd(cmd);
1569		spin_lock_irq(q->queue_lock);
1570		if(rtn) {
1571			/* we're refusing the command; because of
1572			 * the way locks get dropped, we need to
1573			 * check here if plugging is required */
1574			if(sdev->device_busy == 0)
1575				blk_plug_device(q);
1576
1577			break;
1578		}
1579	}
1580
1581	goto out;
1582
1583 not_ready:
1584	spin_unlock_irq(shost->host_lock);
1585
1586	/*
1587	 * lock q, handle tag, requeue req, and decrement device_busy. We
1588	 * must return with queue_lock held.
1589	 *
1590	 * Decrementing device_busy without checking it is OK, as all such
1591	 * cases (host limits or settings) should run the queue at some
1592	 * later time.
1593	 */
1594	spin_lock_irq(q->queue_lock);
1595	blk_requeue_request(q, req);
1596	sdev->device_busy--;
1597	if(sdev->device_busy == 0)
1598		blk_plug_device(q);
1599 out:
1600	/* must be careful here...if we trigger the ->remove() function
1601	 * we cannot be holding the q lock */
1602	spin_unlock_irq(q->queue_lock);
1603	put_device(&sdev->sdev_gendev);
1604	spin_lock_irq(q->queue_lock);
1605}
1606
1607u64 scsi_calculate_bounce_limit(struct Scsi_Host *shost)
1608{
1609	struct device *host_dev;
1610	u64 bounce_limit = 0xffffffff;
1611
1612	if (shost->unchecked_isa_dma)
1613		return BLK_BOUNCE_ISA;
1614	/*
1615	 * Platforms with virtual-DMA translation
1616	 * hardware have no practical limit.
1617	 */
1618	if (!PCI_DMA_BUS_IS_PHYS)
1619		return BLK_BOUNCE_ANY;
1620
1621	host_dev = scsi_get_device(shost);
1622	if (host_dev && host_dev->dma_mask)
1623		bounce_limit = *host_dev->dma_mask;
1624
1625	return bounce_limit;
1626}
1627EXPORT_SYMBOL(scsi_calculate_bounce_limit);
1628
1629struct request_queue *__scsi_alloc_queue(struct Scsi_Host *shost,
1630					 request_fn_proc *request_fn)
1631{
1632	struct request_queue *q;
1633	struct device *dev = shost->shost_gendev.parent;
1634
1635	q = blk_init_queue(request_fn, NULL);
1636	if (!q)
1637		return NULL;
1638
1639	/*
1640	 * this limit is imposed by hardware restrictions
1641	 */
1642	blk_queue_max_hw_segments(q, shost->sg_tablesize);
1643	blk_queue_max_phys_segments(q, SCSI_MAX_SG_CHAIN_SEGMENTS);
1644
1645	blk_queue_max_sectors(q, shost->max_sectors);
1646	blk_queue_bounce_limit(q, scsi_calculate_bounce_limit(shost));
1647	blk_queue_segment_boundary(q, shost->dma_boundary);
1648	dma_set_seg_boundary(dev, shost->dma_boundary);
1649
1650	blk_queue_max_segment_size(q, dma_get_max_seg_size(dev));
1651
1652	/* New queue, no concurrency on queue_flags */
1653	if (!shost->use_clustering)
1654		queue_flag_clear_unlocked(QUEUE_FLAG_CLUSTER, q);
1655
1656	/*
1657	 * set a reasonable default alignment on word boundaries: the
1658	 * host and device may alter it using
1659	 * blk_queue_update_dma_alignment() later.
1660	 */
1661	blk_queue_dma_alignment(q, 0x03);
1662
1663	return q;
1664}
1665EXPORT_SYMBOL(__scsi_alloc_queue);
1666
1667struct request_queue *scsi_alloc_queue(struct scsi_device *sdev)
1668{
1669	struct request_queue *q;
1670
1671	q = __scsi_alloc_queue(sdev->host, scsi_request_fn);
1672	if (!q)
1673		return NULL;
1674
1675	blk_queue_prep_rq(q, scsi_prep_fn);
1676	blk_queue_softirq_done(q, scsi_softirq_done);
1677	return q;
1678}
1679
1680void scsi_free_queue(struct request_queue *q)
1681{
1682	blk_cleanup_queue(q);
1683}
1684
1685/*
1686 * Function:    scsi_block_requests()
1687 *
1688 * Purpose:     Utility function used by low-level drivers to prevent further
1689 *		commands from being queued to the device.
