iommu sg merging: call blk_queue_segment_boundary in __scsi_alloc_queue
[linux-2.6] / drivers / scsi / scsi_lib.c
1 /*
2  *  scsi_lib.c Copyright (C) 1999 Eric Youngdale
3  *
4  *  SCSI queueing library.
5  *      Initial versions: Eric Youngdale (eric@andante.org).
6  *                        Based upon conversations with large numbers
7  *                        of people at Linux Expo.
8  */
9
10 #include <linux/bio.h>
11 #include <linux/bitops.h>
12 #include <linux/blkdev.h>
13 #include <linux/completion.h>
14 #include <linux/kernel.h>
15 #include <linux/mempool.h>
16 #include <linux/slab.h>
17 #include <linux/init.h>
18 #include <linux/pci.h>
19 #include <linux/delay.h>
20 #include <linux/hardirq.h>
21 #include <linux/scatterlist.h>
22
23 #include <scsi/scsi.h>
24 #include <scsi/scsi_cmnd.h>
25 #include <scsi/scsi_dbg.h>
26 #include <scsi/scsi_device.h>
27 #include <scsi/scsi_driver.h>
28 #include <scsi/scsi_eh.h>
29 #include <scsi/scsi_host.h>
30
31 #include "scsi_priv.h"
32 #include "scsi_logging.h"
33
34
35 #define SG_MEMPOOL_NR           ARRAY_SIZE(scsi_sg_pools)
36 #define SG_MEMPOOL_SIZE         2
37
38 struct scsi_host_sg_pool {
39         size_t          size;
40         char            *name;
41         struct kmem_cache       *slab;
42         mempool_t       *pool;
43 };
44
45 #define SP(x) { x, "sgpool-" __stringify(x) }
46 #if (SCSI_MAX_SG_SEGMENTS < 32)
47 #error SCSI_MAX_SG_SEGMENTS is too small (must be 32 or greater)
48 #endif
49 static struct scsi_host_sg_pool scsi_sg_pools[] = {
50         SP(8),
51         SP(16),
52 #if (SCSI_MAX_SG_SEGMENTS > 32)
53         SP(32),
54 #if (SCSI_MAX_SG_SEGMENTS > 64)
55         SP(64),
56 #if (SCSI_MAX_SG_SEGMENTS > 128)
57         SP(128),
58 #if (SCSI_MAX_SG_SEGMENTS > 256)
59 #error SCSI_MAX_SG_SEGMENTS is too large (256 MAX)
60 #endif
61 #endif
62 #endif
63 #endif
64         SP(SCSI_MAX_SG_SEGMENTS)
65 };
66 #undef SP
67
68 static struct kmem_cache *scsi_bidi_sdb_cache;
69
70 static void scsi_run_queue(struct request_queue *q);
71
72 /*
73  * Function:    scsi_unprep_request()
74  *
75  * Purpose:     Remove all preparation done for a request, including its
76  *              associated scsi_cmnd, so that it can be requeued.
77  *
78  * Arguments:   req     - request to unprepare
79  *
80  * Lock status: Assumed that no locks are held upon entry.
81  *
82  * Returns:     Nothing.
83  */
84 static void scsi_unprep_request(struct request *req)
85 {
86         struct scsi_cmnd *cmd = req->special;
87
88         req->cmd_flags &= ~REQ_DONTPREP;
89         req->special = NULL;
90
91         scsi_put_command(cmd);
92 }
93
94 /*
95  * Function:    scsi_queue_insert()
96  *
97  * Purpose:     Insert a command in the midlevel queue.
98  *
99  * Arguments:   cmd    - command that we are adding to queue.
100  *              reason - why we are inserting command to queue.
101  *
102  * Lock status: Assumed that lock is not held upon entry.
103  *
104  * Returns:     Nothing.
105  *
106  * Notes:       We do this for one of two cases.  Either the host is busy
107  *              and it cannot accept any more commands for the time being,
108  *              or the device returned QUEUE_FULL and can accept no more
109  *              commands.
110  * Notes:       This could be called either from an interrupt context or a
111  *              normal process context.
112  */
113 int scsi_queue_insert(struct scsi_cmnd *cmd, int reason)
114 {
115         struct Scsi_Host *host = cmd->device->host;
116         struct scsi_device *device = cmd->device;
117         struct request_queue *q = device->request_queue;
118         unsigned long flags;
119
120         SCSI_LOG_MLQUEUE(1,
121                  printk("Inserting command %p into mlqueue\n", cmd));
122
123         /*
124          * Set the appropriate busy bit for the device/host.
125          *
126          * If the host/device isn't busy, assume that something actually
127          * completed, and that we should be able to queue a command now.
128          *
129          * Note that the prior mid-layer assumption that any host could
130          * always queue at least one command is now broken.  The mid-layer
131          * will implement a user specifiable stall (see
132          * scsi_host.max_host_blocked and scsi_device.max_device_blocked)
133          * if a command is requeued with no other commands outstanding
134          * either for the device or for the host.
135          */
136         if (reason == SCSI_MLQUEUE_HOST_BUSY)
137                 host->host_blocked = host->max_host_blocked;
138         else if (reason == SCSI_MLQUEUE_DEVICE_BUSY)
139                 device->device_blocked = device->max_device_blocked;
140
141         /*
142          * Decrement the counters, since these commands are no longer
143          * active on the host/device.
144          */
145         scsi_device_unbusy(device);
146
147         /*
148          * Requeue this command.  It will go before all other commands
149          * that are already in the queue.
150          *
151          * NOTE: there is magic here about the way the queue is plugged if
152          * we have no outstanding commands.
153          * 
154          * Although we *don't* plug the queue, we call the request
155          * function.  The SCSI request function detects the blocked condition
156          * and plugs the queue appropriately.
157          */
158         spin_lock_irqsave(q->queue_lock, flags);
159         blk_requeue_request(q, cmd->request);
160         spin_unlock_irqrestore(q->queue_lock, flags);
161
162         scsi_run_queue(q);
163
164         return 0;
165 }
166
167 /**
168  * scsi_execute - insert request and wait for the result
169  * @sdev:       scsi device
170  * @cmd:        scsi command
171  * @data_direction: data direction
172  * @buffer:     data buffer
173  * @bufflen:    len of buffer
174  * @sense:      optional sense buffer
175  * @timeout:    request timeout in seconds
176  * @retries:    number of times to retry request
177  * @flags:      or into request flags;
178  *
179  * returns the req->errors value which is the scsi_cmnd result
180  * field.
181  */
182 int scsi_execute(struct scsi_device *sdev, const unsigned char *cmd,
183                  int data_direction, void *buffer, unsigned bufflen,
184                  unsigned char *sense, int timeout, int retries, int flags)
185 {
186         struct request *req;
187         int write = (data_direction == DMA_TO_DEVICE);
188         int ret = DRIVER_ERROR << 24;
189
190         req = blk_get_request(sdev->request_queue, write, __GFP_WAIT);
191
192         if (bufflen &&  blk_rq_map_kern(sdev->request_queue, req,
193                                         buffer, bufflen, __GFP_WAIT))
194                 goto out;
195
196         req->cmd_len = COMMAND_SIZE(cmd[0]);
197         memcpy(req->cmd, cmd, req->cmd_len);
198         req->sense = sense;
199         req->sense_len = 0;
200         req->retries = retries;
201         req->timeout = timeout;
202         req->cmd_type = REQ_TYPE_BLOCK_PC;
203         req->cmd_flags |= flags | REQ_QUIET | REQ_PREEMPT;
204
205         /*
206          * head injection *required* here otherwise quiesce won't work
207          */
208         blk_execute_rq(req->q, NULL, req, 1);
209
210         ret = req->errors;
211  out:
212         blk_put_request(req);
213
214         return ret;
215 }
216 EXPORT_SYMBOL(scsi_execute);
217
218
219 int scsi_execute_req(struct scsi_device *sdev, const unsigned char *cmd,
220                      int data_direction, void *buffer, unsigned bufflen,
221                      struct scsi_sense_hdr *sshdr, int timeout, int retries)
222 {
223         char *sense = NULL;
224         int result;
225         
226         if (sshdr) {
227                 sense = kzalloc(SCSI_SENSE_BUFFERSIZE, GFP_NOIO);
228                 if (!sense)
229                         return DRIVER_ERROR << 24;
230         }
231         result = scsi_execute(sdev, cmd, data_direction, buffer, bufflen,
232                               sense, timeout, retries, 0);
233         if (sshdr)
234                 scsi_normalize_sense(sense, SCSI_SENSE_BUFFERSIZE, sshdr);
235
236         kfree(sense);
237         return result;
238 }
239 EXPORT_SYMBOL(scsi_execute_req);
240
241 struct scsi_io_context {
242         void *data;
243         void (*done)(void *data, char *sense, int result, int resid);
244         char sense[SCSI_SENSE_BUFFERSIZE];
245 };
246
247 static struct kmem_cache *scsi_io_context_cache;
248
249 static void scsi_end_async(struct request *req, int uptodate)
250 {
251         struct scsi_io_context *sioc = req->end_io_data;
252
253         if (sioc->done)
254                 sioc->done(sioc->data, sioc->sense, req->errors, req->data_len);
255
256         kmem_cache_free(scsi_io_context_cache, sioc);
257         __blk_put_request(req->q, req);
258 }
259
260 static int scsi_merge_bio(struct request *rq, struct bio *bio)
261 {
262         struct request_queue *q = rq->q;
263
264         bio->bi_flags &= ~(1 << BIO_SEG_VALID);
265         if (rq_data_dir(rq) == WRITE)
266                 bio->bi_rw |= (1 << BIO_RW);
267         blk_queue_bounce(q, &bio);
268
269         return blk_rq_append_bio(q, rq, bio);
270 }
271
272 static void scsi_bi_endio(struct bio *bio, int error)
273 {
274         bio_put(bio);
275 }
276
277 /**
278  * scsi_req_map_sg - map a scatterlist into a request
279  * @rq:         request to fill
280  * @sgl:        scatterlist
281  * @nsegs:      number of elements
282  * @bufflen:    len of buffer
283  * @gfp:        memory allocation flags
284  *
285  * scsi_req_map_sg maps a scatterlist into a request so that the
286  * request can be sent to the block layer. We do not trust the scatterlist
287  * sent to use, as some ULDs use that struct to only organize the pages.
