Pull bugzilla-9494 into release branch
[linux-2.6] / drivers / scsi / scsi_lib.c
1 /*
2  *  scsi_lib.c Copyright (C) 1999 Eric Youngdale
3  *
4  *  SCSI queueing library.
5  *      Initial versions: Eric Youngdale (eric@andante.org).
6  *                        Based upon conversations with large numbers
7  *                        of people at Linux Expo.
8  */
9
10 #include <linux/bio.h>
11 #include <linux/blkdev.h>
12 #include <linux/completion.h>
13 #include <linux/kernel.h>
14 #include <linux/mempool.h>
15 #include <linux/slab.h>
16 #include <linux/init.h>
17 #include <linux/pci.h>
18 #include <linux/delay.h>
19 #include <linux/hardirq.h>
20 #include <linux/scatterlist.h>
21
22 #include <scsi/scsi.h>
23 #include <scsi/scsi_cmnd.h>
24 #include <scsi/scsi_dbg.h>
25 #include <scsi/scsi_device.h>
26 #include <scsi/scsi_driver.h>
27 #include <scsi/scsi_eh.h>
28 #include <scsi/scsi_host.h>
29
30 #include "scsi_priv.h"
31 #include "scsi_logging.h"
32
33
34 #define SG_MEMPOOL_NR           ARRAY_SIZE(scsi_sg_pools)
35 #define SG_MEMPOOL_SIZE         2
36
37 /*
38  * The maximum number of SG segments that we will put inside a scatterlist
39  * (unless chaining is used). Should ideally fit inside a single page, to
40  * avoid a higher order allocation.
41  */
42 #define SCSI_MAX_SG_SEGMENTS    128
43
44 struct scsi_host_sg_pool {
45         size_t          size;
46         char            *name;
47         struct kmem_cache       *slab;
48         mempool_t       *pool;
49 };
50
51 #define SP(x) { x, "sgpool-" #x }
52 static struct scsi_host_sg_pool scsi_sg_pools[] = {
53         SP(8),
54         SP(16),
55 #if (SCSI_MAX_SG_SEGMENTS > 16)
56         SP(32),
57 #if (SCSI_MAX_SG_SEGMENTS > 32)
58         SP(64),
59 #if (SCSI_MAX_SG_SEGMENTS > 64)
60         SP(128),
61 #endif
62 #endif
63 #endif
64 };
65 #undef SP
66
67 static void scsi_run_queue(struct request_queue *q);
68
69 /*
70  * Function:    scsi_unprep_request()
71  *
72  * Purpose:     Remove all preparation done for a request, including its
73  *              associated scsi_cmnd, so that it can be requeued.
74  *
75  * Arguments:   req     - request to unprepare
76  *
77  * Lock status: Assumed that no locks are held upon entry.
78  *
79  * Returns:     Nothing.
80  */
81 static void scsi_unprep_request(struct request *req)
82 {
83         struct scsi_cmnd *cmd = req->special;
84
85         req->cmd_flags &= ~REQ_DONTPREP;
86         req->special = NULL;
87
88         scsi_put_command(cmd);
89 }
90
91 /*
92  * Function:    scsi_queue_insert()
93  *
94  * Purpose:     Insert a command in the midlevel queue.
95  *
96  * Arguments:   cmd    - command that we are adding to queue.
97  *              reason - why we are inserting command to queue.
98  *
99  * Lock status: Assumed that lock is not held upon entry.
100  *
101  * Returns:     Nothing.
102  *
103  * Notes:       We do this for one of two cases.  Either the host is busy
104  *              and it cannot accept any more commands for the time being,
105  *              or the device returned QUEUE_FULL and can accept no more
106  *              commands.
107  * Notes:       This could be called either from an interrupt context or a
108  *              normal process context.
109  */
110 int scsi_queue_insert(struct scsi_cmnd *cmd, int reason)
111 {
112         struct Scsi_Host *host = cmd->device->host;
113         struct scsi_device *device = cmd->device;
114         struct request_queue *q = device->request_queue;
115         unsigned long flags;
116
117         SCSI_LOG_MLQUEUE(1,
118                  printk("Inserting command %p into mlqueue\n", cmd));
119
120         /*
121          * Set the appropriate busy bit for the device/host.
122          *
123          * If the host/device isn't busy, assume that something actually
124          * completed, and that we should be able to queue a command now.
125          *
126          * Note that the prior mid-layer assumption that any host could
127          * always queue at least one command is now broken.  The mid-layer
128          * will implement a user specifiable stall (see
129          * scsi_host.max_host_blocked and scsi_device.max_device_blocked)
130          * if a command is requeued with no other commands outstanding
131          * either for the device or for the host.
132          */
133         if (reason == SCSI_MLQUEUE_HOST_BUSY)
134                 host->host_blocked = host->max_host_blocked;
135         else if (reason == SCSI_MLQUEUE_DEVICE_BUSY)
136                 device->device_blocked = device->max_device_blocked;
137
138         /*
139          * Decrement the counters, since these commands are no longer
140          * active on the host/device.
141          */
142         scsi_device_unbusy(device);
143
144         /*
145          * Requeue this command.  It will go before all other commands
146          * that are already in the queue.
147          *
148          * NOTE: there is magic here about the way the queue is plugged if
149          * we have no outstanding commands.
150          * 
151          * Although we *don't* plug the queue, we call the request
152          * function.  The SCSI request function detects the blocked condition
153          * and plugs the queue appropriately.
154          */
155         spin_lock_irqsave(q->queue_lock, flags);
156         blk_requeue_request(q, cmd->request);
157         spin_unlock_irqrestore(q->queue_lock, flags);
158
159         scsi_run_queue(q);
160
161         return 0;
162 }
163
164 /**
165  * scsi_execute - insert request and wait for the result
166  * @sdev:       scsi device
167  * @cmd:        scsi command
168  * @data_direction: data direction
169  * @buffer:     data buffer
170  * @bufflen:    len of buffer
171  * @sense:      optional sense buffer
172  * @timeout:    request timeout in seconds
173  * @retries:    number of times to retry request
174  * @flags:      or into request flags;
175  *
176  * returns the req->errors value which is the scsi_cmnd result
177  * field.
178  **/
179 int scsi_execute(struct scsi_device *sdev, const unsigned char *cmd,
180                  int data_direction, void *buffer, unsigned bufflen,
181                  unsigned char *sense, int timeout, int retries, int flags)
182 {
183         struct request *req;
184         int write = (data_direction == DMA_TO_DEVICE);
185         int ret = DRIVER_ERROR << 24;
186
187         req = blk_get_request(sdev->request_queue, write, __GFP_WAIT);
188
189         if (bufflen &&  blk_rq_map_kern(sdev->request_queue, req,
190                                         buffer, bufflen, __GFP_WAIT))
191                 goto out;
192
193         req->cmd_len = COMMAND_SIZE(cmd[0]);
194         memcpy(req->cmd, cmd, req->cmd_len);
195         req->sense = sense;
196         req->sense_len = 0;
197         req->retries = retries;
198         req->timeout = timeout;
199         req->cmd_type = REQ_TYPE_BLOCK_PC;
200         req->cmd_flags |= flags | REQ_QUIET | REQ_PREEMPT;
201
202         /*
203          * head injection *required* here otherwise quiesce won't work
204          */
205         blk_execute_rq(req->q, NULL, req, 1);
206
207         ret = req->errors;
208  out:
209         blk_put_request(req);
210
211         return ret;
212 }
213 EXPORT_SYMBOL(scsi_execute);
214
215
216 int scsi_execute_req(struct scsi_device *sdev, const unsigned char *cmd,
217                      int data_direction, void *buffer, unsigned bufflen,
218                      struct scsi_sense_hdr *sshdr, int timeout, int retries)
219 {
220         char *sense = NULL;
221         int result;
222         
223         if (sshdr) {
224                 sense = kzalloc(SCSI_SENSE_BUFFERSIZE, GFP_NOIO);
225                 if (!sense)
226                         return DRIVER_ERROR << 24;
227         }
228         result = scsi_execute(sdev, cmd, data_direction, buffer, bufflen,
229                               sense, timeout, retries, 0);
230         if (sshdr)
231                 scsi_normalize_sense(sense, SCSI_SENSE_BUFFERSIZE, sshdr);
232
233         kfree(sense);
234         return result;
235 }
236 EXPORT_SYMBOL(scsi_execute_req);
237
238 struct scsi_io_context {
239         void *data;
240         void (*done)(void *data, char *sense, int result, int resid);
241         char sense[SCSI_SENSE_BUFFERSIZE];
242 };
243
244 static struct kmem_cache *scsi_io_context_cache;
245
246 static void scsi_end_async(struct request *req, int uptodate)
247 {
248         struct scsi_io_context *sioc = req->end_io_data;
249
250         if (sioc->done)
251                 sioc->done(sioc->data, sioc->sense, req->errors, req->data_len);
252
253         kmem_cache_free(scsi_io_context_cache, sioc);
254         __blk_put_request(req->q, req);
255 }
256
257 static int scsi_merge_bio(struct request *rq, struct bio *bio)
258 {
259         struct request_queue *q = rq->q;
260
261         bio->bi_flags &= ~(1 << BIO_SEG_VALID);
262         if (rq_data_dir(rq) == WRITE)
263                 bio->bi_rw |= (1 << BIO_RW);
264         blk_queue_bounce(q, &bio);
265
266         return blk_rq_append_bio(q, rq, bio);
267 }
268
269 static void scsi_bi_endio(struct bio *bio, int error)
270 {
271         bio_put(bio);
272 }
273
274 /**
275  * scsi_req_map_sg - map a scatterlist into a request
276  * @rq:         request to fill
277  * @sg:         scatterlist
278  * @nsegs:      number of elements
279  * @bufflen:    len of buffer
280  * @gfp:        memory allocation flags
281  *
282  * scsi_req_map_sg maps a scatterlist into a request so that the
283  * request can be sent to the block layer. We do not trust the scatterlist
284  * sent to use, as some ULDs use that struct to only organize the pages.
