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