Pull bugzilla-5764 into release branch
[linux-2.6] / drivers / scsi / scsi_lib.c
1 /*
2  *  scsi_lib.c Copyright (C) 1999 Eric Youngdale
3  *
4  *  SCSI queueing library.
5  *      Initial versions: Eric Youngdale (eric@andante.org).
6  *                        Based upon conversations with large numbers
7  *                        of people at Linux Expo.
8  */
9
10 #include <linux/bio.h>
11 #include <linux/blkdev.h>
12 #include <linux/completion.h>
13 #include <linux/kernel.h>
14 #include <linux/mempool.h>
15 #include <linux/slab.h>
16 #include <linux/init.h>
17 #include <linux/pci.h>
18 #include <linux/delay.h>
19 #include <linux/hardirq.h>
20
21 #include <scsi/scsi.h>
22 #include <scsi/scsi_dbg.h>
23 #include <scsi/scsi_device.h>
24 #include <scsi/scsi_driver.h>
25 #include <scsi/scsi_eh.h>
26 #include <scsi/scsi_host.h>
27 #include <scsi/scsi_request.h>
28
29 #include "scsi_priv.h"
30 #include "scsi_logging.h"
31
32
33 #define SG_MEMPOOL_NR           (sizeof(scsi_sg_pools)/sizeof(struct scsi_host_sg_pool))
34 #define SG_MEMPOOL_SIZE         32
35
36 struct scsi_host_sg_pool {
37         size_t          size;
38         char            *name; 
39         kmem_cache_t    *slab;
40         mempool_t       *pool;
41 };
42
43 #if (SCSI_MAX_PHYS_SEGMENTS < 32)
44 #error SCSI_MAX_PHYS_SEGMENTS is too small
45 #endif
46
47 #define SP(x) { x, "sgpool-" #x } 
48 static struct scsi_host_sg_pool scsi_sg_pools[] = {
49         SP(8),
50         SP(16),
51         SP(32),
52 #if (SCSI_MAX_PHYS_SEGMENTS > 32)
53         SP(64),
54 #if (SCSI_MAX_PHYS_SEGMENTS > 64)
55         SP(128),
56 #if (SCSI_MAX_PHYS_SEGMENTS > 128)
57         SP(256),
58 #if (SCSI_MAX_PHYS_SEGMENTS > 256)
59 #error SCSI_MAX_PHYS_SEGMENTS is too large
60 #endif
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->flags &= ~REQ_DONTPREP;
86         req->special = (req->flags & REQ_SPECIAL) ? cmd->sc_request : 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  * Function:    scsi_do_req
166  *
167  * Purpose:     Queue a SCSI request
168  *
169  * Arguments:   sreq      - command descriptor.
170  *              cmnd      - actual SCSI command to be performed.
171  *              buffer    - data buffer.
172  *              bufflen   - size of data buffer.
173  *              done      - completion function to be run.
174  *              timeout   - how long to let it run before timeout.
175  *              retries   - number of retries we allow.
176  *
177  * Lock status: No locks held upon entry.
178  *
179  * Returns:     Nothing.
180  *
181  * Notes:       This function is only used for queueing requests for things
182  *              like ioctls and character device requests - this is because
183  *              we essentially just inject a request into the queue for the
184  *              device.
185  *
186  *              In order to support the scsi_device_quiesce function, we
187  *              now inject requests on the *head* of the device queue
188  *              rather than the tail.
189  */
190 void scsi_do_req(struct scsi_request *sreq, const void *cmnd,
191                  void *buffer, unsigned bufflen,
192                  void (*done)(struct scsi_cmnd *),
193                  int timeout, int retries)
194 {
195         /*
196          * If the upper level driver is reusing these things, then
197          * we should release the low-level block now.  Another one will
198          * be allocated later when this request is getting queued.
199          */
200         __scsi_release_request(sreq);
201
202         /*
203          * Our own function scsi_done (which marks the host as not busy,
204          * disables the timeout counter, etc) will be called by us or by the
205          * scsi_hosts[host].queuecommand() function needs to also call
206          * the completion function for the high level driver.
207          */
208         memcpy(sreq->sr_cmnd, cmnd, sizeof(sreq->sr_cmnd));
209         sreq->sr_bufflen = bufflen;
210         sreq->sr_buffer = buffer;
211         sreq->sr_allowed = retries;
212         sreq->sr_done = done;
213         sreq->sr_timeout_per_command = timeout;
214
215         if (sreq->sr_cmd_len == 0)
216                 sreq->sr_cmd_len = COMMAND_SIZE(sreq->sr_cmnd[0]);
217
218         /*
219          * head injection *required* here otherwise quiesce won't work
220          *
221          * Because users of this function are apt to reuse requests with no
222          * modification, we have to sanitise the request flags here
223          */
224         sreq->sr_request->flags &= ~REQ_DONTPREP;
225         blk_insert_request(sreq->sr_device->request_queue, sreq->sr_request,
226                            1, sreq);
227 }
228 EXPORT_SYMBOL(scsi_do_req);
229
230 /**
231  * scsi_execute - insert request and wait for the result
232  * @sdev:       scsi device
233  * @cmd:        scsi command
234  * @data_direction: data direction
235  * @buffer:     data buffer
236  * @bufflen:    len of buffer
237  * @sense:      optional sense buffer
238  * @timeout:    request timeout in seconds
239  * @retries:    number of times to retry request
240  * @flags:      or into request flags;
241  *
242  * returns the req->errors value which is the the scsi_cmnd result
243  * field.
244  **/
245 int scsi_execute(struct scsi_device *sdev, const unsigned char *cmd,
246                  int data_direction, void *buffer, unsigned bufflen,
247                  unsigned char *sense, int timeout, int retries, int flags)
248 {
249         struct request *req;
250         int write = (data_direction == DMA_TO_DEVICE);
251         int ret = DRIVER_ERROR << 24;
252
253         req = blk_get_request(sdev->request_queue, write, __GFP_WAIT);
254
255         if (bufflen &&  blk_rq_map_kern(sdev->request_queue, req,
256                                         buffer, bufflen, __GFP_WAIT))
257                 goto out;
258
259         req->cmd_len = COMMAND_SIZE(cmd[0]);
260         memcpy(req->cmd, cmd, req->cmd_len);
261         req->sense = sense;
262         req->sense_len = 0;
263         req->retries = retries;
264         req->timeout = timeout;
265         req->flags |= flags | REQ_BLOCK_PC | REQ_SPECIAL | REQ_QUIET;
266
267         /*
268          * head injection *required* here otherwise quiesce won't work
269          */
270         blk_execute_rq(req->q, NULL, req, 1);
271
272         ret = req->errors;
273  out:
274         blk_put_request(req);
275
276         return ret;
277 }
278 EXPORT_SYMBOL(scsi_execute);
279
280
281 int scsi_execute_req(struct scsi_device *sdev, const unsigned char *cmd,
282                      int data_direction, void *buffer, unsigned bufflen,
283                      struct scsi_sense_hdr *sshdr, int timeout, int retries)
284 {
285         char *sense = NULL;
286         int result;
287         
288         if (sshdr) {
289                 sense = kzalloc(SCSI_SENSE_BUFFERSIZE, GFP_NOIO);
290                 if (!sense)
291                         return DRIVER_ERROR << 24;
292         }
293         result = scsi_execute(sdev, cmd, data_direction, buffer, bufflen,
294                               sense, timeout, retries, 0);
295         if (sshdr)
296                 scsi_normalize_sense(sense, SCSI_SENSE_BUFFERSIZE, sshdr);
297
298         kfree(sense);
299         return result;
300 }
301 EXPORT_SYMBOL(scsi_execute_req);
302
303 struct scsi_io_context {
304         void *data;
305         void (*done)(void *data, char *sense, int result, int resid);
306         char sense[SCSI_SENSE_BUFFERSIZE];
307 };
308
309 static kmem_cache_t *scsi_io_context_cache;
310
311 static void scsi_end_async(struct request *req, int uptodate)
312 {
313         struct scsi_io_context *sioc = req->end_io_data;
314
315         if (sioc->done)
316                 sioc->done(sioc->data, sioc->sense, req->errors, req->data_len);
317
318         kmem_cache_free(scsi_io_context_cache, sioc);
319         __blk_put_request(req->q, req);
320 }
321
322 static int scsi_merge_bio(struct request *rq, struct bio *bio)
323 {
324         struct request_queue *q = rq->q;
325
326         bio->bi_flags &= ~(1 << BIO_SEG_VALID);
327         if (rq_data_dir(rq) == WRITE)
328                 bio->bi_rw |= (1 << BIO_RW);
329         blk_queue_bounce(q, &bio);
330
331         if (!rq->bio)
332                 blk_rq_bio_prep(q, rq, bio);
333         else if (!q->back_merge_fn(q, rq, bio))
334                 return -EINVAL;
335         else {
336                 rq->biotail->bi_next = bio;
337                 rq->biotail = bio;
338                 rq->hard_nr_sectors += bio_sectors(bio);
339                 rq->nr_sectors = rq->hard_nr_sectors;
340         }
341
342         return 0;
343 }
344
345 static int scsi_bi_endio(struct bio *bio, unsigned int bytes_done, int error)
346 {
347         if (bio->bi_size)
348                 return 1;
349
350         bio_put(bio);
351         return 0;
352 }
353
354 /**
355  * scsi_req_map_sg - map a scatterlist into a request
356  * @rq:         request to fill
357  * @sg:         scatterlist
358  * @nsegs:      number of elements
359  * @bufflen:    len of buffer
360  * @gfp:        memory allocation flags
361  *
362  * scsi_req_map_sg maps a scatterlist into a request so that the
363  * request can be sent to the block layer. We do not trust the scatterlist
364  * sent to use, as some ULDs use that struct to only organize the pages.
