blk_end_request: changing scsi (take 4)
[linux-2.6.git] / drivers / scsi / scsi_lib.c
1 /*
2  *  scsi_lib.c Copyright (C) 1999 Eric Youngdale
3  *
4  *  SCSI queueing library.
5  *      Initial versions: Eric Youngdale (eric@andante.org).
6  *                        Based upon conversations with large numbers
7  *                        of people at Linux Expo.
8  */
9
10 #include <linux/bio.h>
11 #include <linux/blkdev.h>
12 #include <linux/completion.h>
13 #include <linux/kernel.h>
14 #include <linux/mempool.h>
15 #include <linux/slab.h>
16 #include <linux/init.h>
17 #include <linux/pci.h>
18 #include <linux/delay.h>
19 #include <linux/hardirq.h>
20 #include <linux/scatterlist.h>
21
22 #include <scsi/scsi.h>
23 #include <scsi/scsi_cmnd.h>
24 #include <scsi/scsi_dbg.h>
25 #include <scsi/scsi_device.h>
26 #include <scsi/scsi_driver.h>
27 #include <scsi/scsi_eh.h>
28 #include <scsi/scsi_host.h>
29
30 #include "scsi_priv.h"
31 #include "scsi_logging.h"
32
33
34 #define SG_MEMPOOL_NR           ARRAY_SIZE(scsi_sg_pools)
35 #define SG_MEMPOOL_SIZE         2
36
37 /*
38  * The maximum number of SG segments that we will put inside a scatterlist
39  * (unless chaining is used). Should ideally fit inside a single page, to
40  * avoid a higher order allocation.
41  */
42 #define SCSI_MAX_SG_SEGMENTS    128
43
44 struct scsi_host_sg_pool {
45         size_t          size;
46         char            *name;
47         struct kmem_cache       *slab;
48         mempool_t       *pool;
49 };
50
51 #define SP(x) { x, "sgpool-" #x }
52 static struct scsi_host_sg_pool scsi_sg_pools[] = {
53         SP(8),
54         SP(16),
55 #if (SCSI_MAX_SG_SEGMENTS > 16)
56         SP(32),
57 #if (SCSI_MAX_SG_SEGMENTS > 32)
58         SP(64),
59 #if (SCSI_MAX_SG_SEGMENTS > 64)
60         SP(128),
61 #endif
62 #endif
63 #endif
64 };
65 #undef SP
66
67 static void scsi_run_queue(struct request_queue *q);
68
69 /*
70  * Function:    scsi_unprep_request()
71  *
72  * Purpose:     Remove all preparation done for a request, including its
73  *              associated scsi_cmnd, so that it can be requeued.
74  *
75  * Arguments:   req     - request to unprepare
76  *
77  * Lock status: Assumed that no locks are held upon entry.
78  *
79  * Returns:     Nothing.
80  */
81 static void scsi_unprep_request(struct request *req)
82 {
83         struct scsi_cmnd *cmd = req->special;
84
85         req->cmd_flags &= ~REQ_DONTPREP;
86         req->special = NULL;
87
88         scsi_put_command(cmd);
89 }
90
91 /*
92  * Function:    scsi_queue_insert()
93  *
94  * Purpose:     Insert a command in the midlevel queue.
95  *
96  * Arguments:   cmd    - command that we are adding to queue.
97  *              reason - why we are inserting command to queue.
98  *
99  * Lock status: Assumed that lock is not held upon entry.
100  *
101  * Returns:     Nothing.
102  *
103  * Notes:       We do this for one of two cases.  Either the host is busy
104  *              and it cannot accept any more commands for the time being,
105  *              or the device returned QUEUE_FULL and can accept no more
106  *              commands.
107  * Notes:       This could be called either from an interrupt context or a
108  *              normal process context.
109  */
110 int scsi_queue_insert(struct scsi_cmnd *cmd, int reason)
111 {
112         struct Scsi_Host *host = cmd->device->host;
113         struct scsi_device *device = cmd->device;
114         struct request_queue *q = device->request_queue;
115         unsigned long flags;
116
117         SCSI_LOG_MLQUEUE(1,
118                  printk("Inserting command %p into mlqueue\n", cmd));
119
120         /*
121          * Set the appropriate busy bit for the device/host.
122          *
123          * If the host/device isn't busy, assume that something actually
124          * completed, and that we should be able to queue a command now.
125          *
126          * Note that the prior mid-layer assumption that any host could
127          * always queue at least one command is now broken.  The mid-layer
128          * will implement a user specifiable stall (see
129          * scsi_host.max_host_blocked and scsi_device.max_device_blocked)
130          * if a command is requeued with no other commands outstanding
131          * either for the device or for the host.
132          */
133         if (reason == SCSI_MLQUEUE_HOST_BUSY)
134                 host->host_blocked = host->max_host_blocked;
135         else if (reason == SCSI_MLQUEUE_DEVICE_BUSY)
136                 device->device_blocked = device->max_device_blocked;
137
138         /*
139          * Decrement the counters, since these commands are no longer
140          * active on the host/device.
141          */
142         scsi_device_unbusy(device);
143
144         /*
145          * Requeue this command.  It will go before all other commands
146          * that are already in the queue.
147          *
148          * NOTE: there is magic here about the way the queue is plugged if
149          * we have no outstanding commands.
150          * 
151          * Although we *don't* plug the queue, we call the request
152          * function.  The SCSI request function detects the blocked condition
153          * and plugs the queue appropriately.
154          */
155         spin_lock_irqsave(q->queue_lock, flags);
156         blk_requeue_request(q, cmd->request);
157         spin_unlock_irqrestore(q->queue_lock, flags);
158
159         scsi_run_queue(q);
160
161         return 0;
162 }
163
164 /**
165  * scsi_execute - insert request and wait for the result
166  * @sdev:       scsi device
167  * @cmd:        scsi command
168  * @data_direction: data direction
169  * @buffer:     data buffer
170  * @bufflen:    len of buffer
171  * @sense:      optional sense buffer
172  * @timeout:    request timeout in seconds
173  * @retries:    number of times to retry request
174  * @flags:      or into request flags;
175  *
176  * returns the req->errors value which is the scsi_cmnd result
177  * field.
178  */
179 int scsi_execute(struct scsi_device *sdev, const unsigned char *cmd,
180                  int data_direction, void *buffer, unsigned bufflen,
181                  unsigned char *sense, int timeout, int retries, int flags)
182 {
183         struct request *req;
184         int write = (data_direction == DMA_TO_DEVICE);
185         int ret = DRIVER_ERROR << 24;
186
187         req = blk_get_request(sdev->request_queue, write, __GFP_WAIT);
188
189         if (bufflen &&  blk_rq_map_kern(sdev->request_queue, req,
190                                         buffer, bufflen, __GFP_WAIT))
191                 goto out;
192
193         req->cmd_len = COMMAND_SIZE(cmd[0]);
194         memcpy(req->cmd, cmd, req->cmd_len);
195         req->sense = sense;
196         req->sense_len = 0;
197         req->retries = retries;
198         req->timeout = timeout;
199         req->cmd_type = REQ_TYPE_BLOCK_PC;
200         req->cmd_flags |= flags | REQ_QUIET | REQ_PREEMPT;
201
202         /*
203          * head injection *required* here otherwise quiesce won't work
204          */
205         blk_execute_rq(req->q, NULL, req, 1);
206
207         ret = req->errors;
208  out:
209         blk_put_request(req);
210
211         return ret;
212 }
213 EXPORT_SYMBOL(scsi_execute);
214
215
216 int scsi_execute_req(struct scsi_device *sdev, const unsigned char *cmd,
217                      int data_direction, void *buffer, unsigned bufflen,
218                      struct scsi_sense_hdr *sshdr, int timeout, int retries)
219 {
220         char *sense = NULL;
221         int result;
222         
223         if (sshdr) {
224                 sense = kzalloc(SCSI_SENSE_BUFFERSIZE, GFP_NOIO);
225                 if (!sense)
226                         return DRIVER_ERROR << 24;
227         }
228         result = scsi_execute(sdev, cmd, data_direction, buffer, bufflen,
229                               sense, timeout, retries, 0);
230         if (sshdr)
231                 scsi_normalize_sense(sense, SCSI_SENSE_BUFFERSIZE, sshdr);
232
233         kfree(sense);
234         return result;
235 }
236 EXPORT_SYMBOL(scsi_execute_req);
237
238 struct scsi_io_context {
239         void *data;
240         void (*done)(void *data, char *sense, int result, int resid);
241         char sense[SCSI_SENSE_BUFFERSIZE];
242 };
243
244 static struct kmem_cache *scsi_io_context_cache;
245
246 static void scsi_end_async(struct request *req, int uptodate)
247 {
248         struct scsi_io_context *sioc = req->end_io_data;
249
250         if (sioc->done)
251                 sioc->done(sioc->data, sioc->sense, req->errors, req->data_len);
252
253         kmem_cache_free(scsi_io_context_cache, sioc);
254         __blk_put_request(req->q, req);
255 }
256
257 static int scsi_merge_bio(struct request *rq, struct bio *bio)
258 {
259         struct request_queue *q = rq->q;
260
261         bio->bi_flags &= ~(1 << BIO_SEG_VALID);
262         if (rq_data_dir(rq) == WRITE)
263                 bio->bi_rw |= (1 << BIO_RW);
264         blk_queue_bounce(q, &bio);
265
266         return blk_rq_append_bio(q, rq, bio);
267 }
268
269 static void scsi_bi_endio(struct bio *bio, int error)
270 {
271         bio_put(bio);
272 }
273
274 /**
275  * scsi_req_map_sg - map a scatterlist into a request
276  * @rq:         request to fill
277  * @sgl:        scatterlist
278  * @nsegs:      number of elements
279  * @bufflen:    len of buffer
280  * @gfp:        memory allocation flags
281  *
282  * scsi_req_map_sg maps a scatterlist into a request so that the
283  * request can be sent to the block layer. We do not trust the scatterlist
284  * sent to use, as some ULDs use that struct to only organize the pages.
