Merge branch 'master' into upstream-fixes
[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         32
35
36 struct scsi_host_sg_pool {
37         size_t          size;
38         char            *name; 
39         kmem_cache_t    *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 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 kmem_cache_t *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 (!q->back_merge_fn(q, rq, bio))
269                 return -EINVAL;
270         else {
271                 rq->biotail->bi_next = bio;
272                 rq->biotail = bio;
273                 rq->hard_nr_sectors += bio_sectors(bio);
274                 rq->nr_sectors = rq->hard_nr_sectors;
275         }
276
277         return 0;
278 }
279
280 static int scsi_bi_endio(struct bio *bio, unsigned int bytes_done, int error)
281 {
282         if (bio->bi_size)
283                 return 1;
284
285         bio_put(bio);
286         return 0;
287 }
288
289 /**
290  * scsi_req_map_sg - map a scatterlist into a request
291  * @rq:         request to fill
292  * @sg:         scatterlist
293  * @nsegs:      number of elements
294  * @bufflen:    len of buffer
295  * @gfp:        memory allocation flags
296  *
297  * scsi_req_map_sg maps a scatterlist into a request so that the
298  * request can be sent to the block layer. We do not trust the scatterlist
299  * sent to use, as some ULDs use that struct to only organize the pages.
300  */
301 static int scsi_req_map_sg(struct request *rq, struct scatterlist *sgl,
302                            int nsegs, unsigned bufflen, gfp_t gfp)
303 {
304         struct request_queue *q = rq->q;
305         int nr_pages = (bufflen + sgl[0].offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
306         unsigned int data_len = 0, len, bytes, off;
307         struct page *page;
308         struct bio *bio = NULL;
309         int i, err, nr_vecs = 0;
310
311         for (i = 0; i < nsegs; i++) {
312                 page = sgl[i].page;
313                 off = sgl[i].offset;
314                 len = sgl[i].length;
315                 data_len += len;
316
317                 while (len > 0) {
318                         bytes = min_t(unsigned int, len, PAGE_SIZE - off);
319
320                         if (!bio) {
321                                 nr_vecs = min_t(int, BIO_MAX_PAGES, nr_pages);
322                                 nr_pages -= nr_vecs;
323
324                                 bio = bio_alloc(gfp, nr_vecs);
325                                 if (!bio) {
326                                         err = -ENOMEM;
327                                         goto free_bios;
328                                 }
329                                 bio->bi_end_io = scsi_bi_endio;
330                         }
331
332                         if (bio_add_pc_page(q, bio, page, bytes, off) !=
333                             bytes) {
334                                 bio_put(bio);
335                                 err = -EINVAL;
336                                 goto free_bios;
337                         }
338
339                         if (bio->bi_vcnt >= nr_vecs) {
340                                 err = scsi_merge_bio(rq, bio);
341                                 if (err) {
342                                         bio_endio(bio, bio->bi_size, 0);
343                                         goto free_bios;
344                                 }
345                                 bio = NULL;
346                         }
347
348                         page++;
349                         len -= bytes;
350                         off = 0;
351                 }
352         }
353
354         rq->buffer = rq->data = NULL;
355         rq->data_len = data_len;
356         return 0;
357
358 free_bios:
359         while ((bio = rq->bio) != NULL) {
360                 rq->bio = bio->bi_next;
361                 /*
362                  * call endio instead of bio_put incase it was bounced
363                  */
364                 bio_endio(bio, bio->bi_size, 0);
365         }
366
367         return err;
368 }
369
370 /**
371  * scsi_execute_async - insert request
372  * @sdev:       scsi device
373  * @cmd:        scsi command
374  * @cmd_len:    length of scsi cdb
375  * @data_direction: data direction
376  * @buffer:     data buffer (this can be a kernel buffer or scatterlist)
377  * @bufflen:    len of buffer
378  * @use_sg:     if buffer is a scatterlist this is the number of elements
379  * @timeout:    request timeout in seconds
380  * @retries:    number of times to retry request
381  * @flags:      or into request flags
382  **/
383 int scsi_execute_async(struct scsi_device *sdev, const unsigned char *cmd,
384                        int cmd_len, int data_direction, void *buffer, unsigned bufflen,
385                        int use_sg, int timeout, int retries, void *privdata,
386                        void (*done)(void *, char *, int, int), gfp_t gfp)
387 {
388         struct request *req;
389         struct scsi_io_context *sioc;
390         int err = 0;
391         int write = (data_direction == DMA_TO_DEVICE);
392
393         sioc = kmem_cache_alloc(scsi_io_context_cache, gfp);
394         if (!sioc)
395                 return DRIVER_ERROR << 24;
396         memset(sioc, 0, sizeof(*sioc));
397
398         req = blk_get_request(sdev->request_queue, write, gfp);
399         if (!req)
400                 goto free_sense;
401         req->cmd_type = REQ_TYPE_BLOCK_PC;
402         req->cmd_flags |= REQ_QUIET;
403
404         if (use_sg)
405                 err = scsi_req_map_sg(req, buffer, use_sg, bufflen, gfp);
406         else if (bufflen)
407                 err = blk_rq_map_kern(req->q, req, buffer, bufflen, gfp);
408
409         if (err)
410                 goto free_req;
411
412         req->cmd_len = cmd_len;
413         memcpy(req->cmd, cmd, req->cmd_len);
414         req->sense = sioc->sense;
415         req->sense_len = 0;
416         req->timeout = timeout;
417         req->retries = retries;
418         req->end_io_data = sioc;
419
420         sioc->data = privdata;
421         sioc->done = done;
422
423         blk_execute_rq_nowait(req->q, NULL, req, 1, scsi_end_async);
424         return 0;
425
426 free_req:
427         blk_put_request(req);
428 free_sense:
429         kmem_cache_free(scsi_io_context_cache, sioc);
430         return DRIVER_ERROR << 24;
431 }
432 EXPORT_SYMBOL_GPL(scsi_execute_async);
433
434 /*
435  * Function:    scsi_init_cmd_errh()
436  *
437  * Purpose:     Initialize cmd fields related to error handling.
438  *
439  * Arguments:   cmd     - command that is ready to be queued.
440  *
441  * Notes:       This function has the job of initializing a number of
442  *              fields related to error handling.   Typically this will
443  *              be called once for each command, as required.
444  */
445 static void scsi_init_cmd_errh(struct scsi_cmnd *cmd)
446 {
447         cmd->serial_number = 0;
448         memset(cmd->sense_buffer, 0, sizeof cmd->sense_buffer);
449         if (cmd->cmd_len == 0)
450                 cmd->cmd_len = COMMAND_SIZE(cmd->cmnd[0]);
451 }
452
453 void scsi_device_unbusy(struct scsi_device *sdev)
454 {
455         struct Scsi_Host *shost = sdev->host;
456         unsigned long flags;
457
458         spin_lock_irqsave(shost->host_lock, flags);
459         shost->host_busy--;
460         if (unlikely(scsi_host_in_recovery(shost) &&
461                      (shost->host_failed || shost->host_eh_scheduled)))
462                 scsi_eh_wakeup(shost);
463         spin_unlock(shost->host_lock);
464         spin_lock(sdev->request_queue->queue_lock);
465         sdev->device_busy--;
466         spin_unlock_irqrestore(sdev->request_queue->queue_lock, flags);
467 }
468
469 /*
470  * Called for single_lun devices on IO completion. Clear starget_sdev_user,
471  * and call blk_run_queue for all the scsi_devices on the target -
472  * including current_sdev first.
473  *
474  * Called with *no* scsi locks held.
475  */
476 static void scsi_single_lun_run(struct scsi_device *current_sdev)
477 {
478         struct Scsi_Host *shost = current_sdev->host;
479         struct scsi_device *sdev, *tmp;
480         struct scsi_target *starget = scsi_target(current_sdev);
481         unsigned long flags;
482
483         spin_lock_irqsave(shost->host_lock, flags);
484         starget->starget_sdev_user = NULL;
485         spin_unlock_irqrestore(shost->host_lock, flags);
486
487         /*
488          * Call blk_run_queue for all LUNs on the target, starting with
489          * current_sdev. We race with others (to set starget_sdev_user),
490          * but in most cases, we will be first. Ideally, each LU on the
491          * target would get some limited time or requests on the target.
492          */
493         blk_run_queue(current_sdev->request_queue);
494
495         spin_lock_irqsave(shost->host_lock, flags);
496         if (starget->starget_sdev_user)
497                 goto out;
498         list_for_each_entry_safe(sdev, tmp, &starget->devices,
499                         same_target_siblings) {
500                 if (sdev == current_sdev)
501                         continue;
502                 if (scsi_device_get(sdev))
503                         continue;
504
505                 spin_unlock_irqrestore(shost->host_lock, flags);
506                 blk_run_queue(sdev->request_queue);
507                 spin_lock_irqsave(shost->host_lock, flags);
508         
509                 scsi_device_put(sdev);
510         }
511  out:
512         spin_unlock_irqrestore(shost->host_lock, flags);
513 }
514
515 /*
516  * Function:    scsi_run_queue()
517  *
518  * Purpose:     Select a proper request queue to serve next
519  *
520  * Arguments:   q       - last request's queue
521  *
522  * Returns:     Nothing
523  *
524  * Notes:       The previous command was completely finished, start
525  *              a new one if possible.
