New function blk_req_append_bio
[linux-2.6.git] / 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         return blk_rq_append_bio(q, rq, bio);
267 }
268
269 static int scsi_bi_endio(struct bio *bio, unsigned int bytes_done, int error)
270 {
271         if (bio->bi_size)
272                 return 1;
273
274         bio_put(bio);
275         return 0;
276 }
277
278 /**
279  * scsi_req_map_sg - map a scatterlist into a request
280  * @rq:         request to fill
281  * @sg:         scatterlist
282  * @nsegs:      number of elements
283  * @bufflen:    len of buffer
284  * @gfp:        memory allocation flags
285  *
286  * scsi_req_map_sg maps a scatterlist into a request so that the
287  * request can be sent to the block layer. We do not trust the scatterlist
288  * sent to use, as some ULDs use that struct to only organize the pages.
289  */
290 static int scsi_req_map_sg(struct request *rq, struct scatterlist *sgl,
291                            int nsegs, unsigned bufflen, gfp_t gfp)
292 {
293         struct request_queue *q = rq->q;
294         int nr_pages = (bufflen + sgl[0].offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
295         unsigned int data_len = 0, len, bytes, off;
296         struct page *page;
297         struct bio *bio = NULL;
298         int i, err, nr_vecs = 0;
299
300         for (i = 0; i < nsegs; i++) {
301                 page = sgl[i].page;
302                 off = sgl[i].offset;
303                 len = sgl[i].length;
304                 data_len += len;
305
306                 while (len > 0) {
307                         bytes = min_t(unsigned int, len, PAGE_SIZE - off);
308
309                         if (!bio) {
310                                 nr_vecs = min_t(int, BIO_MAX_PAGES, nr_pages);
311                                 nr_pages -= nr_vecs;
312
313                                 bio = bio_alloc(gfp, nr_vecs);
314                                 if (!bio) {
315                                         err = -ENOMEM;
316                                         goto free_bios;
317                                 }
318                                 bio->bi_end_io = scsi_bi_endio;
319                         }
320
321                         if (bio_add_pc_page(q, bio, page, bytes, off) !=
322                             bytes) {
323                                 bio_put(bio);
324                                 err = -EINVAL;
325                                 goto free_bios;
326                         }
327
328                         if (bio->bi_vcnt >= nr_vecs) {
329                                 err = scsi_merge_bio(rq, bio);
330                                 if (err) {
331                                         bio_endio(bio, bio->bi_size, 0);
332                                         goto free_bios;
333                                 }
334                                 bio = NULL;
335                         }
336
337                         page++;
338                         len -= bytes;
339                         off = 0;
340                 }
341         }
342
343         rq->buffer = rq->data = NULL;
344         rq->data_len = data_len;
345         return 0;
346
347 free_bios:
348         while ((bio = rq->bio) != NULL) {
349                 rq->bio = bio->bi_next;
350                 /*
351                  * call endio instead of bio_put incase it was bounced
352                  */
353                 bio_endio(bio, bio->bi_size, 0);
354         }
355
356         return err;
357 }
358
359 /**
360  * scsi_execute_async - insert request
361  * @sdev:       scsi device
362  * @cmd:        scsi command
363  * @cmd_len:    length of scsi cdb
364  * @data_direction: data direction
365  * @buffer:     data buffer (this can be a kernel buffer or scatterlist)
366  * @bufflen:    len of buffer
367  * @use_sg:     if buffer is a scatterlist this is the number of elements
368  * @timeout:    request timeout in seconds
369  * @retries:    number of times to retry request
370  * @flags:      or into request flags
371  **/
372 int scsi_execute_async(struct scsi_device *sdev, const unsigned char *cmd,
373                        int cmd_len, int data_direction, void *buffer, unsigned bufflen,
374                        int use_sg, int timeout, int retries, void *privdata,
375                        void (*done)(void *, char *, int, int), gfp_t gfp)
376 {
377         struct request *req;
378         struct scsi_io_context *sioc;
379         int err = 0;
380         int write = (data_direction == DMA_TO_DEVICE);
381
382         sioc = kmem_cache_zalloc(scsi_io_context_cache, gfp);
383         if (!sioc)
384                 return DRIVER_ERROR << 24;
385
386         req = blk_get_request(sdev->request_queue, write, gfp);
387         if (!req)
388                 goto free_sense;
389         req->cmd_type = REQ_TYPE_BLOCK_PC;
390         req->cmd_flags |= REQ_QUIET;
391
392         if (use_sg)
393                 err = scsi_req_map_sg(req, buffer, use_sg, bufflen, gfp);
394         else if (bufflen)
395                 err = blk_rq_map_kern(req->q, req, buffer, bufflen, gfp);
396
397         if (err)
398                 goto free_req;
399
400         req->cmd_len = cmd_len;
401         memset(req->cmd, 0, BLK_MAX_CDB); /* ATAPI hates garbage after CDB */
402         memcpy(req->cmd, cmd, req->cmd_len);
403         req->sense = sioc->sense;
404         req->sense_len = 0;
405         req->timeout = timeout;
406         req->retries = retries;
407         req->end_io_data = sioc;
408
409         sioc->data = privdata;
410         sioc->done = done;
411
412         blk_execute_rq_nowait(req->q, NULL, req, 1, scsi_end_async);
413         return 0;
414
415 free_req:
416         blk_put_request(req);
417 free_sense:
418         kmem_cache_free(scsi_io_context_cache, sioc);
419         return DRIVER_ERROR << 24;
420 }
421 EXPORT_SYMBOL_GPL(scsi_execute_async);
422
423 /*
424  * Function:    scsi_init_cmd_errh()
425  *
426  * Purpose:     Initialize cmd fields related to error handling.
427  *
428  * Arguments:   cmd     - command that is ready to be queued.
429  *
430  * Notes:       This function has the job of initializing a number of
431  *              fields related to error handling.   Typically this will
432  *              be called once for each command, as required.
433  */
434 static void scsi_init_cmd_errh(struct scsi_cmnd *cmd)
435 {
436         cmd->serial_number = 0;
437         memset(cmd->sense_buffer, 0, sizeof cmd->sense_buffer);
438         if (cmd->cmd_len == 0)
439                 cmd->cmd_len = COMMAND_SIZE(cmd->cmnd[0]);
440 }
441
442 void scsi_device_unbusy(struct scsi_device *sdev)
443 {
444         struct Scsi_Host *shost = sdev->host;
445         unsigned long flags;
446
447         spin_lock_irqsave(shost->host_lock, flags);
448         shost->host_busy--;
449         if (unlikely(scsi_host_in_recovery(shost) &&
450                      (shost->host_failed || shost->host_eh_scheduled)))
451                 scsi_eh_wakeup(shost);
452         spin_unlock(shost->host_lock);
453         spin_lock(sdev->request_queue->queue_lock);
454         sdev->device_busy--;
455         spin_unlock_irqrestore(sdev->request_queue->queue_lock, flags);
456 }
457
458 /*
459  * Called for single_lun devices on IO completion. Clear starget_sdev_user,
460  * and call blk_run_queue for all the scsi_devices on the target -
461  * including current_sdev first.
462  *
463  * Called with *no* scsi locks held.
464  */
465 static void scsi_single_lun_run(struct scsi_device *current_sdev)
466 {
467         struct Scsi_Host *shost = current_sdev->host;
468         struct scsi_device *sdev, *tmp;
469         struct scsi_target *starget = scsi_target(current_sdev);
470         unsigned long flags;
471
472         spin_lock_irqsave(shost->host_lock, flags);
473         starget->starget_sdev_user = NULL;
474         spin_unlock_irqrestore(shost->host_lock, flags);
475
476         /*
477          * Call blk_run_queue for all LUNs on the target, starting with
478          * current_sdev. We race with others (to set starget_sdev_user),
479          * but in most cases, we will be first. Ideally, each LU on the
480          * target would get some limited time or requests on the target.
481          */
482         blk_run_queue(current_sdev->request_queue);
483
484         spin_lock_irqsave(shost->host_lock, flags);
485         if (starget->starget_sdev_user)
486                 goto out;
487         list_for_each_entry_safe(sdev, tmp, &starget->devices,
488                         same_target_siblings) {
489                 if (sdev == current_sdev)
490                         continue;
491                 if (scsi_device_get(sdev))
492                         continue;
493
494                 spin_unlock_irqrestore(shost->host_lock, flags);
495                 blk_run_queue(sdev->request_queue);
496                 spin_lock_irqsave(shost->host_lock, flags);
497         
498                 scsi_device_put(sdev);
499         }
500  out:
501         spin_unlock_irqrestore(shost->host_lock, flags);
502 }
503
504 /*
505  * Function:    scsi_run_queue()
506  *
507  * Purpose:     Select a proper request queue to serve next
508  *
509  * Arguments:   q       - last request's queue
510  *
511  * Returns:     Nothing
512  *
513  * Notes:       The previous command was completely finished, start
514  *              a new one if possible.
