Merge git://git.kernel.org/pub/scm/linux/kernel/git/lethal/sh-2.6
[linux-2.6] / drivers / scsi / scsi_lib.c
1 /*
2  *  scsi_lib.c Copyright (C) 1999 Eric Youngdale
3  *
4  *  SCSI queueing library.
5  *      Initial versions: Eric Youngdale (eric@andante.org).
6  *                        Based upon conversations with large numbers
7  *                        of people at Linux Expo.
8  */
9
10 #include <linux/bio.h>
11 #include <linux/blkdev.h>
12 #include <linux/completion.h>
13 #include <linux/kernel.h>
14 #include <linux/mempool.h>
15 #include <linux/slab.h>
16 #include <linux/init.h>
17 #include <linux/pci.h>
18 #include <linux/delay.h>
19 #include <linux/hardirq.h>
20 #include <linux/scatterlist.h>
21
22 #include <scsi/scsi.h>
23 #include <scsi/scsi_cmnd.h>
24 #include <scsi/scsi_dbg.h>
25 #include <scsi/scsi_device.h>
26 #include <scsi/scsi_driver.h>
27 #include <scsi/scsi_eh.h>
28 #include <scsi/scsi_host.h>
29
30 #include "scsi_priv.h"
31 #include "scsi_logging.h"
32
33
34 #define SG_MEMPOOL_NR           ARRAY_SIZE(scsi_sg_pools)
35 #define SG_MEMPOOL_SIZE         2
36
37 /*
38  * The maximum number of SG segments that we will put inside a scatterlist
39  * (unless chaining is used). Should ideally fit inside a single page, to
40  * avoid a higher order allocation.
41  */
42 #define SCSI_MAX_SG_SEGMENTS    128
43
44 struct scsi_host_sg_pool {
45         size_t          size;
46         char            *name;
47         struct kmem_cache       *slab;
48         mempool_t       *pool;
49 };
50
51 #define SP(x) { x, "sgpool-" #x }
52 static struct scsi_host_sg_pool scsi_sg_pools[] = {
53         SP(8),
54         SP(16),
55 #if (SCSI_MAX_SG_SEGMENTS > 16)
56         SP(32),
57 #if (SCSI_MAX_SG_SEGMENTS > 32)
58         SP(64),
59 #if (SCSI_MAX_SG_SEGMENTS > 64)
60         SP(128),
61 #endif
62 #endif
63 #endif
64 };
65 #undef SP
66
67 static void scsi_run_queue(struct request_queue *q);
68
69 /*
70  * Function:    scsi_unprep_request()
71  *
72  * Purpose:     Remove all preparation done for a request, including its
73  *              associated scsi_cmnd, so that it can be requeued.
74  *
75  * Arguments:   req     - request to unprepare
76  *
77  * Lock status: Assumed that no locks are held upon entry.
78  *
79  * Returns:     Nothing.
80  */
81 static void scsi_unprep_request(struct request *req)
82 {
83         struct scsi_cmnd *cmd = req->special;
84
85         req->cmd_flags &= ~REQ_DONTPREP;
86         req->special = NULL;
87
88         scsi_put_command(cmd);
89 }
90
91 /*
92  * Function:    scsi_queue_insert()
93  *
94  * Purpose:     Insert a command in the midlevel queue.
95  *
96  * Arguments:   cmd    - command that we are adding to queue.
97  *              reason - why we are inserting command to queue.
98  *
99  * Lock status: Assumed that lock is not held upon entry.
100  *
101  * Returns:     Nothing.
102  *
103  * Notes:       We do this for one of two cases.  Either the host is busy
104  *              and it cannot accept any more commands for the time being,
105  *              or the device returned QUEUE_FULL and can accept no more
106  *              commands.
107  * Notes:       This could be called either from an interrupt context or a
108  *              normal process context.
109  */
110 int scsi_queue_insert(struct scsi_cmnd *cmd, int reason)
111 {
112         struct Scsi_Host *host = cmd->device->host;
113         struct scsi_device *device = cmd->device;
114         struct request_queue *q = device->request_queue;
115         unsigned long flags;
116
117         SCSI_LOG_MLQUEUE(1,
118                  printk("Inserting command %p into mlqueue\n", cmd));
119
120         /*
121          * Set the appropriate busy bit for the device/host.
122          *
123          * If the host/device isn't busy, assume that something actually
124          * completed, and that we should be able to queue a command now.
125          *
126          * Note that the prior mid-layer assumption that any host could
127          * always queue at least one command is now broken.  The mid-layer
128          * will implement a user specifiable stall (see
129          * scsi_host.max_host_blocked and scsi_device.max_device_blocked)
130          * if a command is requeued with no other commands outstanding
131          * either for the device or for the host.
132          */
133         if (reason == SCSI_MLQUEUE_HOST_BUSY)
134                 host->host_blocked = host->max_host_blocked;
135         else if (reason == SCSI_MLQUEUE_DEVICE_BUSY)
136                 device->device_blocked = device->max_device_blocked;
137
138         /*
139          * Decrement the counters, since these commands are no longer
140          * active on the host/device.
141          */
142         scsi_device_unbusy(device);
143
144         /*
145          * Requeue this command.  It will go before all other commands
146          * that are already in the queue.
147          *
148          * NOTE: there is magic here about the way the queue is plugged if
149          * we have no outstanding commands.
150          * 
151          * Although we *don't* plug the queue, we call the request
152          * function.  The SCSI request function detects the blocked condition
153          * and plugs the queue appropriately.
154          */
155         spin_lock_irqsave(q->queue_lock, flags);
156         blk_requeue_request(q, cmd->request);
157         spin_unlock_irqrestore(q->queue_lock, flags);
158
159         scsi_run_queue(q);
160
161         return 0;
162 }
163
164 /**
165  * scsi_execute - insert request and wait for the result
166  * @sdev:       scsi device
167  * @cmd:        scsi command
168  * @data_direction: data direction
169  * @buffer:     data buffer
170  * @bufflen:    len of buffer
171  * @sense:      optional sense buffer
172  * @timeout:    request timeout in seconds
173  * @retries:    number of times to retry request
174  * @flags:      or into request flags;
175  *
176  * returns the req->errors value which is the scsi_cmnd result
177  * field.
178  */
179 int scsi_execute(struct scsi_device *sdev, const unsigned char *cmd,
180                  int data_direction, void *buffer, unsigned bufflen,
181                  unsigned char *sense, int timeout, int retries, int flags)
182 {
183         struct request *req;
184         int write = (data_direction == DMA_TO_DEVICE);
185         int ret = DRIVER_ERROR << 24;
186
187         req = blk_get_request(sdev->request_queue, write, __GFP_WAIT);
188
189         if (bufflen &&  blk_rq_map_kern(sdev->request_queue, req,
190                                         buffer, bufflen, __GFP_WAIT))
191                 goto out;
192
193         req->cmd_len = COMMAND_SIZE(cmd[0]);
194         memcpy(req->cmd, cmd, req->cmd_len);
195         req->sense = sense;
196         req->sense_len = 0;
197         req->retries = retries;
198         req->timeout = timeout;
199         req->cmd_type = REQ_TYPE_BLOCK_PC;
200         req->cmd_flags |= flags | REQ_QUIET | REQ_PREEMPT;
201
202         /*
203          * head injection *required* here otherwise quiesce won't work
204          */
205         blk_execute_rq(req->q, NULL, req, 1);
206
207         ret = req->errors;
208  out:
209         blk_put_request(req);
210
211         return ret;
212 }
213 EXPORT_SYMBOL(scsi_execute);
214
215
216 int scsi_execute_req(struct scsi_device *sdev, const unsigned char *cmd,
217                      int data_direction, void *buffer, unsigned bufflen,
218                      struct scsi_sense_hdr *sshdr, int timeout, int retries)
219 {
220         char *sense = NULL;
221         int result;
222         
223         if (sshdr) {
224                 sense = kzalloc(SCSI_SENSE_BUFFERSIZE, GFP_NOIO);
225                 if (!sense)
226                         return DRIVER_ERROR << 24;
227         }
228         result = scsi_execute(sdev, cmd, data_direction, buffer, bufflen,
229                               sense, timeout, retries, 0);
230         if (sshdr)
231                 scsi_normalize_sense(sense, SCSI_SENSE_BUFFERSIZE, sshdr);
232
233         kfree(sense);
234         return result;
235 }
236 EXPORT_SYMBOL(scsi_execute_req);
237
238 struct scsi_io_context {
239         void *data;
240         void (*done)(void *data, char *sense, int result, int resid);
241         char sense[SCSI_SENSE_BUFFERSIZE];
242 };
243
244 static struct kmem_cache *scsi_io_context_cache;
245
246 static void scsi_end_async(struct request *req, int uptodate)
247 {
248         struct scsi_io_context *sioc = req->end_io_data;
249
250         if (sioc->done)
251                 sioc->done(sioc->data, sioc->sense, req->errors, req->data_len);
252
253         kmem_cache_free(scsi_io_context_cache, sioc);
254         __blk_put_request(req->q, req);
255 }
256
257 static int scsi_merge_bio(struct request *rq, struct bio *bio)
258 {
259         struct request_queue *q = rq->q;
260
261         bio->bi_flags &= ~(1 << BIO_SEG_VALID);
262         if (rq_data_dir(rq) == WRITE)
263                 bio->bi_rw |= (1 << BIO_RW);
264         blk_queue_bounce(q, &bio);
265
266         return blk_rq_append_bio(q, rq, bio);
267 }
268
269 static void scsi_bi_endio(struct bio *bio, int error)
270 {
271         bio_put(bio);
272 }
273
274 /**
275  * scsi_req_map_sg - map a scatterlist into a request
276  * @rq:         request to fill
277  * @sgl:        scatterlist
278  * @nsegs:      number of elements
279  * @bufflen:    len of buffer
280  * @gfp:        memory allocation flags
281  *
282  * scsi_req_map_sg maps a scatterlist into a request so that the
283  * request can be sent to the block layer. We do not trust the scatterlist
284  * sent to use, as some ULDs use that struct to only organize the pages.