288  */
289 static int scsi_req_map_sg(struct request *rq, struct scatterlist *sgl,
290                            int nsegs, unsigned bufflen, gfp_t gfp)
291 {
292         struct request_queue *q = rq->q;
293         int nr_pages = (bufflen + sgl[0].offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
294         unsigned int data_len = bufflen, len, bytes, off;
295         struct scatterlist *sg;
296         struct page *page;
297         struct bio *bio = NULL;
298         int i, err, nr_vecs = 0;
299
300         for_each_sg(sgl, sg, nsegs, i) {
301                 page = sg_page(sg);
302                 off = sg->offset;
303                 len = sg->length;
304                 data_len += len;
305
306                 while (len > 0 && data_len > 0) {
307                         /*
308                          * sg sends a scatterlist that is larger than
309                          * the data_len it wants transferred for certain
310                          * IO sizes
311                          */
312                         bytes = min_t(unsigned int, len, PAGE_SIZE - off);
313                         bytes = min(bytes, data_len);
314
315                         if (!bio) {
316                                 nr_vecs = min_t(int, BIO_MAX_PAGES, nr_pages);
317                                 nr_pages -= nr_vecs;
318
319                                 bio = bio_alloc(gfp, nr_vecs);
320                                 if (!bio) {
321                                         err = -ENOMEM;
322                                         goto free_bios;
323                                 }
324                                 bio->bi_end_io = scsi_bi_endio;
325                         }
326
327                         if (bio_add_pc_page(q, bio, page, bytes, off) !=
328                             bytes) {
329                                 bio_put(bio);
330                                 err = -EINVAL;
331                                 goto free_bios;
332                         }
333
334                         if (bio->bi_vcnt >= nr_vecs) {
335                                 err = scsi_merge_bio(rq, bio);
336                                 if (err) {
337                                         bio_endio(bio, 0);
338                                         goto free_bios;
339                                 }
340                                 bio = NULL;
341                         }
342
343                         page++;
344                         len -= bytes;
345                         data_len -=bytes;
346                         off = 0;
347                 }
348         }
349
350         rq->buffer = rq->data = NULL;
351         rq->data_len = bufflen;
352         return 0;
353
354 free_bios:
355         while ((bio = rq->bio) != NULL) {
356                 rq->bio = bio->bi_next;
357                 /*
358                  * call endio instead of bio_put incase it was bounced
359                  */
360                 bio_endio(bio, 0);
361         }
362
363         return err;
364 }
365
366 /**
367  * scsi_execute_async - insert request
368  * @sdev:       scsi device
369  * @cmd:        scsi command
370  * @cmd_len:    length of scsi cdb
371  * @data_direction: DMA_TO_DEVICE, DMA_FROM_DEVICE, or DMA_NONE
372  * @buffer:     data buffer (this can be a kernel buffer or scatterlist)
373  * @bufflen:    len of buffer
374  * @use_sg:     if buffer is a scatterlist this is the number of elements
375  * @timeout:    request timeout in seconds
376  * @retries:    number of times to retry request
377  * @privdata:   data passed to done()
378  * @done:       callback function when done
379  * @gfp:        memory allocation flags
380  */
381 int scsi_execute_async(struct scsi_device *sdev, const unsigned char *cmd,
382                        int cmd_len, int data_direction, void *buffer, unsigned bufflen,
383                        int use_sg, int timeout, int retries, void *privdata,
384                        void (*done)(void *, char *, int, int), gfp_t gfp)
385 {
386         struct request *req;
387         struct scsi_io_context *sioc;
388         int err = 0;
389         int write = (data_direction == DMA_TO_DEVICE);
390
391         sioc = kmem_cache_zalloc(scsi_io_context_cache, gfp);
392         if (!sioc)
393                 return DRIVER_ERROR << 24;
394
395         req = blk_get_request(sdev->request_queue, write, gfp);
396         if (!req)
397                 goto free_sense;
398         req->cmd_type = REQ_TYPE_BLOCK_PC;
399         req->cmd_flags |= REQ_QUIET;
400
401         if (use_sg)
402                 err = scsi_req_map_sg(req, buffer, use_sg, bufflen, gfp);
403         else if (bufflen)
404                 err = blk_rq_map_kern(req->q, req, buffer, bufflen, gfp);
405
406         if (err)
407                 goto free_req;
408
409         req->cmd_len = cmd_len;
410         memset(req->cmd, 0, BLK_MAX_CDB); /* ATAPI hates garbage after CDB */
411         memcpy(req->cmd, cmd, req->cmd_len);
412         req->sense = sioc->sense;
413         req->sense_len = 0;
414         req->timeout = timeout;
415         req->retries = retries;
416         req->end_io_data = sioc;
417
418         sioc->data = privdata;
419         sioc->done = done;
420
421         blk_execute_rq_nowait(req->q, NULL, req, 1, scsi_end_async);
422         return 0;
423
424 free_req:
425         blk_put_request(req);
426 free_sense:
427         kmem_cache_free(scsi_io_context_cache, sioc);
428         return DRIVER_ERROR << 24;
429 }
430 EXPORT_SYMBOL_GPL(scsi_execute_async);
431
432 /*
433  * Function:    scsi_init_cmd_errh()
434  *
435  * Purpose:     Initialize cmd fields related to error handling.
436  *
437  * Arguments:   cmd     - command that is ready to be queued.
438  *
439  * Notes:       This function has the job of initializing a number of
440  *              fields related to error handling.   Typically this will
441  *              be called once for each command, as required.
442  */
443 static void scsi_init_cmd_errh(struct scsi_cmnd *cmd)
444 {
445         cmd->serial_number = 0;
446         scsi_set_resid(cmd, 0);
447         memset(cmd->sense_buffer, 0, SCSI_SENSE_BUFFERSIZE);
448         if (cmd->cmd_len == 0)
449                 cmd->cmd_len = COMMAND_SIZE(cmd->cmnd[0]);
450 }
451
452 void scsi_device_unbusy(struct scsi_device *sdev)
453 {
454         struct Scsi_Host *shost = sdev->host;
455         unsigned long flags;
456
457         spin_lock_irqsave(shost->host_lock, flags);
458         shost->host_busy--;
459         if (unlikely(scsi_host_in_recovery(shost) &&
460                      (shost->host_failed || shost->host_eh_scheduled)))
461                 scsi_eh_wakeup(shost);
462         spin_unlock(shost->host_lock);
463         spin_lock(sdev->request_queue->queue_lock);
464         sdev->device_busy--;
465         spin_unlock_irqrestore(sdev->request_queue->queue_lock, flags);
466 }
467
468 /*
469  * Called for single_lun devices on IO completion. Clear starget_sdev_user,
470  * and call blk_run_queue for all the scsi_devices on the target -
471  * including current_sdev first.
472  *
473  * Called with *no* scsi locks held.
474  */
475 static void scsi_single_lun_run(struct scsi_device *current_sdev)
476 {
477         struct Scsi_Host *shost = current_sdev->host;
478         struct scsi_device *sdev, *tmp;
479         struct scsi_target *starget = scsi_target(current_sdev);
480         unsigned long flags;
481
482         spin_lock_irqsave(shost->host_lock, flags);
483         starget->starget_sdev_user = NULL;
484         spin_unlock_irqrestore(shost->host_lock, flags);
485
486         /*
487          * Call blk_run_queue for all LUNs on the target, starting with
488          * current_sdev. We race with others (to set starget_sdev_user),
489          * but in most cases, we will be first. Ideally, each LU on the
490          * target would get some limited time or requests on the target.
491          */
492         blk_run_queue(current_sdev->request_queue);
493
494         spin_lock_irqsave(shost->host_lock, flags);
495         if (starget->starget_sdev_user)
496                 goto out;
497         list_for_each_entry_safe(sdev, tmp, &starget->devices,
498                         same_target_siblings) {
499                 if (sdev == current_sdev)
500                         continue;
501                 if (scsi_device_get(sdev))
502                         continue;
503
504                 spin_unlock_irqrestore(shost->host_lock, flags);
505                 blk_run_queue(sdev->request_queue);
506                 spin_lock_irqsave(shost->host_lock, flags);
507         
508                 scsi_device_put(sdev);
509         }
510  out:
511         spin_unlock_irqrestore(shost->host_lock, flags);
512 }
513
514 /*
515  * Function:    scsi_run_queue()
516  *
517  * Purpose:     Select a proper request queue to serve next
518  *
519  * Arguments:   q       - last request's queue
520  *
521  * Returns:     Nothing
522  *
523  * Notes:       The previous command was completely finished, start
524  *              a new one if possible.
525  */
526 static void scsi_run_queue(struct request_queue *q)
527 {
528         struct scsi_device *sdev = q->queuedata;
529         struct Scsi_Host *shost = sdev->host;
530         unsigned long flags;
531
532         if (scsi_target(sdev)->single_lun)
533                 scsi_single_lun_run(sdev);
534
535         spin_lock_irqsave(shost->host_lock, flags);
536         while (!list_empty(&shost->starved_list) &&
537                !shost->host_blocked && !shost->host_self_blocked &&
538                 !((shost->can_queue > 0) &&
539                   (shost->host_busy >= shost->can_queue))) {
540                 /*
541                  * As long as shost is accepting commands and we have
542                  * starved queues, call blk_run_queue. scsi_request_fn
543                  * drops the queue_lock and can add us back to the
544                  * starved_list.
545                  *
546                  * host_lock protects the starved_list and starved_entry.
547                  * scsi_request_fn must get the host_lock before checking
548                  * or modifying starved_list or starved_entry.
549                  */
550                 sdev = list_entry(shost->starved_list.next,
551                                           struct scsi_device, starved_entry);
552                 list_del_init(&sdev->starved_entry);
553                 spin_unlock_irqrestore(shost->host_lock, flags);
554
555
556                 if (test_bit(QUEUE_FLAG_REENTER, &q->queue_flags) &&
557                     !test_and_set_bit(QUEUE_FLAG_REENTER,
558                                       &sdev->request_queue->queue_flags)) {
559                         blk_run_queue(sdev->request_queue);
560                         clear_bit(QUEUE_FLAG_REENTER,
561                                   &sdev->request_queue->queue_flags);
562                 } else
563                         blk_run_queue(sdev->request_queue);
564
565                 spin_lock_irqsave(shost->host_lock, flags);
566                 if (unlikely(!list_empty(&sdev->starved_entry)))
567                         /*
568                          * sdev lost a race, and was put back on the
569                          * starved list. This is unlikely but without this
570                          * in theory we could loop forever.
571                          */
572                         break;
573         }
574         spin_unlock_irqrestore(shost->host_lock, flags);
575
576         blk_run_queue(q);
577 }
578
579 /*
580  * Function:    scsi_requeue_command()
581  *
582  * Purpose:     Handle post-processing of completed commands.
583  *
584  * Arguments:   q       - queue to operate on
585  *              cmd     - command that may need to be requeued.
586  *
587  * Returns:     Nothing
588  *
589  * Notes:       After command completion, there may be blocks left
590  *              over which weren't finished by the previous command
591  *              this can be for a number of reasons - the main one is
592  *              I/O errors in the middle of the request, in which case
593  *              we need to request the blocks that come after the bad
594  *              sector.
595  * Notes:       Upon return, cmd is a stale pointer.
596  */
597 static void scsi_requeue_command(struct request_queue *q, struct scsi_cmnd *cmd)
598 {
599         struct request *req = cmd->request;
600         unsigned long flags;
601
602         scsi_unprep_request(req);
603         spin_lock_irqsave(q->queue_lock, flags);
604         blk_requeue_request(q, req);
605         spin_unlock_irqrestore(q->queue_lock, flags);
606
607         scsi_run_queue(q);
608 }
609
610 void scsi_next_command(struct scsi_cmnd *cmd)
611 {
612         struct scsi_device *sdev = cmd->device;
613         struct request_queue *q = sdev->request_queue;
614
615         /* need to hold a reference on the device before we let go of the cmd */
616         get_device(&sdev->sdev_gendev);
617
618         scsi_put_command(cmd);
619         scsi_run_queue(q);
620
621         /* ok to remove device now */
622         put_device(&sdev->sdev_gendev);
623 }
624
625 void scsi_run_host_queues(struct Scsi_Host *shost)
626 {
627         struct scsi_device *sdev;
628
629         shost_for_each_device(sdev, shost)
630                 scsi_run_queue(sdev->request_queue);
631 }
632
633 /*
634  * Function:    scsi_end_request()
635  *
636  * Purpose:     Post-processing of completed commands (usually invoked at end
637  *              of upper level post-processing and scsi_io_completion).