515  */
516 static void scsi_run_queue(struct request_queue *q)
517 {
518         struct scsi_device *sdev = q->queuedata;
519         struct Scsi_Host *shost = sdev->host;
520         unsigned long flags;
521
522         if (sdev->single_lun)
523                 scsi_single_lun_run(sdev);
524
525         spin_lock_irqsave(shost->host_lock, flags);
526         while (!list_empty(&shost->starved_list) &&
527                !shost->host_blocked && !shost->host_self_blocked &&
528                 !((shost->can_queue > 0) &&
529                   (shost->host_busy >= shost->can_queue))) {
530                 /*
531                  * As long as shost is accepting commands and we have
532                  * starved queues, call blk_run_queue. scsi_request_fn
533                  * drops the queue_lock and can add us back to the
534                  * starved_list.
535                  *
536                  * host_lock protects the starved_list and starved_entry.
537                  * scsi_request_fn must get the host_lock before checking
538                  * or modifying starved_list or starved_entry.
539                  */
540                 sdev = list_entry(shost->starved_list.next,
541                                           struct scsi_device, starved_entry);
542                 list_del_init(&sdev->starved_entry);
543                 spin_unlock_irqrestore(shost->host_lock, flags);
544
545
546                 if (test_bit(QUEUE_FLAG_REENTER, &q->queue_flags) &&
547                     !test_and_set_bit(QUEUE_FLAG_REENTER,
548                                       &sdev->request_queue->queue_flags)) {
549                         blk_run_queue(sdev->request_queue);
550                         clear_bit(QUEUE_FLAG_REENTER,
551                                   &sdev->request_queue->queue_flags);
552                 } else
553                         blk_run_queue(sdev->request_queue);
554
555                 spin_lock_irqsave(shost->host_lock, flags);
556                 if (unlikely(!list_empty(&sdev->starved_entry)))
557                         /*
558                          * sdev lost a race, and was put back on the
559                          * starved list. This is unlikely but without this
560                          * in theory we could loop forever.
561                          */
562                         break;
563         }
564         spin_unlock_irqrestore(shost->host_lock, flags);
565
566         blk_run_queue(q);
567 }
568
569 /*
570  * Function:    scsi_requeue_command()
571  *
572  * Purpose:     Handle post-processing of completed commands.
573  *
574  * Arguments:   q       - queue to operate on
575  *              cmd     - command that may need to be requeued.
576  *
577  * Returns:     Nothing
578  *
579  * Notes:       After command completion, there may be blocks left
580  *              over which weren't finished by the previous command
581  *              this can be for a number of reasons - the main one is
582  *              I/O errors in the middle of the request, in which case
583  *              we need to request the blocks that come after the bad
584  *              sector.
585  * Notes:       Upon return, cmd is a stale pointer.
586  */
587 static void scsi_requeue_command(struct request_queue *q, struct scsi_cmnd *cmd)
588 {
589         struct request *req = cmd->request;
590         unsigned long flags;
591
592         scsi_unprep_request(req);
593         spin_lock_irqsave(q->queue_lock, flags);
594         blk_requeue_request(q, req);
595         spin_unlock_irqrestore(q->queue_lock, flags);
596
597         scsi_run_queue(q);
598 }
599
600 void scsi_next_command(struct scsi_cmnd *cmd)
601 {
602         struct scsi_device *sdev = cmd->device;
603         struct request_queue *q = sdev->request_queue;
604
605         /* need to hold a reference on the device before we let go of the cmd */
606         get_device(&sdev->sdev_gendev);
607
608         scsi_put_command(cmd);
609         scsi_run_queue(q);
610
611         /* ok to remove device now */
612         put_device(&sdev->sdev_gendev);
613 }
614
615 void scsi_run_host_queues(struct Scsi_Host *shost)
616 {
617         struct scsi_device *sdev;
618
619         shost_for_each_device(sdev, shost)
620                 scsi_run_queue(sdev->request_queue);
621 }
622
623 /*
624  * Function:    scsi_end_request()
625  *
626  * Purpose:     Post-processing of completed commands (usually invoked at end
627  *              of upper level post-processing and scsi_io_completion).
628  *
629  * Arguments:   cmd      - command that is complete.
630  *              uptodate - 1 if I/O indicates success, <= 0 for I/O error.
631  *              bytes    - number of bytes of completed I/O
632  *              requeue  - indicates whether we should requeue leftovers.
633  *
634  * Lock status: Assumed that lock is not held upon entry.
635  *
636  * Returns:     cmd if requeue required, NULL otherwise.
637  *
638  * Notes:       This is called for block device requests in order to
639  *              mark some number of sectors as complete.
640  * 
641  *              We are guaranteeing that the request queue will be goosed
642  *              at some point during this call.
643  * Notes:       If cmd was requeued, upon return it will be a stale pointer.
644  */
645 static struct scsi_cmnd *scsi_end_request(struct scsi_cmnd *cmd, int uptodate,
646                                           int bytes, int requeue)
647 {
648         struct request_queue *q = cmd->device->request_queue;
649         struct request *req = cmd->request;
650         unsigned long flags;
651
652         /*
653          * If there are blocks left over at the end, set up the command
654          * to queue the remainder of them.
655          */
656         if (end_that_request_chunk(req, uptodate, bytes)) {
657                 int leftover = (req->hard_nr_sectors << 9);
658
659                 if (blk_pc_request(req))
660                         leftover = req->data_len;
661
662                 /* kill remainder if no retrys */
663                 if (!uptodate && blk_noretry_request(req))
664                         end_that_request_chunk(req, 0, leftover);
665                 else {
666                         if (requeue) {
667                                 /*
668                                  * Bleah.  Leftovers again.  Stick the
669                                  * leftovers in the front of the
670                                  * queue, and goose the queue again.
671                                  */
672                                 scsi_requeue_command(q, cmd);
673                                 cmd = NULL;
674                         }
675                         return cmd;
676                 }
677         }
678
679         add_disk_randomness(req->rq_disk);
680
681         spin_lock_irqsave(q->queue_lock, flags);
682         if (blk_rq_tagged(req))
683                 blk_queue_end_tag(q, req);
684         end_that_request_last(req, uptodate);
685         spin_unlock_irqrestore(q->queue_lock, flags);
686
687         /*
688          * This will goose the queue request function at the end, so we don't
689          * need to worry about launching another command.
690          */
691         scsi_next_command(cmd);
692         return NULL;
693 }
694
695 struct scatterlist *scsi_alloc_sgtable(struct scsi_cmnd *cmd, gfp_t gfp_mask)
696 {
697         struct scsi_host_sg_pool *sgp;
698         struct scatterlist *sgl;
699
700         BUG_ON(!cmd->use_sg);
701
702         switch (cmd->use_sg) {
703         case 1 ... 8:
704                 cmd->sglist_len = 0;
705                 break;
706         case 9 ... 16:
707                 cmd->sglist_len = 1;
708                 break;
709         case 17 ... 32:
710                 cmd->sglist_len = 2;
711                 break;
712 #if (SCSI_MAX_PHYS_SEGMENTS > 32)
713         case 33 ... 64:
714                 cmd->sglist_len = 3;
715                 break;
716 #if (SCSI_MAX_PHYS_SEGMENTS > 64)
717         case 65 ... 128:
718                 cmd->sglist_len = 4;
719                 break;
720 #if (SCSI_MAX_PHYS_SEGMENTS  > 128)
721         case 129 ... 256:
722                 cmd->sglist_len = 5;
723                 break;
724 #endif
725 #endif
726 #endif
727         default:
728                 return NULL;
729         }
730
731         sgp = scsi_sg_pools + cmd->sglist_len;
732         sgl = mempool_alloc(sgp->pool, gfp_mask);
733         return sgl;
734 }
735
736 EXPORT_SYMBOL(scsi_alloc_sgtable);
737
738 void scsi_free_sgtable(struct scatterlist *sgl, int index)
739 {
740         struct scsi_host_sg_pool *sgp;
741
742         BUG_ON(index >= SG_MEMPOOL_NR);
743
744         sgp = scsi_sg_pools + index;
745         mempool_free(sgl, sgp->pool);
746 }
747
748 EXPORT_SYMBOL(scsi_free_sgtable);
749
750 /*
751  * Function:    scsi_release_buffers()
752  *
753  * Purpose:     Completion processing for block device I/O requests.
754  *
755  * Arguments:   cmd     - command that we are bailing.
756  *
757  * Lock status: Assumed that no lock is held upon entry.
758  *
759  * Returns:     Nothing
760  *
761  * Notes:       In the event that an upper level driver rejects a
762  *              command, we must release resources allocated during
763  *              the __init_io() function.  Primarily this would involve
764  *              the scatter-gather table, and potentially any bounce
765  *              buffers.
766  */
767 static void scsi_release_buffers(struct scsi_cmnd *cmd)
768 {
769         if (cmd->use_sg)
770                 scsi_free_sgtable(cmd->request_buffer, cmd->sglist_len);
771
772         /*
773          * Zero these out.  They now point to freed memory, and it is
774          * dangerous to hang onto the pointers.
