69fd62183bab4e97bd5cfce98ed85aa4f014ef59
[linux-2.6.git] / drivers / scsi / scsi_lib.c
1 /*
2  *  scsi_lib.c Copyright (C) 1999 Eric Youngdale
3  *
4  *  SCSI queueing library.
5  *      Initial versions: Eric Youngdale (eric@andante.org).
6  *                        Based upon conversations with large numbers
7  *                        of people at Linux Expo.
8  */
9
10 #include <linux/bio.h>
11 #include <linux/blkdev.h>
12 #include <linux/completion.h>
13 #include <linux/kernel.h>
14 #include <linux/mempool.h>
15 #include <linux/slab.h>
16 #include <linux/init.h>
17 #include <linux/pci.h>
18 #include <linux/delay.h>
19 #include <linux/hardirq.h>
20 #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 static 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 static void scsi_free_sgtable(struct scsi_cmnd *cmd)
766 {
767         __sg_free_table(&cmd->sg_table, SCSI_MAX_SG_SEGMENTS, scsi_sg_free);
768 }
769
770 /*
771  * Function:    scsi_release_buffers()
772  *
773  * Purpose:     Completion processing for block device I/O requests.
774  *
775  * Arguments:   cmd     - command that we are bailing.
776  *
777  * Lock status: Assumed that no lock is held upon entry.
778  *
779  * Returns:     Nothing
780  *
781  * Notes:       In the event that an upper level driver rejects a
782  *              command, we must release resources allocated during
783  *              the __init_io() function.  Primarily this would involve
784  *              the scatter-gather table, and potentially any bounce
785  *              buffers.
786  */
787 void scsi_release_buffers(struct scsi_cmnd *cmd)
788 {
789         if (cmd->use_sg)
790                 scsi_free_sgtable(cmd);
791
792         /*
793          * Zero these out.  They now point to freed memory, and it is
794          * dangerous to hang onto the pointers.
795          */
796         cmd->request_buffer = NULL;
797         cmd->request_bufflen = 0;
798 }
799 EXPORT_SYMBOL(scsi_release_buffers);
800
801 /*
802  * Function:    scsi_io_completion()
803  *
804  * Purpose:     Completion processing for block device I/O requests.
805  *
806  * Arguments:   cmd   - command that is finished.
807  *
808  * Lock status: Assumed that no lock is held upon entry.
809  *
810  * Returns:     Nothing
811  *
812  * Notes:       This function is matched in terms of capabilities to
813  *              the function that created the scatter-gather list.
814  *              In other words, if there are no bounce buffers
815  *              (the normal case for most drivers), we don't need
816  *              the logic to deal with cleaning up afterwards.
817  *
818  *              We must do one of several things here:
819  *
820  *              a) Call scsi_end_request.  This will finish off the
821  *                 specified number of sectors.  If we are done, the
822  *                 command block will be released, and the queue
823  *                 function will be goosed.  If we are not done, then
824  *                 scsi_end_request will directly goose the queue.
825  *
826  *              b) We can just use scsi_requeue_command() here.  This would
827  *                 be used if we just wanted to retry, for example.
828  */
829 void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes)
830 {
831         int result = cmd->result;
832         int this_count = cmd->request_bufflen;
833         struct request_queue *q = cmd->device->request_queue;
834         struct request *req = cmd->request;
835         int clear_errors = 1;
836         struct scsi_sense_hdr sshdr;
837         int sense_valid = 0;
838         int sense_deferred = 0;
839
840         scsi_release_buffers(cmd);
841
842         if (result) {
843                 sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
844                 if (sense_valid)
845                         sense_deferred = scsi_sense_is_deferred(&sshdr);
846         }
847
848         if (blk_pc_request(req)) { /* SG_IO ioctl from block level */
849                 req->errors = result;
850                 if (result) {
851                         clear_errors = 0;
852                         if (sense_valid && req->sense) {
853                                 /*
854                                  * SG_IO wants current and deferred errors
855                                  */
856                                 int len = 8 + cmd->sense_buffer[7];
857
858                                 if (len > SCSI_SENSE_BUFFERSIZE)
859                                         len = SCSI_SENSE_BUFFERSIZE;
860                                 memcpy(req->sense, cmd->sense_buffer,  len);
861                                 req->sense_len = len;
862                         }
863                 }
864                 req->data_len = cmd->resid;
865         }
866
867         /*
868          * Next deal with any sectors which we were able to correctly
869          * handle.
870          */
871         SCSI_LOG_HLCOMPLETE(1, printk("%ld sectors total, "
872                                       "%d bytes done.\n",
873                                       req->nr_sectors, good_bytes));
874         SCSI_LOG_HLCOMPLETE(1, printk("use_sg is %d\n", cmd->use_sg));
875
876         if (clear_errors)
877                 req->errors = 0;
878
879         /* A number of bytes were successfully read.  If there
880          * are leftovers and there is some kind of error
881          * (result != 0), retry the rest.
882          */
883         if (scsi_end_request(cmd, 0, good_bytes, result == 0) == NULL)
884                 return;
885
886         /* good_bytes = 0, or (inclusive) there were leftovers and
887          * result = 0, so scsi_end_request couldn't retry.
888          */
889         if (sense_valid && !sense_deferred) {
890                 switch (sshdr.sense_key) {
891                 case UNIT_ATTENTION:
892                         if (cmd->device->removable) {
893                                 /* Detected disc change.  Set a bit
894                                  * and quietly refuse further access.
895                                  */
896                                 cmd->device->changed = 1;
897                                 scsi_end_request(cmd, -EIO, this_count, 1);
898                                 return;
899                         } else {
900                                 /* Must have been a power glitch, or a
901                                  * bus reset.  Could not have been a
902                                  * media change, so we just retry the
903                                  * request and see what happens.
904                                  */
905                                 scsi_requeue_command(q, cmd);
906                                 return;
907                         }
908                         break;
909                 case ILLEGAL_REQUEST:
910                         /* If we had an ILLEGAL REQUEST returned, then
911                          * we may have performed an unsupported
912                          * command.  The only thing this should be
913                          * would be a ten byte read where only a six
914                          * byte read was supported.  Also, on a system
915                          * where READ CAPACITY failed, we may have
916                          * read past the end of the disk.
917                          */
918                         if ((cmd->device->use_10_for_rw &&
919                             sshdr.asc == 0x20 && sshdr.ascq == 0x00) &&
920                             (cmd->cmnd[0] == READ_10 ||
921                              cmd->cmnd[0] == WRITE_10)) {
922                                 cmd->device->use_10_for_rw = 0;
923                                 /* This will cause a retry with a
924                                  * 6-byte command.
925                                  */
926                                 scsi_requeue_command(q, cmd);
927                                 return;
928                         } else {
929                                 scsi_end_request(cmd, -EIO, this_count, 1);
930                                 return;
931                         }
932                         break;
933                 case NOT_READY:
934                         /* If the device is in the process of becoming
935                          * ready, or has a temporary blockage, retry.