285  */
286 static int scsi_req_map_sg(struct request *rq, struct scatterlist *sgl,
287                            int nsegs, unsigned bufflen, gfp_t gfp)
288 {
289         struct request_queue *q = rq->q;
290         int nr_pages = (bufflen + sgl[0].offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
291         unsigned int data_len = bufflen, len, bytes, off;
292         struct scatterlist *sg;
293         struct page *page;
294         struct bio *bio = NULL;
295         int i, err, nr_vecs = 0;
296
297         for_each_sg(sgl, sg, nsegs, i) {
298                 page = sg_page(sg);
299                 off = sg->offset;
300                 len = sg->length;
301                 data_len += len;
302
303                 while (len > 0 && data_len > 0) {
304                         /*
305                          * sg sends a scatterlist that is larger than
306                          * the data_len it wants transferred for certain
307                          * IO sizes
308                          */
309                         bytes = min_t(unsigned int, len, PAGE_SIZE - off);
310                         bytes = min(bytes, data_len);
311
312                         if (!bio) {
313                                 nr_vecs = min_t(int, BIO_MAX_PAGES, nr_pages);
314                                 nr_pages -= nr_vecs;
315
316                                 bio = bio_alloc(gfp, nr_vecs);
317                                 if (!bio) {
318                                         err = -ENOMEM;
319                                         goto free_bios;
320                                 }
321                                 bio->bi_end_io = scsi_bi_endio;
322                         }
323
324                         if (bio_add_pc_page(q, bio, page, bytes, off) !=
325                             bytes) {
326                                 bio_put(bio);
327                                 err = -EINVAL;
328                                 goto free_bios;
329                         }
330
331                         if (bio->bi_vcnt >= nr_vecs) {
332                                 err = scsi_merge_bio(rq, bio);
333                                 if (err) {
334                                         bio_endio(bio, 0);
335                                         goto free_bios;
336                                 }
337                                 bio = NULL;
338                         }
339
340                         page++;
341                         len -= bytes;
342                         data_len -=bytes;
343                         off = 0;
344                 }
345         }
346
347         rq->buffer = rq->data = NULL;
348         rq->data_len = bufflen;
349         return 0;
350
351 free_bios:
352         while ((bio = rq->bio) != NULL) {
353                 rq->bio = bio->bi_next;
354                 /*
355                  * call endio instead of bio_put incase it was bounced
356                  */
357                 bio_endio(bio, 0);
358         }
359
360         return err;
361 }
362
363 /**
364  * scsi_execute_async - insert request
365  * @sdev:       scsi device
366  * @cmd:        scsi command
367  * @cmd_len:    length of scsi cdb
368  * @data_direction: DMA_TO_DEVICE, DMA_FROM_DEVICE, or DMA_NONE
369  * @buffer:     data buffer (this can be a kernel buffer or scatterlist)
370  * @bufflen:    len of buffer
371  * @use_sg:     if buffer is a scatterlist this is the number of elements
372  * @timeout:    request timeout in seconds
373  * @retries:    number of times to retry request
374  * @privdata:   data passed to done()
375  * @done:       callback function when done
376  * @gfp:        memory allocation flags
377  */
378 int scsi_execute_async(struct scsi_device *sdev, const unsigned char *cmd,
379                        int cmd_len, int data_direction, void *buffer, unsigned bufflen,
380                        int use_sg, int timeout, int retries, void *privdata,
381                        void (*done)(void *, char *, int, int), gfp_t gfp)
382 {
383         struct request *req;
384         struct scsi_io_context *sioc;
385         int err = 0;
386         int write = (data_direction == DMA_TO_DEVICE);
387
388         sioc = kmem_cache_zalloc(scsi_io_context_cache, gfp);
389         if (!sioc)
390                 return DRIVER_ERROR << 24;
391
392         req = blk_get_request(sdev->request_queue, write, gfp);
393         if (!req)
394                 goto free_sense;
395         req->cmd_type = REQ_TYPE_BLOCK_PC;
396         req->cmd_flags |= REQ_QUIET;
397
398         if (use_sg)
399                 err = scsi_req_map_sg(req, buffer, use_sg, bufflen, gfp);
400         else if (bufflen)
401                 err = blk_rq_map_kern(req->q, req, buffer, bufflen, gfp);
402
403         if (err)
404                 goto free_req;
405
406         req->cmd_len = cmd_len;
407         memset(req->cmd, 0, BLK_MAX_CDB); /* ATAPI hates garbage after CDB */
408         memcpy(req->cmd, cmd, req->cmd_len);
409         req->sense = sioc->sense;
410         req->sense_len = 0;
411         req->timeout = timeout;
412         req->retries = retries;
413         req->end_io_data = sioc;
414
415         sioc->data = privdata;
416         sioc->done = done;
417
418         blk_execute_rq_nowait(req->q, NULL, req, 1, scsi_end_async);
419         return 0;
420
421 free_req:
422         blk_put_request(req);
423 free_sense:
424         kmem_cache_free(scsi_io_context_cache, sioc);
425         return DRIVER_ERROR << 24;
426 }
427 EXPORT_SYMBOL_GPL(scsi_execute_async);
428
429 /*
430  * Function:    scsi_init_cmd_errh()
431  *
432  * Purpose:     Initialize cmd fields related to error handling.
433  *
434  * Arguments:   cmd     - command that is ready to be queued.
435  *
436  * Notes:       This function has the job of initializing a number of
437  *              fields related to error handling.   Typically this will
438  *              be called once for each command, as required.
439  */
440 static void scsi_init_cmd_errh(struct scsi_cmnd *cmd)
441 {
442         cmd->serial_number = 0;
443         cmd->resid = 0;
444         memset(cmd->sense_buffer, 0, SCSI_SENSE_BUFFERSIZE);
445         if (cmd->cmd_len == 0)
446                 cmd->cmd_len = COMMAND_SIZE(cmd->cmnd[0]);
447 }
448
449 void scsi_device_unbusy(struct scsi_device *sdev)
450 {
451         struct Scsi_Host *shost = sdev->host;
452         unsigned long flags;
453
454         spin_lock_irqsave(shost->host_lock, flags);
455         shost->host_busy--;
456         if (unlikely(scsi_host_in_recovery(shost) &&
457                      (shost->host_failed || shost->host_eh_scheduled)))
458                 scsi_eh_wakeup(shost);
459         spin_unlock(shost->host_lock);
460         spin_lock(sdev->request_queue->queue_lock);
461         sdev->device_busy--;
462         spin_unlock_irqrestore(sdev->request_queue->queue_lock, flags);
463 }
464
465 /*
466  * Called for single_lun devices on IO completion. Clear starget_sdev_user,
467  * and call blk_run_queue for all the scsi_devices on the target -
468  * including current_sdev first.
469  *
470  * Called with *no* scsi locks held.
471  */
472 static void scsi_single_lun_run(struct scsi_device *current_sdev)
473 {
474         struct Scsi_Host *shost = current_sdev->host;
475         struct scsi_device *sdev, *tmp;
476         struct scsi_target *starget = scsi_target(current_sdev);
477         unsigned long flags;
478
479         spin_lock_irqsave(shost->host_lock, flags);
480         starget->starget_sdev_user = NULL;
481         spin_unlock_irqrestore(shost->host_lock, flags);
482
483         /*
484          * Call blk_run_queue for all LUNs on the target, starting with
485          * current_sdev. We race with others (to set starget_sdev_user),
486          * but in most cases, we will be first. Ideally, each LU on the
487          * target would get some limited time or requests on the target.
488          */
489         blk_run_queue(current_sdev->request_queue);
490
491         spin_lock_irqsave(shost->host_lock, flags);
492         if (starget->starget_sdev_user)
493                 goto out;
494         list_for_each_entry_safe(sdev, tmp, &starget->devices,
495                         same_target_siblings) {
496                 if (sdev == current_sdev)
497                         continue;
498                 if (scsi_device_get(sdev))
499                         continue;
500
501                 spin_unlock_irqrestore(shost->host_lock, flags);
502                 blk_run_queue(sdev->request_queue);
503                 spin_lock_irqsave(shost->host_lock, flags);
504         
505                 scsi_device_put(sdev);
506         }
507  out:
508         spin_unlock_irqrestore(shost->host_lock, flags);
509 }
510
511 /*
512  * Function:    scsi_run_queue()
513  *
514  * Purpose:     Select a proper request queue to serve next
515  *
516  * Arguments:   q       - last request's queue
517  *
518  * Returns:     Nothing
519  *
520  * Notes:       The previous command was completely finished, start
521  *              a new one if possible.
522  */
523 static void scsi_run_queue(struct request_queue *q)
524 {
525         struct scsi_device *sdev = q->queuedata;
526         struct Scsi_Host *shost = sdev->host;
527         unsigned long flags;
528
529         if (scsi_target(sdev)->single_lun)
530                 scsi_single_lun_run(sdev);
531
532         spin_lock_irqsave(shost->host_lock, flags);
533         while (!list_empty(&shost->starved_list) &&
534                !shost->host_blocked && !shost->host_self_blocked &&
535                 !((shost->can_queue > 0) &&
536                   (shost->host_busy >= shost->can_queue))) {
537                 /*
538                  * As long as shost is accepting commands and we have
539                  * starved queues, call blk_run_queue. scsi_request_fn
540                  * drops the queue_lock and can add us back to the
541                  * starved_list.
542                  *
543                  * host_lock protects the starved_list and starved_entry.
544                  * scsi_request_fn must get the host_lock before checking
545                  * or modifying starved_list or starved_entry.
546                  */
547                 sdev = list_entry(shost->starved_list.next,
548                                           struct scsi_device, starved_entry);
549                 list_del_init(&sdev->starved_entry);
550                 spin_unlock_irqrestore(shost->host_lock, flags);
551
552
553                 if (test_bit(QUEUE_FLAG_REENTER, &q->queue_flags) &&
554                     !test_and_set_bit(QUEUE_FLAG_REENTER,
555                                       &sdev->request_queue->queue_flags)) {
556                         blk_run_queue(sdev->request_queue);
557                         clear_bit(QUEUE_FLAG_REENTER,
558                                   &sdev->request_queue->queue_flags);
559                 } else
560                         blk_run_queue(sdev->request_queue);
561
562                 spin_lock_irqsave(shost->host_lock, flags);
563                 if (unlikely(!list_empty(&sdev->starved_entry)))
564                         /*
565                          * sdev lost a race, and was put back on the
566                          * starved list. This is unlikely but without this
567                          * in theory we could loop forever.
568                          */
569                         break;
570         }
571         spin_unlock_irqrestore(shost->host_lock, flags);
572
573         blk_run_queue(q);
574 }
575
576 /*
577  * Function:    scsi_requeue_command()
578  *
579  * Purpose:     Handle post-processing of completed commands.
580  *
581  * Arguments:   q       - queue to operate on
582  *              cmd     - command that may need to be requeued.
583  *
584  * Returns:     Nothing
585  *
586  * Notes:       After command completion, there may be blocks left
587  *              over which weren't finished by the previous command
588  *              this can be for a number of reasons - the main one is
589  *              I/O errors in the middle of the request, in which case
590  *              we need to request the blocks that come after the bad
591  *              sector.
592  * Notes:       Upon return, cmd is a stale pointer.
593  */
594 static void scsi_requeue_command(struct request_queue *q, struct scsi_cmnd *cmd)
595 {
596         struct request *req = cmd->request;
597         unsigned long flags;
598
599         scsi_unprep_request(req);
600         spin_lock_irqsave(q->queue_lock, flags);
601         blk_requeue_request(q, req);
602         spin_unlock_irqrestore(q->queue_lock, flags);
603
604         scsi_run_queue(q);
605 }
606
607 void scsi_next_command(struct scsi_cmnd *cmd)
608 {
609         struct scsi_device *sdev = cmd->device;
610         struct request_queue *q = sdev->request_queue;
611
612         /* need to hold a reference on the device before we let go of the cmd */
613         get_device(&sdev->sdev_gendev);
614
615         scsi_put_command(cmd);
616         scsi_run_queue(q);
617
618         /* ok to remove device now */
619         put_device(&sdev->sdev_gendev);
620 }
621
622 void scsi_run_host_queues(struct Scsi_Host *shost)
623 {
624         struct scsi_device *sdev;
625
626         shost_for_each_device(sdev, shost)
627                 scsi_run_queue(sdev->request_queue);
628 }
629
630 /*
631  * Function:    scsi_end_request()
632  *
633  * Purpose:     Post-processing of completed commands (usually invoked at end
634  *              of upper level post-processing and scsi_io_completion).