775          */
776         cmd->request_buffer = NULL;
777         cmd->request_bufflen = 0;
778 }
779
780 /*
781  * Function:    scsi_io_completion()
782  *
783  * Purpose:     Completion processing for block device I/O requests.
784  *
785  * Arguments:   cmd   - command that is finished.
786  *
787  * Lock status: Assumed that no lock is held upon entry.
788  *
789  * Returns:     Nothing
790  *
791  * Notes:       This function is matched in terms of capabilities to
792  *              the function that created the scatter-gather list.
793  *              In other words, if there are no bounce buffers
794  *              (the normal case for most drivers), we don't need
795  *              the logic to deal with cleaning up afterwards.
796  *
797  *              We must do one of several things here:
798  *
799  *              a) Call scsi_end_request.  This will finish off the
800  *                 specified number of sectors.  If we are done, the
801  *                 command block will be released, and the queue
802  *                 function will be goosed.  If we are not done, then
803  *                 scsi_end_request will directly goose the queue.
804  *
805  *              b) We can just use scsi_requeue_command() here.  This would
806  *                 be used if we just wanted to retry, for example.
807  */
808 void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes)
809 {
810         int result = cmd->result;
811         int this_count = cmd->request_bufflen;
812         struct request_queue *q = cmd->device->request_queue;
813         struct request *req = cmd->request;
814         int clear_errors = 1;
815         struct scsi_sense_hdr sshdr;
816         int sense_valid = 0;
817         int sense_deferred = 0;
818
819         scsi_release_buffers(cmd);
820
821         if (result) {
822                 sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
823                 if (sense_valid)
824                         sense_deferred = scsi_sense_is_deferred(&sshdr);
825         }
826
827         if (blk_pc_request(req)) { /* SG_IO ioctl from block level */
828                 req->errors = result;
829                 if (result) {
830                         clear_errors = 0;
831                         if (sense_valid && req->sense) {
832                                 /*
833                                  * SG_IO wants current and deferred errors
834                                  */
835                                 int len = 8 + cmd->sense_buffer[7];
836
837                                 if (len > SCSI_SENSE_BUFFERSIZE)
838                                         len = SCSI_SENSE_BUFFERSIZE;
839                                 memcpy(req->sense, cmd->sense_buffer,  len);
840                                 req->sense_len = len;
841                         }
842                 }
843                 req->data_len = cmd->resid;
844         }
845
846         /*
847          * Next deal with any sectors which we were able to correctly
848          * handle.
849          */
850         SCSI_LOG_HLCOMPLETE(1, printk("%ld sectors total, "
851                                       "%d bytes done.\n",
852                                       req->nr_sectors, good_bytes));
853         SCSI_LOG_HLCOMPLETE(1, printk("use_sg is %d\n", cmd->use_sg));
854
855         if (clear_errors)
856                 req->errors = 0;
857
858         /* A number of bytes were successfully read.  If there
859          * are leftovers and there is some kind of error
860          * (result != 0), retry the rest.
861          */
862         if (scsi_end_request(cmd, 1, good_bytes, result == 0) == NULL)
863                 return;
864
865         /* good_bytes = 0, or (inclusive) there were leftovers and
866          * result = 0, so scsi_end_request couldn't retry.
867          */
868         if (sense_valid && !sense_deferred) {
869                 switch (sshdr.sense_key) {
870                 case UNIT_ATTENTION:
871                         if (cmd->device->removable) {
872                                 /* Detected disc change.  Set a bit
873                                  * and quietly refuse further access.
874                                  */
875                                 cmd->device->changed = 1;
876                                 scsi_end_request(cmd, 0, this_count, 1);
877                                 return;
878                         } else {
879                                 /* Must have been a power glitch, or a
880                                  * bus reset.  Could not have been a
881                                  * media change, so we just retry the
882                                  * request and see what happens.
883                                  */
884                                 scsi_requeue_command(q, cmd);
885                                 return;
886                         }
887                         break;
888                 case ILLEGAL_REQUEST:
889                         /* If we had an ILLEGAL REQUEST returned, then
890                          * we may have performed an unsupported
891                          * command.  The only thing this should be
892                          * would be a ten byte read where only a six
893                          * byte read was supported.  Also, on a system
894                          * where READ CAPACITY failed, we may have
895                          * read past the end of the disk.
896                          */
897                         if ((cmd->device->use_10_for_rw &&
898                             sshdr.asc == 0x20 && sshdr.ascq == 0x00) &&
899                             (cmd->cmnd[0] == READ_10 ||
900                              cmd->cmnd[0] == WRITE_10)) {
901                                 cmd->device->use_10_for_rw = 0;
902                                 /* This will cause a retry with a
903                                  * 6-byte command.
904                                  */
905                                 scsi_requeue_command(q, cmd);
906                                 return;
907                         } else {
908                                 scsi_end_request(cmd, 0, this_count, 1);
909                                 return;
910                         }
911                         break;
912                 case NOT_READY:
913                         /* If the device is in the process of becoming
914                          * ready, or has a temporary blockage, retry.
915                          */
916                         if (sshdr.asc == 0x04) {
917                                 switch (sshdr.ascq) {
918                                 case 0x01: /* becoming ready */
919                                 case 0x04: /* format in progress */
920                                 case 0x05: /* rebuild in progress */
921                                 case 0x06: /* recalculation in progress */
922                                 case 0x07: /* operation in progress */
923                                 case 0x08: /* Long write in progress */
924                                 case 0x09: /* self test in progress */
925                                         scsi_requeue_command(q, cmd);
926                                         return;
927                                 default:
928                                         break;
929                                 }
930                         }
931                         if (!(req->cmd_flags & REQ_QUIET)) {
932                                 scmd_printk(KERN_INFO, cmd,
933                                             "Device not ready: ");
934                                 scsi_print_sense_hdr("", &sshdr);
935                         }
936                         scsi_end_request(cmd, 0, this_count, 1);
937                         return;
938                 case VOLUME_OVERFLOW:
939                         if (!(req->cmd_flags & REQ_QUIET)) {
940                                 scmd_printk(KERN_INFO, cmd,
941                                             "Volume overflow, CDB: ");
942                                 __scsi_print_command(cmd->cmnd);
943                                 scsi_print_sense("", cmd);
944                         }
945                         /* See SSC3rXX or current. */
946                         scsi_end_request(cmd, 0, this_count, 1);
947                         return;
948                 default:
949                         break;
950                 }
951         }
952         if (host_byte(result) == DID_RESET) {
953                 /* Third party bus reset or reset for error recovery
954                  * reasons.  Just retry the request and see what
955                  * happens.
956                  */
957                 scsi_requeue_command(q, cmd);
958                 return;
959         }
960         if (result) {
961                 if (!(req->cmd_flags & REQ_QUIET)) {
962                         scsi_print_result(cmd);
963                         if (driver_byte(result) & DRIVER_SENSE)
964                                 scsi_print_sense("", cmd);
965                 }
966         }
967         scsi_end_request(cmd, 0, this_count, !result);
968 }
969 EXPORT_SYMBOL(scsi_io_completion);
970
971 /*
972  * Function:    scsi_init_io()
973  *
974  * Purpose:     SCSI I/O initialize function.
975  *
976  * Arguments:   cmd   - Command descriptor we wish to initialize
977  *
978  * Returns:     0 on success
979  *              BLKPREP_DEFER if the failure is retryable
980  *              BLKPREP_KILL if the failure is fatal
981  */
982 static int scsi_init_io(struct scsi_cmnd *cmd)
983 {
984         struct request     *req = cmd->request;
985         struct scatterlist *sgpnt;
986         int                count;
987
988         /*
989          * We used to not use scatter-gather for single segment request,
990          * but now we do (it makes highmem I/O easier to support without
991          * kmapping pages)
992          */
993         cmd->use_sg = req->nr_phys_segments;
994
995         /*
996          * If sg table allocation fails, requeue request later.
997          */
998         sgpnt = scsi_alloc_sgtable(cmd, GFP_ATOMIC);
999         if (unlikely(!sgpnt)) {
1000                 scsi_unprep_request(req);
1001                 return BLKPREP_DEFER;
1002         }
1003
1004         req->buffer = NULL;
1005         cmd->request_buffer = (char *) sgpnt;
1006         if (blk_pc_request(req))
1007                 cmd->request_bufflen = req->data_len;
1008         else
1009                 cmd->request_bufflen = req->nr_sectors << 9;
1010
1011         /* 
1012          * Next, walk the list, and fill in the addresses and sizes of
1013          * each segment.