285  */
286 static int scsi_req_map_sg(struct request *rq, struct scatterlist *sgl,
287                            int nsegs, unsigned bufflen, gfp_t gfp)
288 {
289         struct request_queue *q = rq->q;
290         int nr_pages = (bufflen + sgl[0].offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
291         unsigned int data_len = bufflen, len, bytes, off;
292         struct scatterlist *sg;
293         struct page *page;
294         struct bio *bio = NULL;
295         int i, err, nr_vecs = 0;
296
297         for_each_sg(sgl, sg, nsegs, i) {
298                 page = sg_page(sg);
299                 off = sg->offset;
300                 len = sg->length;
301                 data_len += len;
302
303                 while (len > 0 && data_len > 0) {
304                         /*
305                          * sg sends a scatterlist that is larger than
306                          * the data_len it wants transferred for certain
307                          * IO sizes
308                          */
309                         bytes = min_t(unsigned int, len, PAGE_SIZE - off);
310                         bytes = min(bytes, data_len);
311
312                         if (!bio) {
313                                 nr_vecs = min_t(int, BIO_MAX_PAGES, nr_pages);
314                                 nr_pages -= nr_vecs;
315
316                                 bio = bio_alloc(gfp, nr_vecs);
317                                 if (!bio) {
318                                         err = -ENOMEM;
319                                         goto free_bios;
320                                 }
321                                 bio->bi_end_io = scsi_bi_endio;
322                         }
323
324                         if (bio_add_pc_page(q, bio, page, bytes, off) !=
325                             bytes) {
326                                 bio_put(bio);
327                                 err = -EINVAL;
328                                 goto free_bios;
329                         }
330
331                         if (bio->bi_vcnt >= nr_vecs) {
332                                 err = scsi_merge_bio(rq, bio);
333                                 if (err) {
334                                         bio_endio(bio, 0);
335                                         goto free_bios;
336                                 }
337                                 bio = NULL;
338                         }
339
340                         page++;
341                         len -= bytes;
342                         data_len -=bytes;
343                         off = 0;
344                 }
345         }
346
347         rq->buffer = rq->data = NULL;
348         rq->data_len = bufflen;
349         return 0;
350
351 free_bios:
352         while ((bio = rq->bio) != NULL) {
353                 rq->bio = bio->bi_next;
354                 /*
355                  * call endio instead of bio_put incase it was bounced
356                  */
357                 bio_endio(bio, 0);
358         }
359
360         return err;
361 }
362
363 /**
364  * scsi_execute_async - insert request
365  * @sdev:       scsi device
366  * @cmd:        scsi command
367  * @cmd_len:    length of scsi cdb
368  * @data_direction: DMA_TO_DEVICE, DMA_FROM_DEVICE, or DMA_NONE
369  * @buffer:     data buffer (this can be a kernel buffer or scatterlist)
370  * @bufflen:    len of buffer
371  * @use_sg:     if buffer is a scatterlist this is the number of elements
372  * @timeout:    request timeout in seconds
373  * @retries:    number of times to retry request
374  * @privdata:   data passed to done()
375  * @done:       callback function when done
376  * @gfp:        memory allocation flags
377  */
378 int scsi_execute_async(struct scsi_device *sdev, const unsigned char *cmd,
379                        int cmd_len, int data_direction, void *buffer, unsigned bufflen,
380                        int use_sg, int timeout, int retries, void *privdata,
381                        void (*done)(void *, char *, int, int), gfp_t gfp)
382 {
383         struct request *req;
384         struct scsi_io_context *sioc;
385         int err = 0;
386         int write = (data_direction == DMA_TO_DEVICE);
387
388         sioc = kmem_cache_zalloc(scsi_io_context_cache, gfp);
389         if (!sioc)
390                 return DRIVER_ERROR << 24;
391
392         req = blk_get_request(sdev->request_queue, write, gfp);
393         if (!req)
394                 goto free_sense;
395         req->cmd_type = REQ_TYPE_BLOCK_PC;
396         req->cmd_flags |= REQ_QUIET;
397
398         if (use_sg)
399                 err = scsi_req_map_sg(req, buffer, use_sg, bufflen, gfp);
400         else if (bufflen)
401                 err = blk_rq_map_kern(req->q, req, buffer, bufflen, gfp);
402
403         if (err)
404                 goto free_req;
405
406         req->cmd_len = cmd_len;
407         memset(req->cmd, 0, BLK_MAX_CDB); /* ATAPI hates garbage after CDB */
408         memcpy(req->cmd, cmd, req->cmd_len);
409         req->sense = sioc->sense;
410         req->sense_len = 0;
411         req->timeout = timeout;
412         req->retries = retries;
413         req->end_io_data = sioc;
414
415         sioc->data = privdata;
416         sioc->done = done;
417
418         blk_execute_rq_nowait(req->q, NULL, req, 1, scsi_end_async);
419         return 0;
420
421 free_req:
422         blk_put_request(req);
423 free_sense:
424         kmem_cache_free(scsi_io_context_cache, sioc);
425         return DRIVER_ERROR << 24;
426 }
427 EXPORT_SYMBOL_GPL(scsi_execute_async);
428
429 /*
430  * Function:    scsi_init_cmd_errh()
431  *
432  * Purpose:     Initialize cmd fields related to error handling.
433  *
434  * Arguments:   cmd     - command that is ready to be queued.
435  *
436  * Notes:       This function has the job of initializing a number of
437  *              fields related to error handling.   Typically this will
438  *              be called once for each command, as required.
439  */
440 static void scsi_init_cmd_errh(struct scsi_cmnd *cmd)
441 {
442         cmd->serial_number = 0;
443         cmd->resid = 0;
444         memset(cmd->sense_buffer, 0, SCSI_SENSE_BUFFERSIZE);
445         if (cmd->cmd_len == 0)
446                 cmd->cmd_len = COMMAND_SIZE(cmd->cmnd[0]);
447 }
448
449 void scsi_device_unbusy(struct scsi_device *sdev)
450 {
451         struct Scsi_Host *shost = sdev->host;
452         unsigned long flags;
453
454         spin_lock_irqsave(shost->host_lock, flags);
455         shost->host_busy--;
456         if (unlikely(scsi_host_in_recovery(shost) &&
457                      (shost->host_failed || shost->host_eh_scheduled)))
458                 scsi_eh_wakeup(shost);
459         spin_unlock(shost->host_lock);
460         spin_lock(sdev->request_queue->queue_lock);
461         sdev->device_busy--;
462         spin_unlock_irqrestore(sdev->request_queue->queue_lock, flags);
463 }
464
465 /*
466  * Called for single_lun devices on IO completion. Clear starget_sdev_user,
467  * and call blk_run_queue for all the scsi_devices on the target -
468  * including current_sdev first.
469  *
470  * Called with *no* scsi locks held.
471  */
472 static void scsi_single_lun_run(struct scsi_device *current_sdev)
473 {
474         struct Scsi_Host *shost = current_sdev->host;
475         struct scsi_device *sdev, *tmp;
476         struct scsi_target *starget = scsi_target(current_sdev);
477         unsigned long flags;
478
479         spin_lock_irqsave(shost->host_lock, flags);
480         starget->starget_sdev_user = NULL;
481         spin_unlock_irqrestore(shost->host_lock, flags);
482
483         /*
484          * Call blk_run_queue for all LUNs on the target, starting with
485          * current_sdev. We race with others (to set starget_sdev_user),
486          * but in most cases, we will be first. Ideally, each LU on the
487          * target would get some limited time or requests on the target.
488          */
489         blk_run_queue(current_sdev->request_queue);
490
491         spin_lock_irqsave(shost->host_lock, flags);
492         if (starget->starget_sdev_user)
493                 goto out;
494         list_for_each_entry_safe(sdev, tmp, &starget->devices,
495                         same_target_siblings) {
496                 if (sdev == current_sdev)
497                         continue;
498                 if (scsi_device_get(sdev))
499                         continue;
500
501                 spin_unlock_irqrestore(shost->host_lock, flags);
502                 blk_run_queue(sdev->request_queue);
503                 spin_lock_irqsave(shost->host_lock, flags);
504         
505                 scsi_device_put(sdev);
506         }
507  out:
508         spin_unlock_irqrestore(shost->host_lock, flags);
509 }
510
511 /*
512  * Function:    scsi_run_queue()
513  *
514  * Purpose:     Select a proper request queue to serve next
515  *
516  * Arguments:   q       - last request's queue
517  *
518  * Returns:     Nothing
519  *
520  * Notes:       The previous command was completely finished, start
521  *              a new one if possible.
522  */
523 static void scsi_run_queue(struct request_queue *q)
524 {
525         struct scsi_device *sdev = q->queuedata;
526         struct Scsi_Host *shost = sdev->host;
527         unsigned long flags;
528
529         if (scsi_target(sdev)->single_lun)
530                 scsi_single_lun_run(sdev);
531
532         spin_lock_irqsave(shost->host_lock, flags);
533         while (!list_empty(&shost->starved_list) &&
534                !shost->host_blocked && !shost->host_self_blocked &&
535                 !((shost->can_queue > 0) &&
536                   (shost->host_busy >= shost->can_queue))) {
537                 /*
538                  * As long as shost is accepting commands and we have
539                  * starved queues, call blk_run_queue. scsi_request_fn
540                  * drops the queue_lock and can add us back to the
541                  * starved_list.
542                  *
543                  * host_lock protects the starved_list and starved_entry.
544                  * scsi_request_fn must get the host_lock before checking
545                  * or modifying starved_list or starved_entry.
546                  */
547                 sdev = list_entry(shost->starved_list.next,
548                                           struct scsi_device, starved_entry);
549                 list_del_init(&sdev->starved_entry);
550                 spin_unlock_irqrestore(shost->host_lock, flags);
551
552
553                 if (test_bit(QUEUE_FLAG_REENTER, &q->queue_flags) &&
554                     !test_and_set_bit(QUEUE_FLAG_REENTER,
555                                       &sdev->request_queue->queue_flags)) {
556                         blk_run_queue(sdev->request_queue);
557                         clear_bit(QUEUE_FLAG_REENTER,
558                                   &sdev->request_queue->queue_flags);
559                 } else
560                         blk_run_queue(sdev->request_queue);
561
562                 spin_lock_irqsave(shost->host_lock, flags);
563                 if (unlikely(!list_empty(&sdev->starved_entry)))
564                         /*
565                          * sdev lost a race, and was put back on the
566                          * starved list. This is unlikely but without this
567                          * in theory we could loop forever.
568                          */
569                         break;
570         }
571         spin_unlock_irqrestore(shost->host_lock, flags);
572
573         blk_run_queue(q);
574 }
575
576 /*
577  * Function:    scsi_requeue_command()
578  *
579  * Purpose:     Handle post-processing of completed commands.
580  *
581  * Arguments:   q       - queue to operate on
582  *              cmd     - command that may need to be requeued.
583  *
584  * Returns:     Nothing
585  *
586  * Notes:       After command completion, there may be blocks left
587  *              over which weren't finished by the previous command
588  *              this can be for a number of reasons - the main one is
589  *              I/O errors in the middle of the request, in which case
590  *              we need to request the blocks that come after the bad
591  *              sector.
592  * Notes:       Upon return, cmd is a stale pointer.
