[SCSI] scsi_dh: add infrastructure for SCSI Device Handlers
[linux-2.6.git] / drivers / scsi / scsi_lib.c
1 /*
2  *  scsi_lib.c Copyright (C) 1999 Eric Youngdale
3  *
4  *  SCSI queueing library.
5  *      Initial versions: Eric Youngdale (eric@andante.org).
6  *                        Based upon conversations with large numbers
7  *                        of people at Linux Expo.
8  */
9
10 #include <linux/bio.h>
11 #include <linux/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
305                 while (len > 0 && data_len > 0) {
306                         /*
307                          * sg sends a scatterlist that is larger than
308                          * the data_len it wants transferred for certain
309                          * IO sizes
310                          */
311                         bytes = min_t(unsigned int, len, PAGE_SIZE - off);
312                         bytes = min(bytes, data_len);
313
314                         if (!bio) {
315                                 nr_vecs = min_t(int, BIO_MAX_PAGES, nr_pages);
316                                 nr_pages -= nr_vecs;
317
318                                 bio = bio_alloc(gfp, nr_vecs);
319                                 if (!bio) {
320                                         err = -ENOMEM;
321                                         goto free_bios;
322                                 }
323                                 bio->bi_end_io = scsi_bi_endio;
324                         }
325
326                         if (bio_add_pc_page(q, bio, page, bytes, off) !=
327                             bytes) {
328                                 bio_put(bio);
329                                 err = -EINVAL;
330                                 goto free_bios;
331                         }
332
333                         if (bio->bi_vcnt >= nr_vecs) {
334                                 err = scsi_merge_bio(rq, bio);
335                                 if (err) {
336                                         bio_endio(bio, 0);
337                                         goto free_bios;
338                                 }
339                                 bio = NULL;
340                         }
341
342                         page++;
343                         len -= bytes;
344                         data_len -=bytes;
345                         off = 0;
346                 }
347         }
348
349         rq->buffer = rq->data = NULL;
350         rq->data_len = bufflen;
351         return 0;
352
353 free_bios:
354         while ((bio = rq->bio) != NULL) {
355                 rq->bio = bio->bi_next;
356                 /*
357                  * call endio instead of bio_put incase it was bounced
358                  */
359                 bio_endio(bio, 0);
360         }
361
362         return err;
363 }
364
365 /**
366  * scsi_execute_async - insert request
367  * @sdev:       scsi device
368  * @cmd:        scsi command
369  * @cmd_len:    length of scsi cdb
370  * @data_direction: DMA_TO_DEVICE, DMA_FROM_DEVICE, or DMA_NONE
371  * @buffer:     data buffer (this can be a kernel buffer or scatterlist)
372  * @bufflen:    len of buffer
373  * @use_sg:     if buffer is a scatterlist this is the number of elements
374  * @timeout:    request timeout in seconds
375  * @retries:    number of times to retry request
376  * @privdata:   data passed to done()
377  * @done:       callback function when done
378  * @gfp:        memory allocation flags
379  */
380 int scsi_execute_async(struct scsi_device *sdev, const unsigned char *cmd,
381                        int cmd_len, int data_direction, void *buffer, unsigned bufflen,
382                        int use_sg, int timeout, int retries, void *privdata,
383                        void (*done)(void *, char *, int, int), gfp_t gfp)
384 {
385         struct request *req;
386         struct scsi_io_context *sioc;
387         int err = 0;
388         int write = (data_direction == DMA_TO_DEVICE);
389
390         sioc = kmem_cache_zalloc(scsi_io_context_cache, gfp);
391         if (!sioc)
392                 return DRIVER_ERROR << 24;
393
394         req = blk_get_request(sdev->request_queue, write, gfp);
395         if (!req)
396                 goto free_sense;
397         req->cmd_type = REQ_TYPE_BLOCK_PC;
398         req->cmd_flags |= REQ_QUIET;
399
400         if (use_sg)
401                 err = scsi_req_map_sg(req, buffer, use_sg, bufflen, gfp);
402         else if (bufflen)
403                 err = blk_rq_map_kern(req->q, req, buffer, bufflen, gfp);
404
405         if (err)
406                 goto free_req;
407
408         req->cmd_len = cmd_len;
409         memset(req->cmd, 0, BLK_MAX_CDB); /* ATAPI hates garbage after CDB */
410         memcpy(req->cmd, cmd, req->cmd_len);
411         req->sense = sioc->sense;
412         req->sense_len = 0;
413         req->timeout = timeout;
414         req->retries = retries;
415         req->end_io_data = sioc;
416
417         sioc->data = privdata;
418         sioc->done = done;
419
420         blk_execute_rq_nowait(req->q, NULL, req, 1, scsi_end_async);
421         return 0;
422
423 free_req:
424         blk_put_request(req);
425 free_sense:
426         kmem_cache_free(scsi_io_context_cache, sioc);
427         return DRIVER_ERROR << 24;
428 }
429 EXPORT_SYMBOL_GPL(scsi_execute_async);
430
431 /*
432  * Function:    scsi_init_cmd_errh()
433  *
434  * Purpose:     Initialize cmd fields related to error handling.
435  *
436  * Arguments:   cmd     - command that is ready to be queued.
437  *
438  * Notes:       This function has the job of initializing a number of
439  *              fields related to error handling.   Typically this will
440  *              be called once for each command, as required.
441  */
442 static void scsi_init_cmd_errh(struct scsi_cmnd *cmd)
443 {
444         cmd->serial_number = 0;
445         scsi_set_resid(cmd, 0);
446         memset(cmd->sense_buffer, 0, SCSI_SENSE_BUFFERSIZE);
447         if (cmd->cmd_len == 0)
448                 cmd->cmd_len = scsi_command_size(cmd->cmnd);
449 }
450
451 void scsi_device_unbusy(struct scsi_device *sdev)
452 {
453         struct Scsi_Host *shost = sdev->host;
454         unsigned long flags;
455
456         spin_lock_irqsave(shost->host_lock, flags);
457         shost->host_busy--;
458         if (unlikely(scsi_host_in_recovery(shost) &&
459                      (shost->host_failed || shost->host_eh_scheduled)))
460                 scsi_eh_wakeup(shost);
461         spin_unlock(shost->host_lock);
462         spin_lock(sdev->request_queue->queue_lock);
463         sdev->device_busy--;
464         spin_unlock_irqrestore(sdev->request_queue->queue_lock, flags);
465 }
466
467 /*
468  * Called for single_lun devices on IO completion. Clear starget_sdev_user,
469  * and call blk_run_queue for all the scsi_devices on the target -
470  * including current_sdev first.
471  *
472  * Called with *no* scsi locks held.
473  */
474 static void scsi_single_lun_run(struct scsi_device *current_sdev)
475 {
476         struct Scsi_Host *shost = current_sdev->host;
477         struct scsi_device *sdev, *tmp;
478         struct scsi_target *starget = scsi_target(current_sdev);
479         unsigned long flags;
480
481         spin_lock_irqsave(shost->host_lock, flags);
482         starget->starget_sdev_user = NULL;
483         spin_unlock_irqrestore(shost->host_lock, flags);
484
485         /*
486          * Call blk_run_queue for all LUNs on the target, starting with
487          * current_sdev. We race with others (to set starget_sdev_user),
488          * but in most cases, we will be first. Ideally, each LU on the
489          * target would get some limited time or requests on the target.
490          */
491         blk_run_queue(current_sdev->request_queue);
492
493         spin_lock_irqsave(shost->host_lock, flags);
494         if (starget->starget_sdev_user)
495                 goto out;
496         list_for_each_entry_safe(sdev, tmp, &starget->devices,
497                         same_target_siblings) {
498                 if (sdev == current_sdev)
499                         continue;
500                 if (scsi_device_get(sdev))
501                         continue;
502
503                 spin_unlock_irqrestore(shost->host_lock, flags);
504                 blk_run_queue(sdev->request_queue);
505                 spin_lock_irqsave(shost->host_lock, flags);
506         
507                 scsi_device_put(sdev);
508         }
509  out:
510         spin_unlock_irqrestore(shost->host_lock, flags);
511 }
512
513 /*
514  * Function:    scsi_run_queue()
515  *
516  * Purpose:     Select a proper request queue to serve next
517  *
518  * Arguments:   q       - last request's queue
519  *
520  * Returns:     Nothing
521  *
522  * Notes:       The previous command was completely finished, start
523  *              a new one if possible.
524  */
525 static void scsi_run_queue(struct request_queue *q)
526 {
527         struct scsi_device *sdev = q->queuedata;
528         struct Scsi_Host *shost = sdev->host;
529         unsigned long flags;
530
531         if (scsi_target(sdev)->single_lun)
532                 scsi_single_lun_run(sdev);
533
534         spin_lock_irqsave(shost->host_lock, flags);
535         while (!list_empty(&shost->starved_list) &&
536                !shost->host_blocked && !shost->host_self_blocked &&
537                 !((shost->can_queue > 0) &&
538                   (shost->host_busy >= shost->can_queue))) {
539
540                 int flagset;
541
542                 /*
543                  * As long as shost is accepting commands and we have
544                  * starved queues, call blk_run_queue. scsi_request_fn
545                  * drops the queue_lock and can add us back to the
546                  * starved_list.
547                  *
548                  * host_lock protects the starved_list and starved_entry.
549                  * scsi_request_fn must get the host_lock before checking
550                  * or modifying starved_list or starved_entry.
551                  */
552                 sdev = list_entry(shost->starved_list.next,
553                                           struct scsi_device, starved_entry);
554                 list_del_init(&sdev->starved_entry);
555                 spin_unlock(shost->host_lock);
556
557                 spin_lock(sdev->request_queue->queue_lock);
558                 flagset = test_bit(QUEUE_FLAG_REENTER, &q->queue_flags) &&
559                                 !test_bit(QUEUE_FLAG_REENTER,
560                                         &sdev->request_queue->queue_flags);
561                 if (flagset)
562                         queue_flag_set(QUEUE_FLAG_REENTER, sdev->request_queue);
563                 __blk_run_queue(sdev->request_queue);
564                 if (flagset)
565                         queue_flag_clear(QUEUE_FLAG_REENTER, sdev->request_queue);
566                 spin_unlock(sdev->request_queue->queue_lock);
567
568                 spin_lock(shost->host_lock);
569                 if (unlikely(!list_empty(&sdev->starved_entry)))
570                         /*
571                          * sdev lost a race, and was put back on the
572                          * starved list. This is unlikely but without this
573                          * in theory we could loop forever.
