[SCSI] stale residual returned on write following BUSY retry
[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 = bufflen, 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
314                 while (len > 0 && data_len > 0) {
315                         /*
316                          * sg sends a scatterlist that is larger than
317                          * the data_len it wants transferred for certain
318                          * IO sizes
319                          */
320                         bytes = min_t(unsigned int, len, PAGE_SIZE - off);
321                         bytes = min(bytes, data_len);
322
323                         if (!bio) {
324                                 nr_vecs = min_t(int, BIO_MAX_PAGES, nr_pages);
325                                 nr_pages -= nr_vecs;
326
327                                 bio = bio_alloc(gfp, nr_vecs);
328                                 if (!bio) {
329                                         err = -ENOMEM;
330                                         goto free_bios;
331                                 }
332                                 bio->bi_end_io = scsi_bi_endio;
333                         }
334
335                         if (bio_add_pc_page(q, bio, page, bytes, off) !=
336                             bytes) {
337                                 bio_put(bio);
338                                 err = -EINVAL;
339                                 goto free_bios;
340                         }
341
342                         if (bio->bi_vcnt >= nr_vecs) {
343                                 err = scsi_merge_bio(rq, bio);
344                                 if (err) {
345                                         bio_endio(bio, bio->bi_size, 0);
346                                         goto free_bios;
347                                 }
348                                 bio = NULL;
349                         }
350
351                         page++;
352                         len -= bytes;
353                         data_len -=bytes;
354                         off = 0;
355                 }
356         }
357
358         rq->buffer = rq->data = NULL;
359         rq->data_len = bufflen;
360         return 0;
361
362 free_bios:
363         while ((bio = rq->bio) != NULL) {
364                 rq->bio = bio->bi_next;
365                 /*
366                  * call endio instead of bio_put incase it was bounced
367                  */
368                 bio_endio(bio, bio->bi_size, 0);
369         }
370
371         return err;
372 }
373
374 /**
375  * scsi_execute_async - insert request
376  * @sdev:       scsi device
377  * @cmd:        scsi command
378  * @cmd_len:    length of scsi cdb
379  * @data_direction: data direction
380  * @buffer:     data buffer (this can be a kernel buffer or scatterlist)
381  * @bufflen:    len of buffer
382  * @use_sg:     if buffer is a scatterlist this is the number of elements
383  * @timeout:    request timeout in seconds
384  * @retries:    number of times to retry request
385  * @flags:      or into request flags
386  **/
387 int scsi_execute_async(struct scsi_device *sdev, const unsigned char *cmd,
388                        int cmd_len, int data_direction, void *buffer, unsigned bufflen,
389                        int use_sg, int timeout, int retries, void *privdata,
390                        void (*done)(void *, char *, int, int), gfp_t gfp)
391 {
392         struct request *req;
393         struct scsi_io_context *sioc;
394         int err = 0;
395         int write = (data_direction == DMA_TO_DEVICE);
396
397         sioc = kmem_cache_zalloc(scsi_io_context_cache, gfp);
398         if (!sioc)
399                 return DRIVER_ERROR << 24;
400
401         req = blk_get_request(sdev->request_queue, write, gfp);
402         if (!req)
403                 goto free_sense;
404         req->cmd_type = REQ_TYPE_BLOCK_PC;
405         req->cmd_flags |= REQ_QUIET;
406
407         if (use_sg)
408                 err = scsi_req_map_sg(req, buffer, use_sg, bufflen, gfp);
409         else if (bufflen)
410                 err = blk_rq_map_kern(req->q, req, buffer, bufflen, gfp);
411
412         if (err)
413                 goto free_req;
414
415         req->cmd_len = cmd_len;
416         memset(req->cmd, 0, BLK_MAX_CDB); /* ATAPI hates garbage after CDB */
417         memcpy(req->cmd, cmd, req->cmd_len);
418         req->sense = sioc->sense;
419         req->sense_len = 0;
420         req->timeout = timeout;
421         req->retries = retries;
422         req->end_io_data = sioc;
423
424         sioc->data = privdata;
425         sioc->done = done;
426
427         blk_execute_rq_nowait(req->q, NULL, req, 1, scsi_end_async);
428         return 0;
429
430 free_req:
431         blk_put_request(req);
432 free_sense:
433         kmem_cache_free(scsi_io_context_cache, sioc);
434         return DRIVER_ERROR << 24;
435 }
436 EXPORT_SYMBOL_GPL(scsi_execute_async);
437
438 /*
439  * Function:    scsi_init_cmd_errh()
440  *
441  * Purpose:     Initialize cmd fields related to error handling.
442  *
443  * Arguments:   cmd     - command that is ready to be queued.
444  *
445  * Notes:       This function has the job of initializing a number of
446  *              fields related to error handling.   Typically this will
447  *              be called once for each command, as required.
448  */
449 static void scsi_init_cmd_errh(struct scsi_cmnd *cmd)
450 {
451         cmd->serial_number = 0;
452         cmd->resid = 0;
453         memset(cmd->sense_buffer, 0, sizeof cmd->sense_buffer);
454         if (cmd->cmd_len == 0)
455                 cmd->cmd_len = COMMAND_SIZE(cmd->cmnd[0]);
456 }
457
458 void scsi_device_unbusy(struct scsi_device *sdev)
459 {
460         struct Scsi_Host *shost = sdev->host;
461         unsigned long flags;
462
463         spin_lock_irqsave(shost->host_lock, flags);
464         shost->host_busy--;
465         if (unlikely(scsi_host_in_recovery(shost) &&
466                      (shost->host_failed || shost->host_eh_scheduled)))
467                 scsi_eh_wakeup(shost);
468         spin_unlock(shost->host_lock);
469         spin_lock(sdev->request_queue->queue_lock);
470         sdev->device_busy--;
471         spin_unlock_irqrestore(sdev->request_queue->queue_lock, flags);
472 }
473
474 /*
475  * Called for single_lun devices on IO completion. Clear starget_sdev_user,
476  * and call blk_run_queue for all the scsi_devices on the target -
477  * including current_sdev first.
478  *
479  * Called with *no* scsi locks held.
480  */
481 static void scsi_single_lun_run(struct scsi_device *current_sdev)
482 {
483         struct Scsi_Host *shost = current_sdev->host;
484         struct scsi_device *sdev, *tmp;
485         struct scsi_target *starget = scsi_target(current_sdev);
486         unsigned long flags;
487
488         spin_lock_irqsave(shost->host_lock, flags);
489         starget->starget_sdev_user = NULL;
490         spin_unlock_irqrestore(shost->host_lock, flags);
491
492         /*
493          * Call blk_run_queue for all LUNs on the target, starting with
494          * current_sdev. We race with others (to set starget_sdev_user),
495          * but in most cases, we will be first. Ideally, each LU on the
496          * target would get some limited time or requests on the target.
497          */
498         blk_run_queue(current_sdev->request_queue);
499
500         spin_lock_irqsave(shost->host_lock, flags);
501         if (starget->starget_sdev_user)
502                 goto out;
503         list_for_each_entry_safe(sdev, tmp, &starget->devices,
504                         same_target_siblings) {
505                 if (sdev == current_sdev)
506                         continue;
507                 if (scsi_device_get(sdev))
508                         continue;
509
510                 spin_unlock_irqrestore(shost->host_lock, flags);
511                 blk_run_queue(sdev->request_queue);
512                 spin_lock_irqsave(shost->host_lock, flags);
513         
514                 scsi_device_put(sdev);
515         }
516  out:
517         spin_unlock_irqrestore(shost->host_lock, flags);
518 }
519
520 /*
521  * Function:    scsi_run_queue()
522  *
523  * Purpose:     Select a proper request queue to serve next
524  *
525  * Arguments:   q       - last request's queue
526  *
527  * Returns:     Nothing
528  *
529  * Notes:       The previous command was completely finished, start
530  *              a new one if possible.
531  */
532 static void scsi_run_queue(struct request_queue *q)
533 {
534         struct scsi_device *sdev = q->queuedata;
535         struct Scsi_Host *shost = sdev->host;
536         unsigned long flags;
537
538         if (sdev->single_lun)
539                 scsi_single_lun_run(sdev);
540
541         spin_lock_irqsave(shost->host_lock, flags);
542         while (!list_empty(&shost->starved_list) &&
543                !shost->host_blocked && !shost->host_self_blocked &&
544                 !((shost->can_queue > 0) &&
545                   (shost->host_busy >= shost->can_queue))) {
546                 /*
547                  * As long as shost is accepting commands and we have
548                  * starved queues, call blk_run_queue. scsi_request_fn
549                  * drops the queue_lock and can add us back to the
550                  * starved_list.
551                  *
552                  * host_lock protects the starved_list and starved_entry.
553                  * scsi_request_fn must get the host_lock before checking
554                  * or modifying starved_list or starved_entry.