1690 *
1691 * Arguments:   shost       - Host in question
1692 *
1693 * Returns:     Nothing
1694 *
1695 * Lock status: No locks are assumed held.
1696 *
1697 * Notes:       There is no timer nor any other means by which the requests
1698 *		get unblocked other than the low-level driver calling
1699 *		scsi_unblock_requests().
1700 */
1701void scsi_block_requests(struct Scsi_Host *shost)
1702{
1703	shost->host_self_blocked = 1;
1704}
1705EXPORT_SYMBOL(scsi_block_requests);
1706
1707/*
1708 * Function:    scsi_unblock_requests()
1709 *
1710 * Purpose:     Utility function used by low-level drivers to allow further
1711 *		commands from being queued to the device.
1712 *
1713 * Arguments:   shost       - Host in question
1714 *
1715 * Returns:     Nothing
1716 *
1717 * Lock status: No locks are assumed held.
1718 *
1719 * Notes:       There is no timer nor any other means by which the requests
1720 *		get unblocked other than the low-level driver calling
1721 *		scsi_unblock_requests().
1722 *
1723 *		This is done as an API function so that changes to the
1724 *		internals of the scsi mid-layer won't require wholesale
1725 *		changes to drivers that use this feature.
1726 */
1727void scsi_unblock_requests(struct Scsi_Host *shost)
1728{
1729	shost->host_self_blocked = 0;
1730	scsi_run_host_queues(shost);
1731}
1732EXPORT_SYMBOL(scsi_unblock_requests);
1733
1734int __init scsi_init_queue(void)
1735{
1736	int i;
1737
1738	scsi_io_context_cache = kmem_cache_create("scsi_io_context",
1739					sizeof(struct scsi_io_context),
1740					0, 0, NULL);
1741	if (!scsi_io_context_cache) {
1742		printk(KERN_ERR "SCSI: can't init scsi io context cache\n");
1743		return -ENOMEM;
1744	}
1745
1746	scsi_sdb_cache = kmem_cache_create("scsi_data_buffer",
1747					   sizeof(struct scsi_data_buffer),
1748					   0, 0, NULL);
1749	if (!scsi_sdb_cache) {
1750		printk(KERN_ERR "SCSI: can't init scsi sdb cache\n");
1751		goto cleanup_io_context;
1752	}
1753
1754	for (i = 0; i < SG_MEMPOOL_NR; i++) {
1755		struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1756		int size = sgp->size * sizeof(struct scatterlist);
1757
1758		sgp->slab = kmem_cache_create(sgp->name, size, 0,
1759				SLAB_HWCACHE_ALIGN, NULL);
1760		if (!sgp->slab) {
1761			printk(KERN_ERR "SCSI: can't init sg slab %s\n",
1762					sgp->name);
1763			goto cleanup_sdb;
1764		}
1765
1766		sgp->pool = mempool_create_slab_pool(SG_MEMPOOL_SIZE,
1767						     sgp->slab);
1768		if (!sgp->pool) {
1769			printk(KERN_ERR "SCSI: can't init sg mempool %s\n",
1770					sgp->name);
1771			goto cleanup_sdb;
1772		}
1773	}
1774
1775	return 0;
1776
1777cleanup_sdb:
1778	for (i = 0; i < SG_MEMPOOL_NR; i++) {
1779		struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1780		if (sgp->pool)
1781			mempool_destroy(sgp->pool);
1782		if (sgp->slab)
1783			kmem_cache_destroy(sgp->slab);
1784	}
1785	kmem_cache_destroy(scsi_sdb_cache);
1786cleanup_io_context:
1787	kmem_cache_destroy(scsi_io_context_cache);
1788
1789	return -ENOMEM;
1790}
1791
1792void scsi_exit_queue(void)
1793{
1794	int i;
1795
1796	kmem_cache_destroy(scsi_io_context_cache);
1797	kmem_cache_destroy(scsi_sdb_cache);
1798
1799	for (i = 0; i < SG_MEMPOOL_NR; i++) {
1800		struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1801		mempool_destroy(sgp->pool);
1802		kmem_cache_destroy(sgp->slab);
1803	}
1804}
1805
1806/**
1807 *	scsi_mode_select - issue a mode select
1808 *	@sdev:	SCSI device to be queried
1809 *	@pf:	Page format bit (1 == standard, 0 == vendor specific)
1810 *	@sp:	Save page bit (0 == don't save, 1 == save)
1811 *	@modepage: mode page being requested
1812 *	@buffer: request buffer (may not be smaller than eight bytes)
1813 *	@len:	length of request buffer.