574                          */
575                         break;
576         }
577         spin_unlock_irqrestore(shost->host_lock, flags);
578
579         blk_run_queue(q);
580 }
581
582 /*
583  * Function:    scsi_requeue_command()
584  *
585  * Purpose:     Handle post-processing of completed commands.
586  *
587  * Arguments:   q       - queue to operate on
588  *              cmd     - command that may need to be requeued.
589  *
590  * Returns:     Nothing
591  *
592  * Notes:       After command completion, there may be blocks left
593  *              over which weren't finished by the previous command
594  *              this can be for a number of reasons - the main one is
595  *              I/O errors in the middle of the request, in which case
596  *              we need to request the blocks that come after the bad
597  *              sector.
598  * Notes:       Upon return, cmd is a stale pointer.
599  */
600 static void scsi_requeue_command(struct request_queue *q, struct scsi_cmnd *cmd)
601 {
602         struct request *req = cmd->request;
603         unsigned long flags;
604
605         scsi_unprep_request(req);
606         spin_lock_irqsave(q->queue_lock, flags);
607         blk_requeue_request(q, req);
608         spin_unlock_irqrestore(q->queue_lock, flags);
609
610         scsi_run_queue(q);
611 }
612
613 void scsi_next_command(struct scsi_cmnd *cmd)
614 {
615         struct scsi_device *sdev = cmd->device;
616         struct request_queue *q = sdev->request_queue;
617
618         /* need to hold a reference on the device before we let go of the cmd */
619         get_device(&sdev->sdev_gendev);
620
621         scsi_put_command(cmd);
622         scsi_run_queue(q);
623
624         /* ok to remove device now */
625         put_device(&sdev->sdev_gendev);
626 }
627
628 void scsi_run_host_queues(struct Scsi_Host *shost)
629 {
630         struct scsi_device *sdev;
631
632         shost_for_each_device(sdev, shost)
633                 scsi_run_queue(sdev->request_queue);
634 }
635
636 /*
637  * Function:    scsi_end_request()
638  *
639  * Purpose:     Post-processing of completed commands (usually invoked at end
640  *              of upper level post-processing and scsi_io_completion).
641  *
642  * Arguments:   cmd      - command that is complete.
643  *              error    - 0 if I/O indicates success, < 0 for I/O error.
644  *              bytes    - number of bytes of completed I/O
645  *              requeue  - indicates whether we should requeue leftovers.
646  *
647  * Lock status: Assumed that lock is not held upon entry.
648  *
649  * Returns:     cmd if requeue required, NULL otherwise.
650  *
651  * Notes:       This is called for block device requests in order to
652  *              mark some number of sectors as complete.
653  * 
654  *              We are guaranteeing that the request queue will be goosed
655  *              at some point during this call.
656  * Notes:       If cmd was requeued, upon return it will be a stale pointer.
657  */
658 static struct scsi_cmnd *scsi_end_request(struct scsi_cmnd *cmd, int error,
659                                           int bytes, int requeue)
660 {
661         struct request_queue *q = cmd->device->request_queue;
662         struct request *req = cmd->request;
663
664         /*
665          * If there are blocks left over at the end, set up the command
666          * to queue the remainder of them.
667          */
668         if (blk_end_request(req, error, bytes)) {
669                 int leftover = (req->hard_nr_sectors << 9);
670
671                 if (blk_pc_request(req))
672                         leftover = req->data_len;
673
674                 /* kill remainder if no retrys */
675                 if (error && blk_noretry_request(req))
676                         blk_end_request(req, error, leftover);
677                 else {
678                         if (requeue) {
679                                 /*
680                                  * Bleah.  Leftovers again.  Stick the
681                                  * leftovers in the front of the
682                                  * queue, and goose the queue again.
683                                  */
684                                 scsi_requeue_command(q, cmd);
685                                 cmd = NULL;
686                         }
687                         return cmd;
688                 }
689         }
690
691         /*
692          * This will goose the queue request function at the end, so we don't
693          * need to worry about launching another command.
694          */
695         scsi_next_command(cmd);
696         return NULL;
697 }
698
699 static inline unsigned int scsi_sgtable_index(unsigned short nents)
700 {
701         unsigned int index;
702
703         BUG_ON(nents > SCSI_MAX_SG_SEGMENTS);
704
705         if (nents <= 8)
706                 index = 0;
707         else
708                 index = get_count_order(nents) - 3;
709
710         return index;
711 }
712
713 static void scsi_sg_free(struct scatterlist *sgl, unsigned int nents)
714 {
715         struct scsi_host_sg_pool *sgp;
716
717         sgp = scsi_sg_pools + scsi_sgtable_index(nents);
718         mempool_free(sgl, sgp->pool);
719 }
720
721 static struct scatterlist *scsi_sg_alloc(unsigned int nents, gfp_t gfp_mask)
722 {
723         struct scsi_host_sg_pool *sgp;
724
725         sgp = scsi_sg_pools + scsi_sgtable_index(nents);
726         return mempool_alloc(sgp->pool, gfp_mask);
727 }
728
729 static int scsi_alloc_sgtable(struct scsi_data_buffer *sdb, int nents,
730                               gfp_t gfp_mask)
731 {
732         int ret;
733
734         BUG_ON(!nents);
735
736         ret = __sg_alloc_table(&sdb->table, nents, SCSI_MAX_SG_SEGMENTS,
737                                gfp_mask, scsi_sg_alloc);
738         if (unlikely(ret))
739                 __sg_free_table(&sdb->table, SCSI_MAX_SG_SEGMENTS,
740                                 scsi_sg_free);
741
742         return ret;
743 }
744
745 static void scsi_free_sgtable(struct scsi_data_buffer *sdb)
746 {
747         __sg_free_table(&sdb->table, SCSI_MAX_SG_SEGMENTS, scsi_sg_free);
748 }
749
750 /*
751  * Function:    scsi_release_buffers()
752  *
753  * Purpose:     Completion processing for block device I/O requests.
754  *
755  * Arguments:   cmd     - command that we are bailing.
756  *
757  * Lock status: Assumed that no lock is held upon entry.
758  *
759  * Returns:     Nothing
760  *
761  * Notes:       In the event that an upper level driver rejects a
762  *              command, we must release resources allocated during
763  *              the __init_io() function.  Primarily this would involve
764  *              the scatter-gather table, and potentially any bounce
765  *              buffers.
766  */
767 void scsi_release_buffers(struct scsi_cmnd *cmd)
768 {
769         if (cmd->sdb.table.nents)
770                 scsi_free_sgtable(&cmd->sdb);
771
772         memset(&cmd->sdb, 0, sizeof(cmd->sdb));
773
774         if (scsi_bidi_cmnd(cmd)) {
775                 struct scsi_data_buffer *bidi_sdb =
776                         cmd->request->next_rq->special;
777                 scsi_free_sgtable(bidi_sdb);
778                 kmem_cache_free(scsi_bidi_sdb_cache, bidi_sdb);
779                 cmd->request->next_rq->special = NULL;
780         }
781 }
782 EXPORT_SYMBOL(scsi_release_buffers);
783
784 /*
785  * Bidi commands Must be complete as a whole, both sides at once.
786  * If part of the bytes were written and lld returned
787  * scsi_in()->resid and/or scsi_out()->resid this information will be left
788  * in req->data_len and req->next_rq->data_len. The upper-layer driver can
789  * decide what to do with this information.
790  */
791 static void scsi_end_bidi_request(struct scsi_cmnd *cmd)
792 {
793         struct request *req = cmd->request;
794         unsigned int dlen = req->data_len;
795         unsigned int next_dlen = req->next_rq->data_len;
796
797         req->data_len = scsi_out(cmd)->resid;
798         req->next_rq->data_len = scsi_in(cmd)->resid;
799
800         /* The req and req->next_rq have not been completed */
801         BUG_ON(blk_end_bidi_request(req, 0, dlen, next_dlen));
802
803         scsi_release_buffers(cmd);
804
805         /*
806          * This will goose the queue request function at the end, so we don't
807          * need to worry about launching another command.
808          */
809         scsi_next_command(cmd);
810 }
811
812 /*
813  * Function:    scsi_io_completion()
814  *
815  * Purpose:     Completion processing for block device I/O requests.
816  *
817  * Arguments:   cmd   - command that is finished.
818  *
819  * Lock status: Assumed that no lock is held upon entry.
820  *
821  * Returns:     Nothing
822  *
823  * Notes:       This function is matched in terms of capabilities to
824  *              the function that created the scatter-gather list.
825  *              In other words, if there are no bounce buffers
826  *              (the normal case for most drivers), we don't need
827  *              the logic to deal with cleaning up afterwards.
828  *
829  *              We must do one of several things here:
830  *
831  *              a) Call scsi_end_request.  This will finish off the
832  *                 specified number of sectors.  If we are done, the
833  *                 command block will be released, and the queue
834  *                 function will be goosed.  If we are not done, then
835  *                 scsi_end_request will directly goose the queue.
836  *
837  *              b) We can just use scsi_requeue_command() here.  This would
838  *                 be used if we just wanted to retry, for example.