555                  */
556                 sdev = list_entry(shost->starved_list.next,
557                                           struct scsi_device, starved_entry);
558                 list_del_init(&sdev->starved_entry);
559                 spin_unlock_irqrestore(shost->host_lock, flags);
560
561
562                 if (test_bit(QUEUE_FLAG_REENTER, &q->queue_flags) &&
563                     !test_and_set_bit(QUEUE_FLAG_REENTER,
564                                       &sdev->request_queue->queue_flags)) {
565                         blk_run_queue(sdev->request_queue);
566                         clear_bit(QUEUE_FLAG_REENTER,
567                                   &sdev->request_queue->queue_flags);
568                 } else
569                         blk_run_queue(sdev->request_queue);
570
571                 spin_lock_irqsave(shost->host_lock, flags);
572                 if (unlikely(!list_empty(&sdev->starved_entry)))
573                         /*
574                          * sdev lost a race, and was put back on the
575                          * starved list. This is unlikely but without this
576                          * in theory we could loop forever.
577                          */
578                         break;
579         }
580         spin_unlock_irqrestore(shost->host_lock, flags);
581
582         blk_run_queue(q);
583 }
584
585 /*
586  * Function:    scsi_requeue_command()
587  *
588  * Purpose:     Handle post-processing of completed commands.
589  *
590  * Arguments:   q       - queue to operate on
591  *              cmd     - command that may need to be requeued.
592  *
593  * Returns:     Nothing
594  *
595  * Notes:       After command completion, there may be blocks left
596  *              over which weren't finished by the previous command
597  *              this can be for a number of reasons - the main one is
598  *              I/O errors in the middle of the request, in which case
599  *              we need to request the blocks that come after the bad
600  *              sector.
601  * Notes:       Upon return, cmd is a stale pointer.
602  */
603 static void scsi_requeue_command(struct request_queue *q, struct scsi_cmnd *cmd)
604 {
605         struct request *req = cmd->request;
606         unsigned long flags;
607
608         scsi_unprep_request(req);
609         spin_lock_irqsave(q->queue_lock, flags);
610         blk_requeue_request(q, req);
611         spin_unlock_irqrestore(q->queue_lock, flags);
612
613         scsi_run_queue(q);
614 }
615
616 void scsi_next_command(struct scsi_cmnd *cmd)
617 {
618         struct scsi_device *sdev = cmd->device;
619         struct request_queue *q = sdev->request_queue;
620
621         /* need to hold a reference on the device before we let go of the cmd */
622         get_device(&sdev->sdev_gendev);
623
624         scsi_put_command(cmd);
625         scsi_run_queue(q);
626
627         /* ok to remove device now */
628         put_device(&sdev->sdev_gendev);
629 }
630
631 void scsi_run_host_queues(struct Scsi_Host *shost)
632 {
633         struct scsi_device *sdev;
634
635         shost_for_each_device(sdev, shost)
636                 scsi_run_queue(sdev->request_queue);
637 }
638
639 /*
640  * Function:    scsi_end_request()
641  *
642  * Purpose:     Post-processing of completed commands (usually invoked at end
643  *              of upper level post-processing and scsi_io_completion).
644  *
645  * Arguments:   cmd      - command that is complete.
646  *              uptodate - 1 if I/O indicates success, <= 0 for I/O error.
647  *              bytes    - number of bytes of completed I/O
648  *              requeue  - indicates whether we should requeue leftovers.
649  *
650  * Lock status: Assumed that lock is not held upon entry.
651  *
652  * Returns:     cmd if requeue required, NULL otherwise.
653  *
654  * Notes:       This is called for block device requests in order to
655  *              mark some number of sectors as complete.
656  * 
657  *              We are guaranteeing that the request queue will be goosed
658  *              at some point during this call.
659  * Notes:       If cmd was requeued, upon return it will be a stale pointer.
660  */
661 static struct scsi_cmnd *scsi_end_request(struct scsi_cmnd *cmd, int uptodate,
662                                           int bytes, int requeue)
663 {
664         struct request_queue *q = cmd->device->request_queue;
665         struct request *req = cmd->request;
666         unsigned long flags;
667
668         /*
669          * If there are blocks left over at the end, set up the command
670          * to queue the remainder of them.
671          */
672         if (end_that_request_chunk(req, uptodate, bytes)) {
673                 int leftover = (req->hard_nr_sectors << 9);
674
675                 if (blk_pc_request(req))
676                         leftover = req->data_len;
677
678                 /* kill remainder if no retrys */
679                 if (!uptodate && blk_noretry_request(req))
680                         end_that_request_chunk(req, 0, leftover);
681                 else {
682                         if (requeue) {
683                                 /*
684                                  * Bleah.  Leftovers again.  Stick the
685                                  * leftovers in the front of the
686                                  * queue, and goose the queue again.
687                                  */
688                                 scsi_requeue_command(q, cmd);
689                                 cmd = NULL;
690                         }
691                         return cmd;
692                 }
693         }
694
695         add_disk_randomness(req->rq_disk);
696
697         spin_lock_irqsave(q->queue_lock, flags);
698         if (blk_rq_tagged(req))
699                 blk_queue_end_tag(q, req);
700         end_that_request_last(req, uptodate);
701         spin_unlock_irqrestore(q->queue_lock, flags);
702
703         /*
704          * This will goose the queue request function at the end, so we don't
705          * need to worry about launching another command.
706          */
707         scsi_next_command(cmd);
708         return NULL;
709 }
710
711 struct scatterlist *scsi_alloc_sgtable(struct scsi_cmnd *cmd, gfp_t gfp_mask)
712 {
713         struct scsi_host_sg_pool *sgp;
714         struct scatterlist *sgl;
715
716         BUG_ON(!cmd->use_sg);
717
718         switch (cmd->use_sg) {
719         case 1 ... 8:
720                 cmd->sglist_len = 0;
721                 break;
722         case 9 ... 16:
723                 cmd->sglist_len = 1;
724                 break;
725         case 17 ... 32:
726                 cmd->sglist_len = 2;
727                 break;
728 #if (SCSI_MAX_PHYS_SEGMENTS > 32)
729         case 33 ... 64:
730                 cmd->sglist_len = 3;
731                 break;
732 #if (SCSI_MAX_PHYS_SEGMENTS > 64)
733         case 65 ... 128:
734                 cmd->sglist_len = 4;
735                 break;
736 #if (SCSI_MAX_PHYS_SEGMENTS  > 128)
737         case 129 ... 256:
738                 cmd->sglist_len = 5;
739                 break;
740 #endif
741 #endif
742 #endif
743         default:
744                 return NULL;
745         }
746
747         sgp = scsi_sg_pools + cmd->sglist_len;
748         sgl = mempool_alloc(sgp->pool, gfp_mask);
749         return sgl;
750 }
751
752 EXPORT_SYMBOL(scsi_alloc_sgtable);
753
754 void scsi_free_sgtable(struct scatterlist *sgl, int index)
755 {
756         struct scsi_host_sg_pool *sgp;
757
758         BUG_ON(index >= SG_MEMPOOL_NR);
759
760         sgp = scsi_sg_pools + index;
761         mempool_free(sgl, sgp->pool);
762 }
763
764 EXPORT_SYMBOL(scsi_free_sgtable);
765
766 /*
767  * Function:    scsi_release_buffers()
768  *
769  * Purpose:     Completion processing for block device I/O requests.
770  *
771  * Arguments:   cmd     - command that we are bailing.
772  *
773  * Lock status: Assumed that no lock is held upon entry.
774  *
775  * Returns:     Nothing
776  *
777  * Notes:       In the event that an upper level driver rejects a
778  *              command, we must release resources allocated during
779  *              the __init_io() function.  Primarily this would involve
780  *              the scatter-gather table, and potentially any bounce
781  *              buffers.
782  */
783 static void scsi_release_buffers(struct scsi_cmnd *cmd)
784 {
785         if (cmd->use_sg)
786                 scsi_free_sgtable(cmd->request_buffer, cmd->sglist_len);
787
788         /*
789          * Zero these out.  They now point to freed memory, and it is
790          * dangerous to hang onto the pointers.
791          */
792         cmd->request_buffer = NULL;
793         cmd->request_bufflen = 0;
794 }
795
796 /*
797  * Function:    scsi_io_completion()
798  *
799  * Purpose:     Completion processing for block device I/O requests.
800  *
801  * Arguments:   cmd   - command that is finished.
802  *
803  * Lock status: Assumed that no lock is held upon entry.
804  *
805  * Returns:     Nothing
806  *
807  * Notes:       This function is matched in terms of capabilities to
808  *              the function that created the scatter-gather list.
809  *              In other words, if there are no bounce buffers
810  *              (the normal case for most drivers), we don't need
811  *              the logic to deal with cleaning up afterwards.