593  */
594 static void scsi_requeue_command(struct request_queue *q, struct scsi_cmnd *cmd)
595 {
596         struct request *req = cmd->request;
597         unsigned long flags;
598
599         scsi_unprep_request(req);
600         spin_lock_irqsave(q->queue_lock, flags);
601         blk_requeue_request(q, req);
602         spin_unlock_irqrestore(q->queue_lock, flags);
603
604         scsi_run_queue(q);
605 }
606
607 void scsi_next_command(struct scsi_cmnd *cmd)
608 {
609         struct scsi_device *sdev = cmd->device;
610         struct request_queue *q = sdev->request_queue;
611
612         /* need to hold a reference on the device before we let go of the cmd */
613         get_device(&sdev->sdev_gendev);
614
615         scsi_put_command(cmd);
616         scsi_run_queue(q);
617
618         /* ok to remove device now */
619         put_device(&sdev->sdev_gendev);
620 }
621
622 void scsi_run_host_queues(struct Scsi_Host *shost)
623 {
624         struct scsi_device *sdev;
625
626         shost_for_each_device(sdev, shost)
627                 scsi_run_queue(sdev->request_queue);
628 }
629
630 /*
631  * Function:    scsi_end_request()
632  *
633  * Purpose:     Post-processing of completed commands (usually invoked at end
634  *              of upper level post-processing and scsi_io_completion).
635  *
636  * Arguments:   cmd      - command that is complete.
637  *              error    - 0 if I/O indicates success, < 0 for I/O error.
638  *              bytes    - number of bytes of completed I/O
639  *              requeue  - indicates whether we should requeue leftovers.
640  *
641  * Lock status: Assumed that lock is not held upon entry.
642  *
643  * Returns:     cmd if requeue required, NULL otherwise.
644  *
645  * Notes:       This is called for block device requests in order to
646  *              mark some number of sectors as complete.
647  * 
648  *              We are guaranteeing that the request queue will be goosed
649  *              at some point during this call.
650  * Notes:       If cmd was requeued, upon return it will be a stale pointer.
651  */
652 static struct scsi_cmnd *scsi_end_request(struct scsi_cmnd *cmd, int error,
653                                           int bytes, int requeue)
654 {
655         struct request_queue *q = cmd->device->request_queue;
656         struct request *req = cmd->request;
657
658         /*
659          * If there are blocks left over at the end, set up the command
660          * to queue the remainder of them.
661          */
662         if (blk_end_request(req, error, bytes)) {
663                 int leftover = (req->hard_nr_sectors << 9);
664
665                 if (blk_pc_request(req))
666                         leftover = req->data_len;
667
668                 /* kill remainder if no retrys */
669                 if (error && blk_noretry_request(req))
670                         blk_end_request(req, error, leftover);
671                 else {
672                         if (requeue) {
673                                 /*
674                                  * Bleah.  Leftovers again.  Stick the
675                                  * leftovers in the front of the
676                                  * queue, and goose the queue again.
677                                  */
678                                 scsi_requeue_command(q, cmd);
679                                 cmd = NULL;
680                         }
681                         return cmd;
682                 }
683         }
684
685         /*
686          * This will goose the queue request function at the end, so we don't
687          * need to worry about launching another command.
688          */
689         scsi_next_command(cmd);
690         return NULL;
691 }
692
693 /*
694  * Like SCSI_MAX_SG_SEGMENTS, but for archs that have sg chaining. This limit
695  * is totally arbitrary, a setting of 2048 will get you at least 8mb ios.
696  */
697 #define SCSI_MAX_SG_CHAIN_SEGMENTS      2048
698
699 static inline unsigned int scsi_sgtable_index(unsigned short nents)
700 {
701         unsigned int index;
702
703         switch (nents) {
704         case 1 ... 8:
705                 index = 0;
706                 break;
707         case 9 ... 16:
708                 index = 1;
709                 break;
710 #if (SCSI_MAX_SG_SEGMENTS > 16)
711         case 17 ... 32:
712                 index = 2;
713                 break;
714 #if (SCSI_MAX_SG_SEGMENTS > 32)
715         case 33 ... 64:
716                 index = 3;
717                 break;
718 #if (SCSI_MAX_SG_SEGMENTS > 64)
719         case 65 ... 128:
720                 index = 4;
721                 break;
722 #endif
723 #endif
724 #endif
725         default:
726                 printk(KERN_ERR "scsi: bad segment count=%d\n", nents);
727                 BUG();
728         }
729
730         return index;
731 }
732
733 static void scsi_sg_free(struct scatterlist *sgl, unsigned int nents)
734 {
735         struct scsi_host_sg_pool *sgp;
736
737         sgp = scsi_sg_pools + scsi_sgtable_index(nents);
738         mempool_free(sgl, sgp->pool);
739 }
740
741 static struct scatterlist *scsi_sg_alloc(unsigned int nents, gfp_t gfp_mask)
742 {
743         struct scsi_host_sg_pool *sgp;
744
745         sgp = scsi_sg_pools + scsi_sgtable_index(nents);
746         return mempool_alloc(sgp->pool, gfp_mask);
747 }
748
749 int scsi_alloc_sgtable(struct scsi_cmnd *cmd, gfp_t gfp_mask)
750 {
751         int ret;
752
753         BUG_ON(!cmd->use_sg);
754
755         ret = __sg_alloc_table(&cmd->sg_table, cmd->use_sg,
756                                SCSI_MAX_SG_SEGMENTS, gfp_mask, scsi_sg_alloc);
757         if (unlikely(ret))
758                 __sg_free_table(&cmd->sg_table, SCSI_MAX_SG_SEGMENTS,
759                                 scsi_sg_free);
760
761         cmd->request_buffer = cmd->sg_table.sgl;
762         return ret;
763 }
764
765 EXPORT_SYMBOL(scsi_alloc_sgtable);
766
767 void scsi_free_sgtable(struct scsi_cmnd *cmd)
768 {
769         __sg_free_table(&cmd->sg_table, SCSI_MAX_SG_SEGMENTS, scsi_sg_free);
770 }
771
772 EXPORT_SYMBOL(scsi_free_sgtable);
773
774 /*
775  * Function:    scsi_release_buffers()
776  *
777  * Purpose:     Completion processing for block device I/O requests.
778  *
779  * Arguments:   cmd     - command that we are bailing.
780  *
781  * Lock status: Assumed that no lock is held upon entry.
782  *
783  * Returns:     Nothing
784  *
785  * Notes:       In the event that an upper level driver rejects a
786  *              command, we must release resources allocated during
787  *              the __init_io() function.  Primarily this would involve
788  *              the scatter-gather table, and potentially any bounce
789  *              buffers.
790  */
791 static void scsi_release_buffers(struct scsi_cmnd *cmd)
792 {
793         if (cmd->use_sg)
794                 scsi_free_sgtable(cmd);
795
796         /*
797          * Zero these out.  They now point to freed memory, and it is
798          * dangerous to hang onto the pointers.
799          */
800         cmd->request_buffer = NULL;
801         cmd->request_bufflen = 0;
802 }
803
804 /*
805  * Function:    scsi_io_completion()
806  *
807  * Purpose:     Completion processing for block device I/O requests.
808  *
809  * Arguments:   cmd   - command that is finished.
810  *
811  * Lock status: Assumed that no lock is held upon entry.
812  *
813  * Returns:     Nothing
814  *
815  * Notes:       This function is matched in terms of capabilities to
816  *              the function that created the scatter-gather list.
817  *              In other words, if there are no bounce buffers
818  *              (the normal case for most drivers), we don't need
819  *              the logic to deal with cleaning up afterwards.
820  *
821  *              We must do one of several things here:
822  *
823  *              a) Call scsi_end_request.  This will finish off the
824  *                 specified number of sectors.  If we are done, the
825  *                 command block will be released, and the queue
826  *                 function will be goosed.  If we are not done, then
827  *                 scsi_end_request will directly goose the queue.
828  *
829  *              b) We can just use scsi_requeue_command() here.  This would
830  *                 be used if we just wanted to retry, for example.
831  */
832 void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes)
833 {
834         int result = cmd->result;
835         int this_count = cmd->request_bufflen;
836         struct request_queue *q = cmd->device->request_queue;
837         struct request *req = cmd->request;
838         int clear_errors = 1;
839         struct scsi_sense_hdr sshdr;
840         int sense_valid = 0;
841         int sense_deferred = 0;
842
843         scsi_release_buffers(cmd);
844
845         if (result) {
846                 sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
847                 if (sense_valid)
848                         sense_deferred = scsi_sense_is_deferred(&sshdr);
849         }
850
851         if (blk_pc_request(req)) { /* SG_IO ioctl from block level */
852                 req->errors = result;
853                 if (result) {
854                         clear_errors = 0;
855                         if (sense_valid && req->sense) {
856                                 /*
857                                  * SG_IO wants current and deferred errors
858                                  */
859                                 int len = 8 + cmd->sense_buffer[7];
860
861                                 if (len > SCSI_SENSE_BUFFERSIZE)
862                                         len = SCSI_SENSE_BUFFERSIZE;
863                                 memcpy(req->sense, cmd->sense_buffer,  len);
864                                 req->sense_len = len;
865                         }
866                 }
867                 req->data_len = cmd->resid;
868         }
869
870         /*
871          * Next deal with any sectors which we were able to correctly
872          * handle.
873          */
874         SCSI_LOG_HLCOMPLETE(1, printk("%ld sectors total, "
875                                       "%d bytes done.\n",
876                                       req->nr_sectors, good_bytes));
877         SCSI_LOG_HLCOMPLETE(1, printk("use_sg is %d\n", cmd->use_sg));
878
879         if (clear_errors)
880                 req->errors = 0;
881
882         /* A number of bytes were successfully read.  If there
883          * are leftovers and there is some kind of error
884          * (result != 0), retry the rest.
885          */
886         if (scsi_end_request(cmd, 0, good_bytes, result == 0) == NULL)
887                 return;
888
889         /* good_bytes = 0, or (inclusive) there were leftovers and
890          * result = 0, so scsi_end_request couldn't retry.
891          */
892         if (sense_valid && !sense_deferred) {
893                 switch (sshdr.sense_key) {
894                 case UNIT_ATTENTION:
895                         if (cmd->device->removable) {
896                                 /* Detected disc change.  Set a bit
897                                  * and quietly refuse further access.
898                                  */
899                                 cmd->device->changed = 1;
900                                 scsi_end_request(cmd, -EIO, this_count, 1);
901                                 return;
902                         } else {
903                                 /* Must have been a power glitch, or a
904                                  * bus reset.  Could not have been a
905                                  * media change, so we just retry the
906                                  * request and see what happens.
907                                  */
908                                 scsi_requeue_command(q, cmd);
909                                 return;
910                         }
911                         break;
912                 case ILLEGAL_REQUEST:
913                         /* If we had an ILLEGAL REQUEST returned, then
914                          * we may have performed an unsupported
915                          * command.  The only thing this should be
916                          * would be a ten byte read where only a six
917                          * byte read was supported.  Also, on a system
918                          * where READ CAPACITY failed, we may have
919                          * read past the end of the disk.