285  */
286 static int scsi_req_map_sg(struct request *rq, struct scatterlist *sgl,
287                            int nsegs, unsigned bufflen, gfp_t gfp)
288 {
289         struct request_queue *q = rq->q;
290         int nr_pages = (bufflen + sgl[0].offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
291         unsigned int data_len = bufflen, len, bytes, off;
292         struct scatterlist *sg;
293         struct page *page;
294         struct bio *bio = NULL;
295         int i, err, nr_vecs = 0;
296
297         for_each_sg(sgl, sg, nsegs, i) {
298                 page = sg_page(sg);
299                 off = sg->offset;
300                 len = sg->length;
301                 data_len += len;
302
303                 while (len > 0 && data_len > 0) {
304                         /*
305                          * sg sends a scatterlist that is larger than
306                          * the data_len it wants transferred for certain
307                          * IO sizes
308                          */
309                         bytes = min_t(unsigned int, len, PAGE_SIZE - off);
310                         bytes = min(bytes, data_len);
311
312                         if (!bio) {
313                                 nr_vecs = min_t(int, BIO_MAX_PAGES, nr_pages);
314                                 nr_pages -= nr_vecs;
315
316                                 bio = bio_alloc(gfp, nr_vecs);
317                                 if (!bio) {
318                                         err = -ENOMEM;
319                                         goto free_bios;
320                                 }
321                                 bio->bi_end_io = scsi_bi_endio;
322                         }
323
324                         if (bio_add_pc_page(q, bio, page, bytes, off) !=
325                             bytes) {
326                                 bio_put(bio);
327                                 err = -EINVAL;
328                                 goto free_bios;
329                         }
330
331                         if (bio->bi_vcnt >= nr_vecs) {
332                                 err = scsi_merge_bio(rq, bio);
333                                 if (err) {
334                                         bio_endio(bio, 0);
335                                         goto free_bios;
336                                 }
337                                 bio = NULL;
338                         }
339
340                         page++;
341                         len -= bytes;
342                         data_len -=bytes;
343                         off = 0;
344                 }
345         }
346
347         rq->buffer = rq->data = NULL;
348         rq->data_len = bufflen;
349         return 0;
350
351 free_bios:
352         while ((bio = rq->bio) != NULL) {
353                 rq->bio = bio->bi_next;
354                 /*
355                  * call endio instead of bio_put incase it was bounced
356                  */
357                 bio_endio(bio, 0);
358         }
359
360         return err;
361 }
362
363 /**
364  * scsi_execute_async - insert request
365  * @sdev:       scsi device
366  * @cmd:        scsi command
367  * @cmd_len:    length of scsi cdb
368  * @data_direction: data direction
369  * @buffer:     data buffer (this can be a kernel buffer or scatterlist)
370  * @bufflen:    len of buffer
371  * @use_sg:     if buffer is a scatterlist this is the number of elements
372  * @timeout:    request timeout in seconds
373  * @retries:    number of times to retry request
374  * @flags:      or into request flags
375  **/
376 int scsi_execute_async(struct scsi_device *sdev, const unsigned char *cmd,
377                        int cmd_len, int data_direction, void *buffer, unsigned bufflen,
378                        int use_sg, int timeout, int retries, void *privdata,
379                        void (*done)(void *, char *, int, int), gfp_t gfp)
380 {
381         struct request *req;
382         struct scsi_io_context *sioc;
383         int err = 0;
384         int write = (data_direction == DMA_TO_DEVICE);
385
386         sioc = kmem_cache_zalloc(scsi_io_context_cache, gfp);
387         if (!sioc)
388                 return DRIVER_ERROR << 24;
389
390         req = blk_get_request(sdev->request_queue, write, gfp);
391         if (!req)
392                 goto free_sense;
393         req->cmd_type = REQ_TYPE_BLOCK_PC;
394         req->cmd_flags |= REQ_QUIET;
395
396         if (use_sg)
397                 err = scsi_req_map_sg(req, buffer, use_sg, bufflen, gfp);
398         else if (bufflen)
399                 err = blk_rq_map_kern(req->q, req, buffer, bufflen, gfp);
400
401         if (err)
402                 goto free_req;
403
404         req->cmd_len = cmd_len;
405         memset(req->cmd, 0, BLK_MAX_CDB); /* ATAPI hates garbage after CDB */
406         memcpy(req->cmd, cmd, req->cmd_len);
407         req->sense = sioc->sense;
408         req->sense_len = 0;
409         req->timeout = timeout;
410         req->retries = retries;
411         req->end_io_data = sioc;
412
413         sioc->data = privdata;
414         sioc->done = done;
415
416         blk_execute_rq_nowait(req->q, NULL, req, 1, scsi_end_async);
417         return 0;
418
419 free_req:
420         blk_put_request(req);
421 free_sense:
422         kmem_cache_free(scsi_io_context_cache, sioc);
423         return DRIVER_ERROR << 24;
424 }
425 EXPORT_SYMBOL_GPL(scsi_execute_async);
426
427 /*
428  * Function:    scsi_init_cmd_errh()
429  *
430  * Purpose:     Initialize cmd fields related to error handling.
431  *
432  * Arguments:   cmd     - command that is ready to be queued.
433  *
434  * Notes:       This function has the job of initializing a number of
435  *              fields related to error handling.   Typically this will
436  *              be called once for each command, as required.
437  */
438 static void scsi_init_cmd_errh(struct scsi_cmnd *cmd)
439 {
440         cmd->serial_number = 0;
441         cmd->resid = 0;
442         memset(cmd->sense_buffer, 0, sizeof cmd->sense_buffer);
443         if (cmd->cmd_len == 0)
444                 cmd->cmd_len = COMMAND_SIZE(cmd->cmnd[0]);
445 }
446
447 void scsi_device_unbusy(struct scsi_device *sdev)
448 {
449         struct Scsi_Host *shost = sdev->host;
450         unsigned long flags;
451
452         spin_lock_irqsave(shost->host_lock, flags);
453         shost->host_busy--;
454         if (unlikely(scsi_host_in_recovery(shost) &&
455                      (shost->host_failed || shost->host_eh_scheduled)))
456                 scsi_eh_wakeup(shost);
457         spin_unlock(shost->host_lock);
458         spin_lock(sdev->request_queue->queue_lock);
459         sdev->device_busy--;
460         spin_unlock_irqrestore(sdev->request_queue->queue_lock, flags);
461 }
462
463 /*
464  * Called for single_lun devices on IO completion. Clear starget_sdev_user,
465  * and call blk_run_queue for all the scsi_devices on the target -
466  * including current_sdev first.
467  *
468  * Called with *no* scsi locks held.
469  */
470 static void scsi_single_lun_run(struct scsi_device *current_sdev)
471 {
472         struct Scsi_Host *shost = current_sdev->host;
473         struct scsi_device *sdev, *tmp;
474         struct scsi_target *starget = scsi_target(current_sdev);
475         unsigned long flags;
476
477         spin_lock_irqsave(shost->host_lock, flags);
478         starget->starget_sdev_user = NULL;
479         spin_unlock_irqrestore(shost->host_lock, flags);
480
481         /*
482          * Call blk_run_queue for all LUNs on the target, starting with
483          * current_sdev. We race with others (to set starget_sdev_user),
484          * but in most cases, we will be first. Ideally, each LU on the
485          * target would get some limited time or requests on the target.
486          */
487         blk_run_queue(current_sdev->request_queue);
488
489         spin_lock_irqsave(shost->host_lock, flags);
490         if (starget->starget_sdev_user)
491                 goto out;
492         list_for_each_entry_safe(sdev, tmp, &starget->devices,
493                         same_target_siblings) {
494                 if (sdev == current_sdev)
495                         continue;
496                 if (scsi_device_get(sdev))
497                         continue;
498
499                 spin_unlock_irqrestore(shost->host_lock, flags);
500                 blk_run_queue(sdev->request_queue);
501                 spin_lock_irqsave(shost->host_lock, flags);
502         
503                 scsi_device_put(sdev);
504         }
505  out:
506         spin_unlock_irqrestore(shost->host_lock, flags);
507 }
508
509 /*
510  * Function:    scsi_run_queue()
511  *
512  * Purpose:     Select a proper request queue to serve next
513  *
514  * Arguments:   q       - last request's queue
515  *
516  * Returns:     Nothing
517  *
518  * Notes:       The previous command was completely finished, start
519  *              a new one if possible.
520  */
521 static void scsi_run_queue(struct request_queue *q)
522 {
523         struct scsi_device *sdev = q->queuedata;
524         struct Scsi_Host *shost = sdev->host;
525         unsigned long flags;
526
527         if (sdev->single_lun)
528                 scsi_single_lun_run(sdev);
529
530         spin_lock_irqsave(shost->host_lock, flags);
531         while (!list_empty(&shost->starved_list) &&
532                !shost->host_blocked && !shost->host_self_blocked &&
533                 !((shost->can_queue > 0) &&
534                   (shost->host_busy >= shost->can_queue))) {
535                 /*
536                  * As long as shost is accepting commands and we have
537                  * starved queues, call blk_run_queue. scsi_request_fn
538                  * drops the queue_lock and can add us back to the
539                  * starved_list.
540                  *
541                  * host_lock protects the starved_list and starved_entry.
542                  * scsi_request_fn must get the host_lock before checking
543                  * or modifying starved_list or starved_entry.
544                  */
545                 sdev = list_entry(shost->starved_list.next,
546                                           struct scsi_device, starved_entry);
547                 list_del_init(&sdev->starved_entry);
548                 spin_unlock_irqrestore(shost->host_lock, flags);
549
550
551                 if (test_bit(QUEUE_FLAG_REENTER, &q->queue_flags) &&
552                     !test_and_set_bit(QUEUE_FLAG_REENTER,
553                                       &sdev->request_queue->queue_flags)) {
554                         blk_run_queue(sdev->request_queue);
555                         clear_bit(QUEUE_FLAG_REENTER,
556                                   &sdev->request_queue->queue_flags);
557                 } else
558                         blk_run_queue(sdev->request_queue);
559
560                 spin_lock_irqsave(shost->host_lock, flags);
561                 if (unlikely(!list_empty(&sdev->starved_entry)))
562                         /*
563                          * sdev lost a race, and was put back on the
564                          * starved list. This is unlikely but without this
565                          * in theory we could loop forever.
566                          */
567                         break;
568         }
569         spin_unlock_irqrestore(shost->host_lock, flags);
570
571         blk_run_queue(q);
572 }
573
574 /*
575  * Function:    scsi_requeue_command()
576  *
577  * Purpose:     Handle post-processing of completed commands.
578  *
579  * Arguments:   q       - queue to operate on
580  *              cmd     - command that may need to be requeued.
581  *
582  * Returns:     Nothing
583  *
584  * Notes:       After command completion, there may be blocks left
585  *              over which weren't finished by the previous command
586  *              this can be for a number of reasons - the main one is
587  *              I/O errors in the middle of the request, in which case
588  *              we need to request the blocks that come after the bad
589  *              sector.
590  * Notes:       Upon return, cmd is a stale pointer.
591  */
592 static void scsi_requeue_command(struct request_queue *q, struct scsi_cmnd *cmd)
593 {
594         struct request *req = cmd->request;
595         unsigned long flags;
596
597         scsi_unprep_request(req);
598         spin_lock_irqsave(q->queue_lock, flags);
599         blk_requeue_request(q, req);
600         spin_unlock_irqrestore(q->queue_lock, flags);
601
602         scsi_run_queue(q);
603 }
604
605 void scsi_next_command(struct scsi_cmnd *cmd)
606 {
607         struct scsi_device *sdev = cmd->device;
608         struct request_queue *q = sdev->request_queue;
609
610         /* need to hold a reference on the device before we let go of the cmd */
611         get_device(&sdev->sdev_gendev);
612
613         scsi_put_command(cmd);
614         scsi_run_queue(q);
615
616         /* ok to remove device now */
617         put_device(&sdev->sdev_gendev);
618 }
619
620 void scsi_run_host_queues(struct Scsi_Host *shost)
621 {
622         struct scsi_device *sdev;
623
624         shost_for_each_device(sdev, shost)
625                 scsi_run_queue(sdev->request_queue);
626 }
627
628 /*
629  * Function:    scsi_end_request()
630  *
631  * Purpose:     Post-processing of completed commands (usually invoked at end
632  *              of upper level post-processing and scsi_io_completion).