365  */
366 static int scsi_req_map_sg(struct request *rq, struct scatterlist *sgl,
367                            int nsegs, unsigned bufflen, gfp_t gfp)
368 {
369         struct request_queue *q = rq->q;
370         int nr_pages = (bufflen + sgl[0].offset + PAGE_SIZE - 1) >> PAGE_SHIFT;
371         unsigned int data_len = 0, len, bytes, off;
372         struct page *page;
373         struct bio *bio = NULL;
374         int i, err, nr_vecs = 0;
375
376         for (i = 0; i < nsegs; i++) {
377                 page = sgl[i].page;
378                 off = sgl[i].offset;
379                 len = sgl[i].length;
380                 data_len += len;
381
382                 while (len > 0) {
383                         bytes = min_t(unsigned int, len, PAGE_SIZE - off);
384
385                         if (!bio) {
386                                 nr_vecs = min_t(int, BIO_MAX_PAGES, nr_pages);
387                                 nr_pages -= nr_vecs;
388
389                                 bio = bio_alloc(gfp, nr_vecs);
390                                 if (!bio) {
391                                         err = -ENOMEM;
392                                         goto free_bios;
393                                 }
394                                 bio->bi_end_io = scsi_bi_endio;
395                         }
396
397                         if (bio_add_pc_page(q, bio, page, bytes, off) !=
398                             bytes) {
399                                 bio_put(bio);
400                                 err = -EINVAL;
401                                 goto free_bios;
402                         }
403
404                         if (bio->bi_vcnt >= nr_vecs) {
405                                 err = scsi_merge_bio(rq, bio);
406                                 if (err) {
407                                         bio_endio(bio, bio->bi_size, 0);
408                                         goto free_bios;
409                                 }
410                                 bio = NULL;
411                         }
412
413                         page++;
414                         len -= bytes;
415                         off = 0;
416                 }
417         }
418
419         rq->buffer = rq->data = NULL;
420         rq->data_len = data_len;
421         return 0;
422
423 free_bios:
424         while ((bio = rq->bio) != NULL) {
425                 rq->bio = bio->bi_next;
426                 /*
427                  * call endio instead of bio_put incase it was bounced
428                  */
429                 bio_endio(bio, bio->bi_size, 0);
430         }
431
432         return err;
433 }
434
435 /**
436  * scsi_execute_async - insert request
437  * @sdev:       scsi device
438  * @cmd:        scsi command
439  * @cmd_len:    length of scsi cdb
440  * @data_direction: data direction
441  * @buffer:     data buffer (this can be a kernel buffer or scatterlist)
442  * @bufflen:    len of buffer
443  * @use_sg:     if buffer is a scatterlist this is the number of elements
444  * @timeout:    request timeout in seconds
445  * @retries:    number of times to retry request
446  * @flags:      or into request flags
447  **/
448 int scsi_execute_async(struct scsi_device *sdev, const unsigned char *cmd,
449                        int cmd_len, int data_direction, void *buffer, unsigned bufflen,
450                        int use_sg, int timeout, int retries, void *privdata,
451                        void (*done)(void *, char *, int, int), gfp_t gfp)
452 {
453         struct request *req;
454         struct scsi_io_context *sioc;
455         int err = 0;
456         int write = (data_direction == DMA_TO_DEVICE);
457
458         sioc = kmem_cache_alloc(scsi_io_context_cache, gfp);
459         if (!sioc)
460                 return DRIVER_ERROR << 24;
461         memset(sioc, 0, sizeof(*sioc));
462
463         req = blk_get_request(sdev->request_queue, write, gfp);
464         if (!req)
465                 goto free_sense;
466         req->flags |= REQ_BLOCK_PC | REQ_QUIET;
467
468         if (use_sg)
469                 err = scsi_req_map_sg(req, buffer, use_sg, bufflen, gfp);
470         else if (bufflen)
471                 err = blk_rq_map_kern(req->q, req, buffer, bufflen, gfp);
472
473         if (err)
474                 goto free_req;
475
476         req->cmd_len = cmd_len;
477         memcpy(req->cmd, cmd, req->cmd_len);
478         req->sense = sioc->sense;
479         req->sense_len = 0;
480         req->timeout = timeout;
481         req->retries = retries;
482         req->end_io_data = sioc;
483
484         sioc->data = privdata;
485         sioc->done = done;
486
487         blk_execute_rq_nowait(req->q, NULL, req, 1, scsi_end_async);
488         return 0;
489
490 free_req:
491         blk_put_request(req);
492 free_sense:
493         kfree(sioc);
494         return DRIVER_ERROR << 24;
495 }
496 EXPORT_SYMBOL_GPL(scsi_execute_async);
497
498 /*
499  * Function:    scsi_init_cmd_errh()
500  *
501  * Purpose:     Initialize cmd fields related to error handling.
502  *
503  * Arguments:   cmd     - command that is ready to be queued.
504  *
505  * Returns:     Nothing
506  *
507  * Notes:       This function has the job of initializing a number of
508  *              fields related to error handling.   Typically this will
509  *              be called once for each command, as required.
510  */
511 static int scsi_init_cmd_errh(struct scsi_cmnd *cmd)
512 {
513         cmd->serial_number = 0;
514
515         memset(cmd->sense_buffer, 0, sizeof cmd->sense_buffer);
516
517         if (cmd->cmd_len == 0)
518                 cmd->cmd_len = COMMAND_SIZE(cmd->cmnd[0]);
519
520         /*
521          * We need saved copies of a number of fields - this is because
522          * error handling may need to overwrite these with different values
523          * to run different commands, and once error handling is complete,
524          * we will need to restore these values prior to running the actual
525          * command.
526          */
527         cmd->old_use_sg = cmd->use_sg;
528         cmd->old_cmd_len = cmd->cmd_len;
529         cmd->sc_old_data_direction = cmd->sc_data_direction;
530         cmd->old_underflow = cmd->underflow;
531         memcpy(cmd->data_cmnd, cmd->cmnd, sizeof(cmd->cmnd));
532         cmd->buffer = cmd->request_buffer;
533         cmd->bufflen = cmd->request_bufflen;
534
535         return 1;
536 }
537
538 /*
539  * Function:   scsi_setup_cmd_retry()
540  *
541  * Purpose:    Restore the command state for a retry
542  *
543  * Arguments:  cmd      - command to be restored
544  *
545  * Returns:    Nothing
546  *
547  * Notes:      Immediately prior to retrying a command, we need
548  *             to restore certain fields that we saved above.
549  */
550 void scsi_setup_cmd_retry(struct scsi_cmnd *cmd)
551 {
552         memcpy(cmd->cmnd, cmd->data_cmnd, sizeof(cmd->data_cmnd));
553         cmd->request_buffer = cmd->buffer;
554         cmd->request_bufflen = cmd->bufflen;
555         cmd->use_sg = cmd->old_use_sg;
556         cmd->cmd_len = cmd->old_cmd_len;
557         cmd->sc_data_direction = cmd->sc_old_data_direction;
558         cmd->underflow = cmd->old_underflow;
559 }
560
561 void scsi_device_unbusy(struct scsi_device *sdev)
562 {
563         struct Scsi_Host *shost = sdev->host;
564         unsigned long flags;
565
566         spin_lock_irqsave(shost->host_lock, flags);
567         shost->host_busy--;
568         if (unlikely(scsi_host_in_recovery(shost) &&
569                      shost->host_failed))
570                 scsi_eh_wakeup(shost);
571         spin_unlock(shost->host_lock);
572         spin_lock(sdev->request_queue->queue_lock);
573         sdev->device_busy--;
574         spin_unlock_irqrestore(sdev->request_queue->queue_lock, flags);
575 }
576
577 /*
578  * Called for single_lun devices on IO completion. Clear starget_sdev_user,
579  * and call blk_run_queue for all the scsi_devices on the target -
580  * including current_sdev first.
581  *
582  * Called with *no* scsi locks held.