936                          */
937                         if (sshdr.asc == 0x04) {
938                                 switch (sshdr.ascq) {
939                                 case 0x01: /* becoming ready */
940                                 case 0x04: /* format in progress */
941                                 case 0x05: /* rebuild in progress */
942                                 case 0x06: /* recalculation in progress */
943                                 case 0x07: /* operation in progress */
944                                 case 0x08: /* Long write in progress */
945                                 case 0x09: /* self test in progress */
946                                         scsi_requeue_command(q, cmd);
947                                         return;
948                                 default:
949                                         break;
950                                 }
951                         }
952                         if (!(req->cmd_flags & REQ_QUIET))
953                                 scsi_cmd_print_sense_hdr(cmd,
954                                                          "Device not ready",
955                                                          &sshdr);
956
957                         scsi_end_request(cmd, -EIO, this_count, 1);
958                         return;
959                 case VOLUME_OVERFLOW:
960                         if (!(req->cmd_flags & REQ_QUIET)) {
961                                 scmd_printk(KERN_INFO, cmd,
962                                             "Volume overflow, CDB: ");
963                                 __scsi_print_command(cmd->cmnd);
964                                 scsi_print_sense("", cmd);
965                         }
966                         /* See SSC3rXX or current. */
967                         scsi_end_request(cmd, -EIO, this_count, 1);
968                         return;
969                 default:
970                         break;
971                 }
972         }
973         if (host_byte(result) == DID_RESET) {
974                 /* Third party bus reset or reset for error recovery
975                  * reasons.  Just retry the request and see what
976                  * happens.
977                  */
978                 scsi_requeue_command(q, cmd);
979                 return;
980         }
981         if (result) {
982                 if (!(req->cmd_flags & REQ_QUIET)) {
983                         scsi_print_result(cmd);
984                         if (driver_byte(result) & DRIVER_SENSE)
985                                 scsi_print_sense("", cmd);
986                 }
987         }
988         scsi_end_request(cmd, -EIO, this_count, !result);
989 }
990
991 /*
992  * Function:    scsi_init_io()
993  *
994  * Purpose:     SCSI I/O initialize function.
995  *
996  * Arguments:   cmd   - Command descriptor we wish to initialize
997  *
998  * Returns:     0 on success
999  *              BLKPREP_DEFER if the failure is retryable
1000  */
1001 int scsi_init_io(struct scsi_cmnd *cmd, gfp_t gfp_mask)
1002 {
1003         struct request     *req = cmd->request;
1004         int                count;
1005
1006         /*
1007          * We used to not use scatter-gather for single segment request,
1008          * but now we do (it makes highmem I/O easier to support without
1009          * kmapping pages)
1010          */
1011         cmd->use_sg = req->nr_phys_segments;
1012
1013         /*
1014          * If sg table allocation fails, requeue request later.
1015          */
1016         if (unlikely(scsi_alloc_sgtable(cmd, gfp_mask))) {
1017                 scsi_unprep_request(req);
1018                 return BLKPREP_DEFER;
1019         }
1020
1021         req->buffer = NULL;
1022         if (blk_pc_request(req))
1023                 cmd->request_bufflen = req->data_len;
1024         else
1025                 cmd->request_bufflen = req->nr_sectors << 9;
1026
1027         /* 
1028          * Next, walk the list, and fill in the addresses and sizes of
1029          * each segment.
1030          */
1031         count = blk_rq_map_sg(req->q, req, cmd->request_buffer);
1032         BUG_ON(count > cmd->use_sg);
1033         cmd->use_sg = count;
1034         return BLKPREP_OK;
1035 }
1036 EXPORT_SYMBOL(scsi_init_io);
1037
1038 static struct scsi_cmnd *scsi_get_cmd_from_req(struct scsi_device *sdev,
1039                 struct request *req)
1040 {
1041         struct scsi_cmnd *cmd;
1042
1043         if (!req->special) {
1044                 cmd = scsi_get_command(sdev, GFP_ATOMIC);
1045                 if (unlikely(!cmd))
1046                         return NULL;
1047                 req->special = cmd;
1048         } else {
1049                 cmd = req->special;
1050         }
1051
1052         /* pull a tag out of the request if we have one */
1053         cmd->tag = req->tag;
1054         cmd->request = req;
1055
1056         return cmd;
1057 }
1058
1059 int scsi_setup_blk_pc_cmnd(struct scsi_device *sdev, struct request *req)
1060 {
1061         struct scsi_cmnd *cmd;
1062         int ret = scsi_prep_state_check(sdev, req);
1063
1064         if (ret != BLKPREP_OK)
1065                 return ret;
1066
1067         cmd = scsi_get_cmd_from_req(sdev, req);
1068         if (unlikely(!cmd))
1069                 return BLKPREP_DEFER;
1070
1071         /*
1072          * BLOCK_PC requests may transfer data, in which case they must
1073          * a bio attached to them.  Or they might contain a SCSI command
1074          * that does not transfer data, in which case they may optionally
1075          * submit a request without an attached bio.
1076          */
1077         if (req->bio) {
1078                 int ret;
1079
1080                 BUG_ON(!req->nr_phys_segments);
1081
1082                 ret = scsi_init_io(cmd, GFP_ATOMIC);
1083                 if (unlikely(ret))
1084                         return ret;
1085         } else {
1086                 BUG_ON(req->data_len);
1087                 BUG_ON(req->data);
1088
1089                 cmd->request_bufflen = 0;
1090                 cmd->request_buffer = NULL;
1091                 cmd->use_sg = 0;
1092                 req->buffer = NULL;
1093         }
1094
1095         BUILD_BUG_ON(sizeof(req->cmd) > sizeof(cmd->cmnd));
1096         memcpy(cmd->cmnd, req->cmd, sizeof(cmd->cmnd));
1097         cmd->cmd_len = req->cmd_len;
1098         if (!req->data_len)
1099                 cmd->sc_data_direction = DMA_NONE;
1100         else if (rq_data_dir(req) == WRITE)
1101                 cmd->sc_data_direction = DMA_TO_DEVICE;
1102         else
1103                 cmd->sc_data_direction = DMA_FROM_DEVICE;
1104         
1105         cmd->transfersize = req->data_len;
1106         cmd->allowed = req->retries;
1107         cmd->timeout_per_command = req->timeout;
1108         return BLKPREP_OK;
1109 }
1110 EXPORT_SYMBOL(scsi_setup_blk_pc_cmnd);
1111
1112 /*
1113  * Setup a REQ_TYPE_FS command.  These are simple read/write request
1114  * from filesystems that still need to be translated to SCSI CDBs from
1115  * the ULD.
1116  */
1117 int scsi_setup_fs_cmnd(struct scsi_device *sdev, struct request *req)
1118 {
1119         struct scsi_cmnd *cmd;
1120         int ret = scsi_prep_state_check(sdev, req);
1121
1122         if (ret != BLKPREP_OK)
1123                 return ret;
1124         /*
1125          * Filesystem requests must transfer data.
1126          */
1127         BUG_ON(!req->nr_phys_segments);
1128
1129         cmd = scsi_get_cmd_from_req(sdev, req);
1130         if (unlikely(!cmd))
1131                 return BLKPREP_DEFER;
1132
1133         return scsi_init_io(cmd, GFP_ATOMIC);
1134 }
1135 EXPORT_SYMBOL(scsi_setup_fs_cmnd);
1136
1137 int scsi_prep_state_check(struct scsi_device *sdev, struct request *req)
1138 {
1139         int ret = BLKPREP_OK;
1140
1141         /*
1142          * If the device is not in running state we will reject some
1143          * or all commands.
1144          */
1145         if (unlikely(sdev->sdev_state != SDEV_RUNNING)) {
1146                 switch (sdev->sdev_state) {
1147                 case SDEV_OFFLINE:
1148                         /*
1149                          * If the device is offline we refuse to process any
1150                          * commands.  The device must be brought online
1151                          * before trying any recovery commands.
1152                          */
1153                         sdev_printk(KERN_ERR, sdev,
1154                                     "rejecting I/O to offline device\n");
1155                         ret = BLKPREP_KILL;
1156                         break;
1157                 case SDEV_DEL:
1158                         /*
1159                          * If the device is fully deleted, we refuse to
1160                          * process any commands as well.