920                          */
921                         if ((cmd->device->use_10_for_rw &&
922                             sshdr.asc == 0x20 && sshdr.ascq == 0x00) &&
923                             (cmd->cmnd[0] == READ_10 ||
924                              cmd->cmnd[0] == WRITE_10)) {
925                                 cmd->device->use_10_for_rw = 0;
926                                 /* This will cause a retry with a
927                                  * 6-byte command.
928                                  */
929                                 scsi_requeue_command(q, cmd);
930                                 return;
931                         } else {
932                                 scsi_end_request(cmd, -EIO, this_count, 1);
933                                 return;
934                         }
935                         break;
936                 case NOT_READY:
937                         /* If the device is in the process of becoming
938                          * ready, or has a temporary blockage, retry.
939                          */
940                         if (sshdr.asc == 0x04) {
941                                 switch (sshdr.ascq) {
942                                 case 0x01: /* becoming ready */
943                                 case 0x04: /* format in progress */
944                                 case 0x05: /* rebuild in progress */
945                                 case 0x06: /* recalculation in progress */
946                                 case 0x07: /* operation in progress */
947                                 case 0x08: /* Long write in progress */
948                                 case 0x09: /* self test in progress */
949                                         scsi_requeue_command(q, cmd);
950                                         return;
951                                 default:
952                                         break;
953                                 }
954                         }
955                         if (!(req->cmd_flags & REQ_QUIET))
956                                 scsi_cmd_print_sense_hdr(cmd,
957                                                          "Device not ready",
958                                                          &sshdr);
959
960                         scsi_end_request(cmd, -EIO, this_count, 1);
961                         return;
962                 case VOLUME_OVERFLOW:
963                         if (!(req->cmd_flags & REQ_QUIET)) {
964                                 scmd_printk(KERN_INFO, cmd,
965                                             "Volume overflow, CDB: ");
966                                 __scsi_print_command(cmd->cmnd);
967                                 scsi_print_sense("", cmd);
968                         }
969                         /* See SSC3rXX or current. */
970                         scsi_end_request(cmd, -EIO, this_count, 1);
971                         return;
972                 default:
973                         break;
974                 }
975         }
976         if (host_byte(result) == DID_RESET) {
977                 /* Third party bus reset or reset for error recovery
978                  * reasons.  Just retry the request and see what
979                  * happens.
980                  */
981                 scsi_requeue_command(q, cmd);
982                 return;
983         }
984         if (result) {
985                 if (!(req->cmd_flags & REQ_QUIET)) {
986                         scsi_print_result(cmd);
987                         if (driver_byte(result) & DRIVER_SENSE)
988                                 scsi_print_sense("", cmd);
989                 }
990         }
991         scsi_end_request(cmd, -EIO, this_count, !result);
992 }
993
994 /*
995  * Function:    scsi_init_io()
996  *
997  * Purpose:     SCSI I/O initialize function.
998  *
999  * Arguments:   cmd   - Command descriptor we wish to initialize
1000  *
1001  * Returns:     0 on success
1002  *              BLKPREP_DEFER if the failure is retryable
1003  */
1004 static int scsi_init_io(struct scsi_cmnd *cmd)
1005 {
1006         struct request     *req = cmd->request;
1007         int                count;
1008
1009         /*
1010          * We used to not use scatter-gather for single segment request,
1011          * but now we do (it makes highmem I/O easier to support without
1012          * kmapping pages)
1013          */
1014         cmd->use_sg = req->nr_phys_segments;
1015
1016         /*
1017          * If sg table allocation fails, requeue request later.
1018          */
1019         if (unlikely(scsi_alloc_sgtable(cmd, GFP_ATOMIC))) {
1020                 scsi_unprep_request(req);
1021                 return BLKPREP_DEFER;
1022         }
1023
1024         req->buffer = NULL;
1025         if (blk_pc_request(req))
1026                 cmd->request_bufflen = req->data_len;
1027         else
1028                 cmd->request_bufflen = req->nr_sectors << 9;
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         BUG_ON(count > cmd->use_sg);
1036         cmd->use_sg = count;
1037         return BLKPREP_OK;
1038 }
1039
1040 static struct scsi_cmnd *scsi_get_cmd_from_req(struct scsi_device *sdev,
1041                 struct request *req)
1042 {
1043         struct scsi_cmnd *cmd;
1044
1045         if (!req->special) {
1046                 cmd = scsi_get_command(sdev, GFP_ATOMIC);
1047                 if (unlikely(!cmd))
1048                         return NULL;
1049                 req->special = cmd;
1050         } else {
1051                 cmd = req->special;
1052         }
1053
1054         /* pull a tag out of the request if we have one */
1055         cmd->tag = req->tag;
1056         cmd->request = req;
1057
1058         return cmd;
1059 }
1060
1061 int scsi_setup_blk_pc_cmnd(struct scsi_device *sdev, struct request *req)
1062 {
1063         struct scsi_cmnd *cmd;
1064         int ret = scsi_prep_state_check(sdev, req);
1065
1066         if (ret != BLKPREP_OK)
1067                 return ret;
1068
1069         cmd = scsi_get_cmd_from_req(sdev, req);
1070         if (unlikely(!cmd))
1071                 return BLKPREP_DEFER;
1072
1073         /*
1074          * BLOCK_PC requests may transfer data, in which case they must
1075          * a bio attached to them.  Or they might contain a SCSI command
1076          * that does not transfer data, in which case they may optionally
1077          * submit a request without an attached bio.
1078          */
1079         if (req->bio) {
1080                 int ret;
1081
1082                 BUG_ON(!req->nr_phys_segments);
1083
1084                 ret = scsi_init_io(cmd);
1085                 if (unlikely(ret))
1086                         return ret;
1087         } else {
1088                 BUG_ON(req->data_len);
1089                 BUG_ON(req->data);
1090
1091                 cmd->request_bufflen = 0;
1092                 cmd->request_buffer = NULL;
1093                 cmd->use_sg = 0;
1094                 req->buffer = NULL;
1095         }
1096
1097         BUILD_BUG_ON(sizeof(req->cmd) > sizeof(cmd->cmnd));
1098         memcpy(cmd->cmnd, req->cmd, sizeof(cmd->cmnd));
1099         cmd->cmd_len = req->cmd_len;
1100         if (!req->data_len)
1101                 cmd->sc_data_direction = DMA_NONE;
1102         else if (rq_data_dir(req) == WRITE)
1103                 cmd->sc_data_direction = DMA_TO_DEVICE;
1104         else
1105                 cmd->sc_data_direction = DMA_FROM_DEVICE;
1106         
1107         cmd->transfersize = req->data_len;
1108         cmd->allowed = req->retries;
1109         cmd->timeout_per_command = req->timeout;
1110         return BLKPREP_OK;
1111 }
1112 EXPORT_SYMBOL(scsi_setup_blk_pc_cmnd);
1113
1114 /*
1115  * Setup a REQ_TYPE_FS command.  These are simple read/write request
1116  * from filesystems that still need to be translated to SCSI CDBs from
1117  * the ULD.
1118  */
1119 int scsi_setup_fs_cmnd(struct scsi_device *sdev, struct request *req)
1120 {
1121         struct scsi_cmnd *cmd;
1122         int ret = scsi_prep_state_check(sdev, req);
1123
1124         if (ret != BLKPREP_OK)
1125                 return ret;
1126         /*
1127          * Filesystem requests must transfer data.
1128          */
1129         BUG_ON(!req->nr_phys_segments);
1130
1131         cmd = scsi_get_cmd_from_req(sdev, req);
1132         if (unlikely(!cmd))
1133                 return BLKPREP_DEFER;
1134
1135         return scsi_init_io(cmd);
1136 }
1137 EXPORT_SYMBOL(scsi_setup_fs_cmnd);
1138
1139 int scsi_prep_state_check(struct scsi_device *sdev, struct request *req)
1140 {
1141         int ret = BLKPREP_OK;
1142
1143         /*
1144          * If the device is not in running state we will reject some
1145          * or all commands.
1146          */
1147         if (unlikely(sdev->sdev_state != SDEV_RUNNING)) {
1148                 switch (sdev->sdev_state) {
1149                 case SDEV_OFFLINE:
1150                         /*
1151                          * If the device is offline we refuse to process any
1152                          * commands.  The device must be brought online
1153                          * before trying any recovery commands.
1154                          */
1155                         sdev_printk(KERN_ERR, sdev,
1156                                     "rejecting I/O to offline device\n");
1157                         ret = BLKPREP_KILL;
1158                         break;
1159                 case SDEV_DEL:
1160                         /*
1161                          * If the device is fully deleted, we refuse to
1162                          * process any commands as well.
1163                          */
1164                         sdev_printk(KERN_ERR, sdev,
1165                                     "rejecting I/O to dead device\n");
1166                         ret = BLKPREP_KILL;
1167                         break;
1168                 case SDEV_QUIESCE:
1169                 case SDEV_BLOCK:
1170                         /*
1171                          * If the devices is blocked we defer normal commands.
1172                          */
1173                         if (!(req->cmd_flags & REQ_PREEMPT))
1174                                 ret = BLKPREP_DEFER;
1175                         break;
1176                 default:
1177                         /*
1178                          * For any other not fully online state we only allow
1179                          * special commands.  In particular any user initiated
1180                          * command is not allowed.
1181                          */
1182                         if (!(req->cmd_flags & REQ_PREEMPT))
1183                                 ret = BLKPREP_KILL;
1184                         break;
1185                 }
1186         }
1187         return ret;
1188 }
1189 EXPORT_SYMBOL(scsi_prep_state_check);
1190
1191 int scsi_prep_return(struct request_queue *q, struct request *req, int ret)
1192 {
1193         struct scsi_device *sdev = q->queuedata;
1194
1195         switch (ret) {
1196         case BLKPREP_KILL:
1197                 req->errors = DID_NO_CONNECT << 16;
1198                 /* release the command and kill it */
1199                 if (req->special) {
1200                         struct scsi_cmnd *cmd = req->special;
1201                         scsi_release_buffers(cmd);
1202                         scsi_put_command(cmd);
1203                         req->special = NULL;
1204                 }
1205                 break;
1206         case BLKPREP_DEFER:
1207                 /*
1208                  * If we defer, the elv_next_request() returns NULL, but the
1209                  * queue must be restarted, so we plug here if no returning
1210                  * command will automatically do that.
1211                  */
1212                 if (sdev->device_busy == 0)
1213                         blk_plug_device(q);
1214                 break;
1215         default:
1216                 req->cmd_flags |= REQ_DONTPREP;
1217         }
1218
1219         return ret;
1220 }
1221 EXPORT_SYMBOL(scsi_prep_return);
1222
1223 int scsi_prep_fn(struct request_queue *q, struct request *req)
1224 {
1225         struct scsi_device *sdev = q->queuedata;
1226         int ret = BLKPREP_KILL;
1227
1228         if (req->cmd_type == REQ_TYPE_BLOCK_PC)
1229                 ret = scsi_setup_blk_pc_cmnd(sdev, req);
1230         return scsi_prep_return(q, req, ret);
1231 }
1232
1233 /*
1234  * scsi_dev_queue_ready: if we can send requests to sdev, return 1 else
1235  * return 0.