583  */
584 static void scsi_single_lun_run(struct scsi_device *current_sdev)
585 {
586         struct Scsi_Host *shost = current_sdev->host;
587         struct scsi_device *sdev, *tmp;
588         struct scsi_target *starget = scsi_target(current_sdev);
589         unsigned long flags;
590
591         spin_lock_irqsave(shost->host_lock, flags);
592         starget->starget_sdev_user = NULL;
593         spin_unlock_irqrestore(shost->host_lock, flags);
594
595         /*
596          * Call blk_run_queue for all LUNs on the target, starting with
597          * current_sdev. We race with others (to set starget_sdev_user),
598          * but in most cases, we will be first. Ideally, each LU on the
599          * target would get some limited time or requests on the target.
600          */
601         blk_run_queue(current_sdev->request_queue);
602
603         spin_lock_irqsave(shost->host_lock, flags);
604         if (starget->starget_sdev_user)
605                 goto out;
606         list_for_each_entry_safe(sdev, tmp, &starget->devices,
607                         same_target_siblings) {
608                 if (sdev == current_sdev)
609                         continue;
610                 if (scsi_device_get(sdev))
611                         continue;
612
613                 spin_unlock_irqrestore(shost->host_lock, flags);
614                 blk_run_queue(sdev->request_queue);
615                 spin_lock_irqsave(shost->host_lock, flags);
616         
617                 scsi_device_put(sdev);
618         }
619  out:
620         spin_unlock_irqrestore(shost->host_lock, flags);
621 }
622
623 /*
624  * Function:    scsi_run_queue()
625  *
626  * Purpose:     Select a proper request queue to serve next
627  *
628  * Arguments:   q       - last request's queue
629  *
630  * Returns:     Nothing
631  *
632  * Notes:       The previous command was completely finished, start
633  *              a new one if possible.
634  */
635 static void scsi_run_queue(struct request_queue *q)
636 {
637         struct scsi_device *sdev = q->queuedata;
638         struct Scsi_Host *shost = sdev->host;
639         unsigned long flags;
640
641         if (sdev->single_lun)
642                 scsi_single_lun_run(sdev);
643
644         spin_lock_irqsave(shost->host_lock, flags);
645         while (!list_empty(&shost->starved_list) &&
646                !shost->host_blocked && !shost->host_self_blocked &&
647                 !((shost->can_queue > 0) &&
648                   (shost->host_busy >= shost->can_queue))) {
649                 /*
650                  * As long as shost is accepting commands and we have
651                  * starved queues, call blk_run_queue. scsi_request_fn
652                  * drops the queue_lock and can add us back to the
653                  * starved_list.
654                  *
655                  * host_lock protects the starved_list and starved_entry.
656                  * scsi_request_fn must get the host_lock before checking
657                  * or modifying starved_list or starved_entry.
658                  */
659                 sdev = list_entry(shost->starved_list.next,
660                                           struct scsi_device, starved_entry);
661                 list_del_init(&sdev->starved_entry);
662                 spin_unlock_irqrestore(shost->host_lock, flags);
663
664                 blk_run_queue(sdev->request_queue);
665
666                 spin_lock_irqsave(shost->host_lock, flags);
667                 if (unlikely(!list_empty(&sdev->starved_entry)))
668                         /*
669                          * sdev lost a race, and was put back on the
670                          * starved list. This is unlikely but without this
671                          * in theory we could loop forever.
672                          */
673                         break;
674         }
675         spin_unlock_irqrestore(shost->host_lock, flags);
676
677         blk_run_queue(q);
678 }
679
680 /*
681  * Function:    scsi_requeue_command()
682  *
683  * Purpose:     Handle post-processing of completed commands.
684  *
685  * Arguments:   q       - queue to operate on
686  *              cmd     - command that may need to be requeued.
687  *
688  * Returns:     Nothing
689  *
690  * Notes:       After command completion, there may be blocks left
691  *              over which weren't finished by the previous command
692  *              this can be for a number of reasons - the main one is
693  *              I/O errors in the middle of the request, in which case
694  *              we need to request the blocks that come after the bad
695  *              sector.
696  * Notes:       Upon return, cmd is a stale pointer.
697  */
698 static void scsi_requeue_command(struct request_queue *q, struct scsi_cmnd *cmd)
699 {
700         struct request *req = cmd->request;
701         unsigned long flags;
702
703         scsi_unprep_request(req);
704         spin_lock_irqsave(q->queue_lock, flags);
705         blk_requeue_request(q, req);
706         spin_unlock_irqrestore(q->queue_lock, flags);
707
708         scsi_run_queue(q);
709 }
710
711 void scsi_next_command(struct scsi_cmnd *cmd)
712 {
713         struct scsi_device *sdev = cmd->device;
714         struct request_queue *q = sdev->request_queue;
715
716         /* need to hold a reference on the device before we let go of the cmd */
717         get_device(&sdev->sdev_gendev);
718
719         scsi_put_command(cmd);
720         scsi_run_queue(q);
721
722         /* ok to remove device now */
723         put_device(&sdev->sdev_gendev);
724 }
725
726 void scsi_run_host_queues(struct Scsi_Host *shost)
727 {
728         struct scsi_device *sdev;
729
730         shost_for_each_device(sdev, shost)
731                 scsi_run_queue(sdev->request_queue);
732 }
733
734 /*
735  * Function:    scsi_end_request()
736  *
737  * Purpose:     Post-processing of completed commands (usually invoked at end
738  *              of upper level post-processing and scsi_io_completion).
739  *
740  * Arguments:   cmd      - command that is complete.
741  *              uptodate - 1 if I/O indicates success, <= 0 for I/O error.
742  *              bytes    - number of bytes of completed I/O
743  *              requeue  - indicates whether we should requeue leftovers.
744  *
745  * Lock status: Assumed that lock is not held upon entry.
746  *
747  * Returns:     cmd if requeue required, NULL otherwise.
748  *
749  * Notes:       This is called for block device requests in order to
750  *              mark some number of sectors as complete.
751  * 
752  *              We are guaranteeing that the request queue will be goosed
753  *              at some point during this call.
754  * Notes:       If cmd was requeued, upon return it will be a stale pointer.
755  */
756 static struct scsi_cmnd *scsi_end_request(struct scsi_cmnd *cmd, int uptodate,
757                                           int bytes, int requeue)
758 {
759         request_queue_t *q = cmd->device->request_queue;
760         struct request *req = cmd->request;
761         unsigned long flags;
762
763         /*
764          * If there are blocks left over at the end, set up the command
765          * to queue the remainder of them.
766          */
767         if (end_that_request_chunk(req, uptodate, bytes)) {
768                 int leftover = (req->hard_nr_sectors << 9);
769
770                 if (blk_pc_request(req))
771                         leftover = req->data_len;
772
773                 /* kill remainder if no retrys */
774                 if (!uptodate && blk_noretry_request(req))
775                         end_that_request_chunk(req, 0, leftover);
776                 else {
777                         if (requeue) {
778                                 /*
779                                  * Bleah.  Leftovers again.  Stick the
780                                  * leftovers in the front of the
781                                  * queue, and goose the queue again.
782                                  */
783                                 scsi_requeue_command(q, cmd);
784                                 cmd = NULL;
785                         }
786                         return cmd;
787                 }
788         }
789
790         add_disk_randomness(req->rq_disk);
791
792         spin_lock_irqsave(q->queue_lock, flags);
793         if (blk_rq_tagged(req))
794                 blk_queue_end_tag(q, req);
795         end_that_request_last(req, uptodate);
796         spin_unlock_irqrestore(q->queue_lock, flags);
797
798         /*
799          * This will goose the queue request function at the end, so we don't
800          * need to worry about launching another command.
801          */
802         scsi_next_command(cmd);
803         return NULL;
804 }
805
806 static struct scatterlist *scsi_alloc_sgtable(struct scsi_cmnd *cmd, gfp_t gfp_mask)
807 {
808         struct scsi_host_sg_pool *sgp;
809         struct scatterlist *sgl;
810
811         BUG_ON(!cmd->use_sg);
812
813         switch (cmd->use_sg) {
814         case 1 ... 8:
815                 cmd->sglist_len = 0;
816                 break;
817         case 9 ... 16:
818                 cmd->sglist_len = 1;
819                 break;
820         case 17 ... 32:
821                 cmd->sglist_len = 2;
822                 break;
823 #if (SCSI_MAX_PHYS_SEGMENTS > 32)
824         case 33 ... 64:
825                 cmd->sglist_len = 3;
826                 break;
827 #if (SCSI_MAX_PHYS_SEGMENTS > 64)
828         case 65 ... 128:
829                 cmd->sglist_len = 4;
830                 break;
831 #if (SCSI_MAX_PHYS_SEGMENTS  > 128)
832         case 129 ... 256:
833                 cmd->sglist_len = 5;
834                 break;
835 #endif
836 #endif
837 #endif
838         default:
839                 return NULL;
840         }
841
842         sgp = scsi_sg_pools + cmd->sglist_len;
843         sgl = mempool_alloc(sgp->pool, gfp_mask);
844         return sgl;
845 }
846
847 static void scsi_free_sgtable(struct scatterlist *sgl, int index)
848 {
849         struct scsi_host_sg_pool *sgp;
850
851         BUG_ON(index >= SG_MEMPOOL_NR);
852
853         sgp = scsi_sg_pools + index;
854         mempool_free(sgl, sgp->pool);
855 }
856
857 /*
858  * Function:    scsi_release_buffers()
859  *
860  * Purpose:     Completion processing for block device I/O requests.
