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