526  */
527 static void scsi_run_queue(struct request_queue *q)
528 {
529         struct scsi_device *sdev = q->queuedata;
530         struct Scsi_Host *shost = sdev->host;
531         unsigned long flags;
532
533         if (sdev->single_lun)
534                 scsi_single_lun_run(sdev);
535
536         spin_lock_irqsave(shost->host_lock, flags);
537         while (!list_empty(&shost->starved_list) &&
538                !shost->host_blocked && !shost->host_self_blocked &&
539                 !((shost->can_queue > 0) &&
540                   (shost->host_busy >= shost->can_queue))) {
541                 /*
542                  * As long as shost is accepting commands and we have
543                  * starved queues, call blk_run_queue. scsi_request_fn
544                  * drops the queue_lock and can add us back to the
545                  * starved_list.
546                  *
547                  * host_lock protects the starved_list and starved_entry.
548                  * scsi_request_fn must get the host_lock before checking
549                  * or modifying starved_list or starved_entry.
550                  */
551                 sdev = list_entry(shost->starved_list.next,
552                                           struct scsi_device, starved_entry);
553                 list_del_init(&sdev->starved_entry);
554                 spin_unlock_irqrestore(shost->host_lock, flags);
555
556
557                 if (test_bit(QUEUE_FLAG_REENTER, &q->queue_flags) &&
558                     !test_and_set_bit(QUEUE_FLAG_REENTER,
559                                       &sdev->request_queue->queue_flags)) {
560                         blk_run_queue(sdev->request_queue);
561                         clear_bit(QUEUE_FLAG_REENTER,
562                                   &sdev->request_queue->queue_flags);
563                 } else
564                         blk_run_queue(sdev->request_queue);
565
566                 spin_lock_irqsave(shost->host_lock, flags);
567                 if (unlikely(!list_empty(&sdev->starved_entry)))
568                         /*
569                          * sdev lost a race, and was put back on the
570                          * starved list. This is unlikely but without this
571                          * in theory we could loop forever.
572                          */
573                         break;
574         }
575         spin_unlock_irqrestore(shost->host_lock, flags);
576
577         blk_run_queue(q);
578 }
579
580 /*
581  * Function:    scsi_requeue_command()
582  *
583  * Purpose:     Handle post-processing of completed commands.
584  *
585  * Arguments:   q       - queue to operate on
586  *              cmd     - command that may need to be requeued.
587  *
588  * Returns:     Nothing
589  *
590  * Notes:       After command completion, there may be blocks left
591  *              over which weren't finished by the previous command
592  *              this can be for a number of reasons - the main one is
593  *              I/O errors in the middle of the request, in which case
594  *              we need to request the blocks that come after the bad
595  *              sector.
596  * Notes:       Upon return, cmd is a stale pointer.
597  */
598 static void scsi_requeue_command(struct request_queue *q, struct scsi_cmnd *cmd)
599 {
600         struct request *req = cmd->request;
601         unsigned long flags;
602
603         scsi_unprep_request(req);
604         spin_lock_irqsave(q->queue_lock, flags);
605         blk_requeue_request(q, req);
606         spin_unlock_irqrestore(q->queue_lock, flags);
607
608         scsi_run_queue(q);
609 }
610
611 void scsi_next_command(struct scsi_cmnd *cmd)
612 {
613         struct scsi_device *sdev = cmd->device;
614         struct request_queue *q = sdev->request_queue;
615
616         /* need to hold a reference on the device before we let go of the cmd */
617         get_device(&sdev->sdev_gendev);
618
619         scsi_put_command(cmd);
620         scsi_run_queue(q);
621
622         /* ok to remove device now */
623         put_device(&sdev->sdev_gendev);
624 }
625
626 void scsi_run_host_queues(struct Scsi_Host *shost)
627 {
628         struct scsi_device *sdev;
629
630         shost_for_each_device(sdev, shost)
631                 scsi_run_queue(sdev->request_queue);
632 }
633
634 /*
635  * Function:    scsi_end_request()
636  *
637  * Purpose:     Post-processing of completed commands (usually invoked at end
638  *              of upper level post-processing and scsi_io_completion).
639  *
640  * Arguments:   cmd      - command that is complete.
641  *              uptodate - 1 if I/O indicates success, <= 0 for I/O error.
642  *              bytes    - number of bytes of completed I/O
643  *              requeue  - indicates whether we should requeue leftovers.
644  *
645  * Lock status: Assumed that lock is not held upon entry.
646  *
647  * Returns:     cmd if requeue required, NULL otherwise.
648  *
649  * Notes:       This is called for block device requests in order to
650  *              mark some number of sectors as complete.
651  * 
652  *              We are guaranteeing that the request queue will be goosed
653  *              at some point during this call.
654  * Notes:       If cmd was requeued, upon return it will be a stale pointer.
655  */
656 static struct scsi_cmnd *scsi_end_request(struct scsi_cmnd *cmd, int uptodate,
657                                           int bytes, int requeue)
658 {
659         request_queue_t *q = cmd->device->request_queue;
660         struct request *req = cmd->request;
661         unsigned long flags;
662
663         /*
664          * If there are blocks left over at the end, set up the command
665          * to queue the remainder of them.
666          */
667         if (end_that_request_chunk(req, uptodate, bytes)) {
668                 int leftover = (req->hard_nr_sectors << 9);
669
670                 if (blk_pc_request(req))
671                         leftover = req->data_len;
672
673                 /* kill remainder if no retrys */
674                 if (!uptodate && blk_noretry_request(req))
675                         end_that_request_chunk(req, 0, leftover);
676                 else {
677                         if (requeue) {
678                                 /*
679                                  * Bleah.  Leftovers again.  Stick the
680                                  * leftovers in the front of the
681                                  * queue, and goose the queue again.
682                                  */
683                                 scsi_requeue_command(q, cmd);
684                                 cmd = NULL;
685                         }
686                         return cmd;
687                 }
688         }
689
690         add_disk_randomness(req->rq_disk);
691
692         spin_lock_irqsave(q->queue_lock, flags);
693         if (blk_rq_tagged(req))
694                 blk_queue_end_tag(q, req);
695         end_that_request_last(req, uptodate);
696         spin_unlock_irqrestore(q->queue_lock, flags);
697
698         /*
699          * This will goose the queue request function at the end, so we don't
700          * need to worry about launching another command.
701          */
702         scsi_next_command(cmd);
703         return NULL;
704 }
705
706 static struct scatterlist *scsi_alloc_sgtable(struct scsi_cmnd *cmd, gfp_t gfp_mask)
707 {
708         struct scsi_host_sg_pool *sgp;
709         struct scatterlist *sgl;
710
711         BUG_ON(!cmd->use_sg);
712
713         switch (cmd->use_sg) {
714         case 1 ... 8:
715                 cmd->sglist_len = 0;
716                 break;
717         case 9 ... 16:
718                 cmd->sglist_len = 1;
719                 break;
720         case 17 ... 32:
721                 cmd->sglist_len = 2;
722                 break;
723 #if (SCSI_MAX_PHYS_SEGMENTS > 32)
724         case 33 ... 64:
725                 cmd->sglist_len = 3;
726                 break;
727 #if (SCSI_MAX_PHYS_SEGMENTS > 64)
728         case 65 ... 128:
729                 cmd->sglist_len = 4;
730                 break;
731 #if (SCSI_MAX_PHYS_SEGMENTS  > 128)
732         case 129 ... 256:
733                 cmd->sglist_len = 5;
734                 break;
735 #endif
736 #endif
737 #endif
738         default:
739                 return NULL;
740         }
741
742         sgp = scsi_sg_pools + cmd->sglist_len;
743         sgl = mempool_alloc(sgp->pool, gfp_mask);
744         return sgl;
745 }
746
747 static void scsi_free_sgtable(struct scatterlist *sgl, int index)
748 {
749         struct scsi_host_sg_pool *sgp;
750
751         BUG_ON(index >= SG_MEMPOOL_NR);
752
753         sgp = scsi_sg_pools + index;
754         mempool_free(sgl, sgp->pool);
755 }
756
757 /*
758  * Function:    scsi_release_buffers()
759  *
760  * Purpose:     Completion processing for block device I/O requests.
761  *
762  * Arguments:   cmd     - command that we are bailing.
763  *
764  * Lock status: Assumed that no lock is held upon entry.
765  *
766  * Returns:     Nothing
767  *
768  * Notes:       In the event that an upper level driver rejects a
769  *              command, we must release resources allocated during
770  *              the __init_io() function.  Primarily this would involve
771  *              the scatter-gather table, and potentially any bounce
772  *              buffers.