812  *
813  *              We must do one of several things here:
814  *
815  *              a) Call scsi_end_request.  This will finish off the
816  *                 specified number of sectors.  If we are done, the
817  *                 command block will be released, and the queue
818  *                 function will be goosed.  If we are not done, then
819  *                 scsi_end_request will directly goose the queue.
820  *
821  *              b) We can just use scsi_requeue_command() here.  This would
822  *                 be used if we just wanted to retry, for example.
823  */
824 void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes)
825 {
826         int result = cmd->result;
827         int this_count = cmd->request_bufflen;
828         struct request_queue *q = cmd->device->request_queue;
829         struct request *req = cmd->request;
830         int clear_errors = 1;
831         struct scsi_sense_hdr sshdr;
832         int sense_valid = 0;
833         int sense_deferred = 0;
834
835         scsi_release_buffers(cmd);
836
837         if (result) {
838                 sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
839                 if (sense_valid)
840                         sense_deferred = scsi_sense_is_deferred(&sshdr);
841         }
842
843         if (blk_pc_request(req)) { /* SG_IO ioctl from block level */
844                 req->errors = result;
845                 if (result) {
846                         clear_errors = 0;
847                         if (sense_valid && req->sense) {
848                                 /*
849                                  * SG_IO wants current and deferred errors
850                                  */
851                                 int len = 8 + cmd->sense_buffer[7];
852
853                                 if (len > SCSI_SENSE_BUFFERSIZE)
854                                         len = SCSI_SENSE_BUFFERSIZE;
855                                 memcpy(req->sense, cmd->sense_buffer,  len);
856                                 req->sense_len = len;
857                         }
858                 }
859                 req->data_len = cmd->resid;
860         }
861
862         /*
863          * Next deal with any sectors which we were able to correctly
864          * handle.
865          */
866         SCSI_LOG_HLCOMPLETE(1, printk("%ld sectors total, "
867                                       "%d bytes done.\n",
868                                       req->nr_sectors, good_bytes));
869         SCSI_LOG_HLCOMPLETE(1, printk("use_sg is %d\n", cmd->use_sg));
870
871         if (clear_errors)
872                 req->errors = 0;
873
874         /* A number of bytes were successfully read.  If there
875          * are leftovers and there is some kind of error
876          * (result != 0), retry the rest.
877          */
878         if (scsi_end_request(cmd, 1, good_bytes, result == 0) == NULL)
879                 return;
880
881         /* good_bytes = 0, or (inclusive) there were leftovers and
882          * result = 0, so scsi_end_request couldn't retry.
883          */
884         if (sense_valid && !sense_deferred) {
885                 switch (sshdr.sense_key) {
886                 case UNIT_ATTENTION:
887                         if (cmd->device->removable) {
888                                 /* Detected disc change.  Set a bit
889                                  * and quietly refuse further access.
890                                  */
891                                 cmd->device->changed = 1;
892                                 scsi_end_request(cmd, 0, this_count, 1);
893                                 return;
894                         } else {
895                                 /* Must have been a power glitch, or a
896                                  * bus reset.  Could not have been a
897                                  * media change, so we just retry the
898                                  * request and see what happens.
899                                  */
900                                 scsi_requeue_command(q, cmd);
901                                 return;
902                         }
903                         break;
904                 case ILLEGAL_REQUEST:
905                         /* If we had an ILLEGAL REQUEST returned, then
906                          * we may have performed an unsupported
907                          * command.  The only thing this should be
908                          * would be a ten byte read where only a six
909                          * byte read was supported.  Also, on a system
910                          * where READ CAPACITY failed, we may have
911                          * read past the end of the disk.
912                          */
913                         if ((cmd->device->use_10_for_rw &&
914                             sshdr.asc == 0x20 && sshdr.ascq == 0x00) &&
915                             (cmd->cmnd[0] == READ_10 ||
916                              cmd->cmnd[0] == WRITE_10)) {
917                                 cmd->device->use_10_for_rw = 0;
918                                 /* This will cause a retry with a
919                                  * 6-byte command.
920                                  */
921                                 scsi_requeue_command(q, cmd);
922                                 return;
923                         } else {
924                                 scsi_end_request(cmd, 0, this_count, 1);
925                                 return;
926                         }
927                         break;
928                 case NOT_READY:
929                         /* If the device is in the process of becoming
930                          * ready, or has a temporary blockage, retry.
931                          */
932                         if (sshdr.asc == 0x04) {
933                                 switch (sshdr.ascq) {
934                                 case 0x01: /* becoming ready */
935                                 case 0x04: /* format in progress */
936                                 case 0x05: /* rebuild in progress */
937                                 case 0x06: /* recalculation in progress */
938                                 case 0x07: /* operation in progress */
939                                 case 0x08: /* Long write in progress */
940                                 case 0x09: /* self test in progress */
941                                         scsi_requeue_command(q, cmd);
942                                         return;
943                                 default:
944                                         break;
945                                 }
946                         }
947                         if (!(req->cmd_flags & REQ_QUIET)) {
948                                 scmd_printk(KERN_INFO, cmd,
949                                             "Device not ready: ");
950                                 scsi_print_sense_hdr("", &sshdr);
951                         }
952                         scsi_end_request(cmd, 0, this_count, 1);
953                         return;
954                 case VOLUME_OVERFLOW:
955                         if (!(req->cmd_flags & REQ_QUIET)) {
956                                 scmd_printk(KERN_INFO, cmd,
957                                             "Volume overflow, CDB: ");
958                                 __scsi_print_command(cmd->cmnd);
959                                 scsi_print_sense("", cmd);
960                         }
961                         /* See SSC3rXX or current. */
962                         scsi_end_request(cmd, 0, this_count, 1);
963                         return;
964                 default:
965                         break;
966                 }
967         }
968         if (host_byte(result) == DID_RESET) {
969                 /* Third party bus reset or reset for error recovery
970                  * reasons.  Just retry the request and see what
971                  * happens.
972                  */
973                 scsi_requeue_command(q, cmd);
974                 return;
975         }
976         if (result) {
977                 if (!(req->cmd_flags & REQ_QUIET)) {
978                         scsi_print_result(cmd);
979                         if (driver_byte(result) & DRIVER_SENSE)
980                                 scsi_print_sense("", cmd);
981                 }
982         }
983         scsi_end_request(cmd, 0, this_count, !result);
984 }
985 EXPORT_SYMBOL(scsi_io_completion);
986
987 /*
988  * Function:    scsi_init_io()
989  *
990  * Purpose:     SCSI I/O initialize function.
991  *
992  * Arguments:   cmd   - Command descriptor we wish to initialize
993  *
994  * Returns:     0 on success
995  *              BLKPREP_DEFER if the failure is retryable
996  *              BLKPREP_KILL if the failure is fatal
997  */
998 static int scsi_init_io(struct scsi_cmnd *cmd)
999 {
1000         struct request     *req = cmd->request;
1001         struct scatterlist *sgpnt;
1002         int                count;
1003
1004         /*
1005          * We used to not use scatter-gather for single segment request,
1006          * but now we do (it makes highmem I/O easier to support without
1007          * kmapping pages)
1008          */
1009         cmd->use_sg = req->nr_phys_segments;
1010
1011         /*
1012          * If sg table allocation fails, requeue request later.
1013          */
1014         sgpnt = scsi_alloc_sgtable(cmd, GFP_ATOMIC);
1015         if (unlikely(!sgpnt)) {
1016                 scsi_unprep_request(req);
1017                 return BLKPREP_DEFER;
1018         }
1019
1020         req->buffer = NULL;
1021         cmd->request_buffer = (char *) sgpnt;
1022         if (blk_pc_request(req))
1023                 cmd->request_bufflen = req->data_len;
1024         else
1025                 cmd->request_bufflen = req->nr_sectors << 9;
1026
1027         /* 
1028          * Next, walk the list, and fill in the addresses and sizes of
1029          * each segment.