861  *
862  * Arguments:   cmd     - command that we are bailing.
863  *
864  * Lock status: Assumed that no lock is held upon entry.
865  *
866  * Returns:     Nothing
867  *
868  * Notes:       In the event that an upper level driver rejects a
869  *              command, we must release resources allocated during
870  *              the __init_io() function.  Primarily this would involve
871  *              the scatter-gather table, and potentially any bounce
872  *              buffers.
873  */
874 static void scsi_release_buffers(struct scsi_cmnd *cmd)
875 {
876         struct request *req = cmd->request;
877
878         /*
879          * Free up any indirection buffers we allocated for DMA purposes. 
880          */
881         if (cmd->use_sg)
882                 scsi_free_sgtable(cmd->request_buffer, cmd->sglist_len);
883         else if (cmd->request_buffer != req->buffer)
884                 kfree(cmd->request_buffer);
885
886         /*
887          * Zero these out.  They now point to freed memory, and it is
888          * dangerous to hang onto the pointers.
889          */
890         cmd->buffer  = NULL;
891         cmd->bufflen = 0;
892         cmd->request_buffer = NULL;
893         cmd->request_bufflen = 0;
894 }
895
896 /*
897  * Function:    scsi_io_completion()
898  *
899  * Purpose:     Completion processing for block device I/O requests.
900  *
901  * Arguments:   cmd   - command that is finished.
902  *
903  * Lock status: Assumed that no lock is held upon entry.
904  *
905  * Returns:     Nothing
906  *
907  * Notes:       This function is matched in terms of capabilities to
908  *              the function that created the scatter-gather list.
909  *              In other words, if there are no bounce buffers
910  *              (the normal case for most drivers), we don't need
911  *              the logic to deal with cleaning up afterwards.
912  *
913  *              We must do one of several things here:
914  *
915  *              a) Call scsi_end_request.  This will finish off the
916  *                 specified number of sectors.  If we are done, the
917  *                 command block will be released, and the queue
918  *                 function will be goosed.  If we are not done, then
919  *                 scsi_end_request will directly goose the queue.
920  *
921  *              b) We can just use scsi_requeue_command() here.  This would
922  *                 be used if we just wanted to retry, for example.
923  */
924 void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes,
925                         unsigned int block_bytes)
926 {
927         int result = cmd->result;
928         int this_count = cmd->bufflen;
929         request_queue_t *q = cmd->device->request_queue;
930         struct request *req = cmd->request;
931         int clear_errors = 1;
932         struct scsi_sense_hdr sshdr;
933         int sense_valid = 0;
934         int sense_deferred = 0;
935
936         /*
937          * Free up any indirection buffers we allocated for DMA purposes. 
938          * For the case of a READ, we need to copy the data out of the
939          * bounce buffer and into the real buffer.
940          */
941         if (cmd->use_sg)
942                 scsi_free_sgtable(cmd->buffer, cmd->sglist_len);
943         else if (cmd->buffer != req->buffer) {
944                 if (rq_data_dir(req) == READ) {
945                         unsigned long flags;
946                         char *to = bio_kmap_irq(req->bio, &flags);
947                         memcpy(to, cmd->buffer, cmd->bufflen);
948                         bio_kunmap_irq(to, &flags);
949                 }
950                 kfree(cmd->buffer);
951         }
952
953         if (result) {
954                 sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
955                 if (sense_valid)
956                         sense_deferred = scsi_sense_is_deferred(&sshdr);
957         }
958         if (blk_pc_request(req)) { /* SG_IO ioctl from block level */
959                 req->errors = result;
960                 if (result) {
961                         clear_errors = 0;
962                         if (sense_valid && req->sense) {
963                                 /*
964                                  * SG_IO wants current and deferred errors
965                                  */
966                                 int len = 8 + cmd->sense_buffer[7];
967
968                                 if (len > SCSI_SENSE_BUFFERSIZE)
969                                         len = SCSI_SENSE_BUFFERSIZE;
970                                 memcpy(req->sense, cmd->sense_buffer,  len);
971                                 req->sense_len = len;
972                         }
973                 } else
974                         req->data_len = cmd->resid;
975         }
976
977         /*
978          * Zero these out.  They now point to freed memory, and it is
979          * dangerous to hang onto the pointers.
980          */
981         cmd->buffer  = NULL;
982         cmd->bufflen = 0;
983         cmd->request_buffer = NULL;
984         cmd->request_bufflen = 0;
985
986         /*
987          * Next deal with any sectors which we were able to correctly
988          * handle.
989          */
990         if (good_bytes >= 0) {
991                 SCSI_LOG_HLCOMPLETE(1, printk("%ld sectors total, %d bytes done.\n",
992                                               req->nr_sectors, good_bytes));
993                 SCSI_LOG_HLCOMPLETE(1, printk("use_sg is %d\n", cmd->use_sg));
994
995                 if (clear_errors)
996                         req->errors = 0;
997                 /*
998                  * If multiple sectors are requested in one buffer, then
999                  * they will have been finished off by the first command.
1000                  * If not, then we have a multi-buffer command.
1001                  *
1002                  * If block_bytes != 0, it means we had a medium error
1003                  * of some sort, and that we want to mark some number of
1004                  * sectors as not uptodate.  Thus we want to inhibit
1005                  * requeueing right here - we will requeue down below
1006                  * when we handle the bad sectors.
1007                  */
1008
1009                 /*
1010                  * If the command completed without error, then either
1011                  * finish off the rest of the command, or start a new one.
1012                  */
1013                 if (scsi_end_request(cmd, 1, good_bytes, result == 0) == NULL)
1014                         return;
1015         }
1016         /*
1017          * Now, if we were good little boys and girls, Santa left us a request
1018          * sense buffer.  We can extract information from this, so we
1019          * can choose a block to remap, etc.
1020          */
1021         if (sense_valid && !sense_deferred) {
1022                 switch (sshdr.sense_key) {
1023                 case UNIT_ATTENTION:
1024                         if (cmd->device->removable) {
1025                                 /* detected disc change.  set a bit 
1026                                  * and quietly refuse further access.
1027                                  */
1028                                 cmd->device->changed = 1;
1029                                 scsi_end_request(cmd, 0,
1030                                                 this_count, 1);
1031                                 return;
1032                         } else {
1033                                 /*
1034                                 * Must have been a power glitch, or a
1035                                 * bus reset.  Could not have been a
1036                                 * media change, so we just retry the
1037                                 * request and see what happens.  
1038                                 */
1039                                 scsi_requeue_command(q, cmd);
1040                                 return;
1041                         }
1042                         break;
1043                 case ILLEGAL_REQUEST:
1044                         /*
1045                         * If we had an ILLEGAL REQUEST returned, then we may
1046                         * have performed an unsupported command.  The only
1047                         * thing this should be would be a ten byte read where
1048                         * only a six byte read was supported.  Also, on a
1049                         * system where READ CAPACITY failed, we may have read
1050                         * past the end of the disk.
1051                         */
1052                         if ((cmd->device->use_10_for_rw &&
1053                             sshdr.asc == 0x20 && sshdr.ascq == 0x00) &&
1054                             (cmd->cmnd[0] == READ_10 ||
1055                              cmd->cmnd[0] == WRITE_10)) {
1056                                 cmd->device->use_10_for_rw = 0;
1057                                 /*
1058                                  * This will cause a retry with a 6-byte
1059                                  * command.
1060                                  */
1061                                 scsi_requeue_command(q, cmd);
1062                                 result = 0;
1063                         } else {
1064                                 scsi_end_request(cmd, 0, this_count, 1);
1065                                 return;
1066                         }
1067                         break;
1068                 case NOT_READY:
1069                         /*
1070                          * If the device is in the process of becoming
1071                          * ready, or has a temporary blockage, retry.