773  */
774 static void scsi_release_buffers(struct scsi_cmnd *cmd)
775 {
776         if (cmd->use_sg)
777                 scsi_free_sgtable(cmd->request_buffer, cmd->sglist_len);
778
779         /*
780          * Zero these out.  They now point to freed memory, and it is
781          * dangerous to hang onto the pointers.
782          */
783         cmd->request_buffer = NULL;
784         cmd->request_bufflen = 0;
785 }
786
787 /*
788  * Function:    scsi_io_completion()
789  *
790  * Purpose:     Completion processing for block device I/O requests.
791  *
792  * Arguments:   cmd   - command that is finished.
793  *
794  * Lock status: Assumed that no lock is held upon entry.
795  *
796  * Returns:     Nothing
797  *
798  * Notes:       This function is matched in terms of capabilities to
799  *              the function that created the scatter-gather list.
800  *              In other words, if there are no bounce buffers
801  *              (the normal case for most drivers), we don't need
802  *              the logic to deal with cleaning up afterwards.
803  *
804  *              We must do one of several things here:
805  *
806  *              a) Call scsi_end_request.  This will finish off the
807  *                 specified number of sectors.  If we are done, the
808  *                 command block will be released, and the queue
809  *                 function will be goosed.  If we are not done, then
810  *                 scsi_end_request will directly goose the queue.
811  *
812  *              b) We can just use scsi_requeue_command() here.  This would
813  *                 be used if we just wanted to retry, for example.
814  */
815 void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes)
816 {
817         int result = cmd->result;
818         int this_count = cmd->request_bufflen;
819         request_queue_t *q = cmd->device->request_queue;
820         struct request *req = cmd->request;
821         int clear_errors = 1;
822         struct scsi_sense_hdr sshdr;
823         int sense_valid = 0;
824         int sense_deferred = 0;
825
826         scsi_release_buffers(cmd);
827
828         if (result) {
829                 sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
830                 if (sense_valid)
831                         sense_deferred = scsi_sense_is_deferred(&sshdr);
832         }
833
834         if (blk_pc_request(req)) { /* SG_IO ioctl from block level */
835                 req->errors = result;
836                 if (result) {
837                         clear_errors = 0;
838                         if (sense_valid && req->sense) {
839                                 /*
840                                  * SG_IO wants current and deferred errors
841                                  */
842                                 int len = 8 + cmd->sense_buffer[7];
843
844                                 if (len > SCSI_SENSE_BUFFERSIZE)
845                                         len = SCSI_SENSE_BUFFERSIZE;
846                                 memcpy(req->sense, cmd->sense_buffer,  len);
847                                 req->sense_len = len;
848                         }
849                 } else
850                         req->data_len = cmd->resid;
851         }
852
853         /*
854          * Next deal with any sectors which we were able to correctly
855          * handle.
856          */
857         SCSI_LOG_HLCOMPLETE(1, printk("%ld sectors total, "
858                                       "%d bytes done.\n",
859                                       req->nr_sectors, good_bytes));
860         SCSI_LOG_HLCOMPLETE(1, printk("use_sg is %d\n", cmd->use_sg));
861
862         if (clear_errors)
863                 req->errors = 0;
864
865         /* A number of bytes were successfully read.  If there
866          * are leftovers and there is some kind of error
867          * (result != 0), retry the rest.
868          */
869         if (scsi_end_request(cmd, 1, good_bytes, result == 0) == NULL)
870                 return;
871
872         /* good_bytes = 0, or (inclusive) there were leftovers and
873          * result = 0, so scsi_end_request couldn't retry.
874          */
875         if (sense_valid && !sense_deferred) {
876                 switch (sshdr.sense_key) {
877                 case UNIT_ATTENTION:
878                         if (cmd->device->removable) {
879                                 /* Detected disc change.  Set a bit
880                                  * and quietly refuse further access.
881                                  */
882                                 cmd->device->changed = 1;
883                                 scsi_end_request(cmd, 0, this_count, 1);
884                                 return;
885                         } else {
886                                 /* Must have been a power glitch, or a
887                                  * bus reset.  Could not have been a
888                                  * media change, so we just retry the
889                                  * request and see what happens.
890                                  */
891                                 scsi_requeue_command(q, cmd);
892                                 return;
893                         }
894                         break;
895                 case ILLEGAL_REQUEST:
896                         /* If we had an ILLEGAL REQUEST returned, then
897                          * we may have performed an unsupported
898                          * command.  The only thing this should be
899                          * would be a ten byte read where only a six
900                          * byte read was supported.  Also, on a system
901                          * where READ CAPACITY failed, we may have
902                          * read past the end of the disk.
903                          */
904                         if ((cmd->device->use_10_for_rw &&
905                             sshdr.asc == 0x20 && sshdr.ascq == 0x00) &&
906                             (cmd->cmnd[0] == READ_10 ||
907                              cmd->cmnd[0] == WRITE_10)) {
908                                 cmd->device->use_10_for_rw = 0;
909                                 /* This will cause a retry with a
910                                  * 6-byte command.
911                                  */
912                                 scsi_requeue_command(q, cmd);
913                                 return;
914                         } else {
915                                 scsi_end_request(cmd, 0, this_count, 1);
916                                 return;
917                         }
918                         break;
919                 case NOT_READY:
920                         /* If the device is in the process of becoming
921                          * ready, or has a temporary blockage, retry.
922                          */
923                         if (sshdr.asc == 0x04) {
924                                 switch (sshdr.ascq) {
925                                 case 0x01: /* becoming ready */
926                                 case 0x04: /* format in progress */
927                                 case 0x05: /* rebuild in progress */
928                                 case 0x06: /* recalculation in progress */
929                                 case 0x07: /* operation in progress */
930                                 case 0x08: /* Long write in progress */
931                                 case 0x09: /* self test in progress */
932                                         scsi_requeue_command(q, cmd);
933                                         return;
934                                 default:
935                                         break;
936                                 }
937                         }
938                         if (!(req->cmd_flags & REQ_QUIET)) {
939                                 scmd_printk(KERN_INFO, cmd,
940                                             "Device not ready: ");
941                                 scsi_print_sense_hdr("", &sshdr);
942                         }
943                         scsi_end_request(cmd, 0, this_count, 1);
944                         return;
945                 case VOLUME_OVERFLOW:
946                         if (!(req->cmd_flags & REQ_QUIET)) {
947                                 scmd_printk(KERN_INFO, cmd,
948                                             "Volume overflow, CDB: ");
949                                 __scsi_print_command(cmd->cmnd);
950                                 scsi_print_sense("", cmd);
951                         }
952                         /* See SSC3rXX or current. */
953                         scsi_end_request(cmd, 0, this_count, 1);
954                         return;
955                 default:
956                         break;
957                 }
958         }
959         if (host_byte(result) == DID_RESET) {
960                 /* Third party bus reset or reset for error recovery
961                  * reasons.  Just retry the request and see what
962                  * happens.
963                  */
964                 scsi_requeue_command(q, cmd);
965                 return;
966         }
967         if (result) {
968                 if (!(req->cmd_flags & REQ_QUIET)) {
969                         scmd_printk(KERN_INFO, cmd,
970                                     "SCSI error: return code = 0x%08x\n",
971                                     result);
972                         if (driver_byte(result) & DRIVER_SENSE)
973                                 scsi_print_sense("", cmd);
974                 }
975         }
976         scsi_end_request(cmd, 0, this_count, !result);
977 }
978 EXPORT_SYMBOL(scsi_io_completion);
979
980 /*
981  * Function:    scsi_init_io()
982  *
983  * Purpose:     SCSI I/O initialize function.
984  *
985  * Arguments:   cmd   - Command descriptor we wish to initialize
986  *
987  * Returns:     0 on success
988  *              BLKPREP_DEFER if the failure is retryable
989  *              BLKPREP_KILL if the failure is fatal
990  */
991 static int scsi_init_io(struct scsi_cmnd *cmd)
992 {
993         struct request     *req = cmd->request;
994         struct scatterlist *sgpnt;
995         int                count;
996
997         /*
998          * if this is a rq->data based REQ_BLOCK_PC, setup for a non-sg xfer
999          */
1000         if (blk_pc_request(req) && !req->bio) {
1001                 cmd->request_bufflen = req->data_len;
1002                 cmd->request_buffer = req->data;
1003                 req->buffer = req->data;
1004                 cmd->use_sg = 0;
1005                 return 0;
1006         }
1007
1008         /*
1009          * we used to not use scatter-gather for single segment request,
1010          * but now we do (it makes highmem I/O easier to support without
1011          * kmapping pages)
1012          */
1013         cmd->use_sg = req->nr_phys_segments;
1014
1015         /*
1016          * if sg table allocation fails, requeue request later.
1017          */
1018         sgpnt = scsi_alloc_sgtable(cmd, GFP_ATOMIC);
1019         if (unlikely(!sgpnt)) {
1020                 scsi_unprep_request(req);
1021                 return BLKPREP_DEFER;
1022         }
1023
1024         cmd->request_buffer = (char *) sgpnt;
1025         cmd->request_bufflen = req->nr_sectors << 9;
1026         if (blk_pc_request(req))
1027                 cmd->request_bufflen = req->data_len;
1028         req->buffer = NULL;
1029
1030         /* 
1031          * Next, walk the list, and fill in the addresses and sizes of
1032          * each segment.