1014          */
1015         count = blk_rq_map_sg(req->q, req, cmd->request_buffer);
1016         if (likely(count <= cmd->use_sg)) {
1017                 cmd->use_sg = count;
1018                 return BLKPREP_OK;
1019         }
1020
1021         printk(KERN_ERR "Incorrect number of segments after building list\n");
1022         printk(KERN_ERR "counted %d, received %d\n", count, cmd->use_sg);
1023         printk(KERN_ERR "req nr_sec %lu, cur_nr_sec %u\n", req->nr_sectors,
1024                         req->current_nr_sectors);
1025
1026         /* release the command and kill it */
1027         scsi_release_buffers(cmd);
1028         scsi_put_command(cmd);
1029         return BLKPREP_KILL;
1030 }
1031
1032 static struct scsi_cmnd *scsi_get_cmd_from_req(struct scsi_device *sdev,
1033                 struct request *req)
1034 {
1035         struct scsi_cmnd *cmd;
1036
1037         if (!req->special) {
1038                 cmd = scsi_get_command(sdev, GFP_ATOMIC);
1039                 if (unlikely(!cmd))
1040                         return NULL;
1041                 req->special = cmd;
1042         } else {
1043                 cmd = req->special;
1044         }
1045
1046         /* pull a tag out of the request if we have one */
1047         cmd->tag = req->tag;
1048         cmd->request = req;
1049
1050         return cmd;
1051 }
1052
1053 static void scsi_blk_pc_done(struct scsi_cmnd *cmd)
1054 {
1055         BUG_ON(!blk_pc_request(cmd->request));
1056         /*
1057          * This will complete the whole command with uptodate=1 so
1058          * as far as the block layer is concerned the command completed
1059          * successfully. Since this is a REQ_BLOCK_PC command the
1060          * caller should check the request's errors value
1061          */
1062         scsi_io_completion(cmd, cmd->request_bufflen);
1063 }
1064
1065 static int scsi_setup_blk_pc_cmnd(struct scsi_device *sdev, struct request *req)
1066 {
1067         struct scsi_cmnd *cmd;
1068
1069         cmd = scsi_get_cmd_from_req(sdev, req);
1070         if (unlikely(!cmd))
1071                 return BLKPREP_DEFER;
1072
1073         /*
1074          * BLOCK_PC requests may transfer data, in which case they must
1075          * a bio attached to them.  Or they might contain a SCSI command
1076          * that does not transfer data, in which case they may optionally
1077          * submit a request without an attached bio.
1078          */
1079         if (req->bio) {
1080                 int ret;
1081
1082                 BUG_ON(!req->nr_phys_segments);
1083
1084                 ret = scsi_init_io(cmd);
1085                 if (unlikely(ret))
1086                         return ret;
1087         } else {
1088                 BUG_ON(req->data_len);
1089                 BUG_ON(req->data);
1090
1091                 cmd->request_bufflen = 0;
1092                 cmd->request_buffer = NULL;
1093                 cmd->use_sg = 0;
1094                 req->buffer = NULL;
1095         }
1096
1097         BUILD_BUG_ON(sizeof(req->cmd) > sizeof(cmd->cmnd));
1098         memcpy(cmd->cmnd, req->cmd, sizeof(cmd->cmnd));
1099         cmd->cmd_len = req->cmd_len;
1100         if (!req->data_len)
1101                 cmd->sc_data_direction = DMA_NONE;
1102         else if (rq_data_dir(req) == WRITE)
1103                 cmd->sc_data_direction = DMA_TO_DEVICE;
1104         else
1105                 cmd->sc_data_direction = DMA_FROM_DEVICE;
1106         
1107         cmd->transfersize = req->data_len;
1108         cmd->allowed = req->retries;
1109         cmd->timeout_per_command = req->timeout;
1110         cmd->done = scsi_blk_pc_done;
1111         return BLKPREP_OK;
1112 }
1113
1114 /*
1115  * Setup a REQ_TYPE_FS command.  These are simple read/write request
1116  * from filesystems that still need to be translated to SCSI CDBs from
1117  * the ULD.
1118  */
1119 static int scsi_setup_fs_cmnd(struct scsi_device *sdev, struct request *req)
1120 {
1121         struct scsi_cmnd *cmd;
1122         struct scsi_driver *drv;
1123         int ret;
1124
1125         /*
1126          * Filesystem requests must transfer data.
1127          */
1128         BUG_ON(!req->nr_phys_segments);
1129
1130         cmd = scsi_get_cmd_from_req(sdev, req);
1131         if (unlikely(!cmd))
1132                 return BLKPREP_DEFER;
1133
1134         ret = scsi_init_io(cmd);
1135         if (unlikely(ret))
1136                 return ret;
1137
1138         /*
1139          * Initialize the actual SCSI command for this request.
1140          */
1141         drv = *(struct scsi_driver **)req->rq_disk->private_data;
1142         if (unlikely(!drv->init_command(cmd))) {
1143                 scsi_release_buffers(cmd);
1144                 scsi_put_command(cmd);
1145                 return BLKPREP_KILL;
1146         }
1147
1148         return BLKPREP_OK;
1149 }
1150
1151 static int scsi_prep_fn(struct request_queue *q, struct request *req)
1152 {
1153         struct scsi_device *sdev = q->queuedata;
1154         int ret = BLKPREP_OK;
1155
1156         /*
1157          * If the device is not in running state we will reject some
1158          * or all commands.
1159          */
1160         if (unlikely(sdev->sdev_state != SDEV_RUNNING)) {
1161                 switch (sdev->sdev_state) {
1162                 case SDEV_OFFLINE:
1163                         /*
1164                          * If the device is offline we refuse to process any
1165                          * commands.  The device must be brought online
1166                          * before trying any recovery commands.
1167                          */
1168                         sdev_printk(KERN_ERR, sdev,
1169                                     "rejecting I/O to offline device\n");
1170                         ret = BLKPREP_KILL;
1171                         break;
1172                 case SDEV_DEL:
1173                         /*
1174                          * If the device is fully deleted, we refuse to
1175                          * process any commands as well.
1176                          */
1177                         sdev_printk(KERN_ERR, sdev,
1178                                     "rejecting I/O to dead device\n");
1179                         ret = BLKPREP_KILL;
1180                         break;
1181                 case SDEV_QUIESCE:
1182                 case SDEV_BLOCK:
1183                         /*
1184                          * If the devices is blocked we defer normal commands.
1185                          */
1186                         if (!(req->cmd_flags & REQ_PREEMPT))
1187                                 ret = BLKPREP_DEFER;
1188                         break;
1189                 default:
1190                         /*
1191                          * For any other not fully online state we only allow
1192                          * special commands.  In particular any user initiated
1193                          * command is not allowed.
1194                          */
1195                         if (!(req->cmd_flags & REQ_PREEMPT))
1196                                 ret = BLKPREP_KILL;
1197                         break;
1198                 }
1199
1200                 if (ret != BLKPREP_OK)
1201                         goto out;
1202         }
1203
1204         switch (req->cmd_type) {
1205         case REQ_TYPE_BLOCK_PC:
1206                 ret = scsi_setup_blk_pc_cmnd(sdev, req);
1207                 break;
1208         case REQ_TYPE_FS:
1209                 ret = scsi_setup_fs_cmnd(sdev, req);
1210                 break;
1211         default:
1212                 /*
1213                  * All other command types are not supported.
1214                  *
1215                  * Note that these days the SCSI subsystem does not use
1216                  * REQ_TYPE_SPECIAL requests anymore.  These are only used
1217                  * (directly or via blk_insert_request) by non-SCSI drivers.
1218                  */
1219                 blk_dump_rq_flags(req, "SCSI bad req");
1220                 ret = BLKPREP_KILL;
1221                 break;
1222         }
1223
1224  out:
1225         switch (ret) {
1226         case BLKPREP_KILL:
1227                 req->errors = DID_NO_CONNECT << 16;
1228                 break;
1229         case BLKPREP_DEFER:
1230                 /*
1231                  * If we defer, the elv_next_request() returns NULL, but the
1232                  * queue must be restarted, so we plug here if no returning
1233                  * command will automatically do that.
1234                  */
1235                 if (sdev->device_busy == 0)
1236                         blk_plug_device(q);
1237                 break;
1238         default:
1239                 req->cmd_flags |= REQ_DONTPREP;
1240         }
1241
1242         return ret;
1243 }
1244
1245 /*
1246  * scsi_dev_queue_ready: if we can send requests to sdev, return 1 else
1247  * return 0.
1248  *
1249  * Called with the queue_lock held.
1250  */
1251 static inline int scsi_dev_queue_ready(struct request_queue *q,
1252                                   struct scsi_device *sdev)
1253 {
1254         if (sdev->device_busy >= sdev->queue_depth)
1255                 return 0;
1256         if (sdev->device_busy == 0 && sdev->device_blocked) {
1257                 /*
1258                  * unblock after device_blocked iterates to zero
1259                  */
1260                 if (--sdev->device_blocked == 0) {
1261                         SCSI_LOG_MLQUEUE(3,
1262                                    sdev_printk(KERN_INFO, sdev,
1263                                    "unblocking device at zero depth\n"));
1264                 } else {
1265                         blk_plug_device(q);
1266                         return 0;
1267                 }
1268         }
1269         if (sdev->device_blocked)
1270                 return 0;
1271
1272         return 1;
1273 }
1274
1275 /*
1276  * scsi_host_queue_ready: if we can send requests to shost, return 1 else
1277  * return 0. We must end up running the queue again whenever 0 is
1278  * returned, else IO can hang.