1072                          */
1073                         if (sshdr.asc == 0x04) {
1074                                 switch (sshdr.ascq) {
1075                                 case 0x01: /* becoming ready */
1076                                 case 0x04: /* format in progress */
1077                                 case 0x05: /* rebuild in progress */
1078                                 case 0x06: /* recalculation in progress */
1079                                 case 0x07: /* operation in progress */
1080                                 case 0x08: /* Long write in progress */
1081                                 case 0x09: /* self test in progress */
1082                                         scsi_requeue_command(q, cmd);
1083                                         return;
1084                                 default:
1085                                         break;
1086                                 }
1087                         }
1088                         if (!(req->flags & REQ_QUIET)) {
1089                                 scmd_printk(KERN_INFO, cmd,
1090                                            "Device not ready: ");
1091                                 scsi_print_sense_hdr("", &sshdr);
1092                         }
1093                         scsi_end_request(cmd, 0, this_count, 1);
1094                         return;
1095                 case VOLUME_OVERFLOW:
1096                         if (!(req->flags & REQ_QUIET)) {
1097                                 scmd_printk(KERN_INFO, cmd,
1098                                            "Volume overflow, CDB: ");
1099                                 __scsi_print_command(cmd->data_cmnd);
1100                                 scsi_print_sense("", cmd);
1101                         }
1102                         scsi_end_request(cmd, 0, block_bytes, 1);
1103                         return;
1104                 default:
1105                         break;
1106                 }
1107         }                       /* driver byte != 0 */
1108         if (host_byte(result) == DID_RESET) {
1109                 /*
1110                  * Third party bus reset or reset for error
1111                  * recovery reasons.  Just retry the request
1112                  * and see what happens.  
1113                  */
1114                 scsi_requeue_command(q, cmd);
1115                 return;
1116         }
1117         if (result) {
1118                 if (!(req->flags & REQ_QUIET)) {
1119                         scmd_printk(KERN_INFO, cmd,
1120                                    "SCSI error: return code = 0x%x\n", result);
1121
1122                         if (driver_byte(result) & DRIVER_SENSE)
1123                                 scsi_print_sense("", cmd);
1124                 }
1125                 /*
1126                  * Mark a single buffer as not uptodate.  Queue the remainder.
1127                  * We sometimes get this cruft in the event that a medium error
1128                  * isn't properly reported.
1129                  */
1130                 block_bytes = req->hard_cur_sectors << 9;
1131                 if (!block_bytes)
1132                         block_bytes = req->data_len;
1133                 scsi_end_request(cmd, 0, block_bytes, 1);
1134         }
1135 }
1136 EXPORT_SYMBOL(scsi_io_completion);
1137
1138 /*
1139  * Function:    scsi_init_io()
1140  *
1141  * Purpose:     SCSI I/O initialize function.
1142  *
1143  * Arguments:   cmd   - Command descriptor we wish to initialize
1144  *
1145  * Returns:     0 on success
1146  *              BLKPREP_DEFER if the failure is retryable
1147  *              BLKPREP_KILL if the failure is fatal
1148  */
1149 static int scsi_init_io(struct scsi_cmnd *cmd)
1150 {
1151         struct request     *req = cmd->request;
1152         struct scatterlist *sgpnt;
1153         int                count;
1154
1155         /*
1156          * if this is a rq->data based REQ_BLOCK_PC, setup for a non-sg xfer
1157          */
1158         if ((req->flags & REQ_BLOCK_PC) && !req->bio) {
1159                 cmd->request_bufflen = req->data_len;
1160                 cmd->request_buffer = req->data;
1161                 req->buffer = req->data;
1162                 cmd->use_sg = 0;
1163                 return 0;
1164         }
1165
1166         /*
1167          * we used to not use scatter-gather for single segment request,
1168          * but now we do (it makes highmem I/O easier to support without
1169          * kmapping pages)
1170          */
1171         cmd->use_sg = req->nr_phys_segments;
1172
1173         /*
1174          * if sg table allocation fails, requeue request later.
1175          */
1176         sgpnt = scsi_alloc_sgtable(cmd, GFP_ATOMIC);
1177         if (unlikely(!sgpnt)) {
1178                 scsi_unprep_request(req);
1179                 return BLKPREP_DEFER;
1180         }
1181
1182         cmd->request_buffer = (char *) sgpnt;
1183         cmd->request_bufflen = req->nr_sectors << 9;
1184         if (blk_pc_request(req))
1185                 cmd->request_bufflen = req->data_len;
1186         req->buffer = NULL;
1187
1188         /* 
1189          * Next, walk the list, and fill in the addresses and sizes of
1190          * each segment.
1191          */
1192         count = blk_rq_map_sg(req->q, req, cmd->request_buffer);
1193
1194         /*
1195          * mapped well, send it off
1196          */
1197         if (likely(count <= cmd->use_sg)) {
1198                 cmd->use_sg = count;
1199                 return 0;
1200         }
1201
1202         printk(KERN_ERR "Incorrect number of segments after building list\n");
1203         printk(KERN_ERR "counted %d, received %d\n", count, cmd->use_sg);
1204         printk(KERN_ERR "req nr_sec %lu, cur_nr_sec %u\n", req->nr_sectors,
1205                         req->current_nr_sectors);
1206
1207         /* release the command and kill it */
1208         scsi_release_buffers(cmd);
1209         scsi_put_command(cmd);
1210         return BLKPREP_KILL;
1211 }
1212
1213 static int scsi_issue_flush_fn(request_queue_t *q, struct gendisk *disk,
1214                                sector_t *error_sector)
1215 {
1216         struct scsi_device *sdev = q->queuedata;
1217         struct scsi_driver *drv;
1218
1219         if (sdev->sdev_state != SDEV_RUNNING)
1220                 return -ENXIO;
1221
1222         drv = *(struct scsi_driver **) disk->private_data;
1223         if (drv->issue_flush)
1224                 return drv->issue_flush(&sdev->sdev_gendev, error_sector);
1225
1226         return -EOPNOTSUPP;
1227 }
1228
1229 static void scsi_blk_pc_done(struct scsi_cmnd *cmd)
1230 {
1231         BUG_ON(!blk_pc_request(cmd->request));
1232         /*
1233          * This will complete the whole command with uptodate=1 so
1234          * as far as the block layer is concerned the command completed
1235          * successfully. Since this is a REQ_BLOCK_PC command the
1236          * caller should check the request's errors value
1237          */
1238         scsi_io_completion(cmd, cmd->bufflen, 0);
1239 }
1240
1241 static void scsi_setup_blk_pc_cmnd(struct scsi_cmnd *cmd)
1242 {
1243         struct request *req = cmd->request;
1244
1245         BUG_ON(sizeof(req->cmd) > sizeof(cmd->cmnd));
1246         memcpy(cmd->cmnd, req->cmd, sizeof(cmd->cmnd));
1247         cmd->cmd_len = req->cmd_len;
1248         if (!req->data_len)
1249                 cmd->sc_data_direction = DMA_NONE;
1250         else if (rq_data_dir(req) == WRITE)
1251                 cmd->sc_data_direction = DMA_TO_DEVICE;
1252         else
1253                 cmd->sc_data_direction = DMA_FROM_DEVICE;
1254         
1255         cmd->transfersize = req->data_len;
1256         cmd->allowed = req->retries;
1257         cmd->timeout_per_command = req->timeout;
1258         cmd->done = scsi_blk_pc_done;
1259 }
1260
1261 static int scsi_prep_fn(struct request_queue *q, struct request *req)
1262 {
1263         struct scsi_device *sdev = q->queuedata;
1264         struct scsi_cmnd *cmd;
1265         int specials_only = 0;
1266
1267         /*
1268          * Just check to see if the device is online.  If it isn't, we
1269          * refuse to process any commands.  The device must be brought
1270          * online before trying any recovery commands
1271          */
1272         if (unlikely(!scsi_device_online(sdev))) {
1273                 sdev_printk(KERN_ERR, sdev,
1274                             "rejecting I/O to offline device\n");
1275                 goto kill;
1276         }
1277         if (unlikely(sdev->sdev_state != SDEV_RUNNING)) {
1278                 /* OK, we're not in a running state don't prep
1279                  * user commands */
1280                 if (sdev->sdev_state == SDEV_DEL) {
1281                         /* Device is fully deleted, no commands
1282                          * at all allowed down */
1283                         sdev_printk(KERN_ERR, sdev,
1284                                     "rejecting I/O to dead device\n");
1285                         goto kill;
1286                 }
1287                 /* OK, we only allow special commands (i.e. not
1288                  * user initiated ones */
1289                 specials_only = sdev->sdev_state;
1290         }
1291
1292         /*
1293          * Find the actual device driver associated with this command.
1294          * The SPECIAL requests are things like character device or
1295          * ioctls, which did not originate from ll_rw_blk.  Note that
1296          * the special field is also used to indicate the cmd for
1297          * the remainder of a partially fulfilled request that can 
1298          * come up when there is a medium error.  We have to treat
1299          * these two cases differently.  We differentiate by looking
1300          * at request->cmd, as this tells us the real story.
1301          */
1302         if (req->flags & REQ_SPECIAL && req->special) {
1303                 struct scsi_request *sreq = req->special;
1304
1305                 if (sreq->sr_magic == SCSI_REQ_MAGIC) {
1306                         cmd = scsi_get_command(sreq->sr_device, GFP_ATOMIC);
1307                         if (unlikely(!cmd))
1308                                 goto defer;
1309                         scsi_init_cmd_from_req(cmd, sreq);
1310                 } else
1311                         cmd = req->special;
1312         } else if (req->flags & (REQ_CMD | REQ_BLOCK_PC)) {
1313
1314                 if(unlikely(specials_only) && !(req->flags & REQ_SPECIAL)) {
1315                         if(specials_only == SDEV_QUIESCE ||
1316                                         specials_only == SDEV_BLOCK)
1317                                 goto defer;
1318                         
1319                         sdev_printk(KERN_ERR, sdev,
1320                                     "rejecting I/O to device being removed\n");
1321                         goto kill;
1322                 }
1323                         
1324                         
1325                 /*
1326                  * Now try and find a command block that we can use.
