Merge branch 'master' of git://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux-2.6
[linux-2.6] / drivers / scsi / scsi_lib.c
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/blkdev.h>
12 #include <linux/completion.h>
13 #include <linux/kernel.h>
14 #include <linux/mempool.h>
15 #include <linux/slab.h>
16 #include <linux/init.h>
17 #include <linux/pci.h>
18 #include <linux/delay.h>
19 #include <linux/hardirq.h>
20
21 #include <scsi/scsi.h>
22 #include <scsi/scsi_cmnd.h>
23 #include <scsi/scsi_dbg.h>
24 #include <scsi/scsi_device.h>
25 #include <scsi/scsi_driver.h>
26 #include <scsi/scsi_eh.h>
27 #include <scsi/scsi_host.h>
28
29 #include "scsi_priv.h"
30 #include "scsi_logging.h"
31
32
33 #define SG_MEMPOOL_NR           ARRAY_SIZE(scsi_sg_pools)
34 #define SG_MEMPOOL_SIZE         2
35
36 struct scsi_host_sg_pool {
37         size_t          size;
38         char            *name; 
39         struct kmem_cache       *slab;
40         mempool_t       *pool;
41 };
42
43 #if (SCSI_MAX_PHYS_SEGMENTS < 32)
44 #error SCSI_MAX_PHYS_SEGMENTS is too small
45 #endif
46
47 #define SP(x) { x, "sgpool-" #x } 
48 static struct scsi_host_sg_pool scsi_sg_pools[] = {
49         SP(8),
50         SP(16),
51         SP(32),
52 #if (SCSI_MAX_PHYS_SEGMENTS > 32)
53         SP(64),
54 #if (SCSI_MAX_PHYS_SEGMENTS > 64)
55         SP(128),
56 #if (SCSI_MAX_PHYS_SEGMENTS > 128)
57         SP(256),
58 #if (SCSI_MAX_PHYS_SEGMENTS > 256)
59 #error SCSI_MAX_PHYS_SEGMENTS is too large
60 #endif
61 #endif
62 #endif
63 #endif
64 };      
65 #undef SP
66
67 static void scsi_run_queue(struct request_queue *q);
68
69 /*
70  * Function:    scsi_unprep_request()
71  *
72  * Purpose:     Remove all preparation done for a request, including its
73  *              associated scsi_cmnd, so that it can be requeued.
74  *
75  * Arguments:   req     - request to unprepare
76  *
77  * Lock status: Assumed that no locks are held upon entry.
78  *
79  * Returns:     Nothing.
80  */
81 static void scsi_unprep_request(struct request *req)
82 {
83         struct scsi_cmnd *cmd = req->special;
84
85         req->cmd_flags &= ~REQ_DONTPREP;
86         req->special = NULL;
87
88         scsi_put_command(cmd);
89 }
90
91 /*
92  * Function:    scsi_queue_insert()
93  *
94  * Purpose:     Insert a command in the midlevel queue.
95  *
96  * Arguments:   cmd    - command that we are adding to queue.
97  *              reason - why we are inserting command to queue.
98  *
99  * Lock status: Assumed that lock is not held upon entry.
100  *
101  * Returns:     Nothing.
102  *
103  * Notes:       We do this for one of two cases.  Either the host is busy
104  *              and it cannot accept any more commands for the time being,
105  *              or the device returned QUEUE_FULL and can accept no more
106  *              commands.
107  * Notes:       This could be called either from an interrupt context or a
108  *              normal process context.
109  */
110 int scsi_queue_insert(struct scsi_cmnd *cmd, int reason)
111 {
112         struct Scsi_Host *host = cmd->device->host;
113         struct scsi_device *device = cmd->device;
114         struct request_queue *q = device->request_queue;
115         unsigned long flags;
116
117         SCSI_LOG_MLQUEUE(1,
118                  printk("Inserting command %p into mlqueue\n", cmd));
119
120         /*
121          * Set the appropriate busy bit for the device/host.
122          *
123          * If the host/device isn't busy, assume that something actually
124          * completed, and that we should be able to queue a command now.
125          *
126          * Note that the prior mid-layer assumption that any host could
127          * always queue at least one command is now broken.  The mid-layer
128          * will implement a user specifiable stall (see
129          * scsi_host.max_host_blocked and scsi_device.max_device_blocked)
130          * if a command is requeued with no other commands outstanding
131          * either for the device or for the host.
132          */
133         if (reason == SCSI_MLQUEUE_HOST_BUSY)
134                 host->host_blocked = host->max_host_blocked;
135         else if (reason == SCSI_MLQUEUE_DEVICE_BUSY)
136                 device->device_blocked = device->max_device_blocked;
137
138         /*
139          * Decrement the counters, since these commands are no longer
140          * active on the host/device.
141          */
142         scsi_device_unbusy(device);
143
144         /*
145          * Requeue this command.  It will go before all other commands
146          * that are already in the queue.
147          *
148          * NOTE: there is magic here about the way the queue is plugged if
149          * we have no outstanding commands.
150          * 
151          * Although we *don't* plug the queue, we call the request
152          * function.  The SCSI request function detects the blocked condition
153          * and plugs the queue appropriately.
154          */
155         spin_lock_irqsave(q->queue_lock, flags);
156         blk_requeue_request(q, cmd->request);
157         spin_unlock_irqrestore(q->queue_lock, flags);
158
159         scsi_run_queue(q);
160
161         return 0;
162 }
163
164 /**
165  * scsi_execute - insert request and wait for the result
166  * @sdev:       scsi device
167  * @cmd:        scsi command
168  * @data_direction: data direction
169  * @buffer:     data buffer
170  * @bufflen:    len of buffer
171  * @sense:      optional sense buffer
172  * @timeout:    request timeout in seconds
173  * @retries:    number of times to retry request
174  * @flags:      or into request flags;
175  *
176  * returns the req->errors value which is the scsi_cmnd result
177  * field.
178  **/
179 int scsi_execute(struct scsi_device *sdev, const unsigned char *cmd,
180                  int data_direction, void *buffer, unsigned bufflen,
181                  unsigned char *sense, int timeout, int retries, int flags)
182 {
183         struct request *req;
184         int write = (data_direction == DMA_TO_DEVICE);
185         int ret = DRIVER_ERROR << 24;
186
187         req = blk_get_request(sdev->request_queue, write, __GFP_WAIT);
188
189         if (bufflen &&  blk_rq_map_kern(sdev->request_queue, req,
190                                         buffer, bufflen, __GFP_WAIT))
191                 goto out;
192
193         req->cmd_len = COMMAND_SIZE(cmd[0]);
194         memcpy(req->cmd, cmd, req->cmd_len);
195         req->sense = sense;
196         req->sense_len = 0;
197         req->retries = retries;
198         req->timeout = timeout;
199         req->cmd_type = REQ_TYPE_BLOCK_PC;
200         req->cmd_flags |= flags | REQ_QUIET | REQ_PREEMPT;
201
202         /*
203          * head injection *required* here otherwise quiesce won't work
204          */
205         blk_execute_rq(req->q, NULL, req, 1);
206
207         ret = req->errors;
208  out:
209         blk_put_request(req);
210
211         return ret;
212 }
213 EXPORT_SYMBOL(scsi_execute);
214
215
216 int scsi_execute_req(struct scsi_device *sdev, const unsigned char *cmd,
217                      int data_direction, void *buffer, unsigned bufflen,
218                      struct scsi_sense_hdr *sshdr, int timeout, int retries)
219 {
220         char *sense = NULL;
221         int result;
222         
223         if (sshdr) {
224                 sense = kzalloc(SCSI_SENSE_BUFFERSIZE, GFP_NOIO);
225                 if (!sense)
226                         return DRIVER_ERROR << 24;
227         }
228         result = scsi_execute(sdev, cmd, data_direction, buffer, bufflen,
229                               sense, timeout, retries, 0);
230         if (sshdr)
231                 scsi_normalize_sense(sense, SCSI_SENSE_BUFFERSIZE, sshdr);
232
233         kfree(sense);
234         return result;
235 }
236 EXPORT_SYMBOL(scsi_execute_req);
237
238 struct scsi_io_context {
239         void *data;
240         void (*done)(void *data, char *sense, int result, int resid);
241         char sense[SCSI_SENSE_BUFFERSIZE];
242 };
243
244 static struct kmem_cache *scsi_io_context_cache;
245
246 static void scsi_end_async(struct request *req, int uptodate)
247 {
248         struct scsi_io_context *sioc = req->end_io_data;
249
250         if (sioc->done)
251                 sioc->done(sioc->data, sioc->sense, req->errors, req->data_len);
252
253         kmem_cache_free(scsi_io_context_cache, sioc);
254         __blk_put_request(req->q, req);
255 }
256
257 static int scsi_merge_bio(struct request *rq, struct bio *bio)
258 {
259         struct request_queue *q = rq->q;
260
261         bio->bi_flags &= ~(1 << BIO_SEG_VALID);
262         if (rq_data_dir(rq) == WRITE)
263                 bio->bi_rw |= (1 << BIO_RW);
264         blk_queue_bounce(q, &bio);
265
266         if (!rq->bio)
267                 blk_rq_bio_prep(q, rq, bio);
268         else if (!ll_back_merge_fn(q, rq, bio))
269                 return -EINVAL;
270         else {
271                 rq->biotail->bi_next = bio;
272                 rq->biotail = bio;
273         }
274
275         return 0;
276 }
277
278 static int scsi_bi_endio(struct bio *bio, unsigned int bytes_done, int error)
279 {
280         if (bio->bi_size)
281                 return 1;
282
283         bio_put(bio);
284         return 0;
285 }
286
287 /**
288  * scsi_req_map_sg - map a scatterlist into a request
289  * @rq:         request to fill
290  * @sg:         scatterlist
291  * @nsegs:      number of elements
292  * @bufflen:    len of buffer
293  * @gfp:        memory allocation flags
294  *
295  * scsi_req_map_sg maps a scatterlist into a request so that the
296  * request can be sent to the block layer. We do not trust the scatterlist
297  * sent to use, as some ULDs use that struct to only organize the pages.