1814 *	@timeout: command timeout
1815 *	@retries: number of retries before failing
1816 *	@data: returns a structure abstracting the mode header data
1817 *	@sshdr: place to put sense data (or NULL if no sense to be collected).
1818 *		must be SCSI_SENSE_BUFFERSIZE big.
1819 *
1820 *	Returns zero if successful; negative error number or scsi
1821 *	status on error
1822 *
1823 */
1824int
1825scsi_mode_select(struct scsi_device *sdev, int pf, int sp, int modepage,
1826		 unsigned char *buffer, int len, int timeout, int retries,
1827		 struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1828{
1829	unsigned char cmd[10];
1830	unsigned char *real_buffer;
1831	int ret;
1832
1833	memset(cmd, 0, sizeof(cmd));
1834	cmd[1] = (pf ? 0x10 : 0) | (sp ? 0x01 : 0);
1835
1836	if (sdev->use_10_for_ms) {
1837		if (len > 65535)
1838			return -EINVAL;
1839		real_buffer = kmalloc(8 + len, GFP_KERNEL);
1840		if (!real_buffer)
1841			return -ENOMEM;
1842		memcpy(real_buffer + 8, buffer, len);
1843		len += 8;
1844		real_buffer[0] = 0;
1845		real_buffer[1] = 0;
1846		real_buffer[2] = data->medium_type;
1847		real_buffer[3] = data->device_specific;
1848		real_buffer[4] = data->longlba ? 0x01 : 0;
1849		real_buffer[5] = 0;
1850		real_buffer[6] = data->block_descriptor_length >> 8;
1851		real_buffer[7] = data->block_descriptor_length;
1852
1853		cmd[0] = MODE_SELECT_10;
1854		cmd[7] = len >> 8;
1855		cmd[8] = len;
1856	} else {
1857		if (len > 255 || data->block_descriptor_length > 255 ||
1858		    data->longlba)
1859			return -EINVAL;
1860
1861		real_buffer = kmalloc(4 + len, GFP_KERNEL);
1862		if (!real_buffer)
1863			return -ENOMEM;
1864		memcpy(real_buffer + 4, buffer, len);
1865		len += 4;
1866		real_buffer[0] = 0;
1867		real_buffer[1] = data->medium_type;
1868		real_buffer[2] = data->device_specific;
1869		real_buffer[3] = data->block_descriptor_length;
1870
1871
1872		cmd[0] = MODE_SELECT;
1873		cmd[4] = len;
1874	}
1875
1876	ret = scsi_execute_req(sdev, cmd, DMA_TO_DEVICE, real_buffer, len,
1877			       sshdr, timeout, retries);
1878	kfree(real_buffer);
1879	return ret;
1880}
1881EXPORT_SYMBOL_GPL(scsi_mode_select);
1882
1883/**
1884 *	scsi_mode_sense - issue a mode sense, falling back from 10 to six bytes if necessary.
1885 *	@sdev:	SCSI device to be queried
1886 *	@dbd:	set if mode sense will allow block descriptors to be returned
1887 *	@modepage: mode page being requested
1888 *	@buffer: request buffer (may not be smaller than eight bytes)
1889 *	@len:	length of request buffer.
1890 *	@timeout: command timeout
1891 *	@retries: number of retries before failing
1892 *	@data: returns a structure abstracting the mode header data
1893 *	@sshdr: place to put sense data (or NULL if no sense to be collected).
1894 *		must be SCSI_SENSE_BUFFERSIZE big.
1895 *
1896 *	Returns zero if unsuccessful, or the header offset (either 4
1897 *	or 8 depending on whether a six or ten byte command was
1898 *	issued) if successful.