633  *
634  * Arguments:   cmd      - command that is complete.
635  *              uptodate - 1 if I/O indicates success, <= 0 for I/O error.
636  *              bytes    - number of bytes of completed I/O
637  *              requeue  - indicates whether we should requeue leftovers.
638  *
639  * Lock status: Assumed that lock is not held upon entry.
640  *
641  * Returns:     cmd if requeue required, NULL otherwise.
642  *
643  * Notes:       This is called for block device requests in order to
644  *              mark some number of sectors as complete.
645  * 
646  *              We are guaranteeing that the request queue will be goosed
647  *              at some point during this call.
648  * Notes:       If cmd was requeued, upon return it will be a stale pointer.
649  */
650 static struct scsi_cmnd *scsi_end_request(struct scsi_cmnd *cmd, int uptodate,
651                                           int bytes, int requeue)
652 {
653         struct request_queue *q = cmd->device->request_queue;
654         struct request *req = cmd->request;
655         unsigned long flags;
656
657         /*
658          * If there are blocks left over at the end, set up the command
659          * to queue the remainder of them.
660          */
661         if (end_that_request_chunk(req, uptodate, bytes)) {
662                 int leftover = (req->hard_nr_sectors << 9);
663
664                 if (blk_pc_request(req))
665                         leftover = req->data_len;
666
667                 /* kill remainder if no retrys */
668                 if (!uptodate && blk_noretry_request(req))
669                         end_that_request_chunk(req, 0, leftover);
670                 else {
671                         if (requeue) {
672                                 /*
673                                  * Bleah.  Leftovers again.  Stick the
674                                  * leftovers in the front of the
675                                  * queue, and goose the queue again.
676                                  */
677                                 scsi_requeue_command(q, cmd);
678                                 cmd = NULL;
679                         }
680                         return cmd;
681                 }
682         }
683
684         add_disk_randomness(req->rq_disk);
685
686         spin_lock_irqsave(q->queue_lock, flags);
687         if (blk_rq_tagged(req))
688                 blk_queue_end_tag(q, req);
689         end_that_request_last(req, uptodate);
690         spin_unlock_irqrestore(q->queue_lock, flags);
691
692         /*
693          * This will goose the queue request function at the end, so we don't
694          * need to worry about launching another command.
695          */
696         scsi_next_command(cmd);
697         return NULL;
698 }
699
700 /*
701  * Like SCSI_MAX_SG_SEGMENTS, but for archs that have sg chaining. This limit
702  * is totally arbitrary, a setting of 2048 will get you at least 8mb ios.
703  */
704 #define SCSI_MAX_SG_CHAIN_SEGMENTS      2048
705
706 static inline unsigned int scsi_sgtable_index(unsigned short nents)
707 {
708         unsigned int index;
709
710         switch (nents) {
711         case 1 ... 8:
712                 index = 0;
713                 break;
714         case 9 ... 16:
715                 index = 1;
716                 break;
717 #if (SCSI_MAX_SG_SEGMENTS > 16)
718         case 17 ... 32:
719                 index = 2;
720                 break;
721 #if (SCSI_MAX_SG_SEGMENTS > 32)
722         case 33 ... 64:
723                 index = 3;
724                 break;
725 #if (SCSI_MAX_SG_SEGMENTS > 64)
726         case 65 ... 128:
727                 index = 4;
728                 break;
729 #endif
730 #endif
731 #endif
732         default:
733                 printk(KERN_ERR "scsi: bad segment count=%d\n", nents);
734                 BUG();
735         }
736
737         return index;
738 }
739
740 struct scatterlist *scsi_alloc_sgtable(struct scsi_cmnd *cmd, gfp_t gfp_mask)
741 {
742         struct scsi_host_sg_pool *sgp;
743         struct scatterlist *sgl, *prev, *ret;
744         unsigned int index;
745         int this, left;
746
747         BUG_ON(!cmd->use_sg);
748
749         left = cmd->use_sg;
750         ret = prev = NULL;
751         do {
752                 this = left;
753                 if (this > SCSI_MAX_SG_SEGMENTS) {
754                         this = SCSI_MAX_SG_SEGMENTS - 1;
755                         index = SG_MEMPOOL_NR - 1;
756                 } else
757                         index = scsi_sgtable_index(this);
758
759                 left -= this;
760
761                 sgp = scsi_sg_pools + index;
762
763                 sgl = mempool_alloc(sgp->pool, gfp_mask);
764                 if (unlikely(!sgl))
765                         goto enomem;
766
767                 sg_init_table(sgl, sgp->size);
768
769                 /*
770                  * first loop through, set initial index and return value
771                  */
772                 if (!ret)
773                         ret = sgl;
774
775                 /*
776                  * chain previous sglist, if any. we know the previous
777                  * sglist must be the biggest one, or we would not have
778                  * ended up doing another loop.
779                  */
780                 if (prev)
781                         sg_chain(prev, SCSI_MAX_SG_SEGMENTS, sgl);
782
783                 /*
784                  * if we have nothing left, mark the last segment as
785                  * end-of-list
786                  */
787                 if (!left)
788                         sg_mark_end(&sgl[this - 1]);
789
790                 /*
791                  * don't allow subsequent mempool allocs to sleep, it would
792                  * violate the mempool principle.
793                  */
794                 gfp_mask &= ~__GFP_WAIT;
795                 gfp_mask |= __GFP_HIGH;
796                 prev = sgl;
797         } while (left);
798
799         /*
800          * ->use_sg may get modified after dma mapping has potentially
801          * shrunk the number of segments, so keep a copy of it for free.
802          */
803         cmd->__use_sg = cmd->use_sg;
804         return ret;
805 enomem:
806         if (ret) {
807                 /*
808                  * Free entries chained off ret. Since we were trying to
809                  * allocate another sglist, we know that all entries are of
810                  * the max size.
811                  */
812                 sgp = scsi_sg_pools + SG_MEMPOOL_NR - 1;
813                 prev = ret;
814                 ret = &ret[SCSI_MAX_SG_SEGMENTS - 1];
815
816                 while ((sgl = sg_chain_ptr(ret)) != NULL) {
817                         ret = &sgl[SCSI_MAX_SG_SEGMENTS - 1];
818                         mempool_free(sgl, sgp->pool);
819                 }
820
821                 mempool_free(prev, sgp->pool);
822         }
823         return NULL;
824 }
825
826 EXPORT_SYMBOL(scsi_alloc_sgtable);
827
828 void scsi_free_sgtable(struct scsi_cmnd *cmd)
829 {
830         struct scatterlist *sgl = cmd->request_buffer;
831         struct scsi_host_sg_pool *sgp;
832
833         /*
834          * if this is the biggest size sglist, check if we have
835          * chained parts we need to free
836          */
837         if (cmd->__use_sg > SCSI_MAX_SG_SEGMENTS) {
838                 unsigned short this, left;
839                 struct scatterlist *next;
840                 unsigned int index;
841
842                 left = cmd->__use_sg - (SCSI_MAX_SG_SEGMENTS - 1);
843                 next = sg_chain_ptr(&sgl[SCSI_MAX_SG_SEGMENTS - 1]);
844                 while (left && next) {
845                         sgl = next;
846                         this = left;
847                         if (this > SCSI_MAX_SG_SEGMENTS) {
848                                 this = SCSI_MAX_SG_SEGMENTS - 1;
849                                 index = SG_MEMPOOL_NR - 1;
850                         } else
851                                 index = scsi_sgtable_index(this);
852
853                         left -= this;
854
855                         sgp = scsi_sg_pools + index;
856
857                         if (left)
858                                 next = sg_chain_ptr(&sgl[sgp->size - 1]);
859
860                         mempool_free(sgl, sgp->pool);
861                 }
862
863                 /*
864                  * Restore original, will be freed below
865                  */
866                 sgl = cmd->request_buffer;
867                 sgp = scsi_sg_pools + SG_MEMPOOL_NR - 1;
868         } else
869                 sgp = scsi_sg_pools + scsi_sgtable_index(cmd->__use_sg);
870
871         mempool_free(sgl, sgp->pool);
872 }
873
874 EXPORT_SYMBOL(scsi_free_sgtable);
875
876 /*
877  * Function:    scsi_release_buffers()
878  *
879  * Purpose:     Completion processing for block device I/O requests.
880  *
881  * Arguments:   cmd     - command that we are bailing.
882  *
883  * Lock status: Assumed that no lock is held upon entry.
884  *
885  * Returns:     Nothing
886  *
887  * Notes:       In the event that an upper level driver rejects a
888  *              command, we must release resources allocated during
889  *              the __init_io() function.  Primarily this would involve
890  *              the scatter-gather table, and potentially any bounce
891  *              buffers.
892  */
893 static void scsi_release_buffers(struct scsi_cmnd *cmd)
894 {
895         if (cmd->use_sg)
896                 scsi_free_sgtable(cmd);
897
898         /*
899          * Zero these out.  They now point to freed memory, and it is
900          * dangerous to hang onto the pointers.
901          */
902         cmd->request_buffer = NULL;
903         cmd->request_bufflen = 0;
904 }
905
906 /*
907  * Function:    scsi_io_completion()
908  *
909  * Purpose:     Completion processing for block device I/O requests.
910  *
911  * Arguments:   cmd   - command that is finished.
912  *
913  * Lock status: Assumed that no lock is held upon entry.
914  *
915  * Returns:     Nothing
916  *
917  * Notes:       This function is matched in terms of capabilities to
918  *              the function that created the scatter-gather list.
919  *              In other words, if there are no bounce buffers
920  *              (the normal case for most drivers), we don't need
921  *              the logic to deal with cleaning up afterwards.
922  *
923  *              We must do one of several things here:
924  *
925  *              a) Call scsi_end_request.  This will finish off the
926  *                 specified number of sectors.  If we are done, the
927  *                 command block will be released, and the queue
928  *                 function will be goosed.  If we are not done, then
929  *                 scsi_end_request will directly goose the queue.
930  *
931  *              b) We can just use scsi_requeue_command() here.  This would
932  *                 be used if we just wanted to retry, for example.
933  */
934 void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes)
935 {
936         int result = cmd->result;
937         int this_count = cmd->request_bufflen;
938         struct request_queue *q = cmd->device->request_queue;
939         struct request *req = cmd->request;
940         int clear_errors = 1;
941         struct scsi_sense_hdr sshdr;
942         int sense_valid = 0;
943         int sense_deferred = 0;
944
945         scsi_release_buffers(cmd);
946
947         if (result) {
948                 sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
949                 if (sense_valid)
950                         sense_deferred = scsi_sense_is_deferred(&sshdr);
951         }
952
953         if (blk_pc_request(req)) { /* SG_IO ioctl from block level */
954                 req->errors = result;
955                 if (result) {
956                         clear_errors = 0;
957                         if (sense_valid && req->sense) {
958                                 /*
959                                  * SG_IO wants current and deferred errors
960                                  */
961                                 int len = 8 + cmd->sense_buffer[7];
962
963                                 if (len > SCSI_SENSE_BUFFERSIZE)
964                                         len = SCSI_SENSE_BUFFERSIZE;
965                                 memcpy(req->sense, cmd->sense_buffer,  len);
966                                 req->sense_len = len;
967                         }
968                 }
969                 req->data_len = cmd->resid;
970         }
971
972         /*
973          * Next deal with any sectors which we were able to correctly
974          * handle.