638  *
639  * Arguments:   cmd      - command that is complete.
640  *              error    - 0 if I/O indicates success, < 0 for I/O error.
641  *              bytes    - number of bytes of completed I/O
642  *              requeue  - indicates whether we should requeue leftovers.
643  *
644  * Lock status: Assumed that lock is not held upon entry.
645  *
646  * Returns:     cmd if requeue required, NULL otherwise.
647  *
648  * Notes:       This is called for block device requests in order to
649  *              mark some number of sectors as complete.
650  * 
651  *              We are guaranteeing that the request queue will be goosed
652  *              at some point during this call.
653  * Notes:       If cmd was requeued, upon return it will be a stale pointer.
654  */
655 static struct scsi_cmnd *scsi_end_request(struct scsi_cmnd *cmd, int error,
656                                           int bytes, int requeue)
657 {
658         struct request_queue *q = cmd->device->request_queue;
659         struct request *req = cmd->request;
660
661         /*
662          * If there are blocks left over at the end, set up the command
663          * to queue the remainder of them.
664          */
665         if (blk_end_request(req, error, bytes)) {
666                 int leftover = (req->hard_nr_sectors << 9);
667
668                 if (blk_pc_request(req))
669                         leftover = req->data_len;
670
671                 /* kill remainder if no retrys */
672                 if (error && blk_noretry_request(req))
673                         blk_end_request(req, error, leftover);
674                 else {
675                         if (requeue) {
676                                 /*
677                                  * Bleah.  Leftovers again.  Stick the
678                                  * leftovers in the front of the
679                                  * queue, and goose the queue again.
680                                  */
681                                 scsi_requeue_command(q, cmd);
682                                 cmd = NULL;
683                         }
684                         return cmd;
685                 }
686         }
687
688         /*
689          * This will goose the queue request function at the end, so we don't
690          * need to worry about launching another command.
691          */
692         scsi_next_command(cmd);
693         return NULL;
694 }
695
696 static inline unsigned int scsi_sgtable_index(unsigned short nents)
697 {
698         unsigned int index;
699
700         BUG_ON(nents > SCSI_MAX_SG_SEGMENTS);
701
702         if (nents <= 8)
703                 index = 0;
704         else
705                 index = get_count_order(nents) - 3;
706
707         return index;
708 }
709
710 static void scsi_sg_free(struct scatterlist *sgl, unsigned int nents)
711 {
712         struct scsi_host_sg_pool *sgp;
713
714         sgp = scsi_sg_pools + scsi_sgtable_index(nents);
715         mempool_free(sgl, sgp->pool);
716 }
717
718 static struct scatterlist *scsi_sg_alloc(unsigned int nents, gfp_t gfp_mask)
719 {
720         struct scsi_host_sg_pool *sgp;
721
722         sgp = scsi_sg_pools + scsi_sgtable_index(nents);
723         return mempool_alloc(sgp->pool, gfp_mask);
724 }
725
726 static int scsi_alloc_sgtable(struct scsi_data_buffer *sdb, int nents,
727                               gfp_t gfp_mask)
728 {
729         int ret;
730
731         BUG_ON(!nents);
732
733         ret = __sg_alloc_table(&sdb->table, nents, SCSI_MAX_SG_SEGMENTS,
734                                gfp_mask, scsi_sg_alloc);
735         if (unlikely(ret))
736                 __sg_free_table(&sdb->table, SCSI_MAX_SG_SEGMENTS,
737                                 scsi_sg_free);
738
739         return ret;
740 }
741
742 static void scsi_free_sgtable(struct scsi_data_buffer *sdb)
743 {
744         __sg_free_table(&sdb->table, SCSI_MAX_SG_SEGMENTS, scsi_sg_free);
745 }
746
747 /*
748  * Function:    scsi_release_buffers()
749  *
750  * Purpose:     Completion processing for block device I/O requests.
751  *
752  * Arguments:   cmd     - command that we are bailing.
753  *
754  * Lock status: Assumed that no lock is held upon entry.
755  *
756  * Returns:     Nothing
757  *
758  * Notes:       In the event that an upper level driver rejects a
759  *              command, we must release resources allocated during
760  *              the __init_io() function.  Primarily this would involve
761  *              the scatter-gather table, and potentially any bounce
762  *              buffers.
763  */
764 void scsi_release_buffers(struct scsi_cmnd *cmd)
765 {
766         if (cmd->sdb.table.nents)
767                 scsi_free_sgtable(&cmd->sdb);
768
769         memset(&cmd->sdb, 0, sizeof(cmd->sdb));
770
771         if (scsi_bidi_cmnd(cmd)) {
772                 struct scsi_data_buffer *bidi_sdb =
773                         cmd->request->next_rq->special;
774                 scsi_free_sgtable(bidi_sdb);
775                 kmem_cache_free(scsi_bidi_sdb_cache, bidi_sdb);
776                 cmd->request->next_rq->special = NULL;
777         }
778 }
779 EXPORT_SYMBOL(scsi_release_buffers);
780
781 /*
782  * Bidi commands Must be complete as a whole, both sides at once.
783  * If part of the bytes were written and lld returned
784  * scsi_in()->resid and/or scsi_out()->resid this information will be left
785  * in req->data_len and req->next_rq->data_len. The upper-layer driver can
786  * decide what to do with this information.
787  */
788 void scsi_end_bidi_request(struct scsi_cmnd *cmd)
789 {
790         struct request *req = cmd->request;
791         unsigned int dlen = req->data_len;
792         unsigned int next_dlen = req->next_rq->data_len;
793
794         req->data_len = scsi_out(cmd)->resid;
795         req->next_rq->data_len = scsi_in(cmd)->resid;
796
797         /* The req and req->next_rq have not been completed */
798         BUG_ON(blk_end_bidi_request(req, 0, dlen, next_dlen));
799
800         scsi_release_buffers(cmd);
801
802         /*
803          * This will goose the queue request function at the end, so we don't
804          * need to worry about launching another command.
805          */
806         scsi_next_command(cmd);
807 }
808
809 /*
810  * Function:    scsi_io_completion()
811  *
812  * Purpose:     Completion processing for block device I/O requests.
813  *
814  * Arguments:   cmd   - command that is finished.
815  *
816  * Lock status: Assumed that no lock is held upon entry.
817  *
818  * Returns:     Nothing
819  *
820  * Notes:       This function is matched in terms of capabilities to
821  *              the function that created the scatter-gather list.
822  *              In other words, if there are no bounce buffers
823  *              (the normal case for most drivers), we don't need
824  *              the logic to deal with cleaning up afterwards.
825  *
826  *              We must do one of several things here:
827  *
828  *              a) Call scsi_end_request.  This will finish off the
829  *                 specified number of sectors.  If we are done, the
830  *                 command block will be released, and the queue
831  *                 function will be goosed.  If we are not done, then
832  *                 scsi_end_request will directly goose the queue.
833  *
834  *              b) We can just use scsi_requeue_command() here.  This would
835  *                 be used if we just wanted to retry, for example.
836  */
837 void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes)
838 {
839         int result = cmd->result;
840         int this_count = scsi_bufflen(cmd);
841         struct request_queue *q = cmd->device->request_queue;
842         struct request *req = cmd->request;
843         int clear_errors = 1;
844         struct scsi_sense_hdr sshdr;
845         int sense_valid = 0;
846         int sense_deferred = 0;
847
848         if (result) {
849                 sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
850                 if (sense_valid)
851                         sense_deferred = scsi_sense_is_deferred(&sshdr);
852         }
853
854         if (blk_pc_request(req)) { /* SG_IO ioctl from block level */
855                 req->errors = result;
856                 if (result) {
857                         clear_errors = 0;
858                         if (sense_valid && req->sense) {
859                                 /*
860                                  * SG_IO wants current and deferred errors
861                                  */
862                                 int len = 8 + cmd->sense_buffer[7];
863
864                                 if (len > SCSI_SENSE_BUFFERSIZE)
865                                         len = SCSI_SENSE_BUFFERSIZE;
866                                 memcpy(req->sense, cmd->sense_buffer,  len);
867                                 req->sense_len = len;
868                         }
869                 }
870                 if (scsi_bidi_cmnd(cmd)) {
871                         /* will also release_buffers */
872                         scsi_end_bidi_request(cmd);
873                         return;
874                 }
875                 req->data_len = scsi_get_resid(cmd);
876         }
877
878         BUG_ON(blk_bidi_rq(req)); /* bidi not support for !blk_pc_request yet */
879         scsi_release_buffers(cmd);
880
881         /*
882          * Next deal with any sectors which we were able to correctly
883          * handle.
884          */
885         SCSI_LOG_HLCOMPLETE(1, printk("%ld sectors total, "
886                                       "%d bytes done.\n",
887                                       req->nr_sectors, good_bytes));
888
889         if (clear_errors)
890                 req->errors = 0;
891
892         /* A number of bytes were successfully read.  If there
893          * are leftovers and there is some kind of error
894          * (result != 0), retry the rest.
895          */
896         if (scsi_end_request(cmd, 0, good_bytes, result == 0) == NULL)
897                 return;
898
899         /* good_bytes = 0, or (inclusive) there were leftovers and
900          * result = 0, so scsi_end_request couldn't retry.
901          */
902         if (sense_valid && !sense_deferred) {
903                 switch (sshdr.sense_key) {
904                 case UNIT_ATTENTION:
905                         if (cmd->device->removable) {
906                                 /* Detected disc change.  Set a bit
907                                  * and quietly refuse further access.
908                                  */
909                                 cmd->device->changed = 1;
910                                 scsi_end_request(cmd, -EIO, this_count, 1);
911                                 return;
912                         } else {
913                                 /* Must have been a power glitch, or a
914                                  * bus reset.  Could not have been a
915                                  * media change, so we just retry the
916                                  * request and see what happens.
917                                  */
918                                 scsi_requeue_command(q, cmd);
919                                 return;
920                         }
921                         break;
922                 case ILLEGAL_REQUEST:
923                         /* If we had an ILLEGAL REQUEST returned, then
924                          * we may have performed an unsupported
925                          * command.  The only thing this should be
926                          * would be a ten byte read where only a six
927                          * byte read was supported.  Also, on a system
928                          * where READ CAPACITY failed, we may have
929                          * read past the end of the disk.
930                          */
931                         if ((cmd->device->use_10_for_rw &&
932                             sshdr.asc == 0x20 && sshdr.ascq == 0x00) &&
933                             (cmd->cmnd[0] == READ_10 ||
934                              cmd->cmnd[0] == WRITE_10)) {
935                                 cmd->device->use_10_for_rw = 0;
936                                 /* This will cause a retry with a
937                                  * 6-byte command.