1899 */
1900int
1901scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
1902		  unsigned char *buffer, int len, int timeout, int retries,
1903		  struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1904{
1905	unsigned char cmd[12];
1906	int use_10_for_ms;
1907	int header_length;
1908	int result;
1909	struct scsi_sense_hdr my_sshdr;
1910
1911	memset(data, 0, sizeof(*data));
1912	memset(&cmd[0], 0, 12);
1913	cmd[1] = dbd & 0x18;	/* allows DBD and LLBA bits */
1914	cmd[2] = modepage;
1915
1916	/* caller might not be interested in sense, but we need it */
1917	if (!sshdr)
1918		sshdr = &my_sshdr;
1919
1920 retry:
1921	use_10_for_ms = sdev->use_10_for_ms;
1922
1923	if (use_10_for_ms) {
1924		if (len < 8)
1925			len = 8;
1926
1927		cmd[0] = MODE_SENSE_10;
1928		cmd[8] = len;
1929		header_length = 8;
1930	} else {
1931		if (len < 4)
1932			len = 4;
1933
1934		cmd[0] = MODE_SENSE;
1935		cmd[4] = len;
1936		header_length = 4;
1937	}
1938
1939	memset(buffer, 0, len);
1940
1941	result = scsi_execute_req(sdev, cmd, DMA_FROM_DEVICE, buffer, len,
1942				  sshdr, timeout, retries);
1943
1944	/* This code looks awful: what it's doing is making sure an
1945	 * ILLEGAL REQUEST sense return identifies the actual command
1946	 * byte as the problem.  MODE_SENSE commands can return
1947	 * ILLEGAL REQUEST if the code page isn't supported */
1948
1949	if (use_10_for_ms && !scsi_status_is_good(result) &&
1950	    (driver_byte(result) & DRIVER_SENSE)) {
1951		if (scsi_sense_valid(sshdr)) {
1952			if ((sshdr->sense_key == ILLEGAL_REQUEST) &&
1953			    (sshdr->asc == 0x20) && (sshdr->ascq == 0)) {
1954				/*
1955				 * Invalid command operation code
1956				 */
1957				sdev->use_10_for_ms = 0;
1958				goto retry;
1959			}
1960		}
1961	}
1962
1963	if(scsi_status_is_good(result)) {
1964		if (unlikely(buffer[0] == 0x86 && buffer[1] == 0x0b &&
1965			     (modepage == 6 || modepage == 8))) {
1966			/* Initio breakage? */
1967			header_length = 0;
1968			data->length = 13;
1969			data->medium_type = 0;
1970			data->device_specific = 0;
1971			data->longlba = 0;
1972			data->block_descriptor_length = 0;
1973		} else if(use_10_for_ms) {
1974			data->length = buffer[0]*256 + buffer[1] + 2;
1975			data->medium_type = buffer[2];
1976			data->device_specific = buffer[3];
1977			data->longlba = buffer[4] & 0x01;
1978			data->block_descriptor_length = buffer[6]*256
1979				+ buffer[7];
1980		} else {
1981			data->length = buffer[0] + 1;
1982			data->medium_type = buffer[1];
1983			data->device_specific = buffer[2];
1984			data->block_descriptor_length = buffer[3];
1985		}
1986		data->header_length = header_length;
1987	}
1988
1989	return result;
1990}
1991EXPORT_SYMBOL(scsi_mode_sense);
1992
1993/**
1994 *	scsi_test_unit_ready - test if unit is ready
1995 *	@sdev:	scsi device to change the state of.
1996 *	@timeout: command timeout
1997 *	@retries: number of retries before failing
1998 *	@sshdr_external: Optional pointer to struct scsi_sense_hdr for
1999 *		returning sense. Make sure that this is cleared before passing
2000 *		in.
2001 *
2002 *	Returns zero if unsuccessful or an error if TUR failed.  For
2003 *	removable media, a return of NOT_READY or UNIT_ATTENTION is
2004 *	translated to success, with the ->changed flag updated.
2005 **/
2006int
2007scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries,
2008		     struct scsi_sense_hdr *sshdr_external)
2009{
2010	char cmd[] = {
2011		TEST_UNIT_READY, 0, 0, 0, 0, 0,
2012	};
2013	struct scsi_sense_hdr *sshdr;
2014	int result;
2015
2016	if (!sshdr_external)
2017		sshdr = kzalloc(sizeof(*sshdr), GFP_KERNEL);
2018	else
2019		sshdr = sshdr_external;
2020
2021	/* try to eat the UNIT_ATTENTION if there are enough retries */
2022	do {
2023		result = scsi_execute_req(sdev, cmd, DMA_NONE, NULL, 0, sshdr,
2024					  timeout, retries);
2025	} while ((driver_byte(result) & DRIVER_SENSE) &&
2026		 sshdr && sshdr->sense_key == UNIT_ATTENTION &&
2027		 --retries);
2028
2029	if (!sshdr)
2030		/* could not allocate sense buffer, so can't process it */
2031		return result;
2032
2033	if ((driver_byte(result) & DRIVER_SENSE) && sdev->removable) {
2034
2035		if ((scsi_sense_valid(sshdr)) &&
2036		    ((sshdr->sense_key == UNIT_ATTENTION) ||
2037		     (sshdr->sense_key == NOT_READY))) {
2038			sdev->changed = 1;
2039			result = 0;
2040		}
2041	}
2042	if (!sshdr_external)
2043		kfree(sshdr);
2044	return result;
2045}
2046EXPORT_SYMBOL(scsi_test_unit_ready);
2047
2048/**
2049 *	scsi_device_set_state - Take the given device through the device state model.