1033          */
1034         count = blk_rq_map_sg(req->q, req, cmd->request_buffer);
1035
1036         /*
1037          * mapped well, send it off
1038          */
1039         if (likely(count <= cmd->use_sg)) {
1040                 cmd->use_sg = count;
1041                 return 0;
1042         }
1043
1044         printk(KERN_ERR "Incorrect number of segments after building list\n");
1045         printk(KERN_ERR "counted %d, received %d\n", count, cmd->use_sg);
1046         printk(KERN_ERR "req nr_sec %lu, cur_nr_sec %u\n", req->nr_sectors,
1047                         req->current_nr_sectors);
1048
1049         /* release the command and kill it */
1050         scsi_release_buffers(cmd);
1051         scsi_put_command(cmd);
1052         return BLKPREP_KILL;
1053 }
1054
1055 static int scsi_issue_flush_fn(request_queue_t *q, struct gendisk *disk,
1056                                sector_t *error_sector)
1057 {
1058         struct scsi_device *sdev = q->queuedata;
1059         struct scsi_driver *drv;
1060
1061         if (sdev->sdev_state != SDEV_RUNNING)
1062                 return -ENXIO;
1063
1064         drv = *(struct scsi_driver **) disk->private_data;
1065         if (drv->issue_flush)
1066                 return drv->issue_flush(&sdev->sdev_gendev, error_sector);
1067
1068         return -EOPNOTSUPP;
1069 }
1070
1071 static void scsi_blk_pc_done(struct scsi_cmnd *cmd)
1072 {
1073         BUG_ON(!blk_pc_request(cmd->request));
1074         /*
1075          * This will complete the whole command with uptodate=1 so
1076          * as far as the block layer is concerned the command completed
1077          * successfully. Since this is a REQ_BLOCK_PC command the
1078          * caller should check the request's errors value
1079          */
1080         scsi_io_completion(cmd, cmd->request_bufflen);
1081 }
1082
1083 static void scsi_setup_blk_pc_cmnd(struct scsi_cmnd *cmd)
1084 {
1085         struct request *req = cmd->request;
1086
1087         BUG_ON(sizeof(req->cmd) > sizeof(cmd->cmnd));
1088         memcpy(cmd->cmnd, req->cmd, sizeof(cmd->cmnd));
1089         cmd->cmd_len = req->cmd_len;
1090         if (!req->data_len)
1091                 cmd->sc_data_direction = DMA_NONE;
1092         else if (rq_data_dir(req) == WRITE)
1093                 cmd->sc_data_direction = DMA_TO_DEVICE;
1094         else
1095                 cmd->sc_data_direction = DMA_FROM_DEVICE;
1096         
1097         cmd->transfersize = req->data_len;
1098         cmd->allowed = req->retries;
1099         cmd->timeout_per_command = req->timeout;
1100         cmd->done = scsi_blk_pc_done;
1101 }
1102
1103 static int scsi_prep_fn(struct request_queue *q, struct request *req)
1104 {
1105         struct scsi_device *sdev = q->queuedata;
1106         struct scsi_cmnd *cmd;
1107         int specials_only = 0;
1108
1109         /*
1110          * Just check to see if the device is online.  If it isn't, we
1111          * refuse to process any commands.  The device must be brought
1112          * online before trying any recovery commands
1113          */
1114         if (unlikely(!scsi_device_online(sdev))) {
1115                 sdev_printk(KERN_ERR, sdev,
1116                             "rejecting I/O to offline device\n");
1117                 goto kill;
1118         }
1119         if (unlikely(sdev->sdev_state != SDEV_RUNNING)) {
1120                 /* OK, we're not in a running state don't prep
1121                  * user commands */
1122                 if (sdev->sdev_state == SDEV_DEL) {
1123                         /* Device is fully deleted, no commands
1124                          * at all allowed down */
1125                         sdev_printk(KERN_ERR, sdev,
1126                                     "rejecting I/O to dead device\n");
1127                         goto kill;
1128                 }
1129                 /* OK, we only allow special commands (i.e. not
1130                  * user initiated ones */
1131                 specials_only = sdev->sdev_state;
1132         }
1133
1134         /*
1135          * Find the actual device driver associated with this command.
1136          * The SPECIAL requests are things like character device or
1137          * ioctls, which did not originate from ll_rw_blk.  Note that
1138          * the special field is also used to indicate the cmd for
1139          * the remainder of a partially fulfilled request that can 
1140          * come up when there is a medium error.  We have to treat
1141          * these two cases differently.  We differentiate by looking
1142          * at request->cmd, as this tells us the real story.
1143          */
1144         if (blk_special_request(req) && req->special)
1145                 cmd = req->special;
1146         else if (blk_pc_request(req) || blk_fs_request(req)) {
1147                 if (unlikely(specials_only) && !(req->cmd_flags & REQ_PREEMPT)){
1148                         if (specials_only == SDEV_QUIESCE ||
1149                             specials_only == SDEV_BLOCK)
1150                                 goto defer;
1151                         
1152                         sdev_printk(KERN_ERR, sdev,
1153                                     "rejecting I/O to device being removed\n");
1154                         goto kill;
1155                 }
1156                         
1157                 /*
1158                  * Now try and find a command block that we can use.
1159                  */
1160                 if (!req->special) {
1161                         cmd = scsi_get_command(sdev, GFP_ATOMIC);
1162                         if (unlikely(!cmd))
1163                                 goto defer;
1164                 } else
1165                         cmd = req->special;
1166                 
1167                 /* pull a tag out of the request if we have one */
1168                 cmd->tag = req->tag;
1169         } else {
1170                 blk_dump_rq_flags(req, "SCSI bad req");
1171                 goto kill;
1172         }
1173         
1174         /* note the overloading of req->special.  When the tag
1175          * is active it always means cmd.  If the tag goes
1176          * back for re-queueing, it may be reset */
1177         req->special = cmd;
1178         cmd->request = req;
1179         
1180         /*
1181          * FIXME: drop the lock here because the functions below
1182          * expect to be called without the queue lock held.  Also,
1183          * previously, we dequeued the request before dropping the
1184          * lock.  We hope REQ_STARTED prevents anything untoward from
1185          * happening now.
1186          */
1187         if (blk_fs_request(req) || blk_pc_request(req)) {
1188                 int ret;
1189
1190                 /*
1191                  * This will do a couple of things:
1192                  *  1) Fill in the actual SCSI command.
1193                  *  2) Fill in any other upper-level specific fields
1194                  * (timeout).
1195                  *
1196                  * If this returns 0, it means that the request failed
1197                  * (reading past end of disk, reading offline device,
1198                  * etc).   This won't actually talk to the device, but
1199                  * some kinds of consistency checking may cause the     
1200                  * request to be rejected immediately.
1201                  */
1202
1203                 /* 
1204                  * This sets up the scatter-gather table (allocating if
1205                  * required).
1206                  */
1207                 ret = scsi_init_io(cmd);
1208                 switch(ret) {
1209                         /* For BLKPREP_KILL/DEFER the cmd was released */
1210                 case BLKPREP_KILL:
1211                         goto kill;
1212                 case BLKPREP_DEFER:
1213                         goto defer;
1214                 }
1215                 
1216                 /*
1217                  * Initialize the actual SCSI command for this request.
1218                  */
1219                 if (blk_pc_request(req)) {
1220                         scsi_setup_blk_pc_cmnd(cmd);
1221                 } else if (req->rq_disk) {
1222                         struct scsi_driver *drv;
1223
1224                         drv = *(struct scsi_driver **)req->rq_disk->private_data;
1225                         if (unlikely(!drv->init_command(cmd))) {
1226                                 scsi_release_buffers(cmd);
1227                                 scsi_put_command(cmd);
1228                                 goto kill;
1229                         }
1230                 }
1231         }
1232
1233         /*
1234          * The request is now prepped, no need to come back here
1235          */
1236         req->cmd_flags |= REQ_DONTPREP;
1237         return BLKPREP_OK;
1238
1239  defer:
1240         /* If we defer, the elv_next_request() returns NULL, but the
1241          * queue must be restarted, so we plug here if no returning
1242          * command will automatically do that. */
1243         if (sdev->device_busy == 0)
1244                 blk_plug_device(q);
1245         return BLKPREP_DEFER;
1246  kill:
1247         req->errors = DID_NO_CONNECT << 16;
1248         return BLKPREP_KILL;
1249 }
1250
1251 /*
1252  * scsi_dev_queue_ready: if we can send requests to sdev, return 1 else
1253  * return 0.
1254  *
1255  * Called with the queue_lock held.