839  */
840 void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes)
841 {
842         int result = cmd->result;
843         int this_count = scsi_bufflen(cmd);
844         struct request_queue *q = cmd->device->request_queue;
845         struct request *req = cmd->request;
846         int error = 0;
847         struct scsi_sense_hdr sshdr;
848         int sense_valid = 0;
849         int sense_deferred = 0;
850
851         if (result) {
852                 sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
853                 if (sense_valid)
854                         sense_deferred = scsi_sense_is_deferred(&sshdr);
855         }
856
857         if (blk_pc_request(req)) { /* SG_IO ioctl from block level */
858                 req->errors = result;
859                 if (result) {
860                         if (sense_valid && req->sense) {
861                                 /*
862                                  * SG_IO wants current and deferred errors
863                                  */
864                                 int len = 8 + cmd->sense_buffer[7];
865
866                                 if (len > SCSI_SENSE_BUFFERSIZE)
867                                         len = SCSI_SENSE_BUFFERSIZE;
868                                 memcpy(req->sense, cmd->sense_buffer,  len);
869                                 req->sense_len = len;
870                         }
871                         if (!sense_deferred)
872                                 error = -EIO;
873                 }
874                 if (scsi_bidi_cmnd(cmd)) {
875                         /* will also release_buffers */
876                         scsi_end_bidi_request(cmd);
877                         return;
878                 }
879                 req->data_len = scsi_get_resid(cmd);
880         }
881
882         BUG_ON(blk_bidi_rq(req)); /* bidi not support for !blk_pc_request yet */
883         scsi_release_buffers(cmd);
884
885         /*
886          * Next deal with any sectors which we were able to correctly
887          * handle.
888          */
889         SCSI_LOG_HLCOMPLETE(1, printk("%ld sectors total, "
890                                       "%d bytes done.\n",
891                                       req->nr_sectors, good_bytes));
892
893         /* A number of bytes were successfully read.  If there
894          * are leftovers and there is some kind of error
895          * (result != 0), retry the rest.
896          */
897         if (scsi_end_request(cmd, error, good_bytes, result == 0) == NULL)
898                 return;
899
900         /* good_bytes = 0, or (inclusive) there were leftovers and
901          * result = 0, so scsi_end_request couldn't retry.
902          */
903         if (sense_valid && !sense_deferred) {
904                 switch (sshdr.sense_key) {
905                 case UNIT_ATTENTION:
906                         if (cmd->device->removable) {
907                                 /* Detected disc change.  Set a bit
908                                  * and quietly refuse further access.
909                                  */
910                                 cmd->device->changed = 1;
911                                 scsi_end_request(cmd, -EIO, this_count, 1);
912                                 return;
913                         } else {
914                                 /* Must have been a power glitch, or a
915                                  * bus reset.  Could not have been a
916                                  * media change, so we just retry the
917                                  * request and see what happens.
918                                  */
919                                 scsi_requeue_command(q, cmd);
920                                 return;
921                         }
922                         break;
923                 case ILLEGAL_REQUEST:
924                         /* If we had an ILLEGAL REQUEST returned, then
925                          * we may have performed an unsupported
926                          * command.  The only thing this should be
927                          * would be a ten byte read where only a six
928                          * byte read was supported.  Also, on a system
929                          * where READ CAPACITY failed, we may have
930                          * read past the end of the disk.
931                          */
932                         if ((cmd->device->use_10_for_rw &&
933                             sshdr.asc == 0x20 && sshdr.ascq == 0x00) &&
934                             (cmd->cmnd[0] == READ_10 ||
935                              cmd->cmnd[0] == WRITE_10)) {
936                                 cmd->device->use_10_for_rw = 0;
937                                 /* This will cause a retry with a
938                                  * 6-byte command.
939                                  */
940                                 scsi_requeue_command(q, cmd);
941                                 return;
942                         } else {
943                                 scsi_end_request(cmd, -EIO, this_count, 1);
944                                 return;
945                         }
946                         break;
947                 case NOT_READY:
948                         /* If the device is in the process of becoming
949                          * ready, or has a temporary blockage, retry.
950                          */
951                         if (sshdr.asc == 0x04) {
952                                 switch (sshdr.ascq) {
953                                 case 0x01: /* becoming ready */
954                                 case 0x04: /* format in progress */
955                                 case 0x05: /* rebuild in progress */
956                                 case 0x06: /* recalculation in progress */
957                                 case 0x07: /* operation in progress */
958                                 case 0x08: /* Long write in progress */
959                                 case 0x09: /* self test in progress */
960                                         scsi_requeue_command(q, cmd);
961                                         return;
962                                 default:
963                                         break;
964                                 }
965                         }
966                         if (!(req->cmd_flags & REQ_QUIET))
967                                 scsi_cmd_print_sense_hdr(cmd,
968                                                          "Device not ready",
969                                                          &sshdr);
970
971                         scsi_end_request(cmd, -EIO, this_count, 1);
972                         return;
973                 case VOLUME_OVERFLOW:
974                         if (!(req->cmd_flags & REQ_QUIET)) {
975                                 scmd_printk(KERN_INFO, cmd,
976                                             "Volume overflow, CDB: ");
977                                 __scsi_print_command(cmd->cmnd);
978                                 scsi_print_sense("", cmd);
979                         }
980                         /* See SSC3rXX or current. */
981                         scsi_end_request(cmd, -EIO, this_count, 1);
982                         return;
983                 default:
984                         break;
985                 }
986         }
987         if (host_byte(result) == DID_RESET) {
988                 /* Third party bus reset or reset for error recovery
989                  * reasons.  Just retry the request and see what
990                  * happens.
991                  */
992                 scsi_requeue_command(q, cmd);
993                 return;
994         }
995         if (result) {
996                 if (!(req->cmd_flags & REQ_QUIET)) {
997                         scsi_print_result(cmd);
998                         if (driver_byte(result) & DRIVER_SENSE)
999                                 scsi_print_sense("", cmd);
1000                 }
1001         }
1002         scsi_end_request(cmd, -EIO, this_count, !result);
1003 }
1004
1005 static int scsi_init_sgtable(struct request *req, struct scsi_data_buffer *sdb,
1006                              gfp_t gfp_mask)
1007 {
1008         int count;
1009
1010         /*
1011          * If sg table allocation fails, requeue request later.
1012          */
1013         if (unlikely(scsi_alloc_sgtable(sdb, req->nr_phys_segments,
1014                                         gfp_mask))) {
1015                 return BLKPREP_DEFER;
1016         }
1017
1018         req->buffer = NULL;
1019
1020         /* 
1021          * Next, walk the list, and fill in the addresses and sizes of
1022          * each segment.
1023          */
1024         count = blk_rq_map_sg(req->q, req, sdb->table.sgl);
1025         BUG_ON(count > sdb->table.nents);
1026         sdb->table.nents = count;
1027         if (blk_pc_request(req))
1028                 sdb->length = req->data_len;
1029         else
1030                 sdb->length = req->nr_sectors << 9;
1031         return BLKPREP_OK;
1032 }
1033
1034 /*
1035  * Function:    scsi_init_io()
1036  *
1037  * Purpose:     SCSI I/O initialize function.
1038  *
1039  * Arguments:   cmd   - Command descriptor we wish to initialize
1040  *
1041  * Returns:     0 on success
1042  *              BLKPREP_DEFER if the failure is retryable
1043  *              BLKPREP_KILL if the failure is fatal
1044  */
1045 int scsi_init_io(struct scsi_cmnd *cmd, gfp_t gfp_mask)
1046 {
1047         int error = scsi_init_sgtable(cmd->request, &cmd->sdb, gfp_mask);
1048         if (error)
1049                 goto err_exit;
1050
1051         if (blk_bidi_rq(cmd->request)) {
1052                 struct scsi_data_buffer *bidi_sdb = kmem_cache_zalloc(
1053                         scsi_bidi_sdb_cache, GFP_ATOMIC);
1054                 if (!bidi_sdb) {
1055                         error = BLKPREP_DEFER;
1056                         goto err_exit;
1057                 }
1058
1059                 cmd->request->next_rq->special = bidi_sdb;
1060                 error = scsi_init_sgtable(cmd->request->next_rq, bidi_sdb,
1061                                                                     GFP_ATOMIC);
1062                 if (error)
1063                         goto err_exit;
1064         }
1065
1066         return BLKPREP_OK ;
1067
1068 err_exit:
1069         scsi_release_buffers(cmd);
1070         if (error == BLKPREP_KILL)
1071                 scsi_put_command(cmd);
1072         else /* BLKPREP_DEFER */
1073                 scsi_unprep_request(cmd->request);
1074
1075         return error;
1076 }
1077 EXPORT_SYMBOL(scsi_init_io);
1078
1079 static struct scsi_cmnd *scsi_get_cmd_from_req(struct scsi_device *sdev,
1080                 struct request *req)
1081 {
1082         struct scsi_cmnd *cmd;
1083
1084         if (!req->special) {
1085                 cmd = scsi_get_command(sdev, GFP_ATOMIC);
1086                 if (unlikely(!cmd))
1087                         return NULL;
1088                 req->special = cmd;
1089         } else {
1090                 cmd = req->special;
1091         }
1092
1093         /* pull a tag out of the request if we have one */
1094         cmd->tag = req->tag;
1095         cmd->request = req;
1096
1097         cmd->cmnd = req->cmd;
1098
1099         return cmd;
1100 }
1101
1102 int scsi_setup_blk_pc_cmnd(struct scsi_device *sdev, struct request *req)
1103 {
1104         struct scsi_cmnd *cmd;
1105         int ret = scsi_prep_state_check(sdev, req);
1106
1107         if (ret != BLKPREP_OK)
1108                 return ret;
1109
1110         cmd = scsi_get_cmd_from_req(sdev, req);
1111         if (unlikely(!cmd))
1112                 return BLKPREP_DEFER;
1113
1114         /*
1115          * BLOCK_PC requests may transfer data, in which case they must
1116          * a bio attached to them.  Or they might contain a SCSI command
1117          * that does not transfer data, in which case they may optionally
1118          * submit a request without an attached bio.