1279  *
1280  * Called with host_lock held.
1281  */
1282 static inline int scsi_host_queue_ready(struct request_queue *q,
1283                                    struct Scsi_Host *shost,
1284                                    struct scsi_device *sdev)
1285 {
1286         if (scsi_host_in_recovery(shost))
1287                 return 0;
1288         if (shost->host_busy == 0 && shost->host_blocked) {
1289                 /*
1290                  * unblock after host_blocked iterates to zero
1291                  */
1292                 if (--shost->host_blocked == 0) {
1293                         SCSI_LOG_MLQUEUE(3,
1294                                 printk("scsi%d unblocking host at zero depth\n",
1295                                         shost->host_no));
1296                 } else {
1297                         blk_plug_device(q);
1298                         return 0;
1299                 }
1300         }
1301         if ((shost->can_queue > 0 && shost->host_busy >= shost->can_queue) ||
1302             shost->host_blocked || shost->host_self_blocked) {
1303                 if (list_empty(&sdev->starved_entry))
1304                         list_add_tail(&sdev->starved_entry, &shost->starved_list);
1305                 return 0;
1306         }
1307
1308         /* We're OK to process the command, so we can't be starved */
1309         if (!list_empty(&sdev->starved_entry))
1310                 list_del_init(&sdev->starved_entry);
1311
1312         return 1;
1313 }
1314
1315 /*
1316  * Kill a request for a dead device
1317  */
1318 static void scsi_kill_request(struct request *req, struct request_queue *q)
1319 {
1320         struct scsi_cmnd *cmd = req->special;
1321         struct scsi_device *sdev = cmd->device;
1322         struct Scsi_Host *shost = sdev->host;
1323
1324         blkdev_dequeue_request(req);
1325
1326         if (unlikely(cmd == NULL)) {
1327                 printk(KERN_CRIT "impossible request in %s.\n",
1328                                  __FUNCTION__);
1329                 BUG();
1330         }
1331
1332         scsi_init_cmd_errh(cmd);
1333         cmd->result = DID_NO_CONNECT << 16;
1334         atomic_inc(&cmd->device->iorequest_cnt);
1335
1336         /*
1337          * SCSI request completion path will do scsi_device_unbusy(),
1338          * bump busy counts.  To bump the counters, we need to dance
1339          * with the locks as normal issue path does.
1340          */
1341         sdev->device_busy++;
1342         spin_unlock(sdev->request_queue->queue_lock);
1343         spin_lock(shost->host_lock);
1344         shost->host_busy++;
1345         spin_unlock(shost->host_lock);
1346         spin_lock(sdev->request_queue->queue_lock);
1347
1348         __scsi_done(cmd);
1349 }
1350
1351 static void scsi_softirq_done(struct request *rq)
1352 {
1353         struct scsi_cmnd *cmd = rq->completion_data;
1354         unsigned long wait_for = (cmd->allowed + 1) * cmd->timeout_per_command;
1355         int disposition;
1356
1357         INIT_LIST_HEAD(&cmd->eh_entry);
1358
1359         disposition = scsi_decide_disposition(cmd);
1360         if (disposition != SUCCESS &&
1361             time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) {
1362                 sdev_printk(KERN_ERR, cmd->device,
1363                             "timing out command, waited %lus\n",
1364                             wait_for/HZ);
1365                 disposition = SUCCESS;
1366         }
1367                         
1368         scsi_log_completion(cmd, disposition);
1369
1370         switch (disposition) {
1371                 case SUCCESS:
1372                         scsi_finish_command(cmd);
1373                         break;
1374                 case NEEDS_RETRY:
1375                         scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY);
1376                         break;
1377                 case ADD_TO_MLQUEUE:
1378                         scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY);
1379                         break;
1380                 default:
1381                         if (!scsi_eh_scmd_add(cmd, 0))
1382                                 scsi_finish_command(cmd);
1383         }
1384 }
1385
1386 /*
1387  * Function:    scsi_request_fn()
1388  *
1389  * Purpose:     Main strategy routine for SCSI.
1390  *
1391  * Arguments:   q       - Pointer to actual queue.
1392  *
1393  * Returns:     Nothing
1394  *
1395  * Lock status: IO request lock assumed to be held when called.
1396  */
1397 static void scsi_request_fn(struct request_queue *q)
1398 {
1399         struct scsi_device *sdev = q->queuedata;
1400         struct Scsi_Host *shost;
1401         struct scsi_cmnd *cmd;
1402         struct request *req;
1403
1404         if (!sdev) {
1405                 printk("scsi: killing requests for dead queue\n");
1406                 while ((req = elv_next_request(q)) != NULL)
1407                         scsi_kill_request(req, q);
1408                 return;
1409         }
1410
1411         if(!get_device(&sdev->sdev_gendev))
1412                 /* We must be tearing the block queue down already */
1413                 return;
1414
1415         /*
1416          * To start with, we keep looping until the queue is empty, or until
1417          * the host is no longer able to accept any more requests.
1418          */
1419         shost = sdev->host;
1420         while (!blk_queue_plugged(q)) {
1421                 int rtn;
1422                 /*
1423                  * get next queueable request.  We do this early to make sure
1424                  * that the request is fully prepared even if we cannot 
1425                  * accept it.
1426                  */
1427                 req = elv_next_request(q);
1428                 if (!req || !scsi_dev_queue_ready(q, sdev))
1429                         break;
1430
1431                 if (unlikely(!scsi_device_online(sdev))) {
1432                         sdev_printk(KERN_ERR, sdev,
1433                                     "rejecting I/O to offline device\n");
1434                         scsi_kill_request(req, q);
1435                         continue;
1436                 }
1437
1438
1439                 /*
1440                  * Remove the request from the request list.
1441                  */
1442                 if (!(blk_queue_tagged(q) && !blk_queue_start_tag(q, req)))
1443                         blkdev_dequeue_request(req);
1444                 sdev->device_busy++;
1445
1446                 spin_unlock(q->queue_lock);
1447                 cmd = req->special;
1448                 if (unlikely(cmd == NULL)) {
1449                         printk(KERN_CRIT "impossible request in %s.\n"
1450                                          "please mail a stack trace to "
1451                                          "linux-scsi@vger.kernel.org\n",
1452                                          __FUNCTION__);
1453                         blk_dump_rq_flags(req, "foo");
1454                         BUG();
1455                 }
1456                 spin_lock(shost->host_lock);
1457
1458                 if (!scsi_host_queue_ready(q, shost, sdev))
1459                         goto not_ready;
1460                 if (sdev->single_lun) {
1461                         if (scsi_target(sdev)->starget_sdev_user &&
1462                             scsi_target(sdev)->starget_sdev_user != sdev)
1463                                 goto not_ready;
1464                         scsi_target(sdev)->starget_sdev_user = sdev;
1465                 }
1466                 shost->host_busy++;
1467
1468                 /*
1469                  * XXX(hch): This is rather suboptimal, scsi_dispatch_cmd will
1470                  *              take the lock again.
1471                  */
1472                 spin_unlock_irq(shost->host_lock);
1473
1474                 /*
1475                  * Finally, initialize any error handling parameters, and set up
1476                  * the timers for timeouts.
1477                  */
1478                 scsi_init_cmd_errh(cmd);
1479
1480                 /*
1481                  * Dispatch the command to the low-level driver.
1482                  */
1483                 rtn = scsi_dispatch_cmd(cmd);
1484                 spin_lock_irq(q->queue_lock);
1485                 if(rtn) {
1486                         /* we're refusing the command; because of
1487                          * the way locks get dropped, we need to 
1488                          * check here if plugging is required */
1489                         if(sdev->device_busy == 0)
1490                                 blk_plug_device(q);
1491
1492                         break;
1493                 }
1494         }
1495
1496         goto out;
1497
1498  not_ready:
1499         spin_unlock_irq(shost->host_lock);
1500
1501         /*
1502          * lock q, handle tag, requeue req, and decrement device_busy. We
1503          * must return with queue_lock held.
1504          *
1505          * Decrementing device_busy without checking it is OK, as all such
1506          * cases (host limits or settings) should run the queue at some
1507          * later time.
1508          */
1509         spin_lock_irq(q->queue_lock);
1510         blk_requeue_request(q, req);
1511         sdev->device_busy--;
1512         if(sdev->device_busy == 0)
1513                 blk_plug_device(q);
1514  out:
1515         /* must be careful here...if we trigger the ->remove() function
1516          * we cannot be holding the q lock */
1517         spin_unlock_irq(q->queue_lock);
1518         put_device(&sdev->sdev_gendev);
1519         spin_lock_irq(q->queue_lock);
1520 }
1521
1522 u64 scsi_calculate_bounce_limit(struct Scsi_Host *shost)
1523 {
1524         struct device *host_dev;
1525         u64 bounce_limit = 0xffffffff;
1526
1527         if (shost->unchecked_isa_dma)
1528                 return BLK_BOUNCE_ISA;
1529         /*
1530          * Platforms with virtual-DMA translation
1531          * hardware have no practical limit.