635  *
636  * Arguments:   cmd      - command that is complete.
637  *              error    - 0 if I/O indicates success, < 0 for I/O error.
638  *              bytes    - number of bytes of completed I/O
639  *              requeue  - indicates whether we should requeue leftovers.
640  *
641  * Lock status: Assumed that lock is not held upon entry.
642  *
643  * Returns:     cmd if requeue required, NULL otherwise.
644  *
645  * Notes:       This is called for block device requests in order to
646  *              mark some number of sectors as complete.
647  * 
648  *              We are guaranteeing that the request queue will be goosed
649  *              at some point during this call.
650  * Notes:       If cmd was requeued, upon return it will be a stale pointer.
651  */
652 static struct scsi_cmnd *scsi_end_request(struct scsi_cmnd *cmd, int error,
653                                           int bytes, int requeue)
654 {
655         struct request_queue *q = cmd->device->request_queue;
656         struct request *req = cmd->request;
657
658         /*
659          * If there are blocks left over at the end, set up the command
660          * to queue the remainder of them.
661          */
662         if (blk_end_request(req, error, bytes)) {
663                 int leftover = (req->hard_nr_sectors << 9);
664
665                 if (blk_pc_request(req))
666                         leftover = req->data_len;
667
668                 /* kill remainder if no retrys */
669                 if (error && blk_noretry_request(req))
670                         blk_end_request(req, error, leftover);
671                 else {
672                         if (requeue) {
673                                 /*
674                                  * Bleah.  Leftovers again.  Stick the
675                                  * leftovers in the front of the
676                                  * queue, and goose the queue again.
677                                  */
678                                 scsi_requeue_command(q, cmd);
679                                 cmd = NULL;
680                         }
681                         return cmd;
682                 }
683         }
684
685         /*
686          * This will goose the queue request function at the end, so we don't
687          * need to worry about launching another command.
688          */
689         scsi_next_command(cmd);
690         return NULL;
691 }
692
693 /*
694  * Like SCSI_MAX_SG_SEGMENTS, but for archs that have sg chaining. This limit
695  * is totally arbitrary, a setting of 2048 will get you at least 8mb ios.
696  */
697 #define SCSI_MAX_SG_CHAIN_SEGMENTS      2048
698
699 static inline unsigned int scsi_sgtable_index(unsigned short nents)
700 {
701         unsigned int index;
702
703         switch (nents) {
704         case 1 ... 8:
705                 index = 0;
706                 break;
707         case 9 ... 16:
708                 index = 1;
709                 break;
710 #if (SCSI_MAX_SG_SEGMENTS > 16)
711         case 17 ... 32:
712                 index = 2;
713                 break;
714 #if (SCSI_MAX_SG_SEGMENTS > 32)
715         case 33 ... 64:
716                 index = 3;
717                 break;
718 #if (SCSI_MAX_SG_SEGMENTS > 64)
719         case 65 ... 128:
720                 index = 4;
721                 break;
722 #endif
723 #endif
724 #endif
725         default:
726                 printk(KERN_ERR "scsi: bad segment count=%d\n", nents);
727                 BUG();
728         }
729
730         return index;
731 }
732
733 struct scatterlist *scsi_alloc_sgtable(struct scsi_cmnd *cmd, gfp_t gfp_mask)
734 {
735         struct scsi_host_sg_pool *sgp;
736         struct scatterlist *sgl, *prev, *ret;
737         unsigned int index;
738         int this, left;
739
740         BUG_ON(!cmd->use_sg);
741
742         left = cmd->use_sg;
743         ret = prev = NULL;
744         do {
745                 this = left;
746                 if (this > SCSI_MAX_SG_SEGMENTS) {
747                         this = SCSI_MAX_SG_SEGMENTS - 1;
748                         index = SG_MEMPOOL_NR - 1;
749                 } else
750                         index = scsi_sgtable_index(this);
751
752                 left -= this;
753
754                 sgp = scsi_sg_pools + index;
755
756                 sgl = mempool_alloc(sgp->pool, gfp_mask);
757                 if (unlikely(!sgl))
758                         goto enomem;
759
760                 sg_init_table(sgl, sgp->size);
761
762                 /*
763                  * first loop through, set initial index and return value
764                  */
765                 if (!ret)
766                         ret = sgl;
767
768                 /*
769                  * chain previous sglist, if any. we know the previous
770                  * sglist must be the biggest one, or we would not have
771                  * ended up doing another loop.
772                  */
773                 if (prev)
774                         sg_chain(prev, SCSI_MAX_SG_SEGMENTS, sgl);
775
776                 /*
777                  * if we have nothing left, mark the last segment as
778                  * end-of-list
779                  */
780                 if (!left)
781                         sg_mark_end(&sgl[this - 1]);
782
783                 /*
784                  * don't allow subsequent mempool allocs to sleep, it would
785                  * violate the mempool principle.
786                  */
787                 gfp_mask &= ~__GFP_WAIT;
788                 gfp_mask |= __GFP_HIGH;
789                 prev = sgl;
790         } while (left);
791
792         /*
793          * ->use_sg may get modified after dma mapping has potentially
794          * shrunk the number of segments, so keep a copy of it for free.
795          */
796         cmd->__use_sg = cmd->use_sg;
797         return ret;
798 enomem:
799         if (ret) {
800                 /*
801                  * Free entries chained off ret. Since we were trying to
802                  * allocate another sglist, we know that all entries are of
803                  * the max size.
804                  */
805                 sgp = scsi_sg_pools + SG_MEMPOOL_NR - 1;
806                 prev = ret;
807                 ret = &ret[SCSI_MAX_SG_SEGMENTS - 1];
808
809                 while ((sgl = sg_chain_ptr(ret)) != NULL) {
810                         ret = &sgl[SCSI_MAX_SG_SEGMENTS - 1];
811                         mempool_free(sgl, sgp->pool);
812                 }
813
814                 mempool_free(prev, sgp->pool);
815         }
816         return NULL;
817 }
818
819 EXPORT_SYMBOL(scsi_alloc_sgtable);
820
821 void scsi_free_sgtable(struct scsi_cmnd *cmd)
822 {
823         struct scatterlist *sgl = cmd->request_buffer;
824         struct scsi_host_sg_pool *sgp;
825
826         /*
827          * if this is the biggest size sglist, check if we have
828          * chained parts we need to free
829          */
830         if (cmd->__use_sg > SCSI_MAX_SG_SEGMENTS) {
831                 unsigned short this, left;
832                 struct scatterlist *next;
833                 unsigned int index;
834
835                 left = cmd->__use_sg - (SCSI_MAX_SG_SEGMENTS - 1);
836                 next = sg_chain_ptr(&sgl[SCSI_MAX_SG_SEGMENTS - 1]);
837                 while (left && next) {
838                         sgl = next;
839                         this = left;
840                         if (this > SCSI_MAX_SG_SEGMENTS) {
841                                 this = SCSI_MAX_SG_SEGMENTS - 1;
842                                 index = SG_MEMPOOL_NR - 1;
843                         } else
844                                 index = scsi_sgtable_index(this);
845
846                         left -= this;
847
848                         sgp = scsi_sg_pools + index;
849
850                         if (left)
851                                 next = sg_chain_ptr(&sgl[sgp->size - 1]);
852
853                         mempool_free(sgl, sgp->pool);
854                 }
855
856                 /*
857                  * Restore original, will be freed below
858                  */
859                 sgl = cmd->request_buffer;
860                 sgp = scsi_sg_pools + SG_MEMPOOL_NR - 1;
861         } else
862                 sgp = scsi_sg_pools + scsi_sgtable_index(cmd->__use_sg);
863
864         mempool_free(sgl, sgp->pool);
865 }
866
867 EXPORT_SYMBOL(scsi_free_sgtable);
868
869 /*
870  * Function:    scsi_release_buffers()
871  *
872  * Purpose:     Completion processing for block device I/O requests.
873  *
874  * Arguments:   cmd     - command that we are bailing.
875  *
876  * Lock status: Assumed that no lock is held upon entry.
877  *
878  * Returns:     Nothing
879  *
880  * Notes:       In the event that an upper level driver rejects a
881  *              command, we must release resources allocated during
882  *              the __init_io() function.  Primarily this would involve
883  *              the scatter-gather table, and potentially any bounce
884  *              buffers.
885  */
886 static void scsi_release_buffers(struct scsi_cmnd *cmd)
887 {
888         if (cmd->use_sg)
889                 scsi_free_sgtable(cmd);
890
891         /*
892          * Zero these out.  They now point to freed memory, and it is
893          * dangerous to hang onto the pointers.
894          */
895         cmd->request_buffer = NULL;
896         cmd->request_bufflen = 0;
897 }
898
899 /*
900  * Function:    scsi_io_completion()
901  *
902  * Purpose:     Completion processing for block device I/O requests.
903  *
904  * Arguments:   cmd   - command that is finished.
905  *
906  * Lock status: Assumed that no lock is held upon entry.
907  *
908  * Returns:     Nothing
909  *
910  * Notes:       This function is matched in terms of capabilities to
911  *              the function that created the scatter-gather list.
912  *              In other words, if there are no bounce buffers
913  *              (the normal case for most drivers), we don't need
914  *              the logic to deal with cleaning up afterwards.
915  *
916  *              We must do one of several things here:
917  *
918  *              a) Call scsi_end_request.  This will finish off the
919  *                 specified number of sectors.  If we are done, the
920  *                 command block will be released, and the queue
921  *                 function will be goosed.  If we are not done, then
922  *                 scsi_end_request will directly goose the queue.
923  *
924  *              b) We can just use scsi_requeue_command() here.  This would
925  *                 be used if we just wanted to retry, for example.
926  */
927 void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes)
928 {
929         int result = cmd->result;
930         int this_count = cmd->request_bufflen;
931         struct request_queue *q = cmd->device->request_queue;
932         struct request *req = cmd->request;
933         int clear_errors = 1;
934         struct scsi_sense_hdr sshdr;
935         int sense_valid = 0;
936         int sense_deferred = 0;
937
938         scsi_release_buffers(cmd);
939
940         if (result) {
941                 sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
942                 if (sense_valid)
943                         sense_deferred = scsi_sense_is_deferred(&sshdr);
944         }
945
946         if (blk_pc_request(req)) { /* SG_IO ioctl from block level */
947                 req->errors = result;
948                 if (result) {
949                         clear_errors = 0;
950                         if (sense_valid && req->sense) {
951                                 /*
952                                  * SG_IO wants current and deferred errors
953                                  */
954                                 int len = 8 + cmd->sense_buffer[7];
955
956                                 if (len > SCSI_SENSE_BUFFERSIZE)
957                                         len = SCSI_SENSE_BUFFERSIZE;
958                                 memcpy(req->sense, cmd->sense_buffer,  len);
959                                 req->sense_len = len;
960                         }
961                 }
962                 req->data_len = cmd->resid;
963         }
964
965         /*
966          * Next deal with any sectors which we were able to correctly
967          * handle.
968          */
969         SCSI_LOG_HLCOMPLETE(1, printk("%ld sectors total, "
970                                       "%d bytes done.\n",
971                                       req->nr_sectors, good_bytes));
972         SCSI_LOG_HLCOMPLETE(1, printk("use_sg is %d\n", cmd->use_sg));
973
974         if (clear_errors)
975                 req->errors = 0;
976
977         /* A number of bytes were successfully read.  If there
978          * are leftovers and there is some kind of error
979          * (result != 0), retry the rest.