1256  */
1257 static inline int scsi_dev_queue_ready(struct request_queue *q,
1258                                   struct scsi_device *sdev)
1259 {
1260         if (sdev->device_busy >= sdev->queue_depth)
1261                 return 0;
1262         if (sdev->device_busy == 0 && sdev->device_blocked) {
1263                 /*
1264                  * unblock after device_blocked iterates to zero
1265                  */
1266                 if (--sdev->device_blocked == 0) {
1267                         SCSI_LOG_MLQUEUE(3,
1268                                    sdev_printk(KERN_INFO, sdev,
1269                                    "unblocking device at zero depth\n"));
1270                 } else {
1271                         blk_plug_device(q);
1272                         return 0;
1273                 }
1274         }
1275         if (sdev->device_blocked)
1276                 return 0;
1277
1278         return 1;
1279 }
1280
1281 /*
1282  * scsi_host_queue_ready: if we can send requests to shost, return 1 else
1283  * return 0. We must end up running the queue again whenever 0 is
1284  * returned, else IO can hang.
1285  *
1286  * Called with host_lock held.
1287  */
1288 static inline int scsi_host_queue_ready(struct request_queue *q,
1289                                    struct Scsi_Host *shost,
1290                                    struct scsi_device *sdev)
1291 {
1292         if (scsi_host_in_recovery(shost))
1293                 return 0;
1294         if (shost->host_busy == 0 && shost->host_blocked) {
1295                 /*
1296                  * unblock after host_blocked iterates to zero
1297                  */
1298                 if (--shost->host_blocked == 0) {
1299                         SCSI_LOG_MLQUEUE(3,
1300                                 printk("scsi%d unblocking host at zero depth\n",
1301                                         shost->host_no));
1302                 } else {
1303                         blk_plug_device(q);
1304                         return 0;
1305                 }
1306         }
1307         if ((shost->can_queue > 0 && shost->host_busy >= shost->can_queue) ||
1308             shost->host_blocked || shost->host_self_blocked) {
1309                 if (list_empty(&sdev->starved_entry))
1310                         list_add_tail(&sdev->starved_entry, &shost->starved_list);
1311                 return 0;
1312         }
1313
1314         /* We're OK to process the command, so we can't be starved */
1315         if (!list_empty(&sdev->starved_entry))
1316                 list_del_init(&sdev->starved_entry);
1317
1318         return 1;
1319 }
1320
1321 /*
1322  * Kill a request for a dead device
1323  */
1324 static void scsi_kill_request(struct request *req, request_queue_t *q)
1325 {
1326         struct scsi_cmnd *cmd = req->special;
1327         struct scsi_device *sdev = cmd->device;
1328         struct Scsi_Host *shost = sdev->host;
1329
1330         blkdev_dequeue_request(req);
1331
1332         if (unlikely(cmd == NULL)) {
1333                 printk(KERN_CRIT "impossible request in %s.\n",
1334                                  __FUNCTION__);
1335                 BUG();
1336         }
1337
1338         scsi_init_cmd_errh(cmd);
1339         cmd->result = DID_NO_CONNECT << 16;
1340         atomic_inc(&cmd->device->iorequest_cnt);
1341
1342         /*
1343          * SCSI request completion path will do scsi_device_unbusy(),
1344          * bump busy counts.  To bump the counters, we need to dance
1345          * with the locks as normal issue path does.
1346          */
1347         sdev->device_busy++;
1348         spin_unlock(sdev->request_queue->queue_lock);
1349         spin_lock(shost->host_lock);
1350         shost->host_busy++;
1351         spin_unlock(shost->host_lock);
1352         spin_lock(sdev->request_queue->queue_lock);
1353
1354         __scsi_done(cmd);
1355 }
1356
1357 static void scsi_softirq_done(struct request *rq)
1358 {
1359         struct scsi_cmnd *cmd = rq->completion_data;
1360         unsigned long wait_for = (cmd->allowed + 1) * cmd->timeout_per_command;
1361         int disposition;
1362
1363         INIT_LIST_HEAD(&cmd->eh_entry);
1364
1365         disposition = scsi_decide_disposition(cmd);
1366         if (disposition != SUCCESS &&
1367             time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) {
1368                 sdev_printk(KERN_ERR, cmd->device,
1369                             "timing out command, waited %lus\n",
1370                             wait_for/HZ);
1371                 disposition = SUCCESS;
1372         }
1373                         
1374         scsi_log_completion(cmd, disposition);
1375
1376         switch (disposition) {
1377                 case SUCCESS:
1378                         scsi_finish_command(cmd);
1379                         break;
1380                 case NEEDS_RETRY:
1381                         scsi_retry_command(cmd);
1382                         break;
1383                 case ADD_TO_MLQUEUE:
1384                         scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY);
1385                         break;
1386                 default:
1387                         if (!scsi_eh_scmd_add(cmd, 0))
1388                                 scsi_finish_command(cmd);
1389         }
1390 }
1391
1392 /*
1393  * Function:    scsi_request_fn()
1394  *
1395  * Purpose:     Main strategy routine for SCSI.
1396  *
1397  * Arguments:   q       - Pointer to actual queue.
1398  *
1399  * Returns:     Nothing
1400  *
1401  * Lock status: IO request lock assumed to be held when called.
1402  */
1403 static void scsi_request_fn(struct request_queue *q)
1404 {
1405         struct scsi_device *sdev = q->queuedata;
1406         struct Scsi_Host *shost;
1407         struct scsi_cmnd *cmd;
1408         struct request *req;
1409
1410         if (!sdev) {
1411                 printk("scsi: killing requests for dead queue\n");
1412                 while ((req = elv_next_request(q)) != NULL)
1413                         scsi_kill_request(req, q);
1414                 return;
1415         }
1416
1417         if(!get_device(&sdev->sdev_gendev))
1418                 /* We must be tearing the block queue down already */
1419                 return;
1420
1421         /*
1422          * To start with, we keep looping until the queue is empty, or until
1423          * the host is no longer able to accept any more requests.
1424          */
1425         shost = sdev->host;
1426         while (!blk_queue_plugged(q)) {
1427                 int rtn;
1428                 /*
1429                  * get next queueable request.  We do this early to make sure
1430                  * that the request is fully prepared even if we cannot 
1431                  * accept it.
1432                  */
1433                 req = elv_next_request(q);
1434                 if (!req || !scsi_dev_queue_ready(q, sdev))
1435                         break;
1436
1437                 if (unlikely(!scsi_device_online(sdev))) {
1438                         sdev_printk(KERN_ERR, sdev,
1439                                     "rejecting I/O to offline device\n");
1440                         scsi_kill_request(req, q);
1441                         continue;
1442                 }
1443
1444
1445                 /*
1446                  * Remove the request from the request list.
1447                  */
1448                 if (!(blk_queue_tagged(q) && !blk_queue_start_tag(q, req)))
1449                         blkdev_dequeue_request(req);
1450                 sdev->device_busy++;
1451
1452                 spin_unlock(q->queue_lock);
1453                 cmd = req->special;
1454                 if (unlikely(cmd == NULL)) {
1455                         printk(KERN_CRIT "impossible request in %s.\n"
1456                                          "please mail a stack trace to "
1457                                          "linux-scsi@vger.kernel.org\n",
1458                                          __FUNCTION__);
1459                         blk_dump_rq_flags(req, "foo");
1460                         BUG();
1461                 }
1462                 spin_lock(shost->host_lock);
1463
1464                 if (!scsi_host_queue_ready(q, shost, sdev))
1465                         goto not_ready;
1466                 if (sdev->single_lun) {
1467                         if (scsi_target(sdev)->starget_sdev_user &&
1468                             scsi_target(sdev)->starget_sdev_user != sdev)
1469                                 goto not_ready;
1470                         scsi_target(sdev)->starget_sdev_user = sdev;
1471                 }
1472                 shost->host_busy++;
1473
1474                 /*
1475                  * XXX(hch): This is rather suboptimal, scsi_dispatch_cmd will
1476                  *              take the lock again.
1477                  */
1478                 spin_unlock_irq(shost->host_lock);
1479
1480                 /*
1481                  * Finally, initialize any error handling parameters, and set up
1482                  * the timers for timeouts.
1483                  */
1484                 scsi_init_cmd_errh(cmd);
1485
1486                 /*
1487                  * Dispatch the command to the low-level driver.
1488                  */
1489                 rtn = scsi_dispatch_cmd(cmd);
1490                 spin_lock_irq(q->queue_lock);
1491                 if(rtn) {
1492                         /* we're refusing the command; because of
1493                          * the way locks get dropped, we need to 
1494                          * check here if plugging is required */
1495                         if(sdev->device_busy == 0)
1496                                 blk_plug_device(q);
1497
1498                         break;
1499                 }
1500         }
1501
1502         goto out;
1503
1504  not_ready:
1505         spin_unlock_irq(shost->host_lock);
1506
1507         /*
1508          * lock q, handle tag, requeue req, and decrement device_busy. We
1509          * must return with queue_lock held.
1510          *
1511          * Decrementing device_busy without checking it is OK, as all such
1512          * cases (host limits or settings) should run the queue at some
1513          * later time.