1119          */
1120         if (req->bio) {
1121                 int ret;
1122
1123                 BUG_ON(!req->nr_phys_segments);
1124
1125                 ret = scsi_init_io(cmd, GFP_ATOMIC);
1126                 if (unlikely(ret))
1127                         return ret;
1128         } else {
1129                 BUG_ON(req->data_len);
1130                 BUG_ON(req->data);
1131
1132                 memset(&cmd->sdb, 0, sizeof(cmd->sdb));
1133                 req->buffer = NULL;
1134         }
1135
1136         cmd->cmd_len = req->cmd_len;
1137         if (!req->data_len)
1138                 cmd->sc_data_direction = DMA_NONE;
1139         else if (rq_data_dir(req) == WRITE)
1140                 cmd->sc_data_direction = DMA_TO_DEVICE;
1141         else
1142                 cmd->sc_data_direction = DMA_FROM_DEVICE;
1143         
1144         cmd->transfersize = req->data_len;
1145         cmd->allowed = req->retries;
1146         cmd->timeout_per_command = req->timeout;
1147         return BLKPREP_OK;
1148 }
1149 EXPORT_SYMBOL(scsi_setup_blk_pc_cmnd);
1150
1151 /*
1152  * Setup a REQ_TYPE_FS command.  These are simple read/write request
1153  * from filesystems that still need to be translated to SCSI CDBs from
1154  * the ULD.
1155  */
1156 int scsi_setup_fs_cmnd(struct scsi_device *sdev, struct request *req)
1157 {
1158         struct scsi_cmnd *cmd;
1159         int ret = scsi_prep_state_check(sdev, req);
1160
1161         if (ret != BLKPREP_OK)
1162                 return ret;
1163
1164         if (unlikely(sdev->scsi_dh_data && sdev->scsi_dh_data->scsi_dh
1165                          && sdev->scsi_dh_data->scsi_dh->prep_fn)) {
1166                 ret = sdev->scsi_dh_data->scsi_dh->prep_fn(sdev, req);
1167                 if (ret != BLKPREP_OK)
1168                         return ret;
1169         }
1170
1171         /*
1172          * Filesystem requests must transfer data.
1173          */
1174         BUG_ON(!req->nr_phys_segments);
1175
1176         cmd = scsi_get_cmd_from_req(sdev, req);
1177         if (unlikely(!cmd))
1178                 return BLKPREP_DEFER;
1179
1180         memset(cmd->cmnd, 0, BLK_MAX_CDB);
1181         return scsi_init_io(cmd, GFP_ATOMIC);
1182 }
1183 EXPORT_SYMBOL(scsi_setup_fs_cmnd);
1184
1185 int scsi_prep_state_check(struct scsi_device *sdev, struct request *req)
1186 {
1187         int ret = BLKPREP_OK;
1188
1189         /*
1190          * If the device is not in running state we will reject some
1191          * or all commands.
1192          */
1193         if (unlikely(sdev->sdev_state != SDEV_RUNNING)) {
1194                 switch (sdev->sdev_state) {
1195                 case SDEV_OFFLINE:
1196                         /*
1197                          * If the device is offline we refuse to process any
1198                          * commands.  The device must be brought online
1199                          * before trying any recovery commands.
1200                          */
1201                         sdev_printk(KERN_ERR, sdev,
1202                                     "rejecting I/O to offline device\n");
1203                         ret = BLKPREP_KILL;
1204                         break;
1205                 case SDEV_DEL:
1206                         /*
1207                          * If the device is fully deleted, we refuse to
1208                          * process any commands as well.
1209                          */
1210                         sdev_printk(KERN_ERR, sdev,
1211                                     "rejecting I/O to dead device\n");
1212                         ret = BLKPREP_KILL;
1213                         break;
1214                 case SDEV_QUIESCE:
1215                 case SDEV_BLOCK:
1216                         /*
1217                          * If the devices is blocked we defer normal commands.
1218                          */
1219                         if (!(req->cmd_flags & REQ_PREEMPT))
1220                                 ret = BLKPREP_DEFER;
1221                         break;
1222                 default:
1223                         /*
1224                          * For any other not fully online state we only allow
1225                          * special commands.  In particular any user initiated
1226                          * command is not allowed.
1227                          */
1228                         if (!(req->cmd_flags & REQ_PREEMPT))
1229                                 ret = BLKPREP_KILL;
1230                         break;
1231                 }
1232         }
1233         return ret;
1234 }
1235 EXPORT_SYMBOL(scsi_prep_state_check);
1236
1237 int scsi_prep_return(struct request_queue *q, struct request *req, int ret)
1238 {
1239         struct scsi_device *sdev = q->queuedata;
1240
1241         switch (ret) {
1242         case BLKPREP_KILL:
1243                 req->errors = DID_NO_CONNECT << 16;
1244                 /* release the command and kill it */
1245                 if (req->special) {
1246                         struct scsi_cmnd *cmd = req->special;
1247                         scsi_release_buffers(cmd);
1248                         scsi_put_command(cmd);
1249                         req->special = NULL;
1250                 }
1251                 break;
1252         case BLKPREP_DEFER:
1253                 /*
1254                  * If we defer, the elv_next_request() returns NULL, but the
1255                  * queue must be restarted, so we plug here if no returning
1256                  * command will automatically do that.
1257                  */
1258                 if (sdev->device_busy == 0)
1259                         blk_plug_device(q);
1260                 break;
1261         default:
1262                 req->cmd_flags |= REQ_DONTPREP;
1263         }
1264
1265         return ret;
1266 }
1267 EXPORT_SYMBOL(scsi_prep_return);
1268
1269 int scsi_prep_fn(struct request_queue *q, struct request *req)
1270 {
1271         struct scsi_device *sdev = q->queuedata;
1272         int ret = BLKPREP_KILL;
1273
1274         if (req->cmd_type == REQ_TYPE_BLOCK_PC)
1275                 ret = scsi_setup_blk_pc_cmnd(sdev, req);
1276         return scsi_prep_return(q, req, ret);
1277 }
1278
1279 /*
1280  * scsi_dev_queue_ready: if we can send requests to sdev, return 1 else
1281  * return 0.
1282  *
1283  * Called with the queue_lock held.
1284  */
1285 static inline int scsi_dev_queue_ready(struct request_queue *q,
1286                                   struct scsi_device *sdev)
1287 {
1288         if (sdev->device_busy >= sdev->queue_depth)
1289                 return 0;
1290         if (sdev->device_busy == 0 && sdev->device_blocked) {
1291                 /*
1292                  * unblock after device_blocked iterates to zero
1293                  */
1294                 if (--sdev->device_blocked == 0) {
1295                         SCSI_LOG_MLQUEUE(3,
1296                                    sdev_printk(KERN_INFO, sdev,
1297                                    "unblocking device at zero depth\n"));
1298                 } else {
1299                         blk_plug_device(q);
1300                         return 0;
1301                 }
1302         }
1303         if (sdev->device_blocked)
1304                 return 0;
1305
1306         return 1;
1307 }
1308
1309 /*
1310  * scsi_host_queue_ready: if we can send requests to shost, return 1 else
1311  * return 0. We must end up running the queue again whenever 0 is
1312  * returned, else IO can hang.
1313  *
1314  * Called with host_lock held.
1315  */
1316 static inline int scsi_host_queue_ready(struct request_queue *q,
1317                                    struct Scsi_Host *shost,
1318                                    struct scsi_device *sdev)
1319 {
1320         if (scsi_host_in_recovery(shost))
1321                 return 0;
1322         if (shost->host_busy == 0 && shost->host_blocked) {
1323                 /*
1324                  * unblock after host_blocked iterates to zero
1325                  */
1326                 if (--shost->host_blocked == 0) {
1327                         SCSI_LOG_MLQUEUE(3,
1328                                 printk("scsi%d unblocking host at zero depth\n",
1329                                         shost->host_no));
1330                 } else {
1331                         blk_plug_device(q);
1332                         return 0;
1333                 }
1334         }
1335         if ((shost->can_queue > 0 && shost->host_busy >= shost->can_queue) ||
1336             shost->host_blocked || shost->host_self_blocked) {
1337                 if (list_empty(&sdev->starved_entry))
1338                         list_add_tail(&sdev->starved_entry, &shost->starved_list);
1339                 return 0;
1340         }
1341
1342         /* We're OK to process the command, so we can't be starved */
1343         if (!list_empty(&sdev->starved_entry))
1344                 list_del_init(&sdev->starved_entry);
1345
1346         return 1;
1347 }
1348
1349 /*
1350  * Kill a request for a dead device
1351  */
1352 static void scsi_kill_request(struct request *req, struct request_queue *q)
1353 {
1354         struct scsi_cmnd *cmd = req->special;
1355         struct scsi_device *sdev = cmd->device;
1356         struct Scsi_Host *shost = sdev->host;
1357
1358         blkdev_dequeue_request(req);
1359
1360         if (unlikely(cmd == NULL)) {
1361                 printk(KERN_CRIT "impossible request in %s.\n",
1362                                  __FUNCTION__);
1363                 BUG();
1364         }
1365
1366         scsi_init_cmd_errh(cmd);
1367         cmd->result = DID_NO_CONNECT << 16;
1368         atomic_inc(&cmd->device->iorequest_cnt);
1369
1370         /*
1371          * SCSI request completion path will do scsi_device_unbusy(),
1372          * bump busy counts.  To bump the counters, we need to dance
1373          * with the locks as normal issue path does.
1374          */
1375         sdev->device_busy++;
1376         spin_unlock(sdev->request_queue->queue_lock);
1377         spin_lock(shost->host_lock);
1378         shost->host_busy++;
1379         spin_unlock(shost->host_lock);
1380         spin_lock(sdev->request_queue->queue_lock);
1381
1382         __scsi_done(cmd);
1383 }
1384
1385 static void scsi_softirq_done(struct request *rq)
1386 {
1387         struct scsi_cmnd *cmd = rq->completion_data;
1388         unsigned long wait_for = (cmd->allowed + 1) * cmd->timeout_per_command;
1389         int disposition;
1390
1391         INIT_LIST_HEAD(&cmd->eh_entry);
1392
1393         disposition = scsi_decide_disposition(cmd);
1394         if (disposition != SUCCESS &&
1395             time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) {
1396                 sdev_printk(KERN_ERR, cmd->device,
1397                             "timing out command, waited %lus\n",
1398                             wait_for/HZ);
1399                 disposition = SUCCESS;
1400         }
1401                         
1402         scsi_log_completion(cmd, disposition);
1403
1404         switch (disposition) {
1405                 case SUCCESS:
1406                         scsi_finish_command(cmd);
1407                         break;
1408                 case NEEDS_RETRY:
1409                         scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY);
1410                         break;
1411                 case ADD_TO_MLQUEUE:
1412                         scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY);
1413                         break;
1414                 default:
1415                         if (!scsi_eh_scmd_add(cmd, 0))
1416                                 scsi_finish_command(cmd);
1417         }
1418 }
1419
1420 /*
1421  * Function:    scsi_request_fn()
1422  *
1423  * Purpose:     Main strategy routine for SCSI.