1327                  */
1328                 if (!req->special) {
1329                         cmd = scsi_get_command(sdev, GFP_ATOMIC);
1330                         if (unlikely(!cmd))
1331                                 goto defer;
1332                 } else
1333                         cmd = req->special;
1334                 
1335                 /* pull a tag out of the request if we have one */
1336                 cmd->tag = req->tag;
1337         } else {
1338                 blk_dump_rq_flags(req, "SCSI bad req");
1339                 goto kill;
1340         }
1341         
1342         /* note the overloading of req->special.  When the tag
1343          * is active it always means cmd.  If the tag goes
1344          * back for re-queueing, it may be reset */
1345         req->special = cmd;
1346         cmd->request = req;
1347         
1348         /*
1349          * FIXME: drop the lock here because the functions below
1350          * expect to be called without the queue lock held.  Also,
1351          * previously, we dequeued the request before dropping the
1352          * lock.  We hope REQ_STARTED prevents anything untoward from
1353          * happening now.
1354          */
1355         if (req->flags & (REQ_CMD | REQ_BLOCK_PC)) {
1356                 int ret;
1357
1358                 /*
1359                  * This will do a couple of things:
1360                  *  1) Fill in the actual SCSI command.
1361                  *  2) Fill in any other upper-level specific fields
1362                  * (timeout).
1363                  *
1364                  * If this returns 0, it means that the request failed
1365                  * (reading past end of disk, reading offline device,
1366                  * etc).   This won't actually talk to the device, but
1367                  * some kinds of consistency checking may cause the     
1368                  * request to be rejected immediately.
1369                  */
1370
1371                 /* 
1372                  * This sets up the scatter-gather table (allocating if
1373                  * required).
1374                  */
1375                 ret = scsi_init_io(cmd);
1376                 switch(ret) {
1377                         /* For BLKPREP_KILL/DEFER the cmd was released */
1378                 case BLKPREP_KILL:
1379                         goto kill;
1380                 case BLKPREP_DEFER:
1381                         goto defer;
1382                 }
1383                 
1384                 /*
1385                  * Initialize the actual SCSI command for this request.
1386                  */
1387                 if (req->flags & REQ_BLOCK_PC) {
1388                         scsi_setup_blk_pc_cmnd(cmd);
1389                 } else if (req->rq_disk) {
1390                         struct scsi_driver *drv;
1391
1392                         drv = *(struct scsi_driver **)req->rq_disk->private_data;
1393                         if (unlikely(!drv->init_command(cmd))) {
1394                                 scsi_release_buffers(cmd);
1395                                 scsi_put_command(cmd);
1396                                 goto kill;
1397                         }
1398                 }
1399         }
1400
1401         /*
1402          * The request is now prepped, no need to come back here
1403          */
1404         req->flags |= REQ_DONTPREP;
1405         return BLKPREP_OK;
1406
1407  defer:
1408         /* If we defer, the elv_next_request() returns NULL, but the
1409          * queue must be restarted, so we plug here if no returning
1410          * command will automatically do that. */
1411         if (sdev->device_busy == 0)
1412                 blk_plug_device(q);
1413         return BLKPREP_DEFER;
1414  kill:
1415         req->errors = DID_NO_CONNECT << 16;
1416         return BLKPREP_KILL;
1417 }
1418
1419 /*
1420  * scsi_dev_queue_ready: if we can send requests to sdev, return 1 else
1421  * return 0.
1422  *
1423  * Called with the queue_lock held.
1424  */
1425 static inline int scsi_dev_queue_ready(struct request_queue *q,
1426                                   struct scsi_device *sdev)
1427 {
1428         if (sdev->device_busy >= sdev->queue_depth)
1429                 return 0;
1430         if (sdev->device_busy == 0 && sdev->device_blocked) {
1431                 /*
1432                  * unblock after device_blocked iterates to zero
1433                  */
1434                 if (--sdev->device_blocked == 0) {
1435                         SCSI_LOG_MLQUEUE(3,
1436                                    sdev_printk(KERN_INFO, sdev,
1437                                    "unblocking device at zero depth\n"));
1438                 } else {
1439                         blk_plug_device(q);
1440                         return 0;
1441                 }
1442         }
1443         if (sdev->device_blocked)
1444                 return 0;
1445
1446         return 1;
1447 }
1448
1449 /*
1450  * scsi_host_queue_ready: if we can send requests to shost, return 1 else
1451  * return 0. We must end up running the queue again whenever 0 is
1452  * returned, else IO can hang.
1453  *
1454  * Called with host_lock held.
1455  */
1456 static inline int scsi_host_queue_ready(struct request_queue *q,
1457                                    struct Scsi_Host *shost,
1458                                    struct scsi_device *sdev)
1459 {
1460         if (scsi_host_in_recovery(shost))
1461                 return 0;
1462         if (shost->host_busy == 0 && shost->host_blocked) {
1463                 /*
1464                  * unblock after host_blocked iterates to zero
1465                  */
1466                 if (--shost->host_blocked == 0) {
1467                         SCSI_LOG_MLQUEUE(3,
1468                                 printk("scsi%d unblocking host at zero depth\n",
1469                                         shost->host_no));
1470                 } else {
1471                         blk_plug_device(q);
1472                         return 0;
1473                 }
1474         }
1475         if ((shost->can_queue > 0 && shost->host_busy >= shost->can_queue) ||
1476             shost->host_blocked || shost->host_self_blocked) {
1477                 if (list_empty(&sdev->starved_entry))
1478                         list_add_tail(&sdev->starved_entry, &shost->starved_list);
1479                 return 0;
1480         }
1481
1482         /* We're OK to process the command, so we can't be starved */
1483         if (!list_empty(&sdev->starved_entry))
1484                 list_del_init(&sdev->starved_entry);
1485
1486         return 1;
1487 }
1488
1489 /*
1490  * Kill a request for a dead device
1491  */
1492 static void scsi_kill_request(struct request *req, request_queue_t *q)
1493 {
1494         struct scsi_cmnd *cmd = req->special;
1495         struct scsi_device *sdev = cmd->device;
1496         struct Scsi_Host *shost = sdev->host;
1497
1498         blkdev_dequeue_request(req);
1499
1500         if (unlikely(cmd == NULL)) {
1501                 printk(KERN_CRIT "impossible request in %s.\n",
1502                                  __FUNCTION__);
1503                 BUG();
1504         }
1505
1506         scsi_init_cmd_errh(cmd);
1507         cmd->result = DID_NO_CONNECT << 16;
1508         atomic_inc(&cmd->device->iorequest_cnt);
1509
1510         /*
1511          * SCSI request completion path will do scsi_device_unbusy(),
1512          * bump busy counts.  To bump the counters, we need to dance
1513          * with the locks as normal issue path does.
1514          */
1515         sdev->device_busy++;
1516         spin_unlock(sdev->request_queue->queue_lock);
1517         spin_lock(shost->host_lock);
1518         shost->host_busy++;
1519         spin_unlock(shost->host_lock);
1520         spin_lock(sdev->request_queue->queue_lock);
1521
1522         __scsi_done(cmd);
1523 }
1524
1525 static void scsi_softirq_done(struct request *rq)
1526 {
1527         struct scsi_cmnd *cmd = rq->completion_data;
1528         unsigned long wait_for = (cmd->allowed + 1) * cmd->timeout_per_command;
1529         int disposition;
1530
1531         INIT_LIST_HEAD(&cmd->eh_entry);
1532
1533         disposition = scsi_decide_disposition(cmd);
1534         if (disposition != SUCCESS &&
1535             time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) {
1536                 sdev_printk(KERN_ERR, cmd->device,
1537                             "timing out command, waited %lus\n",
1538                             wait_for/HZ);
1539                 disposition = SUCCESS;
1540         }
1541                         
1542         scsi_log_completion(cmd, disposition);
1543
1544         switch (disposition) {
1545                 case SUCCESS:
1546                         scsi_finish_command(cmd);
1547                         break;
1548                 case NEEDS_RETRY:
1549                         scsi_retry_command(cmd);
1550                         break;
1551                 case ADD_TO_MLQUEUE:
1552                         scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY);
1553                         break;
1554                 default:
1555                         if (!scsi_eh_scmd_add(cmd, 0))
1556                                 scsi_finish_command(cmd);
1557         }
1558 }
1559
1560 /*
1561  * Function:    scsi_request_fn()
1562  *
1563  * Purpose:     Main strategy routine for SCSI.
1564  *
1565  * Arguments:   q       - Pointer to actual queue.