1532          */
1533         if (!PCI_DMA_BUS_IS_PHYS)
1534                 return BLK_BOUNCE_ANY;
1535
1536         host_dev = scsi_get_device(shost);
1537         if (host_dev && host_dev->dma_mask)
1538                 bounce_limit = *host_dev->dma_mask;
1539
1540         return bounce_limit;
1541 }
1542 EXPORT_SYMBOL(scsi_calculate_bounce_limit);
1543
1544 struct request_queue *__scsi_alloc_queue(struct Scsi_Host *shost,
1545                                          request_fn_proc *request_fn)
1546 {
1547         struct request_queue *q;
1548
1549         q = blk_init_queue(request_fn, NULL);
1550         if (!q)
1551                 return NULL;
1552
1553         blk_queue_max_hw_segments(q, shost->sg_tablesize);
1554         blk_queue_max_phys_segments(q, SCSI_MAX_PHYS_SEGMENTS);
1555         blk_queue_max_sectors(q, shost->max_sectors);
1556         blk_queue_bounce_limit(q, scsi_calculate_bounce_limit(shost));
1557         blk_queue_segment_boundary(q, shost->dma_boundary);
1558
1559         if (!shost->use_clustering)
1560                 clear_bit(QUEUE_FLAG_CLUSTER, &q->queue_flags);
1561         return q;
1562 }
1563 EXPORT_SYMBOL(__scsi_alloc_queue);
1564
1565 struct request_queue *scsi_alloc_queue(struct scsi_device *sdev)
1566 {
1567         struct request_queue *q;
1568
1569         q = __scsi_alloc_queue(sdev->host, scsi_request_fn);
1570         if (!q)
1571                 return NULL;
1572
1573         blk_queue_prep_rq(q, scsi_prep_fn);
1574         blk_queue_softirq_done(q, scsi_softirq_done);
1575         return q;
1576 }
1577
1578 void scsi_free_queue(struct request_queue *q)
1579 {
1580         blk_cleanup_queue(q);
1581 }
1582
1583 /*
1584  * Function:    scsi_block_requests()
1585  *
1586  * Purpose:     Utility function used by low-level drivers to prevent further
1587  *              commands from being queued to the device.
1588  *
1589  * Arguments:   shost       - Host in question
1590  *
1591  * Returns:     Nothing
1592  *
1593  * Lock status: No locks are assumed held.
1594  *
1595  * Notes:       There is no timer nor any other means by which the requests
1596  *              get unblocked other than the low-level driver calling
1597  *              scsi_unblock_requests().
1598  */
1599 void scsi_block_requests(struct Scsi_Host *shost)
1600 {
1601         shost->host_self_blocked = 1;
1602 }
1603 EXPORT_SYMBOL(scsi_block_requests);
1604
1605 /*
1606  * Function:    scsi_unblock_requests()
1607  *
1608  * Purpose:     Utility function used by low-level drivers to allow further
1609  *              commands from being queued to the device.
1610  *
1611  * Arguments:   shost       - Host in question
1612  *
1613  * Returns:     Nothing
1614  *
1615  * Lock status: No locks are assumed held.
1616  *
1617  * Notes:       There is no timer nor any other means by which the requests
1618  *              get unblocked other than the low-level driver calling
1619  *              scsi_unblock_requests().
1620  *
1621  *              This is done as an API function so that changes to the
1622  *              internals of the scsi mid-layer won't require wholesale
1623  *              changes to drivers that use this feature.
1624  */
1625 void scsi_unblock_requests(struct Scsi_Host *shost)
1626 {
1627         shost->host_self_blocked = 0;
1628         scsi_run_host_queues(shost);
1629 }
1630 EXPORT_SYMBOL(scsi_unblock_requests);
1631
1632 int __init scsi_init_queue(void)
1633 {
1634         int i;
1635
1636         scsi_io_context_cache = kmem_cache_create("scsi_io_context",
1637                                         sizeof(struct scsi_io_context),
1638                                         0, 0, NULL);
1639         if (!scsi_io_context_cache) {
1640                 printk(KERN_ERR "SCSI: can't init scsi io context cache\n");
1641                 return -ENOMEM;
1642         }
1643
1644         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1645                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1646                 int size = sgp->size * sizeof(struct scatterlist);
1647
1648                 sgp->slab = kmem_cache_create(sgp->name, size, 0,
1649                                 SLAB_HWCACHE_ALIGN, NULL);
1650                 if (!sgp->slab) {
1651                         printk(KERN_ERR "SCSI: can't init sg slab %s\n",
1652                                         sgp->name);
1653                 }
1654
1655                 sgp->pool = mempool_create_slab_pool(SG_MEMPOOL_SIZE,
1656                                                      sgp->slab);
1657                 if (!sgp->pool) {
1658                         printk(KERN_ERR "SCSI: can't init sg mempool %s\n",
1659                                         sgp->name);
1660                 }
1661         }
1662
1663         return 0;
1664 }
1665
1666 void scsi_exit_queue(void)
1667 {
1668         int i;
1669
1670         kmem_cache_destroy(scsi_io_context_cache);
1671
1672         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1673                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1674                 mempool_destroy(sgp->pool);
1675                 kmem_cache_destroy(sgp->slab);
1676         }
1677 }
1678
1679 /**
1680  *      scsi_mode_select - issue a mode select
1681  *      @sdev:  SCSI device to be queried
1682  *      @pf:    Page format bit (1 == standard, 0 == vendor specific)
1683  *      @sp:    Save page bit (0 == don't save, 1 == save)
1684  *      @modepage: mode page being requested
1685  *      @buffer: request buffer (may not be smaller than eight bytes)
1686  *      @len:   length of request buffer.
1687  *      @timeout: command timeout
1688  *      @retries: number of retries before failing
1689  *      @data: returns a structure abstracting the mode header data
1690  *      @sense: place to put sense data (or NULL if no sense to be collected).
1691  *              must be SCSI_SENSE_BUFFERSIZE big.
1692  *
1693  *      Returns zero if successful; negative error number or scsi
1694  *      status on error
1695  *
1696  */
1697 int
1698 scsi_mode_select(struct scsi_device *sdev, int pf, int sp, int modepage,
1699                  unsigned char *buffer, int len, int timeout, int retries,
1700                  struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1701 {
1702         unsigned char cmd[10];
1703         unsigned char *real_buffer;
1704         int ret;
1705
1706         memset(cmd, 0, sizeof(cmd));
1707         cmd[1] = (pf ? 0x10 : 0) | (sp ? 0x01 : 0);
1708
1709         if (sdev->use_10_for_ms) {
1710                 if (len > 65535)
1711                         return -EINVAL;
1712                 real_buffer = kmalloc(8 + len, GFP_KERNEL);
1713                 if (!real_buffer)
1714                         return -ENOMEM;
1715                 memcpy(real_buffer + 8, buffer, len);
1716                 len += 8;
1717                 real_buffer[0] = 0;
1718                 real_buffer[1] = 0;
1719                 real_buffer[2] = data->medium_type;
1720                 real_buffer[3] = data->device_specific;
1721                 real_buffer[4] = data->longlba ? 0x01 : 0;
1722                 real_buffer[5] = 0;
1723                 real_buffer[6] = data->block_descriptor_length >> 8;
1724                 real_buffer[7] = data->block_descriptor_length;
1725
1726                 cmd[0] = MODE_SELECT_10;
1727                 cmd[7] = len >> 8;
1728                 cmd[8] = len;
1729         } else {
1730                 if (len > 255 || data->block_descriptor_length > 255 ||
1731                     data->longlba)
1732                         return -EINVAL;
1733
1734                 real_buffer = kmalloc(4 + len, GFP_KERNEL);
1735                 if (!real_buffer)
1736                         return -ENOMEM;
1737                 memcpy(real_buffer + 4, buffer, len);
1738                 len += 4;
1739                 real_buffer[0] = 0;
1740                 real_buffer[1] = data->medium_type;
1741                 real_buffer[2] = data->device_specific;
1742                 real_buffer[3] = data->block_descriptor_length;
1743                 
1744
1745                 cmd[0] = MODE_SELECT;
1746                 cmd[4] = len;
1747         }
1748
1749         ret = scsi_execute_req(sdev, cmd, DMA_TO_DEVICE, real_buffer, len,
1750                                sshdr, timeout, retries);
1751         kfree(real_buffer);
1752         return ret;
1753 }
1754 EXPORT_SYMBOL_GPL(scsi_mode_select);
1755
1756 /**
1757  *      scsi_mode_sense - issue a mode sense, falling back from 10 to 
1758  *              six bytes if necessary.