980          */
981         if (scsi_end_request(cmd, 0, good_bytes, result == 0) == NULL)
982                 return;
983
984         /* good_bytes = 0, or (inclusive) there were leftovers and
985          * result = 0, so scsi_end_request couldn't retry.
986          */
987         if (sense_valid && !sense_deferred) {
988                 switch (sshdr.sense_key) {
989                 case UNIT_ATTENTION:
990                         if (cmd->device->removable) {
991                                 /* Detected disc change.  Set a bit
992                                  * and quietly refuse further access.
993                                  */
994                                 cmd->device->changed = 1;
995                                 scsi_end_request(cmd, -EIO, this_count, 1);
996                                 return;
997                         } else {
998                                 /* Must have been a power glitch, or a
999                                  * bus reset.  Could not have been a
1000                                  * media change, so we just retry the
1001                                  * request and see what happens.
1002                                  */
1003                                 scsi_requeue_command(q, cmd);
1004                                 return;
1005                         }
1006                         break;
1007                 case ILLEGAL_REQUEST:
1008                         /* If we had an ILLEGAL REQUEST returned, then
1009                          * we may have performed an unsupported
1010                          * command.  The only thing this should be
1011                          * would be a ten byte read where only a six
1012                          * byte read was supported.  Also, on a system
1013                          * where READ CAPACITY failed, we may have
1014                          * read past the end of the disk.
1015                          */
1016                         if ((cmd->device->use_10_for_rw &&
1017                             sshdr.asc == 0x20 && sshdr.ascq == 0x00) &&
1018                             (cmd->cmnd[0] == READ_10 ||
1019                              cmd->cmnd[0] == WRITE_10)) {
1020                                 cmd->device->use_10_for_rw = 0;
1021                                 /* This will cause a retry with a
1022                                  * 6-byte command.
1023                                  */
1024                                 scsi_requeue_command(q, cmd);
1025                                 return;
1026                         } else {
1027                                 scsi_end_request(cmd, -EIO, this_count, 1);
1028                                 return;
1029                         }
1030                         break;
1031                 case NOT_READY:
1032                         /* If the device is in the process of becoming
1033                          * ready, or has a temporary blockage, retry.
1034                          */
1035                         if (sshdr.asc == 0x04) {
1036                                 switch (sshdr.ascq) {
1037                                 case 0x01: /* becoming ready */
1038                                 case 0x04: /* format in progress */
1039                                 case 0x05: /* rebuild in progress */
1040                                 case 0x06: /* recalculation in progress */
1041                                 case 0x07: /* operation in progress */
1042                                 case 0x08: /* Long write in progress */
1043                                 case 0x09: /* self test in progress */
1044                                         scsi_requeue_command(q, cmd);
1045                                         return;
1046                                 default:
1047                                         break;
1048                                 }
1049                         }
1050                         if (!(req->cmd_flags & REQ_QUIET))
1051                                 scsi_cmd_print_sense_hdr(cmd,
1052                                                          "Device not ready",
1053                                                          &sshdr);
1054
1055                         scsi_end_request(cmd, -EIO, this_count, 1);
1056                         return;
1057                 case VOLUME_OVERFLOW:
1058                         if (!(req->cmd_flags & REQ_QUIET)) {
1059                                 scmd_printk(KERN_INFO, cmd,
1060                                             "Volume overflow, CDB: ");
1061                                 __scsi_print_command(cmd->cmnd);
1062                                 scsi_print_sense("", cmd);
1063                         }
1064                         /* See SSC3rXX or current. */
1065                         scsi_end_request(cmd, -EIO, this_count, 1);
1066                         return;
1067                 default:
1068                         break;
1069                 }
1070         }
1071         if (host_byte(result) == DID_RESET) {
1072                 /* Third party bus reset or reset for error recovery
1073                  * reasons.  Just retry the request and see what
1074                  * happens.
1075                  */
1076                 scsi_requeue_command(q, cmd);
1077                 return;
1078         }
1079         if (result) {
1080                 if (!(req->cmd_flags & REQ_QUIET)) {
1081                         scsi_print_result(cmd);
1082                         if (driver_byte(result) & DRIVER_SENSE)
1083                                 scsi_print_sense("", cmd);
1084                 }
1085         }
1086         scsi_end_request(cmd, -EIO, this_count, !result);
1087 }
1088
1089 /*
1090  * Function:    scsi_init_io()
1091  *
1092  * Purpose:     SCSI I/O initialize function.
1093  *
1094  * Arguments:   cmd   - Command descriptor we wish to initialize
1095  *
1096  * Returns:     0 on success
1097  *              BLKPREP_DEFER if the failure is retryable
1098  */
1099 static int scsi_init_io(struct scsi_cmnd *cmd)
1100 {
1101         struct request     *req = cmd->request;
1102         int                count;
1103
1104         /*
1105          * We used to not use scatter-gather for single segment request,
1106          * but now we do (it makes highmem I/O easier to support without
1107          * kmapping pages)
1108          */
1109         cmd->use_sg = req->nr_phys_segments;
1110
1111         /*
1112          * If sg table allocation fails, requeue request later.
1113          */
1114         cmd->request_buffer = scsi_alloc_sgtable(cmd, GFP_ATOMIC);
1115         if (unlikely(!cmd->request_buffer)) {
1116                 scsi_unprep_request(req);
1117                 return BLKPREP_DEFER;
1118         }
1119
1120         req->buffer = NULL;
1121         if (blk_pc_request(req))
1122                 cmd->request_bufflen = req->data_len;
1123         else
1124                 cmd->request_bufflen = req->nr_sectors << 9;
1125
1126         /* 
1127          * Next, walk the list, and fill in the addresses and sizes of
1128          * each segment.
1129          */
1130         count = blk_rq_map_sg(req->q, req, cmd->request_buffer);
1131         BUG_ON(count > cmd->use_sg);
1132         cmd->use_sg = count;
1133         return BLKPREP_OK;
1134 }
1135
1136 static struct scsi_cmnd *scsi_get_cmd_from_req(struct scsi_device *sdev,
1137                 struct request *req)
1138 {
1139         struct scsi_cmnd *cmd;
1140
1141         if (!req->special) {
1142                 cmd = scsi_get_command(sdev, GFP_ATOMIC);
1143                 if (unlikely(!cmd))
1144                         return NULL;
1145                 req->special = cmd;
1146         } else {
1147                 cmd = req->special;
1148         }
1149
1150         /* pull a tag out of the request if we have one */
1151         cmd->tag = req->tag;
1152         cmd->request = req;
1153
1154         return cmd;
1155 }
1156
1157 int scsi_setup_blk_pc_cmnd(struct scsi_device *sdev, struct request *req)
1158 {
1159         struct scsi_cmnd *cmd;
1160         int ret = scsi_prep_state_check(sdev, req);
1161
1162         if (ret != BLKPREP_OK)
1163                 return ret;
1164
1165         cmd = scsi_get_cmd_from_req(sdev, req);
1166         if (unlikely(!cmd))
1167                 return BLKPREP_DEFER;
1168
1169         /*
1170          * BLOCK_PC requests may transfer data, in which case they must
1171          * a bio attached to them.  Or they might contain a SCSI command
1172          * that does not transfer data, in which case they may optionally
1173          * submit a request without an attached bio.
1174          */
1175         if (req->bio) {
1176                 int ret;
1177
1178                 BUG_ON(!req->nr_phys_segments);
1179
1180                 ret = scsi_init_io(cmd);
1181                 if (unlikely(ret))
1182                         return ret;
1183         } else {
1184                 BUG_ON(req->data_len);
1185                 BUG_ON(req->data);
1186
1187                 cmd->request_bufflen = 0;
1188                 cmd->request_buffer = NULL;
1189                 cmd->use_sg = 0;
1190                 req->buffer = NULL;
1191         }
1192
1193         BUILD_BUG_ON(sizeof(req->cmd) > sizeof(cmd->cmnd));
1194         memcpy(cmd->cmnd, req->cmd, sizeof(cmd->cmnd));
1195         cmd->cmd_len = req->cmd_len;
1196         if (!req->data_len)
1197                 cmd->sc_data_direction = DMA_NONE;
1198         else if (rq_data_dir(req) == WRITE)
1199                 cmd->sc_data_direction = DMA_TO_DEVICE;
1200         else
1201                 cmd->sc_data_direction = DMA_FROM_DEVICE;
1202         
1203         cmd->transfersize = req->data_len;
1204         cmd->allowed = req->retries;
1205         cmd->timeout_per_command = req->timeout;
1206         return BLKPREP_OK;
1207 }
1208 EXPORT_SYMBOL(scsi_setup_blk_pc_cmnd);
1209
1210 /*
1211  * Setup a REQ_TYPE_FS command.  These are simple read/write request
1212  * from filesystems that still need to be translated to SCSI CDBs from
1213  * the ULD.
1214  */
1215 int scsi_setup_fs_cmnd(struct scsi_device *sdev, struct request *req)
1216 {
1217         struct scsi_cmnd *cmd;
1218         int ret = scsi_prep_state_check(sdev, req);
1219
1220         if (ret != BLKPREP_OK)
1221                 return ret;
1222         /*
1223          * Filesystem requests must transfer data.
1224          */
1225         BUG_ON(!req->nr_phys_segments);
1226
1227         cmd = scsi_get_cmd_from_req(sdev, req);
1228         if (unlikely(!cmd))
1229                 return BLKPREP_DEFER;
1230
1231         return scsi_init_io(cmd);
1232 }
1233 EXPORT_SYMBOL(scsi_setup_fs_cmnd);
1234
1235 int scsi_prep_state_check(struct scsi_device *sdev, struct request *req)
1236 {
1237         int ret = BLKPREP_OK;
1238
1239         /*
1240          * If the device is not in running state we will reject some
1241          * or all commands.
1242          */
1243         if (unlikely(sdev->sdev_state != SDEV_RUNNING)) {
1244                 switch (sdev->sdev_state) {
1245                 case SDEV_OFFLINE:
1246                         /*
1247                          * If the device is offline we refuse to process any
1248                          * commands.  The device must be brought online
1249                          * before trying any recovery commands.
1250                          */
1251                         sdev_printk(KERN_ERR, sdev,
1252                                     "rejecting I/O to offline device\n");
1253                         ret = BLKPREP_KILL;
1254                         break;
1255                 case SDEV_DEL:
1256                         /*
1257                          * If the device is fully deleted, we refuse to
1258                          * process any commands as well.
1259                          */
1260                         sdev_printk(KERN_ERR, sdev,
1261                                     "rejecting I/O to dead device\n");
1262                         ret = BLKPREP_KILL;
1263                         break;
1264                 case SDEV_QUIESCE:
1265                 case SDEV_BLOCK:
1266                         /*
1267                          * If the devices is blocked we defer normal commands.
1268                          */
1269                         if (!(req->cmd_flags & REQ_PREEMPT))
1270                                 ret = BLKPREP_DEFER;
1271                         break;
1272                 default:
1273                         /*
1274                          * For any other not fully online state we only allow
1275                          * special commands.  In particular any user initiated
1276                          * command is not allowed.