938                                  */
939                                 scsi_requeue_command(q, cmd);
940                                 return;
941                         } else {
942                                 scsi_end_request(cmd, -EIO, this_count, 1);
943                                 return;
944                         }
945                         break;
946                 case NOT_READY:
947                         /* If the device is in the process of becoming
948                          * ready, or has a temporary blockage, retry.
949                          */
950                         if (sshdr.asc == 0x04) {
951                                 switch (sshdr.ascq) {
952                                 case 0x01: /* becoming ready */
953                                 case 0x04: /* format in progress */
954                                 case 0x05: /* rebuild in progress */
955                                 case 0x06: /* recalculation in progress */
956                                 case 0x07: /* operation in progress */
957                                 case 0x08: /* Long write in progress */
958                                 case 0x09: /* self test in progress */
959                                         scsi_requeue_command(q, cmd);
960                                         return;
961                                 default:
962                                         break;
963                                 }
964                         }
965                         if (!(req->cmd_flags & REQ_QUIET))
966                                 scsi_cmd_print_sense_hdr(cmd,
967                                                          "Device not ready",
968                                                          &sshdr);
969
970                         scsi_end_request(cmd, -EIO, this_count, 1);
971                         return;
972                 case VOLUME_OVERFLOW:
973                         if (!(req->cmd_flags & REQ_QUIET)) {
974                                 scmd_printk(KERN_INFO, cmd,
975                                             "Volume overflow, CDB: ");
976                                 __scsi_print_command(cmd->cmnd);
977                                 scsi_print_sense("", cmd);
978                         }
979                         /* See SSC3rXX or current. */
980                         scsi_end_request(cmd, -EIO, this_count, 1);
981                         return;
982                 default:
983                         break;
984                 }
985         }
986         if (host_byte(result) == DID_RESET) {
987                 /* Third party bus reset or reset for error recovery
988                  * reasons.  Just retry the request and see what
989                  * happens.
990                  */
991                 scsi_requeue_command(q, cmd);
992                 return;
993         }
994         if (result) {
995                 if (!(req->cmd_flags & REQ_QUIET)) {
996                         scsi_print_result(cmd);
997                         if (driver_byte(result) & DRIVER_SENSE)
998                                 scsi_print_sense("", cmd);
999                 }
1000         }
1001         scsi_end_request(cmd, -EIO, this_count, !result);
1002 }
1003
1004 static int scsi_init_sgtable(struct request *req, struct scsi_data_buffer *sdb,
1005                              gfp_t gfp_mask)
1006 {
1007         int count;
1008
1009         /*
1010          * If sg table allocation fails, requeue request later.
1011          */
1012         if (unlikely(scsi_alloc_sgtable(sdb, req->nr_phys_segments,
1013                                         gfp_mask))) {
1014                 return BLKPREP_DEFER;
1015         }
1016
1017         req->buffer = NULL;
1018         if (blk_pc_request(req))
1019                 sdb->length = req->data_len;
1020         else
1021                 sdb->length = req->nr_sectors << 9;
1022
1023         /* 
1024          * Next, walk the list, and fill in the addresses and sizes of
1025          * each segment.
1026          */
1027         count = blk_rq_map_sg(req->q, req, sdb->table.sgl);
1028         BUG_ON(count > sdb->table.nents);
1029         sdb->table.nents = count;
1030         return BLKPREP_OK;
1031 }
1032
1033 /*
1034  * Function:    scsi_init_io()
1035  *
1036  * Purpose:     SCSI I/O initialize function.
1037  *
1038  * Arguments:   cmd   - Command descriptor we wish to initialize
1039  *
1040  * Returns:     0 on success
1041  *              BLKPREP_DEFER if the failure is retryable
1042  *              BLKPREP_KILL if the failure is fatal
1043  */
1044 int scsi_init_io(struct scsi_cmnd *cmd, gfp_t gfp_mask)
1045 {
1046         int error = scsi_init_sgtable(cmd->request, &cmd->sdb, gfp_mask);
1047         if (error)
1048                 goto err_exit;
1049
1050         if (blk_bidi_rq(cmd->request)) {
1051                 struct scsi_data_buffer *bidi_sdb = kmem_cache_zalloc(
1052                         scsi_bidi_sdb_cache, GFP_ATOMIC);
1053                 if (!bidi_sdb) {
1054                         error = BLKPREP_DEFER;
1055                         goto err_exit;
1056                 }
1057
1058                 cmd->request->next_rq->special = bidi_sdb;
1059                 error = scsi_init_sgtable(cmd->request->next_rq, bidi_sdb,
1060                                                                     GFP_ATOMIC);
1061                 if (error)
1062                         goto err_exit;
1063         }
1064
1065         return BLKPREP_OK ;
1066
1067 err_exit:
1068         scsi_release_buffers(cmd);
1069         if (error == BLKPREP_KILL)
1070                 scsi_put_command(cmd);
1071         else /* BLKPREP_DEFER */
1072                 scsi_unprep_request(cmd->request);
1073
1074         return error;
1075 }
1076 EXPORT_SYMBOL(scsi_init_io);
1077
1078 static struct scsi_cmnd *scsi_get_cmd_from_req(struct scsi_device *sdev,
1079                 struct request *req)
1080 {
1081         struct scsi_cmnd *cmd;
1082
1083         if (!req->special) {
1084                 cmd = scsi_get_command(sdev, GFP_ATOMIC);
1085                 if (unlikely(!cmd))
1086                         return NULL;
1087                 req->special = cmd;
1088         } else {
1089                 cmd = req->special;
1090         }
1091
1092         /* pull a tag out of the request if we have one */
1093         cmd->tag = req->tag;
1094         cmd->request = req;
1095
1096         return cmd;
1097 }
1098
1099 int scsi_setup_blk_pc_cmnd(struct scsi_device *sdev, struct request *req)
1100 {
1101         struct scsi_cmnd *cmd;
1102         int ret = scsi_prep_state_check(sdev, req);
1103
1104         if (ret != BLKPREP_OK)
1105                 return ret;
1106
1107         cmd = scsi_get_cmd_from_req(sdev, req);
1108         if (unlikely(!cmd))
1109                 return BLKPREP_DEFER;
1110
1111         /*
1112          * BLOCK_PC requests may transfer data, in which case they must
1113          * a bio attached to them.  Or they might contain a SCSI command
1114          * that does not transfer data, in which case they may optionally
1115          * submit a request without an attached bio.
1116          */
1117         if (req->bio) {
1118                 int ret;
1119
1120                 BUG_ON(!req->nr_phys_segments);
1121
1122                 ret = scsi_init_io(cmd, GFP_ATOMIC);
1123                 if (unlikely(ret))
1124                         return ret;
1125         } else {
1126                 BUG_ON(req->data_len);
1127                 BUG_ON(req->data);
1128
1129                 memset(&cmd->sdb, 0, sizeof(cmd->sdb));
1130                 req->buffer = NULL;
1131         }
1132
1133         BUILD_BUG_ON(sizeof(req->cmd) > sizeof(cmd->cmnd));
1134         memcpy(cmd->cmnd, req->cmd, sizeof(cmd->cmnd));
1135         cmd->cmd_len = req->cmd_len;
1136         if (!req->data_len)
1137                 cmd->sc_data_direction = DMA_NONE;
1138         else if (rq_data_dir(req) == WRITE)
1139                 cmd->sc_data_direction = DMA_TO_DEVICE;
1140         else
1141                 cmd->sc_data_direction = DMA_FROM_DEVICE;
1142         
1143         cmd->transfersize = req->data_len;
1144         cmd->allowed = req->retries;
1145         cmd->timeout_per_command = req->timeout;
1146         return BLKPREP_OK;
1147 }
1148 EXPORT_SYMBOL(scsi_setup_blk_pc_cmnd);
1149
1150 /*
1151  * Setup a REQ_TYPE_FS command.  These are simple read/write request
1152  * from filesystems that still need to be translated to SCSI CDBs from
1153  * the ULD.
1154  */
1155 int scsi_setup_fs_cmnd(struct scsi_device *sdev, struct request *req)
1156 {
1157         struct scsi_cmnd *cmd;
1158         int ret = scsi_prep_state_check(sdev, req);
1159
1160         if (ret != BLKPREP_OK)
1161                 return ret;
1162         /*
1163          * Filesystem requests must transfer data.
1164          */
1165         BUG_ON(!req->nr_phys_segments);
1166
1167         cmd = scsi_get_cmd_from_req(sdev, req);
1168         if (unlikely(!cmd))
1169                 return BLKPREP_DEFER;
1170
1171         return scsi_init_io(cmd, GFP_ATOMIC);
1172 }
1173 EXPORT_SYMBOL(scsi_setup_fs_cmnd);
1174
1175 int scsi_prep_state_check(struct scsi_device *sdev, struct request *req)
1176 {
1177         int ret = BLKPREP_OK;
1178
1179         /*
1180          * If the device is not in running state we will reject some
1181          * or all commands.
1182          */
1183         if (unlikely(sdev->sdev_state != SDEV_RUNNING)) {
1184                 switch (sdev->sdev_state) {
1185                 case SDEV_OFFLINE:
1186                         /*
1187                          * If the device is offline we refuse to process any
1188                          * commands.  The device must be brought online
1189                          * before trying any recovery commands.
1190                          */
1191                         sdev_printk(KERN_ERR, sdev,
1192                                     "rejecting I/O to offline device\n");
1193                         ret = BLKPREP_KILL;
1194                         break;
1195                 case SDEV_DEL:
1196                         /*
1197                          * If the device is fully deleted, we refuse to
1198                          * process any commands as well.
1199                          */
1200                         sdev_printk(KERN_ERR, sdev,
1201                                     "rejecting I/O to dead device\n");
1202                         ret = BLKPREP_KILL;
1203                         break;
1204                 case SDEV_QUIESCE:
1205                 case SDEV_BLOCK:
1206                         /*
1207                          * If the devices is blocked we defer normal commands.
1208                          */
1209                         if (!(req->cmd_flags & REQ_PREEMPT))
1210                                 ret = BLKPREP_DEFER;
1211                         break;
1212                 default:
1213                         /*
1214                          * For any other not fully online state we only allow
1215                          * special commands.  In particular any user initiated
1216                          * command is not allowed.
1217                          */
1218                         if (!(req->cmd_flags & REQ_PREEMPT))
1219                                 ret = BLKPREP_KILL;
1220                         break;
1221                 }
1222         }
1223         return ret;
1224 }
1225 EXPORT_SYMBOL(scsi_prep_state_check);
1226
1227 int scsi_prep_return(struct request_queue *q, struct request *req, int ret)
1228 {
1229         struct scsi_device *sdev = q->queuedata;
1230
1231         switch (ret) {
1232         case BLKPREP_KILL:
1233                 req->errors = DID_NO_CONNECT << 16;
1234                 /* release the command and kill it */
1235                 if (req->special) {
1236                         struct scsi_cmnd *cmd = req->special;
1237                         scsi_release_buffers(cmd);
1238                         scsi_put_command(cmd);
1239                         req->special = NULL;
1240                 }
1241                 break;
1242         case BLKPREP_DEFER:
1243                 /*
1244                  * If we defer, the elv_next_request() returns NULL, but the
1245                  * queue must be restarted, so we plug here if no returning
1246                  * command will automatically do that.