1759  *      @sdev:  SCSI device to be queried
1760  *      @dbd:   set if mode sense will allow block descriptors to be returned
1761  *      @modepage: mode page being requested
1762  *      @buffer: request buffer (may not be smaller than eight bytes)
1763  *      @len:   length of request buffer.
1764  *      @timeout: command timeout
1765  *      @retries: number of retries before failing
1766  *      @data: returns a structure abstracting the mode header data
1767  *      @sense: place to put sense data (or NULL if no sense to be collected).
1768  *              must be SCSI_SENSE_BUFFERSIZE big.
1769  *
1770  *      Returns zero if unsuccessful, or the header offset (either 4
1771  *      or 8 depending on whether a six or ten byte command was
1772  *      issued) if successful.
1773  **/
1774 int
1775 scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
1776                   unsigned char *buffer, int len, int timeout, int retries,
1777                   struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1778 {
1779         unsigned char cmd[12];
1780         int use_10_for_ms;
1781         int header_length;
1782         int result;
1783         struct scsi_sense_hdr my_sshdr;
1784
1785         memset(data, 0, sizeof(*data));
1786         memset(&cmd[0], 0, 12);
1787         cmd[1] = dbd & 0x18;    /* allows DBD and LLBA bits */
1788         cmd[2] = modepage;
1789
1790         /* caller might not be interested in sense, but we need it */
1791         if (!sshdr)
1792                 sshdr = &my_sshdr;
1793
1794  retry:
1795         use_10_for_ms = sdev->use_10_for_ms;
1796
1797         if (use_10_for_ms) {
1798                 if (len < 8)
1799                         len = 8;
1800
1801                 cmd[0] = MODE_SENSE_10;
1802                 cmd[8] = len;
1803                 header_length = 8;
1804         } else {
1805                 if (len < 4)
1806                         len = 4;
1807
1808                 cmd[0] = MODE_SENSE;
1809                 cmd[4] = len;
1810                 header_length = 4;
1811         }
1812
1813         memset(buffer, 0, len);
1814
1815         result = scsi_execute_req(sdev, cmd, DMA_FROM_DEVICE, buffer, len,
1816                                   sshdr, timeout, retries);
1817
1818         /* This code looks awful: what it's doing is making sure an
1819          * ILLEGAL REQUEST sense return identifies the actual command
1820          * byte as the problem.  MODE_SENSE commands can return
1821          * ILLEGAL REQUEST if the code page isn't supported */
1822
1823         if (use_10_for_ms && !scsi_status_is_good(result) &&
1824             (driver_byte(result) & DRIVER_SENSE)) {
1825                 if (scsi_sense_valid(sshdr)) {
1826                         if ((sshdr->sense_key == ILLEGAL_REQUEST) &&
1827                             (sshdr->asc == 0x20) && (sshdr->ascq == 0)) {
1828                                 /* 
1829                                  * Invalid command operation code
1830                                  */
1831                                 sdev->use_10_for_ms = 0;
1832                                 goto retry;
1833                         }
1834                 }
1835         }
1836
1837         if(scsi_status_is_good(result)) {
1838                 if (unlikely(buffer[0] == 0x86 && buffer[1] == 0x0b &&
1839                              (modepage == 6 || modepage == 8))) {
1840                         /* Initio breakage? */
1841                         header_length = 0;
1842                         data->length = 13;
1843                         data->medium_type = 0;
1844                         data->device_specific = 0;
1845                         data->longlba = 0;
1846                         data->block_descriptor_length = 0;
1847                 } else if(use_10_for_ms) {
1848                         data->length = buffer[0]*256 + buffer[1] + 2;
1849                         data->medium_type = buffer[2];
1850                         data->device_specific = buffer[3];
1851                         data->longlba = buffer[4] & 0x01;
1852                         data->block_descriptor_length = buffer[6]*256
1853                                 + buffer[7];
1854                 } else {
1855                         data->length = buffer[0] + 1;
1856                         data->medium_type = buffer[1];
1857                         data->device_specific = buffer[2];
1858                         data->block_descriptor_length = buffer[3];
1859                 }
1860                 data->header_length = header_length;
1861         }
1862
1863         return result;
1864 }
1865 EXPORT_SYMBOL(scsi_mode_sense);
1866
1867 int
1868 scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries)
1869 {
1870         char cmd[] = {
1871                 TEST_UNIT_READY, 0, 0, 0, 0, 0,
1872         };
1873         struct scsi_sense_hdr sshdr;
1874         int result;
1875         
1876         result = scsi_execute_req(sdev, cmd, DMA_NONE, NULL, 0, &sshdr,
1877                                   timeout, retries);
1878
1879         if ((driver_byte(result) & DRIVER_SENSE) && sdev->removable) {
1880
1881                 if ((scsi_sense_valid(&sshdr)) &&
1882                     ((sshdr.sense_key == UNIT_ATTENTION) ||
1883                      (sshdr.sense_key == NOT_READY))) {
1884                         sdev->changed = 1;
1885                         result = 0;
1886                 }
1887         }
1888         return result;
1889 }
1890 EXPORT_SYMBOL(scsi_test_unit_ready);
1891
1892 /**
1893  *      scsi_device_set_state - Take the given device through the device
1894  *              state model.
1895  *      @sdev:  scsi device to change the state of.
1896  *      @state: state to change to.
1897  *
1898  *      Returns zero if unsuccessful or an error if the requested 
1899  *      transition is illegal.
1900  **/
1901 int
1902 scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
1903 {
1904         enum scsi_device_state oldstate = sdev->sdev_state;
1905
1906         if (state == oldstate)
1907                 return 0;
1908
1909         switch (state) {
1910         case SDEV_CREATED:
1911                 /* There are no legal states that come back to
1912                  * created.  This is the manually initialised start
1913                  * state */
1914                 goto illegal;
1915                         
1916         case SDEV_RUNNING:
1917                 switch (oldstate) {
1918                 case SDEV_CREATED:
1919                 case SDEV_OFFLINE:
1920                 case SDEV_QUIESCE:
1921                 case SDEV_BLOCK:
1922                         break;
1923                 default:
1924                         goto illegal;
1925                 }
1926                 break;
1927
1928         case SDEV_QUIESCE:
1929                 switch (oldstate) {
1930                 case SDEV_RUNNING:
1931                 case SDEV_OFFLINE:
1932                         break;
1933                 default:
1934                         goto illegal;
1935                 }
1936                 break;
1937
1938         case SDEV_OFFLINE:
1939                 switch (oldstate) {
1940                 case SDEV_CREATED:
1941                 case SDEV_RUNNING:
1942                 case SDEV_QUIESCE:
1943                 case SDEV_BLOCK:
1944                         break;
1945                 default:
1946                         goto illegal;
1947                 }
1948                 break;
1949
1950         case SDEV_BLOCK:
1951                 switch (oldstate) {
1952                 case SDEV_CREATED:
1953                 case SDEV_RUNNING:
1954                         break;
1955                 default:
1956                         goto illegal;
1957                 }
1958                 break;
1959
1960         case SDEV_CANCEL:
1961                 switch (oldstate) {
1962                 case SDEV_CREATED:
1963                 case SDEV_RUNNING:
1964                 case SDEV_QUIESCE:
1965                 case SDEV_OFFLINE:
1966                 case SDEV_BLOCK:
1967                         break;
1968                 default:
1969                         goto illegal;
1970                 }
1971                 break;
1972
1973         case SDEV_DEL:
1974                 switch (oldstate) {
1975                 case SDEV_CREATED:
1976                 case SDEV_RUNNING:
1977                 case SDEV_OFFLINE:
1978                 case SDEV_CANCEL:
1979                         break;
1980                 default:
1981                         goto illegal;
1982                 }
1983                 break;
1984
1985         }
1986         sdev->sdev_state = state;
1987         return 0;
1988
1989  illegal:
1990         SCSI_LOG_ERROR_RECOVERY(1, 
1991                                 sdev_printk(KERN_ERR, sdev,
1992                                             "Illegal state transition %s->%s\n",
1993                                             scsi_device_state_name(oldstate),
1994                                             scsi_device_state_name(state))
1995                                 );
1996         return -EINVAL;
1997 }
1998 EXPORT_SYMBOL(scsi_device_set_state);
1999
2000 /**
2001  *      scsi_device_quiesce - Block user issued commands.
2002  *      @sdev:  scsi device to quiesce.
2003  *
2004  *      This works by trying to transition to the SDEV_QUIESCE state
2005  *      (which must be a legal transition).  When the device is in this
2006  *      state, only special requests will be accepted, all others will
2007  *      be deferred.  Since special requests may also be requeued requests,
2008  *      a successful return doesn't guarantee the device will be 
2009  *      totally quiescent.
2010  *
2011  *      Must be called with user context, may sleep.
2012  *
2013  *      Returns zero if unsuccessful or an error if not.