298  */
299 static int scsi_req_map_sg(struct request *rq, struct scatterlist *sgl,
300                            int nsegs, unsigned bufflen, gfp_t gfp)
301 {
302         struct request_queue *q = rq->q;
303         int nr_pages = (bufflen + sgl[0].offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
304         unsigned int data_len = 0, len, bytes, off;
305         struct page *page;
306         struct bio *bio = NULL;
307         int i, err, nr_vecs = 0;
308
309         for (i = 0; i < nsegs; i++) {
310                 page = sgl[i].page;
311                 off = sgl[i].offset;
312                 len = sgl[i].length;
313                 data_len += len;
314
315                 while (len > 0) {
316                         bytes = min_t(unsigned int, len, PAGE_SIZE - off);
317
318                         if (!bio) {
319                                 nr_vecs = min_t(int, BIO_MAX_PAGES, nr_pages);
320                                 nr_pages -= nr_vecs;
321
322                                 bio = bio_alloc(gfp, nr_vecs);
323                                 if (!bio) {
324                                         err = -ENOMEM;
325                                         goto free_bios;
326                                 }
327                                 bio->bi_end_io = scsi_bi_endio;
328                         }
329
330                         if (bio_add_pc_page(q, bio, page, bytes, off) !=
331                             bytes) {
332                                 bio_put(bio);
333                                 err = -EINVAL;
334                                 goto free_bios;
335                         }
336
337                         if (bio->bi_vcnt >= nr_vecs) {
338                                 err = scsi_merge_bio(rq, bio);
339                                 if (err) {
340                                         bio_endio(bio, bio->bi_size, 0);
341                                         goto free_bios;
342                                 }
343                                 bio = NULL;
344                         }
345
346                         page++;
347                         len -= bytes;
348                         off = 0;
349                 }
350         }
351
352         rq->buffer = rq->data = NULL;
353         rq->data_len = data_len;
354         return 0;
355
356 free_bios:
357         while ((bio = rq->bio) != NULL) {
358                 rq->bio = bio->bi_next;
359                 /*
360                  * call endio instead of bio_put incase it was bounced
361                  */
362                 bio_endio(bio, bio->bi_size, 0);
363         }
364
365         return err;
366 }
367
368 /**
369  * scsi_execute_async - insert request
370  * @sdev:       scsi device
371  * @cmd:        scsi command
372  * @cmd_len:    length of scsi cdb
373  * @data_direction: data direction
374  * @buffer:     data buffer (this can be a kernel buffer or scatterlist)
375  * @bufflen:    len of buffer
376  * @use_sg:     if buffer is a scatterlist this is the number of elements
377  * @timeout:    request timeout in seconds
378  * @retries:    number of times to retry request
379  * @flags:      or into request flags
380  **/
381 int scsi_execute_async(struct scsi_device *sdev, const unsigned char *cmd,
382                        int cmd_len, int data_direction, void *buffer, unsigned bufflen,
383                        int use_sg, int timeout, int retries, void *privdata,
384                        void (*done)(void *, char *, int, int), gfp_t gfp)
385 {
386         struct request *req;
387         struct scsi_io_context *sioc;
388         int err = 0;
389         int write = (data_direction == DMA_TO_DEVICE);
390
391         sioc = kmem_cache_zalloc(scsi_io_context_cache, gfp);
392         if (!sioc)
393                 return DRIVER_ERROR << 24;
394
395         req = blk_get_request(sdev->request_queue, write, gfp);
396         if (!req)
397                 goto free_sense;
398         req->cmd_type = REQ_TYPE_BLOCK_PC;
399         req->cmd_flags |= REQ_QUIET;
400
401         if (use_sg)
402                 err = scsi_req_map_sg(req, buffer, use_sg, bufflen, gfp);
403         else if (bufflen)
404                 err = blk_rq_map_kern(req->q, req, buffer, bufflen, gfp);
405
406         if (err)
407                 goto free_req;
408
409         req->cmd_len = cmd_len;
410         memset(req->cmd, 0, BLK_MAX_CDB); /* ATAPI hates garbage after CDB */
411         memcpy(req->cmd, cmd, req->cmd_len);
412         req->sense = sioc->sense;
413         req->sense_len = 0;
414         req->timeout = timeout;
415         req->retries = retries;
416         req->end_io_data = sioc;
417
418         sioc->data = privdata;
419         sioc->done = done;
420
421         blk_execute_rq_nowait(req->q, NULL, req, 1, scsi_end_async);
422         return 0;
423
424 free_req:
425         blk_put_request(req);
426 free_sense:
427         kmem_cache_free(scsi_io_context_cache, sioc);
428         return DRIVER_ERROR << 24;
429 }
430 EXPORT_SYMBOL_GPL(scsi_execute_async);
431
432 /*
433  * Function:    scsi_init_cmd_errh()
434  *
435  * Purpose:     Initialize cmd fields related to error handling.
436  *
437  * Arguments:   cmd     - command that is ready to be queued.
438  *
439  * Notes:       This function has the job of initializing a number of
440  *              fields related to error handling.   Typically this will
441  *              be called once for each command, as required.
442  */
443 static void scsi_init_cmd_errh(struct scsi_cmnd *cmd)
444 {
445         cmd->serial_number = 0;
446         memset(cmd->sense_buffer, 0, sizeof cmd->sense_buffer);
447         if (cmd->cmd_len == 0)
448                 cmd->cmd_len = COMMAND_SIZE(cmd->cmnd[0]);
449 }
450
451 void scsi_device_unbusy(struct scsi_device *sdev)
452 {
453         struct Scsi_Host *shost = sdev->host;
454         unsigned long flags;
455
456         spin_lock_irqsave(shost->host_lock, flags);
457         shost->host_busy--;
458         if (unlikely(scsi_host_in_recovery(shost) &&
459                      (shost->host_failed || shost->host_eh_scheduled)))
460                 scsi_eh_wakeup(shost);
461         spin_unlock(shost->host_lock);
462         spin_lock(sdev->request_queue->queue_lock);
463         sdev->device_busy--;
464         spin_unlock_irqrestore(sdev->request_queue->queue_lock, flags);
465 }
466
467 /*
468  * Called for single_lun devices on IO completion. Clear starget_sdev_user,
469  * and call blk_run_queue for all the scsi_devices on the target -
470  * including current_sdev first.
471  *
472  * Called with *no* scsi locks held.
473  */
474 static void scsi_single_lun_run(struct scsi_device *current_sdev)
475 {
476         struct Scsi_Host *shost = current_sdev->host;
477         struct scsi_device *sdev, *tmp;
478         struct scsi_target *starget = scsi_target(current_sdev);
479         unsigned long flags;
480
481         spin_lock_irqsave(shost->host_lock, flags);
482         starget->starget_sdev_user = NULL;
483         spin_unlock_irqrestore(shost->host_lock, flags);
484
485         /*
486          * Call blk_run_queue for all LUNs on the target, starting with
487          * current_sdev. We race with others (to set starget_sdev_user),
488          * but in most cases, we will be first. Ideally, each LU on the
489          * target would get some limited time or requests on the target.
490          */
491         blk_run_queue(current_sdev->request_queue);
492
493         spin_lock_irqsave(shost->host_lock, flags);
494         if (starget->starget_sdev_user)
495                 goto out;
496         list_for_each_entry_safe(sdev, tmp, &starget->devices,
497                         same_target_siblings) {
498                 if (sdev == current_sdev)
499                         continue;
500                 if (scsi_device_get(sdev))
501                         continue;
502
503                 spin_unlock_irqrestore(shost->host_lock, flags);
504                 blk_run_queue(sdev->request_queue);
505                 spin_lock_irqsave(shost->host_lock, flags);
506         
507                 scsi_device_put(sdev);
508         }
509  out:
510         spin_unlock_irqrestore(shost->host_lock, flags);
511 }
512
513 /*
514  * Function:    scsi_run_queue()
515  *
516  * Purpose:     Select a proper request queue to serve next
517  *
518  * Arguments:   q       - last request's queue
519  *
520  * Returns:     Nothing
521  *
522  * Notes:       The previous command was completely finished, start
523  *              a new one if possible.
524  */
525 static void scsi_run_queue(struct request_queue *q)
526 {
527         struct scsi_device *sdev = q->queuedata;
528         struct Scsi_Host *shost = sdev->host;
529         unsigned long flags;
530
531         if (sdev->single_lun)
532                 scsi_single_lun_run(sdev);
533
534         spin_lock_irqsave(shost->host_lock, flags);
535         while (!list_empty(&shost->starved_list) &&
536                !shost->host_blocked && !shost->host_self_blocked &&
537                 !((shost->can_queue > 0) &&
538                   (shost->host_busy >= shost->can_queue))) {
539                 /*
540                  * As long as shost is accepting commands and we have
541                  * starved queues, call blk_run_queue. scsi_request_fn
542                  * drops the queue_lock and can add us back to the
543                  * starved_list.
544                  *
545                  * host_lock protects the starved_list and starved_entry.
546                  * scsi_request_fn must get the host_lock before checking
547                  * or modifying starved_list or starved_entry.
548                  */
549                 sdev = list_entry(shost->starved_list.next,
550                                           struct scsi_device, starved_entry);
551                 list_del_init(&sdev->starved_entry);
552                 spin_unlock_irqrestore(shost->host_lock, flags);
553
554
555                 if (test_bit(QUEUE_FLAG_REENTER, &q->queue_flags) &&
556                     !test_and_set_bit(QUEUE_FLAG_REENTER,
557                                       &sdev->request_queue->queue_flags)) {
558                         blk_run_queue(sdev->request_queue);
559                         clear_bit(QUEUE_FLAG_REENTER,
560                                   &sdev->request_queue->queue_flags);
561                 } else
562                         blk_run_queue(sdev->request_queue);
563
564                 spin_lock_irqsave(shost->host_lock, flags);
565                 if (unlikely(!list_empty(&sdev->starved_entry)))
566                         /*
567                          * sdev lost a race, and was put back on the
568                          * starved list. This is unlikely but without this
569                          * in theory we could loop forever.