1247                  */
1248                 if (sdev->device_busy == 0)
1249                         blk_plug_device(q);
1250                 break;
1251         default:
1252                 req->cmd_flags |= REQ_DONTPREP;
1253         }
1254
1255         return ret;
1256 }
1257 EXPORT_SYMBOL(scsi_prep_return);
1258
1259 int scsi_prep_fn(struct request_queue *q, struct request *req)
1260 {
1261         struct scsi_device *sdev = q->queuedata;
1262         int ret = BLKPREP_KILL;
1263
1264         if (req->cmd_type == REQ_TYPE_BLOCK_PC)
1265                 ret = scsi_setup_blk_pc_cmnd(sdev, req);
1266         return scsi_prep_return(q, req, ret);
1267 }
1268
1269 /*
1270  * scsi_dev_queue_ready: if we can send requests to sdev, return 1 else
1271  * return 0.
1272  *
1273  * Called with the queue_lock held.
1274  */
1275 static inline int scsi_dev_queue_ready(struct request_queue *q,
1276                                   struct scsi_device *sdev)
1277 {
1278         if (sdev->device_busy >= sdev->queue_depth)
1279                 return 0;
1280         if (sdev->device_busy == 0 && sdev->device_blocked) {
1281                 /*
1282                  * unblock after device_blocked iterates to zero
1283                  */
1284                 if (--sdev->device_blocked == 0) {
1285                         SCSI_LOG_MLQUEUE(3,
1286                                    sdev_printk(KERN_INFO, sdev,
1287                                    "unblocking device at zero depth\n"));
1288                 } else {
1289                         blk_plug_device(q);
1290                         return 0;
1291                 }
1292         }
1293         if (sdev->device_blocked)
1294                 return 0;
1295
1296         return 1;
1297 }
1298
1299 /*
1300  * scsi_host_queue_ready: if we can send requests to shost, return 1 else
1301  * return 0. We must end up running the queue again whenever 0 is
1302  * returned, else IO can hang.
1303  *
1304  * Called with host_lock held.
1305  */
1306 static inline int scsi_host_queue_ready(struct request_queue *q,
1307                                    struct Scsi_Host *shost,
1308                                    struct scsi_device *sdev)
1309 {
1310         if (scsi_host_in_recovery(shost))
1311                 return 0;
1312         if (shost->host_busy == 0 && shost->host_blocked) {
1313                 /*
1314                  * unblock after host_blocked iterates to zero
1315                  */
1316                 if (--shost->host_blocked == 0) {
1317                         SCSI_LOG_MLQUEUE(3,
1318                                 printk("scsi%d unblocking host at zero depth\n",
1319                                         shost->host_no));
1320                 } else {
1321                         blk_plug_device(q);
1322                         return 0;
1323                 }
1324         }
1325         if ((shost->can_queue > 0 && shost->host_busy >= shost->can_queue) ||
1326             shost->host_blocked || shost->host_self_blocked) {
1327                 if (list_empty(&sdev->starved_entry))
1328                         list_add_tail(&sdev->starved_entry, &shost->starved_list);
1329                 return 0;
1330         }
1331
1332         /* We're OK to process the command, so we can't be starved */
1333         if (!list_empty(&sdev->starved_entry))
1334                 list_del_init(&sdev->starved_entry);
1335
1336         return 1;
1337 }
1338
1339 /*
1340  * Kill a request for a dead device
1341  */
1342 static void scsi_kill_request(struct request *req, struct request_queue *q)
1343 {
1344         struct scsi_cmnd *cmd = req->special;
1345         struct scsi_device *sdev = cmd->device;
1346         struct Scsi_Host *shost = sdev->host;
1347
1348         blkdev_dequeue_request(req);
1349
1350         if (unlikely(cmd == NULL)) {
1351                 printk(KERN_CRIT "impossible request in %s.\n",
1352                                  __FUNCTION__);
1353                 BUG();
1354         }
1355
1356         scsi_init_cmd_errh(cmd);
1357         cmd->result = DID_NO_CONNECT << 16;
1358         atomic_inc(&cmd->device->iorequest_cnt);
1359
1360         /*
1361          * SCSI request completion path will do scsi_device_unbusy(),
1362          * bump busy counts.  To bump the counters, we need to dance
1363          * with the locks as normal issue path does.
1364          */
1365         sdev->device_busy++;
1366         spin_unlock(sdev->request_queue->queue_lock);
1367         spin_lock(shost->host_lock);
1368         shost->host_busy++;
1369         spin_unlock(shost->host_lock);
1370         spin_lock(sdev->request_queue->queue_lock);
1371
1372         __scsi_done(cmd);
1373 }
1374
1375 static void scsi_softirq_done(struct request *rq)
1376 {
1377         struct scsi_cmnd *cmd = rq->completion_data;
1378         unsigned long wait_for = (cmd->allowed + 1) * cmd->timeout_per_command;
1379         int disposition;
1380
1381         INIT_LIST_HEAD(&cmd->eh_entry);
1382
1383         disposition = scsi_decide_disposition(cmd);
1384         if (disposition != SUCCESS &&
1385             time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) {
1386                 sdev_printk(KERN_ERR, cmd->device,
1387                             "timing out command, waited %lus\n",
1388                             wait_for/HZ);
1389                 disposition = SUCCESS;
1390         }
1391                         
1392         scsi_log_completion(cmd, disposition);
1393
1394         switch (disposition) {
1395                 case SUCCESS:
1396                         scsi_finish_command(cmd);
1397                         break;
1398                 case NEEDS_RETRY:
1399                         scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY);
1400                         break;
1401                 case ADD_TO_MLQUEUE:
1402                         scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY);
1403                         break;
1404                 default:
1405                         if (!scsi_eh_scmd_add(cmd, 0))
1406                                 scsi_finish_command(cmd);
1407         }
1408 }
1409
1410 /*
1411  * Function:    scsi_request_fn()
1412  *
1413  * Purpose:     Main strategy routine for SCSI.
1414  *
1415  * Arguments:   q       - Pointer to actual queue.
1416  *
1417  * Returns:     Nothing
1418  *
1419  * Lock status: IO request lock assumed to be held when called.
1420  */
1421 static void scsi_request_fn(struct request_queue *q)
1422 {
1423         struct scsi_device *sdev = q->queuedata;
1424         struct Scsi_Host *shost;
1425         struct scsi_cmnd *cmd;
1426         struct request *req;
1427
1428         if (!sdev) {
1429                 printk("scsi: killing requests for dead queue\n");
1430                 while ((req = elv_next_request(q)) != NULL)
1431                         scsi_kill_request(req, q);
1432                 return;
1433         }
1434
1435         if(!get_device(&sdev->sdev_gendev))
1436                 /* We must be tearing the block queue down already */
1437                 return;
1438
1439         /*
1440          * To start with, we keep looping until the queue is empty, or until
1441          * the host is no longer able to accept any more requests.
1442          */
1443         shost = sdev->host;
1444         while (!blk_queue_plugged(q)) {
1445                 int rtn;
1446                 /*
1447                  * get next queueable request.  We do this early to make sure
1448                  * that the request is fully prepared even if we cannot 
1449                  * accept it.
1450                  */
1451                 req = elv_next_request(q);
1452                 if (!req || !scsi_dev_queue_ready(q, sdev))
1453                         break;
1454
1455                 if (unlikely(!scsi_device_online(sdev))) {
1456                         sdev_printk(KERN_ERR, sdev,
1457                                     "rejecting I/O to offline device\n");
1458                         scsi_kill_request(req, q);
1459                         continue;
1460                 }
1461
1462
1463                 /*
1464                  * Remove the request from the request list.
1465                  */
1466                 if (!(blk_queue_tagged(q) && !blk_queue_start_tag(q, req)))
1467                         blkdev_dequeue_request(req);
1468                 sdev->device_busy++;
1469
1470                 spin_unlock(q->queue_lock);
1471                 cmd = req->special;
1472                 if (unlikely(cmd == NULL)) {
1473                         printk(KERN_CRIT "impossible request in %s.\n"
1474                                          "please mail a stack trace to "
1475                                          "linux-scsi@vger.kernel.org\n",
1476                                          __FUNCTION__);
1477                         blk_dump_rq_flags(req, "foo");
1478                         BUG();
1479                 }
1480                 spin_lock(shost->host_lock);
1481
1482                 if (!scsi_host_queue_ready(q, shost, sdev))
1483                         goto not_ready;
1484                 if (scsi_target(sdev)->single_lun) {
1485                         if (scsi_target(sdev)->starget_sdev_user &&
1486                             scsi_target(sdev)->starget_sdev_user != sdev)
1487                                 goto not_ready;
1488                         scsi_target(sdev)->starget_sdev_user = sdev;
1489                 }
1490                 shost->host_busy++;
1491
1492                 /*
1493                  * XXX(hch): This is rather suboptimal, scsi_dispatch_cmd will
1494                  *              take the lock again.
1495                  */
1496                 spin_unlock_irq(shost->host_lock);
1497
1498                 /*
1499                  * Finally, initialize any error handling parameters, and set up
1500                  * the timers for timeouts.
1501                  */
1502                 scsi_init_cmd_errh(cmd);
1503
1504                 /*
1505                  * Dispatch the command to the low-level driver.
1506                  */
1507                 rtn = scsi_dispatch_cmd(cmd);
1508                 spin_lock_irq(q->queue_lock);
1509                 if(rtn) {
1510                         /* we're refusing the command; because of
1511                          * the way locks get dropped, we need to 
1512                          * check here if plugging is required */
1513                         if(sdev->device_busy == 0)
1514                                 blk_plug_device(q);
1515
1516                         break;
1517                 }
1518         }
1519
1520         goto out;
1521
1522  not_ready:
1523         spin_unlock_irq(shost->host_lock);
1524
1525         /*
1526          * lock q, handle tag, requeue req, and decrement device_busy. We
1527          * must return with queue_lock held.
1528          *
1529          * Decrementing device_busy without checking it is OK, as all such
1530          * cases (host limits or settings) should run the queue at some
1531          * later time.
1532          */
1533         spin_lock_irq(q->queue_lock);
1534         blk_requeue_request(q, req);
1535         sdev->device_busy--;
1536         if(sdev->device_busy == 0)
1537                 blk_plug_device(q);
1538  out:
1539         /* must be careful here...if we trigger the ->remove() function
1540          * we cannot be holding the q lock */
1541         spin_unlock_irq(q->queue_lock);
1542         put_device(&sdev->sdev_gendev);
1543         spin_lock_irq(q->queue_lock);
1544 }
1545
1546 u64 scsi_calculate_bounce_limit(struct Scsi_Host *shost)
1547 {
1548         struct device *host_dev;
1549         u64 bounce_limit = 0xffffffff;
1550
1551         if (shost->unchecked_isa_dma)
1552                 return BLK_BOUNCE_ISA;
1553         /*
1554          * Platforms with virtual-DMA translation
1555          * hardware have no practical limit.