1236  *
1237  * Called with the queue_lock held.
1238  */
1239 static inline int scsi_dev_queue_ready(struct request_queue *q,
1240                                   struct scsi_device *sdev)
1241 {
1242         if (sdev->device_busy >= sdev->queue_depth)
1243                 return 0;
1244         if (sdev->device_busy == 0 && sdev->device_blocked) {
1245                 /*
1246                  * unblock after device_blocked iterates to zero
1247                  */
1248                 if (--sdev->device_blocked == 0) {
1249                         SCSI_LOG_MLQUEUE(3,
1250                                    sdev_printk(KERN_INFO, sdev,
1251                                    "unblocking device at zero depth\n"));
1252                 } else {
1253                         blk_plug_device(q);
1254                         return 0;
1255                 }
1256         }
1257         if (sdev->device_blocked)
1258                 return 0;
1259
1260         return 1;
1261 }
1262
1263 /*
1264  * scsi_host_queue_ready: if we can send requests to shost, return 1 else
1265  * return 0. We must end up running the queue again whenever 0 is
1266  * returned, else IO can hang.
1267  *
1268  * Called with host_lock held.
1269  */
1270 static inline int scsi_host_queue_ready(struct request_queue *q,
1271                                    struct Scsi_Host *shost,
1272                                    struct scsi_device *sdev)
1273 {
1274         if (scsi_host_in_recovery(shost))
1275                 return 0;
1276         if (shost->host_busy == 0 && shost->host_blocked) {
1277                 /*
1278                  * unblock after host_blocked iterates to zero
1279                  */
1280                 if (--shost->host_blocked == 0) {
1281                         SCSI_LOG_MLQUEUE(3,
1282                                 printk("scsi%d unblocking host at zero depth\n",
1283                                         shost->host_no));
1284                 } else {
1285                         blk_plug_device(q);
1286                         return 0;
1287                 }
1288         }
1289         if ((shost->can_queue > 0 && shost->host_busy >= shost->can_queue) ||
1290             shost->host_blocked || shost->host_self_blocked) {
1291                 if (list_empty(&sdev->starved_entry))
1292                         list_add_tail(&sdev->starved_entry, &shost->starved_list);
1293                 return 0;
1294         }
1295
1296         /* We're OK to process the command, so we can't be starved */
1297         if (!list_empty(&sdev->starved_entry))
1298                 list_del_init(&sdev->starved_entry);
1299
1300         return 1;
1301 }
1302
1303 /*
1304  * Kill a request for a dead device
1305  */
1306 static void scsi_kill_request(struct request *req, struct request_queue *q)
1307 {
1308         struct scsi_cmnd *cmd = req->special;
1309         struct scsi_device *sdev = cmd->device;
1310         struct Scsi_Host *shost = sdev->host;
1311
1312         blkdev_dequeue_request(req);
1313
1314         if (unlikely(cmd == NULL)) {
1315                 printk(KERN_CRIT "impossible request in %s.\n",
1316                                  __FUNCTION__);
1317                 BUG();
1318         }
1319
1320         scsi_init_cmd_errh(cmd);
1321         cmd->result = DID_NO_CONNECT << 16;
1322         atomic_inc(&cmd->device->iorequest_cnt);
1323
1324         /*
1325          * SCSI request completion path will do scsi_device_unbusy(),
1326          * bump busy counts.  To bump the counters, we need to dance
1327          * with the locks as normal issue path does.
1328          */
1329         sdev->device_busy++;
1330         spin_unlock(sdev->request_queue->queue_lock);
1331         spin_lock(shost->host_lock);
1332         shost->host_busy++;
1333         spin_unlock(shost->host_lock);
1334         spin_lock(sdev->request_queue->queue_lock);
1335
1336         __scsi_done(cmd);
1337 }
1338
1339 static void scsi_softirq_done(struct request *rq)
1340 {
1341         struct scsi_cmnd *cmd = rq->completion_data;
1342         unsigned long wait_for = (cmd->allowed + 1) * cmd->timeout_per_command;
1343         int disposition;
1344
1345         INIT_LIST_HEAD(&cmd->eh_entry);
1346
1347         disposition = scsi_decide_disposition(cmd);
1348         if (disposition != SUCCESS &&
1349             time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) {
1350                 sdev_printk(KERN_ERR, cmd->device,
1351                             "timing out command, waited %lus\n",
1352                             wait_for/HZ);
1353                 disposition = SUCCESS;
1354         }
1355                         
1356         scsi_log_completion(cmd, disposition);
1357
1358         switch (disposition) {
1359                 case SUCCESS:
1360                         scsi_finish_command(cmd);
1361                         break;
1362                 case NEEDS_RETRY:
1363                         scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY);
1364                         break;
1365                 case ADD_TO_MLQUEUE:
1366                         scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY);
1367                         break;
1368                 default:
1369                         if (!scsi_eh_scmd_add(cmd, 0))
1370                                 scsi_finish_command(cmd);
1371         }
1372 }
1373
1374 /*
1375  * Function:    scsi_request_fn()
1376  *
1377  * Purpose:     Main strategy routine for SCSI.
1378  *
1379  * Arguments:   q       - Pointer to actual queue.
1380  *
1381  * Returns:     Nothing
1382  *
1383  * Lock status: IO request lock assumed to be held when called.
1384  */
1385 static void scsi_request_fn(struct request_queue *q)
1386 {
1387         struct scsi_device *sdev = q->queuedata;
1388         struct Scsi_Host *shost;
1389         struct scsi_cmnd *cmd;
1390         struct request *req;
1391
1392         if (!sdev) {
1393                 printk("scsi: killing requests for dead queue\n");
1394                 while ((req = elv_next_request(q)) != NULL)
1395                         scsi_kill_request(req, q);
1396                 return;
1397         }
1398
1399         if(!get_device(&sdev->sdev_gendev))
1400                 /* We must be tearing the block queue down already */
1401                 return;
1402
1403         /*
1404          * To start with, we keep looping until the queue is empty, or until
1405          * the host is no longer able to accept any more requests.
1406          */
1407         shost = sdev->host;
1408         while (!blk_queue_plugged(q)) {
1409                 int rtn;
1410                 /*
1411                  * get next queueable request.  We do this early to make sure
1412                  * that the request is fully prepared even if we cannot 
1413                  * accept it.
1414                  */
1415                 req = elv_next_request(q);
1416                 if (!req || !scsi_dev_queue_ready(q, sdev))
1417                         break;
1418
1419                 if (unlikely(!scsi_device_online(sdev))) {
1420                         sdev_printk(KERN_ERR, sdev,
1421                                     "rejecting I/O to offline device\n");
1422                         scsi_kill_request(req, q);
1423                         continue;
1424                 }
1425
1426
1427                 /*
1428                  * Remove the request from the request list.
1429                  */
1430                 if (!(blk_queue_tagged(q) && !blk_queue_start_tag(q, req)))
1431                         blkdev_dequeue_request(req);
1432                 sdev->device_busy++;
1433
1434                 spin_unlock(q->queue_lock);
1435                 cmd = req->special;
1436                 if (unlikely(cmd == NULL)) {
1437                         printk(KERN_CRIT "impossible request in %s.\n"
1438                                          "please mail a stack trace to "
1439                                          "linux-scsi@vger.kernel.org\n",
1440                                          __FUNCTION__);
1441                         blk_dump_rq_flags(req, "foo");
1442                         BUG();
1443                 }
1444                 spin_lock(shost->host_lock);
1445
1446                 if (!scsi_host_queue_ready(q, shost, sdev))
1447                         goto not_ready;
1448                 if (scsi_target(sdev)->single_lun) {
1449                         if (scsi_target(sdev)->starget_sdev_user &&
1450                             scsi_target(sdev)->starget_sdev_user != sdev)
1451                                 goto not_ready;
1452                         scsi_target(sdev)->starget_sdev_user = sdev;
1453                 }
1454                 shost->host_busy++;
1455
1456                 /*
1457                  * XXX(hch): This is rather suboptimal, scsi_dispatch_cmd will
1458                  *              take the lock again.
1459                  */
1460                 spin_unlock_irq(shost->host_lock);
1461
1462                 /*
1463                  * Finally, initialize any error handling parameters, and set up
1464                  * the timers for timeouts.
1465                  */
1466                 scsi_init_cmd_errh(cmd);
1467
1468                 /*
1469                  * Dispatch the command to the low-level driver.
1470                  */
1471                 rtn = scsi_dispatch_cmd(cmd);
1472                 spin_lock_irq(q->queue_lock);
1473                 if(rtn) {
1474                         /* we're refusing the command; because of
1475                          * the way locks get dropped, we need to 
1476                          * check here if plugging is required */
1477                         if(sdev->device_busy == 0)
1478                                 blk_plug_device(q);
1479
1480                         break;
1481                 }
1482         }
1483
1484         goto out;
1485
1486  not_ready:
1487         spin_unlock_irq(shost->host_lock);
1488
1489         /*
1490          * lock q, handle tag, requeue req, and decrement device_busy. We
1491          * must return with queue_lock held.
1492          *
1493          * Decrementing device_busy without checking it is OK, as all such
1494          * cases (host limits or settings) should run the queue at some
1495          * later time.
1496          */
1497         spin_lock_irq(q->queue_lock);
1498         blk_requeue_request(q, req);
1499         sdev->device_busy--;
1500         if(sdev->device_busy == 0)
1501                 blk_plug_device(q);
1502  out:
1503         /* must be careful here...if we trigger the ->remove() function
1504          * we cannot be holding the q lock */
1505         spin_unlock_irq(q->queue_lock);
1506         put_device(&sdev->sdev_gendev);
1507         spin_lock_irq(q->queue_lock);
1508 }
1509
1510 u64 scsi_calculate_bounce_limit(struct Scsi_Host *shost)
1511 {
1512         struct device *host_dev;
1513         u64 bounce_limit = 0xffffffff;
1514
1515         if (shost->unchecked_isa_dma)
1516                 return BLK_BOUNCE_ISA;
1517         /*
1518          * Platforms with virtual-DMA translation
1519          * hardware have no practical limit.
1520          */
1521         if (!PCI_DMA_BUS_IS_PHYS)
1522                 return BLK_BOUNCE_ANY;
1523
1524         host_dev = scsi_get_device(shost);
1525         if (host_dev && host_dev->dma_mask)
1526                 bounce_limit = *host_dev->dma_mask;
1527
1528         return bounce_limit;
1529 }
1530 EXPORT_SYMBOL(scsi_calculate_bounce_limit);
1531
1532 struct request_queue *__scsi_alloc_queue(struct Scsi_Host *shost,
1533                                          request_fn_proc *request_fn)
1534 {
1535         struct request_queue *q;
1536
1537         q = blk_init_queue(request_fn, NULL);
1538         if (!q)
1539                 return NULL;
1540
1541         /*
1542          * this limit is imposed by hardware restrictions
1543          */
1544         blk_queue_max_hw_segments(q, shost->sg_tablesize);
1545
1546         /*
1547          * In the future, sg chaining support will be mandatory and this
1548          * ifdef can then go away. Right now we don't have all archs
1549          * converted, so better keep it safe.