570                          */
571                         break;
572         }
573         spin_unlock_irqrestore(shost->host_lock, flags);
574
575         blk_run_queue(q);
576 }
577
578 /*
579  * Function:    scsi_requeue_command()
580  *
581  * Purpose:     Handle post-processing of completed commands.
582  *
583  * Arguments:   q       - queue to operate on
584  *              cmd     - command that may need to be requeued.
585  *
586  * Returns:     Nothing
587  *
588  * Notes:       After command completion, there may be blocks left
589  *              over which weren't finished by the previous command
590  *              this can be for a number of reasons - the main one is
591  *              I/O errors in the middle of the request, in which case
592  *              we need to request the blocks that come after the bad
593  *              sector.
594  * Notes:       Upon return, cmd is a stale pointer.
595  */
596 static void scsi_requeue_command(struct request_queue *q, struct scsi_cmnd *cmd)
597 {
598         struct request *req = cmd->request;
599         unsigned long flags;
600
601         scsi_unprep_request(req);
602         spin_lock_irqsave(q->queue_lock, flags);
603         blk_requeue_request(q, req);
604         spin_unlock_irqrestore(q->queue_lock, flags);
605
606         scsi_run_queue(q);
607 }
608
609 void scsi_next_command(struct scsi_cmnd *cmd)
610 {
611         struct scsi_device *sdev = cmd->device;
612         struct request_queue *q = sdev->request_queue;
613
614         /* need to hold a reference on the device before we let go of the cmd */
615         get_device(&sdev->sdev_gendev);
616
617         scsi_put_command(cmd);
618         scsi_run_queue(q);
619
620         /* ok to remove device now */
621         put_device(&sdev->sdev_gendev);
622 }
623
624 void scsi_run_host_queues(struct Scsi_Host *shost)
625 {
626         struct scsi_device *sdev;
627
628         shost_for_each_device(sdev, shost)
629                 scsi_run_queue(sdev->request_queue);
630 }
631
632 /*
633  * Function:    scsi_end_request()
634  *
635  * Purpose:     Post-processing of completed commands (usually invoked at end
636  *              of upper level post-processing and scsi_io_completion).
637  *
638  * Arguments:   cmd      - command that is complete.
639  *              uptodate - 1 if I/O indicates success, <= 0 for I/O error.
640  *              bytes    - number of bytes of completed I/O
641  *              requeue  - indicates whether we should requeue leftovers.
642  *
643  * Lock status: Assumed that lock is not held upon entry.
644  *
645  * Returns:     cmd if requeue required, NULL otherwise.
646  *
647  * Notes:       This is called for block device requests in order to
648  *              mark some number of sectors as complete.
649  * 
650  *              We are guaranteeing that the request queue will be goosed
651  *              at some point during this call.
652  * Notes:       If cmd was requeued, upon return it will be a stale pointer.
653  */
654 static struct scsi_cmnd *scsi_end_request(struct scsi_cmnd *cmd, int uptodate,
655                                           int bytes, int requeue)
656 {
657         struct request_queue *q = cmd->device->request_queue;
658         struct request *req = cmd->request;
659         unsigned long flags;
660
661         /*
662          * If there are blocks left over at the end, set up the command
663          * to queue the remainder of them.
664          */
665         if (end_that_request_chunk(req, uptodate, bytes)) {
666                 int leftover = (req->hard_nr_sectors << 9);
667
668                 if (blk_pc_request(req))
669                         leftover = req->data_len;
670
671                 /* kill remainder if no retrys */
672                 if (!uptodate && blk_noretry_request(req))
673                         end_that_request_chunk(req, 0, leftover);
674                 else {
675                         if (requeue) {
676                                 /*
677                                  * Bleah.  Leftovers again.  Stick the
678                                  * leftovers in the front of the
679                                  * queue, and goose the queue again.
680                                  */
681                                 scsi_requeue_command(q, cmd);
682                                 cmd = NULL;
683                         }
684                         return cmd;
685                 }
686         }
687
688         add_disk_randomness(req->rq_disk);
689
690         spin_lock_irqsave(q->queue_lock, flags);
691         if (blk_rq_tagged(req))
692                 blk_queue_end_tag(q, req);
693         end_that_request_last(req, uptodate);
694         spin_unlock_irqrestore(q->queue_lock, flags);
695
696         /*
697          * This will goose the queue request function at the end, so we don't
698          * need to worry about launching another command.
699          */
700         scsi_next_command(cmd);
701         return NULL;
702 }
703
704 struct scatterlist *scsi_alloc_sgtable(struct scsi_cmnd *cmd, gfp_t gfp_mask)
705 {
706         struct scsi_host_sg_pool *sgp;
707         struct scatterlist *sgl;
708
709         BUG_ON(!cmd->use_sg);
710
711         switch (cmd->use_sg) {
712         case 1 ... 8:
713                 cmd->sglist_len = 0;
714                 break;
715         case 9 ... 16:
716                 cmd->sglist_len = 1;
717                 break;
718         case 17 ... 32:
719                 cmd->sglist_len = 2;
720                 break;
721 #if (SCSI_MAX_PHYS_SEGMENTS > 32)
722         case 33 ... 64:
723                 cmd->sglist_len = 3;
724                 break;
725 #if (SCSI_MAX_PHYS_SEGMENTS > 64)
726         case 65 ... 128:
727                 cmd->sglist_len = 4;
728                 break;
729 #if (SCSI_MAX_PHYS_SEGMENTS  > 128)
730         case 129 ... 256:
731                 cmd->sglist_len = 5;
732                 break;
733 #endif
734 #endif
735 #endif
736         default:
737                 return NULL;
738         }
739
740         sgp = scsi_sg_pools + cmd->sglist_len;
741         sgl = mempool_alloc(sgp->pool, gfp_mask);
742         return sgl;
743 }
744
745 EXPORT_SYMBOL(scsi_alloc_sgtable);
746
747 void scsi_free_sgtable(struct scatterlist *sgl, int index)
748 {
749         struct scsi_host_sg_pool *sgp;
750
751         BUG_ON(index >= SG_MEMPOOL_NR);
752
753         sgp = scsi_sg_pools + index;
754         mempool_free(sgl, sgp->pool);
755 }
756
757 EXPORT_SYMBOL(scsi_free_sgtable);
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  */
776 static void scsi_release_buffers(struct scsi_cmnd *cmd)
777 {
778         if (cmd->use_sg)
779                 scsi_free_sgtable(cmd->request_buffer, cmd->sglist_len);
780
781         /*
782          * Zero these out.  They now point to freed memory, and it is
783          * dangerous to hang onto the pointers.
784          */
785         cmd->request_buffer = NULL;
786         cmd->request_bufflen = 0;
787 }
788
789 /*
790  * Function:    scsi_io_completion()
791  *
792  * Purpose:     Completion processing for block device I/O requests.
793  *
794  * Arguments:   cmd   - command that is finished.
795  *
796  * Lock status: Assumed that no lock is held upon entry.
797  *
798  * Returns:     Nothing
799  *
800  * Notes:       This function is matched in terms of capabilities to
801  *              the function that created the scatter-gather list.
802  *              In other words, if there are no bounce buffers
803  *              (the normal case for most drivers), we don't need
804  *              the logic to deal with cleaning up afterwards.
805  *
806  *              We must do one of several things here:
807  *
808  *              a) Call scsi_end_request.  This will finish off the
809  *                 specified number of sectors.  If we are done, the
810  *                 command block will be released, and the queue
811  *                 function will be goosed.  If we are not done, then
812  *                 scsi_end_request will directly goose the queue.
813  *
814  *              b) We can just use scsi_requeue_command() here.  This would
815  *                 be used if we just wanted to retry, for example.
816  */
817 void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes)
818 {
819         int result = cmd->result;
820         int this_count = cmd->request_bufflen;
821         struct request_queue *q = cmd->device->request_queue;
822         struct request *req = cmd->request;
823         int clear_errors = 1;
824         struct scsi_sense_hdr sshdr;
825         int sense_valid = 0;
826         int sense_deferred = 0;
827
828         scsi_release_buffers(cmd);
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 (blk_pc_request(req)) { /* SG_IO ioctl from block level */
837                 req->errors = result;
838                 if (result) {
839                         clear_errors = 0;
840                         if (sense_valid && req->sense) {
841                                 /*
842                                  * SG_IO wants current and deferred errors
843                                  */
844                                 int len = 8 + cmd->sense_buffer[7];
845
846                                 if (len > SCSI_SENSE_BUFFERSIZE)
847                                         len = SCSI_SENSE_BUFFERSIZE;
848                                 memcpy(req->sense, cmd->sense_buffer,  len);
849                                 req->sense_len = len;
850                         }
851                 }
852                 req->data_len = cmd->resid;
853         }
854
855         /*
856          * Next deal with any sectors which we were able to correctly
857          * handle.
858          */
859         SCSI_LOG_HLCOMPLETE(1, printk("%ld sectors total, "
860                                       "%d bytes done.\n",
861                                       req->nr_sectors, good_bytes));
862         SCSI_LOG_HLCOMPLETE(1, printk("use_sg is %d\n", cmd->use_sg));
863
864         if (clear_errors)
865                 req->errors = 0;
866
867         /* A number of bytes were successfully read.  If there
868          * are leftovers and there is some kind of error
869          * (result != 0), retry the rest.
870          */
871         if (scsi_end_request(cmd, 1, good_bytes, result == 0) == NULL)
872                 return;
873
874         /* good_bytes = 0, or (inclusive) there were leftovers and
875          * result = 0, so scsi_end_request couldn't retry.
876          */
877         if (sense_valid && !sense_deferred) {
878                 switch (sshdr.sense_key) {
879                 case UNIT_ATTENTION:
880                         if (cmd->device->removable) {
881                                 /* Detected disc change.  Set a bit
882                                  * and quietly refuse further access.