1556          */
1557         if (!PCI_DMA_BUS_IS_PHYS)
1558                 return BLK_BOUNCE_ANY;
1559
1560         host_dev = scsi_get_device(shost);
1561         if (host_dev && host_dev->dma_mask)
1562                 bounce_limit = *host_dev->dma_mask;
1563
1564         return bounce_limit;
1565 }
1566 EXPORT_SYMBOL(scsi_calculate_bounce_limit);
1567
1568 struct request_queue *__scsi_alloc_queue(struct Scsi_Host *shost,
1569                                          request_fn_proc *request_fn)
1570 {
1571         struct request_queue *q;
1572         struct device *dev = shost->shost_gendev.parent;
1573
1574         q = blk_init_queue(request_fn, NULL);
1575         if (!q)
1576                 return NULL;
1577
1578         /*
1579          * this limit is imposed by hardware restrictions
1580          */
1581         blk_queue_max_hw_segments(q, shost->sg_tablesize);
1582         blk_queue_max_phys_segments(q, SCSI_MAX_SG_CHAIN_SEGMENTS);
1583
1584         blk_queue_max_sectors(q, shost->max_sectors);
1585         blk_queue_bounce_limit(q, scsi_calculate_bounce_limit(shost));
1586         blk_queue_segment_boundary(q, shost->dma_boundary);
1587
1588         blk_queue_max_segment_size(q, dma_get_max_seg_size(dev));
1589
1590         if (!shost->use_clustering)
1591                 clear_bit(QUEUE_FLAG_CLUSTER, &q->queue_flags);
1592
1593         /*
1594          * set a reasonable default alignment on word boundaries: the
1595          * host and device may alter it using
1596          * blk_queue_update_dma_alignment() later.
1597          */
1598         blk_queue_dma_alignment(q, 0x03);
1599
1600         return q;
1601 }
1602 EXPORT_SYMBOL(__scsi_alloc_queue);
1603
1604 struct request_queue *scsi_alloc_queue(struct scsi_device *sdev)
1605 {
1606         struct request_queue *q;
1607
1608         q = __scsi_alloc_queue(sdev->host, scsi_request_fn);
1609         if (!q)
1610                 return NULL;
1611
1612         blk_queue_prep_rq(q, scsi_prep_fn);
1613         blk_queue_softirq_done(q, scsi_softirq_done);
1614         return q;
1615 }
1616
1617 void scsi_free_queue(struct request_queue *q)
1618 {
1619         blk_cleanup_queue(q);
1620 }
1621
1622 /*
1623  * Function:    scsi_block_requests()
1624  *
1625  * Purpose:     Utility function used by low-level drivers to prevent further
1626  *              commands from being queued to the device.
1627  *
1628  * Arguments:   shost       - Host in question
1629  *
1630  * Returns:     Nothing
1631  *
1632  * Lock status: No locks are assumed held.
1633  *
1634  * Notes:       There is no timer nor any other means by which the requests
1635  *              get unblocked other than the low-level driver calling
1636  *              scsi_unblock_requests().
1637  */
1638 void scsi_block_requests(struct Scsi_Host *shost)
1639 {
1640         shost->host_self_blocked = 1;
1641 }
1642 EXPORT_SYMBOL(scsi_block_requests);
1643
1644 /*
1645  * Function:    scsi_unblock_requests()
1646  *
1647  * Purpose:     Utility function used by low-level drivers to allow further
1648  *              commands from being queued to the device.
1649  *
1650  * Arguments:   shost       - Host in question
1651  *
1652  * Returns:     Nothing
1653  *
1654  * Lock status: No locks are assumed held.
1655  *
1656  * Notes:       There is no timer nor any other means by which the requests
1657  *              get unblocked other than the low-level driver calling
1658  *              scsi_unblock_requests().
1659  *
1660  *              This is done as an API function so that changes to the
1661  *              internals of the scsi mid-layer won't require wholesale
1662  *              changes to drivers that use this feature.
1663  */
1664 void scsi_unblock_requests(struct Scsi_Host *shost)
1665 {
1666         shost->host_self_blocked = 0;
1667         scsi_run_host_queues(shost);
1668 }
1669 EXPORT_SYMBOL(scsi_unblock_requests);
1670
1671 int __init scsi_init_queue(void)
1672 {
1673         int i;
1674
1675         scsi_io_context_cache = kmem_cache_create("scsi_io_context",
1676                                         sizeof(struct scsi_io_context),
1677                                         0, 0, NULL);
1678         if (!scsi_io_context_cache) {
1679                 printk(KERN_ERR "SCSI: can't init scsi io context cache\n");
1680                 return -ENOMEM;
1681         }
1682
1683         scsi_bidi_sdb_cache = kmem_cache_create("scsi_bidi_sdb",
1684                                         sizeof(struct scsi_data_buffer),
1685                                         0, 0, NULL);
1686         if (!scsi_bidi_sdb_cache) {
1687                 printk(KERN_ERR "SCSI: can't init scsi bidi sdb cache\n");
1688                 goto cleanup_io_context;
1689         }
1690
1691         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1692                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1693                 int size = sgp->size * sizeof(struct scatterlist);
1694
1695                 sgp->slab = kmem_cache_create(sgp->name, size, 0,
1696                                 SLAB_HWCACHE_ALIGN, NULL);
1697                 if (!sgp->slab) {
1698                         printk(KERN_ERR "SCSI: can't init sg slab %s\n",
1699                                         sgp->name);
1700                         goto cleanup_bidi_sdb;
1701                 }
1702
1703                 sgp->pool = mempool_create_slab_pool(SG_MEMPOOL_SIZE,
1704                                                      sgp->slab);
1705                 if (!sgp->pool) {
1706                         printk(KERN_ERR "SCSI: can't init sg mempool %s\n",
1707                                         sgp->name);
1708                         goto cleanup_bidi_sdb;
1709                 }
1710         }
1711
1712         return 0;
1713
1714 cleanup_bidi_sdb:
1715         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1716                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1717                 if (sgp->pool)
1718                         mempool_destroy(sgp->pool);
1719                 if (sgp->slab)
1720                         kmem_cache_destroy(sgp->slab);
1721         }
1722         kmem_cache_destroy(scsi_bidi_sdb_cache);
1723 cleanup_io_context:
1724         kmem_cache_destroy(scsi_io_context_cache);
1725
1726         return -ENOMEM;
1727 }
1728
1729 void scsi_exit_queue(void)
1730 {
1731         int i;
1732
1733         kmem_cache_destroy(scsi_io_context_cache);
1734         kmem_cache_destroy(scsi_bidi_sdb_cache);
1735
1736         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1737                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1738                 mempool_destroy(sgp->pool);
1739                 kmem_cache_destroy(sgp->slab);
1740         }
1741 }
1742
1743 /**
1744  *      scsi_mode_select - issue a mode select
1745  *      @sdev:  SCSI device to be queried
1746  *      @pf:    Page format bit (1 == standard, 0 == vendor specific)
1747  *      @sp:    Save page bit (0 == don't save, 1 == save)
1748  *      @modepage: mode page being requested
1749  *      @buffer: request buffer (may not be smaller than eight bytes)
1750  *      @len:   length of request buffer.
1751  *      @timeout: command timeout
1752  *      @retries: number of retries before failing
1753  *      @data: returns a structure abstracting the mode header data
1754  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
1755  *              must be SCSI_SENSE_BUFFERSIZE big.
1756  *
1757  *      Returns zero if successful; negative error number or scsi
1758  *      status on error
1759  *
1760  */
1761 int
1762 scsi_mode_select(struct scsi_device *sdev, int pf, int sp, int modepage,
1763                  unsigned char *buffer, int len, int timeout, int retries,
1764                  struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1765 {
1766         unsigned char cmd[10];
1767         unsigned char *real_buffer;
1768         int ret;
1769
1770         memset(cmd, 0, sizeof(cmd));
1771         cmd[1] = (pf ? 0x10 : 0) | (sp ? 0x01 : 0);
1772
1773         if (sdev->use_10_for_ms) {
1774                 if (len > 65535)
1775                         return -EINVAL;
1776                 real_buffer = kmalloc(8 + len, GFP_KERNEL);
1777                 if (!real_buffer)
1778                         return -ENOMEM;
1779                 memcpy(real_buffer + 8, buffer, len);
1780                 len += 8;
1781                 real_buffer[0] = 0;
1782                 real_buffer[1] = 0;
1783                 real_buffer[2] = data->medium_type;
1784                 real_buffer[3] = data->device_specific;
1785                 real_buffer[4] = data->longlba ? 0x01 : 0;
1786                 real_buffer[5] = 0;
1787                 real_buffer[6] = data->block_descriptor_length >> 8;
1788                 real_buffer[7] = data->block_descriptor_length;
1789
1790                 cmd[0] = MODE_SELECT_10;
1791                 cmd[7] = len >> 8;
1792                 cmd[8] = len;
1793         } else {
1794                 if (len > 255 || data->block_descriptor_length > 255 ||
1795                     data->longlba)
1796                         return -EINVAL;
1797
1798                 real_buffer = kmalloc(4 + len, GFP_KERNEL);
1799                 if (!real_buffer)
1800                         return -ENOMEM;
1801                 memcpy(real_buffer + 4, buffer, len);
1802                 len += 4;
1803                 real_buffer[0] = 0;
1804                 real_buffer[1] = data->medium_type;
1805                 real_buffer[2] = data->device_specific;
1806                 real_buffer[3] = data->block_descriptor_length;
1807                 
1808
1809                 cmd[0] = MODE_SELECT;
1810                 cmd[4] = len;
1811         }
1812
1813         ret = scsi_execute_req(sdev, cmd, DMA_TO_DEVICE, real_buffer, len,
1814                                sshdr, timeout, retries);
1815         kfree(real_buffer);
1816         return ret;
1817 }
1818 EXPORT_SYMBOL_GPL(scsi_mode_select);
1819
1820 /**
1821  *      scsi_mode_sense - issue a mode sense, falling back from 10 to six bytes if necessary.
1822  *      @sdev:  SCSI device to be queried
1823  *      @dbd:   set if mode sense will allow block descriptors to be returned
1824  *      @modepage: mode page being requested
1825  *      @buffer: request buffer (may not be smaller than eight bytes)
1826  *      @len:   length of request buffer.
1827  *      @timeout: command timeout
1828  *      @retries: number of retries before failing
1829  *      @data: returns a structure abstracting the mode header data
1830  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
1831  *              must be SCSI_SENSE_BUFFERSIZE big.
1832  *
1833  *      Returns zero if unsuccessful, or the header offset (either 4
1834  *      or 8 depending on whether a six or ten byte command was
1835  *      issued) if successful.