1424  *
1425  * Arguments:   q       - Pointer to actual queue.
1426  *
1427  * Returns:     Nothing
1428  *
1429  * Lock status: IO request lock assumed to be held when called.
1430  */
1431 static void scsi_request_fn(struct request_queue *q)
1432 {
1433         struct scsi_device *sdev = q->queuedata;
1434         struct Scsi_Host *shost;
1435         struct scsi_cmnd *cmd;
1436         struct request *req;
1437
1438         if (!sdev) {
1439                 printk("scsi: killing requests for dead queue\n");
1440                 while ((req = elv_next_request(q)) != NULL)
1441                         scsi_kill_request(req, q);
1442                 return;
1443         }
1444
1445         if(!get_device(&sdev->sdev_gendev))
1446                 /* We must be tearing the block queue down already */
1447                 return;
1448
1449         /*
1450          * To start with, we keep looping until the queue is empty, or until
1451          * the host is no longer able to accept any more requests.
1452          */
1453         shost = sdev->host;
1454         while (!blk_queue_plugged(q)) {
1455                 int rtn;
1456                 /*
1457                  * get next queueable request.  We do this early to make sure
1458                  * that the request is fully prepared even if we cannot 
1459                  * accept it.
1460                  */
1461                 req = elv_next_request(q);
1462                 if (!req || !scsi_dev_queue_ready(q, sdev))
1463                         break;
1464
1465                 if (unlikely(!scsi_device_online(sdev))) {
1466                         sdev_printk(KERN_ERR, sdev,
1467                                     "rejecting I/O to offline device\n");
1468                         scsi_kill_request(req, q);
1469                         continue;
1470                 }
1471
1472
1473                 /*
1474                  * Remove the request from the request list.
1475                  */
1476                 if (!(blk_queue_tagged(q) && !blk_queue_start_tag(q, req)))
1477                         blkdev_dequeue_request(req);
1478                 sdev->device_busy++;
1479
1480                 spin_unlock(q->queue_lock);
1481                 cmd = req->special;
1482                 if (unlikely(cmd == NULL)) {
1483                         printk(KERN_CRIT "impossible request in %s.\n"
1484                                          "please mail a stack trace to "
1485                                          "linux-scsi@vger.kernel.org\n",
1486                                          __FUNCTION__);
1487                         blk_dump_rq_flags(req, "foo");
1488                         BUG();
1489                 }
1490                 spin_lock(shost->host_lock);
1491
1492                 if (!scsi_host_queue_ready(q, shost, sdev))
1493                         goto not_ready;
1494                 if (scsi_target(sdev)->single_lun) {
1495                         if (scsi_target(sdev)->starget_sdev_user &&
1496                             scsi_target(sdev)->starget_sdev_user != sdev)
1497                                 goto not_ready;
1498                         scsi_target(sdev)->starget_sdev_user = sdev;
1499                 }
1500                 shost->host_busy++;
1501
1502                 /*
1503                  * XXX(hch): This is rather suboptimal, scsi_dispatch_cmd will
1504                  *              take the lock again.
1505                  */
1506                 spin_unlock_irq(shost->host_lock);
1507
1508                 /*
1509                  * Finally, initialize any error handling parameters, and set up
1510                  * the timers for timeouts.
1511                  */
1512                 scsi_init_cmd_errh(cmd);
1513
1514                 /*
1515                  * Dispatch the command to the low-level driver.
1516                  */
1517                 rtn = scsi_dispatch_cmd(cmd);
1518                 spin_lock_irq(q->queue_lock);
1519                 if(rtn) {
1520                         /* we're refusing the command; because of
1521                          * the way locks get dropped, we need to 
1522                          * check here if plugging is required */
1523                         if(sdev->device_busy == 0)
1524                                 blk_plug_device(q);
1525
1526                         break;
1527                 }
1528         }
1529
1530         goto out;
1531
1532  not_ready:
1533         spin_unlock_irq(shost->host_lock);
1534
1535         /*
1536          * lock q, handle tag, requeue req, and decrement device_busy. We
1537          * must return with queue_lock held.
1538          *
1539          * Decrementing device_busy without checking it is OK, as all such
1540          * cases (host limits or settings) should run the queue at some
1541          * later time.
1542          */
1543         spin_lock_irq(q->queue_lock);
1544         blk_requeue_request(q, req);
1545         sdev->device_busy--;
1546         if(sdev->device_busy == 0)
1547                 blk_plug_device(q);
1548  out:
1549         /* must be careful here...if we trigger the ->remove() function
1550          * we cannot be holding the q lock */
1551         spin_unlock_irq(q->queue_lock);
1552         put_device(&sdev->sdev_gendev);
1553         spin_lock_irq(q->queue_lock);
1554 }
1555
1556 u64 scsi_calculate_bounce_limit(struct Scsi_Host *shost)
1557 {
1558         struct device *host_dev;
1559         u64 bounce_limit = 0xffffffff;
1560
1561         if (shost->unchecked_isa_dma)
1562                 return BLK_BOUNCE_ISA;
1563         /*
1564          * Platforms with virtual-DMA translation
1565          * hardware have no practical limit.
1566          */
1567         if (!PCI_DMA_BUS_IS_PHYS)
1568                 return BLK_BOUNCE_ANY;
1569
1570         host_dev = scsi_get_device(shost);
1571         if (host_dev && host_dev->dma_mask)
1572                 bounce_limit = *host_dev->dma_mask;
1573
1574         return bounce_limit;
1575 }
1576 EXPORT_SYMBOL(scsi_calculate_bounce_limit);
1577
1578 struct request_queue *__scsi_alloc_queue(struct Scsi_Host *shost,
1579                                          request_fn_proc *request_fn)
1580 {
1581         struct request_queue *q;
1582         struct device *dev = shost->shost_gendev.parent;
1583
1584         q = blk_init_queue(request_fn, NULL);
1585         if (!q)
1586                 return NULL;
1587
1588         /*
1589          * this limit is imposed by hardware restrictions
1590          */
1591         blk_queue_max_hw_segments(q, shost->sg_tablesize);
1592         blk_queue_max_phys_segments(q, SCSI_MAX_SG_CHAIN_SEGMENTS);
1593
1594         blk_queue_max_sectors(q, shost->max_sectors);
1595         blk_queue_bounce_limit(q, scsi_calculate_bounce_limit(shost));
1596         blk_queue_segment_boundary(q, shost->dma_boundary);
1597         dma_set_seg_boundary(dev, shost->dma_boundary);
1598
1599         blk_queue_max_segment_size(q, dma_get_max_seg_size(dev));
1600
1601         /* New queue, no concurrency on queue_flags */
1602         if (!shost->use_clustering)
1603                 queue_flag_clear_unlocked(QUEUE_FLAG_CLUSTER, q);
1604
1605         /*
1606          * set a reasonable default alignment on word boundaries: the
1607          * host and device may alter it using
1608          * blk_queue_update_dma_alignment() later.
1609          */
1610         blk_queue_dma_alignment(q, 0x03);
1611
1612         return q;
1613 }
1614 EXPORT_SYMBOL(__scsi_alloc_queue);
1615
1616 struct request_queue *scsi_alloc_queue(struct scsi_device *sdev)
1617 {
1618         struct request_queue *q;
1619
1620         q = __scsi_alloc_queue(sdev->host, scsi_request_fn);
1621         if (!q)
1622                 return NULL;
1623
1624         blk_queue_prep_rq(q, scsi_prep_fn);
1625         blk_queue_softirq_done(q, scsi_softirq_done);
1626         return q;
1627 }
1628
1629 void scsi_free_queue(struct request_queue *q)
1630 {
1631         blk_cleanup_queue(q);
1632 }
1633
1634 /*
1635  * Function:    scsi_block_requests()
1636  *
1637  * Purpose:     Utility function used by low-level drivers to prevent further
1638  *              commands from being queued to the device.
1639  *
1640  * Arguments:   shost       - Host in question
1641  *
1642  * Returns:     Nothing
1643  *
1644  * Lock status: No locks are assumed held.
1645  *
1646  * Notes:       There is no timer nor any other means by which the requests
1647  *              get unblocked other than the low-level driver calling
1648  *              scsi_unblock_requests().
1649  */
1650 void scsi_block_requests(struct Scsi_Host *shost)
1651 {
1652         shost->host_self_blocked = 1;
1653 }
1654 EXPORT_SYMBOL(scsi_block_requests);
1655
1656 /*
1657  * Function:    scsi_unblock_requests()
1658  *
1659  * Purpose:     Utility function used by low-level drivers to allow further
1660  *              commands from being queued to the device.
1661  *
1662  * Arguments:   shost       - Host in question
1663  *
1664  * Returns:     Nothing
1665  *
1666  * Lock status: No locks are assumed held.
1667  *
1668  * Notes:       There is no timer nor any other means by which the requests
1669  *              get unblocked other than the low-level driver calling
1670  *              scsi_unblock_requests().
1671  *
1672  *              This is done as an API function so that changes to the
1673  *              internals of the scsi mid-layer won't require wholesale
1674  *              changes to drivers that use this feature.