1566  *
1567  * Returns:     Nothing
1568  *
1569  * Lock status: IO request lock assumed to be held when called.
1570  */
1571 static void scsi_request_fn(struct request_queue *q)
1572 {
1573         struct scsi_device *sdev = q->queuedata;
1574         struct Scsi_Host *shost;
1575         struct scsi_cmnd *cmd;
1576         struct request *req;
1577
1578         if (!sdev) {
1579                 printk("scsi: killing requests for dead queue\n");
1580                 while ((req = elv_next_request(q)) != NULL)
1581                         scsi_kill_request(req, q);
1582                 return;
1583         }
1584
1585         if(!get_device(&sdev->sdev_gendev))
1586                 /* We must be tearing the block queue down already */
1587                 return;
1588
1589         /*
1590          * To start with, we keep looping until the queue is empty, or until
1591          * the host is no longer able to accept any more requests.
1592          */
1593         shost = sdev->host;
1594         while (!blk_queue_plugged(q)) {
1595                 int rtn;
1596                 /*
1597                  * get next queueable request.  We do this early to make sure
1598                  * that the request is fully prepared even if we cannot 
1599                  * accept it.
1600                  */
1601                 req = elv_next_request(q);
1602                 if (!req || !scsi_dev_queue_ready(q, sdev))
1603                         break;
1604
1605                 if (unlikely(!scsi_device_online(sdev))) {
1606                         sdev_printk(KERN_ERR, sdev,
1607                                     "rejecting I/O to offline device\n");
1608                         scsi_kill_request(req, q);
1609                         continue;
1610                 }
1611
1612
1613                 /*
1614                  * Remove the request from the request list.
1615                  */
1616                 if (!(blk_queue_tagged(q) && !blk_queue_start_tag(q, req)))
1617                         blkdev_dequeue_request(req);
1618                 sdev->device_busy++;
1619
1620                 spin_unlock(q->queue_lock);
1621                 cmd = req->special;
1622                 if (unlikely(cmd == NULL)) {
1623                         printk(KERN_CRIT "impossible request in %s.\n"
1624                                          "please mail a stack trace to "
1625                                          "linux-scsi@vger.kernel.org",
1626                                          __FUNCTION__);
1627                         BUG();
1628                 }
1629                 spin_lock(shost->host_lock);
1630
1631                 if (!scsi_host_queue_ready(q, shost, sdev))
1632                         goto not_ready;
1633                 if (sdev->single_lun) {
1634                         if (scsi_target(sdev)->starget_sdev_user &&
1635                             scsi_target(sdev)->starget_sdev_user != sdev)
1636                                 goto not_ready;
1637                         scsi_target(sdev)->starget_sdev_user = sdev;
1638                 }
1639                 shost->host_busy++;
1640
1641                 /*
1642                  * XXX(hch): This is rather suboptimal, scsi_dispatch_cmd will
1643                  *              take the lock again.
1644                  */
1645                 spin_unlock_irq(shost->host_lock);
1646
1647                 /*
1648                  * Finally, initialize any error handling parameters, and set up
1649                  * the timers for timeouts.
1650                  */
1651                 scsi_init_cmd_errh(cmd);
1652
1653                 /*
1654                  * Dispatch the command to the low-level driver.
1655                  */
1656                 rtn = scsi_dispatch_cmd(cmd);
1657                 spin_lock_irq(q->queue_lock);
1658                 if(rtn) {
1659                         /* we're refusing the command; because of
1660                          * the way locks get dropped, we need to 
1661                          * check here if plugging is required */
1662                         if(sdev->device_busy == 0)
1663                                 blk_plug_device(q);
1664
1665                         break;
1666                 }
1667         }
1668
1669         goto out;
1670
1671  not_ready:
1672         spin_unlock_irq(shost->host_lock);
1673
1674         /*
1675          * lock q, handle tag, requeue req, and decrement device_busy. We
1676          * must return with queue_lock held.
1677          *
1678          * Decrementing device_busy without checking it is OK, as all such
1679          * cases (host limits or settings) should run the queue at some
1680          * later time.
1681          */
1682         spin_lock_irq(q->queue_lock);
1683         blk_requeue_request(q, req);
1684         sdev->device_busy--;
1685         if(sdev->device_busy == 0)
1686                 blk_plug_device(q);
1687  out:
1688         /* must be careful here...if we trigger the ->remove() function
1689          * we cannot be holding the q lock */
1690         spin_unlock_irq(q->queue_lock);
1691         put_device(&sdev->sdev_gendev);
1692         spin_lock_irq(q->queue_lock);
1693 }
1694
1695 u64 scsi_calculate_bounce_limit(struct Scsi_Host *shost)
1696 {
1697         struct device *host_dev;
1698         u64 bounce_limit = 0xffffffff;
1699
1700         if (shost->unchecked_isa_dma)
1701                 return BLK_BOUNCE_ISA;
1702         /*
1703          * Platforms with virtual-DMA translation
1704          * hardware have no practical limit.
1705          */
1706         if (!PCI_DMA_BUS_IS_PHYS)
1707                 return BLK_BOUNCE_ANY;
1708
1709         host_dev = scsi_get_device(shost);
1710         if (host_dev && host_dev->dma_mask)
1711                 bounce_limit = *host_dev->dma_mask;
1712
1713         return bounce_limit;
1714 }
1715 EXPORT_SYMBOL(scsi_calculate_bounce_limit);
1716
1717 struct request_queue *scsi_alloc_queue(struct scsi_device *sdev)
1718 {
1719         struct Scsi_Host *shost = sdev->host;
1720         struct request_queue *q;
1721
1722         q = blk_init_queue(scsi_request_fn, NULL);
1723         if (!q)
1724                 return NULL;
1725
1726         blk_queue_prep_rq(q, scsi_prep_fn);
1727
1728         blk_queue_max_hw_segments(q, shost->sg_tablesize);
1729         blk_queue_max_phys_segments(q, SCSI_MAX_PHYS_SEGMENTS);
1730         blk_queue_max_sectors(q, shost->max_sectors);
1731         blk_queue_bounce_limit(q, scsi_calculate_bounce_limit(shost));
1732         blk_queue_segment_boundary(q, shost->dma_boundary);
1733         blk_queue_issue_flush_fn(q, scsi_issue_flush_fn);
1734         blk_queue_softirq_done(q, scsi_softirq_done);
1735
1736         if (!shost->use_clustering)
1737                 clear_bit(QUEUE_FLAG_CLUSTER, &q->queue_flags);
1738         return q;
1739 }
1740
1741 void scsi_free_queue(struct request_queue *q)
1742 {
1743         blk_cleanup_queue(q);
1744 }
1745
1746 /*
1747  * Function:    scsi_block_requests()
1748  *
1749  * Purpose:     Utility function used by low-level drivers to prevent further
1750  *              commands from being queued to the device.
1751  *
1752  * Arguments:   shost       - Host in question
1753  *
1754  * Returns:     Nothing
1755  *
1756  * Lock status: No locks are assumed held.
1757  *
1758  * Notes:       There is no timer nor any other means by which the requests
1759  *              get unblocked other than the low-level driver calling
1760  *              scsi_unblock_requests().
1761  */
1762 void scsi_block_requests(struct Scsi_Host *shost)
1763 {
1764         shost->host_self_blocked = 1;
1765 }
1766 EXPORT_SYMBOL(scsi_block_requests);
1767
1768 /*
1769  * Function:    scsi_unblock_requests()
1770  *
1771  * Purpose:     Utility function used by low-level drivers to allow further
1772  *              commands from being queued to the device.
1773  *
1774  * Arguments:   shost       - Host in question
1775  *
1776  * Returns:     Nothing
1777  *
1778  * Lock status: No locks are assumed held.
1779  *
1780  * Notes:       There is no timer nor any other means by which the requests
1781  *              get unblocked other than the low-level driver calling
1782  *              scsi_unblock_requests().
1783  *
1784  *              This is done as an API function so that changes to the
1785  *              internals of the scsi mid-layer won't require wholesale
1786  *              changes to drivers that use this feature.