975          */
976         SCSI_LOG_HLCOMPLETE(1, printk("%ld sectors total, "
977                                       "%d bytes done.\n",
978                                       req->nr_sectors, good_bytes));
979         SCSI_LOG_HLCOMPLETE(1, printk("use_sg is %d\n", cmd->use_sg));
980
981         if (clear_errors)
982                 req->errors = 0;
983
984         /* A number of bytes were successfully read.  If there
985          * are leftovers and there is some kind of error
986          * (result != 0), retry the rest.
987          */
988         if (scsi_end_request(cmd, 1, good_bytes, result == 0) == NULL)
989                 return;
990
991         /* good_bytes = 0, or (inclusive) there were leftovers and
992          * result = 0, so scsi_end_request couldn't retry.
993          */
994         if (sense_valid && !sense_deferred) {
995                 switch (sshdr.sense_key) {
996                 case UNIT_ATTENTION:
997                         if (cmd->device->removable) {
998                                 /* Detected disc change.  Set a bit
999                                  * and quietly refuse further access.
1000                                  */
1001                                 cmd->device->changed = 1;
1002                                 scsi_end_request(cmd, 0, this_count, 1);
1003                                 return;
1004                         } else {
1005                                 /* Must have been a power glitch, or a
1006                                  * bus reset.  Could not have been a
1007                                  * media change, so we just retry the
1008                                  * request and see what happens.
1009                                  */
1010                                 scsi_requeue_command(q, cmd);
1011                                 return;
1012                         }
1013                         break;
1014                 case ILLEGAL_REQUEST:
1015                         /* If we had an ILLEGAL REQUEST returned, then
1016                          * we may have performed an unsupported
1017                          * command.  The only thing this should be
1018                          * would be a ten byte read where only a six
1019                          * byte read was supported.  Also, on a system
1020                          * where READ CAPACITY failed, we may have
1021                          * read past the end of the disk.
1022                          */
1023                         if ((cmd->device->use_10_for_rw &&
1024                             sshdr.asc == 0x20 && sshdr.ascq == 0x00) &&
1025                             (cmd->cmnd[0] == READ_10 ||
1026                              cmd->cmnd[0] == WRITE_10)) {
1027                                 cmd->device->use_10_for_rw = 0;
1028                                 /* This will cause a retry with a
1029                                  * 6-byte command.
1030                                  */
1031                                 scsi_requeue_command(q, cmd);
1032                                 return;
1033                         } else {
1034                                 scsi_end_request(cmd, 0, this_count, 1);
1035                                 return;
1036                         }
1037                         break;
1038                 case NOT_READY:
1039                         /* If the device is in the process of becoming
1040                          * ready, or has a temporary blockage, retry.
1041                          */
1042                         if (sshdr.asc == 0x04) {
1043                                 switch (sshdr.ascq) {
1044                                 case 0x01: /* becoming ready */
1045                                 case 0x04: /* format in progress */
1046                                 case 0x05: /* rebuild in progress */
1047                                 case 0x06: /* recalculation in progress */
1048                                 case 0x07: /* operation in progress */
1049                                 case 0x08: /* Long write in progress */
1050                                 case 0x09: /* self test in progress */
1051                                         scsi_requeue_command(q, cmd);
1052                                         return;
1053                                 default:
1054                                         break;
1055                                 }
1056                         }
1057                         if (!(req->cmd_flags & REQ_QUIET))
1058                                 scsi_cmd_print_sense_hdr(cmd,
1059                                                          "Device not ready",
1060                                                          &sshdr);
1061
1062                         scsi_end_request(cmd, 0, this_count, 1);
1063                         return;
1064                 case VOLUME_OVERFLOW:
1065                         if (!(req->cmd_flags & REQ_QUIET)) {
1066                                 scmd_printk(KERN_INFO, cmd,
1067                                             "Volume overflow, CDB: ");
1068                                 __scsi_print_command(cmd->cmnd);
1069                                 scsi_print_sense("", cmd);
1070                         }
1071                         /* See SSC3rXX or current. */
1072                         scsi_end_request(cmd, 0, this_count, 1);
1073                         return;
1074                 default:
1075                         break;
1076                 }
1077         }
1078         if (host_byte(result) == DID_RESET) {
1079                 /* Third party bus reset or reset for error recovery
1080                  * reasons.  Just retry the request and see what
1081                  * happens.
1082                  */
1083                 scsi_requeue_command(q, cmd);
1084                 return;
1085         }
1086         if (result) {
1087                 if (!(req->cmd_flags & REQ_QUIET)) {
1088                         scsi_print_result(cmd);
1089                         if (driver_byte(result) & DRIVER_SENSE)
1090                                 scsi_print_sense("", cmd);
1091                 }
1092         }
1093         scsi_end_request(cmd, 0, this_count, !result);
1094 }
1095
1096 /*
1097  * Function:    scsi_init_io()
1098  *
1099  * Purpose:     SCSI I/O initialize function.
1100  *
1101  * Arguments:   cmd   - Command descriptor we wish to initialize
1102  *
1103  * Returns:     0 on success
1104  *              BLKPREP_DEFER if the failure is retryable
1105  *              BLKPREP_KILL if the failure is fatal
1106  */
1107 static int scsi_init_io(struct scsi_cmnd *cmd)
1108 {
1109         struct request     *req = cmd->request;
1110         int                count;
1111
1112         /*
1113          * We used to not use scatter-gather for single segment request,
1114          * but now we do (it makes highmem I/O easier to support without
1115          * kmapping pages)
1116          */
1117         cmd->use_sg = req->nr_phys_segments;
1118
1119         /*
1120          * If sg table allocation fails, requeue request later.
1121          */
1122         cmd->request_buffer = scsi_alloc_sgtable(cmd, GFP_ATOMIC);
1123         if (unlikely(!cmd->request_buffer)) {
1124                 scsi_unprep_request(req);
1125                 return BLKPREP_DEFER;
1126         }
1127
1128         req->buffer = NULL;
1129         if (blk_pc_request(req))
1130                 cmd->request_bufflen = req->data_len;
1131         else
1132                 cmd->request_bufflen = req->nr_sectors << 9;
1133
1134         /* 
1135          * Next, walk the list, and fill in the addresses and sizes of
1136          * each segment.
1137          */
1138         count = blk_rq_map_sg(req->q, req, cmd->request_buffer);
1139         if (likely(count <= cmd->use_sg)) {
1140                 cmd->use_sg = count;
1141                 return BLKPREP_OK;
1142         }
1143
1144         printk(KERN_ERR "Incorrect number of segments after building list\n");
1145         printk(KERN_ERR "counted %d, received %d\n", count, cmd->use_sg);
1146         printk(KERN_ERR "req nr_sec %lu, cur_nr_sec %u\n", req->nr_sectors,
1147                         req->current_nr_sectors);
1148
1149         return BLKPREP_KILL;
1150 }
1151
1152 static struct scsi_cmnd *scsi_get_cmd_from_req(struct scsi_device *sdev,
1153                 struct request *req)
1154 {
1155         struct scsi_cmnd *cmd;
1156
1157         if (!req->special) {
1158                 cmd = scsi_get_command(sdev, GFP_ATOMIC);
1159                 if (unlikely(!cmd))
1160                         return NULL;
1161                 req->special = cmd;
1162         } else {
1163                 cmd = req->special;
1164         }
1165
1166         /* pull a tag out of the request if we have one */
1167         cmd->tag = req->tag;
1168         cmd->request = req;
1169
1170         return cmd;
1171 }
1172
1173 int scsi_setup_blk_pc_cmnd(struct scsi_device *sdev, struct request *req)
1174 {
1175         struct scsi_cmnd *cmd;
1176         int ret = scsi_prep_state_check(sdev, req);
1177
1178         if (ret != BLKPREP_OK)
1179                 return ret;
1180
1181         cmd = scsi_get_cmd_from_req(sdev, req);
1182         if (unlikely(!cmd))
1183                 return BLKPREP_DEFER;
1184
1185         /*
1186          * BLOCK_PC requests may transfer data, in which case they must
1187          * a bio attached to them.  Or they might contain a SCSI command
1188          * that does not transfer data, in which case they may optionally
1189          * submit a request without an attached bio.
1190          */
1191         if (req->bio) {
1192                 int ret;
1193
1194                 BUG_ON(!req->nr_phys_segments);
1195
1196                 ret = scsi_init_io(cmd);
1197                 if (unlikely(ret))
1198                         return ret;
1199         } else {
1200                 BUG_ON(req->data_len);
1201                 BUG_ON(req->data);
1202
1203                 cmd->request_bufflen = 0;
1204                 cmd->request_buffer = NULL;
1205                 cmd->use_sg = 0;
1206                 req->buffer = NULL;
1207         }
1208
1209         BUILD_BUG_ON(sizeof(req->cmd) > sizeof(cmd->cmnd));
1210         memcpy(cmd->cmnd, req->cmd, sizeof(cmd->cmnd));
1211         cmd->cmd_len = req->cmd_len;
1212         if (!req->data_len)
1213                 cmd->sc_data_direction = DMA_NONE;
1214         else if (rq_data_dir(req) == WRITE)
1215                 cmd->sc_data_direction = DMA_TO_DEVICE;
1216         else
1217                 cmd->sc_data_direction = DMA_FROM_DEVICE;
1218         
1219         cmd->transfersize = req->data_len;
1220         cmd->allowed = req->retries;
1221         cmd->timeout_per_command = req->timeout;
1222         return BLKPREP_OK;
1223 }
1224 EXPORT_SYMBOL(scsi_setup_blk_pc_cmnd);
1225
1226 /*
1227  * Setup a REQ_TYPE_FS command.  These are simple read/write request
1228  * from filesystems that still need to be translated to SCSI CDBs from
1229  * the ULD.
1230  */
1231 int scsi_setup_fs_cmnd(struct scsi_device *sdev, struct request *req)
1232 {
1233         struct scsi_cmnd *cmd;
1234         int ret = scsi_prep_state_check(sdev, req);
1235
1236         if (ret != BLKPREP_OK)
1237                 return ret;
1238         /*
1239          * Filesystem requests must transfer data.
1240          */
1241         BUG_ON(!req->nr_phys_segments);
1242
1243         cmd = scsi_get_cmd_from_req(sdev, req);
1244         if (unlikely(!cmd))
1245                 return BLKPREP_DEFER;
1246
1247         return scsi_init_io(cmd);
1248 }
1249 EXPORT_SYMBOL(scsi_setup_fs_cmnd);
1250
1251 int scsi_prep_state_check(struct scsi_device *sdev, struct request *req)
1252 {
1253         int ret = BLKPREP_OK;
1254
1255         /*
1256          * If the device is not in running state we will reject some
1257          * or all commands.
1258          */
1259         if (unlikely(sdev->sdev_state != SDEV_RUNNING)) {
1260                 switch (sdev->sdev_state) {
1261                 case SDEV_OFFLINE:
1262                         /*
1263                          * If the device is offline we refuse to process any
1264                          * commands.  The device must be brought online
1265                          * before trying any recovery commands.