1675  */
1676 void scsi_unblock_requests(struct Scsi_Host *shost)
1677 {
1678         shost->host_self_blocked = 0;
1679         scsi_run_host_queues(shost);
1680 }
1681 EXPORT_SYMBOL(scsi_unblock_requests);
1682
1683 int __init scsi_init_queue(void)
1684 {
1685         int i;
1686
1687         scsi_io_context_cache = kmem_cache_create("scsi_io_context",
1688                                         sizeof(struct scsi_io_context),
1689                                         0, 0, NULL);
1690         if (!scsi_io_context_cache) {
1691                 printk(KERN_ERR "SCSI: can't init scsi io context cache\n");
1692                 return -ENOMEM;
1693         }
1694
1695         scsi_bidi_sdb_cache = kmem_cache_create("scsi_bidi_sdb",
1696                                         sizeof(struct scsi_data_buffer),
1697                                         0, 0, NULL);
1698         if (!scsi_bidi_sdb_cache) {
1699                 printk(KERN_ERR "SCSI: can't init scsi bidi sdb cache\n");
1700                 goto cleanup_io_context;
1701         }
1702
1703         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1704                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1705                 int size = sgp->size * sizeof(struct scatterlist);
1706
1707                 sgp->slab = kmem_cache_create(sgp->name, size, 0,
1708                                 SLAB_HWCACHE_ALIGN, NULL);
1709                 if (!sgp->slab) {
1710                         printk(KERN_ERR "SCSI: can't init sg slab %s\n",
1711                                         sgp->name);
1712                         goto cleanup_bidi_sdb;
1713                 }
1714
1715                 sgp->pool = mempool_create_slab_pool(SG_MEMPOOL_SIZE,
1716                                                      sgp->slab);
1717                 if (!sgp->pool) {
1718                         printk(KERN_ERR "SCSI: can't init sg mempool %s\n",
1719                                         sgp->name);
1720                         goto cleanup_bidi_sdb;
1721                 }
1722         }
1723
1724         return 0;
1725
1726 cleanup_bidi_sdb:
1727         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1728                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1729                 if (sgp->pool)
1730                         mempool_destroy(sgp->pool);
1731                 if (sgp->slab)
1732                         kmem_cache_destroy(sgp->slab);
1733         }
1734         kmem_cache_destroy(scsi_bidi_sdb_cache);
1735 cleanup_io_context:
1736         kmem_cache_destroy(scsi_io_context_cache);
1737
1738         return -ENOMEM;
1739 }
1740
1741 void scsi_exit_queue(void)
1742 {
1743         int i;
1744
1745         kmem_cache_destroy(scsi_io_context_cache);
1746         kmem_cache_destroy(scsi_bidi_sdb_cache);
1747
1748         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1749                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1750                 mempool_destroy(sgp->pool);
1751                 kmem_cache_destroy(sgp->slab);
1752         }
1753 }
1754
1755 /**
1756  *      scsi_mode_select - issue a mode select
1757  *      @sdev:  SCSI device to be queried
1758  *      @pf:    Page format bit (1 == standard, 0 == vendor specific)
1759  *      @sp:    Save page bit (0 == don't save, 1 == save)
1760  *      @modepage: mode page being requested
1761  *      @buffer: request buffer (may not be smaller than eight bytes)
1762  *      @len:   length of request buffer.
1763  *      @timeout: command timeout
1764  *      @retries: number of retries before failing
1765  *      @data: returns a structure abstracting the mode header data
1766  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
1767  *              must be SCSI_SENSE_BUFFERSIZE big.
1768  *
1769  *      Returns zero if successful; negative error number or scsi
1770  *      status on error
1771  *
1772  */
1773 int
1774 scsi_mode_select(struct scsi_device *sdev, int pf, int sp, int modepage,
1775                  unsigned char *buffer, int len, int timeout, int retries,
1776                  struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1777 {
1778         unsigned char cmd[10];
1779         unsigned char *real_buffer;
1780         int ret;
1781
1782         memset(cmd, 0, sizeof(cmd));
1783         cmd[1] = (pf ? 0x10 : 0) | (sp ? 0x01 : 0);
1784
1785         if (sdev->use_10_for_ms) {
1786                 if (len > 65535)
1787                         return -EINVAL;
1788                 real_buffer = kmalloc(8 + len, GFP_KERNEL);
1789                 if (!real_buffer)
1790                         return -ENOMEM;
1791                 memcpy(real_buffer + 8, buffer, len);
1792                 len += 8;
1793                 real_buffer[0] = 0;
1794                 real_buffer[1] = 0;
1795                 real_buffer[2] = data->medium_type;
1796                 real_buffer[3] = data->device_specific;
1797                 real_buffer[4] = data->longlba ? 0x01 : 0;
1798                 real_buffer[5] = 0;
1799                 real_buffer[6] = data->block_descriptor_length >> 8;
1800                 real_buffer[7] = data->block_descriptor_length;
1801
1802                 cmd[0] = MODE_SELECT_10;
1803                 cmd[7] = len >> 8;
1804                 cmd[8] = len;
1805         } else {
1806                 if (len > 255 || data->block_descriptor_length > 255 ||
1807                     data->longlba)
1808                         return -EINVAL;
1809
1810                 real_buffer = kmalloc(4 + len, GFP_KERNEL);
1811                 if (!real_buffer)
1812                         return -ENOMEM;
1813                 memcpy(real_buffer + 4, buffer, len);
1814                 len += 4;
1815                 real_buffer[0] = 0;
1816                 real_buffer[1] = data->medium_type;
1817                 real_buffer[2] = data->device_specific;
1818                 real_buffer[3] = data->block_descriptor_length;
1819                 
1820
1821                 cmd[0] = MODE_SELECT;
1822                 cmd[4] = len;
1823         }
1824
1825         ret = scsi_execute_req(sdev, cmd, DMA_TO_DEVICE, real_buffer, len,
1826                                sshdr, timeout, retries);
1827         kfree(real_buffer);
1828         return ret;
1829 }
1830 EXPORT_SYMBOL_GPL(scsi_mode_select);
1831
1832 /**
1833  *      scsi_mode_sense - issue a mode sense, falling back from 10 to six bytes if necessary.
1834  *      @sdev:  SCSI device to be queried
1835  *      @dbd:   set if mode sense will allow block descriptors to be returned
1836  *      @modepage: mode page being requested
1837  *      @buffer: request buffer (may not be smaller than eight bytes)
1838  *      @len:   length of request buffer.
1839  *      @timeout: command timeout
1840  *      @retries: number of retries before failing
1841  *      @data: returns a structure abstracting the mode header data
1842  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
1843  *              must be SCSI_SENSE_BUFFERSIZE big.
1844  *
1845  *      Returns zero if unsuccessful, or the header offset (either 4
1846  *      or 8 depending on whether a six or ten byte command was
1847  *      issued) if successful.
1848  */
1849 int
1850 scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
1851                   unsigned char *buffer, int len, int timeout, int retries,
1852                   struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1853 {
1854         unsigned char cmd[12];
1855         int use_10_for_ms;
1856         int header_length;
1857         int result;
1858         struct scsi_sense_hdr my_sshdr;
1859
1860         memset(data, 0, sizeof(*data));
1861         memset(&cmd[0], 0, 12);
1862         cmd[1] = dbd & 0x18;    /* allows DBD and LLBA bits */
1863         cmd[2] = modepage;
1864
1865         /* caller might not be interested in sense, but we need it */
1866         if (!sshdr)
1867                 sshdr = &my_sshdr;
1868
1869  retry:
1870         use_10_for_ms = sdev->use_10_for_ms;
1871
1872         if (use_10_for_ms) {
1873                 if (len < 8)
1874                         len = 8;
1875
1876                 cmd[0] = MODE_SENSE_10;
1877                 cmd[8] = len;
1878                 header_length = 8;
1879         } else {
1880                 if (len < 4)
1881                         len = 4;
1882
1883                 cmd[0] = MODE_SENSE;
1884                 cmd[4] = len;
1885                 header_length = 4;
1886         }
1887
1888         memset(buffer, 0, len);
1889
1890         result = scsi_execute_req(sdev, cmd, DMA_FROM_DEVICE, buffer, len,
1891                                   sshdr, timeout, retries);
1892
1893         /* This code looks awful: what it's doing is making sure an
1894          * ILLEGAL REQUEST sense return identifies the actual command
1895          * byte as the problem.  MODE_SENSE commands can return
1896          * ILLEGAL REQUEST if the code page isn't supported */
1897
1898         if (use_10_for_ms && !scsi_status_is_good(result) &&
1899             (driver_byte(result) & DRIVER_SENSE)) {
1900                 if (scsi_sense_valid(sshdr)) {
1901                         if ((sshdr->sense_key == ILLEGAL_REQUEST) &&
1902                             (sshdr->asc == 0x20) && (sshdr->ascq == 0)) {
1903                                 /* 
1904                                  * Invalid command operation code
1905                                  */
1906                                 sdev->use_10_for_ms = 0;
1907                                 goto retry;
1908                         }
1909                 }
1910         }
1911
1912         if(scsi_status_is_good(result)) {
1913                 if (unlikely(buffer[0] == 0x86 && buffer[1] == 0x0b &&
1914                              (modepage == 6 || modepage == 8))) {
1915                         /* Initio breakage? */
1916                         header_length = 0;
1917                         data->length = 13;
1918                         data->medium_type = 0;
1919                         data->device_specific = 0;
1920                         data->longlba = 0;
1921                         data->block_descriptor_length = 0;
1922                 } else if(use_10_for_ms) {
1923                         data->length = buffer[0]*256 + buffer[1] + 2;
1924                         data->medium_type = buffer[2];
1925                         data->device_specific = buffer[3];
1926                         data->longlba = buffer[4] & 0x01;
1927                         data->block_descriptor_length = buffer[6]*256
1928                                 + buffer[7];
1929                 } else {
1930                         data->length = buffer[0] + 1;
1931                         data->medium_type = buffer[1];
1932                         data->device_specific = buffer[2];
1933                         data->block_descriptor_length = buffer[3];
1934                 }
1935                 data->header_length = header_length;
1936         }
1937
1938         return result;
1939 }
1940 EXPORT_SYMBOL(scsi_mode_sense);
1941
1942 /**
1943  *      scsi_test_unit_ready - test if unit is ready
1944  *      @sdev:  scsi device to change the state of.