1836  */
1837 int
1838 scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
1839                   unsigned char *buffer, int len, int timeout, int retries,
1840                   struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1841 {
1842         unsigned char cmd[12];
1843         int use_10_for_ms;
1844         int header_length;
1845         int result;
1846         struct scsi_sense_hdr my_sshdr;
1847
1848         memset(data, 0, sizeof(*data));
1849         memset(&cmd[0], 0, 12);
1850         cmd[1] = dbd & 0x18;    /* allows DBD and LLBA bits */
1851         cmd[2] = modepage;
1852
1853         /* caller might not be interested in sense, but we need it */
1854         if (!sshdr)
1855                 sshdr = &my_sshdr;
1856
1857  retry:
1858         use_10_for_ms = sdev->use_10_for_ms;
1859
1860         if (use_10_for_ms) {
1861                 if (len < 8)
1862                         len = 8;
1863
1864                 cmd[0] = MODE_SENSE_10;
1865                 cmd[8] = len;
1866                 header_length = 8;
1867         } else {
1868                 if (len < 4)
1869                         len = 4;
1870
1871                 cmd[0] = MODE_SENSE;
1872                 cmd[4] = len;
1873                 header_length = 4;
1874         }
1875
1876         memset(buffer, 0, len);
1877
1878         result = scsi_execute_req(sdev, cmd, DMA_FROM_DEVICE, buffer, len,
1879                                   sshdr, timeout, retries);
1880
1881         /* This code looks awful: what it's doing is making sure an
1882          * ILLEGAL REQUEST sense return identifies the actual command
1883          * byte as the problem.  MODE_SENSE commands can return
1884          * ILLEGAL REQUEST if the code page isn't supported */
1885
1886         if (use_10_for_ms && !scsi_status_is_good(result) &&
1887             (driver_byte(result) & DRIVER_SENSE)) {
1888                 if (scsi_sense_valid(sshdr)) {
1889                         if ((sshdr->sense_key == ILLEGAL_REQUEST) &&
1890                             (sshdr->asc == 0x20) && (sshdr->ascq == 0)) {
1891                                 /* 
1892                                  * Invalid command operation code
1893                                  */
1894                                 sdev->use_10_for_ms = 0;
1895                                 goto retry;
1896                         }
1897                 }
1898         }
1899
1900         if(scsi_status_is_good(result)) {
1901                 if (unlikely(buffer[0] == 0x86 && buffer[1] == 0x0b &&
1902                              (modepage == 6 || modepage == 8))) {
1903                         /* Initio breakage? */
1904                         header_length = 0;
1905                         data->length = 13;
1906                         data->medium_type = 0;
1907                         data->device_specific = 0;
1908                         data->longlba = 0;
1909                         data->block_descriptor_length = 0;
1910                 } else if(use_10_for_ms) {
1911                         data->length = buffer[0]*256 + buffer[1] + 2;
1912                         data->medium_type = buffer[2];
1913                         data->device_specific = buffer[3];
1914                         data->longlba = buffer[4] & 0x01;
1915                         data->block_descriptor_length = buffer[6]*256
1916                                 + buffer[7];
1917                 } else {
1918                         data->length = buffer[0] + 1;
1919                         data->medium_type = buffer[1];
1920                         data->device_specific = buffer[2];
1921                         data->block_descriptor_length = buffer[3];
1922                 }
1923                 data->header_length = header_length;
1924         }
1925
1926         return result;
1927 }
1928 EXPORT_SYMBOL(scsi_mode_sense);
1929
1930 /**
1931  *      scsi_test_unit_ready - test if unit is ready
1932  *      @sdev:  scsi device to change the state of.
1933  *      @timeout: command timeout
1934  *      @retries: number of retries before failing
1935  *      @sshdr_external: Optional pointer to struct scsi_sense_hdr for
1936  *              returning sense. Make sure that this is cleared before passing
1937  *              in.
1938  *
1939  *      Returns zero if unsuccessful or an error if TUR failed.  For
1940  *      removable media, a return of NOT_READY or UNIT_ATTENTION is
1941  *      translated to success, with the ->changed flag updated.
1942  **/
1943 int
1944 scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries,
1945                      struct scsi_sense_hdr *sshdr_external)
1946 {
1947         char cmd[] = {
1948                 TEST_UNIT_READY, 0, 0, 0, 0, 0,
1949         };
1950         struct scsi_sense_hdr *sshdr;
1951         int result;
1952
1953         if (!sshdr_external)
1954                 sshdr = kzalloc(sizeof(*sshdr), GFP_KERNEL);
1955         else
1956                 sshdr = sshdr_external;
1957
1958         /* try to eat the UNIT_ATTENTION if there are enough retries */
1959         do {
1960                 result = scsi_execute_req(sdev, cmd, DMA_NONE, NULL, 0, sshdr,
1961                                           timeout, retries);
1962         } while ((driver_byte(result) & DRIVER_SENSE) &&
1963                  sshdr && sshdr->sense_key == UNIT_ATTENTION &&
1964                  --retries);
1965
1966         if (!sshdr)
1967                 /* could not allocate sense buffer, so can't process it */
1968                 return result;
1969
1970         if ((driver_byte(result) & DRIVER_SENSE) && sdev->removable) {
1971
1972                 if ((scsi_sense_valid(sshdr)) &&
1973                     ((sshdr->sense_key == UNIT_ATTENTION) ||
1974                      (sshdr->sense_key == NOT_READY))) {
1975                         sdev->changed = 1;
1976                         result = 0;
1977                 }
1978         }
1979         if (!sshdr_external)
1980                 kfree(sshdr);
1981         return result;
1982 }
1983 EXPORT_SYMBOL(scsi_test_unit_ready);
1984
1985 /**
1986  *      scsi_device_set_state - Take the given device through the device state model.
1987  *      @sdev:  scsi device to change the state of.
1988  *      @state: state to change to.
1989  *
1990  *      Returns zero if unsuccessful or an error if the requested 
1991  *      transition is illegal.
1992  */
1993 int
1994 scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
1995 {
1996         enum scsi_device_state oldstate = sdev->sdev_state;
1997
1998         if (state == oldstate)
1999                 return 0;
2000
2001         switch (state) {
2002         case SDEV_CREATED:
2003                 /* There are no legal states that come back to
2004                  * created.  This is the manually initialised start
2005                  * state */
2006                 goto illegal;
2007                         
2008         case SDEV_RUNNING:
2009                 switch (oldstate) {
2010                 case SDEV_CREATED:
2011                 case SDEV_OFFLINE:
2012                 case SDEV_QUIESCE:
2013                 case SDEV_BLOCK:
2014                         break;
2015                 default:
2016                         goto illegal;
2017                 }
2018                 break;
2019
2020         case SDEV_QUIESCE:
2021                 switch (oldstate) {
2022                 case SDEV_RUNNING:
2023                 case SDEV_OFFLINE:
2024                         break;
2025                 default:
2026                         goto illegal;
2027                 }
2028                 break;
2029
2030         case SDEV_OFFLINE:
2031                 switch (oldstate) {
2032                 case SDEV_CREATED:
2033                 case SDEV_RUNNING:
2034                 case SDEV_QUIESCE:
2035                 case SDEV_BLOCK:
2036                         break;
2037                 default:
2038                         goto illegal;
2039                 }
2040                 break;
2041
2042         case SDEV_BLOCK:
2043                 switch (oldstate) {
2044                 case SDEV_CREATED:
2045                 case SDEV_RUNNING:
2046                         break;
2047                 default:
2048                         goto illegal;
2049                 }
2050                 break;
2051
2052         case SDEV_CANCEL:
2053                 switch (oldstate) {
2054                 case SDEV_CREATED:
2055                 case SDEV_RUNNING:
2056                 case SDEV_QUIESCE:
2057                 case SDEV_OFFLINE:
2058                 case SDEV_BLOCK:
2059                         break;
2060                 default:
2061                         goto illegal;
2062                 }
2063                 break;
2064
2065         case SDEV_DEL:
2066                 switch (oldstate) {
2067                 case SDEV_CREATED:
2068                 case SDEV_RUNNING:
2069                 case SDEV_OFFLINE:
2070                 case SDEV_CANCEL:
2071                         break;
2072                 default:
2073                         goto illegal;
2074                 }
2075                 break;
2076
2077         }
2078         sdev->sdev_state = state;
2079         return 0;
2080
2081  illegal:
2082         SCSI_LOG_ERROR_RECOVERY(1, 
2083                                 sdev_printk(KERN_ERR, sdev,
2084                                             "Illegal state transition %s->%s\n",
2085                                             scsi_device_state_name(oldstate),
2086                                             scsi_device_state_name(state))
2087                                 );
2088         return -EINVAL;
2089 }
2090 EXPORT_SYMBOL(scsi_device_set_state);
2091
2092 /**
2093  *      sdev_evt_emit - emit a single SCSI device uevent
2094  *      @sdev: associated SCSI device
2095  *      @evt: event to emit
2096  *
2097  *      Send a single uevent (scsi_event) to the associated scsi_device.
2098  */
2099 static void scsi_evt_emit(struct scsi_device *sdev, struct scsi_event *evt)
2100 {
2101         int idx = 0;
2102         char *envp[3];
2103
2104         switch (evt->evt_type) {
2105         case SDEV_EVT_MEDIA_CHANGE:
2106                 envp[idx++] = "SDEV_MEDIA_CHANGE=1";
2107                 break;
2108
2109         default:
2110                 /* do nothing */
2111                 break;
2112         }
2113
2114         envp[idx++] = NULL;
2115
2116         kobject_uevent_env(&sdev->sdev_gendev.kobj, KOBJ_CHANGE, envp);
2117 }
2118
2119 /**
2120  *      sdev_evt_thread - send a uevent for each scsi event
2121  *      @work: work struct for scsi_device
2122  *
2123  *      Dispatch queued events to their associated scsi_device kobjects
2124  *      as uevents.
2125  */
2126 void scsi_evt_thread(struct work_struct *work)
2127 {
2128         struct scsi_device *sdev;
2129         LIST_HEAD(event_list);
2130
2131         sdev = container_of(work, struct scsi_device, event_work);
2132
2133         while (1) {
2134                 struct scsi_event *evt;
2135                 struct list_head *this, *tmp;
2136                 unsigned long flags;
2137
2138                 spin_lock_irqsave(&sdev->list_lock, flags);
2139                 list_splice_init(&sdev->event_list, &event_list);
2140                 spin_unlock_irqrestore(&sdev->list_lock, flags);
2141
2142                 if (list_empty(&event_list))
2143                         break;
2144
2145                 list_for_each_safe(this, tmp, &event_list) {
2146                         evt = list_entry(this, struct scsi_event, node);
2147                         list_del(&evt->node);
2148                         scsi_evt_emit(sdev, evt);
2149                         kfree(evt);
2150                 }
2151         }
2152 }
2153
2154 /**
2155  *      sdev_evt_send - send asserted event to uevent thread
2156  *      @sdev: scsi_device event occurred on
2157  *      @evt: event to send
2158  *
2159  *      Assert scsi device event asynchronously.
2160  */
2161 void sdev_evt_send(struct scsi_device *sdev, struct scsi_event *evt)
2162 {
2163         unsigned long flags;
2164
2165         if (!test_bit(evt->evt_type, sdev->supported_events)) {
2166                 kfree(evt);
2167                 return;
2168         }
2169
2170         spin_lock_irqsave(&sdev->list_lock, flags);
2171         list_add_tail(&evt->node, &sdev->event_list);
2172         schedule_work(&sdev->event_work);
2173         spin_unlock_irqrestore(&sdev->list_lock, flags);
2174 }
2175 EXPORT_SYMBOL_GPL(sdev_evt_send);
2176
2177 /**
2178  *      sdev_evt_alloc - allocate a new scsi event
2179  *      @evt_type: type of event to allocate
2180  *      @gfpflags: GFP flags for allocation
2181  *
2182  *      Allocates and returns a new scsi_event.