1277                          */
1278                         if (!(req->cmd_flags & REQ_PREEMPT))
1279                                 ret = BLKPREP_KILL;
1280                         break;
1281                 }
1282         }
1283         return ret;
1284 }
1285 EXPORT_SYMBOL(scsi_prep_state_check);
1286
1287 int scsi_prep_return(struct request_queue *q, struct request *req, int ret)
1288 {
1289         struct scsi_device *sdev = q->queuedata;
1290
1291         switch (ret) {
1292         case BLKPREP_KILL:
1293                 req->errors = DID_NO_CONNECT << 16;
1294                 /* release the command and kill it */
1295                 if (req->special) {
1296                         struct scsi_cmnd *cmd = req->special;
1297                         scsi_release_buffers(cmd);
1298                         scsi_put_command(cmd);
1299                         req->special = NULL;
1300                 }
1301                 break;
1302         case BLKPREP_DEFER:
1303                 /*
1304                  * If we defer, the elv_next_request() returns NULL, but the
1305                  * queue must be restarted, so we plug here if no returning
1306                  * command will automatically do that.
1307                  */
1308                 if (sdev->device_busy == 0)
1309                         blk_plug_device(q);
1310                 break;
1311         default:
1312                 req->cmd_flags |= REQ_DONTPREP;
1313         }
1314
1315         return ret;
1316 }
1317 EXPORT_SYMBOL(scsi_prep_return);
1318
1319 int scsi_prep_fn(struct request_queue *q, struct request *req)
1320 {
1321         struct scsi_device *sdev = q->queuedata;
1322         int ret = BLKPREP_KILL;
1323
1324         if (req->cmd_type == REQ_TYPE_BLOCK_PC)
1325                 ret = scsi_setup_blk_pc_cmnd(sdev, req);
1326         return scsi_prep_return(q, req, ret);
1327 }
1328
1329 /*
1330  * scsi_dev_queue_ready: if we can send requests to sdev, return 1 else
1331  * return 0.
1332  *
1333  * Called with the queue_lock held.
1334  */
1335 static inline int scsi_dev_queue_ready(struct request_queue *q,
1336                                   struct scsi_device *sdev)
1337 {
1338         if (sdev->device_busy >= sdev->queue_depth)
1339                 return 0;
1340         if (sdev->device_busy == 0 && sdev->device_blocked) {
1341                 /*
1342                  * unblock after device_blocked iterates to zero
1343                  */
1344                 if (--sdev->device_blocked == 0) {
1345                         SCSI_LOG_MLQUEUE(3,
1346                                    sdev_printk(KERN_INFO, sdev,
1347                                    "unblocking device at zero depth\n"));
1348                 } else {
1349                         blk_plug_device(q);
1350                         return 0;
1351                 }
1352         }
1353         if (sdev->device_blocked)
1354                 return 0;
1355
1356         return 1;
1357 }
1358
1359 /*
1360  * scsi_host_queue_ready: if we can send requests to shost, return 1 else
1361  * return 0. We must end up running the queue again whenever 0 is
1362  * returned, else IO can hang.
1363  *
1364  * Called with host_lock held.
1365  */
1366 static inline int scsi_host_queue_ready(struct request_queue *q,
1367                                    struct Scsi_Host *shost,
1368                                    struct scsi_device *sdev)
1369 {
1370         if (scsi_host_in_recovery(shost))
1371                 return 0;
1372         if (shost->host_busy == 0 && shost->host_blocked) {
1373                 /*
1374                  * unblock after host_blocked iterates to zero
1375                  */
1376                 if (--shost->host_blocked == 0) {
1377                         SCSI_LOG_MLQUEUE(3,
1378                                 printk("scsi%d unblocking host at zero depth\n",
1379                                         shost->host_no));
1380                 } else {
1381                         blk_plug_device(q);
1382                         return 0;
1383                 }
1384         }
1385         if ((shost->can_queue > 0 && shost->host_busy >= shost->can_queue) ||
1386             shost->host_blocked || shost->host_self_blocked) {
1387                 if (list_empty(&sdev->starved_entry))
1388                         list_add_tail(&sdev->starved_entry, &shost->starved_list);
1389                 return 0;
1390         }
1391
1392         /* We're OK to process the command, so we can't be starved */
1393         if (!list_empty(&sdev->starved_entry))
1394                 list_del_init(&sdev->starved_entry);
1395
1396         return 1;
1397 }
1398
1399 /*
1400  * Kill a request for a dead device
1401  */
1402 static void scsi_kill_request(struct request *req, struct request_queue *q)
1403 {
1404         struct scsi_cmnd *cmd = req->special;
1405         struct scsi_device *sdev = cmd->device;
1406         struct Scsi_Host *shost = sdev->host;
1407
1408         blkdev_dequeue_request(req);
1409
1410         if (unlikely(cmd == NULL)) {
1411                 printk(KERN_CRIT "impossible request in %s.\n",
1412                                  __FUNCTION__);
1413                 BUG();
1414         }
1415
1416         scsi_init_cmd_errh(cmd);
1417         cmd->result = DID_NO_CONNECT << 16;
1418         atomic_inc(&cmd->device->iorequest_cnt);
1419
1420         /*
1421          * SCSI request completion path will do scsi_device_unbusy(),
1422          * bump busy counts.  To bump the counters, we need to dance
1423          * with the locks as normal issue path does.
1424          */
1425         sdev->device_busy++;
1426         spin_unlock(sdev->request_queue->queue_lock);
1427         spin_lock(shost->host_lock);
1428         shost->host_busy++;
1429         spin_unlock(shost->host_lock);
1430         spin_lock(sdev->request_queue->queue_lock);
1431
1432         __scsi_done(cmd);
1433 }
1434
1435 static void scsi_softirq_done(struct request *rq)
1436 {
1437         struct scsi_cmnd *cmd = rq->completion_data;
1438         unsigned long wait_for = (cmd->allowed + 1) * cmd->timeout_per_command;
1439         int disposition;
1440
1441         INIT_LIST_HEAD(&cmd->eh_entry);
1442
1443         disposition = scsi_decide_disposition(cmd);
1444         if (disposition != SUCCESS &&
1445             time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) {
1446                 sdev_printk(KERN_ERR, cmd->device,
1447                             "timing out command, waited %lus\n",
1448                             wait_for/HZ);
1449                 disposition = SUCCESS;
1450         }
1451                         
1452         scsi_log_completion(cmd, disposition);
1453
1454         switch (disposition) {
1455                 case SUCCESS:
1456                         scsi_finish_command(cmd);
1457                         break;
1458                 case NEEDS_RETRY:
1459                         scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY);
1460                         break;
1461                 case ADD_TO_MLQUEUE:
1462                         scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY);
1463                         break;
1464                 default:
1465                         if (!scsi_eh_scmd_add(cmd, 0))
1466                                 scsi_finish_command(cmd);
1467         }
1468 }
1469
1470 /*
1471  * Function:    scsi_request_fn()
1472  *
1473  * Purpose:     Main strategy routine for SCSI.
1474  *
1475  * Arguments:   q       - Pointer to actual queue.
1476  *
1477  * Returns:     Nothing
1478  *
1479  * Lock status: IO request lock assumed to be held when called.
1480  */
1481 static void scsi_request_fn(struct request_queue *q)
1482 {
1483         struct scsi_device *sdev = q->queuedata;
1484         struct Scsi_Host *shost;
1485         struct scsi_cmnd *cmd;
1486         struct request *req;
1487
1488         if (!sdev) {
1489                 printk("scsi: killing requests for dead queue\n");
1490                 while ((req = elv_next_request(q)) != NULL)
1491                         scsi_kill_request(req, q);
1492                 return;
1493         }
1494
1495         if(!get_device(&sdev->sdev_gendev))
1496                 /* We must be tearing the block queue down already */
1497                 return;
1498
1499         /*
1500          * To start with, we keep looping until the queue is empty, or until
1501          * the host is no longer able to accept any more requests.
1502          */
1503         shost = sdev->host;
1504         while (!blk_queue_plugged(q)) {
1505                 int rtn;
1506                 /*
1507                  * get next queueable request.  We do this early to make sure
1508                  * that the request is fully prepared even if we cannot 
1509                  * accept it.
1510                  */
1511                 req = elv_next_request(q);
1512                 if (!req || !scsi_dev_queue_ready(q, sdev))
1513                         break;
1514
1515                 if (unlikely(!scsi_device_online(sdev))) {
1516                         sdev_printk(KERN_ERR, sdev,
1517                                     "rejecting I/O to offline device\n");
1518                         scsi_kill_request(req, q);
1519                         continue;
1520                 }
1521
1522
1523                 /*
1524                  * Remove the request from the request list.
1525                  */
1526                 if (!(blk_queue_tagged(q) && !blk_queue_start_tag(q, req)))
1527                         blkdev_dequeue_request(req);
1528                 sdev->device_busy++;
1529
1530                 spin_unlock(q->queue_lock);
1531                 cmd = req->special;
1532                 if (unlikely(cmd == NULL)) {
1533                         printk(KERN_CRIT "impossible request in %s.\n"
1534                                          "please mail a stack trace to "
1535                                          "linux-scsi@vger.kernel.org\n",
1536                                          __FUNCTION__);
1537                         blk_dump_rq_flags(req, "foo");
1538                         BUG();
1539                 }
1540                 spin_lock(shost->host_lock);
1541
1542                 if (!scsi_host_queue_ready(q, shost, sdev))
1543                         goto not_ready;
1544                 if (scsi_target(sdev)->single_lun) {
1545                         if (scsi_target(sdev)->starget_sdev_user &&
1546                             scsi_target(sdev)->starget_sdev_user != sdev)
1547                                 goto not_ready;
1548                         scsi_target(sdev)->starget_sdev_user = sdev;
1549                 }
1550                 shost->host_busy++;
1551
1552                 /*
1553                  * XXX(hch): This is rather suboptimal, scsi_dispatch_cmd will
1554                  *              take the lock again.
1555                  */
1556                 spin_unlock_irq(shost->host_lock);
1557
1558                 /*
1559                  * Finally, initialize any error handling parameters, and set up
1560                  * the timers for timeouts.
1561                  */
1562                 scsi_init_cmd_errh(cmd);
1563
1564                 /*
1565                  * Dispatch the command to the low-level driver.
1566                  */
1567                 rtn = scsi_dispatch_cmd(cmd);
1568                 spin_lock_irq(q->queue_lock);
1569                 if(rtn) {
1570                         /* we're refusing the command; because of
1571                          * the way locks get dropped, we need to 
1572                          * check here if plugging is required */
1573                         if(sdev->device_busy == 0)
1574                                 blk_plug_device(q);
1575
1576                         break;
1577                 }
1578         }
1579
1580         goto out;
1581
1582  not_ready:
1583         spin_unlock_irq(shost->host_lock);
1584
1585         /*
1586          * lock q, handle tag, requeue req, and decrement device_busy. We
1587          * must return with queue_lock held.
1588          *
1589          * Decrementing device_busy without checking it is OK, as all such
1590          * cases (host limits or settings) should run the queue at some
1591          * later time.