1161                          */
1162                         sdev_printk(KERN_ERR, sdev,
1163                                     "rejecting I/O to dead device\n");
1164                         ret = BLKPREP_KILL;
1165                         break;
1166                 case SDEV_QUIESCE:
1167                 case SDEV_BLOCK:
1168                         /*
1169                          * If the devices is blocked we defer normal commands.
1170                          */
1171                         if (!(req->cmd_flags & REQ_PREEMPT))
1172                                 ret = BLKPREP_DEFER;
1173                         break;
1174                 default:
1175                         /*
1176                          * For any other not fully online state we only allow
1177                          * special commands.  In particular any user initiated
1178                          * command is not allowed.
1179                          */
1180                         if (!(req->cmd_flags & REQ_PREEMPT))
1181                                 ret = BLKPREP_KILL;
1182                         break;
1183                 }
1184         }
1185         return ret;
1186 }
1187 EXPORT_SYMBOL(scsi_prep_state_check);
1188
1189 int scsi_prep_return(struct request_queue *q, struct request *req, int ret)
1190 {
1191         struct scsi_device *sdev = q->queuedata;
1192
1193         switch (ret) {
1194         case BLKPREP_KILL:
1195                 req->errors = DID_NO_CONNECT << 16;
1196                 /* release the command and kill it */
1197                 if (req->special) {
1198                         struct scsi_cmnd *cmd = req->special;
1199                         scsi_release_buffers(cmd);
1200                         scsi_put_command(cmd);
1201                         req->special = NULL;
1202                 }
1203                 break;
1204         case BLKPREP_DEFER:
1205                 /*
1206                  * If we defer, the elv_next_request() returns NULL, but the
1207                  * queue must be restarted, so we plug here if no returning
1208                  * command will automatically do that.
1209                  */
1210                 if (sdev->device_busy == 0)
1211                         blk_plug_device(q);
1212                 break;
1213         default:
1214                 req->cmd_flags |= REQ_DONTPREP;
1215         }
1216
1217         return ret;
1218 }
1219 EXPORT_SYMBOL(scsi_prep_return);
1220
1221 int scsi_prep_fn(struct request_queue *q, struct request *req)
1222 {
1223         struct scsi_device *sdev = q->queuedata;
1224         int ret = BLKPREP_KILL;
1225
1226         if (req->cmd_type == REQ_TYPE_BLOCK_PC)
1227                 ret = scsi_setup_blk_pc_cmnd(sdev, req);
1228         return scsi_prep_return(q, req, ret);
1229 }
1230
1231 /*
1232  * scsi_dev_queue_ready: if we can send requests to sdev, return 1 else
1233  * return 0.
1234  *
1235  * Called with the queue_lock held.
1236  */
1237 static inline int scsi_dev_queue_ready(struct request_queue *q,
1238                                   struct scsi_device *sdev)
1239 {
1240         if (sdev->device_busy >= sdev->queue_depth)
1241                 return 0;
1242         if (sdev->device_busy == 0 && sdev->device_blocked) {
1243                 /*
1244                  * unblock after device_blocked iterates to zero
1245                  */
1246                 if (--sdev->device_blocked == 0) {
1247                         SCSI_LOG_MLQUEUE(3,
1248                                    sdev_printk(KERN_INFO, sdev,
1249                                    "unblocking device at zero depth\n"));
1250                 } else {
1251                         blk_plug_device(q);
1252                         return 0;
1253                 }
1254         }
1255         if (sdev->device_blocked)
1256                 return 0;
1257
1258         return 1;
1259 }
1260
1261 /*
1262  * scsi_host_queue_ready: if we can send requests to shost, return 1 else
1263  * return 0. We must end up running the queue again whenever 0 is
1264  * returned, else IO can hang.
1265  *
1266  * Called with host_lock held.
1267  */
1268 static inline int scsi_host_queue_ready(struct request_queue *q,
1269                                    struct Scsi_Host *shost,
1270                                    struct scsi_device *sdev)
1271 {
1272         if (scsi_host_in_recovery(shost))
1273                 return 0;
1274         if (shost->host_busy == 0 && shost->host_blocked) {
1275                 /*
1276                  * unblock after host_blocked iterates to zero
1277                  */
1278                 if (--shost->host_blocked == 0) {
1279                         SCSI_LOG_MLQUEUE(3,
1280                                 printk("scsi%d unblocking host at zero depth\n",
1281                                         shost->host_no));
1282                 } else {
1283                         blk_plug_device(q);
1284                         return 0;
1285                 }
1286         }
1287         if ((shost->can_queue > 0 && shost->host_busy >= shost->can_queue) ||
1288             shost->host_blocked || shost->host_self_blocked) {
1289                 if (list_empty(&sdev->starved_entry))
1290                         list_add_tail(&sdev->starved_entry, &shost->starved_list);
1291                 return 0;
1292         }
1293
1294         /* We're OK to process the command, so we can't be starved */
1295         if (!list_empty(&sdev->starved_entry))
1296                 list_del_init(&sdev->starved_entry);
1297
1298         return 1;
1299 }
1300
1301 /*
1302  * Kill a request for a dead device
1303  */
1304 static void scsi_kill_request(struct request *req, struct request_queue *q)
1305 {
1306         struct scsi_cmnd *cmd = req->special;
1307         struct scsi_device *sdev = cmd->device;
1308         struct Scsi_Host *shost = sdev->host;
1309
1310         blkdev_dequeue_request(req);
1311
1312         if (unlikely(cmd == NULL)) {
1313                 printk(KERN_CRIT "impossible request in %s.\n",
1314                                  __FUNCTION__);
1315                 BUG();
1316         }
1317
1318         scsi_init_cmd_errh(cmd);
1319         cmd->result = DID_NO_CONNECT << 16;
1320         atomic_inc(&cmd->device->iorequest_cnt);
1321
1322         /*
1323          * SCSI request completion path will do scsi_device_unbusy(),
1324          * bump busy counts.  To bump the counters, we need to dance
1325          * with the locks as normal issue path does.
1326          */
1327         sdev->device_busy++;
1328         spin_unlock(sdev->request_queue->queue_lock);
1329         spin_lock(shost->host_lock);
1330         shost->host_busy++;
1331         spin_unlock(shost->host_lock);
1332         spin_lock(sdev->request_queue->queue_lock);
1333
1334         __scsi_done(cmd);
1335 }
1336
1337 static void scsi_softirq_done(struct request *rq)
1338 {
1339         struct scsi_cmnd *cmd = rq->completion_data;
1340         unsigned long wait_for = (cmd->allowed + 1) * cmd->timeout_per_command;
1341         int disposition;
1342
1343         INIT_LIST_HEAD(&cmd->eh_entry);
1344
1345         disposition = scsi_decide_disposition(cmd);
1346         if (disposition != SUCCESS &&
1347             time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) {
1348                 sdev_printk(KERN_ERR, cmd->device,
1349                             "timing out command, waited %lus\n",
1350                             wait_for/HZ);
1351                 disposition = SUCCESS;
1352         }
1353                         
1354         scsi_log_completion(cmd, disposition);
1355
1356         switch (disposition) {
1357                 case SUCCESS:
1358                         scsi_finish_command(cmd);
1359                         break;
1360                 case NEEDS_RETRY:
1361                         scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY);
1362                         break;
1363                 case ADD_TO_MLQUEUE:
1364                         scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY);
1365                         break;
1366                 default:
1367                         if (!scsi_eh_scmd_add(cmd, 0))
1368                                 scsi_finish_command(cmd);
1369         }
1370 }
1371
1372 /*
1373  * Function:    scsi_request_fn()
1374  *
1375  * Purpose:     Main strategy routine for SCSI.
1376  *
1377  * Arguments:   q       - Pointer to actual queue.
1378  *
1379  * Returns:     Nothing
1380  *
1381  * Lock status: IO request lock assumed to be held when called.