1550          */
1551 #ifdef ARCH_HAS_SG_CHAIN
1552         if (shost->use_sg_chaining)
1553                 blk_queue_max_phys_segments(q, SCSI_MAX_SG_CHAIN_SEGMENTS);
1554         else
1555                 blk_queue_max_phys_segments(q, SCSI_MAX_SG_SEGMENTS);
1556 #else
1557         blk_queue_max_phys_segments(q, SCSI_MAX_SG_SEGMENTS);
1558 #endif
1559
1560         blk_queue_max_sectors(q, shost->max_sectors);
1561         blk_queue_bounce_limit(q, scsi_calculate_bounce_limit(shost));
1562         blk_queue_segment_boundary(q, shost->dma_boundary);
1563
1564         if (!shost->use_clustering)
1565                 clear_bit(QUEUE_FLAG_CLUSTER, &q->queue_flags);
1566
1567         /*
1568          * set a reasonable default alignment on word boundaries: the
1569          * host and device may alter it using
1570          * blk_queue_update_dma_alignment() later.
1571          */
1572         blk_queue_dma_alignment(q, 0x03);
1573
1574         return q;
1575 }
1576 EXPORT_SYMBOL(__scsi_alloc_queue);
1577
1578 struct request_queue *scsi_alloc_queue(struct scsi_device *sdev)
1579 {
1580         struct request_queue *q;
1581
1582         q = __scsi_alloc_queue(sdev->host, scsi_request_fn);
1583         if (!q)
1584                 return NULL;
1585
1586         blk_queue_prep_rq(q, scsi_prep_fn);
1587         blk_queue_softirq_done(q, scsi_softirq_done);
1588         return q;
1589 }
1590
1591 void scsi_free_queue(struct request_queue *q)
1592 {
1593         blk_cleanup_queue(q);
1594 }
1595
1596 /*
1597  * Function:    scsi_block_requests()
1598  *
1599  * Purpose:     Utility function used by low-level drivers to prevent further
1600  *              commands from being queued to the device.
1601  *
1602  * Arguments:   shost       - Host in question
1603  *
1604  * Returns:     Nothing
1605  *
1606  * Lock status: No locks are assumed held.
1607  *
1608  * Notes:       There is no timer nor any other means by which the requests
1609  *              get unblocked other than the low-level driver calling
1610  *              scsi_unblock_requests().
1611  */
1612 void scsi_block_requests(struct Scsi_Host *shost)
1613 {
1614         shost->host_self_blocked = 1;
1615 }
1616 EXPORT_SYMBOL(scsi_block_requests);
1617
1618 /*
1619  * Function:    scsi_unblock_requests()
1620  *
1621  * Purpose:     Utility function used by low-level drivers to allow further
1622  *              commands from being queued to the device.
1623  *
1624  * Arguments:   shost       - Host in question
1625  *
1626  * Returns:     Nothing
1627  *
1628  * Lock status: No locks are assumed held.
1629  *
1630  * Notes:       There is no timer nor any other means by which the requests
1631  *              get unblocked other than the low-level driver calling
1632  *              scsi_unblock_requests().
1633  *
1634  *              This is done as an API function so that changes to the
1635  *              internals of the scsi mid-layer won't require wholesale
1636  *              changes to drivers that use this feature.
1637  */
1638 void scsi_unblock_requests(struct Scsi_Host *shost)
1639 {
1640         shost->host_self_blocked = 0;
1641         scsi_run_host_queues(shost);
1642 }
1643 EXPORT_SYMBOL(scsi_unblock_requests);
1644
1645 int __init scsi_init_queue(void)
1646 {
1647         int i;
1648
1649         scsi_io_context_cache = kmem_cache_create("scsi_io_context",
1650                                         sizeof(struct scsi_io_context),
1651                                         0, 0, NULL);
1652         if (!scsi_io_context_cache) {
1653                 printk(KERN_ERR "SCSI: can't init scsi io context cache\n");
1654                 return -ENOMEM;
1655         }
1656
1657         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1658                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1659                 int size = sgp->size * sizeof(struct scatterlist);
1660
1661                 sgp->slab = kmem_cache_create(sgp->name, size, 0,
1662                                 SLAB_HWCACHE_ALIGN, NULL);
1663                 if (!sgp->slab) {
1664                         printk(KERN_ERR "SCSI: can't init sg slab %s\n",
1665                                         sgp->name);
1666                 }
1667
1668                 sgp->pool = mempool_create_slab_pool(SG_MEMPOOL_SIZE,
1669                                                      sgp->slab);
1670                 if (!sgp->pool) {
1671                         printk(KERN_ERR "SCSI: can't init sg mempool %s\n",
1672                                         sgp->name);
1673                 }
1674         }
1675
1676         return 0;
1677 }
1678
1679 void scsi_exit_queue(void)
1680 {
1681         int i;
1682
1683         kmem_cache_destroy(scsi_io_context_cache);
1684
1685         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1686                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1687                 mempool_destroy(sgp->pool);
1688                 kmem_cache_destroy(sgp->slab);
1689         }
1690 }
1691
1692 /**
1693  *      scsi_mode_select - issue a mode select
1694  *      @sdev:  SCSI device to be queried
1695  *      @pf:    Page format bit (1 == standard, 0 == vendor specific)
1696  *      @sp:    Save page bit (0 == don't save, 1 == save)
1697  *      @modepage: mode page being requested
1698  *      @buffer: request buffer (may not be smaller than eight bytes)
1699  *      @len:   length of request buffer.
1700  *      @timeout: command timeout
1701  *      @retries: number of retries before failing
1702  *      @data: returns a structure abstracting the mode header data
1703  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
1704  *              must be SCSI_SENSE_BUFFERSIZE big.
1705  *
1706  *      Returns zero if successful; negative error number or scsi
1707  *      status on error
1708  *
1709  */
1710 int
1711 scsi_mode_select(struct scsi_device *sdev, int pf, int sp, int modepage,
1712                  unsigned char *buffer, int len, int timeout, int retries,
1713                  struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1714 {
1715         unsigned char cmd[10];
1716         unsigned char *real_buffer;
1717         int ret;
1718
1719         memset(cmd, 0, sizeof(cmd));
1720         cmd[1] = (pf ? 0x10 : 0) | (sp ? 0x01 : 0);
1721
1722         if (sdev->use_10_for_ms) {
1723                 if (len > 65535)
1724                         return -EINVAL;
1725                 real_buffer = kmalloc(8 + len, GFP_KERNEL);
1726                 if (!real_buffer)
1727                         return -ENOMEM;
1728                 memcpy(real_buffer + 8, buffer, len);
1729                 len += 8;
1730                 real_buffer[0] = 0;
1731                 real_buffer[1] = 0;
1732                 real_buffer[2] = data->medium_type;
1733                 real_buffer[3] = data->device_specific;
1734                 real_buffer[4] = data->longlba ? 0x01 : 0;
1735                 real_buffer[5] = 0;
1736                 real_buffer[6] = data->block_descriptor_length >> 8;
1737                 real_buffer[7] = data->block_descriptor_length;
1738
1739                 cmd[0] = MODE_SELECT_10;
1740                 cmd[7] = len >> 8;
1741                 cmd[8] = len;
1742         } else {
1743                 if (len > 255 || data->block_descriptor_length > 255 ||
1744                     data->longlba)
1745                         return -EINVAL;
1746
1747                 real_buffer = kmalloc(4 + len, GFP_KERNEL);
1748                 if (!real_buffer)
1749                         return -ENOMEM;
1750                 memcpy(real_buffer + 4, buffer, len);
1751                 len += 4;
1752                 real_buffer[0] = 0;
1753                 real_buffer[1] = data->medium_type;
1754                 real_buffer[2] = data->device_specific;
1755                 real_buffer[3] = data->block_descriptor_length;
1756                 
1757
1758                 cmd[0] = MODE_SELECT;
1759                 cmd[4] = len;
1760         }
1761
1762         ret = scsi_execute_req(sdev, cmd, DMA_TO_DEVICE, real_buffer, len,
1763                                sshdr, timeout, retries);
1764         kfree(real_buffer);
1765         return ret;
1766 }
1767 EXPORT_SYMBOL_GPL(scsi_mode_select);
1768
1769 /**
1770  *      scsi_mode_sense - issue a mode sense, falling back from 10 to six bytes if necessary.
1771  *      @sdev:  SCSI device to be queried
1772  *      @dbd:   set if mode sense will allow block descriptors to be returned
1773  *      @modepage: mode page being requested
1774  *      @buffer: request buffer (may not be smaller than eight bytes)
1775  *      @len:   length of request buffer.
1776  *      @timeout: command timeout
1777  *      @retries: number of retries before failing
1778  *      @data: returns a structure abstracting the mode header data
1779  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
1780  *              must be SCSI_SENSE_BUFFERSIZE big.
1781  *
1782  *      Returns zero if unsuccessful, or the header offset (either 4
1783  *      or 8 depending on whether a six or ten byte command was
1784  *      issued) if successful.