1266                          */
1267                         sdev_printk(KERN_ERR, sdev,
1268                                     "rejecting I/O to offline device\n");
1269                         ret = BLKPREP_KILL;
1270                         break;
1271                 case SDEV_DEL:
1272                         /*
1273                          * If the device is fully deleted, we refuse to
1274                          * process any commands as well.
1275                          */
1276                         sdev_printk(KERN_ERR, sdev,
1277                                     "rejecting I/O to dead device\n");
1278                         ret = BLKPREP_KILL;
1279                         break;
1280                 case SDEV_QUIESCE:
1281                 case SDEV_BLOCK:
1282                         /*
1283                          * If the devices is blocked we defer normal commands.
1284                          */
1285                         if (!(req->cmd_flags & REQ_PREEMPT))
1286                                 ret = BLKPREP_DEFER;
1287                         break;
1288                 default:
1289                         /*
1290                          * For any other not fully online state we only allow
1291                          * special commands.  In particular any user initiated
1292                          * command is not allowed.
1293                          */
1294                         if (!(req->cmd_flags & REQ_PREEMPT))
1295                                 ret = BLKPREP_KILL;
1296                         break;
1297                 }
1298         }
1299         return ret;
1300 }
1301 EXPORT_SYMBOL(scsi_prep_state_check);
1302
1303 int scsi_prep_return(struct request_queue *q, struct request *req, int ret)
1304 {
1305         struct scsi_device *sdev = q->queuedata;
1306
1307         switch (ret) {
1308         case BLKPREP_KILL:
1309                 req->errors = DID_NO_CONNECT << 16;
1310                 /* release the command and kill it */
1311                 if (req->special) {
1312                         struct scsi_cmnd *cmd = req->special;
1313                         scsi_release_buffers(cmd);
1314                         scsi_put_command(cmd);
1315                         req->special = NULL;
1316                 }
1317                 break;
1318         case BLKPREP_DEFER:
1319                 /*
1320                  * If we defer, the elv_next_request() returns NULL, but the
1321                  * queue must be restarted, so we plug here if no returning
1322                  * command will automatically do that.
1323                  */
1324                 if (sdev->device_busy == 0)
1325                         blk_plug_device(q);
1326                 break;
1327         default:
1328                 req->cmd_flags |= REQ_DONTPREP;
1329         }
1330
1331         return ret;
1332 }
1333 EXPORT_SYMBOL(scsi_prep_return);
1334
1335 int scsi_prep_fn(struct request_queue *q, struct request *req)
1336 {
1337         struct scsi_device *sdev = q->queuedata;
1338         int ret = BLKPREP_KILL;
1339
1340         if (req->cmd_type == REQ_TYPE_BLOCK_PC)
1341                 ret = scsi_setup_blk_pc_cmnd(sdev, req);
1342         return scsi_prep_return(q, req, ret);
1343 }
1344
1345 /*
1346  * scsi_dev_queue_ready: if we can send requests to sdev, return 1 else
1347  * return 0.
1348  *
1349  * Called with the queue_lock held.
1350  */
1351 static inline int scsi_dev_queue_ready(struct request_queue *q,
1352                                   struct scsi_device *sdev)
1353 {
1354         if (sdev->device_busy >= sdev->queue_depth)
1355                 return 0;
1356         if (sdev->device_busy == 0 && sdev->device_blocked) {
1357                 /*
1358                  * unblock after device_blocked iterates to zero
1359                  */
1360                 if (--sdev->device_blocked == 0) {
1361                         SCSI_LOG_MLQUEUE(3,
1362                                    sdev_printk(KERN_INFO, sdev,
1363                                    "unblocking device at zero depth\n"));
1364                 } else {
1365                         blk_plug_device(q);
1366                         return 0;
1367                 }
1368         }
1369         if (sdev->device_blocked)
1370                 return 0;
1371
1372         return 1;
1373 }
1374
1375 /*
1376  * scsi_host_queue_ready: if we can send requests to shost, return 1 else
1377  * return 0. We must end up running the queue again whenever 0 is
1378  * returned, else IO can hang.
1379  *
1380  * Called with host_lock held.
1381  */
1382 static inline int scsi_host_queue_ready(struct request_queue *q,
1383                                    struct Scsi_Host *shost,
1384                                    struct scsi_device *sdev)
1385 {
1386         if (scsi_host_in_recovery(shost))
1387                 return 0;
1388         if (shost->host_busy == 0 && shost->host_blocked) {
1389                 /*
1390                  * unblock after host_blocked iterates to zero
1391                  */
1392                 if (--shost->host_blocked == 0) {
1393                         SCSI_LOG_MLQUEUE(3,
1394                                 printk("scsi%d unblocking host at zero depth\n",
1395                                         shost->host_no));
1396                 } else {
1397                         blk_plug_device(q);
1398                         return 0;
1399                 }
1400         }
1401         if ((shost->can_queue > 0 && shost->host_busy >= shost->can_queue) ||
1402             shost->host_blocked || shost->host_self_blocked) {
1403                 if (list_empty(&sdev->starved_entry))
1404                         list_add_tail(&sdev->starved_entry, &shost->starved_list);
1405                 return 0;
1406         }
1407
1408         /* We're OK to process the command, so we can't be starved */
1409         if (!list_empty(&sdev->starved_entry))
1410                 list_del_init(&sdev->starved_entry);
1411
1412         return 1;
1413 }
1414
1415 /*
1416  * Kill a request for a dead device
1417  */
1418 static void scsi_kill_request(struct request *req, struct request_queue *q)
1419 {
1420         struct scsi_cmnd *cmd = req->special;
1421         struct scsi_device *sdev = cmd->device;
1422         struct Scsi_Host *shost = sdev->host;
1423
1424         blkdev_dequeue_request(req);
1425
1426         if (unlikely(cmd == NULL)) {
1427                 printk(KERN_CRIT "impossible request in %s.\n",
1428                                  __FUNCTION__);
1429                 BUG();
1430         }
1431
1432         scsi_init_cmd_errh(cmd);
1433         cmd->result = DID_NO_CONNECT << 16;
1434         atomic_inc(&cmd->device->iorequest_cnt);
1435
1436         /*
1437          * SCSI request completion path will do scsi_device_unbusy(),
1438          * bump busy counts.  To bump the counters, we need to dance
1439          * with the locks as normal issue path does.
1440          */
1441         sdev->device_busy++;
1442         spin_unlock(sdev->request_queue->queue_lock);
1443         spin_lock(shost->host_lock);
1444         shost->host_busy++;
1445         spin_unlock(shost->host_lock);
1446         spin_lock(sdev->request_queue->queue_lock);
1447
1448         __scsi_done(cmd);
1449 }
1450
1451 static void scsi_softirq_done(struct request *rq)
1452 {
1453         struct scsi_cmnd *cmd = rq->completion_data;
1454         unsigned long wait_for = (cmd->allowed + 1) * cmd->timeout_per_command;
1455         int disposition;
1456
1457         INIT_LIST_HEAD(&cmd->eh_entry);
1458
1459         disposition = scsi_decide_disposition(cmd);
1460         if (disposition != SUCCESS &&
1461             time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) {
1462                 sdev_printk(KERN_ERR, cmd->device,
1463                             "timing out command, waited %lus\n",
1464                             wait_for/HZ);
1465                 disposition = SUCCESS;
1466         }
1467                         
1468         scsi_log_completion(cmd, disposition);
1469
1470         switch (disposition) {
1471                 case SUCCESS:
1472                         scsi_finish_command(cmd);
1473                         break;
1474                 case NEEDS_RETRY:
1475                         scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY);
1476                         break;
1477                 case ADD_TO_MLQUEUE:
1478                         scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY);
1479                         break;
1480                 default:
1481                         if (!scsi_eh_scmd_add(cmd, 0))
1482                                 scsi_finish_command(cmd);
1483         }
1484 }
1485
1486 /*
1487  * Function:    scsi_request_fn()
1488  *
1489  * Purpose:     Main strategy routine for SCSI.
1490  *
1491  * Arguments:   q       - Pointer to actual queue.
1492  *
1493  * Returns:     Nothing
1494  *
1495  * Lock status: IO request lock assumed to be held when called.
1496  */
1497 static void scsi_request_fn(struct request_queue *q)
1498 {
1499         struct scsi_device *sdev = q->queuedata;
1500         struct Scsi_Host *shost;
1501         struct scsi_cmnd *cmd;
1502         struct request *req;
1503
1504         if (!sdev) {
1505                 printk("scsi: killing requests for dead queue\n");
1506                 while ((req = elv_next_request(q)) != NULL)
1507                         scsi_kill_request(req, q);
1508                 return;
1509         }
1510
1511         if(!get_device(&sdev->sdev_gendev))
1512                 /* We must be tearing the block queue down already */
1513                 return;
1514
1515         /*
1516          * To start with, we keep looping until the queue is empty, or until
1517          * the host is no longer able to accept any more requests.
1518          */
1519         shost = sdev->host;
1520         while (!blk_queue_plugged(q)) {
1521                 int rtn;
1522                 /*
1523                  * get next queueable request.  We do this early to make sure
1524                  * that the request is fully prepared even if we cannot 
1525                  * accept it.
1526                  */
1527                 req = elv_next_request(q);
1528                 if (!req || !scsi_dev_queue_ready(q, sdev))
1529                         break;
1530
1531                 if (unlikely(!scsi_device_online(sdev))) {
1532                         sdev_printk(KERN_ERR, sdev,
1533                                     "rejecting I/O to offline device\n");
1534                         scsi_kill_request(req, q);
1535                         continue;
1536                 }
1537
1538
1539                 /*
1540                  * Remove the request from the request list.
1541                  */
1542                 if (!(blk_queue_tagged(q) && !blk_queue_start_tag(q, req)))
1543                         blkdev_dequeue_request(req);
1544                 sdev->device_busy++;
1545
1546                 spin_unlock(q->queue_lock);
1547                 cmd = req->special;
1548                 if (unlikely(cmd == NULL)) {
1549                         printk(KERN_CRIT "impossible request in %s.\n"
1550                                          "please mail a stack trace to "
1551                                          "linux-scsi@vger.kernel.org\n",
1552                                          __FUNCTION__);
1553                         blk_dump_rq_flags(req, "foo");
1554                         BUG();
1555                 }
1556                 spin_lock(shost->host_lock);
1557
1558                 if (!scsi_host_queue_ready(q, shost, sdev))
1559                         goto not_ready;
1560                 if (sdev->single_lun) {
1561                         if (scsi_target(sdev)->starget_sdev_user &&
1562                             scsi_target(sdev)->starget_sdev_user != sdev)
1563                                 goto not_ready;
1564                         scsi_target(sdev)->starget_sdev_user = sdev;
1565                 }
1566                 shost->host_busy++;
1567
1568                 /*
1569                  * XXX(hch): This is rather suboptimal, scsi_dispatch_cmd will
1570                  *              take the lock again.
1571                  */
1572                 spin_unlock_irq(shost->host_lock);
1573
1574                 /*
1575                  * Finally, initialize any error handling parameters, and set up
1576                  * the timers for timeouts.
1577                  */
1578                 scsi_init_cmd_errh(cmd);
1579
1580                 /*
1581                  * Dispatch the command to the low-level driver.