1592          */
1593         spin_lock_irq(q->queue_lock);
1594         blk_requeue_request(q, req);
1595         sdev->device_busy--;
1596         if(sdev->device_busy == 0)
1597                 blk_plug_device(q);
1598  out:
1599         /* must be careful here...if we trigger the ->remove() function
1600          * we cannot be holding the q lock */
1601         spin_unlock_irq(q->queue_lock);
1602         put_device(&sdev->sdev_gendev);
1603         spin_lock_irq(q->queue_lock);
1604 }
1605
1606 u64 scsi_calculate_bounce_limit(struct Scsi_Host *shost)
1607 {
1608         struct device *host_dev;
1609         u64 bounce_limit = 0xffffffff;
1610
1611         if (shost->unchecked_isa_dma)
1612                 return BLK_BOUNCE_ISA;
1613         /*
1614          * Platforms with virtual-DMA translation
1615          * hardware have no practical limit.
1616          */
1617         if (!PCI_DMA_BUS_IS_PHYS)
1618                 return BLK_BOUNCE_ANY;
1619
1620         host_dev = scsi_get_device(shost);
1621         if (host_dev && host_dev->dma_mask)
1622                 bounce_limit = *host_dev->dma_mask;
1623
1624         return bounce_limit;
1625 }
1626 EXPORT_SYMBOL(scsi_calculate_bounce_limit);
1627
1628 struct request_queue *__scsi_alloc_queue(struct Scsi_Host *shost,
1629                                          request_fn_proc *request_fn)
1630 {
1631         struct request_queue *q;
1632
1633         q = blk_init_queue(request_fn, NULL);
1634         if (!q)
1635                 return NULL;
1636
1637         /*
1638          * this limit is imposed by hardware restrictions
1639          */
1640         blk_queue_max_hw_segments(q, shost->sg_tablesize);
1641
1642         /*
1643          * In the future, sg chaining support will be mandatory and this
1644          * ifdef can then go away. Right now we don't have all archs
1645          * converted, so better keep it safe.
1646          */
1647 #ifdef ARCH_HAS_SG_CHAIN
1648         if (shost->use_sg_chaining)
1649                 blk_queue_max_phys_segments(q, SCSI_MAX_SG_CHAIN_SEGMENTS);
1650         else
1651                 blk_queue_max_phys_segments(q, SCSI_MAX_SG_SEGMENTS);
1652 #else
1653         blk_queue_max_phys_segments(q, SCSI_MAX_SG_SEGMENTS);
1654 #endif
1655
1656         blk_queue_max_sectors(q, shost->max_sectors);
1657         blk_queue_bounce_limit(q, scsi_calculate_bounce_limit(shost));
1658         blk_queue_segment_boundary(q, shost->dma_boundary);
1659
1660         if (!shost->use_clustering)
1661                 clear_bit(QUEUE_FLAG_CLUSTER, &q->queue_flags);
1662
1663         /*
1664          * set a reasonable default alignment on word boundaries: the
1665          * host and device may alter it using
1666          * blk_queue_update_dma_alignment() later.
1667          */
1668         blk_queue_dma_alignment(q, 0x03);
1669
1670         return q;
1671 }
1672 EXPORT_SYMBOL(__scsi_alloc_queue);
1673
1674 struct request_queue *scsi_alloc_queue(struct scsi_device *sdev)
1675 {
1676         struct request_queue *q;
1677
1678         q = __scsi_alloc_queue(sdev->host, scsi_request_fn);
1679         if (!q)
1680                 return NULL;
1681
1682         blk_queue_prep_rq(q, scsi_prep_fn);
1683         blk_queue_softirq_done(q, scsi_softirq_done);
1684         return q;
1685 }
1686
1687 void scsi_free_queue(struct request_queue *q)
1688 {
1689         blk_cleanup_queue(q);
1690 }
1691
1692 /*
1693  * Function:    scsi_block_requests()
1694  *
1695  * Purpose:     Utility function used by low-level drivers to prevent further
1696  *              commands from being queued to the device.
1697  *
1698  * Arguments:   shost       - Host in question
1699  *
1700  * Returns:     Nothing
1701  *
1702  * Lock status: No locks are assumed held.
1703  *
1704  * Notes:       There is no timer nor any other means by which the requests
1705  *              get unblocked other than the low-level driver calling
1706  *              scsi_unblock_requests().
1707  */
1708 void scsi_block_requests(struct Scsi_Host *shost)
1709 {
1710         shost->host_self_blocked = 1;
1711 }
1712 EXPORT_SYMBOL(scsi_block_requests);
1713
1714 /*
1715  * Function:    scsi_unblock_requests()
1716  *
1717  * Purpose:     Utility function used by low-level drivers to allow further
1718  *              commands from being queued to the device.
1719  *
1720  * Arguments:   shost       - Host in question
1721  *
1722  * Returns:     Nothing
1723  *
1724  * Lock status: No locks are assumed held.
1725  *
1726  * Notes:       There is no timer nor any other means by which the requests
1727  *              get unblocked other than the low-level driver calling
1728  *              scsi_unblock_requests().
1729  *
1730  *              This is done as an API function so that changes to the
1731  *              internals of the scsi mid-layer won't require wholesale
1732  *              changes to drivers that use this feature.
1733  */
1734 void scsi_unblock_requests(struct Scsi_Host *shost)
1735 {
1736         shost->host_self_blocked = 0;
1737         scsi_run_host_queues(shost);
1738 }
1739 EXPORT_SYMBOL(scsi_unblock_requests);
1740
1741 int __init scsi_init_queue(void)
1742 {
1743         int i;
1744
1745         scsi_io_context_cache = kmem_cache_create("scsi_io_context",
1746                                         sizeof(struct scsi_io_context),
1747                                         0, 0, NULL);
1748         if (!scsi_io_context_cache) {
1749                 printk(KERN_ERR "SCSI: can't init scsi io context cache\n");
1750                 return -ENOMEM;
1751         }
1752
1753         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1754                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1755                 int size = sgp->size * sizeof(struct scatterlist);
1756
1757                 sgp->slab = kmem_cache_create(sgp->name, size, 0,
1758                                 SLAB_HWCACHE_ALIGN, NULL);
1759                 if (!sgp->slab) {
1760                         printk(KERN_ERR "SCSI: can't init sg slab %s\n",
1761                                         sgp->name);
1762                 }
1763
1764                 sgp->pool = mempool_create_slab_pool(SG_MEMPOOL_SIZE,
1765                                                      sgp->slab);
1766                 if (!sgp->pool) {
1767                         printk(KERN_ERR "SCSI: can't init sg mempool %s\n",
1768                                         sgp->name);
1769                 }
1770         }
1771
1772         return 0;
1773 }
1774
1775 void scsi_exit_queue(void)
1776 {
1777         int i;
1778
1779         kmem_cache_destroy(scsi_io_context_cache);
1780
1781         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1782                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1783                 mempool_destroy(sgp->pool);
1784                 kmem_cache_destroy(sgp->slab);
1785         }
1786 }
1787
1788 /**
1789  *      scsi_mode_select - issue a mode select
1790  *      @sdev:  SCSI device to be queried
1791  *      @pf:    Page format bit (1 == standard, 0 == vendor specific)
1792  *      @sp:    Save page bit (0 == don't save, 1 == save)
1793  *      @modepage: mode page being requested
1794  *      @buffer: request buffer (may not be smaller than eight bytes)
1795  *      @len:   length of request buffer.
1796  *      @timeout: command timeout
1797  *      @retries: number of retries before failing
1798  *      @data: returns a structure abstracting the mode header data
1799  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
1800  *              must be SCSI_SENSE_BUFFERSIZE big.
1801  *
1802  *      Returns zero if successful; negative error number or scsi
1803  *      status on error
1804  *
1805  */
1806 int
1807 scsi_mode_select(struct scsi_device *sdev, int pf, int sp, int modepage,
1808                  unsigned char *buffer, int len, int timeout, int retries,
1809                  struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1810 {
1811         unsigned char cmd[10];
1812         unsigned char *real_buffer;
1813         int ret;
1814
1815         memset(cmd, 0, sizeof(cmd));
1816         cmd[1] = (pf ? 0x10 : 0) | (sp ? 0x01 : 0);
1817
1818         if (sdev->use_10_for_ms) {
1819                 if (len > 65535)
1820                         return -EINVAL;
1821                 real_buffer = kmalloc(8 + len, GFP_KERNEL);
1822                 if (!real_buffer)
1823                         return -ENOMEM;
1824                 memcpy(real_buffer + 8, buffer, len);
1825                 len += 8;
1826                 real_buffer[0] = 0;
1827                 real_buffer[1] = 0;
1828                 real_buffer[2] = data->medium_type;
1829                 real_buffer[3] = data->device_specific;
1830                 real_buffer[4] = data->longlba ? 0x01 : 0;
1831                 real_buffer[5] = 0;
1832                 real_buffer[6] = data->block_descriptor_length >> 8;
1833                 real_buffer[7] = data->block_descriptor_length;
1834
1835                 cmd[0] = MODE_SELECT_10;
1836                 cmd[7] = len >> 8;
1837                 cmd[8] = len;
1838         } else {
1839                 if (len > 255 || data->block_descriptor_length > 255 ||
1840                     data->longlba)
1841                         return -EINVAL;
1842
1843                 real_buffer = kmalloc(4 + len, GFP_KERNEL);
1844                 if (!real_buffer)
1845                         return -ENOMEM;
1846                 memcpy(real_buffer + 4, buffer, len);
1847                 len += 4;
1848                 real_buffer[0] = 0;
1849                 real_buffer[1] = data->medium_type;
1850                 real_buffer[2] = data->device_specific;
1851                 real_buffer[3] = data->block_descriptor_length;
1852                 
1853
1854                 cmd[0] = MODE_SELECT;
1855                 cmd[4] = len;
1856         }
1857
1858         ret = scsi_execute_req(sdev, cmd, DMA_TO_DEVICE, real_buffer, len,
1859                                sshdr, timeout, retries);
1860         kfree(real_buffer);
1861         return ret;
1862 }
1863 EXPORT_SYMBOL_GPL(scsi_mode_select);
1864
1865 /**
1866  *      scsi_mode_sense - issue a mode sense, falling back from 10 to six bytes if necessary.
1867  *      @sdev:  SCSI device to be queried
1868  *      @dbd:   set if mode sense will allow block descriptors to be returned
1869  *      @modepage: mode page being requested
1870  *      @buffer: request buffer (may not be smaller than eight bytes)
1871  *      @len:   length of request buffer.
1872  *      @timeout: command timeout
1873  *      @retries: number of retries before failing
1874  *      @data: returns a structure abstracting the mode header data
1875  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
1876  *              must be SCSI_SENSE_BUFFERSIZE big.
1877  *
1878  *      Returns zero if unsuccessful, or the header offset (either 4
1879  *      or 8 depending on whether a six or ten byte command was
1880  *      issued) if successful.