1030          */
1031         count = blk_rq_map_sg(req->q, req, cmd->request_buffer);
1032         if (likely(count <= cmd->use_sg)) {
1033                 cmd->use_sg = count;
1034                 return BLKPREP_OK;
1035         }
1036
1037         printk(KERN_ERR "Incorrect number of segments after building list\n");
1038         printk(KERN_ERR "counted %d, received %d\n", count, cmd->use_sg);
1039         printk(KERN_ERR "req nr_sec %lu, cur_nr_sec %u\n", req->nr_sectors,
1040                         req->current_nr_sectors);
1041
1042         /* release the command and kill it */
1043         scsi_release_buffers(cmd);
1044         scsi_put_command(cmd);
1045         return BLKPREP_KILL;
1046 }
1047
1048 static struct scsi_cmnd *scsi_get_cmd_from_req(struct scsi_device *sdev,
1049                 struct request *req)
1050 {
1051         struct scsi_cmnd *cmd;
1052
1053         if (!req->special) {
1054                 cmd = scsi_get_command(sdev, GFP_ATOMIC);
1055                 if (unlikely(!cmd))
1056                         return NULL;
1057                 req->special = cmd;
1058         } else {
1059                 cmd = req->special;
1060         }
1061
1062         /* pull a tag out of the request if we have one */
1063         cmd->tag = req->tag;
1064         cmd->request = req;
1065
1066         return cmd;
1067 }
1068
1069 static void scsi_blk_pc_done(struct scsi_cmnd *cmd)
1070 {
1071         BUG_ON(!blk_pc_request(cmd->request));
1072         /*
1073          * This will complete the whole command with uptodate=1 so
1074          * as far as the block layer is concerned the command completed
1075          * successfully. Since this is a REQ_BLOCK_PC command the
1076          * caller should check the request's errors value
1077          */
1078         scsi_io_completion(cmd, cmd->request_bufflen);
1079 }
1080
1081 static int scsi_setup_blk_pc_cmnd(struct scsi_device *sdev, struct request *req)
1082 {
1083         struct scsi_cmnd *cmd;
1084
1085         cmd = scsi_get_cmd_from_req(sdev, req);
1086         if (unlikely(!cmd))
1087                 return BLKPREP_DEFER;
1088
1089         /*
1090          * BLOCK_PC requests may transfer data, in which case they must
1091          * a bio attached to them.  Or they might contain a SCSI command
1092          * that does not transfer data, in which case they may optionally
1093          * submit a request without an attached bio.
1094          */
1095         if (req->bio) {
1096                 int ret;
1097
1098                 BUG_ON(!req->nr_phys_segments);
1099
1100                 ret = scsi_init_io(cmd);
1101                 if (unlikely(ret))
1102                         return ret;
1103         } else {
1104                 BUG_ON(req->data_len);
1105                 BUG_ON(req->data);
1106
1107                 cmd->request_bufflen = 0;
1108                 cmd->request_buffer = NULL;
1109                 cmd->use_sg = 0;
1110                 req->buffer = NULL;
1111         }
1112
1113         BUILD_BUG_ON(sizeof(req->cmd) > sizeof(cmd->cmnd));
1114         memcpy(cmd->cmnd, req->cmd, sizeof(cmd->cmnd));
1115         cmd->cmd_len = req->cmd_len;
1116         if (!req->data_len)
1117                 cmd->sc_data_direction = DMA_NONE;
1118         else if (rq_data_dir(req) == WRITE)
1119                 cmd->sc_data_direction = DMA_TO_DEVICE;
1120         else
1121                 cmd->sc_data_direction = DMA_FROM_DEVICE;
1122         
1123         cmd->transfersize = req->data_len;
1124         cmd->allowed = req->retries;
1125         cmd->timeout_per_command = req->timeout;
1126         cmd->done = scsi_blk_pc_done;
1127         return BLKPREP_OK;
1128 }
1129
1130 /*
1131  * Setup a REQ_TYPE_FS command.  These are simple read/write request
1132  * from filesystems that still need to be translated to SCSI CDBs from
1133  * the ULD.
1134  */
1135 static int scsi_setup_fs_cmnd(struct scsi_device *sdev, struct request *req)
1136 {
1137         struct scsi_cmnd *cmd;
1138         struct scsi_driver *drv;
1139         int ret;
1140
1141         /*
1142          * Filesystem requests must transfer data.
1143          */
1144         BUG_ON(!req->nr_phys_segments);
1145
1146         cmd = scsi_get_cmd_from_req(sdev, req);
1147         if (unlikely(!cmd))
1148                 return BLKPREP_DEFER;
1149
1150         ret = scsi_init_io(cmd);
1151         if (unlikely(ret))
1152                 return ret;
1153
1154         /*
1155          * Initialize the actual SCSI command for this request.
1156          */
1157         drv = *(struct scsi_driver **)req->rq_disk->private_data;
1158         if (unlikely(!drv->init_command(cmd))) {
1159                 scsi_release_buffers(cmd);
1160                 scsi_put_command(cmd);
1161                 return BLKPREP_KILL;
1162         }
1163
1164         return BLKPREP_OK;
1165 }
1166
1167 static int scsi_prep_fn(struct request_queue *q, struct request *req)
1168 {
1169         struct scsi_device *sdev = q->queuedata;
1170         int ret = BLKPREP_OK;
1171
1172         /*
1173          * If the device is not in running state we will reject some
1174          * or all commands.
1175          */
1176         if (unlikely(sdev->sdev_state != SDEV_RUNNING)) {
1177                 switch (sdev->sdev_state) {
1178                 case SDEV_OFFLINE:
1179                         /*
1180                          * If the device is offline we refuse to process any
1181                          * commands.  The device must be brought online
1182                          * before trying any recovery commands.
1183                          */
1184                         sdev_printk(KERN_ERR, sdev,
1185                                     "rejecting I/O to offline device\n");
1186                         ret = BLKPREP_KILL;
1187                         break;
1188                 case SDEV_DEL:
1189                         /*
1190                          * If the device is fully deleted, we refuse to
1191                          * process any commands as well.
1192                          */
1193                         sdev_printk(KERN_ERR, sdev,
1194                                     "rejecting I/O to dead device\n");
1195                         ret = BLKPREP_KILL;
1196                         break;
1197                 case SDEV_QUIESCE:
1198                 case SDEV_BLOCK:
1199                         /*
1200                          * If the devices is blocked we defer normal commands.
1201                          */
1202                         if (!(req->cmd_flags & REQ_PREEMPT))
1203                                 ret = BLKPREP_DEFER;
1204                         break;
1205                 default:
1206                         /*
1207                          * For any other not fully online state we only allow
1208                          * special commands.  In particular any user initiated
1209                          * command is not allowed.
1210                          */
1211                         if (!(req->cmd_flags & REQ_PREEMPT))
1212                                 ret = BLKPREP_KILL;
1213                         break;
1214                 }
1215
1216                 if (ret != BLKPREP_OK)
1217                         goto out;
1218         }
1219
1220         switch (req->cmd_type) {
1221         case REQ_TYPE_BLOCK_PC:
1222                 ret = scsi_setup_blk_pc_cmnd(sdev, req);
1223                 break;
1224         case REQ_TYPE_FS:
1225                 ret = scsi_setup_fs_cmnd(sdev, req);
1226                 break;
1227         default:
1228                 /*
1229                  * All other command types are not supported.
1230                  *
1231                  * Note that these days the SCSI subsystem does not use
1232                  * REQ_TYPE_SPECIAL requests anymore.  These are only used
1233                  * (directly or via blk_insert_request) by non-SCSI drivers.
1234                  */
1235                 blk_dump_rq_flags(req, "SCSI bad req");
1236                 ret = BLKPREP_KILL;
1237                 break;
1238         }
1239
1240  out:
1241         switch (ret) {
1242         case BLKPREP_KILL:
1243                 req->errors = DID_NO_CONNECT << 16;
1244                 break;
1245         case BLKPREP_DEFER:
1246                 /*
1247                  * If we defer, the elv_next_request() returns NULL, but the
1248                  * queue must be restarted, so we plug here if no returning
1249                  * command will automatically do that.
1250                  */
1251                 if (sdev->device_busy == 0)
1252                         blk_plug_device(q);
1253                 break;
1254         default:
1255                 req->cmd_flags |= REQ_DONTPREP;
1256         }
1257
1258         return ret;
1259 }
1260
1261 /*
1262  * scsi_dev_queue_ready: if we can send requests to sdev, return 1 else
1263  * return 0.
1264  *
1265  * Called with the queue_lock held.
1266  */
1267 static inline int scsi_dev_queue_ready(struct request_queue *q,
1268                                   struct scsi_device *sdev)
1269 {
1270         if (sdev->device_busy >= sdev->queue_depth)
1271                 return 0;
1272         if (sdev->device_busy == 0 && sdev->device_blocked) {
1273                 /*
1274                  * unblock after device_blocked iterates to zero
1275                  */
1276                 if (--sdev->device_blocked == 0) {
1277                         SCSI_LOG_MLQUEUE(3,
1278                                    sdev_printk(KERN_INFO, sdev,
1279                                    "unblocking device at zero depth\n"));
1280                 } else {
1281                         blk_plug_device(q);
1282                         return 0;
1283                 }
1284         }
1285         if (sdev->device_blocked)
1286                 return 0;
1287
1288         return 1;
1289 }
1290
1291 /*
1292  * scsi_host_queue_ready: if we can send requests to shost, return 1 else
1293  * return 0. We must end up running the queue again whenever 0 is
1294  * returned, else IO can hang.
1295  *
1296  * Called with host_lock held.