2050 *	@sdev:	scsi device to change the state of.
2051 *	@state:	state to change to.
2052 *
2053 *	Returns zero if unsuccessful or an error if the requested
2054 *	transition is illegal.
2055 */
2056int
2057scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
2058{
2059	enum scsi_device_state oldstate = sdev->sdev_state;
2060
2061	if (state == oldstate)
2062		return 0;
2063
2064	switch (state) {
2065	case SDEV_CREATED:
2066		/* There are no legal states that come back to
2067		 * created.  This is the manually initialised start
2068		 * state */
2069		goto illegal;
2070
2071	case SDEV_RUNNING:
2072		switch (oldstate) {
2073		case SDEV_CREATED:
2074		case SDEV_OFFLINE:
2075		case SDEV_QUIESCE:
2076		case SDEV_BLOCK:
2077			break;
2078		default:
2079			goto illegal;
2080		}
2081		break;
2082
2083	case SDEV_QUIESCE:
2084		switch (oldstate) {
2085		case SDEV_RUNNING:
2086		case SDEV_OFFLINE:
2087			break;
2088		default:
2089			goto illegal;
2090		}
2091		break;
2092
2093	case SDEV_OFFLINE:
2094		switch (oldstate) {
2095		case SDEV_CREATED:
2096		case SDEV_RUNNING:
2097		case SDEV_QUIESCE:
2098		case SDEV_BLOCK:
2099			break;
2100		default:
2101			goto illegal;
2102		}
2103		break;
2104
2105	case SDEV_BLOCK:
2106		switch (oldstate) {
2107		case SDEV_CREATED:
2108		case SDEV_RUNNING:
2109			break;
2110		default:
2111			goto illegal;
2112		}
2113		break;
2114
2115	case SDEV_CANCEL:
2116		switch (oldstate) {
2117		case SDEV_CREATED:
2118		case SDEV_RUNNING:
2119		case SDEV_QUIESCE:
2120		case SDEV_OFFLINE:
2121		case SDEV_BLOCK:
2122			break;
2123		default:
2124			goto illegal;
2125		}
2126		break;
2127
2128	case SDEV_DEL:
2129		switch (oldstate) {
2130		case SDEV_CREATED:
2131		case SDEV_RUNNING:
2132		case SDEV_OFFLINE:
2133		case SDEV_CANCEL:
2134			break;
2135		default:
2136			goto illegal;
2137		}
2138		break;
2139
2140	}
2141	sdev->sdev_state = state;
2142	return 0;
2143
2144 illegal:
2145	SCSI_LOG_ERROR_RECOVERY(1,
2146				sdev_printk(KERN_ERR, sdev,
2147					    "Illegal state transition %s->%s\n",
2148					    scsi_device_state_name(oldstate),
2149					    scsi_device_state_name(state))
2150				);
2151	return -EINVAL;
2152}
2153EXPORT_SYMBOL(scsi_device_set_state);
2154
2155/**
2156 * 	sdev_evt_emit - emit a single SCSI device uevent
2157 *	@sdev: associated SCSI device
2158 *	@evt: event to emit
2159 *
2160 *	Send a single uevent (scsi_event) to the associated scsi_device.
2161 */
2162static void scsi_evt_emit(struct scsi_device *sdev, struct scsi_event *evt)
2163{
2164	int idx = 0;
2165	char *envp[3];
2166
2167	switch (evt->evt_type) {
2168	case SDEV_EVT_MEDIA_CHANGE:
2169		envp[idx++] = "SDEV_MEDIA_CHANGE=1";
2170		break;
2171
2172	default:
2173		/* do nothing */
2174		break;
2175	}
2176
2177	envp[idx++] = NULL;
2178
2179	kobject_uevent_env(&sdev->sdev_gendev.kobj, KOBJ_CHANGE, envp);
2180}
2181
2182/**
2183 * 	sdev_evt_thread - send a uevent for each scsi event
2184 *	@work: work struct for scsi_device
2185 *
2186 *	Dispatch queued events to their associated scsi_device kobjects
2187 *	as uevents.