1945  *      @timeout: command timeout
1946  *      @retries: number of retries before failing
1947  *      @sshdr_external: Optional pointer to struct scsi_sense_hdr for
1948  *              returning sense. Make sure that this is cleared before passing
1949  *              in.
1950  *
1951  *      Returns zero if unsuccessful or an error if TUR failed.  For
1952  *      removable media, a return of NOT_READY or UNIT_ATTENTION is
1953  *      translated to success, with the ->changed flag updated.
1954  **/
1955 int
1956 scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries,
1957                      struct scsi_sense_hdr *sshdr_external)
1958 {
1959         char cmd[] = {
1960                 TEST_UNIT_READY, 0, 0, 0, 0, 0,
1961         };
1962         struct scsi_sense_hdr *sshdr;
1963         int result;
1964
1965         if (!sshdr_external)
1966                 sshdr = kzalloc(sizeof(*sshdr), GFP_KERNEL);
1967         else
1968                 sshdr = sshdr_external;
1969
1970         /* try to eat the UNIT_ATTENTION if there are enough retries */
1971         do {
1972                 result = scsi_execute_req(sdev, cmd, DMA_NONE, NULL, 0, sshdr,
1973                                           timeout, retries);
1974         } while ((driver_byte(result) & DRIVER_SENSE) &&
1975                  sshdr && sshdr->sense_key == UNIT_ATTENTION &&
1976                  --retries);
1977
1978         if (!sshdr)
1979                 /* could not allocate sense buffer, so can't process it */
1980                 return result;
1981
1982         if ((driver_byte(result) & DRIVER_SENSE) && sdev->removable) {
1983
1984                 if ((scsi_sense_valid(sshdr)) &&
1985                     ((sshdr->sense_key == UNIT_ATTENTION) ||
1986                      (sshdr->sense_key == NOT_READY))) {
1987                         sdev->changed = 1;
1988                         result = 0;
1989                 }
1990         }
1991         if (!sshdr_external)
1992                 kfree(sshdr);
1993         return result;
1994 }
1995 EXPORT_SYMBOL(scsi_test_unit_ready);
1996
1997 /**
1998  *      scsi_device_set_state - Take the given device through the device state model.
1999  *      @sdev:  scsi device to change the state of.
2000  *      @state: state to change to.
2001  *
2002  *      Returns zero if unsuccessful or an error if the requested 
2003  *      transition is illegal.
2004  */
2005 int
2006 scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
2007 {
2008         enum scsi_device_state oldstate = sdev->sdev_state;
2009
2010         if (state == oldstate)
2011                 return 0;
2012
2013         switch (state) {
2014         case SDEV_CREATED:
2015                 /* There are no legal states that come back to
2016                  * created.  This is the manually initialised start
2017                  * state */
2018                 goto illegal;
2019                         
2020         case SDEV_RUNNING:
2021                 switch (oldstate) {
2022                 case SDEV_CREATED:
2023                 case SDEV_OFFLINE:
2024                 case SDEV_QUIESCE:
2025                 case SDEV_BLOCK:
2026                         break;
2027                 default:
2028                         goto illegal;
2029                 }
2030                 break;
2031
2032         case SDEV_QUIESCE:
2033                 switch (oldstate) {
2034                 case SDEV_RUNNING:
2035                 case SDEV_OFFLINE:
2036                         break;
2037                 default:
2038                         goto illegal;
2039                 }
2040                 break;
2041
2042         case SDEV_OFFLINE:
2043                 switch (oldstate) {
2044                 case SDEV_CREATED:
2045                 case SDEV_RUNNING:
2046                 case SDEV_QUIESCE:
2047                 case SDEV_BLOCK:
2048                         break;
2049                 default:
2050                         goto illegal;
2051                 }
2052                 break;
2053
2054         case SDEV_BLOCK:
2055                 switch (oldstate) {
2056                 case SDEV_CREATED:
2057                 case SDEV_RUNNING:
2058                         break;
2059                 default:
2060                         goto illegal;
2061                 }
2062                 break;
2063
2064         case SDEV_CANCEL:
2065                 switch (oldstate) {
2066                 case SDEV_CREATED:
2067                 case SDEV_RUNNING:
2068                 case SDEV_QUIESCE:
2069                 case SDEV_OFFLINE:
2070                 case SDEV_BLOCK:
2071                         break;
2072                 default:
2073                         goto illegal;
2074                 }
2075                 break;
2076
2077         case SDEV_DEL:
2078                 switch (oldstate) {
2079                 case SDEV_CREATED:
2080                 case SDEV_RUNNING:
2081                 case SDEV_OFFLINE:
2082                 case SDEV_CANCEL:
2083                         break;
2084                 default:
2085                         goto illegal;
2086                 }
2087                 break;
2088
2089         }
2090         sdev->sdev_state = state;
2091         return 0;
2092
2093  illegal:
2094         SCSI_LOG_ERROR_RECOVERY(1, 
2095                                 sdev_printk(KERN_ERR, sdev,
2096                                             "Illegal state transition %s->%s\n",
2097                                             scsi_device_state_name(oldstate),
2098                                             scsi_device_state_name(state))
2099                                 );
2100         return -EINVAL;
2101 }
2102 EXPORT_SYMBOL(scsi_device_set_state);
2103
2104 /**
2105  *      sdev_evt_emit - emit a single SCSI device uevent
2106  *      @sdev: associated SCSI device
2107  *      @evt: event to emit
2108  *
2109  *      Send a single uevent (scsi_event) to the associated scsi_device.
2110  */
2111 static void scsi_evt_emit(struct scsi_device *sdev, struct scsi_event *evt)
2112 {
2113         int idx = 0;
2114         char *envp[3];
2115
2116         switch (evt->evt_type) {
2117         case SDEV_EVT_MEDIA_CHANGE:
2118                 envp[idx++] = "SDEV_MEDIA_CHANGE=1";
2119                 break;
2120
2121         default:
2122                 /* do nothing */
2123                 break;
2124         }
2125
2126         envp[idx++] = NULL;
2127
2128         kobject_uevent_env(&sdev->sdev_gendev.kobj, KOBJ_CHANGE, envp);
2129 }
2130
2131 /**
2132  *      sdev_evt_thread - send a uevent for each scsi event
2133  *      @work: work struct for scsi_device
2134  *
2135  *      Dispatch queued events to their associated scsi_device kobjects
2136  *      as uevents.
2137  */
2138 void scsi_evt_thread(struct work_struct *work)
2139 {
2140         struct scsi_device *sdev;
2141         LIST_HEAD(event_list);
2142
2143         sdev = container_of(work, struct scsi_device, event_work);
2144
2145         while (1) {
2146                 struct scsi_event *evt;
2147                 struct list_head *this, *tmp;
2148                 unsigned long flags;
2149
2150                 spin_lock_irqsave(&sdev->list_lock, flags);
2151                 list_splice_init(&sdev->event_list, &event_list);
2152                 spin_unlock_irqrestore(&sdev->list_lock, flags);
2153
2154                 if (list_empty(&event_list))
2155                         break;
2156
2157                 list_for_each_safe(this, tmp, &event_list) {
2158                         evt = list_entry(this, struct scsi_event, node);
2159                         list_del(&evt->node);
2160                         scsi_evt_emit(sdev, evt);
2161                         kfree(evt);
2162                 }
2163         }
2164 }
2165
2166 /**
2167  *      sdev_evt_send - send asserted event to uevent thread
2168  *      @sdev: scsi_device event occurred on
2169  *      @evt: event to send
2170  *
2171  *      Assert scsi device event asynchronously.
2172  */
2173 void sdev_evt_send(struct scsi_device *sdev, struct scsi_event *evt)
2174 {
2175         unsigned long flags;
2176
2177 #if 0
2178         /* FIXME: currently this check eliminates all media change events
2179          * for polled devices.  Need to update to discriminate between AN
2180          * and polled events */
2181         if (!test_bit(evt->evt_type, sdev->supported_events)) {
2182                 kfree(evt);
2183                 return;
2184         }
2185 #endif
2186
2187         spin_lock_irqsave(&sdev->list_lock, flags);
2188         list_add_tail(&evt->node, &sdev->event_list);
2189         schedule_work(&sdev->event_work);
2190         spin_unlock_irqrestore(&sdev->list_lock, flags);
2191 }
2192 EXPORT_SYMBOL_GPL(sdev_evt_send);
2193
2194 /**
2195  *      sdev_evt_alloc - allocate a new scsi event
2196  *      @evt_type: type of event to allocate
2197  *      @gfpflags: GFP flags for allocation
2198  *
2199  *      Allocates and returns a new scsi_event.
2200  */
2201 struct scsi_event *sdev_evt_alloc(enum scsi_device_event evt_type,
2202                                   gfp_t gfpflags)
2203 {
2204         struct scsi_event *evt = kzalloc(sizeof(struct scsi_event), gfpflags);
2205         if (!evt)
2206                 return NULL;
2207
2208         evt->evt_type = evt_type;
2209         INIT_LIST_HEAD(&evt->node);
2210
2211         /* evt_type-specific initialization, if any */
2212         switch (evt_type) {
2213         case SDEV_EVT_MEDIA_CHANGE:
2214         default:
2215                 /* do nothing */
2216                 break;
2217         }
2218
2219         return evt;
2220 }
2221 EXPORT_SYMBOL_GPL(sdev_evt_alloc);
2222
2223 /**
2224  *      sdev_evt_send_simple - send asserted event to uevent thread
2225  *      @sdev: scsi_device event occurred on
2226  *      @evt_type: type of event to send
2227  *      @gfpflags: GFP flags for allocation
2228  *
2229  *      Assert scsi device event asynchronously, given an event type.