1514          */
1515         spin_lock_irq(q->queue_lock);
1516         blk_requeue_request(q, req);
1517         sdev->device_busy--;
1518         if(sdev->device_busy == 0)
1519                 blk_plug_device(q);
1520  out:
1521         /* must be careful here...if we trigger the ->remove() function
1522          * we cannot be holding the q lock */
1523         spin_unlock_irq(q->queue_lock);
1524         put_device(&sdev->sdev_gendev);
1525         spin_lock_irq(q->queue_lock);
1526 }
1527
1528 u64 scsi_calculate_bounce_limit(struct Scsi_Host *shost)
1529 {
1530         struct device *host_dev;
1531         u64 bounce_limit = 0xffffffff;
1532
1533         if (shost->unchecked_isa_dma)
1534                 return BLK_BOUNCE_ISA;
1535         /*
1536          * Platforms with virtual-DMA translation
1537          * hardware have no practical limit.
1538          */
1539         if (!PCI_DMA_BUS_IS_PHYS)
1540                 return BLK_BOUNCE_ANY;
1541
1542         host_dev = scsi_get_device(shost);
1543         if (host_dev && host_dev->dma_mask)
1544                 bounce_limit = *host_dev->dma_mask;
1545
1546         return bounce_limit;
1547 }
1548 EXPORT_SYMBOL(scsi_calculate_bounce_limit);
1549
1550 struct request_queue *scsi_alloc_queue(struct scsi_device *sdev)
1551 {
1552         struct Scsi_Host *shost = sdev->host;
1553         struct request_queue *q;
1554
1555         q = blk_init_queue(scsi_request_fn, NULL);
1556         if (!q)
1557                 return NULL;
1558
1559         blk_queue_prep_rq(q, scsi_prep_fn);
1560
1561         blk_queue_max_hw_segments(q, shost->sg_tablesize);
1562         blk_queue_max_phys_segments(q, SCSI_MAX_PHYS_SEGMENTS);
1563         blk_queue_max_sectors(q, shost->max_sectors);
1564         blk_queue_bounce_limit(q, scsi_calculate_bounce_limit(shost));
1565         blk_queue_segment_boundary(q, shost->dma_boundary);
1566         blk_queue_issue_flush_fn(q, scsi_issue_flush_fn);
1567         blk_queue_softirq_done(q, scsi_softirq_done);
1568
1569         if (!shost->use_clustering)
1570                 clear_bit(QUEUE_FLAG_CLUSTER, &q->queue_flags);
1571         return q;
1572 }
1573
1574 void scsi_free_queue(struct request_queue *q)
1575 {
1576         blk_cleanup_queue(q);
1577 }
1578
1579 /*
1580  * Function:    scsi_block_requests()
1581  *
1582  * Purpose:     Utility function used by low-level drivers to prevent further
1583  *              commands from being queued to the device.
1584  *
1585  * Arguments:   shost       - Host in question
1586  *
1587  * Returns:     Nothing
1588  *
1589  * Lock status: No locks are assumed held.
1590  *
1591  * Notes:       There is no timer nor any other means by which the requests
1592  *              get unblocked other than the low-level driver calling
1593  *              scsi_unblock_requests().
1594  */
1595 void scsi_block_requests(struct Scsi_Host *shost)
1596 {
1597         shost->host_self_blocked = 1;
1598 }
1599 EXPORT_SYMBOL(scsi_block_requests);
1600
1601 /*
1602  * Function:    scsi_unblock_requests()
1603  *
1604  * Purpose:     Utility function used by low-level drivers to allow further
1605  *              commands from being queued to the device.
1606  *
1607  * Arguments:   shost       - Host in question
1608  *
1609  * Returns:     Nothing
1610  *
1611  * Lock status: No locks are assumed held.
1612  *
1613  * Notes:       There is no timer nor any other means by which the requests
1614  *              get unblocked other than the low-level driver calling
1615  *              scsi_unblock_requests().
1616  *
1617  *              This is done as an API function so that changes to the
1618  *              internals of the scsi mid-layer won't require wholesale
1619  *              changes to drivers that use this feature.
1620  */
1621 void scsi_unblock_requests(struct Scsi_Host *shost)
1622 {
1623         shost->host_self_blocked = 0;
1624         scsi_run_host_queues(shost);
1625 }
1626 EXPORT_SYMBOL(scsi_unblock_requests);
1627
1628 int __init scsi_init_queue(void)
1629 {
1630         int i;
1631
1632         scsi_io_context_cache = kmem_cache_create("scsi_io_context",
1633                                         sizeof(struct scsi_io_context),
1634                                         0, 0, NULL, NULL);
1635         if (!scsi_io_context_cache) {
1636                 printk(KERN_ERR "SCSI: can't init scsi io context cache\n");
1637                 return -ENOMEM;
1638         }
1639
1640         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1641                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1642                 int size = sgp->size * sizeof(struct scatterlist);
1643
1644                 sgp->slab = kmem_cache_create(sgp->name, size, 0,
1645                                 SLAB_HWCACHE_ALIGN, NULL, NULL);
1646                 if (!sgp->slab) {
1647                         printk(KERN_ERR "SCSI: can't init sg slab %s\n",
1648                                         sgp->name);
1649                 }
1650
1651                 sgp->pool = mempool_create_slab_pool(SG_MEMPOOL_SIZE,
1652                                                      sgp->slab);
1653                 if (!sgp->pool) {
1654                         printk(KERN_ERR "SCSI: can't init sg mempool %s\n",
1655                                         sgp->name);
1656                 }
1657         }
1658
1659         return 0;
1660 }
1661
1662 void scsi_exit_queue(void)
1663 {
1664         int i;
1665
1666         kmem_cache_destroy(scsi_io_context_cache);
1667
1668         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1669                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1670                 mempool_destroy(sgp->pool);
1671                 kmem_cache_destroy(sgp->slab);
1672         }
1673 }
1674
1675 /**
1676  *      scsi_mode_select - issue a mode select
1677  *      @sdev:  SCSI device to be queried
1678  *      @pf:    Page format bit (1 == standard, 0 == vendor specific)
1679  *      @sp:    Save page bit (0 == don't save, 1 == save)
1680  *      @modepage: mode page being requested
1681  *      @buffer: request buffer (may not be smaller than eight bytes)
1682  *      @len:   length of request buffer.
1683  *      @timeout: command timeout
1684  *      @retries: number of retries before failing
1685  *      @data: returns a structure abstracting the mode header data
1686  *      @sense: place to put sense data (or NULL if no sense to be collected).
1687  *              must be SCSI_SENSE_BUFFERSIZE big.
1688  *
1689  *      Returns zero if successful; negative error number or scsi
1690  *      status on error
1691  *
1692  */
1693 int
1694 scsi_mode_select(struct scsi_device *sdev, int pf, int sp, int modepage,
1695                  unsigned char *buffer, int len, int timeout, int retries,
1696                  struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1697 {
1698         unsigned char cmd[10];
1699         unsigned char *real_buffer;
1700         int ret;
1701
1702         memset(cmd, 0, sizeof(cmd));
1703         cmd[1] = (pf ? 0x10 : 0) | (sp ? 0x01 : 0);
1704
1705         if (sdev->use_10_for_ms) {
1706                 if (len > 65535)
1707                         return -EINVAL;
1708                 real_buffer = kmalloc(8 + len, GFP_KERNEL);
1709                 if (!real_buffer)
1710                         return -ENOMEM;
1711                 memcpy(real_buffer + 8, buffer, len);
1712                 len += 8;
1713                 real_buffer[0] = 0;
1714                 real_buffer[1] = 0;
1715                 real_buffer[2] = data->medium_type;
1716                 real_buffer[3] = data->device_specific;
1717                 real_buffer[4] = data->longlba ? 0x01 : 0;
1718                 real_buffer[5] = 0;
1719                 real_buffer[6] = data->block_descriptor_length >> 8;
1720                 real_buffer[7] = data->block_descriptor_length;
1721
1722                 cmd[0] = MODE_SELECT_10;
1723                 cmd[7] = len >> 8;
1724                 cmd[8] = len;
1725         } else {
1726                 if (len > 255 || data->block_descriptor_length > 255 ||
1727                     data->longlba)
1728                         return -EINVAL;
1729
1730                 real_buffer = kmalloc(4 + len, GFP_KERNEL);
1731                 if (!real_buffer)
1732                         return -ENOMEM;
1733                 memcpy(real_buffer + 4, buffer, len);
1734                 len += 4;
1735                 real_buffer[0] = 0;
1736                 real_buffer[1] = data->medium_type;
1737                 real_buffer[2] = data->device_specific;
1738                 real_buffer[3] = data->block_descriptor_length;
1739                 
1740
1741                 cmd[0] = MODE_SELECT;
1742                 cmd[4] = len;
1743         }
1744
1745         ret = scsi_execute_req(sdev, cmd, DMA_TO_DEVICE, real_buffer, len,
1746                                sshdr, timeout, retries);
1747         kfree(real_buffer);
1748         return ret;
1749 }
1750 EXPORT_SYMBOL_GPL(scsi_mode_select);
1751
1752 /**
1753  *      scsi_mode_sense - issue a mode sense, falling back from 10 to 
1754  *              six bytes if necessary.