2188 */
2189void scsi_evt_thread(struct work_struct *work)
2190{
2191	struct scsi_device *sdev;
2192	LIST_HEAD(event_list);
2193
2194	sdev = container_of(work, struct scsi_device, event_work);
2195
2196	while (1) {
2197		struct scsi_event *evt;
2198		struct list_head *this, *tmp;
2199		unsigned long flags;
2200
2201		spin_lock_irqsave(&sdev->list_lock, flags);
2202		list_splice_init(&sdev->event_list, &event_list);
2203		spin_unlock_irqrestore(&sdev->list_lock, flags);
2204
2205		if (list_empty(&event_list))
2206			break;
2207
2208		list_for_each_safe(this, tmp, &event_list) {
2209			evt = list_entry(this, struct scsi_event, node);
2210			list_del(&evt->node);
2211			scsi_evt_emit(sdev, evt);
2212			kfree(evt);
2213		}
2214	}
2215}
2216
2217/**
2218 * 	sdev_evt_send - send asserted event to uevent thread
2219 *	@sdev: scsi_device event occurred on
2220 *	@evt: event to send
2221 *
2222 *	Assert scsi device event asynchronously.
2223 */
2224void sdev_evt_send(struct scsi_device *sdev, struct scsi_event *evt)
2225{
2226	unsigned long flags;
2227
2228#if 0
2229	/* FIXME: currently this check eliminates all media change events
2230	 * for polled devices.  Need to update to discriminate between AN
2231	 * and polled events */
2232	if (!test_bit(evt->evt_type, sdev->supported_events)) {
2233		kfree(evt);
2234		return;
2235	}
2236#endif
2237
2238	spin_lock_irqsave(&sdev->list_lock, flags);
2239	list_add_tail(&evt->node, &sdev->event_list);
2240	schedule_work(&sdev->event_work);
2241	spin_unlock_irqrestore(&sdev->list_lock, flags);
2242}
2243EXPORT_SYMBOL_GPL(sdev_evt_send);
2244
2245/**
2246 * 	sdev_evt_alloc - allocate a new scsi event
2247 *	@evt_type: type of event to allocate
2248 *	@gfpflags: GFP flags for allocation
2249 *
2250 *	Allocates and returns a new scsi_event.
2251 */
2252struct scsi_event *sdev_evt_alloc(enum scsi_device_event evt_type,
2253				  gfp_t gfpflags)
2254{
2255	struct scsi_event *evt = kzalloc(sizeof(struct scsi_event), gfpflags);
2256	if (!evt)
2257		return NULL;
2258
2259	evt->evt_type = evt_type;
2260	INIT_LIST_HEAD(&evt->node);
2261
2262	/* evt_type-specific initialization, if any */
2263	switch (evt_type) {
2264	case SDEV_EVT_MEDIA_CHANGE:
2265	default:
2266		/* do nothing */
2267		break;
2268	}
2269
2270	return evt;
2271}
2272EXPORT_SYMBOL_GPL(sdev_evt_alloc);
2273
2274/**
2275 * 	sdev_evt_send_simple - send asserted event to uevent thread
2276 *	@sdev: scsi_device event occurred on
2277 *	@evt_type: type of event to send
2278 *	@gfpflags: GFP flags for allocation
2279 *
2280 *	Assert scsi device event asynchronously, given an event type.
2281 */
2282void sdev_evt_send_simple(struct scsi_device *sdev,
2283			  enum scsi_device_event evt_type, gfp_t gfpflags)
2284{
2285	struct scsi_event *evt = sdev_evt_alloc(evt_type, gfpflags);
2286	if (!evt) {
2287		sdev_printk(KERN_ERR, sdev, "event %d eaten due to OOM\n",
2288			    evt_type);
2289		return;
2290	}
2291
2292	sdev_evt_send(sdev, evt);
2293}
2294EXPORT_SYMBOL_GPL(sdev_evt_send_simple);
2295
2296/**
2297 *	scsi_device_quiesce - Block user issued commands.
2298 *	@sdev:	scsi device to quiesce.
2299 *
2300 *	This works by trying to transition to the SDEV_QUIESCE state
2301 *	(which must be a legal transition).  When the device is in this
2302 *	state, only special requests will be accepted, all others will
2303 *	be deferred.  Since special requests may also be requeued requests,
2304 *	a successful return doesn't guarantee the device will be
2305 *	totally quiescent.
2306 *
2307 *	Must be called with user context, may sleep.