883                                  */
884                                 cmd->device->changed = 1;
885                                 scsi_end_request(cmd, 0, this_count, 1);
886                                 return;
887                         } else {
888                                 /* Must have been a power glitch, or a
889                                  * bus reset.  Could not have been a
890                                  * media change, so we just retry the
891                                  * request and see what happens.
892                                  */
893                                 scsi_requeue_command(q, cmd);
894                                 return;
895                         }
896                         break;
897                 case ILLEGAL_REQUEST:
898                         /* If we had an ILLEGAL REQUEST returned, then
899                          * we may have performed an unsupported
900                          * command.  The only thing this should be
901                          * would be a ten byte read where only a six
902                          * byte read was supported.  Also, on a system
903                          * where READ CAPACITY failed, we may have
904                          * read past the end of the disk.
905                          */
906                         if ((cmd->device->use_10_for_rw &&
907                             sshdr.asc == 0x20 && sshdr.ascq == 0x00) &&
908                             (cmd->cmnd[0] == READ_10 ||
909                              cmd->cmnd[0] == WRITE_10)) {
910                                 cmd->device->use_10_for_rw = 0;
911                                 /* This will cause a retry with a
912                                  * 6-byte command.
913                                  */
914                                 scsi_requeue_command(q, cmd);
915                                 return;
916                         } else {
917                                 scsi_end_request(cmd, 0, this_count, 1);
918                                 return;
919                         }
920                         break;
921                 case NOT_READY:
922                         /* If the device is in the process of becoming
923                          * ready, or has a temporary blockage, retry.
924                          */
925                         if (sshdr.asc == 0x04) {
926                                 switch (sshdr.ascq) {
927                                 case 0x01: /* becoming ready */
928                                 case 0x04: /* format in progress */
929                                 case 0x05: /* rebuild in progress */
930                                 case 0x06: /* recalculation in progress */
931                                 case 0x07: /* operation in progress */
932                                 case 0x08: /* Long write in progress */
933                                 case 0x09: /* self test in progress */
934                                         scsi_requeue_command(q, cmd);
935                                         return;
936                                 default:
937                                         break;
938                                 }
939                         }
940                         if (!(req->cmd_flags & REQ_QUIET)) {
941                                 scmd_printk(KERN_INFO, cmd,
942                                             "Device not ready: ");
943                                 scsi_print_sense_hdr("", &sshdr);
944                         }
945                         scsi_end_request(cmd, 0, this_count, 1);
946                         return;
947                 case VOLUME_OVERFLOW:
948                         if (!(req->cmd_flags & REQ_QUIET)) {
949                                 scmd_printk(KERN_INFO, cmd,
950                                             "Volume overflow, CDB: ");
951                                 __scsi_print_command(cmd->cmnd);
952                                 scsi_print_sense("", cmd);
953                         }
954                         /* See SSC3rXX or current. */
955                         scsi_end_request(cmd, 0, this_count, 1);
956                         return;
957                 default:
958                         break;
959                 }
960         }
961         if (host_byte(result) == DID_RESET) {
962                 /* Third party bus reset or reset for error recovery
963                  * reasons.  Just retry the request and see what
964                  * happens.
965                  */
966                 scsi_requeue_command(q, cmd);
967                 return;
968         }
969         if (result) {
970                 if (!(req->cmd_flags & REQ_QUIET)) {
971                         scsi_print_result(cmd);
972                         if (driver_byte(result) & DRIVER_SENSE)
973                                 scsi_print_sense("", cmd);
974                 }
975         }
976         scsi_end_request(cmd, 0, this_count, !result);
977 }
978 EXPORT_SYMBOL(scsi_io_completion);
979
980 /*
981  * Function:    scsi_init_io()
982  *
983  * Purpose:     SCSI I/O initialize function.
984  *
985  * Arguments:   cmd   - Command descriptor we wish to initialize
986  *
987  * Returns:     0 on success
988  *              BLKPREP_DEFER if the failure is retryable
989  *              BLKPREP_KILL if the failure is fatal
990  */
991 static int scsi_init_io(struct scsi_cmnd *cmd)
992 {
993         struct request     *req = cmd->request;
994         struct scatterlist *sgpnt;
995         int                count;
996
997         /*
998          * We used to not use scatter-gather for single segment request,
999          * but now we do (it makes highmem I/O easier to support without
1000          * kmapping pages)
1001          */
1002         cmd->use_sg = req->nr_phys_segments;
1003
1004         /*
1005          * If sg table allocation fails, requeue request later.
1006          */
1007         sgpnt = scsi_alloc_sgtable(cmd, GFP_ATOMIC);
1008         if (unlikely(!sgpnt)) {
1009                 scsi_unprep_request(req);
1010                 return BLKPREP_DEFER;
1011         }
1012
1013         req->buffer = NULL;
1014         cmd->request_buffer = (char *) sgpnt;
1015         if (blk_pc_request(req))
1016                 cmd->request_bufflen = req->data_len;
1017         else
1018                 cmd->request_bufflen = req->nr_sectors << 9;
1019
1020         /* 
1021          * Next, walk the list, and fill in the addresses and sizes of
1022          * each segment.
1023          */
1024         count = blk_rq_map_sg(req->q, req, cmd->request_buffer);
1025         if (likely(count <= cmd->use_sg)) {
1026                 cmd->use_sg = count;
1027                 return BLKPREP_OK;
1028         }
1029
1030         printk(KERN_ERR "Incorrect number of segments after building list\n");
1031         printk(KERN_ERR "counted %d, received %d\n", count, cmd->use_sg);
1032         printk(KERN_ERR "req nr_sec %lu, cur_nr_sec %u\n", req->nr_sectors,
1033                         req->current_nr_sectors);
1034
1035         /* release the command and kill it */
1036         scsi_release_buffers(cmd);
1037         scsi_put_command(cmd);
1038         return BLKPREP_KILL;
1039 }
1040
1041 static struct scsi_cmnd *scsi_get_cmd_from_req(struct scsi_device *sdev,
1042                 struct request *req)
1043 {
1044         struct scsi_cmnd *cmd;
1045
1046         if (!req->special) {
1047                 cmd = scsi_get_command(sdev, GFP_ATOMIC);
1048                 if (unlikely(!cmd))
1049                         return NULL;
1050                 req->special = cmd;
1051         } else {
1052                 cmd = req->special;
1053         }
1054
1055         /* pull a tag out of the request if we have one */
1056         cmd->tag = req->tag;
1057         cmd->request = req;
1058
1059         return cmd;
1060 }
1061
1062 static void scsi_blk_pc_done(struct scsi_cmnd *cmd)
1063 {
1064         BUG_ON(!blk_pc_request(cmd->request));
1065         /*
1066          * This will complete the whole command with uptodate=1 so
1067          * as far as the block layer is concerned the command completed
1068          * successfully. Since this is a REQ_BLOCK_PC command the
1069          * caller should check the request's errors value
1070          */
1071         scsi_io_completion(cmd, cmd->request_bufflen);
1072 }
1073
1074 static int scsi_setup_blk_pc_cmnd(struct scsi_device *sdev, struct request *req)
1075 {
1076         struct scsi_cmnd *cmd;
1077
1078         cmd = scsi_get_cmd_from_req(sdev, req);
1079         if (unlikely(!cmd))
1080                 return BLKPREP_DEFER;
1081
1082         /*
1083          * BLOCK_PC requests may transfer data, in which case they must
1084          * a bio attached to them.  Or they might contain a SCSI command
1085          * that does not transfer data, in which case they may optionally
1086          * submit a request without an attached bio.
1087          */
1088         if (req->bio) {
1089                 int ret;
1090
1091                 BUG_ON(!req->nr_phys_segments);
1092
1093                 ret = scsi_init_io(cmd);
1094                 if (unlikely(ret))
1095                         return ret;
1096         } else {
1097                 BUG_ON(req->data_len);
1098                 BUG_ON(req->data);
1099
1100                 cmd->request_bufflen = 0;
1101                 cmd->request_buffer = NULL;
1102                 cmd->use_sg = 0;
1103                 req->buffer = NULL;
1104         }
1105
1106         BUILD_BUG_ON(sizeof(req->cmd) > sizeof(cmd->cmnd));
1107         memcpy(cmd->cmnd, req->cmd, sizeof(cmd->cmnd));
1108         cmd->cmd_len = req->cmd_len;
1109         if (!req->data_len)
1110                 cmd->sc_data_direction = DMA_NONE;
1111         else if (rq_data_dir(req) == WRITE)
1112                 cmd->sc_data_direction = DMA_TO_DEVICE;
1113         else
1114                 cmd->sc_data_direction = DMA_FROM_DEVICE;
1115         
1116         cmd->transfersize = req->data_len;
1117         cmd->allowed = req->retries;
1118         cmd->timeout_per_command = req->timeout;
1119         cmd->done = scsi_blk_pc_done;
1120         return BLKPREP_OK;
1121 }
1122
1123 /*
1124  * Setup a REQ_TYPE_FS command.  These are simple read/write request
1125  * from filesystems that still need to be translated to SCSI CDBs from
1126  * the ULD.
1127  */
1128 static int scsi_setup_fs_cmnd(struct scsi_device *sdev, struct request *req)
1129 {
1130         struct scsi_cmnd *cmd;
1131         struct scsi_driver *drv;
1132         int ret;
1133
1134         /*
1135          * Filesystem requests must transfer data.
1136          */
1137         BUG_ON(!req->nr_phys_segments);
1138
1139         cmd = scsi_get_cmd_from_req(sdev, req);
1140         if (unlikely(!cmd))
1141                 return BLKPREP_DEFER;
1142
1143         ret = scsi_init_io(cmd);
1144         if (unlikely(ret))
1145                 return ret;
1146
1147         /*
1148          * Initialize the actual SCSI command for this request.