1297  */
1298 static inline int scsi_host_queue_ready(struct request_queue *q,
1299                                    struct Scsi_Host *shost,
1300                                    struct scsi_device *sdev)
1301 {
1302         if (scsi_host_in_recovery(shost))
1303                 return 0;
1304         if (shost->host_busy == 0 && shost->host_blocked) {
1305                 /*
1306                  * unblock after host_blocked iterates to zero
1307                  */
1308                 if (--shost->host_blocked == 0) {
1309                         SCSI_LOG_MLQUEUE(3,
1310                                 printk("scsi%d unblocking host at zero depth\n",
1311                                         shost->host_no));
1312                 } else {
1313                         blk_plug_device(q);
1314                         return 0;
1315                 }
1316         }
1317         if ((shost->can_queue > 0 && shost->host_busy >= shost->can_queue) ||
1318             shost->host_blocked || shost->host_self_blocked) {
1319                 if (list_empty(&sdev->starved_entry))
1320                         list_add_tail(&sdev->starved_entry, &shost->starved_list);
1321                 return 0;
1322         }
1323
1324         /* We're OK to process the command, so we can't be starved */
1325         if (!list_empty(&sdev->starved_entry))
1326                 list_del_init(&sdev->starved_entry);
1327
1328         return 1;
1329 }
1330
1331 /*
1332  * Kill a request for a dead device
1333  */
1334 static void scsi_kill_request(struct request *req, struct request_queue *q)
1335 {
1336         struct scsi_cmnd *cmd = req->special;
1337         struct scsi_device *sdev = cmd->device;
1338         struct Scsi_Host *shost = sdev->host;
1339
1340         blkdev_dequeue_request(req);
1341
1342         if (unlikely(cmd == NULL)) {
1343                 printk(KERN_CRIT "impossible request in %s.\n",
1344                                  __FUNCTION__);
1345                 BUG();
1346         }
1347
1348         scsi_init_cmd_errh(cmd);
1349         cmd->result = DID_NO_CONNECT << 16;
1350         atomic_inc(&cmd->device->iorequest_cnt);
1351
1352         /*
1353          * SCSI request completion path will do scsi_device_unbusy(),
1354          * bump busy counts.  To bump the counters, we need to dance
1355          * with the locks as normal issue path does.
1356          */
1357         sdev->device_busy++;
1358         spin_unlock(sdev->request_queue->queue_lock);
1359         spin_lock(shost->host_lock);
1360         shost->host_busy++;
1361         spin_unlock(shost->host_lock);
1362         spin_lock(sdev->request_queue->queue_lock);
1363
1364         __scsi_done(cmd);
1365 }
1366
1367 static void scsi_softirq_done(struct request *rq)
1368 {
1369         struct scsi_cmnd *cmd = rq->completion_data;
1370         unsigned long wait_for = (cmd->allowed + 1) * cmd->timeout_per_command;
1371         int disposition;
1372
1373         INIT_LIST_HEAD(&cmd->eh_entry);
1374
1375         disposition = scsi_decide_disposition(cmd);
1376         if (disposition != SUCCESS &&
1377             time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) {
1378                 sdev_printk(KERN_ERR, cmd->device,
1379                             "timing out command, waited %lus\n",
1380                             wait_for/HZ);
1381                 disposition = SUCCESS;
1382         }
1383                         
1384         scsi_log_completion(cmd, disposition);
1385
1386         switch (disposition) {
1387                 case SUCCESS:
1388                         scsi_finish_command(cmd);
1389                         break;
1390                 case NEEDS_RETRY:
1391                         scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY);
1392                         break;
1393                 case ADD_TO_MLQUEUE:
1394                         scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY);
1395                         break;
1396                 default:
1397                         if (!scsi_eh_scmd_add(cmd, 0))
1398                                 scsi_finish_command(cmd);
1399         }
1400 }
1401
1402 /*
1403  * Function:    scsi_request_fn()
1404  *
1405  * Purpose:     Main strategy routine for SCSI.
1406  *
1407  * Arguments:   q       - Pointer to actual queue.
1408  *
1409  * Returns:     Nothing
1410  *
1411  * Lock status: IO request lock assumed to be held when called.
1412  */
1413 static void scsi_request_fn(struct request_queue *q)
1414 {
1415         struct scsi_device *sdev = q->queuedata;
1416         struct Scsi_Host *shost;
1417         struct scsi_cmnd *cmd;
1418         struct request *req;
1419
1420         if (!sdev) {
1421                 printk("scsi: killing requests for dead queue\n");
1422                 while ((req = elv_next_request(q)) != NULL)
1423                         scsi_kill_request(req, q);
1424                 return;
1425         }
1426
1427         if(!get_device(&sdev->sdev_gendev))
1428                 /* We must be tearing the block queue down already */
1429                 return;
1430
1431         /*
1432          * To start with, we keep looping until the queue is empty, or until
1433          * the host is no longer able to accept any more requests.
1434          */
1435         shost = sdev->host;
1436         while (!blk_queue_plugged(q)) {
1437                 int rtn;
1438                 /*
1439                  * get next queueable request.  We do this early to make sure
1440                  * that the request is fully prepared even if we cannot 
1441                  * accept it.
1442                  */
1443                 req = elv_next_request(q);
1444                 if (!req || !scsi_dev_queue_ready(q, sdev))
1445                         break;
1446
1447                 if (unlikely(!scsi_device_online(sdev))) {
1448                         sdev_printk(KERN_ERR, sdev,
1449                                     "rejecting I/O to offline device\n");
1450                         scsi_kill_request(req, q);
1451                         continue;
1452                 }
1453
1454
1455                 /*
1456                  * Remove the request from the request list.
1457                  */
1458                 if (!(blk_queue_tagged(q) && !blk_queue_start_tag(q, req)))
1459                         blkdev_dequeue_request(req);
1460                 sdev->device_busy++;
1461
1462                 spin_unlock(q->queue_lock);
1463                 cmd = req->special;
1464                 if (unlikely(cmd == NULL)) {
1465                         printk(KERN_CRIT "impossible request in %s.\n"
1466                                          "please mail a stack trace to "
1467                                          "linux-scsi@vger.kernel.org\n",
1468                                          __FUNCTION__);
1469                         blk_dump_rq_flags(req, "foo");
1470                         BUG();
1471                 }
1472                 spin_lock(shost->host_lock);
1473
1474                 if (!scsi_host_queue_ready(q, shost, sdev))
1475                         goto not_ready;
1476                 if (sdev->single_lun) {
1477                         if (scsi_target(sdev)->starget_sdev_user &&
1478                             scsi_target(sdev)->starget_sdev_user != sdev)
1479                                 goto not_ready;
1480                         scsi_target(sdev)->starget_sdev_user = sdev;
1481                 }
1482                 shost->host_busy++;
1483
1484                 /*
1485                  * XXX(hch): This is rather suboptimal, scsi_dispatch_cmd will
1486                  *              take the lock again.
1487                  */
1488                 spin_unlock_irq(shost->host_lock);
1489
1490                 /*
1491                  * Finally, initialize any error handling parameters, and set up
1492                  * the timers for timeouts.
1493                  */
1494                 scsi_init_cmd_errh(cmd);
1495
1496                 /*
1497                  * Dispatch the command to the low-level driver.
1498                  */
1499                 rtn = scsi_dispatch_cmd(cmd);
1500                 spin_lock_irq(q->queue_lock);
1501                 if(rtn) {
1502                         /* we're refusing the command; because of
1503                          * the way locks get dropped, we need to 
1504                          * check here if plugging is required */
1505                         if(sdev->device_busy == 0)
1506                                 blk_plug_device(q);
1507
1508                         break;
1509                 }
1510         }
1511
1512         goto out;
1513
1514  not_ready:
1515         spin_unlock_irq(shost->host_lock);
1516
1517         /*
1518          * lock q, handle tag, requeue req, and decrement device_busy. We
1519          * must return with queue_lock held.
1520          *
1521          * Decrementing device_busy without checking it is OK, as all such
1522          * cases (host limits or settings) should run the queue at some
1523          * later time.
1524          */
1525         spin_lock_irq(q->queue_lock);
1526         blk_requeue_request(q, req);
1527         sdev->device_busy--;
1528         if(sdev->device_busy == 0)
1529                 blk_plug_device(q);
1530  out:
1531         /* must be careful here...if we trigger the ->remove() function
1532          * we cannot be holding the q lock */
1533         spin_unlock_irq(q->queue_lock);
1534         put_device(&sdev->sdev_gendev);
1535         spin_lock_irq(q->queue_lock);
1536 }
1537
1538 u64 scsi_calculate_bounce_limit(struct Scsi_Host *shost)
1539 {
1540         struct device *host_dev;
1541         u64 bounce_limit = 0xffffffff;
1542
1543         if (shost->unchecked_isa_dma)
1544                 return BLK_BOUNCE_ISA;
1545         /*
1546          * Platforms with virtual-DMA translation
1547          * hardware have no practical limit.