1382  */
1383 static void scsi_request_fn(struct request_queue *q)
1384 {
1385         struct scsi_device *sdev = q->queuedata;
1386         struct Scsi_Host *shost;
1387         struct scsi_cmnd *cmd;
1388         struct request *req;
1389
1390         if (!sdev) {
1391                 printk("scsi: killing requests for dead queue\n");
1392                 while ((req = elv_next_request(q)) != NULL)
1393                         scsi_kill_request(req, q);
1394                 return;
1395         }
1396
1397         if(!get_device(&sdev->sdev_gendev))
1398                 /* We must be tearing the block queue down already */
1399                 return;
1400
1401         /*
1402          * To start with, we keep looping until the queue is empty, or until
1403          * the host is no longer able to accept any more requests.
1404          */
1405         shost = sdev->host;
1406         while (!blk_queue_plugged(q)) {
1407                 int rtn;
1408                 /*
1409                  * get next queueable request.  We do this early to make sure
1410                  * that the request is fully prepared even if we cannot 
1411                  * accept it.
1412                  */
1413                 req = elv_next_request(q);
1414                 if (!req || !scsi_dev_queue_ready(q, sdev))
1415                         break;
1416
1417                 if (unlikely(!scsi_device_online(sdev))) {
1418                         sdev_printk(KERN_ERR, sdev,
1419                                     "rejecting I/O to offline device\n");
1420                         scsi_kill_request(req, q);
1421                         continue;
1422                 }
1423
1424
1425                 /*
1426                  * Remove the request from the request list.
1427                  */
1428                 if (!(blk_queue_tagged(q) && !blk_queue_start_tag(q, req)))
1429                         blkdev_dequeue_request(req);
1430                 sdev->device_busy++;
1431
1432                 spin_unlock(q->queue_lock);
1433                 cmd = req->special;
1434                 if (unlikely(cmd == NULL)) {
1435                         printk(KERN_CRIT "impossible request in %s.\n"
1436                                          "please mail a stack trace to "
1437                                          "linux-scsi@vger.kernel.org\n",
1438                                          __FUNCTION__);
1439                         blk_dump_rq_flags(req, "foo");
1440                         BUG();
1441                 }
1442                 spin_lock(shost->host_lock);
1443
1444                 if (!scsi_host_queue_ready(q, shost, sdev))
1445                         goto not_ready;
1446                 if (scsi_target(sdev)->single_lun) {
1447                         if (scsi_target(sdev)->starget_sdev_user &&
1448                             scsi_target(sdev)->starget_sdev_user != sdev)
1449                                 goto not_ready;
1450                         scsi_target(sdev)->starget_sdev_user = sdev;
1451                 }
1452                 shost->host_busy++;
1453
1454                 /*
1455                  * XXX(hch): This is rather suboptimal, scsi_dispatch_cmd will
1456                  *              take the lock again.
1457                  */
1458                 spin_unlock_irq(shost->host_lock);
1459
1460                 /*
1461                  * Finally, initialize any error handling parameters, and set up
1462                  * the timers for timeouts.
1463                  */
1464                 scsi_init_cmd_errh(cmd);
1465
1466                 /*
1467                  * Dispatch the command to the low-level driver.
1468                  */
1469                 rtn = scsi_dispatch_cmd(cmd);
1470                 spin_lock_irq(q->queue_lock);
1471                 if(rtn) {
1472                         /* we're refusing the command; because of
1473                          * the way locks get dropped, we need to 
1474                          * check here if plugging is required */
1475                         if(sdev->device_busy == 0)
1476                                 blk_plug_device(q);
1477
1478                         break;
1479                 }
1480         }
1481
1482         goto out;
1483
1484  not_ready:
1485         spin_unlock_irq(shost->host_lock);
1486
1487         /*
1488          * lock q, handle tag, requeue req, and decrement device_busy. We
1489          * must return with queue_lock held.
1490          *
1491          * Decrementing device_busy without checking it is OK, as all such
1492          * cases (host limits or settings) should run the queue at some
1493          * later time.
1494          */
1495         spin_lock_irq(q->queue_lock);
1496         blk_requeue_request(q, req);
1497         sdev->device_busy--;
1498         if(sdev->device_busy == 0)
1499                 blk_plug_device(q);
1500  out:
1501         /* must be careful here...if we trigger the ->remove() function
1502          * we cannot be holding the q lock */
1503         spin_unlock_irq(q->queue_lock);
1504         put_device(&sdev->sdev_gendev);
1505         spin_lock_irq(q->queue_lock);
1506 }
1507
1508 u64 scsi_calculate_bounce_limit(struct Scsi_Host *shost)
1509 {
1510         struct device *host_dev;
1511         u64 bounce_limit = 0xffffffff;
1512
1513         if (shost->unchecked_isa_dma)
1514                 return BLK_BOUNCE_ISA;
1515         /*
1516          * Platforms with virtual-DMA translation
1517          * hardware have no practical limit.
1518          */
1519         if (!PCI_DMA_BUS_IS_PHYS)
1520                 return BLK_BOUNCE_ANY;
1521
1522         host_dev = scsi_get_device(shost);
1523         if (host_dev && host_dev->dma_mask)
1524                 bounce_limit = *host_dev->dma_mask;
1525
1526         return bounce_limit;
1527 }
1528 EXPORT_SYMBOL(scsi_calculate_bounce_limit);
1529
1530 struct request_queue *__scsi_alloc_queue(struct Scsi_Host *shost,
1531                                          request_fn_proc *request_fn)
1532 {
1533         struct request_queue *q;
1534
1535         q = blk_init_queue(request_fn, NULL);
1536         if (!q)
1537                 return NULL;
1538
1539         /*
1540          * this limit is imposed by hardware restrictions
1541          */
1542         blk_queue_max_hw_segments(q, shost->sg_tablesize);
1543
1544         /*
1545          * In the future, sg chaining support will be mandatory and this
1546          * ifdef can then go away. Right now we don't have all archs
1547          * converted, so better keep it safe.
1548          */
1549 #ifdef ARCH_HAS_SG_CHAIN
1550         if (shost->use_sg_chaining)
1551                 blk_queue_max_phys_segments(q, SCSI_MAX_SG_CHAIN_SEGMENTS);
1552         else
1553                 blk_queue_max_phys_segments(q, SCSI_MAX_SG_SEGMENTS);
1554 #else
1555         blk_queue_max_phys_segments(q, SCSI_MAX_SG_SEGMENTS);
1556 #endif
1557
1558         blk_queue_max_sectors(q, shost->max_sectors);
1559         blk_queue_bounce_limit(q, scsi_calculate_bounce_limit(shost));
1560         blk_queue_segment_boundary(q, shost->dma_boundary);
1561
1562         if (!shost->use_clustering)
1563                 clear_bit(QUEUE_FLAG_CLUSTER, &q->queue_flags);
1564
1565         /*
1566          * set a reasonable default alignment on word boundaries: the
1567          * host and device may alter it using
1568          * blk_queue_update_dma_alignment() later.
1569          */
1570         blk_queue_dma_alignment(q, 0x03);
1571
1572         return q;
1573 }
1574 EXPORT_SYMBOL(__scsi_alloc_queue);
1575
1576 struct request_queue *scsi_alloc_queue(struct scsi_device *sdev)
1577 {
1578         struct request_queue *q;
1579
1580         q = __scsi_alloc_queue(sdev->host, scsi_request_fn);
1581         if (!q)
1582                 return NULL;
1583
1584         blk_queue_prep_rq(q, scsi_prep_fn);
1585         blk_queue_softirq_done(q, scsi_softirq_done);
1586         return q;
1587 }
1588
1589 void scsi_free_queue(struct request_queue *q)
1590 {
1591         blk_cleanup_queue(q);
1592 }
1593
1594 /*
1595  * Function:    scsi_block_requests()
1596  *
1597  * Purpose:     Utility function used by low-level drivers to prevent further
1598  *              commands from being queued to the device.