1881  */
1882 int
1883 scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
1884                   unsigned char *buffer, int len, int timeout, int retries,
1885                   struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1886 {
1887         unsigned char cmd[12];
1888         int use_10_for_ms;
1889         int header_length;
1890         int result;
1891         struct scsi_sense_hdr my_sshdr;
1892
1893         memset(data, 0, sizeof(*data));
1894         memset(&cmd[0], 0, 12);
1895         cmd[1] = dbd & 0x18;    /* allows DBD and LLBA bits */
1896         cmd[2] = modepage;
1897
1898         /* caller might not be interested in sense, but we need it */
1899         if (!sshdr)
1900                 sshdr = &my_sshdr;
1901
1902  retry:
1903         use_10_for_ms = sdev->use_10_for_ms;
1904
1905         if (use_10_for_ms) {
1906                 if (len < 8)
1907                         len = 8;
1908
1909                 cmd[0] = MODE_SENSE_10;
1910                 cmd[8] = len;
1911                 header_length = 8;
1912         } else {
1913                 if (len < 4)
1914                         len = 4;
1915
1916                 cmd[0] = MODE_SENSE;
1917                 cmd[4] = len;
1918                 header_length = 4;
1919         }
1920
1921         memset(buffer, 0, len);
1922
1923         result = scsi_execute_req(sdev, cmd, DMA_FROM_DEVICE, buffer, len,
1924                                   sshdr, timeout, retries);
1925
1926         /* This code looks awful: what it's doing is making sure an
1927          * ILLEGAL REQUEST sense return identifies the actual command
1928          * byte as the problem.  MODE_SENSE commands can return
1929          * ILLEGAL REQUEST if the code page isn't supported */
1930
1931         if (use_10_for_ms && !scsi_status_is_good(result) &&
1932             (driver_byte(result) & DRIVER_SENSE)) {
1933                 if (scsi_sense_valid(sshdr)) {
1934                         if ((sshdr->sense_key == ILLEGAL_REQUEST) &&
1935                             (sshdr->asc == 0x20) && (sshdr->ascq == 0)) {
1936                                 /* 
1937                                  * Invalid command operation code
1938                                  */
1939                                 sdev->use_10_for_ms = 0;
1940                                 goto retry;
1941                         }
1942                 }
1943         }
1944
1945         if(scsi_status_is_good(result)) {
1946                 if (unlikely(buffer[0] == 0x86 && buffer[1] == 0x0b &&
1947                              (modepage == 6 || modepage == 8))) {
1948                         /* Initio breakage? */
1949                         header_length = 0;
1950                         data->length = 13;
1951                         data->medium_type = 0;
1952                         data->device_specific = 0;
1953                         data->longlba = 0;
1954                         data->block_descriptor_length = 0;
1955                 } else if(use_10_for_ms) {
1956                         data->length = buffer[0]*256 + buffer[1] + 2;
1957                         data->medium_type = buffer[2];
1958                         data->device_specific = buffer[3];
1959                         data->longlba = buffer[4] & 0x01;
1960                         data->block_descriptor_length = buffer[6]*256
1961                                 + buffer[7];
1962                 } else {
1963                         data->length = buffer[0] + 1;
1964                         data->medium_type = buffer[1];
1965                         data->device_specific = buffer[2];
1966                         data->block_descriptor_length = buffer[3];
1967                 }
1968                 data->header_length = header_length;
1969         }
1970
1971         return result;
1972 }
1973 EXPORT_SYMBOL(scsi_mode_sense);
1974
1975 /**
1976  *      scsi_test_unit_ready - test if unit is ready
1977  *      @sdev:  scsi device to change the state of.
1978  *      @timeout: command timeout
1979  *      @retries: number of retries before failing
1980  *      @sshdr_external: Optional pointer to struct scsi_sense_hdr for
1981  *              returning sense. Make sure that this is cleared before passing
1982  *              in.
1983  *
1984  *      Returns zero if unsuccessful or an error if TUR failed.  For
1985  *      removable media, a return of NOT_READY or UNIT_ATTENTION is
1986  *      translated to success, with the ->changed flag updated.
1987  **/
1988 int
1989 scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries,
1990                      struct scsi_sense_hdr *sshdr_external)
1991 {
1992         char cmd[] = {
1993                 TEST_UNIT_READY, 0, 0, 0, 0, 0,
1994         };
1995         struct scsi_sense_hdr *sshdr;
1996         int result;
1997
1998         if (!sshdr_external)
1999                 sshdr = kzalloc(sizeof(*sshdr), GFP_KERNEL);
2000         else
2001                 sshdr = sshdr_external;
2002
2003         /* try to eat the UNIT_ATTENTION if there are enough retries */
2004         do {
2005                 result = scsi_execute_req(sdev, cmd, DMA_NONE, NULL, 0, sshdr,
2006                                           timeout, retries);
2007         } while ((driver_byte(result) & DRIVER_SENSE) &&
2008                  sshdr && sshdr->sense_key == UNIT_ATTENTION &&
2009                  --retries);
2010
2011         if (!sshdr)
2012                 /* could not allocate sense buffer, so can't process it */
2013                 return result;
2014
2015         if ((driver_byte(result) & DRIVER_SENSE) && sdev->removable) {
2016
2017                 if ((scsi_sense_valid(sshdr)) &&
2018                     ((sshdr->sense_key == UNIT_ATTENTION) ||
2019                      (sshdr->sense_key == NOT_READY))) {
2020                         sdev->changed = 1;
2021                         result = 0;
2022                 }
2023         }
2024         if (!sshdr_external)
2025                 kfree(sshdr);
2026         return result;
2027 }
2028 EXPORT_SYMBOL(scsi_test_unit_ready);
2029
2030 /**
2031  *      scsi_device_set_state - Take the given device through the device state model.
2032  *      @sdev:  scsi device to change the state of.
2033  *      @state: state to change to.
2034  *
2035  *      Returns zero if unsuccessful or an error if the requested 
2036  *      transition is illegal.
2037  */
2038 int
2039 scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
2040 {
2041         enum scsi_device_state oldstate = sdev->sdev_state;
2042
2043         if (state == oldstate)
2044                 return 0;
2045
2046         switch (state) {
2047         case SDEV_CREATED:
2048                 /* There are no legal states that come back to
2049                  * created.  This is the manually initialised start
2050                  * state */
2051                 goto illegal;
2052                         
2053         case SDEV_RUNNING:
2054                 switch (oldstate) {
2055                 case SDEV_CREATED:
2056                 case SDEV_OFFLINE:
2057                 case SDEV_QUIESCE:
2058                 case SDEV_BLOCK:
2059                         break;
2060                 default:
2061                         goto illegal;
2062                 }
2063                 break;
2064
2065         case SDEV_QUIESCE:
2066                 switch (oldstate) {
2067                 case SDEV_RUNNING:
2068                 case SDEV_OFFLINE:
2069                         break;
2070                 default:
2071                         goto illegal;
2072                 }
2073                 break;
2074
2075         case SDEV_OFFLINE:
2076                 switch (oldstate) {
2077                 case SDEV_CREATED:
2078                 case SDEV_RUNNING:
2079                 case SDEV_QUIESCE:
2080                 case SDEV_BLOCK:
2081                         break;
2082                 default:
2083                         goto illegal;
2084                 }
2085                 break;
2086
2087         case SDEV_BLOCK:
2088                 switch (oldstate) {
2089                 case SDEV_CREATED:
2090                 case SDEV_RUNNING:
2091                         break;
2092                 default:
2093                         goto illegal;
2094                 }
2095                 break;
2096
2097         case SDEV_CANCEL:
2098                 switch (oldstate) {
2099                 case SDEV_CREATED:
2100                 case SDEV_RUNNING:
2101                 case SDEV_QUIESCE:
2102                 case SDEV_OFFLINE:
2103                 case SDEV_BLOCK:
2104                         break;
2105                 default:
2106                         goto illegal;
2107                 }
2108                 break;
2109
2110         case SDEV_DEL:
2111                 switch (oldstate) {
2112                 case SDEV_CREATED:
2113                 case SDEV_RUNNING:
2114                 case SDEV_OFFLINE:
2115                 case SDEV_CANCEL:
2116                         break;
2117                 default:
2118                         goto illegal;
2119                 }
2120                 break;
2121
2122         }
2123         sdev->sdev_state = state;
2124         return 0;
2125
2126  illegal:
2127         SCSI_LOG_ERROR_RECOVERY(1, 
2128                                 sdev_printk(KERN_ERR, sdev,
2129                                             "Illegal state transition %s->%s\n",
2130                                             scsi_device_state_name(oldstate),
2131                                             scsi_device_state_name(state))
2132                                 );
2133         return -EINVAL;
2134 }
2135 EXPORT_SYMBOL(scsi_device_set_state);
2136
2137 /**
2138  *      sdev_evt_emit - emit a single SCSI device uevent
2139  *      @sdev: associated SCSI device
2140  *      @evt: event to emit
2141  *
2142  *      Send a single uevent (scsi_event) to the associated scsi_device.
2143  */
2144 static void scsi_evt_emit(struct scsi_device *sdev, struct scsi_event *evt)
2145 {
2146         int idx = 0;
2147         char *envp[3];
2148
2149         switch (evt->evt_type) {
2150         case SDEV_EVT_MEDIA_CHANGE:
2151                 envp[idx++] = "SDEV_MEDIA_CHANGE=1";
2152                 break;
2153
2154         default:
2155                 /* do nothing */
2156                 break;
2157         }
2158
2159         envp[idx++] = NULL;
2160
2161         kobject_uevent_env(&sdev->sdev_gendev.kobj, KOBJ_CHANGE, envp);
2162 }
2163
2164 /**
2165  *      sdev_evt_thread - send a uevent for each scsi event
2166  *      @work: work struct for scsi_device
2167  *
2168  *      Dispatch queued events to their associated scsi_device kobjects
2169  *      as uevents.
2170  */
2171 void scsi_evt_thread(struct work_struct *work)
2172 {
2173         struct scsi_device *sdev;
2174         LIST_HEAD(event_list);
2175
2176         sdev = container_of(work, struct scsi_device, event_work);
2177
2178         while (1) {
2179                 struct scsi_event *evt;
2180                 struct list_head *this, *tmp;
2181                 unsigned long flags;
2182
2183                 spin_lock_irqsave(&sdev->list_lock, flags);
2184                 list_splice_init(&sdev->event_list, &event_list);
2185                 spin_unlock_irqrestore(&sdev->list_lock, flags);
2186
2187                 if (list_empty(&event_list))
2188                         break;
2189
2190                 list_for_each_safe(this, tmp, &event_list) {
2191                         evt = list_entry(this, struct scsi_event, node);
2192                         list_del(&evt->node);
2193                         scsi_evt_emit(sdev, evt);
2194                         kfree(evt);
2195                 }
2196         }
2197 }
2198
2199 /**
2200  *      sdev_evt_send - send asserted event to uevent thread
2201  *      @sdev: scsi_device event occurred on
2202  *      @evt: event to send
2203  *
2204  *      Assert scsi device event asynchronously.
2205  */
2206 void sdev_evt_send(struct scsi_device *sdev, struct scsi_event *evt)
2207 {
2208         unsigned long flags;
2209
2210         if (!test_bit(evt->evt_type, sdev->supported_events)) {
2211                 kfree(evt);
2212                 return;
2213         }
2214
2215         spin_lock_irqsave(&sdev->list_lock, flags);
2216         list_add_tail(&evt->node, &sdev->event_list);
2217         schedule_work(&sdev->event_work);
2218         spin_unlock_irqrestore(&sdev->list_lock, flags);
2219 }
2220 EXPORT_SYMBOL_GPL(sdev_evt_send);
2221
2222 /**
2223  *      sdev_evt_alloc - allocate a new scsi event
2224  *      @evt_type: type of event to allocate
2225  *      @gfpflags: GFP flags for allocation
2226  *
2227  *      Allocates and returns a new scsi_event.