2308 *
2309 *	Returns zero if unsuccessful or an error if not.
2310 */
2311int
2312scsi_device_quiesce(struct scsi_device *sdev)
2313{
2314	int err = scsi_device_set_state(sdev, SDEV_QUIESCE);
2315	if (err)
2316		return err;
2317
2318	scsi_run_queue(sdev->request_queue);
2319	while (sdev->device_busy) {
2320		msleep_interruptible(200);
2321		scsi_run_queue(sdev->request_queue);
2322	}
2323	return 0;
2324}
2325EXPORT_SYMBOL(scsi_device_quiesce);
2326
2327/**
2328 *	scsi_device_resume - Restart user issued commands to a quiesced device.
2329 *	@sdev:	scsi device to resume.
2330 *
2331 *	Moves the device from quiesced back to running and restarts the
2332 *	queues.
2333 *
2334 *	Must be called with user context, may sleep.
2335 */
2336void
2337scsi_device_resume(struct scsi_device *sdev)
2338{
2339	if(scsi_device_set_state(sdev, SDEV_RUNNING))
2340		return;
2341	scsi_run_queue(sdev->request_queue);
2342}
2343EXPORT_SYMBOL(scsi_device_resume);
2344
2345static void
2346device_quiesce_fn(struct scsi_device *sdev, void *data)
2347{
2348	scsi_device_quiesce(sdev);
2349}
2350
2351void
2352scsi_target_quiesce(struct scsi_target *starget)
2353{
2354	starget_for_each_device(starget, NULL, device_quiesce_fn);
2355}
2356EXPORT_SYMBOL(scsi_target_quiesce);
2357
2358static void
2359device_resume_fn(struct scsi_device *sdev, void *data)
2360{
2361	scsi_device_resume(sdev);
2362}
2363
2364void
2365scsi_target_resume(struct scsi_target *starget)
2366{
2367	starget_for_each_device(starget, NULL, device_resume_fn);
2368}
2369EXPORT_SYMBOL(scsi_target_resume);
2370
2371/**
2372 * scsi_internal_device_block - internal function to put a device temporarily into the SDEV_BLOCK state
2373 * @sdev:	device to block
2374 *
2375 * Block request made by scsi lld's to temporarily stop all
2376 * scsi commands on the specified device.  Called from interrupt
2377 * or normal process context.
2378 *
2379 * Returns zero if successful or error if not
2380 *
2381 * Notes:
2382 *	This routine transitions the device to the SDEV_BLOCK state
2383 *	(which must be a legal transition).  When the device is in this
2384 *	state, all commands are deferred until the scsi lld reenables
2385 *	the device with scsi_device_unblock or device_block_tmo fires.
2386 *	This routine assumes the host_lock is held on entry.
2387 */
2388int
2389scsi_internal_device_block(struct scsi_device *sdev)
2390{
2391	struct request_queue *q = sdev->request_queue;
2392	unsigned long flags;
2393	int err = 0;
2394
2395	err = scsi_device_set_state(sdev, SDEV_BLOCK);
2396	if (err)
2397		return err;
2398
2399	/*
2400	 * The device has transitioned to SDEV_BLOCK.  Stop the
2401	 * block layer from calling the midlayer with this device's
2402	 * request queue.
2403	 */
2404	spin_lock_irqsave(q->queue_lock, flags);
2405	blk_stop_queue(q);
2406	spin_unlock_irqrestore(q->queue_lock, flags);
2407
2408	return 0;
2409}
2410EXPORT_SYMBOL_GPL(scsi_internal_device_block);
2411
2412/**
2413 * scsi_internal_device_unblock - resume a device after a block request
2414 * @sdev:	device to resume
2415 *
2416 * Called by scsi lld's or the midlayer to restart the device queue
2417 * for the previously suspended scsi device.  Called from interrupt or
2418 * normal process context.
2419 *
2420 * Returns zero if successful or error if not.
2421 *
2422 * Notes:
2423 *	This routine transitions the device to the SDEV_RUNNING state
2424 *	(which must be a legal transition) allowing the midlayer to
2425 *	goose the queue for this device.  This routine assumes the
2426 *	host_lock is held upon entry.
2427 */
2428int
2429scsi_internal_device_unblock(struct scsi_device *sdev)
2430{
2431	struct request_queue *q = sdev->request_queue;
2432	int err;
2433	unsigned long flags;
2434
2435	/*
2436	 * Try to transition the scsi device to SDEV_RUNNING
2437	 * and goose the device queue if successful.