2014  **/
2015 int
2016 scsi_device_quiesce(struct scsi_device *sdev)
2017 {
2018         int err = scsi_device_set_state(sdev, SDEV_QUIESCE);
2019         if (err)
2020                 return err;
2021
2022         scsi_run_queue(sdev->request_queue);
2023         while (sdev->device_busy) {
2024                 msleep_interruptible(200);
2025                 scsi_run_queue(sdev->request_queue);
2026         }
2027         return 0;
2028 }
2029 EXPORT_SYMBOL(scsi_device_quiesce);
2030
2031 /**
2032  *      scsi_device_resume - Restart user issued commands to a quiesced device.
2033  *      @sdev:  scsi device to resume.
2034  *
2035  *      Moves the device from quiesced back to running and restarts the
2036  *      queues.
2037  *
2038  *      Must be called with user context, may sleep.
2039  **/
2040 void
2041 scsi_device_resume(struct scsi_device *sdev)
2042 {
2043         if(scsi_device_set_state(sdev, SDEV_RUNNING))
2044                 return;
2045         scsi_run_queue(sdev->request_queue);
2046 }
2047 EXPORT_SYMBOL(scsi_device_resume);
2048
2049 static void
2050 device_quiesce_fn(struct scsi_device *sdev, void *data)
2051 {
2052         scsi_device_quiesce(sdev);
2053 }
2054
2055 void
2056 scsi_target_quiesce(struct scsi_target *starget)
2057 {
2058         starget_for_each_device(starget, NULL, device_quiesce_fn);
2059 }
2060 EXPORT_SYMBOL(scsi_target_quiesce);
2061
2062 static void
2063 device_resume_fn(struct scsi_device *sdev, void *data)
2064 {
2065         scsi_device_resume(sdev);
2066 }
2067
2068 void
2069 scsi_target_resume(struct scsi_target *starget)
2070 {
2071         starget_for_each_device(starget, NULL, device_resume_fn);
2072 }
2073 EXPORT_SYMBOL(scsi_target_resume);
2074
2075 /**
2076  * scsi_internal_device_block - internal function to put a device
2077  *                              temporarily into the SDEV_BLOCK state
2078  * @sdev:       device to block
2079  *
2080  * Block request made by scsi lld's to temporarily stop all
2081  * scsi commands on the specified device.  Called from interrupt
2082  * or normal process context.
2083  *
2084  * Returns zero if successful or error if not
2085  *
2086  * Notes:       
2087  *      This routine transitions the device to the SDEV_BLOCK state
2088  *      (which must be a legal transition).  When the device is in this
2089  *      state, all commands are deferred until the scsi lld reenables
2090  *      the device with scsi_device_unblock or device_block_tmo fires.
2091  *      This routine assumes the host_lock is held on entry.
2092  **/
2093 int
2094 scsi_internal_device_block(struct scsi_device *sdev)
2095 {
2096         struct request_queue *q = sdev->request_queue;
2097         unsigned long flags;
2098         int err = 0;
2099
2100         err = scsi_device_set_state(sdev, SDEV_BLOCK);
2101         if (err)
2102                 return err;
2103
2104         /* 
2105          * The device has transitioned to SDEV_BLOCK.  Stop the
2106          * block layer from calling the midlayer with this device's
2107          * request queue. 
2108          */
2109         spin_lock_irqsave(q->queue_lock, flags);
2110         blk_stop_queue(q);
2111         spin_unlock_irqrestore(q->queue_lock, flags);
2112
2113         return 0;
2114 }
2115 EXPORT_SYMBOL_GPL(scsi_internal_device_block);
2116  
2117 /**
2118  * scsi_internal_device_unblock - resume a device after a block request
2119  * @sdev:       device to resume
2120  *
2121  * Called by scsi lld's or the midlayer to restart the device queue
2122  * for the previously suspended scsi device.  Called from interrupt or
2123  * normal process context.
2124  *
2125  * Returns zero if successful or error if not.
2126  *
2127  * Notes:       
2128  *      This routine transitions the device to the SDEV_RUNNING state
2129  *      (which must be a legal transition) allowing the midlayer to
2130  *      goose the queue for this device.  This routine assumes the 
2131  *      host_lock is held upon entry.
2132  **/
2133 int
2134 scsi_internal_device_unblock(struct scsi_device *sdev)
2135 {
2136         struct request_queue *q = sdev->request_queue; 
2137         int err;
2138         unsigned long flags;
2139         
2140         /* 
2141          * Try to transition the scsi device to SDEV_RUNNING
2142          * and goose the device queue if successful.  
2143          */
2144         err = scsi_device_set_state(sdev, SDEV_RUNNING);
2145         if (err)
2146                 return err;
2147
2148         spin_lock_irqsave(q->queue_lock, flags);
2149         blk_start_queue(q);
2150         spin_unlock_irqrestore(q->queue_lock, flags);
2151
2152         return 0;
2153 }
2154 EXPORT_SYMBOL_GPL(scsi_internal_device_unblock);
2155
2156 static void
2157 device_block(struct scsi_device *sdev, void *data)
2158 {
2159         scsi_internal_device_block(sdev);
2160 }
2161
2162 static int
2163 target_block(struct device *dev, void *data)
2164 {
2165         if (scsi_is_target_device(dev))
2166                 starget_for_each_device(to_scsi_target(dev), NULL,
2167                                         device_block);
2168         return 0;
2169 }
2170
2171 void
2172 scsi_target_block(struct device *dev)
2173 {
2174         if (scsi_is_target_device(dev))
2175                 starget_for_each_device(to_scsi_target(dev), NULL,
2176                                         device_block);
2177         else
2178                 device_for_each_child(dev, NULL, target_block);
2179 }
2180 EXPORT_SYMBOL_GPL(scsi_target_block);
2181
2182 static void
2183 device_unblock(struct scsi_device *sdev, void *data)
2184 {
2185         scsi_internal_device_unblock(sdev);
2186 }
2187
2188 static int
2189 target_unblock(struct device *dev, void *data)
2190 {
2191         if (scsi_is_target_device(dev))
2192                 starget_for_each_device(to_scsi_target(dev), NULL,
2193                                         device_unblock);
2194         return 0;
2195 }
2196
2197 void
2198 scsi_target_unblock(struct device *dev)
2199 {
2200         if (scsi_is_target_device(dev))
2201                 starget_for_each_device(to_scsi_target(dev), NULL,
2202                                         device_unblock);
2203         else
2204                 device_for_each_child(dev, NULL, target_unblock);
2205 }
2206 EXPORT_SYMBOL_GPL(scsi_target_unblock);
2207
2208 /**
2209  * scsi_kmap_atomic_sg - find and atomically map an sg-elemnt
2210  * @sg:         scatter-gather list
2211  * @sg_count:   number of segments in sg
2212  * @offset:     offset in bytes into sg, on return offset into the mapped area
2213  * @len:        bytes to map, on return number of bytes mapped
2214  *
2215  * Returns virtual address of the start of the mapped page
2216  */
2217 void *scsi_kmap_atomic_sg(struct scatterlist *sg, int sg_count,
2218                           size_t *offset, size_t *len)
2219 {
2220         int i;
2221         size_t sg_len = 0, len_complete = 0;
2222         struct page *page;
2223
2224         WARN_ON(!irqs_disabled());
2225
2226         for (i = 0; i < sg_count; i++) {
2227                 len_complete = sg_len; /* Complete sg-entries */
2228                 sg_len += sg[i].length;
2229                 if (sg_len > *offset)
2230                         break;
2231         }
2232
2233         if (unlikely(i == sg_count)) {
2234                 printk(KERN_ERR "%s: Bytes in sg: %zu, requested offset %zu, "
2235                         "elements %d\n",
2236                        __FUNCTION__, sg_len, *offset, sg_count);
2237                 WARN_ON(1);
2238                 return NULL;
2239         }
2240
2241         /* Offset starting from the beginning of first page in this sg-entry */
2242         *offset = *offset - len_complete + sg[i].offset;
2243
2244         /* Assumption: contiguous pages can be accessed as "page + i" */
2245         page = nth_page(sg[i].page, (*offset >> PAGE_SHIFT));
2246         *offset &= ~PAGE_MASK;
2247
2248         /* Bytes in this sg-entry from *offset to the end of the page */
2249         sg_len = PAGE_SIZE - *offset;
2250         if (*len > sg_len)
2251                 *len = sg_len;
2252
2253         return kmap_atomic(page, KM_BIO_SRC_IRQ);
2254 }
2255 EXPORT_SYMBOL(scsi_kmap_atomic_sg);
2256
2257 /**
2258  * scsi_kunmap_atomic_sg - atomically unmap a virtual address, previously
2259  *                         mapped with scsi_kmap_atomic_sg
2260  * @virt:       virtual address to be unmapped
2261  */
2262 void scsi_kunmap_atomic_sg(void *virt)
2263 {
2264         kunmap_atomic(virt, KM_BIO_SRC_IRQ);
2265 }
2266 EXPORT_SYMBOL(scsi_kunmap_atomic_sg);