1548          */
1549         if (!PCI_DMA_BUS_IS_PHYS)
1550                 return BLK_BOUNCE_ANY;
1551
1552         host_dev = scsi_get_device(shost);
1553         if (host_dev && host_dev->dma_mask)
1554                 bounce_limit = *host_dev->dma_mask;
1555
1556         return bounce_limit;
1557 }
1558 EXPORT_SYMBOL(scsi_calculate_bounce_limit);
1559
1560 struct request_queue *__scsi_alloc_queue(struct Scsi_Host *shost,
1561                                          request_fn_proc *request_fn)
1562 {
1563         struct request_queue *q;
1564
1565         q = blk_init_queue(request_fn, NULL);
1566         if (!q)
1567                 return NULL;
1568
1569         blk_queue_max_hw_segments(q, shost->sg_tablesize);
1570         blk_queue_max_phys_segments(q, SCSI_MAX_PHYS_SEGMENTS);
1571         blk_queue_max_sectors(q, shost->max_sectors);
1572         blk_queue_bounce_limit(q, scsi_calculate_bounce_limit(shost));
1573         blk_queue_segment_boundary(q, shost->dma_boundary);
1574
1575         if (!shost->use_clustering)
1576                 clear_bit(QUEUE_FLAG_CLUSTER, &q->queue_flags);
1577         return q;
1578 }
1579 EXPORT_SYMBOL(__scsi_alloc_queue);
1580
1581 struct request_queue *scsi_alloc_queue(struct scsi_device *sdev)
1582 {
1583         struct request_queue *q;
1584
1585         q = __scsi_alloc_queue(sdev->host, scsi_request_fn);
1586         if (!q)
1587                 return NULL;
1588
1589         blk_queue_prep_rq(q, scsi_prep_fn);
1590         blk_queue_softirq_done(q, scsi_softirq_done);
1591         return q;
1592 }
1593
1594 void scsi_free_queue(struct request_queue *q)
1595 {
1596         blk_cleanup_queue(q);
1597 }
1598
1599 /*
1600  * Function:    scsi_block_requests()
1601  *
1602  * Purpose:     Utility function used by low-level drivers to prevent further
1603  *              commands from being queued to the device.
1604  *
1605  * Arguments:   shost       - Host in question
1606  *
1607  * Returns:     Nothing
1608  *
1609  * Lock status: No locks are assumed held.
1610  *
1611  * Notes:       There is no timer nor any other means by which the requests
1612  *              get unblocked other than the low-level driver calling
1613  *              scsi_unblock_requests().
1614  */
1615 void scsi_block_requests(struct Scsi_Host *shost)
1616 {
1617         shost->host_self_blocked = 1;
1618 }
1619 EXPORT_SYMBOL(scsi_block_requests);
1620
1621 /*
1622  * Function:    scsi_unblock_requests()
1623  *
1624  * Purpose:     Utility function used by low-level drivers to allow further
1625  *              commands from being queued to the device.
1626  *
1627  * Arguments:   shost       - Host in question
1628  *
1629  * Returns:     Nothing
1630  *
1631  * Lock status: No locks are assumed held.
1632  *
1633  * Notes:       There is no timer nor any other means by which the requests
1634  *              get unblocked other than the low-level driver calling
1635  *              scsi_unblock_requests().
1636  *
1637  *              This is done as an API function so that changes to the
1638  *              internals of the scsi mid-layer won't require wholesale
1639  *              changes to drivers that use this feature.
1640  */
1641 void scsi_unblock_requests(struct Scsi_Host *shost)
1642 {
1643         shost->host_self_blocked = 0;
1644         scsi_run_host_queues(shost);
1645 }
1646 EXPORT_SYMBOL(scsi_unblock_requests);
1647
1648 int __init scsi_init_queue(void)
1649 {
1650         int i;
1651
1652         scsi_io_context_cache = kmem_cache_create("scsi_io_context",
1653                                         sizeof(struct scsi_io_context),
1654                                         0, 0, NULL);
1655         if (!scsi_io_context_cache) {
1656                 printk(KERN_ERR "SCSI: can't init scsi io context cache\n");
1657                 return -ENOMEM;
1658         }
1659
1660         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1661                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1662                 int size = sgp->size * sizeof(struct scatterlist);
1663
1664                 sgp->slab = kmem_cache_create(sgp->name, size, 0,
1665                                 SLAB_HWCACHE_ALIGN, NULL);
1666                 if (!sgp->slab) {
1667                         printk(KERN_ERR "SCSI: can't init sg slab %s\n",
1668                                         sgp->name);
1669                 }
1670
1671                 sgp->pool = mempool_create_slab_pool(SG_MEMPOOL_SIZE,
1672                                                      sgp->slab);
1673                 if (!sgp->pool) {
1674                         printk(KERN_ERR "SCSI: can't init sg mempool %s\n",
1675                                         sgp->name);
1676                 }
1677         }
1678
1679         return 0;
1680 }
1681
1682 void scsi_exit_queue(void)
1683 {
1684         int i;
1685
1686         kmem_cache_destroy(scsi_io_context_cache);
1687
1688         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1689                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1690                 mempool_destroy(sgp->pool);
1691                 kmem_cache_destroy(sgp->slab);
1692         }
1693 }
1694
1695 /**
1696  *      scsi_mode_select - issue a mode select
1697  *      @sdev:  SCSI device to be queried
1698  *      @pf:    Page format bit (1 == standard, 0 == vendor specific)
1699  *      @sp:    Save page bit (0 == don't save, 1 == save)
1700  *      @modepage: mode page being requested
1701  *      @buffer: request buffer (may not be smaller than eight bytes)
1702  *      @len:   length of request buffer.
1703  *      @timeout: command timeout
1704  *      @retries: number of retries before failing
1705  *      @data: returns a structure abstracting the mode header data
1706  *      @sense: place to put sense data (or NULL if no sense to be collected).
1707  *              must be SCSI_SENSE_BUFFERSIZE big.
1708  *
1709  *      Returns zero if successful; negative error number or scsi
1710  *      status on error
1711  *
1712  */
1713 int
1714 scsi_mode_select(struct scsi_device *sdev, int pf, int sp, int modepage,
1715                  unsigned char *buffer, int len, int timeout, int retries,
1716                  struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1717 {
1718         unsigned char cmd[10];
1719         unsigned char *real_buffer;
1720         int ret;
1721
1722         memset(cmd, 0, sizeof(cmd));
1723         cmd[1] = (pf ? 0x10 : 0) | (sp ? 0x01 : 0);
1724
1725         if (sdev->use_10_for_ms) {
1726                 if (len > 65535)
1727                         return -EINVAL;
1728                 real_buffer = kmalloc(8 + len, GFP_KERNEL);
1729                 if (!real_buffer)
1730                         return -ENOMEM;
1731                 memcpy(real_buffer + 8, buffer, len);
1732                 len += 8;
1733                 real_buffer[0] = 0;
1734                 real_buffer[1] = 0;
1735                 real_buffer[2] = data->medium_type;
1736                 real_buffer[3] = data->device_specific;
1737                 real_buffer[4] = data->longlba ? 0x01 : 0;
1738                 real_buffer[5] = 0;
1739                 real_buffer[6] = data->block_descriptor_length >> 8;
1740                 real_buffer[7] = data->block_descriptor_length;
1741
1742                 cmd[0] = MODE_SELECT_10;
1743                 cmd[7] = len >> 8;
1744                 cmd[8] = len;
1745         } else {
1746                 if (len > 255 || data->block_descriptor_length > 255 ||
1747                     data->longlba)
1748                         return -EINVAL;
1749
1750                 real_buffer = kmalloc(4 + len, GFP_KERNEL);
1751                 if (!real_buffer)
1752                         return -ENOMEM;
1753                 memcpy(real_buffer + 4, buffer, len);
1754                 len += 4;
1755                 real_buffer[0] = 0;
1756                 real_buffer[1] = data->medium_type;
1757                 real_buffer[2] = data->device_specific;
1758                 real_buffer[3] = data->block_descriptor_length;
1759                 
1760
1761                 cmd[0] = MODE_SELECT;
1762                 cmd[4] = len;
1763         }
1764
1765         ret = scsi_execute_req(sdev, cmd, DMA_TO_DEVICE, real_buffer, len,
1766                                sshdr, timeout, retries);
1767         kfree(real_buffer);
1768         return ret;
1769 }
1770 EXPORT_SYMBOL_GPL(scsi_mode_select);
1771
1772 /**
1773  *      scsi_mode_sense - issue a mode sense, falling back from 10 to 
1774  *              six bytes if necessary.
1775  *      @sdev:  SCSI device to be queried
1776  *      @dbd:   set if mode sense will allow block descriptors to be returned
1777  *      @modepage: mode page being requested
1778  *      @buffer: request buffer (may not be smaller than eight bytes)
1779  *      @len:   length of request buffer.