1599  *
1600  * Arguments:   shost       - Host in question
1601  *
1602  * Returns:     Nothing
1603  *
1604  * Lock status: No locks are assumed held.
1605  *
1606  * Notes:       There is no timer nor any other means by which the requests
1607  *              get unblocked other than the low-level driver calling
1608  *              scsi_unblock_requests().
1609  */
1610 void scsi_block_requests(struct Scsi_Host *shost)
1611 {
1612         shost->host_self_blocked = 1;
1613 }
1614 EXPORT_SYMBOL(scsi_block_requests);
1615
1616 /*
1617  * Function:    scsi_unblock_requests()
1618  *
1619  * Purpose:     Utility function used by low-level drivers to allow further
1620  *              commands from being queued to the device.
1621  *
1622  * Arguments:   shost       - Host in question
1623  *
1624  * Returns:     Nothing
1625  *
1626  * Lock status: No locks are assumed held.
1627  *
1628  * Notes:       There is no timer nor any other means by which the requests
1629  *              get unblocked other than the low-level driver calling
1630  *              scsi_unblock_requests().
1631  *
1632  *              This is done as an API function so that changes to the
1633  *              internals of the scsi mid-layer won't require wholesale
1634  *              changes to drivers that use this feature.
1635  */
1636 void scsi_unblock_requests(struct Scsi_Host *shost)
1637 {
1638         shost->host_self_blocked = 0;
1639         scsi_run_host_queues(shost);
1640 }
1641 EXPORT_SYMBOL(scsi_unblock_requests);
1642
1643 int __init scsi_init_queue(void)
1644 {
1645         int i;
1646
1647         scsi_io_context_cache = kmem_cache_create("scsi_io_context",
1648                                         sizeof(struct scsi_io_context),
1649                                         0, 0, NULL);
1650         if (!scsi_io_context_cache) {
1651                 printk(KERN_ERR "SCSI: can't init scsi io context cache\n");
1652                 return -ENOMEM;
1653         }
1654
1655         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1656                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1657                 int size = sgp->size * sizeof(struct scatterlist);
1658
1659                 sgp->slab = kmem_cache_create(sgp->name, size, 0,
1660                                 SLAB_HWCACHE_ALIGN, NULL);
1661                 if (!sgp->slab) {
1662                         printk(KERN_ERR "SCSI: can't init sg slab %s\n",
1663                                         sgp->name);
1664                 }
1665
1666                 sgp->pool = mempool_create_slab_pool(SG_MEMPOOL_SIZE,
1667                                                      sgp->slab);
1668                 if (!sgp->pool) {
1669                         printk(KERN_ERR "SCSI: can't init sg mempool %s\n",
1670                                         sgp->name);
1671                 }
1672         }
1673
1674         return 0;
1675 }
1676
1677 void scsi_exit_queue(void)
1678 {
1679         int i;
1680
1681         kmem_cache_destroy(scsi_io_context_cache);
1682
1683         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1684                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1685                 mempool_destroy(sgp->pool);
1686                 kmem_cache_destroy(sgp->slab);
1687         }
1688 }
1689
1690 /**
1691  *      scsi_mode_select - issue a mode select
1692  *      @sdev:  SCSI device to be queried
1693  *      @pf:    Page format bit (1 == standard, 0 == vendor specific)
1694  *      @sp:    Save page bit (0 == don't save, 1 == save)
1695  *      @modepage: mode page being requested
1696  *      @buffer: request buffer (may not be smaller than eight bytes)
1697  *      @len:   length of request buffer.
1698  *      @timeout: command timeout
1699  *      @retries: number of retries before failing
1700  *      @data: returns a structure abstracting the mode header data
1701  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
1702  *              must be SCSI_SENSE_BUFFERSIZE big.
1703  *
1704  *      Returns zero if successful; negative error number or scsi
1705  *      status on error
1706  *
1707  */
1708 int
1709 scsi_mode_select(struct scsi_device *sdev, int pf, int sp, int modepage,
1710                  unsigned char *buffer, int len, int timeout, int retries,
1711                  struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1712 {
1713         unsigned char cmd[10];
1714         unsigned char *real_buffer;
1715         int ret;
1716
1717         memset(cmd, 0, sizeof(cmd));
1718         cmd[1] = (pf ? 0x10 : 0) | (sp ? 0x01 : 0);
1719
1720         if (sdev->use_10_for_ms) {
1721                 if (len > 65535)
1722                         return -EINVAL;
1723                 real_buffer = kmalloc(8 + len, GFP_KERNEL);
1724                 if (!real_buffer)
1725                         return -ENOMEM;
1726                 memcpy(real_buffer + 8, buffer, len);
1727                 len += 8;
1728                 real_buffer[0] = 0;
1729                 real_buffer[1] = 0;
1730                 real_buffer[2] = data->medium_type;
1731                 real_buffer[3] = data->device_specific;
1732                 real_buffer[4] = data->longlba ? 0x01 : 0;
1733                 real_buffer[5] = 0;
1734                 real_buffer[6] = data->block_descriptor_length >> 8;
1735                 real_buffer[7] = data->block_descriptor_length;
1736
1737                 cmd[0] = MODE_SELECT_10;
1738                 cmd[7] = len >> 8;
1739                 cmd[8] = len;
1740         } else {
1741                 if (len > 255 || data->block_descriptor_length > 255 ||
1742                     data->longlba)
1743                         return -EINVAL;
1744
1745                 real_buffer = kmalloc(4 + len, GFP_KERNEL);
1746                 if (!real_buffer)
1747                         return -ENOMEM;
1748                 memcpy(real_buffer + 4, buffer, len);
1749                 len += 4;
1750                 real_buffer[0] = 0;
1751                 real_buffer[1] = data->medium_type;
1752                 real_buffer[2] = data->device_specific;
1753                 real_buffer[3] = data->block_descriptor_length;
1754                 
1755
1756                 cmd[0] = MODE_SELECT;
1757                 cmd[4] = len;
1758         }
1759
1760         ret = scsi_execute_req(sdev, cmd, DMA_TO_DEVICE, real_buffer, len,
1761                                sshdr, timeout, retries);
1762         kfree(real_buffer);
1763         return ret;
1764 }
1765 EXPORT_SYMBOL_GPL(scsi_mode_select);
1766
1767 /**
1768  *      scsi_mode_sense - issue a mode sense, falling back from 10 to six bytes if necessary.
1769  *      @sdev:  SCSI device to be queried
1770  *      @dbd:   set if mode sense will allow block descriptors to be returned
1771  *      @modepage: mode page being requested
1772  *      @buffer: request buffer (may not be smaller than eight bytes)
1773  *      @len:   length of request buffer.
1774  *      @timeout: command timeout
1775  *      @retries: number of retries before failing
1776  *      @data: returns a structure abstracting the mode header data
1777  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
1778  *              must be SCSI_SENSE_BUFFERSIZE big.
1779  *
1780  *      Returns zero if unsuccessful, or the header offset (either 4
1781  *      or 8 depending on whether a six or ten byte command was
1782  *      issued) if successful.