2183  */
2184 struct scsi_event *sdev_evt_alloc(enum scsi_device_event evt_type,
2185                                   gfp_t gfpflags)
2186 {
2187         struct scsi_event *evt = kzalloc(sizeof(struct scsi_event), gfpflags);
2188         if (!evt)
2189                 return NULL;
2190
2191         evt->evt_type = evt_type;
2192         INIT_LIST_HEAD(&evt->node);
2193
2194         /* evt_type-specific initialization, if any */
2195         switch (evt_type) {
2196         case SDEV_EVT_MEDIA_CHANGE:
2197         default:
2198                 /* do nothing */
2199                 break;
2200         }
2201
2202         return evt;
2203 }
2204 EXPORT_SYMBOL_GPL(sdev_evt_alloc);
2205
2206 /**
2207  *      sdev_evt_send_simple - send asserted event to uevent thread
2208  *      @sdev: scsi_device event occurred on
2209  *      @evt_type: type of event to send
2210  *      @gfpflags: GFP flags for allocation
2211  *
2212  *      Assert scsi device event asynchronously, given an event type.
2213  */
2214 void sdev_evt_send_simple(struct scsi_device *sdev,
2215                           enum scsi_device_event evt_type, gfp_t gfpflags)
2216 {
2217         struct scsi_event *evt = sdev_evt_alloc(evt_type, gfpflags);
2218         if (!evt) {
2219                 sdev_printk(KERN_ERR, sdev, "event %d eaten due to OOM\n",
2220                             evt_type);
2221                 return;
2222         }
2223
2224         sdev_evt_send(sdev, evt);
2225 }
2226 EXPORT_SYMBOL_GPL(sdev_evt_send_simple);
2227
2228 /**
2229  *      scsi_device_quiesce - Block user issued commands.
2230  *      @sdev:  scsi device to quiesce.
2231  *
2232  *      This works by trying to transition to the SDEV_QUIESCE state
2233  *      (which must be a legal transition).  When the device is in this
2234  *      state, only special requests will be accepted, all others will
2235  *      be deferred.  Since special requests may also be requeued requests,
2236  *      a successful return doesn't guarantee the device will be 
2237  *      totally quiescent.
2238  *
2239  *      Must be called with user context, may sleep.
2240  *
2241  *      Returns zero if unsuccessful or an error if not.
2242  */
2243 int
2244 scsi_device_quiesce(struct scsi_device *sdev)
2245 {
2246         int err = scsi_device_set_state(sdev, SDEV_QUIESCE);
2247         if (err)
2248                 return err;
2249
2250         scsi_run_queue(sdev->request_queue);
2251         while (sdev->device_busy) {
2252                 msleep_interruptible(200);
2253                 scsi_run_queue(sdev->request_queue);
2254         }
2255         return 0;
2256 }
2257 EXPORT_SYMBOL(scsi_device_quiesce);
2258
2259 /**
2260  *      scsi_device_resume - Restart user issued commands to a quiesced device.
2261  *      @sdev:  scsi device to resume.
2262  *
2263  *      Moves the device from quiesced back to running and restarts the
2264  *      queues.
2265  *
2266  *      Must be called with user context, may sleep.
2267  */
2268 void
2269 scsi_device_resume(struct scsi_device *sdev)
2270 {
2271         if(scsi_device_set_state(sdev, SDEV_RUNNING))
2272                 return;
2273         scsi_run_queue(sdev->request_queue);
2274 }
2275 EXPORT_SYMBOL(scsi_device_resume);
2276
2277 static void
2278 device_quiesce_fn(struct scsi_device *sdev, void *data)
2279 {
2280         scsi_device_quiesce(sdev);
2281 }
2282
2283 void
2284 scsi_target_quiesce(struct scsi_target *starget)
2285 {
2286         starget_for_each_device(starget, NULL, device_quiesce_fn);
2287 }
2288 EXPORT_SYMBOL(scsi_target_quiesce);
2289
2290 static void
2291 device_resume_fn(struct scsi_device *sdev, void *data)
2292 {
2293         scsi_device_resume(sdev);
2294 }
2295
2296 void
2297 scsi_target_resume(struct scsi_target *starget)
2298 {
2299         starget_for_each_device(starget, NULL, device_resume_fn);
2300 }
2301 EXPORT_SYMBOL(scsi_target_resume);
2302
2303 /**
2304  * scsi_internal_device_block - internal function to put a device temporarily into the SDEV_BLOCK state
2305  * @sdev:       device to block
2306  *
2307  * Block request made by scsi lld's to temporarily stop all
2308  * scsi commands on the specified device.  Called from interrupt
2309  * or normal process context.
2310  *
2311  * Returns zero if successful or error if not
2312  *
2313  * Notes:       
2314  *      This routine transitions the device to the SDEV_BLOCK state
2315  *      (which must be a legal transition).  When the device is in this
2316  *      state, all commands are deferred until the scsi lld reenables
2317  *      the device with scsi_device_unblock or device_block_tmo fires.
2318  *      This routine assumes the host_lock is held on entry.
2319  */
2320 int
2321 scsi_internal_device_block(struct scsi_device *sdev)
2322 {
2323         struct request_queue *q = sdev->request_queue;
2324         unsigned long flags;
2325         int err = 0;
2326
2327         err = scsi_device_set_state(sdev, SDEV_BLOCK);
2328         if (err)
2329                 return err;
2330
2331         /* 
2332          * The device has transitioned to SDEV_BLOCK.  Stop the
2333          * block layer from calling the midlayer with this device's
2334          * request queue. 
2335          */
2336         spin_lock_irqsave(q->queue_lock, flags);
2337         blk_stop_queue(q);
2338         spin_unlock_irqrestore(q->queue_lock, flags);
2339
2340         return 0;
2341 }
2342 EXPORT_SYMBOL_GPL(scsi_internal_device_block);
2343  
2344 /**
2345  * scsi_internal_device_unblock - resume a device after a block request
2346  * @sdev:       device to resume
2347  *
2348  * Called by scsi lld's or the midlayer to restart the device queue
2349  * for the previously suspended scsi device.  Called from interrupt or
2350  * normal process context.
2351  *
2352  * Returns zero if successful or error if not.
2353  *
2354  * Notes:       
2355  *      This routine transitions the device to the SDEV_RUNNING state
2356  *      (which must be a legal transition) allowing the midlayer to
2357  *      goose the queue for this device.  This routine assumes the 
2358  *      host_lock is held upon entry.
2359  */
2360 int
2361 scsi_internal_device_unblock(struct scsi_device *sdev)
2362 {
2363         struct request_queue *q = sdev->request_queue; 
2364         int err;
2365         unsigned long flags;
2366         
2367         /* 
2368          * Try to transition the scsi device to SDEV_RUNNING
2369          * and goose the device queue if successful.  
2370          */
2371         err = scsi_device_set_state(sdev, SDEV_RUNNING);
2372         if (err)
2373                 return err;
2374
2375         spin_lock_irqsave(q->queue_lock, flags);
2376         blk_start_queue(q);
2377         spin_unlock_irqrestore(q->queue_lock, flags);
2378
2379         return 0;
2380 }
2381 EXPORT_SYMBOL_GPL(scsi_internal_device_unblock);
2382
2383 static void
2384 device_block(struct scsi_device *sdev, void *data)
2385 {
2386         scsi_internal_device_block(sdev);
2387 }
2388
2389 static int
2390 target_block(struct device *dev, void *data)
2391 {
2392         if (scsi_is_target_device(dev))
2393                 starget_for_each_device(to_scsi_target(dev), NULL,
2394                                         device_block);
2395         return 0;
2396 }
2397
2398 void
2399 scsi_target_block(struct device *dev)
2400 {
2401         if (scsi_is_target_device(dev))
2402                 starget_for_each_device(to_scsi_target(dev), NULL,
2403                                         device_block);
2404         else
2405                 device_for_each_child(dev, NULL, target_block);
2406 }
2407 EXPORT_SYMBOL_GPL(scsi_target_block);
2408
2409 static void
2410 device_unblock(struct scsi_device *sdev, void *data)
2411 {
2412         scsi_internal_device_unblock(sdev);
2413 }
2414
2415 static int
2416 target_unblock(struct device *dev, void *data)
2417 {
2418         if (scsi_is_target_device(dev))
2419                 starget_for_each_device(to_scsi_target(dev), NULL,
2420                                         device_unblock);
2421         return 0;
2422 }
2423
2424 void
2425 scsi_target_unblock(struct device *dev)
2426 {
2427         if (scsi_is_target_device(dev))
2428                 starget_for_each_device(to_scsi_target(dev), NULL,
2429                                         device_unblock);
2430         else
2431                 device_for_each_child(dev, NULL, target_unblock);
2432 }
2433 EXPORT_SYMBOL_GPL(scsi_target_unblock);
2434
2435 /**
2436  * scsi_kmap_atomic_sg - find and atomically map an sg-elemnt
2437  * @sgl:        scatter-gather list
2438  * @sg_count:   number of segments in sg
2439  * @offset:     offset in bytes into sg, on return offset into the mapped area
2440  * @len:        bytes to map, on return number of bytes mapped
2441  *
2442  * Returns virtual address of the start of the mapped page
2443  */
2444 void *scsi_kmap_atomic_sg(struct scatterlist *sgl, int sg_count,
2445                           size_t *offset, size_t *len)
2446 {
2447         int i;
2448         size_t sg_len = 0, len_complete = 0;
2449         struct scatterlist *sg;
2450         struct page *page;
2451
2452         WARN_ON(!irqs_disabled());
2453
2454         for_each_sg(sgl, sg, sg_count, i) {
2455                 len_complete = sg_len; /* Complete sg-entries */
2456                 sg_len += sg->length;
2457                 if (sg_len > *offset)
2458                         break;
2459         }
2460
2461         if (unlikely(i == sg_count)) {
2462                 printk(KERN_ERR "%s: Bytes in sg: %zu, requested offset %zu, "
2463                         "elements %d\n",
2464                        __FUNCTION__, sg_len, *offset, sg_count);
2465                 WARN_ON(1);
2466                 return NULL;
2467         }
2468
2469         /* Offset starting from the beginning of first page in this sg-entry */
2470         *offset = *offset - len_complete + sg->offset;
2471
2472         /* Assumption: contiguous pages can be accessed as "page + i" */
2473         page = nth_page(sg_page(sg), (*offset >> PAGE_SHIFT));
2474         *offset &= ~PAGE_MASK;
2475
2476         /* Bytes in this sg-entry from *offset to the end of the page */
2477         sg_len = PAGE_SIZE - *offset;
2478         if (*len > sg_len)
2479                 *len = sg_len;
2480
2481         return kmap_atomic(page, KM_BIO_SRC_IRQ);
2482 }
2483 EXPORT_SYMBOL(scsi_kmap_atomic_sg);
2484
2485 /**
2486  * scsi_kunmap_atomic_sg - atomically unmap a virtual address, previously mapped with scsi_kmap_atomic_sg
2487  * @virt:       virtual address to be unmapped
2488  */
2489 void scsi_kunmap_atomic_sg(void *virt)
2490 {
2491         kunmap_atomic(virt, KM_BIO_SRC_IRQ);
2492 }
2493 EXPORT_SYMBOL(scsi_kunmap_atomic_sg);