1780  *      @timeout: command timeout
1781  *      @retries: number of retries before failing
1782  *      @data: returns a structure abstracting the mode header data
1783  *      @sense: place to put sense data (or NULL if no sense to be collected).
1784  *              must be SCSI_SENSE_BUFFERSIZE big.
1785  *
1786  *      Returns zero if unsuccessful, or the header offset (either 4
1787  *      or 8 depending on whether a six or ten byte command was
1788  *      issued) if successful.
1789  **/
1790 int
1791 scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
1792                   unsigned char *buffer, int len, int timeout, int retries,
1793                   struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1794 {
1795         unsigned char cmd[12];
1796         int use_10_for_ms;
1797         int header_length;
1798         int result;
1799         struct scsi_sense_hdr my_sshdr;
1800
1801         memset(data, 0, sizeof(*data));
1802         memset(&cmd[0], 0, 12);
1803         cmd[1] = dbd & 0x18;    /* allows DBD and LLBA bits */
1804         cmd[2] = modepage;
1805
1806         /* caller might not be interested in sense, but we need it */
1807         if (!sshdr)
1808                 sshdr = &my_sshdr;
1809
1810  retry:
1811         use_10_for_ms = sdev->use_10_for_ms;
1812
1813         if (use_10_for_ms) {
1814                 if (len < 8)
1815                         len = 8;
1816
1817                 cmd[0] = MODE_SENSE_10;
1818                 cmd[8] = len;
1819                 header_length = 8;
1820         } else {
1821                 if (len < 4)
1822                         len = 4;
1823
1824                 cmd[0] = MODE_SENSE;
1825                 cmd[4] = len;
1826                 header_length = 4;
1827         }
1828
1829         memset(buffer, 0, len);
1830
1831         result = scsi_execute_req(sdev, cmd, DMA_FROM_DEVICE, buffer, len,
1832                                   sshdr, timeout, retries);
1833
1834         /* This code looks awful: what it's doing is making sure an
1835          * ILLEGAL REQUEST sense return identifies the actual command
1836          * byte as the problem.  MODE_SENSE commands can return
1837          * ILLEGAL REQUEST if the code page isn't supported */
1838
1839         if (use_10_for_ms && !scsi_status_is_good(result) &&
1840             (driver_byte(result) & DRIVER_SENSE)) {
1841                 if (scsi_sense_valid(sshdr)) {
1842                         if ((sshdr->sense_key == ILLEGAL_REQUEST) &&
1843                             (sshdr->asc == 0x20) && (sshdr->ascq == 0)) {
1844                                 /* 
1845                                  * Invalid command operation code
1846                                  */
1847                                 sdev->use_10_for_ms = 0;
1848                                 goto retry;
1849                         }
1850                 }
1851         }
1852
1853         if(scsi_status_is_good(result)) {
1854                 if (unlikely(buffer[0] == 0x86 && buffer[1] == 0x0b &&
1855                              (modepage == 6 || modepage == 8))) {
1856                         /* Initio breakage? */
1857                         header_length = 0;
1858                         data->length = 13;
1859                         data->medium_type = 0;
1860                         data->device_specific = 0;
1861                         data->longlba = 0;
1862                         data->block_descriptor_length = 0;
1863                 } else if(use_10_for_ms) {
1864                         data->length = buffer[0]*256 + buffer[1] + 2;
1865                         data->medium_type = buffer[2];
1866                         data->device_specific = buffer[3];
1867                         data->longlba = buffer[4] & 0x01;
1868                         data->block_descriptor_length = buffer[6]*256
1869                                 + buffer[7];
1870                 } else {
1871                         data->length = buffer[0] + 1;
1872                         data->medium_type = buffer[1];
1873                         data->device_specific = buffer[2];
1874                         data->block_descriptor_length = buffer[3];
1875                 }
1876                 data->header_length = header_length;
1877         }
1878
1879         return result;
1880 }
1881 EXPORT_SYMBOL(scsi_mode_sense);
1882
1883 int
1884 scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries)
1885 {
1886         char cmd[] = {
1887                 TEST_UNIT_READY, 0, 0, 0, 0, 0,
1888         };
1889         struct scsi_sense_hdr sshdr;
1890         int result;
1891         
1892         result = scsi_execute_req(sdev, cmd, DMA_NONE, NULL, 0, &sshdr,
1893                                   timeout, retries);
1894
1895         if ((driver_byte(result) & DRIVER_SENSE) && sdev->removable) {
1896
1897                 if ((scsi_sense_valid(&sshdr)) &&
1898                     ((sshdr.sense_key == UNIT_ATTENTION) ||
1899                      (sshdr.sense_key == NOT_READY))) {
1900                         sdev->changed = 1;
1901                         result = 0;
1902                 }
1903         }
1904         return result;
1905 }
1906 EXPORT_SYMBOL(scsi_test_unit_ready);
1907
1908 /**
1909  *      scsi_device_set_state - Take the given device through the device
1910  *              state model.
1911  *      @sdev:  scsi device to change the state of.
1912  *      @state: state to change to.
1913  *
1914  *      Returns zero if unsuccessful or an error if the requested 
1915  *      transition is illegal.
1916  **/
1917 int
1918 scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
1919 {
1920         enum scsi_device_state oldstate = sdev->sdev_state;
1921
1922         if (state == oldstate)
1923                 return 0;
1924
1925         switch (state) {
1926         case SDEV_CREATED:
1927                 /* There are no legal states that come back to
1928                  * created.  This is the manually initialised start
1929                  * state */
1930                 goto illegal;
1931                         
1932         case SDEV_RUNNING:
1933                 switch (oldstate) {
1934                 case SDEV_CREATED:
1935                 case SDEV_OFFLINE:
1936                 case SDEV_QUIESCE:
1937                 case SDEV_BLOCK:
1938                         break;
1939                 default:
1940                         goto illegal;
1941                 }
1942                 break;
1943
1944         case SDEV_QUIESCE:
1945                 switch (oldstate) {
1946                 case SDEV_RUNNING:
1947                 case SDEV_OFFLINE:
1948                         break;
1949                 default:
1950                         goto illegal;
1951                 }
1952                 break;
1953
1954         case SDEV_OFFLINE:
1955                 switch (oldstate) {
1956                 case SDEV_CREATED:
1957                 case SDEV_RUNNING:
1958                 case SDEV_QUIESCE:
1959                 case SDEV_BLOCK:
1960                         break;
1961                 default:
1962                         goto illegal;
1963                 }
1964                 break;
1965
1966         case SDEV_BLOCK:
1967                 switch (oldstate) {
1968                 case SDEV_CREATED:
1969                 case SDEV_RUNNING:
1970                         break;
1971                 default:
1972                         goto illegal;
1973                 }
1974                 break;
1975
1976         case SDEV_CANCEL:
1977                 switch (oldstate) {
1978                 case SDEV_CREATED:
1979                 case SDEV_RUNNING:
1980                 case SDEV_QUIESCE:
1981                 case SDEV_OFFLINE:
1982                 case SDEV_BLOCK:
1983                         break;
1984                 default:
1985                         goto illegal;
1986                 }
1987                 break;
1988
1989         case SDEV_DEL:
1990                 switch (oldstate) {
1991                 case SDEV_CREATED:
1992                 case SDEV_RUNNING:
1993                 case SDEV_OFFLINE:
1994                 case SDEV_CANCEL:
1995                         break;
1996                 default:
1997                         goto illegal;
1998                 }
1999                 break;
2000
2001         }
2002         sdev->sdev_state = state;
2003         return 0;
2004
2005  illegal:
2006         SCSI_LOG_ERROR_RECOVERY(1, 
2007                                 sdev_printk(KERN_ERR, sdev,
2008                                             "Illegal state transition %s->%s\n",
2009                                             scsi_device_state_name(oldstate),
2010                                             scsi_device_state_name(state))
2011                                 );
2012         return -EINVAL;
2013 }
2014 EXPORT_SYMBOL(scsi_device_set_state);
2015
2016 /**
2017  *      scsi_device_quiesce - Block user issued commands.
2018  *      @sdev:  scsi device to quiesce.
2019  *
2020  *      This works by trying to transition to the SDEV_QUIESCE state
2021  *      (which must be a legal transition).  When the device is in this
2022  *      state, only special requests will be accepted, all others will
2023  *      be deferred.  Since special requests may also be requeued requests,
2024  *      a successful return doesn't guarantee the device will be 
2025  *      totally quiescent.
2026  *
2027  *      Must be called with user context, may sleep.
2028  *
2029  *      Returns zero if unsuccessful or an error if not.