1149          */
1150         drv = *(struct scsi_driver **)req->rq_disk->private_data;
1151         if (unlikely(!drv->init_command(cmd))) {
1152                 scsi_release_buffers(cmd);
1153                 scsi_put_command(cmd);
1154                 return BLKPREP_KILL;
1155         }
1156
1157         return BLKPREP_OK;
1158 }
1159
1160 static int scsi_prep_fn(struct request_queue *q, struct request *req)
1161 {
1162         struct scsi_device *sdev = q->queuedata;
1163         int ret = BLKPREP_OK;
1164
1165         /*
1166          * If the device is not in running state we will reject some
1167          * or all commands.
1168          */
1169         if (unlikely(sdev->sdev_state != SDEV_RUNNING)) {
1170                 switch (sdev->sdev_state) {
1171                 case SDEV_OFFLINE:
1172                         /*
1173                          * If the device is offline we refuse to process any
1174                          * commands.  The device must be brought online
1175                          * before trying any recovery commands.
1176                          */
1177                         sdev_printk(KERN_ERR, sdev,
1178                                     "rejecting I/O to offline device\n");
1179                         ret = BLKPREP_KILL;
1180                         break;
1181                 case SDEV_DEL:
1182                         /*
1183                          * If the device is fully deleted, we refuse to
1184                          * process any commands as well.
1185                          */
1186                         sdev_printk(KERN_ERR, sdev,
1187                                     "rejecting I/O to dead device\n");
1188                         ret = BLKPREP_KILL;
1189                         break;
1190                 case SDEV_QUIESCE:
1191                 case SDEV_BLOCK:
1192                         /*
1193                          * If the devices is blocked we defer normal commands.
1194                          */
1195                         if (!(req->cmd_flags & REQ_PREEMPT))
1196                                 ret = BLKPREP_DEFER;
1197                         break;
1198                 default:
1199                         /*
1200                          * For any other not fully online state we only allow
1201                          * special commands.  In particular any user initiated
1202                          * command is not allowed.
1203                          */
1204                         if (!(req->cmd_flags & REQ_PREEMPT))
1205                                 ret = BLKPREP_KILL;
1206                         break;
1207                 }
1208
1209                 if (ret != BLKPREP_OK)
1210                         goto out;
1211         }
1212
1213         switch (req->cmd_type) {
1214         case REQ_TYPE_BLOCK_PC:
1215                 ret = scsi_setup_blk_pc_cmnd(sdev, req);
1216                 break;
1217         case REQ_TYPE_FS:
1218                 ret = scsi_setup_fs_cmnd(sdev, req);
1219                 break;
1220         default:
1221                 /*
1222                  * All other command types are not supported.
1223                  *
1224                  * Note that these days the SCSI subsystem does not use
1225                  * REQ_TYPE_SPECIAL requests anymore.  These are only used
1226                  * (directly or via blk_insert_request) by non-SCSI drivers.
1227                  */
1228                 blk_dump_rq_flags(req, "SCSI bad req");
1229                 ret = BLKPREP_KILL;
1230                 break;
1231         }
1232
1233  out:
1234         switch (ret) {
1235         case BLKPREP_KILL:
1236                 req->errors = DID_NO_CONNECT << 16;
1237                 break;
1238         case BLKPREP_DEFER:
1239                 /*
1240                  * If we defer, the elv_next_request() returns NULL, but the
1241                  * queue must be restarted, so we plug here if no returning
1242                  * command will automatically do that.
1243                  */
1244                 if (sdev->device_busy == 0)
1245                         blk_plug_device(q);
1246                 break;
1247         default:
1248                 req->cmd_flags |= REQ_DONTPREP;
1249         }
1250
1251         return ret;
1252 }
1253
1254 /*
1255  * scsi_dev_queue_ready: if we can send requests to sdev, return 1 else
1256  * return 0.
1257  *
1258  * Called with the queue_lock held.
1259  */
1260 static inline int scsi_dev_queue_ready(struct request_queue *q,
1261                                   struct scsi_device *sdev)
1262 {
1263         if (sdev->device_busy >= sdev->queue_depth)
1264                 return 0;
1265         if (sdev->device_busy == 0 && sdev->device_blocked) {
1266                 /*
1267                  * unblock after device_blocked iterates to zero
1268                  */
1269                 if (--sdev->device_blocked == 0) {
1270                         SCSI_LOG_MLQUEUE(3,
1271                                    sdev_printk(KERN_INFO, sdev,
1272                                    "unblocking device at zero depth\n"));
1273                 } else {
1274                         blk_plug_device(q);
1275                         return 0;
1276                 }
1277         }
1278         if (sdev->device_blocked)
1279                 return 0;
1280
1281         return 1;
1282 }
1283
1284 /*
1285  * scsi_host_queue_ready: if we can send requests to shost, return 1 else
1286  * return 0. We must end up running the queue again whenever 0 is
1287  * returned, else IO can hang.
1288  *
1289  * Called with host_lock held.
1290  */
1291 static inline int scsi_host_queue_ready(struct request_queue *q,
1292                                    struct Scsi_Host *shost,
1293                                    struct scsi_device *sdev)
1294 {
1295         if (scsi_host_in_recovery(shost))
1296                 return 0;
1297         if (shost->host_busy == 0 && shost->host_blocked) {
1298                 /*
1299                  * unblock after host_blocked iterates to zero
1300                  */
1301                 if (--shost->host_blocked == 0) {
1302                         SCSI_LOG_MLQUEUE(3,
1303                                 printk("scsi%d unblocking host at zero depth\n",
1304                                         shost->host_no));
1305                 } else {
1306                         blk_plug_device(q);
1307                         return 0;
1308                 }
1309         }
1310         if ((shost->can_queue > 0 && shost->host_busy >= shost->can_queue) ||
1311             shost->host_blocked || shost->host_self_blocked) {
1312                 if (list_empty(&sdev->starved_entry))
1313                         list_add_tail(&sdev->starved_entry, &shost->starved_list);
1314                 return 0;
1315         }
1316
1317         /* We're OK to process the command, so we can't be starved */
1318         if (!list_empty(&sdev->starved_entry))
1319                 list_del_init(&sdev->starved_entry);
1320
1321         return 1;
1322 }
1323
1324 /*
1325  * Kill a request for a dead device
1326  */
1327 static void scsi_kill_request(struct request *req, struct request_queue *q)
1328 {
1329         struct scsi_cmnd *cmd = req->special;
1330         struct scsi_device *sdev = cmd->device;
1331         struct Scsi_Host *shost = sdev->host;
1332
1333         blkdev_dequeue_request(req);
1334
1335         if (unlikely(cmd == NULL)) {
1336                 printk(KERN_CRIT "impossible request in %s.\n",
1337                                  __FUNCTION__);
1338                 BUG();
1339         }
1340
1341         scsi_init_cmd_errh(cmd);
1342         cmd->result = DID_NO_CONNECT << 16;
1343         atomic_inc(&cmd->device->iorequest_cnt);
1344
1345         /*
1346          * SCSI request completion path will do scsi_device_unbusy(),
1347          * bump busy counts.  To bump the counters, we need to dance
1348          * with the locks as normal issue path does.
1349          */
1350         sdev->device_busy++;
1351         spin_unlock(sdev->request_queue->queue_lock);
1352         spin_lock(shost->host_lock);
1353         shost->host_busy++;
1354         spin_unlock(shost->host_lock);
1355         spin_lock(sdev->request_queue->queue_lock);
1356
1357         __scsi_done(cmd);
1358 }
1359
1360 static void scsi_softirq_done(struct request *rq)
1361 {
1362         struct scsi_cmnd *cmd = rq->completion_data;
1363         unsigned long wait_for = (cmd->allowed + 1) * cmd->timeout_per_command;
1364         int disposition;
1365
1366         INIT_LIST_HEAD(&cmd->eh_entry);
1367
1368         disposition = scsi_decide_disposition(cmd);
1369         if (disposition != SUCCESS &&
1370             time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) {
1371                 sdev_printk(KERN_ERR, cmd->device,
1372                             "timing out command, waited %lus\n",
1373                             wait_for/HZ);
1374                 disposition = SUCCESS;
1375         }
1376                         
1377         scsi_log_completion(cmd, disposition);
1378
1379         switch (disposition) {
1380                 case SUCCESS:
1381                         scsi_finish_command(cmd);
1382                         break;
1383                 case NEEDS_RETRY:
1384                         scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY);
1385                         break;
1386                 case ADD_TO_MLQUEUE:
1387                         scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY);
1388                         break;
1389                 default:
1390                         if (!scsi_eh_scmd_add(cmd, 0))
1391                                 scsi_finish_command(cmd);
1392         }
1393 }
1394
1395 /*
1396  * Function:    scsi_request_fn()
1397  *
1398  * Purpose:     Main strategy routine for SCSI.
1399  *
1400  * Arguments:   q       - Pointer to actual queue.
1401  *
1402  * Returns:     Nothing
1403  *
1404  * Lock status: IO request lock assumed to be held when called.
1405  */
1406 static void scsi_request_fn(struct request_queue *q)
1407 {
1408         struct scsi_device *sdev = q->queuedata;
1409         struct Scsi_Host *shost;
1410         struct scsi_cmnd *cmd;
1411         struct request *req;
1412
1413         if (!sdev) {
1414                 printk("scsi: killing requests for dead queue\n");
1415                 while ((req = elv_next_request(q)) != NULL)
1416                         scsi_kill_request(req, q);
1417                 return;
1418         }
1419
1420         if(!get_device(&sdev->sdev_gendev))
1421                 /* We must be tearing the block queue down already */
1422                 return;
1423
1424         /*
1425          * To start with, we keep looping until the queue is empty, or until
1426          * the host is no longer able to accept any more requests.
1427          */
1428         shost = sdev->host;
1429         while (!blk_queue_plugged(q)) {
1430                 int rtn;
1431                 /*
1432                  * get next queueable request.  We do this early to make sure
1433                  * that the request is fully prepared even if we cannot 
1434                  * accept it.