1783  */
1784 int
1785 scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
1786                   unsigned char *buffer, int len, int timeout, int retries,
1787                   struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1788 {
1789         unsigned char cmd[12];
1790         int use_10_for_ms;
1791         int header_length;
1792         int result;
1793         struct scsi_sense_hdr my_sshdr;
1794
1795         memset(data, 0, sizeof(*data));
1796         memset(&cmd[0], 0, 12);
1797         cmd[1] = dbd & 0x18;    /* allows DBD and LLBA bits */
1798         cmd[2] = modepage;
1799
1800         /* caller might not be interested in sense, but we need it */
1801         if (!sshdr)
1802                 sshdr = &my_sshdr;
1803
1804  retry:
1805         use_10_for_ms = sdev->use_10_for_ms;
1806
1807         if (use_10_for_ms) {
1808                 if (len < 8)
1809                         len = 8;
1810
1811                 cmd[0] = MODE_SENSE_10;
1812                 cmd[8] = len;
1813                 header_length = 8;
1814         } else {
1815                 if (len < 4)
1816                         len = 4;
1817
1818                 cmd[0] = MODE_SENSE;
1819                 cmd[4] = len;
1820                 header_length = 4;
1821         }
1822
1823         memset(buffer, 0, len);
1824
1825         result = scsi_execute_req(sdev, cmd, DMA_FROM_DEVICE, buffer, len,
1826                                   sshdr, timeout, retries);
1827
1828         /* This code looks awful: what it's doing is making sure an
1829          * ILLEGAL REQUEST sense return identifies the actual command
1830          * byte as the problem.  MODE_SENSE commands can return
1831          * ILLEGAL REQUEST if the code page isn't supported */
1832
1833         if (use_10_for_ms && !scsi_status_is_good(result) &&
1834             (driver_byte(result) & DRIVER_SENSE)) {
1835                 if (scsi_sense_valid(sshdr)) {
1836                         if ((sshdr->sense_key == ILLEGAL_REQUEST) &&
1837                             (sshdr->asc == 0x20) && (sshdr->ascq == 0)) {
1838                                 /* 
1839                                  * Invalid command operation code
1840                                  */
1841                                 sdev->use_10_for_ms = 0;
1842                                 goto retry;
1843                         }
1844                 }
1845         }
1846
1847         if(scsi_status_is_good(result)) {
1848                 if (unlikely(buffer[0] == 0x86 && buffer[1] == 0x0b &&
1849                              (modepage == 6 || modepage == 8))) {
1850                         /* Initio breakage? */
1851                         header_length = 0;
1852                         data->length = 13;
1853                         data->medium_type = 0;
1854                         data->device_specific = 0;
1855                         data->longlba = 0;
1856                         data->block_descriptor_length = 0;
1857                 } else if(use_10_for_ms) {
1858                         data->length = buffer[0]*256 + buffer[1] + 2;
1859                         data->medium_type = buffer[2];
1860                         data->device_specific = buffer[3];
1861                         data->longlba = buffer[4] & 0x01;
1862                         data->block_descriptor_length = buffer[6]*256
1863                                 + buffer[7];
1864                 } else {
1865                         data->length = buffer[0] + 1;
1866                         data->medium_type = buffer[1];
1867                         data->device_specific = buffer[2];
1868                         data->block_descriptor_length = buffer[3];
1869                 }
1870                 data->header_length = header_length;
1871         }
1872
1873         return result;
1874 }
1875 EXPORT_SYMBOL(scsi_mode_sense);
1876
1877 /**
1878  *      scsi_test_unit_ready - test if unit is ready
1879  *      @sdev:  scsi device to change the state of.
1880  *      @timeout: command timeout
1881  *      @retries: number of retries before failing
1882  *      @sshdr_external: Optional pointer to struct scsi_sense_hdr for
1883  *              returning sense. Make sure that this is cleared before passing
1884  *              in.
1885  *
1886  *      Returns zero if unsuccessful or an error if TUR failed.  For
1887  *      removable media, a return of NOT_READY or UNIT_ATTENTION is
1888  *      translated to success, with the ->changed flag updated.
1889  **/
1890 int
1891 scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries,
1892                      struct scsi_sense_hdr *sshdr_external)
1893 {
1894         char cmd[] = {
1895                 TEST_UNIT_READY, 0, 0, 0, 0, 0,
1896         };
1897         struct scsi_sense_hdr *sshdr;
1898         int result;
1899
1900         if (!sshdr_external)
1901                 sshdr = kzalloc(sizeof(*sshdr), GFP_KERNEL);
1902         else
1903                 sshdr = sshdr_external;
1904
1905         /* try to eat the UNIT_ATTENTION if there are enough retries */
1906         do {
1907                 result = scsi_execute_req(sdev, cmd, DMA_NONE, NULL, 0, sshdr,
1908                                           timeout, retries);
1909         } while ((driver_byte(result) & DRIVER_SENSE) &&
1910                  sshdr && sshdr->sense_key == UNIT_ATTENTION &&
1911                  --retries);
1912
1913         if (!sshdr)
1914                 /* could not allocate sense buffer, so can't process it */
1915                 return result;
1916
1917         if ((driver_byte(result) & DRIVER_SENSE) && sdev->removable) {
1918
1919                 if ((scsi_sense_valid(sshdr)) &&
1920                     ((sshdr->sense_key == UNIT_ATTENTION) ||
1921                      (sshdr->sense_key == NOT_READY))) {
1922                         sdev->changed = 1;
1923                         result = 0;
1924                 }
1925         }
1926         if (!sshdr_external)
1927                 kfree(sshdr);
1928         return result;
1929 }
1930 EXPORT_SYMBOL(scsi_test_unit_ready);
1931
1932 /**
1933  *      scsi_device_set_state - Take the given device through the device state model.
1934  *      @sdev:  scsi device to change the state of.
1935  *      @state: state to change to.
1936  *
1937  *      Returns zero if unsuccessful or an error if the requested 
1938  *      transition is illegal.
1939  */
1940 int
1941 scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
1942 {
1943         enum scsi_device_state oldstate = sdev->sdev_state;
1944
1945         if (state == oldstate)
1946                 return 0;
1947
1948         switch (state) {
1949         case SDEV_CREATED:
1950                 /* There are no legal states that come back to
1951                  * created.  This is the manually initialised start
1952                  * state */
1953                 goto illegal;
1954                         
1955         case SDEV_RUNNING:
1956                 switch (oldstate) {
1957                 case SDEV_CREATED:
1958                 case SDEV_OFFLINE:
1959                 case SDEV_QUIESCE:
1960                 case SDEV_BLOCK:
1961                         break;
1962                 default:
1963                         goto illegal;
1964                 }
1965                 break;
1966
1967         case SDEV_QUIESCE:
1968                 switch (oldstate) {
1969                 case SDEV_RUNNING:
1970                 case SDEV_OFFLINE:
1971                         break;
1972                 default:
1973                         goto illegal;
1974                 }
1975                 break;
1976
1977         case SDEV_OFFLINE:
1978                 switch (oldstate) {
1979                 case SDEV_CREATED:
1980                 case SDEV_RUNNING:
1981                 case SDEV_QUIESCE:
1982                 case SDEV_BLOCK:
1983                         break;
1984                 default:
1985                         goto illegal;
1986                 }
1987                 break;
1988
1989         case SDEV_BLOCK:
1990                 switch (oldstate) {
1991                 case SDEV_CREATED:
1992                 case SDEV_RUNNING:
1993                         break;
1994                 default:
1995                         goto illegal;
1996                 }
1997                 break;
1998
1999         case SDEV_CANCEL:
2000                 switch (oldstate) {
2001                 case SDEV_CREATED:
2002                 case SDEV_RUNNING:
2003                 case SDEV_QUIESCE:
2004                 case SDEV_OFFLINE:
2005                 case SDEV_BLOCK:
2006                         break;
2007                 default:
2008                         goto illegal;
2009                 }
2010                 break;
2011
2012         case SDEV_DEL:
2013                 switch (oldstate) {
2014                 case SDEV_CREATED:
2015                 case SDEV_RUNNING:
2016                 case SDEV_OFFLINE:
2017                 case SDEV_CANCEL:
2018                         break;
2019                 default:
2020                         goto illegal;
2021                 }
2022                 break;
2023
2024         }
2025         sdev->sdev_state = state;
2026         return 0;
2027
2028  illegal:
2029         SCSI_LOG_ERROR_RECOVERY(1, 
2030                                 sdev_printk(KERN_ERR, sdev,
2031                                             "Illegal state transition %s->%s\n",
2032                                             scsi_device_state_name(oldstate),
2033                                             scsi_device_state_name(state))
2034                                 );
2035         return -EINVAL;
2036 }
2037 EXPORT_SYMBOL(scsi_device_set_state);
2038
2039 /**
2040  *      sdev_evt_emit - emit a single SCSI device uevent
2041  *      @sdev: associated SCSI device
2042  *      @evt: event to emit
2043  *
2044  *      Send a single uevent (scsi_event) to the associated scsi_device.