2030  **/
2031 int
2032 scsi_device_quiesce(struct scsi_device *sdev)
2033 {
2034         int err = scsi_device_set_state(sdev, SDEV_QUIESCE);
2035         if (err)
2036                 return err;
2037
2038         scsi_run_queue(sdev->request_queue);
2039         while (sdev->device_busy) {
2040                 msleep_interruptible(200);
2041                 scsi_run_queue(sdev->request_queue);
2042         }
2043         return 0;
2044 }
2045 EXPORT_SYMBOL(scsi_device_quiesce);
2046
2047 /**
2048  *      scsi_device_resume - Restart user issued commands to a quiesced device.
2049  *      @sdev:  scsi device to resume.
2050  *
2051  *      Moves the device from quiesced back to running and restarts the
2052  *      queues.
2053  *
2054  *      Must be called with user context, may sleep.
2055  **/
2056 void
2057 scsi_device_resume(struct scsi_device *sdev)
2058 {
2059         if(scsi_device_set_state(sdev, SDEV_RUNNING))
2060                 return;
2061         scsi_run_queue(sdev->request_queue);
2062 }
2063 EXPORT_SYMBOL(scsi_device_resume);
2064
2065 static void
2066 device_quiesce_fn(struct scsi_device *sdev, void *data)
2067 {
2068         scsi_device_quiesce(sdev);
2069 }
2070
2071 void
2072 scsi_target_quiesce(struct scsi_target *starget)
2073 {
2074         starget_for_each_device(starget, NULL, device_quiesce_fn);
2075 }
2076 EXPORT_SYMBOL(scsi_target_quiesce);
2077
2078 static void
2079 device_resume_fn(struct scsi_device *sdev, void *data)
2080 {
2081         scsi_device_resume(sdev);
2082 }
2083
2084 void
2085 scsi_target_resume(struct scsi_target *starget)
2086 {
2087         starget_for_each_device(starget, NULL, device_resume_fn);
2088 }
2089 EXPORT_SYMBOL(scsi_target_resume);
2090
2091 /**
2092  * scsi_internal_device_block - internal function to put a device
2093  *                              temporarily into the SDEV_BLOCK state
2094  * @sdev:       device to block
2095  *
2096  * Block request made by scsi lld's to temporarily stop all
2097  * scsi commands on the specified device.  Called from interrupt
2098  * or normal process context.
2099  *
2100  * Returns zero if successful or error if not
2101  *
2102  * Notes:       
2103  *      This routine transitions the device to the SDEV_BLOCK state
2104  *      (which must be a legal transition).  When the device is in this
2105  *      state, all commands are deferred until the scsi lld reenables
2106  *      the device with scsi_device_unblock or device_block_tmo fires.
2107  *      This routine assumes the host_lock is held on entry.
2108  **/
2109 int
2110 scsi_internal_device_block(struct scsi_device *sdev)
2111 {
2112         struct request_queue *q = sdev->request_queue;
2113         unsigned long flags;
2114         int err = 0;
2115
2116         err = scsi_device_set_state(sdev, SDEV_BLOCK);
2117         if (err)
2118                 return err;
2119
2120         /* 
2121          * The device has transitioned to SDEV_BLOCK.  Stop the
2122          * block layer from calling the midlayer with this device's
2123          * request queue. 
2124          */
2125         spin_lock_irqsave(q->queue_lock, flags);
2126         blk_stop_queue(q);
2127         spin_unlock_irqrestore(q->queue_lock, flags);
2128
2129         return 0;
2130 }
2131 EXPORT_SYMBOL_GPL(scsi_internal_device_block);
2132  
2133 /**
2134  * scsi_internal_device_unblock - resume a device after a block request
2135  * @sdev:       device to resume
2136  *
2137  * Called by scsi lld's or the midlayer to restart the device queue
2138  * for the previously suspended scsi device.  Called from interrupt or
2139  * normal process context.
2140  *
2141  * Returns zero if successful or error if not.
2142  *
2143  * Notes:       
2144  *      This routine transitions the device to the SDEV_RUNNING state
2145  *      (which must be a legal transition) allowing the midlayer to
2146  *      goose the queue for this device.  This routine assumes the 
2147  *      host_lock is held upon entry.
2148  **/
2149 int
2150 scsi_internal_device_unblock(struct scsi_device *sdev)
2151 {
2152         struct request_queue *q = sdev->request_queue; 
2153         int err;
2154         unsigned long flags;
2155         
2156         /* 
2157          * Try to transition the scsi device to SDEV_RUNNING
2158          * and goose the device queue if successful.  
2159          */
2160         err = scsi_device_set_state(sdev, SDEV_RUNNING);
2161         if (err)
2162                 return err;
2163
2164         spin_lock_irqsave(q->queue_lock, flags);
2165         blk_start_queue(q);
2166         spin_unlock_irqrestore(q->queue_lock, flags);
2167
2168         return 0;
2169 }
2170 EXPORT_SYMBOL_GPL(scsi_internal_device_unblock);
2171
2172 static void
2173 device_block(struct scsi_device *sdev, void *data)
2174 {
2175         scsi_internal_device_block(sdev);
2176 }
2177
2178 static int
2179 target_block(struct device *dev, void *data)
2180 {
2181         if (scsi_is_target_device(dev))
2182                 starget_for_each_device(to_scsi_target(dev), NULL,
2183                                         device_block);
2184         return 0;
2185 }
2186
2187 void
2188 scsi_target_block(struct device *dev)
2189 {
2190         if (scsi_is_target_device(dev))
2191                 starget_for_each_device(to_scsi_target(dev), NULL,
2192                                         device_block);
2193         else
2194                 device_for_each_child(dev, NULL, target_block);
2195 }
2196 EXPORT_SYMBOL_GPL(scsi_target_block);
2197
2198 static void
2199 device_unblock(struct scsi_device *sdev, void *data)
2200 {
2201         scsi_internal_device_unblock(sdev);
2202 }
2203
2204 static int
2205 target_unblock(struct device *dev, void *data)
2206 {
2207         if (scsi_is_target_device(dev))
2208                 starget_for_each_device(to_scsi_target(dev), NULL,
2209                                         device_unblock);
2210         return 0;
2211 }
2212
2213 void
2214 scsi_target_unblock(struct device *dev)
2215 {
2216         if (scsi_is_target_device(dev))
2217                 starget_for_each_device(to_scsi_target(dev), NULL,
2218                                         device_unblock);
2219         else
2220                 device_for_each_child(dev, NULL, target_unblock);
2221 }
2222 EXPORT_SYMBOL_GPL(scsi_target_unblock);
2223
2224 /**
2225  * scsi_kmap_atomic_sg - find and atomically map an sg-elemnt
2226  * @sg:         scatter-gather list
2227  * @sg_count:   number of segments in sg
2228  * @offset:     offset in bytes into sg, on return offset into the mapped area
2229  * @len:        bytes to map, on return number of bytes mapped
2230  *
2231  * Returns virtual address of the start of the mapped page
2232  */
2233 void *scsi_kmap_atomic_sg(struct scatterlist *sg, int sg_count,
2234                           size_t *offset, size_t *len)
2235 {
2236         int i;
2237         size_t sg_len = 0, len_complete = 0;
2238         struct page *page;
2239
2240         WARN_ON(!irqs_disabled());
2241
2242         for (i = 0; i < sg_count; i++) {
2243                 len_complete = sg_len; /* Complete sg-entries */
2244                 sg_len += sg[i].length;
2245                 if (sg_len > *offset)
2246                         break;
2247         }
2248
2249         if (unlikely(i == sg_count)) {
2250                 printk(KERN_ERR "%s: Bytes in sg: %zu, requested offset %zu, "
2251                         "elements %d\n",
2252                        __FUNCTION__, sg_len, *offset, sg_count);
2253                 WARN_ON(1);
2254                 return NULL;
2255         }
2256
2257         /* Offset starting from the beginning of first page in this sg-entry */
2258         *offset = *offset - len_complete + sg[i].offset;
2259
2260         /* Assumption: contiguous pages can be accessed as "page + i" */
2261         page = nth_page(sg[i].page, (*offset >> PAGE_SHIFT));
2262         *offset &= ~PAGE_MASK;
2263
2264         /* Bytes in this sg-entry from *offset to the end of the page */
2265         sg_len = PAGE_SIZE - *offset;
2266         if (*len > sg_len)
2267                 *len = sg_len;
2268
2269         return kmap_atomic(page, KM_BIO_SRC_IRQ);
2270 }
2271 EXPORT_SYMBOL(scsi_kmap_atomic_sg);
2272
2273 /**
2274  * scsi_kunmap_atomic_sg - atomically unmap a virtual address, previously
2275  *                         mapped with scsi_kmap_atomic_sg
2276  * @virt:       virtual address to be unmapped
2277  */
2278 void scsi_kunmap_atomic_sg(void *virt)
2279 {
2280         kunmap_atomic(virt, KM_BIO_SRC_IRQ);
2281 }
2282 EXPORT_SYMBOL(scsi_kunmap_atomic_sg);