1582                  */
1583                 rtn = scsi_dispatch_cmd(cmd);
1584                 spin_lock_irq(q->queue_lock);
1585                 if(rtn) {
1586                         /* we're refusing the command; because of
1587                          * the way locks get dropped, we need to 
1588                          * check here if plugging is required */
1589                         if(sdev->device_busy == 0)
1590                                 blk_plug_device(q);
1591
1592                         break;
1593                 }
1594         }
1595
1596         goto out;
1597
1598  not_ready:
1599         spin_unlock_irq(shost->host_lock);
1600
1601         /*
1602          * lock q, handle tag, requeue req, and decrement device_busy. We
1603          * must return with queue_lock held.
1604          *
1605          * Decrementing device_busy without checking it is OK, as all such
1606          * cases (host limits or settings) should run the queue at some
1607          * later time.
1608          */
1609         spin_lock_irq(q->queue_lock);
1610         blk_requeue_request(q, req);
1611         sdev->device_busy--;
1612         if(sdev->device_busy == 0)
1613                 blk_plug_device(q);
1614  out:
1615         /* must be careful here...if we trigger the ->remove() function
1616          * we cannot be holding the q lock */
1617         spin_unlock_irq(q->queue_lock);
1618         put_device(&sdev->sdev_gendev);
1619         spin_lock_irq(q->queue_lock);
1620 }
1621
1622 u64 scsi_calculate_bounce_limit(struct Scsi_Host *shost)
1623 {
1624         struct device *host_dev;
1625         u64 bounce_limit = 0xffffffff;
1626
1627         if (shost->unchecked_isa_dma)
1628                 return BLK_BOUNCE_ISA;
1629         /*
1630          * Platforms with virtual-DMA translation
1631          * hardware have no practical limit.
1632          */
1633         if (!PCI_DMA_BUS_IS_PHYS)
1634                 return BLK_BOUNCE_ANY;
1635
1636         host_dev = scsi_get_device(shost);
1637         if (host_dev && host_dev->dma_mask)
1638                 bounce_limit = *host_dev->dma_mask;
1639
1640         return bounce_limit;
1641 }
1642 EXPORT_SYMBOL(scsi_calculate_bounce_limit);
1643
1644 struct request_queue *__scsi_alloc_queue(struct Scsi_Host *shost,
1645                                          request_fn_proc *request_fn)
1646 {
1647         struct request_queue *q;
1648
1649         q = blk_init_queue(request_fn, NULL);
1650         if (!q)
1651                 return NULL;
1652
1653         /*
1654          * this limit is imposed by hardware restrictions
1655          */
1656         blk_queue_max_hw_segments(q, shost->sg_tablesize);
1657
1658         /*
1659          * In the future, sg chaining support will be mandatory and this
1660          * ifdef can then go away. Right now we don't have all archs
1661          * converted, so better keep it safe.
1662          */
1663 #ifdef ARCH_HAS_SG_CHAIN
1664         if (shost->use_sg_chaining)
1665                 blk_queue_max_phys_segments(q, SCSI_MAX_SG_CHAIN_SEGMENTS);
1666         else
1667                 blk_queue_max_phys_segments(q, SCSI_MAX_SG_SEGMENTS);
1668 #else
1669         blk_queue_max_phys_segments(q, SCSI_MAX_SG_SEGMENTS);
1670 #endif
1671
1672         blk_queue_max_sectors(q, shost->max_sectors);
1673         blk_queue_bounce_limit(q, scsi_calculate_bounce_limit(shost));
1674         blk_queue_segment_boundary(q, shost->dma_boundary);
1675
1676         if (!shost->use_clustering)
1677                 clear_bit(QUEUE_FLAG_CLUSTER, &q->queue_flags);
1678         return q;
1679 }
1680 EXPORT_SYMBOL(__scsi_alloc_queue);
1681
1682 struct request_queue *scsi_alloc_queue(struct scsi_device *sdev)
1683 {
1684         struct request_queue *q;
1685
1686         q = __scsi_alloc_queue(sdev->host, scsi_request_fn);
1687         if (!q)
1688                 return NULL;
1689
1690         blk_queue_prep_rq(q, scsi_prep_fn);
1691         blk_queue_softirq_done(q, scsi_softirq_done);
1692         return q;
1693 }
1694
1695 void scsi_free_queue(struct request_queue *q)
1696 {
1697         blk_cleanup_queue(q);
1698 }
1699
1700 /*
1701  * Function:    scsi_block_requests()
1702  *
1703  * Purpose:     Utility function used by low-level drivers to prevent further
1704  *              commands from being queued to the device.
1705  *
1706  * Arguments:   shost       - Host in question
1707  *
1708  * Returns:     Nothing
1709  *
1710  * Lock status: No locks are assumed held.
1711  *
1712  * Notes:       There is no timer nor any other means by which the requests
1713  *              get unblocked other than the low-level driver calling
1714  *              scsi_unblock_requests().
1715  */
1716 void scsi_block_requests(struct Scsi_Host *shost)
1717 {
1718         shost->host_self_blocked = 1;
1719 }
1720 EXPORT_SYMBOL(scsi_block_requests);
1721
1722 /*
1723  * Function:    scsi_unblock_requests()
1724  *
1725  * Purpose:     Utility function used by low-level drivers to allow further
1726  *              commands from being queued to the device.
1727  *
1728  * Arguments:   shost       - Host in question
1729  *
1730  * Returns:     Nothing
1731  *
1732  * Lock status: No locks are assumed held.
1733  *
1734  * Notes:       There is no timer nor any other means by which the requests
1735  *              get unblocked other than the low-level driver calling
1736  *              scsi_unblock_requests().
1737  *
1738  *              This is done as an API function so that changes to the
1739  *              internals of the scsi mid-layer won't require wholesale
1740  *              changes to drivers that use this feature.
1741  */
1742 void scsi_unblock_requests(struct Scsi_Host *shost)
1743 {
1744         shost->host_self_blocked = 0;
1745         scsi_run_host_queues(shost);
1746 }
1747 EXPORT_SYMBOL(scsi_unblock_requests);
1748
1749 int __init scsi_init_queue(void)
1750 {
1751         int i;
1752
1753         scsi_io_context_cache = kmem_cache_create("scsi_io_context",
1754                                         sizeof(struct scsi_io_context),
1755                                         0, 0, NULL);
1756         if (!scsi_io_context_cache) {
1757                 printk(KERN_ERR "SCSI: can't init scsi io context cache\n");
1758                 return -ENOMEM;
1759         }
1760
1761         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1762                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1763                 int size = sgp->size * sizeof(struct scatterlist);
1764
1765                 sgp->slab = kmem_cache_create(sgp->name, size, 0,
1766                                 SLAB_HWCACHE_ALIGN, NULL);
1767                 if (!sgp->slab) {
1768                         printk(KERN_ERR "SCSI: can't init sg slab %s\n",
1769                                         sgp->name);
1770                 }
1771
1772                 sgp->pool = mempool_create_slab_pool(SG_MEMPOOL_SIZE,
1773                                                      sgp->slab);
1774                 if (!sgp->pool) {
1775                         printk(KERN_ERR "SCSI: can't init sg mempool %s\n",
1776                                         sgp->name);
1777                 }
1778         }
1779
1780         return 0;
1781 }
1782
1783 void scsi_exit_queue(void)
1784 {
1785         int i;
1786
1787         kmem_cache_destroy(scsi_io_context_cache);
1788
1789         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1790                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1791                 mempool_destroy(sgp->pool);
1792                 kmem_cache_destroy(sgp->slab);
1793         }
1794 }
1795
1796 /**
1797  *      scsi_mode_select - issue a mode select
1798  *      @sdev:  SCSI device to be queried
1799  *      @pf:    Page format bit (1 == standard, 0 == vendor specific)
1800  *      @sp:    Save page bit (0 == don't save, 1 == save)
1801  *      @modepage: mode page being requested
1802  *      @buffer: request buffer (may not be smaller than eight bytes)
1803  *      @len:   length of request buffer.
1804  *      @timeout: command timeout
1805  *      @retries: number of retries before failing
1806  *      @data: returns a structure abstracting the mode header data
1807  *      @sense: place to put sense data (or NULL if no sense to be collected).
1808  *              must be SCSI_SENSE_BUFFERSIZE big.
1809  *
1810  *      Returns zero if successful; negative error number or scsi
1811  *      status on error
1812  *
1813  */
1814 int
1815 scsi_mode_select(struct scsi_device *sdev, int pf, int sp, int modepage,
1816                  unsigned char *buffer, int len, int timeout, int retries,
1817                  struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1818 {
1819         unsigned char cmd[10];
1820         unsigned char *real_buffer;
1821         int ret;
1822
1823         memset(cmd, 0, sizeof(cmd));
1824         cmd[1] = (pf ? 0x10 : 0) | (sp ? 0x01 : 0);
1825
1826         if (sdev->use_10_for_ms) {
1827                 if (len > 65535)
1828                         return -EINVAL;
1829                 real_buffer = kmalloc(8 + len, GFP_KERNEL);
1830                 if (!real_buffer)
1831                         return -ENOMEM;
1832                 memcpy(real_buffer + 8, buffer, len);
1833                 len += 8;
1834                 real_buffer[0] = 0;
1835                 real_buffer[1] = 0;
1836                 real_buffer[2] = data->medium_type;
1837                 real_buffer[3] = data->device_specific;
1838                 real_buffer[4] = data->longlba ? 0x01 : 0;
1839                 real_buffer[5] = 0;
1840                 real_buffer[6] = data->block_descriptor_length >> 8;
1841                 real_buffer[7] = data->block_descriptor_length;
1842
1843                 cmd[0] = MODE_SELECT_10;
1844                 cmd[7] = len >> 8;
1845                 cmd[8] = len;
1846         } else {
1847                 if (len > 255 || data->block_descriptor_length > 255 ||
1848                     data->longlba)
1849                         return -EINVAL;
1850
1851                 real_buffer = kmalloc(4 + len, GFP_KERNEL);
1852                 if (!real_buffer)
1853                         return -ENOMEM;
1854                 memcpy(real_buffer + 4, buffer, len);
1855                 len += 4;
1856                 real_buffer[0] = 0;
1857                 real_buffer[1] = data->medium_type;
1858                 real_buffer[2] = data->device_specific;
1859                 real_buffer[3] = data->block_descriptor_length;
1860                 
1861
1862                 cmd[0] = MODE_SELECT;
1863                 cmd[4] = len;
1864         }
1865
1866         ret = scsi_execute_req(sdev, cmd, DMA_TO_DEVICE, real_buffer, len,
1867                                sshdr, timeout, retries);
1868         kfree(real_buffer);
1869         return ret;
1870 }
1871 EXPORT_SYMBOL_GPL(scsi_mode_select);
1872
1873 /**
1874  *      scsi_mode_sense - issue a mode sense, falling back from 10 to 
1875  *              six bytes if necessary.
1876  *      @sdev:  SCSI device to be queried
1877  *      @dbd:   set if mode sense will allow block descriptors to be returned
1878  *      @modepage: mode page being requested
1879  *      @buffer: request buffer (may not be smaller than eight bytes)
1880  *      @len:   length of request buffer.