2045  */
2046 static void scsi_evt_emit(struct scsi_device *sdev, struct scsi_event *evt)
2047 {
2048         int idx = 0;
2049         char *envp[3];
2050
2051         switch (evt->evt_type) {
2052         case SDEV_EVT_MEDIA_CHANGE:
2053                 envp[idx++] = "SDEV_MEDIA_CHANGE=1";
2054                 break;
2055
2056         default:
2057                 /* do nothing */
2058                 break;
2059         }
2060
2061         envp[idx++] = NULL;
2062
2063         kobject_uevent_env(&sdev->sdev_gendev.kobj, KOBJ_CHANGE, envp);
2064 }
2065
2066 /**
2067  *      sdev_evt_thread - send a uevent for each scsi event
2068  *      @work: work struct for scsi_device
2069  *
2070  *      Dispatch queued events to their associated scsi_device kobjects
2071  *      as uevents.
2072  */
2073 void scsi_evt_thread(struct work_struct *work)
2074 {
2075         struct scsi_device *sdev;
2076         LIST_HEAD(event_list);
2077
2078         sdev = container_of(work, struct scsi_device, event_work);
2079
2080         while (1) {
2081                 struct scsi_event *evt;
2082                 struct list_head *this, *tmp;
2083                 unsigned long flags;
2084
2085                 spin_lock_irqsave(&sdev->list_lock, flags);
2086                 list_splice_init(&sdev->event_list, &event_list);
2087                 spin_unlock_irqrestore(&sdev->list_lock, flags);
2088
2089                 if (list_empty(&event_list))
2090                         break;
2091
2092                 list_for_each_safe(this, tmp, &event_list) {
2093                         evt = list_entry(this, struct scsi_event, node);
2094                         list_del(&evt->node);
2095                         scsi_evt_emit(sdev, evt);
2096                         kfree(evt);
2097                 }
2098         }
2099 }
2100
2101 /**
2102  *      sdev_evt_send - send asserted event to uevent thread
2103  *      @sdev: scsi_device event occurred on
2104  *      @evt: event to send
2105  *
2106  *      Assert scsi device event asynchronously.
2107  */
2108 void sdev_evt_send(struct scsi_device *sdev, struct scsi_event *evt)
2109 {
2110         unsigned long flags;
2111
2112         if (!test_bit(evt->evt_type, sdev->supported_events)) {
2113                 kfree(evt);
2114                 return;
2115         }
2116
2117         spin_lock_irqsave(&sdev->list_lock, flags);
2118         list_add_tail(&evt->node, &sdev->event_list);
2119         schedule_work(&sdev->event_work);
2120         spin_unlock_irqrestore(&sdev->list_lock, flags);
2121 }
2122 EXPORT_SYMBOL_GPL(sdev_evt_send);
2123
2124 /**
2125  *      sdev_evt_alloc - allocate a new scsi event
2126  *      @evt_type: type of event to allocate
2127  *      @gfpflags: GFP flags for allocation
2128  *
2129  *      Allocates and returns a new scsi_event.
2130  */
2131 struct scsi_event *sdev_evt_alloc(enum scsi_device_event evt_type,
2132                                   gfp_t gfpflags)
2133 {
2134         struct scsi_event *evt = kzalloc(sizeof(struct scsi_event), gfpflags);
2135         if (!evt)
2136                 return NULL;
2137
2138         evt->evt_type = evt_type;
2139         INIT_LIST_HEAD(&evt->node);
2140
2141         /* evt_type-specific initialization, if any */
2142         switch (evt_type) {
2143         case SDEV_EVT_MEDIA_CHANGE:
2144         default:
2145                 /* do nothing */
2146                 break;
2147         }
2148
2149         return evt;
2150 }
2151 EXPORT_SYMBOL_GPL(sdev_evt_alloc);
2152
2153 /**
2154  *      sdev_evt_send_simple - send asserted event to uevent thread
2155  *      @sdev: scsi_device event occurred on
2156  *      @evt_type: type of event to send
2157  *      @gfpflags: GFP flags for allocation
2158  *
2159  *      Assert scsi device event asynchronously, given an event type.
2160  */
2161 void sdev_evt_send_simple(struct scsi_device *sdev,
2162                           enum scsi_device_event evt_type, gfp_t gfpflags)
2163 {
2164         struct scsi_event *evt = sdev_evt_alloc(evt_type, gfpflags);
2165         if (!evt) {
2166                 sdev_printk(KERN_ERR, sdev, "event %d eaten due to OOM\n",
2167                             evt_type);
2168                 return;
2169         }
2170
2171         sdev_evt_send(sdev, evt);
2172 }
2173 EXPORT_SYMBOL_GPL(sdev_evt_send_simple);
2174
2175 /**
2176  *      scsi_device_quiesce - Block user issued commands.
2177  *      @sdev:  scsi device to quiesce.
2178  *
2179  *      This works by trying to transition to the SDEV_QUIESCE state
2180  *      (which must be a legal transition).  When the device is in this
2181  *      state, only special requests will be accepted, all others will
2182  *      be deferred.  Since special requests may also be requeued requests,
2183  *      a successful return doesn't guarantee the device will be 
2184  *      totally quiescent.
2185  *
2186  *      Must be called with user context, may sleep.
2187  *
2188  *      Returns zero if unsuccessful or an error if not.
2189  */
2190 int
2191 scsi_device_quiesce(struct scsi_device *sdev)
2192 {
2193         int err = scsi_device_set_state(sdev, SDEV_QUIESCE);
2194         if (err)
2195                 return err;
2196
2197         scsi_run_queue(sdev->request_queue);
2198         while (sdev->device_busy) {
2199                 msleep_interruptible(200);
2200                 scsi_run_queue(sdev->request_queue);
2201         }
2202         return 0;
2203 }
2204 EXPORT_SYMBOL(scsi_device_quiesce);
2205
2206 /**
2207  *      scsi_device_resume - Restart user issued commands to a quiesced device.
2208  *      @sdev:  scsi device to resume.
2209  *
2210  *      Moves the device from quiesced back to running and restarts the
2211  *      queues.
2212  *
2213  *      Must be called with user context, may sleep.