1435                  */
1436                 req = elv_next_request(q);
1437                 if (!req || !scsi_dev_queue_ready(q, sdev))
1438                         break;
1439
1440                 if (unlikely(!scsi_device_online(sdev))) {
1441                         sdev_printk(KERN_ERR, sdev,
1442                                     "rejecting I/O to offline device\n");
1443                         scsi_kill_request(req, q);
1444                         continue;
1445                 }
1446
1447
1448                 /*
1449                  * Remove the request from the request list.
1450                  */
1451                 if (!(blk_queue_tagged(q) && !blk_queue_start_tag(q, req)))
1452                         blkdev_dequeue_request(req);
1453                 sdev->device_busy++;
1454
1455                 spin_unlock(q->queue_lock);
1456                 cmd = req->special;
1457                 if (unlikely(cmd == NULL)) {
1458                         printk(KERN_CRIT "impossible request in %s.\n"
1459                                          "please mail a stack trace to "
1460                                          "linux-scsi@vger.kernel.org\n",
1461                                          __FUNCTION__);
1462                         blk_dump_rq_flags(req, "foo");
1463                         BUG();
1464                 }
1465                 spin_lock(shost->host_lock);
1466
1467                 if (!scsi_host_queue_ready(q, shost, sdev))
1468                         goto not_ready;
1469                 if (sdev->single_lun) {
1470                         if (scsi_target(sdev)->starget_sdev_user &&
1471                             scsi_target(sdev)->starget_sdev_user != sdev)
1472                                 goto not_ready;
1473                         scsi_target(sdev)->starget_sdev_user = sdev;
1474                 }
1475                 shost->host_busy++;
1476
1477                 /*
1478                  * XXX(hch): This is rather suboptimal, scsi_dispatch_cmd will
1479                  *              take the lock again.
1480                  */
1481                 spin_unlock_irq(shost->host_lock);
1482
1483                 /*
1484                  * Finally, initialize any error handling parameters, and set up
1485                  * the timers for timeouts.
1486                  */
1487                 scsi_init_cmd_errh(cmd);
1488
1489                 /*
1490                  * Dispatch the command to the low-level driver.
1491                  */
1492                 rtn = scsi_dispatch_cmd(cmd);
1493                 spin_lock_irq(q->queue_lock);
1494                 if(rtn) {
1495                         /* we're refusing the command; because of
1496                          * the way locks get dropped, we need to 
1497                          * check here if plugging is required */
1498                         if(sdev->device_busy == 0)
1499                                 blk_plug_device(q);
1500
1501                         break;
1502                 }
1503         }
1504
1505         goto out;
1506
1507  not_ready:
1508         spin_unlock_irq(shost->host_lock);
1509
1510         /*
1511          * lock q, handle tag, requeue req, and decrement device_busy. We
1512          * must return with queue_lock held.
1513          *
1514          * Decrementing device_busy without checking it is OK, as all such
1515          * cases (host limits or settings) should run the queue at some
1516          * later time.
1517          */
1518         spin_lock_irq(q->queue_lock);
1519         blk_requeue_request(q, req);
1520         sdev->device_busy--;
1521         if(sdev->device_busy == 0)
1522                 blk_plug_device(q);
1523  out:
1524         /* must be careful here...if we trigger the ->remove() function
1525          * we cannot be holding the q lock */
1526         spin_unlock_irq(q->queue_lock);
1527         put_device(&sdev->sdev_gendev);
1528         spin_lock_irq(q->queue_lock);
1529 }
1530
1531 u64 scsi_calculate_bounce_limit(struct Scsi_Host *shost)
1532 {
1533         struct device *host_dev;
1534         u64 bounce_limit = 0xffffffff;
1535
1536         if (shost->unchecked_isa_dma)
1537                 return BLK_BOUNCE_ISA;
1538         /*
1539          * Platforms with virtual-DMA translation
1540          * hardware have no practical limit.
1541          */
1542         if (!PCI_DMA_BUS_IS_PHYS)
1543                 return BLK_BOUNCE_ANY;
1544
1545         host_dev = scsi_get_device(shost);
1546         if (host_dev && host_dev->dma_mask)
1547                 bounce_limit = *host_dev->dma_mask;
1548
1549         return bounce_limit;
1550 }
1551 EXPORT_SYMBOL(scsi_calculate_bounce_limit);
1552
1553 struct request_queue *__scsi_alloc_queue(struct Scsi_Host *shost,
1554                                          request_fn_proc *request_fn)
1555 {
1556         struct request_queue *q;
1557
1558         q = blk_init_queue(request_fn, NULL);
1559         if (!q)
1560                 return NULL;
1561
1562         blk_queue_max_hw_segments(q, shost->sg_tablesize);
1563         blk_queue_max_phys_segments(q, SCSI_MAX_PHYS_SEGMENTS);
1564         blk_queue_max_sectors(q, shost->max_sectors);
1565         blk_queue_bounce_limit(q, scsi_calculate_bounce_limit(shost));
1566         blk_queue_segment_boundary(q, shost->dma_boundary);
1567
1568         if (!shost->use_clustering)
1569                 clear_bit(QUEUE_FLAG_CLUSTER, &q->queue_flags);
1570         return q;
1571 }
1572 EXPORT_SYMBOL(__scsi_alloc_queue);
1573
1574 struct request_queue *scsi_alloc_queue(struct scsi_device *sdev)
1575 {
1576         struct request_queue *q;
1577
1578         q = __scsi_alloc_queue(sdev->host, scsi_request_fn);
1579         if (!q)
1580                 return NULL;
1581
1582         blk_queue_prep_rq(q, scsi_prep_fn);
1583         blk_queue_softirq_done(q, scsi_softirq_done);
1584         return q;
1585 }
1586
1587 void scsi_free_queue(struct request_queue *q)
1588 {
1589         blk_cleanup_queue(q);
1590 }
1591
1592 /*
1593  * Function:    scsi_block_requests()
1594  *
1595  * Purpose:     Utility function used by low-level drivers to prevent further
1596  *              commands from being queued to the device.
1597  *
1598  * Arguments:   shost       - Host in question
1599  *
1600  * Returns:     Nothing
1601  *
1602  * Lock status: No locks are assumed held.
1603  *
1604  * Notes:       There is no timer nor any other means by which the requests
1605  *              get unblocked other than the low-level driver calling
1606  *              scsi_unblock_requests().
1607  */
1608 void scsi_block_requests(struct Scsi_Host *shost)
1609 {
1610         shost->host_self_blocked = 1;
1611 }
1612 EXPORT_SYMBOL(scsi_block_requests);
1613
1614 /*
1615  * Function:    scsi_unblock_requests()
1616  *
1617  * Purpose:     Utility function used by low-level drivers to allow further
1618  *              commands from being queued to the device.
1619  *
1620  * Arguments:   shost       - Host in question
1621  *
1622  * Returns:     Nothing
1623  *
1624  * Lock status: No locks are assumed held.
1625  *
1626  * Notes:       There is no timer nor any other means by which the requests
1627  *              get unblocked other than the low-level driver calling
1628  *              scsi_unblock_requests().
1629  *
1630  *              This is done as an API function so that changes to the
1631  *              internals of the scsi mid-layer won't require wholesale
1632  *              changes to drivers that use this feature.
1633  */
1634 void scsi_unblock_requests(struct Scsi_Host *shost)
1635 {
1636         shost->host_self_blocked = 0;
1637         scsi_run_host_queues(shost);
1638 }
1639 EXPORT_SYMBOL(scsi_unblock_requests);
1640
1641 int __init scsi_init_queue(void)
1642 {
1643         int i;
1644
1645         scsi_io_context_cache = kmem_cache_create("scsi_io_context",
1646                                         sizeof(struct scsi_io_context),
1647                                         0, 0, NULL);
1648         if (!scsi_io_context_cache) {
1649                 printk(KERN_ERR "SCSI: can't init scsi io context cache\n");
1650                 return -ENOMEM;
1651         }
1652
1653         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1654                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1655                 int size = sgp->size * sizeof(struct scatterlist);
1656
1657                 sgp->slab = kmem_cache_create(sgp->name, size, 0,
1658                                 SLAB_HWCACHE_ALIGN, NULL);
1659                 if (!sgp->slab) {
1660                         printk(KERN_ERR "SCSI: can't init sg slab %s\n",
1661                                         sgp->name);
1662                 }
1663
1664                 sgp->pool = mempool_create_slab_pool(SG_MEMPOOL_SIZE,
1665                                                      sgp->slab);
1666                 if (!sgp->pool) {
1667                         printk(KERN_ERR "SCSI: can't init sg mempool %s\n",
1668                                         sgp->name);
1669                 }
1670         }
1671
1672         return 0;
1673 }
1674
1675 void scsi_exit_queue(void)
1676 {
1677         int i;
1678
1679         kmem_cache_destroy(scsi_io_context_cache);
1680
1681         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1682                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1683                 mempool_destroy(sgp->pool);
1684                 kmem_cache_destroy(sgp->slab);
1685         }
1686 }
1687
1688 /**
1689  *      scsi_mode_select - issue a mode select
1690  *      @sdev:  SCSI device to be queried
1691  *      @pf:    Page format bit (1 == standard, 0 == vendor specific)
1692  *      @sp:    Save page bit (0 == don't save, 1 == save)
1693  *      @modepage: mode page being requested
1694  *      @buffer: request buffer (may not be smaller than eight bytes)
1695  *      @len:   length of request buffer.
1696  *      @timeout: command timeout
1697  *      @retries: number of retries before failing
1698  *      @data: returns a structure abstracting the mode header data
1699  *      @sense: place to put sense data (or NULL if no sense to be collected).
1700  *              must be SCSI_SENSE_BUFFERSIZE big.