2228  */
2229 struct scsi_event *sdev_evt_alloc(enum scsi_device_event evt_type,
2230                                   gfp_t gfpflags)
2231 {
2232         struct scsi_event *evt = kzalloc(sizeof(struct scsi_event), gfpflags);
2233         if (!evt)
2234                 return NULL;
2235
2236         evt->evt_type = evt_type;
2237         INIT_LIST_HEAD(&evt->node);
2238
2239         /* evt_type-specific initialization, if any */
2240         switch (evt_type) {
2241         case SDEV_EVT_MEDIA_CHANGE:
2242         default:
2243                 /* do nothing */
2244                 break;
2245         }
2246
2247         return evt;
2248 }
2249 EXPORT_SYMBOL_GPL(sdev_evt_alloc);
2250
2251 /**
2252  *      sdev_evt_send_simple - send asserted event to uevent thread
2253  *      @sdev: scsi_device event occurred on
2254  *      @evt_type: type of event to send
2255  *      @gfpflags: GFP flags for allocation
2256  *
2257  *      Assert scsi device event asynchronously, given an event type.
2258  */
2259 void sdev_evt_send_simple(struct scsi_device *sdev,
2260                           enum scsi_device_event evt_type, gfp_t gfpflags)
2261 {
2262         struct scsi_event *evt = sdev_evt_alloc(evt_type, gfpflags);
2263         if (!evt) {
2264                 sdev_printk(KERN_ERR, sdev, "event %d eaten due to OOM\n",
2265                             evt_type);
2266                 return;
2267         }
2268
2269         sdev_evt_send(sdev, evt);
2270 }
2271 EXPORT_SYMBOL_GPL(sdev_evt_send_simple);
2272
2273 /**
2274  *      scsi_device_quiesce - Block user issued commands.
2275  *      @sdev:  scsi device to quiesce.
2276  *
2277  *      This works by trying to transition to the SDEV_QUIESCE state
2278  *      (which must be a legal transition).  When the device is in this
2279  *      state, only special requests will be accepted, all others will
2280  *      be deferred.  Since special requests may also be requeued requests,
2281  *      a successful return doesn't guarantee the device will be 
2282  *      totally quiescent.
2283  *
2284  *      Must be called with user context, may sleep.
2285  *
2286  *      Returns zero if unsuccessful or an error if not.
2287  */
2288 int
2289 scsi_device_quiesce(struct scsi_device *sdev)
2290 {
2291         int err = scsi_device_set_state(sdev, SDEV_QUIESCE);
2292         if (err)
2293                 return err;
2294
2295         scsi_run_queue(sdev->request_queue);
2296         while (sdev->device_busy) {
2297                 msleep_interruptible(200);
2298                 scsi_run_queue(sdev->request_queue);
2299         }
2300         return 0;
2301 }
2302 EXPORT_SYMBOL(scsi_device_quiesce);
2303
2304 /**
2305  *      scsi_device_resume - Restart user issued commands to a quiesced device.
2306  *      @sdev:  scsi device to resume.
2307  *
2308  *      Moves the device from quiesced back to running and restarts the
2309  *      queues.
2310  *
2311  *      Must be called with user context, may sleep.
2312  */
2313 void
2314 scsi_device_resume(struct scsi_device *sdev)
2315 {
2316         if(scsi_device_set_state(sdev, SDEV_RUNNING))
2317                 return;
2318         scsi_run_queue(sdev->request_queue);
2319 }
2320 EXPORT_SYMBOL(scsi_device_resume);
2321
2322 static void
2323 device_quiesce_fn(struct scsi_device *sdev, void *data)
2324 {
2325         scsi_device_quiesce(sdev);
2326 }
2327
2328 void
2329 scsi_target_quiesce(struct scsi_target *starget)
2330 {
2331         starget_for_each_device(starget, NULL, device_quiesce_fn);
2332 }
2333 EXPORT_SYMBOL(scsi_target_quiesce);
2334
2335 static void
2336 device_resume_fn(struct scsi_device *sdev, void *data)
2337 {
2338         scsi_device_resume(sdev);
2339 }
2340
2341 void
2342 scsi_target_resume(struct scsi_target *starget)
2343 {
2344         starget_for_each_device(starget, NULL, device_resume_fn);
2345 }
2346 EXPORT_SYMBOL(scsi_target_resume);
2347
2348 /**
2349  * scsi_internal_device_block - internal function to put a device temporarily into the SDEV_BLOCK state
2350  * @sdev:       device to block
2351  *
2352  * Block request made by scsi lld's to temporarily stop all
2353  * scsi commands on the specified device.  Called from interrupt
2354  * or normal process context.
2355  *
2356  * Returns zero if successful or error if not
2357  *
2358  * Notes:       
2359  *      This routine transitions the device to the SDEV_BLOCK state
2360  *      (which must be a legal transition).  When the device is in this
2361  *      state, all commands are deferred until the scsi lld reenables
2362  *      the device with scsi_device_unblock or device_block_tmo fires.
2363  *      This routine assumes the host_lock is held on entry.
2364  */
2365 int
2366 scsi_internal_device_block(struct scsi_device *sdev)
2367 {
2368         struct request_queue *q = sdev->request_queue;
2369         unsigned long flags;
2370         int err = 0;
2371
2372         err = scsi_device_set_state(sdev, SDEV_BLOCK);
2373         if (err)
2374                 return err;
2375
2376         /* 
2377          * The device has transitioned to SDEV_BLOCK.  Stop the
2378          * block layer from calling the midlayer with this device's
2379          * request queue. 
2380          */
2381         spin_lock_irqsave(q->queue_lock, flags);
2382         blk_stop_queue(q);
2383         spin_unlock_irqrestore(q->queue_lock, flags);
2384
2385         return 0;
2386 }
2387 EXPORT_SYMBOL_GPL(scsi_internal_device_block);
2388  
2389 /**
2390  * scsi_internal_device_unblock - resume a device after a block request
2391  * @sdev:       device to resume
2392  *
2393  * Called by scsi lld's or the midlayer to restart the device queue
2394  * for the previously suspended scsi device.  Called from interrupt or
2395  * normal process context.
2396  *
2397  * Returns zero if successful or error if not.
2398  *
2399  * Notes:       
2400  *      This routine transitions the device to the SDEV_RUNNING state
2401  *      (which must be a legal transition) allowing the midlayer to
2402  *      goose the queue for this device.  This routine assumes the 
2403  *      host_lock is held upon entry.
2404  */
2405 int
2406 scsi_internal_device_unblock(struct scsi_device *sdev)
2407 {
2408         struct request_queue *q = sdev->request_queue; 
2409         int err;
2410         unsigned long flags;
2411         
2412         /* 
2413          * Try to transition the scsi device to SDEV_RUNNING
2414          * and goose the device queue if successful.  
2415          */
2416         err = scsi_device_set_state(sdev, SDEV_RUNNING);
2417         if (err)
2418                 return err;
2419
2420         spin_lock_irqsave(q->queue_lock, flags);
2421         blk_start_queue(q);
2422         spin_unlock_irqrestore(q->queue_lock, flags);
2423
2424         return 0;
2425 }
2426 EXPORT_SYMBOL_GPL(scsi_internal_device_unblock);
2427
2428 static void
2429 device_block(struct scsi_device *sdev, void *data)
2430 {
2431         scsi_internal_device_block(sdev);
2432 }
2433
2434 static int
2435 target_block(struct device *dev, void *data)
2436 {
2437         if (scsi_is_target_device(dev))
2438                 starget_for_each_device(to_scsi_target(dev), NULL,
2439                                         device_block);
2440         return 0;
2441 }
2442
2443 void
2444 scsi_target_block(struct device *dev)
2445 {
2446         if (scsi_is_target_device(dev))
2447                 starget_for_each_device(to_scsi_target(dev), NULL,
2448                                         device_block);
2449         else
2450                 device_for_each_child(dev, NULL, target_block);
2451 }
2452 EXPORT_SYMBOL_GPL(scsi_target_block);
2453
2454 static void
2455 device_unblock(struct scsi_device *sdev, void *data)
2456 {
2457         scsi_internal_device_unblock(sdev);
2458 }
2459
2460 static int
2461 target_unblock(struct device *dev, void *data)
2462 {
2463         if (scsi_is_target_device(dev))
2464                 starget_for_each_device(to_scsi_target(dev), NULL,
2465                                         device_unblock);
2466         return 0;
2467 }
2468
2469 void
2470 scsi_target_unblock(struct device *dev)
2471 {
2472         if (scsi_is_target_device(dev))
2473                 starget_for_each_device(to_scsi_target(dev), NULL,
2474                                         device_unblock);
2475         else
2476                 device_for_each_child(dev, NULL, target_unblock);
2477 }
2478 EXPORT_SYMBOL_GPL(scsi_target_unblock);
2479
2480 /**
2481  * scsi_kmap_atomic_sg - find and atomically map an sg-elemnt
2482  * @sgl:        scatter-gather list
2483  * @sg_count:   number of segments in sg
2484  * @offset:     offset in bytes into sg, on return offset into the mapped area
2485  * @len:        bytes to map, on return number of bytes mapped
2486  *
2487  * Returns virtual address of the start of the mapped page
2488  */
2489 void *scsi_kmap_atomic_sg(struct scatterlist *sgl, int sg_count,
2490                           size_t *offset, size_t *len)
2491 {
2492         int i;
2493         size_t sg_len = 0, len_complete = 0;
2494         struct scatterlist *sg;
2495         struct page *page;
2496
2497         WARN_ON(!irqs_disabled());
2498
2499         for_each_sg(sgl, sg, sg_count, i) {
2500                 len_complete = sg_len; /* Complete sg-entries */
2501                 sg_len += sg->length;
2502                 if (sg_len > *offset)
2503                         break;
2504         }
2505
2506         if (unlikely(i == sg_count)) {
2507                 printk(KERN_ERR "%s: Bytes in sg: %zu, requested offset %zu, "
2508                         "elements %d\n",
2509                        __FUNCTION__, sg_len, *offset, sg_count);
2510                 WARN_ON(1);
2511                 return NULL;
2512         }
2513
2514         /* Offset starting from the beginning of first page in this sg-entry */
2515         *offset = *offset - len_complete + sg->offset;
2516
2517         /* Assumption: contiguous pages can be accessed as "page + i" */
2518         page = nth_page(sg_page(sg), (*offset >> PAGE_SHIFT));
2519         *offset &= ~PAGE_MASK;
2520
2521         /* Bytes in this sg-entry from *offset to the end of the page */
2522         sg_len = PAGE_SIZE - *offset;
2523         if (*len > sg_len)
2524                 *len = sg_len;
2525
2526         return kmap_atomic(page, KM_BIO_SRC_IRQ);
2527 }
2528 EXPORT_SYMBOL(scsi_kmap_atomic_sg);
2529
2530 /**
2531  * scsi_kunmap_atomic_sg - atomically unmap a virtual address, previously mapped with scsi_kmap_atomic_sg
2532  * @virt:       virtual address to be unmapped
2533  */
2534 void scsi_kunmap_atomic_sg(void *virt)
2535 {
2536         kunmap_atomic(virt, KM_BIO_SRC_IRQ);
2537 }
2538 EXPORT_SYMBOL(scsi_kunmap_atomic_sg);