2438	 */
2439	err = scsi_device_set_state(sdev, SDEV_RUNNING);
2440	if (err)
2441		return err;
2442
2443	spin_lock_irqsave(q->queue_lock, flags);
2444	blk_start_queue(q);
2445	spin_unlock_irqrestore(q->queue_lock, flags);
2446
2447	return 0;
2448}
2449EXPORT_SYMBOL_GPL(scsi_internal_device_unblock);
2450
2451static void
2452device_block(struct scsi_device *sdev, void *data)
2453{
2454	scsi_internal_device_block(sdev);
2455}
2456
2457static int
2458target_block(struct device *dev, void *data)
2459{
2460	if (scsi_is_target_device(dev))
2461		starget_for_each_device(to_scsi_target(dev), NULL,
2462					device_block);
2463	return 0;
2464}
2465
2466void
2467scsi_target_block(struct device *dev)
2468{
2469	if (scsi_is_target_device(dev))
2470		starget_for_each_device(to_scsi_target(dev), NULL,
2471					device_block);
2472	else
2473		device_for_each_child(dev, NULL, target_block);
2474}
2475EXPORT_SYMBOL_GPL(scsi_target_block);
2476
2477static void
2478device_unblock(struct scsi_device *sdev, void *data)
2479{
2480	scsi_internal_device_unblock(sdev);
2481}
2482
2483static int
2484target_unblock(struct device *dev, void *data)
2485{
2486	if (scsi_is_target_device(dev))
2487		starget_for_each_device(to_scsi_target(dev), NULL,
2488					device_unblock);
2489	return 0;
2490}
2491
2492void
2493scsi_target_unblock(struct device *dev)
2494{
2495	if (scsi_is_target_device(dev))
2496		starget_for_each_device(to_scsi_target(dev), NULL,
2497					device_unblock);
2498	else
2499		device_for_each_child(dev, NULL, target_unblock);
2500}
2501EXPORT_SYMBOL_GPL(scsi_target_unblock);
2502
2503/**
2504 * scsi_kmap_atomic_sg - find and atomically map an sg-elemnt
2505 * @sgl:	scatter-gather list
2506 * @sg_count:	number of segments in sg
2507 * @offset:	offset in bytes into sg, on return offset into the mapped area
2508 * @len:	bytes to map, on return number of bytes mapped
2509 *
2510 * Returns virtual address of the start of the mapped page
2511 */
2512void *scsi_kmap_atomic_sg(struct scatterlist *sgl, int sg_count,
2513			  size_t *offset, size_t *len)
2514{
2515	int i;
2516	size_t sg_len = 0, len_complete = 0;
2517	struct scatterlist *sg;
2518	struct page *page;
2519
2520	WARN_ON(!irqs_disabled());
2521
2522	for_each_sg(sgl, sg, sg_count, i) {
2523		len_complete = sg_len; /* Complete sg-entries */
2524		sg_len += sg->length;
2525		if (sg_len > *offset)
2526			break;
2527	}
2528
2529	if (unlikely(i == sg_count)) {
2530		printk(KERN_ERR "%s: Bytes in sg: %zu, requested offset %zu, "
2531			"elements %d\n",
2532		       __FUNCTION__, sg_len, *offset, sg_count);
2533		WARN_ON(1);
2534		return NULL;
2535	}
2536
2537	/* Offset starting from the beginning of first page in this sg-entry */
2538	*offset = *offset - len_complete + sg->offset;
2539
2540	/* Assumption: contiguous pages can be accessed as "page + i" */
2541	page = nth_page(sg_page(sg), (*offset >> PAGE_SHIFT));
2542	*offset &= ~PAGE_MASK;
2543
2544	/* Bytes in this sg-entry from *offset to the end of the page */
2545	sg_len = PAGE_SIZE - *offset;
2546	if (*len > sg_len)
2547		*len = sg_len;
2548
2549	return kmap_atomic(page, KM_BIO_SRC_IRQ);
2550}
2551EXPORT_SYMBOL(scsi_kmap_atomic_sg);
2552
2553/**
2554 * scsi_kunmap_atomic_sg - atomically unmap a virtual address, previously mapped with scsi_kmap_atomic_sg
2555 * @virt:	virtual address to be unmapped
2556 */
2557void scsi_kunmap_atomic_sg(void *virt)
2558{
2559	kunmap_atomic(virt, KM_BIO_SRC_IRQ);
2560}
2561EXPORT_SYMBOL(scsi_kunmap_atomic_sg);
2562