1701  *
1702  *      Returns zero if successful; negative error number or scsi
1703  *      status on error
1704  *
1705  */
1706 int
1707 scsi_mode_select(struct scsi_device *sdev, int pf, int sp, int modepage,
1708                  unsigned char *buffer, int len, int timeout, int retries,
1709                  struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1710 {
1711         unsigned char cmd[10];
1712         unsigned char *real_buffer;
1713         int ret;
1714
1715         memset(cmd, 0, sizeof(cmd));
1716         cmd[1] = (pf ? 0x10 : 0) | (sp ? 0x01 : 0);
1717
1718         if (sdev->use_10_for_ms) {
1719                 if (len > 65535)
1720                         return -EINVAL;
1721                 real_buffer = kmalloc(8 + len, GFP_KERNEL);
1722                 if (!real_buffer)
1723                         return -ENOMEM;
1724                 memcpy(real_buffer + 8, buffer, len);
1725                 len += 8;
1726                 real_buffer[0] = 0;
1727                 real_buffer[1] = 0;
1728                 real_buffer[2] = data->medium_type;
1729                 real_buffer[3] = data->device_specific;
1730                 real_buffer[4] = data->longlba ? 0x01 : 0;
1731                 real_buffer[5] = 0;
1732                 real_buffer[6] = data->block_descriptor_length >> 8;
1733                 real_buffer[7] = data->block_descriptor_length;
1734
1735                 cmd[0] = MODE_SELECT_10;
1736                 cmd[7] = len >> 8;
1737                 cmd[8] = len;
1738         } else {
1739                 if (len > 255 || data->block_descriptor_length > 255 ||
1740                     data->longlba)
1741                         return -EINVAL;
1742
1743                 real_buffer = kmalloc(4 + len, GFP_KERNEL);
1744                 if (!real_buffer)
1745                         return -ENOMEM;
1746                 memcpy(real_buffer + 4, buffer, len);
1747                 len += 4;
1748                 real_buffer[0] = 0;
1749                 real_buffer[1] = data->medium_type;
1750                 real_buffer[2] = data->device_specific;
1751                 real_buffer[3] = data->block_descriptor_length;
1752                 
1753
1754                 cmd[0] = MODE_SELECT;
1755                 cmd[4] = len;
1756         }
1757
1758         ret = scsi_execute_req(sdev, cmd, DMA_TO_DEVICE, real_buffer, len,
1759                                sshdr, timeout, retries);
1760         kfree(real_buffer);
1761         return ret;
1762 }
1763 EXPORT_SYMBOL_GPL(scsi_mode_select);
1764
1765 /**
1766  *      scsi_mode_sense - issue a mode sense, falling back from 10 to 
1767  *              six bytes if necessary.
1768  *      @sdev:  SCSI device to be queried
1769  *      @dbd:   set if mode sense will allow block descriptors to be returned
1770  *      @modepage: mode page being requested
1771  *      @buffer: request buffer (may not be smaller than eight bytes)
1772  *      @len:   length of request buffer.
1773  *      @timeout: command timeout
1774  *      @retries: number of retries before failing
1775  *      @data: returns a structure abstracting the mode header data
1776  *      @sense: place to put sense data (or NULL if no sense to be collected).
1777  *              must be SCSI_SENSE_BUFFERSIZE big.
1778  *
1779  *      Returns zero if unsuccessful, or the header offset (either 4
1780  *      or 8 depending on whether a six or ten byte command was
1781  *      issued) if successful.
1782  **/
1783 int
1784 scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
1785                   unsigned char *buffer, int len, int timeout, int retries,
1786                   struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1787 {
1788         unsigned char cmd[12];
1789         int use_10_for_ms;
1790         int header_length;
1791         int result;
1792         struct scsi_sense_hdr my_sshdr;
1793
1794         memset(data, 0, sizeof(*data));
1795         memset(&cmd[0], 0, 12);
1796         cmd[1] = dbd & 0x18;    /* allows DBD and LLBA bits */
1797         cmd[2] = modepage;
1798
1799         /* caller might not be interested in sense, but we need it */
1800         if (!sshdr)
1801                 sshdr = &my_sshdr;
1802
1803  retry:
1804         use_10_for_ms = sdev->use_10_for_ms;
1805
1806         if (use_10_for_ms) {
1807                 if (len < 8)
1808                         len = 8;
1809
1810                 cmd[0] = MODE_SENSE_10;
1811                 cmd[8] = len;
1812                 header_length = 8;
1813         } else {
1814                 if (len < 4)
1815                         len = 4;
1816
1817                 cmd[0] = MODE_SENSE;
1818                 cmd[4] = len;
1819                 header_length = 4;
1820         }
1821
1822         memset(buffer, 0, len);
1823
1824         result = scsi_execute_req(sdev, cmd, DMA_FROM_DEVICE, buffer, len,
1825                                   sshdr, timeout, retries);
1826
1827         /* This code looks awful: what it's doing is making sure an
1828          * ILLEGAL REQUEST sense return identifies the actual command
1829          * byte as the problem.  MODE_SENSE commands can return
1830          * ILLEGAL REQUEST if the code page isn't supported */
1831
1832         if (use_10_for_ms && !scsi_status_is_good(result) &&
1833             (driver_byte(result) & DRIVER_SENSE)) {
1834                 if (scsi_sense_valid(sshdr)) {
1835                         if ((sshdr->sense_key == ILLEGAL_REQUEST) &&
1836                             (sshdr->asc == 0x20) && (sshdr->ascq == 0)) {
1837                                 /* 
1838                                  * Invalid command operation code
1839                                  */
1840                                 sdev->use_10_for_ms = 0;
1841                                 goto retry;
1842                         }
1843                 }
1844         }
1845
1846         if(scsi_status_is_good(result)) {
1847                 if (unlikely(buffer[0] == 0x86 && buffer[1] == 0x0b &&
1848                              (modepage == 6 || modepage == 8))) {
1849                         /* Initio breakage? */
1850                         header_length = 0;
1851                         data->length = 13;
1852                         data->medium_type = 0;
1853                         data->device_specific = 0;
1854                         data->longlba = 0;
1855                         data->block_descriptor_length = 0;
1856                 } else if(use_10_for_ms) {
1857                         data->length = buffer[0]*256 + buffer[1] + 2;
1858                         data->medium_type = buffer[2];
1859                         data->device_specific = buffer[3];
1860                         data->longlba = buffer[4] & 0x01;
1861                         data->block_descriptor_length = buffer[6]*256
1862                                 + buffer[7];
1863                 } else {
1864                         data->length = buffer[0] + 1;
1865                         data->medium_type = buffer[1];
1866                         data->device_specific = buffer[2];
1867                         data->block_descriptor_length = buffer[3];
1868                 }
1869                 data->header_length = header_length;
1870         }
1871
1872         return result;
1873 }
1874 EXPORT_SYMBOL(scsi_mode_sense);
1875
1876 int
1877 scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries)
1878 {
1879         char cmd[] = {
1880                 TEST_UNIT_READY, 0, 0, 0, 0, 0,
1881         };
1882         struct scsi_sense_hdr sshdr;
1883         int result;
1884         
1885         result = scsi_execute_req(sdev, cmd, DMA_NONE, NULL, 0, &sshdr,
1886                                   timeout, retries);
1887
1888         if ((driver_byte(result) & DRIVER_SENSE) && sdev->removable) {
1889
1890                 if ((scsi_sense_valid(&sshdr)) &&
1891                     ((sshdr.sense_key == UNIT_ATTENTION) ||
1892                      (sshdr.sense_key == NOT_READY))) {
1893                         sdev->changed = 1;
1894                         result = 0;
1895                 }
1896         }
1897         return result;
1898 }
1899 EXPORT_SYMBOL(scsi_test_unit_ready);
1900
1901 /**
1902  *      scsi_device_set_state - Take the given device through the device
1903  *              state model.
1904  *      @sdev:  scsi device to change the state of.
1905  *      @state: state to change to.
1906  *
1907  *      Returns zero if unsuccessful or an error if the requested 
1908  *      transition is illegal.
1909  **/
1910 int
1911 scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
1912 {
1913         enum scsi_device_state oldstate = sdev->sdev_state;
1914
1915         if (state == oldstate)
1916                 return 0;
1917
1918         switch (state) {
1919         case SDEV_CREATED:
1920                 /* There are no legal states that come back to
1921                  * created.  This is the manually initialised start
1922                  * state */
1923                 goto illegal;
1924                         
1925         case SDEV_RUNNING:
1926                 switch (oldstate) {
1927                 case SDEV_CREATED:
1928                 case SDEV_OFFLINE:
1929                 case SDEV_QUIESCE:
1930                 case SDEV_BLOCK:
1931                         break;
1932                 default:
1933                         goto illegal;
1934                 }
1935                 break;
1936
1937         case SDEV_QUIESCE:
1938                 switch (oldstate) {
1939                 case SDEV_RUNNING:
1940                 case SDEV_OFFLINE:
1941                         break;
1942                 default:
1943                         goto illegal;
1944                 }
1945                 break;
1946
1947         case SDEV_OFFLINE:
1948                 switch (oldstate) {
1949                 case SDEV_CREATED:
1950                 case SDEV_RUNNING:
1951                 case SDEV_QUIESCE:
1952                 case SDEV_BLOCK:
1953                         break;
1954                 default:
1955                         goto illegal;
1956                 }
1957                 break;
1958
1959         case SDEV_BLOCK:
1960                 switch (oldstate) {
1961                 case SDEV_CREATED:
1962                 case SDEV_RUNNING:
1963                         break;
1964                 default:
1965                         goto illegal;
1966                 }
1967                 break;
1968
1969         case SDEV_CANCEL:
1970                 switch (oldstate) {
1971                 case SDEV_CREATED:
1972                 case SDEV_RUNNING:
1973                 case SDEV_QUIESCE:
1974                 case SDEV_OFFLINE:
1975                 case SDEV_BLOCK:
1976                         break;
1977                 default:
1978                         goto illegal;
1979                 }
1980                 break;
1981
1982         case SDEV_DEL:
1983                 switch (oldstate) {
1984                 case SDEV_CREATED:
1985                 case SDEV_RUNNING:
1986                 case SDEV_OFFLINE:
1987                 case SDEV_CANCEL:
1988                         break;
1989                 default:
1990                         goto illegal;
1991                 }
1992                 break;
1993
1994         }
1995         sdev->sdev_state = state;
1996         return 0;
1997
1998  illegal:
1999         SCSI_LOG_ERROR_RECOVERY(1, 
2000                                 sdev_printk(KERN_ERR, sdev,
2001                                             "Illegal state transition %s->%s\n",
2002                                             scsi_device_state_name(oldstate),
2003                                             scsi_device_state_name(state))
2004                                 );
2005         return -EINVAL;
2006 }
2007 EXPORT_SYMBOL(scsi_device_set_state);
2008
2009 /**
2010  *      scsi_device_quiesce - Block user issued commands.