1785  */
1786 int
1787 scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
1788                   unsigned char *buffer, int len, int timeout, int retries,
1789                   struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1790 {
1791         unsigned char cmd[12];
1792         int use_10_for_ms;
1793         int header_length;
1794         int result;
1795         struct scsi_sense_hdr my_sshdr;
1796
1797         memset(data, 0, sizeof(*data));
1798         memset(&cmd[0], 0, 12);
1799         cmd[1] = dbd & 0x18;    /* allows DBD and LLBA bits */
1800         cmd[2] = modepage;
1801
1802         /* caller might not be interested in sense, but we need it */
1803         if (!sshdr)
1804                 sshdr = &my_sshdr;
1805
1806  retry:
1807         use_10_for_ms = sdev->use_10_for_ms;
1808
1809         if (use_10_for_ms) {
1810                 if (len < 8)
1811                         len = 8;
1812
1813                 cmd[0] = MODE_SENSE_10;
1814                 cmd[8] = len;
1815                 header_length = 8;
1816         } else {
1817                 if (len < 4)
1818                         len = 4;
1819
1820                 cmd[0] = MODE_SENSE;
1821                 cmd[4] = len;
1822                 header_length = 4;
1823         }
1824
1825         memset(buffer, 0, len);
1826
1827         result = scsi_execute_req(sdev, cmd, DMA_FROM_DEVICE, buffer, len,
1828                                   sshdr, timeout, retries);
1829
1830         /* This code looks awful: what it's doing is making sure an
1831          * ILLEGAL REQUEST sense return identifies the actual command
1832          * byte as the problem.  MODE_SENSE commands can return
1833          * ILLEGAL REQUEST if the code page isn't supported */
1834
1835         if (use_10_for_ms && !scsi_status_is_good(result) &&
1836             (driver_byte(result) & DRIVER_SENSE)) {
1837                 if (scsi_sense_valid(sshdr)) {
1838                         if ((sshdr->sense_key == ILLEGAL_REQUEST) &&
1839                             (sshdr->asc == 0x20) && (sshdr->ascq == 0)) {
1840                                 /* 
1841                                  * Invalid command operation code
1842                                  */
1843                                 sdev->use_10_for_ms = 0;
1844                                 goto retry;
1845                         }
1846                 }
1847         }
1848
1849         if(scsi_status_is_good(result)) {
1850                 if (unlikely(buffer[0] == 0x86 && buffer[1] == 0x0b &&
1851                              (modepage == 6 || modepage == 8))) {
1852                         /* Initio breakage? */
1853                         header_length = 0;
1854                         data->length = 13;
1855                         data->medium_type = 0;
1856                         data->device_specific = 0;
1857                         data->longlba = 0;
1858                         data->block_descriptor_length = 0;
1859                 } else if(use_10_for_ms) {
1860                         data->length = buffer[0]*256 + buffer[1] + 2;
1861                         data->medium_type = buffer[2];
1862                         data->device_specific = buffer[3];
1863                         data->longlba = buffer[4] & 0x01;
1864                         data->block_descriptor_length = buffer[6]*256
1865                                 + buffer[7];
1866                 } else {
1867                         data->length = buffer[0] + 1;
1868                         data->medium_type = buffer[1];
1869                         data->device_specific = buffer[2];
1870                         data->block_descriptor_length = buffer[3];
1871                 }
1872                 data->header_length = header_length;
1873         }
1874
1875         return result;
1876 }
1877 EXPORT_SYMBOL(scsi_mode_sense);
1878
1879 /**
1880  *      scsi_test_unit_ready - test if unit is ready
1881  *      @sdev:  scsi device to change the state of.
1882  *      @timeout: command timeout
1883  *      @retries: number of retries before failing
1884  *      @sshdr_external: Optional pointer to struct scsi_sense_hdr for
1885  *              returning sense. Make sure that this is cleared before passing
1886  *              in.
1887  *
1888  *      Returns zero if unsuccessful or an error if TUR failed.  For
1889  *      removable media, a return of NOT_READY or UNIT_ATTENTION is
1890  *      translated to success, with the ->changed flag updated.
1891  **/
1892 int
1893 scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries,
1894                      struct scsi_sense_hdr *sshdr_external)
1895 {
1896         char cmd[] = {
1897                 TEST_UNIT_READY, 0, 0, 0, 0, 0,
1898         };
1899         struct scsi_sense_hdr *sshdr;
1900         int result;
1901
1902         if (!sshdr_external)
1903                 sshdr = kzalloc(sizeof(*sshdr), GFP_KERNEL);
1904         else
1905                 sshdr = sshdr_external;
1906
1907         /* try to eat the UNIT_ATTENTION if there are enough retries */
1908         do {
1909                 result = scsi_execute_req(sdev, cmd, DMA_NONE, NULL, 0, sshdr,
1910                                           timeout, retries);
1911         } while ((driver_byte(result) & DRIVER_SENSE) &&
1912                  sshdr && sshdr->sense_key == UNIT_ATTENTION &&
1913                  --retries);
1914
1915         if (!sshdr)
1916                 /* could not allocate sense buffer, so can't process it */
1917                 return result;
1918
1919         if ((driver_byte(result) & DRIVER_SENSE) && sdev->removable) {
1920
1921                 if ((scsi_sense_valid(sshdr)) &&
1922                     ((sshdr->sense_key == UNIT_ATTENTION) ||
1923                      (sshdr->sense_key == NOT_READY))) {
1924                         sdev->changed = 1;
1925                         result = 0;
1926                 }
1927         }
1928         if (!sshdr_external)
1929                 kfree(sshdr);
1930         return result;
1931 }
1932 EXPORT_SYMBOL(scsi_test_unit_ready);
1933
1934 /**
1935  *      scsi_device_set_state - Take the given device through the device state model.
1936  *      @sdev:  scsi device to change the state of.
1937  *      @state: state to change to.
1938  *
1939  *      Returns zero if unsuccessful or an error if the requested 
1940  *      transition is illegal.
1941  */
1942 int
1943 scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
1944 {
1945         enum scsi_device_state oldstate = sdev->sdev_state;
1946
1947         if (state == oldstate)
1948                 return 0;
1949
1950         switch (state) {
1951         case SDEV_CREATED:
1952                 /* There are no legal states that come back to
1953                  * created.  This is the manually initialised start
1954                  * state */
1955                 goto illegal;
1956                         
1957         case SDEV_RUNNING:
1958                 switch (oldstate) {
1959                 case SDEV_CREATED:
1960                 case SDEV_OFFLINE:
1961                 case SDEV_QUIESCE:
1962                 case SDEV_BLOCK:
1963                         break;
1964                 default:
1965                         goto illegal;
1966                 }
1967                 break;
1968
1969         case SDEV_QUIESCE:
1970                 switch (oldstate) {
1971                 case SDEV_RUNNING:
1972                 case SDEV_OFFLINE:
1973                         break;
1974                 default:
1975                         goto illegal;
1976                 }
1977                 break;
1978
1979         case SDEV_OFFLINE:
1980                 switch (oldstate) {
1981                 case SDEV_CREATED:
1982                 case SDEV_RUNNING:
1983                 case SDEV_QUIESCE:
1984                 case SDEV_BLOCK:
1985                         break;
1986                 default:
1987                         goto illegal;
1988                 }
1989                 break;
1990
1991         case SDEV_BLOCK:
1992                 switch (oldstate) {
1993                 case SDEV_CREATED:
1994                 case SDEV_RUNNING:
1995                         break;
1996                 default:
1997                         goto illegal;
1998                 }
1999                 break;
2000
2001         case SDEV_CANCEL:
2002                 switch (oldstate) {
2003                 case SDEV_CREATED:
2004                 case SDEV_RUNNING:
2005                 case SDEV_QUIESCE:
2006                 case SDEV_OFFLINE:
2007                 case SDEV_BLOCK:
2008                         break;
2009                 default:
2010                         goto illegal;
2011                 }
2012                 break;
2013
2014         case SDEV_DEL:
2015                 switch (oldstate) {
2016                 case SDEV_CREATED:
2017                 case SDEV_RUNNING:
2018                 case SDEV_OFFLINE:
2019                 case SDEV_CANCEL:
2020                         break;
2021                 default:
2022                         goto illegal;
2023                 }
2024                 break;
2025
2026         }
2027         sdev->sdev_state = state;
2028         return 0;
2029
2030  illegal:
2031         SCSI_LOG_ERROR_RECOVERY(1, 
2032                                 sdev_printk(KERN_ERR, sdev,
2033                                             "Illegal state transition %s->%s\n",
2034                                             scsi_device_state_name(oldstate),
2035                                             scsi_device_state_name(state))
2036                                 );
2037         return -EINVAL;
2038 }
2039 EXPORT_SYMBOL(scsi_device_set_state);
2040
2041 /**
2042  *      sdev_evt_emit - emit a single SCSI device uevent
2043  *      @sdev: associated SCSI device
2044  *      @evt: event to emit
2045  *
2046  *      Send a single uevent (scsi_event) to the associated scsi_device.
2047  */
2048 static void scsi_evt_emit(struct scsi_device *sdev, struct scsi_event *evt)
2049 {
2050         int idx = 0;
2051         char *envp[3];
2052
2053         switch (evt->evt_type) {
2054         case SDEV_EVT_MEDIA_CHANGE:
2055                 envp[idx++] = "SDEV_MEDIA_CHANGE=1";
2056                 break;
2057
2058         default:
2059                 /* do nothing */
2060                 break;
2061         }
2062
2063         envp[idx++] = NULL;
2064
2065         kobject_uevent_env(&sdev->sdev_gendev.kobj, KOBJ_CHANGE, envp);
2066 }
2067
2068 /**
2069  *      sdev_evt_thread - send a uevent for each scsi event
2070  *      @work: work struct for scsi_device
2071  *
2072  *      Dispatch queued events to their associated scsi_device kobjects
2073  *      as uevents.
2074  */
2075 void scsi_evt_thread(struct work_struct *work)
2076 {
2077         struct scsi_device *sdev;
2078         LIST_HEAD(event_list);
2079
2080         sdev = container_of(work, struct scsi_device, event_work);
2081
2082         while (1) {
2083                 struct scsi_event *evt;
2084                 struct list_head *this, *tmp;
2085                 unsigned long flags;
2086
2087                 spin_lock_irqsave(&sdev->list_lock, flags);
2088                 list_splice_init(&sdev->event_list, &event_list);
2089                 spin_unlock_irqrestore(&sdev->list_lock, flags);
2090
2091                 if (list_empty(&event_list))
2092                         break;
2093
2094                 list_for_each_safe(this, tmp, &event_list) {
2095                         evt = list_entry(this, struct scsi_event, node);
2096                         list_del(&evt->node);
2097                         scsi_evt_emit(sdev, evt);
2098                         kfree(evt);
2099                 }
2100         }
2101 }
2102
2103 /**
2104  *      sdev_evt_send - send asserted event to uevent thread
2105  *      @sdev: scsi_device event occurred on
2106  *      @evt: event to send
2107  *
2108  *      Assert scsi device event asynchronously.
2109  */
2110 void sdev_evt_send(struct scsi_device *sdev, struct scsi_event *evt)
2111 {
2112         unsigned long flags;
2113
2114         if (!test_bit(evt->evt_type, sdev->supported_events)) {
2115                 kfree(evt);
2116                 return;
2117         }
2118
2119         spin_lock_irqsave(&sdev->list_lock, flags);
2120         list_add_tail(&evt->node, &sdev->event_list);
2121         schedule_work(&sdev->event_work);
2122         spin_unlock_irqrestore(&sdev->list_lock, flags);
2123 }
2124 EXPORT_SYMBOL_GPL(sdev_evt_send);
2125
2126 /**
2127  *      sdev_evt_alloc - allocate a new scsi event
2128  *      @evt_type: type of event to allocate
2129  *      @gfpflags: GFP flags for allocation
2130  *
2131  *      Allocates and returns a new scsi_event.