1755  *      @sdev:  SCSI device to be queried
1756  *      @dbd:   set if mode sense will allow block descriptors to be returned
1757  *      @modepage: mode page being requested
1758  *      @buffer: request buffer (may not be smaller than eight bytes)
1759  *      @len:   length of request buffer.
1760  *      @timeout: command timeout
1761  *      @retries: number of retries before failing
1762  *      @data: returns a structure abstracting the mode header data
1763  *      @sense: place to put sense data (or NULL if no sense to be collected).
1764  *              must be SCSI_SENSE_BUFFERSIZE big.
1765  *
1766  *      Returns zero if unsuccessful, or the header offset (either 4
1767  *      or 8 depending on whether a six or ten byte command was
1768  *      issued) if successful.
1769  **/
1770 int
1771 scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
1772                   unsigned char *buffer, int len, int timeout, int retries,
1773                   struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1774 {
1775         unsigned char cmd[12];
1776         int use_10_for_ms;
1777         int header_length;
1778         int result;
1779         struct scsi_sense_hdr my_sshdr;
1780
1781         memset(data, 0, sizeof(*data));
1782         memset(&cmd[0], 0, 12);
1783         cmd[1] = dbd & 0x18;    /* allows DBD and LLBA bits */
1784         cmd[2] = modepage;
1785
1786         /* caller might not be interested in sense, but we need it */
1787         if (!sshdr)
1788                 sshdr = &my_sshdr;
1789
1790  retry:
1791         use_10_for_ms = sdev->use_10_for_ms;
1792
1793         if (use_10_for_ms) {
1794                 if (len < 8)
1795                         len = 8;
1796
1797                 cmd[0] = MODE_SENSE_10;
1798                 cmd[8] = len;
1799                 header_length = 8;
1800         } else {
1801                 if (len < 4)
1802                         len = 4;
1803
1804                 cmd[0] = MODE_SENSE;
1805                 cmd[4] = len;
1806                 header_length = 4;
1807         }
1808
1809         memset(buffer, 0, len);
1810
1811         result = scsi_execute_req(sdev, cmd, DMA_FROM_DEVICE, buffer, len,
1812                                   sshdr, timeout, retries);
1813
1814         /* This code looks awful: what it's doing is making sure an
1815          * ILLEGAL REQUEST sense return identifies the actual command
1816          * byte as the problem.  MODE_SENSE commands can return
1817          * ILLEGAL REQUEST if the code page isn't supported */
1818
1819         if (use_10_for_ms && !scsi_status_is_good(result) &&
1820             (driver_byte(result) & DRIVER_SENSE)) {
1821                 if (scsi_sense_valid(sshdr)) {
1822                         if ((sshdr->sense_key == ILLEGAL_REQUEST) &&
1823                             (sshdr->asc == 0x20) && (sshdr->ascq == 0)) {
1824                                 /* 
1825                                  * Invalid command operation code
1826                                  */
1827                                 sdev->use_10_for_ms = 0;
1828                                 goto retry;
1829                         }
1830                 }
1831         }
1832
1833         if(scsi_status_is_good(result)) {
1834                 if (unlikely(buffer[0] == 0x86 && buffer[1] == 0x0b &&
1835                              (modepage == 6 || modepage == 8))) {
1836                         /* Initio breakage? */
1837                         header_length = 0;
1838                         data->length = 13;
1839                         data->medium_type = 0;
1840                         data->device_specific = 0;
1841                         data->longlba = 0;
1842                         data->block_descriptor_length = 0;
1843                 } else if(use_10_for_ms) {
1844                         data->length = buffer[0]*256 + buffer[1] + 2;
1845                         data->medium_type = buffer[2];
1846                         data->device_specific = buffer[3];
1847                         data->longlba = buffer[4] & 0x01;
1848                         data->block_descriptor_length = buffer[6]*256
1849                                 + buffer[7];
1850                 } else {
1851                         data->length = buffer[0] + 1;
1852                         data->medium_type = buffer[1];
1853                         data->device_specific = buffer[2];
1854                         data->block_descriptor_length = buffer[3];
1855                 }
1856                 data->header_length = header_length;
1857         }
1858
1859         return result;
1860 }
1861 EXPORT_SYMBOL(scsi_mode_sense);
1862
1863 int
1864 scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries)
1865 {
1866         char cmd[] = {
1867                 TEST_UNIT_READY, 0, 0, 0, 0, 0,
1868         };
1869         struct scsi_sense_hdr sshdr;
1870         int result;
1871         
1872         result = scsi_execute_req(sdev, cmd, DMA_NONE, NULL, 0, &sshdr,
1873                                   timeout, retries);
1874
1875         if ((driver_byte(result) & DRIVER_SENSE) && sdev->removable) {
1876
1877                 if ((scsi_sense_valid(&sshdr)) &&
1878                     ((sshdr.sense_key == UNIT_ATTENTION) ||
1879                      (sshdr.sense_key == NOT_READY))) {
1880                         sdev->changed = 1;
1881                         result = 0;
1882                 }
1883         }
1884         return result;
1885 }
1886 EXPORT_SYMBOL(scsi_test_unit_ready);
1887
1888 /**
1889  *      scsi_device_set_state - Take the given device through the device
1890  *              state model.
1891  *      @sdev:  scsi device to change the state of.
1892  *      @state: state to change to.
1893  *
1894  *      Returns zero if unsuccessful or an error if the requested 
1895  *      transition is illegal.
1896  **/
1897 int
1898 scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
1899 {
1900         enum scsi_device_state oldstate = sdev->sdev_state;
1901
1902         if (state == oldstate)
1903                 return 0;
1904
1905         switch (state) {
1906         case SDEV_CREATED:
1907                 /* There are no legal states that come back to
1908                  * created.  This is the manually initialised start
1909                  * state */
1910                 goto illegal;
1911                         
1912         case SDEV_RUNNING:
1913                 switch (oldstate) {
1914                 case SDEV_CREATED:
1915                 case SDEV_OFFLINE:
1916                 case SDEV_QUIESCE:
1917                 case SDEV_BLOCK:
1918                         break;
1919                 default:
1920                         goto illegal;
1921                 }
1922                 break;
1923
1924         case SDEV_QUIESCE:
1925                 switch (oldstate) {
1926                 case SDEV_RUNNING:
1927                 case SDEV_OFFLINE:
1928                         break;
1929                 default:
1930                         goto illegal;
1931                 }
1932                 break;
1933
1934         case SDEV_OFFLINE:
1935                 switch (oldstate) {
1936                 case SDEV_CREATED:
1937                 case SDEV_RUNNING:
1938                 case SDEV_QUIESCE:
1939                 case SDEV_BLOCK:
1940                         break;
1941                 default:
1942                         goto illegal;
1943                 }
1944                 break;
1945
1946         case SDEV_BLOCK:
1947                 switch (oldstate) {
1948                 case SDEV_CREATED:
1949                 case SDEV_RUNNING:
1950                         break;
1951                 default:
1952                         goto illegal;
1953                 }
1954                 break;
1955
1956         case SDEV_CANCEL:
1957                 switch (oldstate) {
1958                 case SDEV_CREATED:
1959                 case SDEV_RUNNING:
1960                 case SDEV_QUIESCE:
1961                 case SDEV_OFFLINE:
1962                 case SDEV_BLOCK:
1963                         break;
1964                 default:
1965                         goto illegal;
1966                 }
1967                 break;
1968
1969         case SDEV_DEL:
1970                 switch (oldstate) {
1971                 case SDEV_CREATED:
1972                 case SDEV_RUNNING:
1973                 case SDEV_OFFLINE:
1974                 case SDEV_CANCEL:
1975                         break;
1976                 default:
1977                         goto illegal;
1978                 }
1979                 break;
1980
1981         }
1982         sdev->sdev_state = state;
1983         return 0;
1984
1985  illegal:
1986         SCSI_LOG_ERROR_RECOVERY(1, 
1987                                 sdev_printk(KERN_ERR, sdev,
1988                                             "Illegal state transition %s->%s\n",
1989                                             scsi_device_state_name(oldstate),
1990                                             scsi_device_state_name(state))
1991                                 );
1992         return -EINVAL;
1993 }
1994 EXPORT_SYMBOL(scsi_device_set_state);
1995
1996 /**
1997  *      scsi_device_quiesce - Block user issued commands.
1998  *      @sdev:  scsi device to quiesce.
1999  *
2000  *      This works by trying to transition to the SDEV_QUIESCE state
2001  *      (which must be a legal transition).  When the device is in this
2002  *      state, only special requests will be accepted, all others will
2003  *      be deferred.  Since special requests may also be requeued requests,
2004  *      a successful return doesn't guarantee the device will be 
2005  *      totally quiescent.
2006  *
2007  *      Must be called with user context, may sleep.
2008  *
2009  *      Returns zero if unsuccessful or an error if not.