2214  */
2215 void
2216 scsi_device_resume(struct scsi_device *sdev)
2217 {
2218         if(scsi_device_set_state(sdev, SDEV_RUNNING))
2219                 return;
2220         scsi_run_queue(sdev->request_queue);
2221 }
2222 EXPORT_SYMBOL(scsi_device_resume);
2223
2224 static void
2225 device_quiesce_fn(struct scsi_device *sdev, void *data)
2226 {
2227         scsi_device_quiesce(sdev);
2228 }
2229
2230 void
2231 scsi_target_quiesce(struct scsi_target *starget)
2232 {
2233         starget_for_each_device(starget, NULL, device_quiesce_fn);
2234 }
2235 EXPORT_SYMBOL(scsi_target_quiesce);
2236
2237 static void
2238 device_resume_fn(struct scsi_device *sdev, void *data)
2239 {
2240         scsi_device_resume(sdev);
2241 }
2242
2243 void
2244 scsi_target_resume(struct scsi_target *starget)
2245 {
2246         starget_for_each_device(starget, NULL, device_resume_fn);
2247 }
2248 EXPORT_SYMBOL(scsi_target_resume);
2249
2250 /**
2251  * scsi_internal_device_block - internal function to put a device temporarily into the SDEV_BLOCK state
2252  * @sdev:       device to block
2253  *
2254  * Block request made by scsi lld's to temporarily stop all
2255  * scsi commands on the specified device.  Called from interrupt
2256  * or normal process context.
2257  *
2258  * Returns zero if successful or error if not
2259  *
2260  * Notes:       
2261  *      This routine transitions the device to the SDEV_BLOCK state
2262  *      (which must be a legal transition).  When the device is in this
2263  *      state, all commands are deferred until the scsi lld reenables
2264  *      the device with scsi_device_unblock or device_block_tmo fires.
2265  *      This routine assumes the host_lock is held on entry.
2266  */
2267 int
2268 scsi_internal_device_block(struct scsi_device *sdev)
2269 {
2270         struct request_queue *q = sdev->request_queue;
2271         unsigned long flags;
2272         int err = 0;
2273
2274         err = scsi_device_set_state(sdev, SDEV_BLOCK);
2275         if (err)
2276                 return err;
2277
2278         /* 
2279          * The device has transitioned to SDEV_BLOCK.  Stop the
2280          * block layer from calling the midlayer with this device's
2281          * request queue. 
2282          */
2283         spin_lock_irqsave(q->queue_lock, flags);
2284         blk_stop_queue(q);
2285         spin_unlock_irqrestore(q->queue_lock, flags);
2286
2287         return 0;
2288 }
2289 EXPORT_SYMBOL_GPL(scsi_internal_device_block);
2290  
2291 /**
2292  * scsi_internal_device_unblock - resume a device after a block request
2293  * @sdev:       device to resume
2294  *
2295  * Called by scsi lld's or the midlayer to restart the device queue
2296  * for the previously suspended scsi device.  Called from interrupt or
2297  * normal process context.
2298  *
2299  * Returns zero if successful or error if not.
2300  *
2301  * Notes:       
2302  *      This routine transitions the device to the SDEV_RUNNING state
2303  *      (which must be a legal transition) allowing the midlayer to
2304  *      goose the queue for this device.  This routine assumes the 
2305  *      host_lock is held upon entry.
2306  */
2307 int
2308 scsi_internal_device_unblock(struct scsi_device *sdev)
2309 {
2310         struct request_queue *q = sdev->request_queue; 
2311         int err;
2312         unsigned long flags;
2313         
2314         /* 
2315          * Try to transition the scsi device to SDEV_RUNNING
2316          * and goose the device queue if successful.  
2317          */
2318         err = scsi_device_set_state(sdev, SDEV_RUNNING);
2319         if (err)
2320                 return err;
2321
2322         spin_lock_irqsave(q->queue_lock, flags);
2323         blk_start_queue(q);
2324         spin_unlock_irqrestore(q->queue_lock, flags);
2325
2326         return 0;
2327 }
2328 EXPORT_SYMBOL_GPL(scsi_internal_device_unblock);
2329
2330 static void
2331 device_block(struct scsi_device *sdev, void *data)
2332 {
2333         scsi_internal_device_block(sdev);
2334 }
2335
2336 static int
2337 target_block(struct device *dev, void *data)
2338 {
2339         if (scsi_is_target_device(dev))
2340                 starget_for_each_device(to_scsi_target(dev), NULL,
2341                                         device_block);
2342         return 0;
2343 }
2344
2345 void
2346 scsi_target_block(struct device *dev)
2347 {
2348         if (scsi_is_target_device(dev))
2349                 starget_for_each_device(to_scsi_target(dev), NULL,
2350                                         device_block);
2351         else
2352                 device_for_each_child(dev, NULL, target_block);
2353 }
2354 EXPORT_SYMBOL_GPL(scsi_target_block);
2355
2356 static void
2357 device_unblock(struct scsi_device *sdev, void *data)
2358 {
2359         scsi_internal_device_unblock(sdev);
2360 }
2361
2362 static int
2363 target_unblock(struct device *dev, void *data)
2364 {
2365         if (scsi_is_target_device(dev))
2366                 starget_for_each_device(to_scsi_target(dev), NULL,
2367                                         device_unblock);
2368         return 0;
2369 }
2370
2371 void
2372 scsi_target_unblock(struct device *dev)
2373 {
2374         if (scsi_is_target_device(dev))
2375                 starget_for_each_device(to_scsi_target(dev), NULL,
2376                                         device_unblock);
2377         else
2378                 device_for_each_child(dev, NULL, target_unblock);
2379 }
2380 EXPORT_SYMBOL_GPL(scsi_target_unblock);
2381
2382 /**
2383  * scsi_kmap_atomic_sg - find and atomically map an sg-elemnt
2384  * @sgl:        scatter-gather list
2385  * @sg_count:   number of segments in sg
2386  * @offset:     offset in bytes into sg, on return offset into the mapped area
2387  * @len:        bytes to map, on return number of bytes mapped
2388  *
2389  * Returns virtual address of the start of the mapped page
2390  */
2391 void *scsi_kmap_atomic_sg(struct scatterlist *sgl, int sg_count,
2392                           size_t *offset, size_t *len)
2393 {
2394         int i;
2395         size_t sg_len = 0, len_complete = 0;
2396         struct scatterlist *sg;
2397         struct page *page;
2398
2399         WARN_ON(!irqs_disabled());
2400
2401         for_each_sg(sgl, sg, sg_count, i) {
2402                 len_complete = sg_len; /* Complete sg-entries */
2403                 sg_len += sg->length;
2404                 if (sg_len > *offset)
2405                         break;
2406         }
2407
2408         if (unlikely(i == sg_count)) {
2409                 printk(KERN_ERR "%s: Bytes in sg: %zu, requested offset %zu, "
2410                         "elements %d\n",
2411                        __FUNCTION__, sg_len, *offset, sg_count);
2412                 WARN_ON(1);
2413                 return NULL;
2414         }
2415
2416         /* Offset starting from the beginning of first page in this sg-entry */
2417         *offset = *offset - len_complete + sg->offset;
2418
2419         /* Assumption: contiguous pages can be accessed as "page + i" */
2420         page = nth_page(sg_page(sg), (*offset >> PAGE_SHIFT));
2421         *offset &= ~PAGE_MASK;
2422
2423         /* Bytes in this sg-entry from *offset to the end of the page */
2424         sg_len = PAGE_SIZE - *offset;
2425         if (*len > sg_len)
2426                 *len = sg_len;
2427
2428         return kmap_atomic(page, KM_BIO_SRC_IRQ);
2429 }
2430 EXPORT_SYMBOL(scsi_kmap_atomic_sg);
2431
2432 /**
2433  * scsi_kunmap_atomic_sg - atomically unmap a virtual address, previously mapped with scsi_kmap_atomic_sg
2434  * @virt:       virtual address to be unmapped
2435  */
2436 void scsi_kunmap_atomic_sg(void *virt)
2437 {
2438         kunmap_atomic(virt, KM_BIO_SRC_IRQ);
2439 }
2440 EXPORT_SYMBOL(scsi_kunmap_atomic_sg);