2132  */
2133 struct scsi_event *sdev_evt_alloc(enum scsi_device_event evt_type,
2134                                   gfp_t gfpflags)
2135 {
2136         struct scsi_event *evt = kzalloc(sizeof(struct scsi_event), gfpflags);
2137         if (!evt)
2138                 return NULL;
2139
2140         evt->evt_type = evt_type;
2141         INIT_LIST_HEAD(&evt->node);
2142
2143         /* evt_type-specific initialization, if any */
2144         switch (evt_type) {
2145         case SDEV_EVT_MEDIA_CHANGE:
2146         default:
2147                 /* do nothing */
2148                 break;
2149         }
2150
2151         return evt;
2152 }
2153 EXPORT_SYMBOL_GPL(sdev_evt_alloc);
2154
2155 /**
2156  *      sdev_evt_send_simple - send asserted event to uevent thread
2157  *      @sdev: scsi_device event occurred on
2158  *      @evt_type: type of event to send
2159  *      @gfpflags: GFP flags for allocation
2160  *
2161  *      Assert scsi device event asynchronously, given an event type.
2162  */
2163 void sdev_evt_send_simple(struct scsi_device *sdev,
2164                           enum scsi_device_event evt_type, gfp_t gfpflags)
2165 {
2166         struct scsi_event *evt = sdev_evt_alloc(evt_type, gfpflags);
2167         if (!evt) {
2168                 sdev_printk(KERN_ERR, sdev, "event %d eaten due to OOM\n",
2169                             evt_type);
2170                 return;
2171         }
2172
2173         sdev_evt_send(sdev, evt);
2174 }
2175 EXPORT_SYMBOL_GPL(sdev_evt_send_simple);
2176
2177 /**
2178  *      scsi_device_quiesce - Block user issued commands.
2179  *      @sdev:  scsi device to quiesce.
2180  *
2181  *      This works by trying to transition to the SDEV_QUIESCE state
2182  *      (which must be a legal transition).  When the device is in this
2183  *      state, only special requests will be accepted, all others will
2184  *      be deferred.  Since special requests may also be requeued requests,
2185  *      a successful return doesn't guarantee the device will be 
2186  *      totally quiescent.
2187  *
2188  *      Must be called with user context, may sleep.
2189  *
2190  *      Returns zero if unsuccessful or an error if not.
2191  */
2192 int
2193 scsi_device_quiesce(struct scsi_device *sdev)
2194 {
2195         int err = scsi_device_set_state(sdev, SDEV_QUIESCE);
2196         if (err)
2197                 return err;
2198
2199         scsi_run_queue(sdev->request_queue);
2200         while (sdev->device_busy) {
2201                 msleep_interruptible(200);
2202                 scsi_run_queue(sdev->request_queue);
2203         }
2204         return 0;
2205 }
2206 EXPORT_SYMBOL(scsi_device_quiesce);
2207
2208 /**
2209  *      scsi_device_resume - Restart user issued commands to a quiesced device.
2210  *      @sdev:  scsi device to resume.
2211  *
2212  *      Moves the device from quiesced back to running and restarts the
2213  *      queues.
2214  *
2215  *      Must be called with user context, may sleep.
2216  */
2217 void
2218 scsi_device_resume(struct scsi_device *sdev)
2219 {
2220         if(scsi_device_set_state(sdev, SDEV_RUNNING))
2221                 return;
2222         scsi_run_queue(sdev->request_queue);
2223 }
2224 EXPORT_SYMBOL(scsi_device_resume);
2225
2226 static void
2227 device_quiesce_fn(struct scsi_device *sdev, void *data)
2228 {
2229         scsi_device_quiesce(sdev);
2230 }
2231
2232 void
2233 scsi_target_quiesce(struct scsi_target *starget)
2234 {
2235         starget_for_each_device(starget, NULL, device_quiesce_fn);
2236 }
2237 EXPORT_SYMBOL(scsi_target_quiesce);
2238
2239 static void
2240 device_resume_fn(struct scsi_device *sdev, void *data)
2241 {
2242         scsi_device_resume(sdev);
2243 }
2244
2245 void
2246 scsi_target_resume(struct scsi_target *starget)
2247 {
2248         starget_for_each_device(starget, NULL, device_resume_fn);
2249 }
2250 EXPORT_SYMBOL(scsi_target_resume);
2251
2252 /**
2253  * scsi_internal_device_block - internal function to put a device temporarily into the SDEV_BLOCK state
2254  * @sdev:       device to block
2255  *
2256  * Block request made by scsi lld's to temporarily stop all
2257  * scsi commands on the specified device.  Called from interrupt
2258  * or normal process context.
2259  *
2260  * Returns zero if successful or error if not
2261  *
2262  * Notes:       
2263  *      This routine transitions the device to the SDEV_BLOCK state
2264  *      (which must be a legal transition).  When the device is in this
2265  *      state, all commands are deferred until the scsi lld reenables
2266  *      the device with scsi_device_unblock or device_block_tmo fires.
2267  *      This routine assumes the host_lock is held on entry.
2268  */
2269 int
2270 scsi_internal_device_block(struct scsi_device *sdev)
2271 {
2272         struct request_queue *q = sdev->request_queue;
2273         unsigned long flags;
2274         int err = 0;
2275
2276         err = scsi_device_set_state(sdev, SDEV_BLOCK);
2277         if (err)
2278                 return err;
2279
2280         /* 
2281          * The device has transitioned to SDEV_BLOCK.  Stop the
2282          * block layer from calling the midlayer with this device's
2283          * request queue. 
2284          */
2285         spin_lock_irqsave(q->queue_lock, flags);
2286         blk_stop_queue(q);
2287         spin_unlock_irqrestore(q->queue_lock, flags);
2288
2289         return 0;
2290 }
2291 EXPORT_SYMBOL_GPL(scsi_internal_device_block);
2292  
2293 /**
2294  * scsi_internal_device_unblock - resume a device after a block request
2295  * @sdev:       device to resume
2296  *
2297  * Called by scsi lld's or the midlayer to restart the device queue
2298  * for the previously suspended scsi device.  Called from interrupt or
2299  * normal process context.
2300  *
2301  * Returns zero if successful or error if not.
2302  *
2303  * Notes:       
2304  *      This routine transitions the device to the SDEV_RUNNING state
2305  *      (which must be a legal transition) allowing the midlayer to
2306  *      goose the queue for this device.  This routine assumes the 
2307  *      host_lock is held upon entry.
2308  */
2309 int
2310 scsi_internal_device_unblock(struct scsi_device *sdev)
2311 {
2312         struct request_queue *q = sdev->request_queue; 
2313         int err;
2314         unsigned long flags;
2315         
2316         /* 
2317          * Try to transition the scsi device to SDEV_RUNNING
2318          * and goose the device queue if successful.  
2319          */
2320         err = scsi_device_set_state(sdev, SDEV_RUNNING);
2321         if (err)
2322                 return err;
2323
2324         spin_lock_irqsave(q->queue_lock, flags);
2325         blk_start_queue(q);
2326         spin_unlock_irqrestore(q->queue_lock, flags);
2327
2328         return 0;
2329 }
2330 EXPORT_SYMBOL_GPL(scsi_internal_device_unblock);
2331
2332 static void
2333 device_block(struct scsi_device *sdev, void *data)
2334 {
2335         scsi_internal_device_block(sdev);
2336 }
2337
2338 static int
2339 target_block(struct device *dev, void *data)
2340 {
2341         if (scsi_is_target_device(dev))
2342                 starget_for_each_device(to_scsi_target(dev), NULL,
2343                                         device_block);
2344         return 0;
2345 }
2346
2347 void
2348 scsi_target_block(struct device *dev)
2349 {
2350         if (scsi_is_target_device(dev))
2351                 starget_for_each_device(to_scsi_target(dev), NULL,
2352                                         device_block);
2353         else
2354                 device_for_each_child(dev, NULL, target_block);
2355 }
2356 EXPORT_SYMBOL_GPL(scsi_target_block);
2357
2358 static void
2359 device_unblock(struct scsi_device *sdev, void *data)
2360 {
2361         scsi_internal_device_unblock(sdev);
2362 }
2363
2364 static int
2365 target_unblock(struct device *dev, void *data)
2366 {
2367         if (scsi_is_target_device(dev))
2368                 starget_for_each_device(to_scsi_target(dev), NULL,
2369                                         device_unblock);
2370         return 0;
2371 }
2372
2373 void
2374 scsi_target_unblock(struct device *dev)
2375 {
2376         if (scsi_is_target_device(dev))
2377                 starget_for_each_device(to_scsi_target(dev), NULL,
2378                                         device_unblock);
2379         else
2380                 device_for_each_child(dev, NULL, target_unblock);
2381 }
2382 EXPORT_SYMBOL_GPL(scsi_target_unblock);
2383
2384 /**
2385  * scsi_kmap_atomic_sg - find and atomically map an sg-elemnt
2386  * @sgl:        scatter-gather list
2387  * @sg_count:   number of segments in sg
2388  * @offset:     offset in bytes into sg, on return offset into the mapped area
2389  * @len:        bytes to map, on return number of bytes mapped
2390  *
2391  * Returns virtual address of the start of the mapped page
2392  */
2393 void *scsi_kmap_atomic_sg(struct scatterlist *sgl, int sg_count,
2394                           size_t *offset, size_t *len)
2395 {
2396         int i;
2397         size_t sg_len = 0, len_complete = 0;
2398         struct scatterlist *sg;
2399         struct page *page;
2400
2401         WARN_ON(!irqs_disabled());
2402
2403         for_each_sg(sgl, sg, sg_count, i) {
2404                 len_complete = sg_len; /* Complete sg-entries */
2405                 sg_len += sg->length;
2406                 if (sg_len > *offset)
2407                         break;
2408         }
2409
2410         if (unlikely(i == sg_count)) {
2411                 printk(KERN_ERR "%s: Bytes in sg: %zu, requested offset %zu, "
2412                         "elements %d\n",
2413                        __FUNCTION__, sg_len, *offset, sg_count);
2414                 WARN_ON(1);
2415                 return NULL;
2416         }
2417
2418         /* Offset starting from the beginning of first page in this sg-entry */
2419         *offset = *offset - len_complete + sg->offset;
2420
2421         /* Assumption: contiguous pages can be accessed as "page + i" */
2422         page = nth_page(sg_page(sg), (*offset >> PAGE_SHIFT));
2423         *offset &= ~PAGE_MASK;
2424
2425         /* Bytes in this sg-entry from *offset to the end of the page */
2426         sg_len = PAGE_SIZE - *offset;
2427         if (*len > sg_len)
2428                 *len = sg_len;
2429
2430         return kmap_atomic(page, KM_BIO_SRC_IRQ);
2431 }
2432 EXPORT_SYMBOL(scsi_kmap_atomic_sg);
2433
2434 /**
2435  * scsi_kunmap_atomic_sg - atomically unmap a virtual address, previously mapped with scsi_kmap_atomic_sg
2436  * @virt:       virtual address to be unmapped
2437  */
2438 void scsi_kunmap_atomic_sg(void *virt)
2439 {
2440         kunmap_atomic(virt, KM_BIO_SRC_IRQ);
2441 }
2442 EXPORT_SYMBOL(scsi_kunmap_atomic_sg);