2010  **/
2011 int
2012 scsi_device_quiesce(struct scsi_device *sdev)
2013 {
2014         int err = scsi_device_set_state(sdev, SDEV_QUIESCE);
2015         if (err)
2016                 return err;
2017
2018         scsi_run_queue(sdev->request_queue);
2019         while (sdev->device_busy) {
2020                 msleep_interruptible(200);
2021                 scsi_run_queue(sdev->request_queue);
2022         }
2023         return 0;
2024 }
2025 EXPORT_SYMBOL(scsi_device_quiesce);
2026
2027 /**
2028  *      scsi_device_resume - Restart user issued commands to a quiesced device.
2029  *      @sdev:  scsi device to resume.
2030  *
2031  *      Moves the device from quiesced back to running and restarts the
2032  *      queues.
2033  *
2034  *      Must be called with user context, may sleep.
2035  **/
2036 void
2037 scsi_device_resume(struct scsi_device *sdev)
2038 {
2039         if(scsi_device_set_state(sdev, SDEV_RUNNING))
2040                 return;
2041         scsi_run_queue(sdev->request_queue);
2042 }
2043 EXPORT_SYMBOL(scsi_device_resume);
2044
2045 static void
2046 device_quiesce_fn(struct scsi_device *sdev, void *data)
2047 {
2048         scsi_device_quiesce(sdev);
2049 }
2050
2051 void
2052 scsi_target_quiesce(struct scsi_target *starget)
2053 {
2054         starget_for_each_device(starget, NULL, device_quiesce_fn);
2055 }
2056 EXPORT_SYMBOL(scsi_target_quiesce);
2057
2058 static void
2059 device_resume_fn(struct scsi_device *sdev, void *data)
2060 {
2061         scsi_device_resume(sdev);
2062 }
2063
2064 void
2065 scsi_target_resume(struct scsi_target *starget)
2066 {
2067         starget_for_each_device(starget, NULL, device_resume_fn);
2068 }
2069 EXPORT_SYMBOL(scsi_target_resume);
2070
2071 /**
2072  * scsi_internal_device_block - internal function to put a device
2073  *                              temporarily into the SDEV_BLOCK state
2074  * @sdev:       device to block
2075  *
2076  * Block request made by scsi lld's to temporarily stop all
2077  * scsi commands on the specified device.  Called from interrupt
2078  * or normal process context.
2079  *
2080  * Returns zero if successful or error if not
2081  *
2082  * Notes:       
2083  *      This routine transitions the device to the SDEV_BLOCK state
2084  *      (which must be a legal transition).  When the device is in this
2085  *      state, all commands are deferred until the scsi lld reenables
2086  *      the device with scsi_device_unblock or device_block_tmo fires.
2087  *      This routine assumes the host_lock is held on entry.
2088  **/
2089 int
2090 scsi_internal_device_block(struct scsi_device *sdev)
2091 {
2092         request_queue_t *q = sdev->request_queue;
2093         unsigned long flags;
2094         int err = 0;
2095
2096         err = scsi_device_set_state(sdev, SDEV_BLOCK);
2097         if (err)
2098                 return err;
2099
2100         /* 
2101          * The device has transitioned to SDEV_BLOCK.  Stop the
2102          * block layer from calling the midlayer with this device's
2103          * request queue. 
2104          */
2105         spin_lock_irqsave(q->queue_lock, flags);
2106         blk_stop_queue(q);
2107         spin_unlock_irqrestore(q->queue_lock, flags);
2108
2109         return 0;
2110 }
2111 EXPORT_SYMBOL_GPL(scsi_internal_device_block);
2112  
2113 /**
2114  * scsi_internal_device_unblock - resume a device after a block request
2115  * @sdev:       device to resume
2116  *
2117  * Called by scsi lld's or the midlayer to restart the device queue
2118  * for the previously suspended scsi device.  Called from interrupt or
2119  * normal process context.
2120  *
2121  * Returns zero if successful or error if not.
2122  *
2123  * Notes:       
2124  *      This routine transitions the device to the SDEV_RUNNING state
2125  *      (which must be a legal transition) allowing the midlayer to
2126  *      goose the queue for this device.  This routine assumes the 
2127  *      host_lock is held upon entry.
2128  **/
2129 int
2130 scsi_internal_device_unblock(struct scsi_device *sdev)
2131 {
2132         request_queue_t *q = sdev->request_queue; 
2133         int err;
2134         unsigned long flags;
2135         
2136         /* 
2137          * Try to transition the scsi device to SDEV_RUNNING
2138          * and goose the device queue if successful.  
2139          */
2140         err = scsi_device_set_state(sdev, SDEV_RUNNING);
2141         if (err)
2142                 return err;
2143
2144         spin_lock_irqsave(q->queue_lock, flags);
2145         blk_start_queue(q);
2146         spin_unlock_irqrestore(q->queue_lock, flags);
2147
2148         return 0;
2149 }
2150 EXPORT_SYMBOL_GPL(scsi_internal_device_unblock);
2151
2152 static void
2153 device_block(struct scsi_device *sdev, void *data)
2154 {
2155         scsi_internal_device_block(sdev);
2156 }
2157
2158 static int
2159 target_block(struct device *dev, void *data)
2160 {
2161         if (scsi_is_target_device(dev))
2162                 starget_for_each_device(to_scsi_target(dev), NULL,
2163                                         device_block);
2164         return 0;
2165 }
2166
2167 void
2168 scsi_target_block(struct device *dev)
2169 {
2170         if (scsi_is_target_device(dev))
2171                 starget_for_each_device(to_scsi_target(dev), NULL,
2172                                         device_block);
2173         else
2174                 device_for_each_child(dev, NULL, target_block);
2175 }
2176 EXPORT_SYMBOL_GPL(scsi_target_block);
2177
2178 static void
2179 device_unblock(struct scsi_device *sdev, void *data)
2180 {
2181         scsi_internal_device_unblock(sdev);
2182 }
2183
2184 static int
2185 target_unblock(struct device *dev, void *data)
2186 {
2187         if (scsi_is_target_device(dev))
2188                 starget_for_each_device(to_scsi_target(dev), NULL,
2189                                         device_unblock);
2190         return 0;
2191 }
2192
2193 void
2194 scsi_target_unblock(struct device *dev)
2195 {
2196         if (scsi_is_target_device(dev))
2197                 starget_for_each_device(to_scsi_target(dev), NULL,
2198                                         device_unblock);
2199         else
2200                 device_for_each_child(dev, NULL, target_unblock);
2201 }
2202 EXPORT_SYMBOL_GPL(scsi_target_unblock);
2203
2204 /**
2205  * scsi_kmap_atomic_sg - find and atomically map an sg-elemnt
2206  * @sg:         scatter-gather list
2207  * @sg_count:   number of segments in sg
2208  * @offset:     offset in bytes into sg, on return offset into the mapped area
2209  * @len:        bytes to map, on return number of bytes mapped
2210  *
2211  * Returns virtual address of the start of the mapped page
2212  */
2213 void *scsi_kmap_atomic_sg(struct scatterlist *sg, int sg_count,
2214                           size_t *offset, size_t *len)
2215 {
2216         int i;
2217         size_t sg_len = 0, len_complete = 0;
2218         struct page *page;
2219
2220         for (i = 0; i < sg_count; i++) {
2221                 len_complete = sg_len; /* Complete sg-entries */
2222                 sg_len += sg[i].length;
2223                 if (sg_len > *offset)
2224                         break;
2225         }
2226
2227         if (unlikely(i == sg_count)) {
2228                 printk(KERN_ERR "%s: Bytes in sg: %zu, requested offset %zu, "
2229                         "elements %d\n",
2230                        __FUNCTION__, sg_len, *offset, sg_count);
2231                 WARN_ON(1);
2232                 return NULL;
2233         }
2234
2235         /* Offset starting from the beginning of first page in this sg-entry */
2236         *offset = *offset - len_complete + sg[i].offset;
2237
2238         /* Assumption: contiguous pages can be accessed as "page + i" */
2239         page = nth_page(sg[i].page, (*offset >> PAGE_SHIFT));
2240         *offset &= ~PAGE_MASK;
2241
2242         /* Bytes in this sg-entry from *offset to the end of the page */
2243         sg_len = PAGE_SIZE - *offset;
2244         if (*len > sg_len)
2245                 *len = sg_len;
2246
2247         return kmap_atomic(page, KM_BIO_SRC_IRQ);
2248 }
2249 EXPORT_SYMBOL(scsi_kmap_atomic_sg);
2250
2251 /**
2252  * scsi_kunmap_atomic_sg - atomically unmap a virtual address, previously
2253  *                         mapped with scsi_kmap_atomic_sg
2254  * @virt:       virtual address to be unmapped
2255  */
2256 void scsi_kunmap_atomic_sg(void *virt)
2257 {
2258         kunmap_atomic(virt, KM_BIO_SRC_IRQ);
2259 }
2260 EXPORT_SYMBOL(scsi_kunmap_atomic_sg);