2230  */
2231 void sdev_evt_send_simple(struct scsi_device *sdev,
2232                           enum scsi_device_event evt_type, gfp_t gfpflags)
2233 {
2234         struct scsi_event *evt = sdev_evt_alloc(evt_type, gfpflags);
2235         if (!evt) {
2236                 sdev_printk(KERN_ERR, sdev, "event %d eaten due to OOM\n",
2237                             evt_type);
2238                 return;
2239         }
2240
2241         sdev_evt_send(sdev, evt);
2242 }
2243 EXPORT_SYMBOL_GPL(sdev_evt_send_simple);
2244
2245 /**
2246  *      scsi_device_quiesce - Block user issued commands.
2247  *      @sdev:  scsi device to quiesce.
2248  *
2249  *      This works by trying to transition to the SDEV_QUIESCE state
2250  *      (which must be a legal transition).  When the device is in this
2251  *      state, only special requests will be accepted, all others will
2252  *      be deferred.  Since special requests may also be requeued requests,
2253  *      a successful return doesn't guarantee the device will be 
2254  *      totally quiescent.
2255  *
2256  *      Must be called with user context, may sleep.
2257  *
2258  *      Returns zero if unsuccessful or an error if not.
2259  */
2260 int
2261 scsi_device_quiesce(struct scsi_device *sdev)
2262 {
2263         int err = scsi_device_set_state(sdev, SDEV_QUIESCE);
2264         if (err)
2265                 return err;
2266
2267         scsi_run_queue(sdev->request_queue);
2268         while (sdev->device_busy) {
2269                 msleep_interruptible(200);
2270                 scsi_run_queue(sdev->request_queue);
2271         }
2272         return 0;
2273 }
2274 EXPORT_SYMBOL(scsi_device_quiesce);
2275
2276 /**
2277  *      scsi_device_resume - Restart user issued commands to a quiesced device.
2278  *      @sdev:  scsi device to resume.
2279  *
2280  *      Moves the device from quiesced back to running and restarts the
2281  *      queues.
2282  *
2283  *      Must be called with user context, may sleep.
2284  */
2285 void
2286 scsi_device_resume(struct scsi_device *sdev)
2287 {
2288         if(scsi_device_set_state(sdev, SDEV_RUNNING))
2289                 return;
2290         scsi_run_queue(sdev->request_queue);
2291 }
2292 EXPORT_SYMBOL(scsi_device_resume);
2293
2294 static void
2295 device_quiesce_fn(struct scsi_device *sdev, void *data)
2296 {
2297         scsi_device_quiesce(sdev);
2298 }
2299
2300 void
2301 scsi_target_quiesce(struct scsi_target *starget)
2302 {
2303         starget_for_each_device(starget, NULL, device_quiesce_fn);
2304 }
2305 EXPORT_SYMBOL(scsi_target_quiesce);
2306
2307 static void
2308 device_resume_fn(struct scsi_device *sdev, void *data)
2309 {
2310         scsi_device_resume(sdev);
2311 }
2312
2313 void
2314 scsi_target_resume(struct scsi_target *starget)
2315 {
2316         starget_for_each_device(starget, NULL, device_resume_fn);
2317 }
2318 EXPORT_SYMBOL(scsi_target_resume);
2319
2320 /**
2321  * scsi_internal_device_block - internal function to put a device temporarily into the SDEV_BLOCK state
2322  * @sdev:       device to block
2323  *
2324  * Block request made by scsi lld's to temporarily stop all
2325  * scsi commands on the specified device.  Called from interrupt
2326  * or normal process context.
2327  *
2328  * Returns zero if successful or error if not
2329  *
2330  * Notes:       
2331  *      This routine transitions the device to the SDEV_BLOCK state
2332  *      (which must be a legal transition).  When the device is in this
2333  *      state, all commands are deferred until the scsi lld reenables
2334  *      the device with scsi_device_unblock or device_block_tmo fires.
2335  *      This routine assumes the host_lock is held on entry.
2336  */
2337 int
2338 scsi_internal_device_block(struct scsi_device *sdev)
2339 {
2340         struct request_queue *q = sdev->request_queue;
2341         unsigned long flags;
2342         int err = 0;
2343
2344         err = scsi_device_set_state(sdev, SDEV_BLOCK);
2345         if (err)
2346                 return err;
2347
2348         /* 
2349          * The device has transitioned to SDEV_BLOCK.  Stop the
2350          * block layer from calling the midlayer with this device's
2351          * request queue. 
2352          */
2353         spin_lock_irqsave(q->queue_lock, flags);
2354         blk_stop_queue(q);
2355         spin_unlock_irqrestore(q->queue_lock, flags);
2356
2357         return 0;
2358 }
2359 EXPORT_SYMBOL_GPL(scsi_internal_device_block);
2360  
2361 /**
2362  * scsi_internal_device_unblock - resume a device after a block request
2363  * @sdev:       device to resume
2364  *
2365  * Called by scsi lld's or the midlayer to restart the device queue
2366  * for the previously suspended scsi device.  Called from interrupt or
2367  * normal process context.
2368  *
2369  * Returns zero if successful or error if not.
2370  *
2371  * Notes:       
2372  *      This routine transitions the device to the SDEV_RUNNING state
2373  *      (which must be a legal transition) allowing the midlayer to
2374  *      goose the queue for this device.  This routine assumes the 
2375  *      host_lock is held upon entry.
2376  */
2377 int
2378 scsi_internal_device_unblock(struct scsi_device *sdev)
2379 {
2380         struct request_queue *q = sdev->request_queue; 
2381         int err;
2382         unsigned long flags;
2383         
2384         /* 
2385          * Try to transition the scsi device to SDEV_RUNNING
2386          * and goose the device queue if successful.  
2387          */
2388         err = scsi_device_set_state(sdev, SDEV_RUNNING);
2389         if (err)
2390                 return err;
2391
2392         spin_lock_irqsave(q->queue_lock, flags);
2393         blk_start_queue(q);
2394         spin_unlock_irqrestore(q->queue_lock, flags);
2395
2396         return 0;
2397 }
2398 EXPORT_SYMBOL_GPL(scsi_internal_device_unblock);
2399
2400 static void
2401 device_block(struct scsi_device *sdev, void *data)
2402 {
2403         scsi_internal_device_block(sdev);
2404 }
2405
2406 static int
2407 target_block(struct device *dev, void *data)
2408 {
2409         if (scsi_is_target_device(dev))
2410                 starget_for_each_device(to_scsi_target(dev), NULL,
2411                                         device_block);
2412         return 0;
2413 }
2414
2415 void
2416 scsi_target_block(struct device *dev)
2417 {
2418         if (scsi_is_target_device(dev))
2419                 starget_for_each_device(to_scsi_target(dev), NULL,
2420                                         device_block);
2421         else
2422                 device_for_each_child(dev, NULL, target_block);
2423 }
2424 EXPORT_SYMBOL_GPL(scsi_target_block);
2425
2426 static void
2427 device_unblock(struct scsi_device *sdev, void *data)
2428 {
2429         scsi_internal_device_unblock(sdev);
2430 }
2431
2432 static int
2433 target_unblock(struct device *dev, void *data)
2434 {
2435         if (scsi_is_target_device(dev))
2436                 starget_for_each_device(to_scsi_target(dev), NULL,
2437                                         device_unblock);
2438         return 0;
2439 }
2440
2441 void
2442 scsi_target_unblock(struct device *dev)
2443 {
2444         if (scsi_is_target_device(dev))
2445                 starget_for_each_device(to_scsi_target(dev), NULL,
2446                                         device_unblock);
2447         else
2448                 device_for_each_child(dev, NULL, target_unblock);
2449 }
2450 EXPORT_SYMBOL_GPL(scsi_target_unblock);
2451
2452 /**
2453  * scsi_kmap_atomic_sg - find and atomically map an sg-elemnt
2454  * @sgl:        scatter-gather list
2455  * @sg_count:   number of segments in sg
2456  * @offset:     offset in bytes into sg, on return offset into the mapped area
2457  * @len:        bytes to map, on return number of bytes mapped
2458  *
2459  * Returns virtual address of the start of the mapped page
2460  */
2461 void *scsi_kmap_atomic_sg(struct scatterlist *sgl, int sg_count,
2462                           size_t *offset, size_t *len)
2463 {
2464         int i;
2465         size_t sg_len = 0, len_complete = 0;
2466         struct scatterlist *sg;
2467         struct page *page;
2468
2469         WARN_ON(!irqs_disabled());
2470
2471         for_each_sg(sgl, sg, sg_count, i) {
2472                 len_complete = sg_len; /* Complete sg-entries */
2473                 sg_len += sg->length;
2474                 if (sg_len > *offset)
2475                         break;
2476         }
2477
2478         if (unlikely(i == sg_count)) {
2479                 printk(KERN_ERR "%s: Bytes in sg: %zu, requested offset %zu, "
2480                         "elements %d\n",
2481                        __FUNCTION__, sg_len, *offset, sg_count);
2482                 WARN_ON(1);
2483                 return NULL;
2484         }
2485
2486         /* Offset starting from the beginning of first page in this sg-entry */
2487         *offset = *offset - len_complete + sg->offset;
2488
2489         /* Assumption: contiguous pages can be accessed as "page + i" */
2490         page = nth_page(sg_page(sg), (*offset >> PAGE_SHIFT));
2491         *offset &= ~PAGE_MASK;
2492
2493         /* Bytes in this sg-entry from *offset to the end of the page */
2494         sg_len = PAGE_SIZE - *offset;
2495         if (*len > sg_len)
2496                 *len = sg_len;
2497
2498         return kmap_atomic(page, KM_BIO_SRC_IRQ);
2499 }
2500 EXPORT_SYMBOL(scsi_kmap_atomic_sg);
2501
2502 /**
2503  * scsi_kunmap_atomic_sg - atomically unmap a virtual address, previously mapped with scsi_kmap_atomic_sg
2504  * @virt:       virtual address to be unmapped
2505  */
2506 void scsi_kunmap_atomic_sg(void *virt)
2507 {
2508         kunmap_atomic(virt, KM_BIO_SRC_IRQ);
2509 }
2510 EXPORT_SYMBOL(scsi_kunmap_atomic_sg);