1881  *      @timeout: command timeout
1882  *      @retries: number of retries before failing
1883  *      @data: returns a structure abstracting the mode header data
1884  *      @sense: place to put sense data (or NULL if no sense to be collected).
1885  *              must be SCSI_SENSE_BUFFERSIZE big.
1886  *
1887  *      Returns zero if unsuccessful, or the header offset (either 4
1888  *      or 8 depending on whether a six or ten byte command was
1889  *      issued) if successful.
1890  **/
1891 int
1892 scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
1893                   unsigned char *buffer, int len, int timeout, int retries,
1894                   struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1895 {
1896         unsigned char cmd[12];
1897         int use_10_for_ms;
1898         int header_length;
1899         int result;
1900         struct scsi_sense_hdr my_sshdr;
1901
1902         memset(data, 0, sizeof(*data));
1903         memset(&cmd[0], 0, 12);
1904         cmd[1] = dbd & 0x18;    /* allows DBD and LLBA bits */
1905         cmd[2] = modepage;
1906
1907         /* caller might not be interested in sense, but we need it */
1908         if (!sshdr)
1909                 sshdr = &my_sshdr;
1910
1911  retry:
1912         use_10_for_ms = sdev->use_10_for_ms;
1913
1914         if (use_10_for_ms) {
1915                 if (len < 8)
1916                         len = 8;
1917
1918                 cmd[0] = MODE_SENSE_10;
1919                 cmd[8] = len;
1920                 header_length = 8;
1921         } else {
1922                 if (len < 4)
1923                         len = 4;
1924
1925                 cmd[0] = MODE_SENSE;
1926                 cmd[4] = len;
1927                 header_length = 4;
1928         }
1929
1930         memset(buffer, 0, len);
1931
1932         result = scsi_execute_req(sdev, cmd, DMA_FROM_DEVICE, buffer, len,
1933                                   sshdr, timeout, retries);
1934
1935         /* This code looks awful: what it's doing is making sure an
1936          * ILLEGAL REQUEST sense return identifies the actual command
1937          * byte as the problem.  MODE_SENSE commands can return
1938          * ILLEGAL REQUEST if the code page isn't supported */
1939
1940         if (use_10_for_ms && !scsi_status_is_good(result) &&
1941             (driver_byte(result) & DRIVER_SENSE)) {
1942                 if (scsi_sense_valid(sshdr)) {
1943                         if ((sshdr->sense_key == ILLEGAL_REQUEST) &&
1944                             (sshdr->asc == 0x20) && (sshdr->ascq == 0)) {
1945                                 /* 
1946                                  * Invalid command operation code
1947                                  */
1948                                 sdev->use_10_for_ms = 0;
1949                                 goto retry;
1950                         }
1951                 }
1952         }
1953
1954         if(scsi_status_is_good(result)) {
1955                 if (unlikely(buffer[0] == 0x86 && buffer[1] == 0x0b &&
1956                              (modepage == 6 || modepage == 8))) {
1957                         /* Initio breakage? */
1958                         header_length = 0;
1959                         data->length = 13;
1960                         data->medium_type = 0;
1961                         data->device_specific = 0;
1962                         data->longlba = 0;
1963                         data->block_descriptor_length = 0;
1964                 } else if(use_10_for_ms) {
1965                         data->length = buffer[0]*256 + buffer[1] + 2;
1966                         data->medium_type = buffer[2];
1967                         data->device_specific = buffer[3];
1968                         data->longlba = buffer[4] & 0x01;
1969                         data->block_descriptor_length = buffer[6]*256
1970                                 + buffer[7];
1971                 } else {
1972                         data->length = buffer[0] + 1;
1973                         data->medium_type = buffer[1];
1974                         data->device_specific = buffer[2];
1975                         data->block_descriptor_length = buffer[3];
1976                 }
1977                 data->header_length = header_length;
1978         }
1979
1980         return result;
1981 }
1982 EXPORT_SYMBOL(scsi_mode_sense);
1983
1984 int
1985 scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries)
1986 {
1987         char cmd[] = {
1988                 TEST_UNIT_READY, 0, 0, 0, 0, 0,
1989         };
1990         struct scsi_sense_hdr sshdr;
1991         int result;
1992         
1993         result = scsi_execute_req(sdev, cmd, DMA_NONE, NULL, 0, &sshdr,
1994                                   timeout, retries);
1995
1996         if ((driver_byte(result) & DRIVER_SENSE) && sdev->removable) {
1997
1998                 if ((scsi_sense_valid(&sshdr)) &&
1999                     ((sshdr.sense_key == UNIT_ATTENTION) ||
2000                      (sshdr.sense_key == NOT_READY))) {
2001                         sdev->changed = 1;
2002                         result = 0;
2003                 }
2004         }
2005         return result;
2006 }
2007 EXPORT_SYMBOL(scsi_test_unit_ready);
2008
2009 /**
2010  *      scsi_device_set_state - Take the given device through the device
2011  *              state model.
2012  *      @sdev:  scsi device to change the state of.
2013  *      @state: state to change to.
2014  *
2015  *      Returns zero if unsuccessful or an error if the requested 
2016  *      transition is illegal.
2017  **/
2018 int
2019 scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
2020 {
2021         enum scsi_device_state oldstate = sdev->sdev_state;
2022
2023         if (state == oldstate)
2024                 return 0;
2025
2026         switch (state) {
2027         case SDEV_CREATED:
2028                 /* There are no legal states that come back to
2029                  * created.  This is the manually initialised start
2030                  * state */
2031                 goto illegal;
2032                         
2033         case SDEV_RUNNING:
2034                 switch (oldstate) {
2035                 case SDEV_CREATED:
2036                 case SDEV_OFFLINE:
2037                 case SDEV_QUIESCE:
2038                 case SDEV_BLOCK:
2039                         break;
2040                 default:
2041                         goto illegal;
2042                 }
2043                 break;
2044
2045         case SDEV_QUIESCE:
2046                 switch (oldstate) {
2047                 case SDEV_RUNNING:
2048                 case SDEV_OFFLINE:
2049                         break;
2050                 default:
2051                         goto illegal;
2052                 }
2053                 break;
2054
2055         case SDEV_OFFLINE:
2056                 switch (oldstate) {
2057                 case SDEV_CREATED:
2058                 case SDEV_RUNNING:
2059                 case SDEV_QUIESCE:
2060                 case SDEV_BLOCK:
2061                         break;
2062                 default:
2063                         goto illegal;
2064                 }
2065                 break;
2066
2067         case SDEV_BLOCK:
2068                 switch (oldstate) {
2069                 case SDEV_CREATED:
2070                 case SDEV_RUNNING:
2071                         break;
2072                 default:
2073                         goto illegal;
2074                 }
2075                 break;
2076
2077         case SDEV_CANCEL:
2078                 switch (oldstate) {
2079                 case SDEV_CREATED:
2080                 case SDEV_RUNNING:
2081                 case SDEV_QUIESCE:
2082                 case SDEV_OFFLINE:
2083                 case SDEV_BLOCK:
2084                         break;
2085                 default:
2086                         goto illegal;
2087                 }
2088                 break;
2089
2090         case SDEV_DEL:
2091                 switch (oldstate) {
2092                 case SDEV_CREATED:
2093                 case SDEV_RUNNING:
2094                 case SDEV_OFFLINE:
2095                 case SDEV_CANCEL:
2096                         break;
2097                 default:
2098                         goto illegal;
2099                 }
2100                 break;
2101
2102         }
2103         sdev->sdev_state = state;
2104         return 0;
2105
2106  illegal:
2107         SCSI_LOG_ERROR_RECOVERY(1, 
2108                                 sdev_printk(KERN_ERR, sdev,
2109                                             "Illegal state transition %s->%s\n",
2110                                             scsi_device_state_name(oldstate),
2111                                             scsi_device_state_name(state))
2112                                 );
2113         return -EINVAL;
2114 }
2115 EXPORT_SYMBOL(scsi_device_set_state);
2116
2117 /**
2118  *      sdev_evt_emit - emit a single SCSI device uevent
2119  *      @sdev: associated SCSI device
2120  *      @evt: event to emit
2121  *
2122  *      Send a single uevent (scsi_event) to the associated scsi_device.
2123  */
2124 static void scsi_evt_emit(struct scsi_device *sdev, struct scsi_event *evt)
2125 {
2126         int idx = 0;
2127         char *envp[3];
2128
2129         switch (evt->evt_type) {
2130         case SDEV_EVT_MEDIA_CHANGE:
2131                 envp[idx++] = "SDEV_MEDIA_CHANGE=1";
2132                 break;
2133
2134         default:
2135                 /* do nothing */
2136                 break;
2137         }
2138
2139         envp[idx++] = NULL;
2140
2141         kobject_uevent_env(&sdev->sdev_gendev.kobj, KOBJ_CHANGE, envp);
2142 }
2143
2144 /**
2145  *      sdev_evt_thread - send a uevent for each scsi event
2146  *      @work: work struct for scsi_device
2147  *
2148  *      Dispatch queued events to their associated scsi_device kobjects
2149  *      as uevents.
2150  */
2151 void scsi_evt_thread(struct work_struct *work)
2152 {
2153         struct scsi_device *sdev;
2154         LIST_HEAD(event_list);
2155
2156         sdev = container_of(work, struct scsi_device, event_work);
2157
2158         while (1) {
2159                 struct scsi_event *evt;
2160                 struct list_head *this, *tmp;
2161                 unsigned long flags;
2162
2163                 spin_lock_irqsave(&sdev->list_lock, flags);
2164                 list_splice_init(&sdev->event_list, &event_list);
2165                 spin_unlock_irqrestore(&sdev->list_lock, flags);
2166
2167                 if (list_empty(&event_list))
2168                         break;
2169
2170                 list_for_each_safe(this, tmp, &event_list) {
2171                         evt = list_entry(this, struct scsi_event, node);
2172                         list_del(&evt->node);
2173                         scsi_evt_emit(sdev, evt);
2174                         kfree(evt);
2175                 }
2176         }
2177 }
2178
2179 /**
2180  *      sdev_evt_send - send asserted event to uevent thread
2181  *      @sdev: scsi_device event occurred on
2182  *      @evt: event to send
2183  *
2184  *      Assert scsi device event asynchronously.
2185  */
2186 void sdev_evt_send(struct scsi_device *sdev, struct scsi_event *evt)
2187 {
2188         unsigned long flags;
2189
2190         if (!test_bit(evt->evt_type, sdev->supported_events)) {
2191                 kfree(evt);
2192                 return;
2193         }
2194
2195         spin_lock_irqsave(&sdev->list_lock, flags);
2196         list_add_tail(&evt->node, &sdev->event_list);
2197         schedule_work(&sdev->event_work);
2198         spin_unlock_irqrestore(&sdev->list_lock, flags);
2199 }
2200 EXPORT_SYMBOL_GPL(sdev_evt_send);
2201
2202 /**
2203  *      sdev_evt_alloc - allocate a new scsi event
2204  *      @evt_type: type of event to allocate
2205  *      @gfpflags: GFP flags for allocation
2206  *
2207  *      Allocates and returns a new scsi_event.