2011  *      @sdev:  scsi device to quiesce.
2012  *
2013  *      This works by trying to transition to the SDEV_QUIESCE state
2014  *      (which must be a legal transition).  When the device is in this
2015  *      state, only special requests will be accepted, all others will
2016  *      be deferred.  Since special requests may also be requeued requests,
2017  *      a successful return doesn't guarantee the device will be 
2018  *      totally quiescent.
2019  *
2020  *      Must be called with user context, may sleep.
2021  *
2022  *      Returns zero if unsuccessful or an error if not.
2023  **/
2024 int
2025 scsi_device_quiesce(struct scsi_device *sdev)
2026 {
2027         int err = scsi_device_set_state(sdev, SDEV_QUIESCE);
2028         if (err)
2029                 return err;
2030
2031         scsi_run_queue(sdev->request_queue);
2032         while (sdev->device_busy) {
2033                 msleep_interruptible(200);
2034                 scsi_run_queue(sdev->request_queue);
2035         }
2036         return 0;
2037 }
2038 EXPORT_SYMBOL(scsi_device_quiesce);
2039
2040 /**
2041  *      scsi_device_resume - Restart user issued commands to a quiesced device.
2042  *      @sdev:  scsi device to resume.
2043  *
2044  *      Moves the device from quiesced back to running and restarts the
2045  *      queues.
2046  *
2047  *      Must be called with user context, may sleep.
2048  **/
2049 void
2050 scsi_device_resume(struct scsi_device *sdev)
2051 {
2052         if(scsi_device_set_state(sdev, SDEV_RUNNING))
2053                 return;
2054         scsi_run_queue(sdev->request_queue);
2055 }
2056 EXPORT_SYMBOL(scsi_device_resume);
2057
2058 static void
2059 device_quiesce_fn(struct scsi_device *sdev, void *data)
2060 {
2061         scsi_device_quiesce(sdev);
2062 }
2063
2064 void
2065 scsi_target_quiesce(struct scsi_target *starget)
2066 {
2067         starget_for_each_device(starget, NULL, device_quiesce_fn);
2068 }
2069 EXPORT_SYMBOL(scsi_target_quiesce);
2070
2071 static void
2072 device_resume_fn(struct scsi_device *sdev, void *data)
2073 {
2074         scsi_device_resume(sdev);
2075 }
2076
2077 void
2078 scsi_target_resume(struct scsi_target *starget)
2079 {
2080         starget_for_each_device(starget, NULL, device_resume_fn);
2081 }
2082 EXPORT_SYMBOL(scsi_target_resume);
2083
2084 /**
2085  * scsi_internal_device_block - internal function to put a device
2086  *                              temporarily into the SDEV_BLOCK state
2087  * @sdev:       device to block
2088  *
2089  * Block request made by scsi lld's to temporarily stop all
2090  * scsi commands on the specified device.  Called from interrupt
2091  * or normal process context.
2092  *
2093  * Returns zero if successful or error if not
2094  *
2095  * Notes:       
2096  *      This routine transitions the device to the SDEV_BLOCK state
2097  *      (which must be a legal transition).  When the device is in this
2098  *      state, all commands are deferred until the scsi lld reenables
2099  *      the device with scsi_device_unblock or device_block_tmo fires.
2100  *      This routine assumes the host_lock is held on entry.
2101  **/
2102 int
2103 scsi_internal_device_block(struct scsi_device *sdev)
2104 {
2105         struct request_queue *q = sdev->request_queue;
2106         unsigned long flags;
2107         int err = 0;
2108
2109         err = scsi_device_set_state(sdev, SDEV_BLOCK);
2110         if (err)
2111                 return err;
2112
2113         /* 
2114          * The device has transitioned to SDEV_BLOCK.  Stop the
2115          * block layer from calling the midlayer with this device's
2116          * request queue. 
2117          */
2118         spin_lock_irqsave(q->queue_lock, flags);
2119         blk_stop_queue(q);
2120         spin_unlock_irqrestore(q->queue_lock, flags);
2121
2122         return 0;
2123 }
2124 EXPORT_SYMBOL_GPL(scsi_internal_device_block);
2125  
2126 /**
2127  * scsi_internal_device_unblock - resume a device after a block request
2128  * @sdev:       device to resume
2129  *
2130  * Called by scsi lld's or the midlayer to restart the device queue
2131  * for the previously suspended scsi device.  Called from interrupt or
2132  * normal process context.
2133  *
2134  * Returns zero if successful or error if not.
2135  *
2136  * Notes:       
2137  *      This routine transitions the device to the SDEV_RUNNING state
2138  *      (which must be a legal transition) allowing the midlayer to
2139  *      goose the queue for this device.  This routine assumes the 
2140  *      host_lock is held upon entry.
2141  **/
2142 int
2143 scsi_internal_device_unblock(struct scsi_device *sdev)
2144 {
2145         struct request_queue *q = sdev->request_queue; 
2146         int err;
2147         unsigned long flags;
2148         
2149         /* 
2150          * Try to transition the scsi device to SDEV_RUNNING
2151          * and goose the device queue if successful.  
2152          */
2153         err = scsi_device_set_state(sdev, SDEV_RUNNING);
2154         if (err)
2155                 return err;
2156
2157         spin_lock_irqsave(q->queue_lock, flags);
2158         blk_start_queue(q);
2159         spin_unlock_irqrestore(q->queue_lock, flags);
2160
2161         return 0;
2162 }
2163 EXPORT_SYMBOL_GPL(scsi_internal_device_unblock);
2164
2165 static void
2166 device_block(struct scsi_device *sdev, void *data)
2167 {
2168         scsi_internal_device_block(sdev);
2169 }
2170
2171 static int
2172 target_block(struct device *dev, void *data)
2173 {
2174         if (scsi_is_target_device(dev))
2175                 starget_for_each_device(to_scsi_target(dev), NULL,
2176                                         device_block);
2177         return 0;
2178 }
2179
2180 void
2181 scsi_target_block(struct device *dev)
2182 {
2183         if (scsi_is_target_device(dev))
2184                 starget_for_each_device(to_scsi_target(dev), NULL,
2185                                         device_block);
2186         else
2187                 device_for_each_child(dev, NULL, target_block);
2188 }
2189 EXPORT_SYMBOL_GPL(scsi_target_block);
2190
2191 static void
2192 device_unblock(struct scsi_device *sdev, void *data)
2193 {
2194         scsi_internal_device_unblock(sdev);
2195 }
2196
2197 static int
2198 target_unblock(struct device *dev, void *data)
2199 {
2200         if (scsi_is_target_device(dev))
2201                 starget_for_each_device(to_scsi_target(dev), NULL,
2202                                         device_unblock);
2203         return 0;
2204 }
2205
2206 void
2207 scsi_target_unblock(struct device *dev)
2208 {
2209         if (scsi_is_target_device(dev))
2210                 starget_for_each_device(to_scsi_target(dev), NULL,
2211                                         device_unblock);
2212         else
2213                 device_for_each_child(dev, NULL, target_unblock);
2214 }
2215 EXPORT_SYMBOL_GPL(scsi_target_unblock);
2216
2217 /**
2218  * scsi_kmap_atomic_sg - find and atomically map an sg-elemnt
2219  * @sg:         scatter-gather list
2220  * @sg_count:   number of segments in sg
2221  * @offset:     offset in bytes into sg, on return offset into the mapped area
2222  * @len:        bytes to map, on return number of bytes mapped
2223  *
2224  * Returns virtual address of the start of the mapped page
2225  */
2226 void *scsi_kmap_atomic_sg(struct scatterlist *sg, int sg_count,
2227                           size_t *offset, size_t *len)
2228 {
2229         int i;
2230         size_t sg_len = 0, len_complete = 0;
2231         struct page *page;
2232
2233         WARN_ON(!irqs_disabled());
2234
2235         for (i = 0; i < sg_count; i++) {
2236                 len_complete = sg_len; /* Complete sg-entries */
2237                 sg_len += sg[i].length;
2238                 if (sg_len > *offset)
2239                         break;
2240         }
2241
2242         if (unlikely(i == sg_count)) {
2243                 printk(KERN_ERR "%s: Bytes in sg: %zu, requested offset %zu, "
2244                         "elements %d\n",
2245                        __FUNCTION__, sg_len, *offset, sg_count);
2246                 WARN_ON(1);
2247                 return NULL;
2248         }
2249
2250         /* Offset starting from the beginning of first page in this sg-entry */
2251         *offset = *offset - len_complete + sg[i].offset;
2252
2253         /* Assumption: contiguous pages can be accessed as "page + i" */
2254         page = nth_page(sg[i].page, (*offset >> PAGE_SHIFT));
2255         *offset &= ~PAGE_MASK;
2256
2257         /* Bytes in this sg-entry from *offset to the end of the page */
2258         sg_len = PAGE_SIZE - *offset;
2259         if (*len > sg_len)
2260                 *len = sg_len;
2261
2262         return kmap_atomic(page, KM_BIO_SRC_IRQ);
2263 }
2264 EXPORT_SYMBOL(scsi_kmap_atomic_sg);
2265
2266 /**
2267  * scsi_kunmap_atomic_sg - atomically unmap a virtual address, previously
2268  *                         mapped with scsi_kmap_atomic_sg
2269  * @virt:       virtual address to be unmapped
2270  */
2271 void scsi_kunmap_atomic_sg(void *virt)
2272 {
2273         kunmap_atomic(virt, KM_BIO_SRC_IRQ);
2274 }
2275 EXPORT_SYMBOL(scsi_kunmap_atomic_sg);