2208  */
2209 struct scsi_event *sdev_evt_alloc(enum scsi_device_event evt_type,
2210                                   gfp_t gfpflags)
2211 {
2212         struct scsi_event *evt = kzalloc(sizeof(struct scsi_event), gfpflags);
2213         if (!evt)
2214                 return NULL;
2215
2216         evt->evt_type = evt_type;
2217         INIT_LIST_HEAD(&evt->node);
2218
2219         /* evt_type-specific initialization, if any */
2220         switch (evt_type) {
2221         case SDEV_EVT_MEDIA_CHANGE:
2222         default:
2223                 /* do nothing */
2224                 break;
2225         }
2226
2227         return evt;
2228 }
2229 EXPORT_SYMBOL_GPL(sdev_evt_alloc);
2230
2231 /**
2232  *      sdev_evt_send_simple - send asserted event to uevent thread
2233  *      @sdev: scsi_device event occurred on
2234  *      @evt_type: type of event to send
2235  *      @gfpflags: GFP flags for allocation
2236  *
2237  *      Assert scsi device event asynchronously, given an event type.
2238  */
2239 void sdev_evt_send_simple(struct scsi_device *sdev,
2240                           enum scsi_device_event evt_type, gfp_t gfpflags)
2241 {
2242         struct scsi_event *evt = sdev_evt_alloc(evt_type, gfpflags);
2243         if (!evt) {
2244                 sdev_printk(KERN_ERR, sdev, "event %d eaten due to OOM\n",
2245                             evt_type);
2246                 return;
2247         }
2248
2249         sdev_evt_send(sdev, evt);
2250 }
2251 EXPORT_SYMBOL_GPL(sdev_evt_send_simple);
2252
2253 /**
2254  *      scsi_device_quiesce - Block user issued commands.
2255  *      @sdev:  scsi device to quiesce.
2256  *
2257  *      This works by trying to transition to the SDEV_QUIESCE state
2258  *      (which must be a legal transition).  When the device is in this
2259  *      state, only special requests will be accepted, all others will
2260  *      be deferred.  Since special requests may also be requeued requests,
2261  *      a successful return doesn't guarantee the device will be 
2262  *      totally quiescent.
2263  *
2264  *      Must be called with user context, may sleep.
2265  *
2266  *      Returns zero if unsuccessful or an error if not.
2267  **/
2268 int
2269 scsi_device_quiesce(struct scsi_device *sdev)
2270 {
2271         int err = scsi_device_set_state(sdev, SDEV_QUIESCE);
2272         if (err)
2273                 return err;
2274
2275         scsi_run_queue(sdev->request_queue);
2276         while (sdev->device_busy) {
2277                 msleep_interruptible(200);
2278                 scsi_run_queue(sdev->request_queue);
2279         }
2280         return 0;
2281 }
2282 EXPORT_SYMBOL(scsi_device_quiesce);
2283
2284 /**
2285  *      scsi_device_resume - Restart user issued commands to a quiesced device.
2286  *      @sdev:  scsi device to resume.
2287  *
2288  *      Moves the device from quiesced back to running and restarts the
2289  *      queues.
2290  *
2291  *      Must be called with user context, may sleep.
2292  **/
2293 void
2294 scsi_device_resume(struct scsi_device *sdev)
2295 {
2296         if(scsi_device_set_state(sdev, SDEV_RUNNING))
2297                 return;
2298         scsi_run_queue(sdev->request_queue);
2299 }
2300 EXPORT_SYMBOL(scsi_device_resume);
2301
2302 static void
2303 device_quiesce_fn(struct scsi_device *sdev, void *data)
2304 {
2305         scsi_device_quiesce(sdev);
2306 }
2307
2308 void
2309 scsi_target_quiesce(struct scsi_target *starget)
2310 {
2311         starget_for_each_device(starget, NULL, device_quiesce_fn);
2312 }
2313 EXPORT_SYMBOL(scsi_target_quiesce);
2314
2315 static void
2316 device_resume_fn(struct scsi_device *sdev, void *data)
2317 {
2318         scsi_device_resume(sdev);
2319 }
2320
2321 void
2322 scsi_target_resume(struct scsi_target *starget)
2323 {
2324         starget_for_each_device(starget, NULL, device_resume_fn);
2325 }
2326 EXPORT_SYMBOL(scsi_target_resume);
2327
2328 /**
2329  * scsi_internal_device_block - internal function to put a device
2330  *                              temporarily into the SDEV_BLOCK state
2331  * @sdev:       device to block
2332  *
2333  * Block request made by scsi lld's to temporarily stop all
2334  * scsi commands on the specified device.  Called from interrupt
2335  * or normal process context.
2336  *
2337  * Returns zero if successful or error if not
2338  *
2339  * Notes:       
2340  *      This routine transitions the device to the SDEV_BLOCK state
2341  *      (which must be a legal transition).  When the device is in this
2342  *      state, all commands are deferred until the scsi lld reenables
2343  *      the device with scsi_device_unblock or device_block_tmo fires.
2344  *      This routine assumes the host_lock is held on entry.
2345  **/
2346 int
2347 scsi_internal_device_block(struct scsi_device *sdev)
2348 {
2349         struct request_queue *q = sdev->request_queue;
2350         unsigned long flags;
2351         int err = 0;
2352
2353         err = scsi_device_set_state(sdev, SDEV_BLOCK);
2354         if (err)
2355                 return err;
2356
2357         /* 
2358          * The device has transitioned to SDEV_BLOCK.  Stop the
2359          * block layer from calling the midlayer with this device's
2360          * request queue. 
2361          */
2362         spin_lock_irqsave(q->queue_lock, flags);
2363         blk_stop_queue(q);
2364         spin_unlock_irqrestore(q->queue_lock, flags);
2365
2366         return 0;
2367 }
2368 EXPORT_SYMBOL_GPL(scsi_internal_device_block);
2369  
2370 /**
2371  * scsi_internal_device_unblock - resume a device after a block request
2372  * @sdev:       device to resume
2373  *
2374  * Called by scsi lld's or the midlayer to restart the device queue
2375  * for the previously suspended scsi device.  Called from interrupt or
2376  * normal process context.
2377  *
2378  * Returns zero if successful or error if not.
2379  *
2380  * Notes:       
2381  *      This routine transitions the device to the SDEV_RUNNING state
2382  *      (which must be a legal transition) allowing the midlayer to
2383  *      goose the queue for this device.  This routine assumes the 
2384  *      host_lock is held upon entry.
2385  **/
2386 int
2387 scsi_internal_device_unblock(struct scsi_device *sdev)
2388 {
2389         struct request_queue *q = sdev->request_queue; 
2390         int err;
2391         unsigned long flags;
2392         
2393         /* 
2394          * Try to transition the scsi device to SDEV_RUNNING
2395          * and goose the device queue if successful.  
2396          */
2397         err = scsi_device_set_state(sdev, SDEV_RUNNING);
2398         if (err)
2399                 return err;
2400
2401         spin_lock_irqsave(q->queue_lock, flags);
2402         blk_start_queue(q);
2403         spin_unlock_irqrestore(q->queue_lock, flags);
2404
2405         return 0;
2406 }
2407 EXPORT_SYMBOL_GPL(scsi_internal_device_unblock);
2408
2409 static void
2410 device_block(struct scsi_device *sdev, void *data)
2411 {
2412         scsi_internal_device_block(sdev);
2413 }
2414
2415 static int
2416 target_block(struct device *dev, void *data)
2417 {
2418         if (scsi_is_target_device(dev))
2419                 starget_for_each_device(to_scsi_target(dev), NULL,
2420                                         device_block);
2421         return 0;
2422 }
2423
2424 void
2425 scsi_target_block(struct device *dev)
2426 {
2427         if (scsi_is_target_device(dev))
2428                 starget_for_each_device(to_scsi_target(dev), NULL,
2429                                         device_block);
2430         else
2431                 device_for_each_child(dev, NULL, target_block);
2432 }
2433 EXPORT_SYMBOL_GPL(scsi_target_block);
2434
2435 static void
2436 device_unblock(struct scsi_device *sdev, void *data)
2437 {
2438         scsi_internal_device_unblock(sdev);
2439 }
2440
2441 static int
2442 target_unblock(struct device *dev, void *data)
2443 {
2444         if (scsi_is_target_device(dev))
2445                 starget_for_each_device(to_scsi_target(dev), NULL,
2446                                         device_unblock);
2447         return 0;
2448 }
2449
2450 void
2451 scsi_target_unblock(struct device *dev)
2452 {
2453         if (scsi_is_target_device(dev))
2454                 starget_for_each_device(to_scsi_target(dev), NULL,
2455                                         device_unblock);
2456         else
2457                 device_for_each_child(dev, NULL, target_unblock);
2458 }
2459 EXPORT_SYMBOL_GPL(scsi_target_unblock);
2460
2461 /**
2462  * scsi_kmap_atomic_sg - find and atomically map an sg-elemnt
2463  * @sg:         scatter-gather list
2464  * @sg_count:   number of segments in sg
2465  * @offset:     offset in bytes into sg, on return offset into the mapped area
2466  * @len:        bytes to map, on return number of bytes mapped
2467  *
2468  * Returns virtual address of the start of the mapped page
2469  */
2470 void *scsi_kmap_atomic_sg(struct scatterlist *sgl, int sg_count,
2471                           size_t *offset, size_t *len)
2472 {
2473         int i;
2474         size_t sg_len = 0, len_complete = 0;
2475         struct scatterlist *sg;
2476         struct page *page;
2477
2478         WARN_ON(!irqs_disabled());
2479
2480         for_each_sg(sgl, sg, sg_count, i) {
2481                 len_complete = sg_len; /* Complete sg-entries */
2482                 sg_len += sg->length;
2483                 if (sg_len > *offset)
2484                         break;
2485         }
2486
2487         if (unlikely(i == sg_count)) {
2488                 printk(KERN_ERR "%s: Bytes in sg: %zu, requested offset %zu, "
2489                         "elements %d\n",
2490                        __FUNCTION__, sg_len, *offset, sg_count);
2491                 WARN_ON(1);
2492                 return NULL;
2493         }
2494
2495         /* Offset starting from the beginning of first page in this sg-entry */
2496         *offset = *offset - len_complete + sg->offset;
2497
2498         /* Assumption: contiguous pages can be accessed as "page + i" */
2499         page = nth_page(sg_page(sg), (*offset >> PAGE_SHIFT));
2500         *offset &= ~PAGE_MASK;
2501
2502         /* Bytes in this sg-entry from *offset to the end of the page */
2503         sg_len = PAGE_SIZE - *offset;
2504         if (*len > sg_len)
2505                 *len = sg_len;
2506
2507         return kmap_atomic(page, KM_BIO_SRC_IRQ);
2508 }
2509 EXPORT_SYMBOL(scsi_kmap_atomic_sg);
2510
2511 /**
2512  * scsi_kunmap_atomic_sg - atomically unmap a virtual address, previously
2513  *                         mapped with scsi_kmap_atomic_sg
2514  * @virt:       virtual address to be unmapped
2515  */
2516 void scsi_kunmap_atomic_sg(void *virt)
2517 {
2518         kunmap_atomic(virt, KM_BIO_SRC_IRQ);
2519 }
2520 EXPORT_SYMBOL(scsi_kunmap_atomic_sg);