1787  */
1788 void scsi_unblock_requests(struct Scsi_Host *shost)
1789 {
1790         shost->host_self_blocked = 0;
1791         scsi_run_host_queues(shost);
1792 }
1793 EXPORT_SYMBOL(scsi_unblock_requests);
1794
1795 int __init scsi_init_queue(void)
1796 {
1797         int i;
1798
1799         scsi_io_context_cache = kmem_cache_create("scsi_io_context",
1800                                         sizeof(struct scsi_io_context),
1801                                         0, 0, NULL, NULL);
1802         if (!scsi_io_context_cache) {
1803                 printk(KERN_ERR "SCSI: can't init scsi io context cache\n");
1804                 return -ENOMEM;
1805         }
1806
1807         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1808                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1809                 int size = sgp->size * sizeof(struct scatterlist);
1810
1811                 sgp->slab = kmem_cache_create(sgp->name, size, 0,
1812                                 SLAB_HWCACHE_ALIGN, NULL, NULL);
1813                 if (!sgp->slab) {
1814                         printk(KERN_ERR "SCSI: can't init sg slab %s\n",
1815                                         sgp->name);
1816                 }
1817
1818                 sgp->pool = mempool_create_slab_pool(SG_MEMPOOL_SIZE,
1819                                                      sgp->slab);
1820                 if (!sgp->pool) {
1821                         printk(KERN_ERR "SCSI: can't init sg mempool %s\n",
1822                                         sgp->name);
1823                 }
1824         }
1825
1826         return 0;
1827 }
1828
1829 void scsi_exit_queue(void)
1830 {
1831         int i;
1832
1833         kmem_cache_destroy(scsi_io_context_cache);
1834
1835         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1836                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1837                 mempool_destroy(sgp->pool);
1838                 kmem_cache_destroy(sgp->slab);
1839         }
1840 }
1841
1842 /**
1843  *      scsi_mode_select - issue a mode select
1844  *      @sdev:  SCSI device to be queried
1845  *      @pf:    Page format bit (1 == standard, 0 == vendor specific)
1846  *      @sp:    Save page bit (0 == don't save, 1 == save)
1847  *      @modepage: mode page being requested
1848  *      @buffer: request buffer (may not be smaller than eight bytes)
1849  *      @len:   length of request buffer.
1850  *      @timeout: command timeout
1851  *      @retries: number of retries before failing
1852  *      @data: returns a structure abstracting the mode header data
1853  *      @sense: place to put sense data (or NULL if no sense to be collected).
1854  *              must be SCSI_SENSE_BUFFERSIZE big.
1855  *
1856  *      Returns zero if successful; negative error number or scsi
1857  *      status on error
1858  *
1859  */
1860 int
1861 scsi_mode_select(struct scsi_device *sdev, int pf, int sp, int modepage,
1862                  unsigned char *buffer, int len, int timeout, int retries,
1863                  struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1864 {
1865         unsigned char cmd[10];
1866         unsigned char *real_buffer;
1867         int ret;
1868
1869         memset(cmd, 0, sizeof(cmd));
1870         cmd[1] = (pf ? 0x10 : 0) | (sp ? 0x01 : 0);
1871
1872         if (sdev->use_10_for_ms) {
1873                 if (len > 65535)
1874                         return -EINVAL;
1875                 real_buffer = kmalloc(8 + len, GFP_KERNEL);
1876                 if (!real_buffer)
1877                         return -ENOMEM;
1878                 memcpy(real_buffer + 8, buffer, len);
1879                 len += 8;
1880                 real_buffer[0] = 0;
1881                 real_buffer[1] = 0;
1882                 real_buffer[2] = data->medium_type;
1883                 real_buffer[3] = data->device_specific;
1884                 real_buffer[4] = data->longlba ? 0x01 : 0;
1885                 real_buffer[5] = 0;
1886                 real_buffer[6] = data->block_descriptor_length >> 8;
1887                 real_buffer[7] = data->block_descriptor_length;
1888
1889                 cmd[0] = MODE_SELECT_10;
1890                 cmd[7] = len >> 8;
1891                 cmd[8] = len;
1892         } else {
1893                 if (len > 255 || data->block_descriptor_length > 255 ||
1894                     data->longlba)
1895                         return -EINVAL;
1896
1897                 real_buffer = kmalloc(4 + len, GFP_KERNEL);
1898                 if (!real_buffer)
1899                         return -ENOMEM;
1900                 memcpy(real_buffer + 4, buffer, len);
1901                 len += 4;
1902                 real_buffer[0] = 0;
1903                 real_buffer[1] = data->medium_type;
1904                 real_buffer[2] = data->device_specific;
1905                 real_buffer[3] = data->block_descriptor_length;
1906                 
1907
1908                 cmd[0] = MODE_SELECT;
1909                 cmd[4] = len;
1910         }
1911
1912         ret = scsi_execute_req(sdev, cmd, DMA_TO_DEVICE, real_buffer, len,
1913                                sshdr, timeout, retries);
1914         kfree(real_buffer);
1915         return ret;
1916 }
1917 EXPORT_SYMBOL_GPL(scsi_mode_select);
1918
1919 /**
1920  *      scsi_mode_sense - issue a mode sense, falling back from 10 to 
1921  *              six bytes if necessary.
1922  *      @sdev:  SCSI device to be queried
1923  *      @dbd:   set if mode sense will allow block descriptors to be returned
1924  *      @modepage: mode page being requested
1925  *      @buffer: request buffer (may not be smaller than eight bytes)
1926  *      @len:   length of request buffer.
1927  *      @timeout: command timeout
1928  *      @retries: number of retries before failing
1929  *      @data: returns a structure abstracting the mode header data
1930  *      @sense: place to put sense data (or NULL if no sense to be collected).
1931  *              must be SCSI_SENSE_BUFFERSIZE big.
1932  *
1933  *      Returns zero if unsuccessful, or the header offset (either 4
1934  *      or 8 depending on whether a six or ten byte command was
1935  *      issued) if successful.
1936  **/
1937 int
1938 scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
1939                   unsigned char *buffer, int len, int timeout, int retries,
1940                   struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1941 {
1942         unsigned char cmd[12];
1943         int use_10_for_ms;
1944         int header_length;
1945         int result;
1946         struct scsi_sense_hdr my_sshdr;
1947
1948         memset(data, 0, sizeof(*data));
1949         memset(&cmd[0], 0, 12);
1950         cmd[1] = dbd & 0x18;    /* allows DBD and LLBA bits */
1951         cmd[2] = modepage;
1952
1953         /* caller might not be interested in sense, but we need it */
1954         if (!sshdr)
1955                 sshdr = &my_sshdr;
1956
1957  retry:
1958         use_10_for_ms = sdev->use_10_for_ms;
1959
1960         if (use_10_for_ms) {
1961                 if (len < 8)
1962                         len = 8;
1963
1964                 cmd[0] = MODE_SENSE_10;
1965                 cmd[8] = len;
1966                 header_length = 8;
1967         } else {
1968                 if (len < 4)
1969                         len = 4;
1970
1971                 cmd[0] = MODE_SENSE;
1972                 cmd[4] = len;
1973                 header_length = 4;
1974         }
1975
1976         memset(buffer, 0, len);
1977
1978         result = scsi_execute_req(sdev, cmd, DMA_FROM_DEVICE, buffer, len,
1979                                   sshdr, timeout, retries);
1980
1981         /* This code looks awful: what it's doing is making sure an
1982          * ILLEGAL REQUEST sense return identifies the actual command
1983          * byte as the problem.  MODE_SENSE commands can return
1984          * ILLEGAL REQUEST if the code page isn't supported */
1985
1986         if (use_10_for_ms && !scsi_status_is_good(result) &&
1987             (driver_byte(result) & DRIVER_SENSE)) {
1988                 if (scsi_sense_valid(sshdr)) {
1989                         if ((sshdr->sense_key == ILLEGAL_REQUEST) &&
1990                             (sshdr->asc == 0x20) && (sshdr->ascq == 0)) {
1991                                 /* 
1992                                  * Invalid command operation code
1993                                  */
1994                                 sdev->use_10_for_ms = 0;
1995                                 goto retry;
1996                         }
1997                 }
1998         }
1999
2000         if(scsi_status_is_good(result)) {
2001                 if (unlikely(buffer[0] == 0x86 && buffer[1] == 0x0b &&
2002                              (modepage == 6 || modepage == 8))) {
2003                         /* Initio breakage? */
2004                         header_length = 0;
2005                         data->length = 13;
2006                         data->medium_type = 0;
2007                         data->device_specific = 0;
2008                         data->longlba = 0;
2009                         data->block_descriptor_length = 0;
2010                 } else if(use_10_for_ms) {
2011                         data->length = buffer[0]*256 + buffer[1] + 2;
2012                         data->medium_type = buffer[2];
2013                         data->device_specific = buffer[3];
2014                         data->longlba = buffer[4] & 0x01;
2015                         data->block_descriptor_length = buffer[6]*256
2016                                 + buffer[7];
2017                 } else {
2018                         data->length = buffer[0] + 1;
2019                         data->medium_type = buffer[1];
2020                         data->device_specific = buffer[2];
2021                         data->block_descriptor_length = buffer[3];
2022                 }
2023                 data->header_length = header_length;
2024         }
2025
2026         return result;
2027 }
2028 EXPORT_SYMBOL(scsi_mode_sense);
2029
2030 int
2031 scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries)
2032 {
2033         char cmd[] = {
2034                 TEST_UNIT_READY, 0, 0, 0, 0, 0,
2035         };
2036         struct scsi_sense_hdr sshdr;
2037         int result;
2038         
2039         result = scsi_execute_req(sdev, cmd, DMA_NONE, NULL, 0, &sshdr,
2040                                   timeout, retries);
2041
2042         if ((driver_byte(result) & DRIVER_SENSE) && sdev->removable) {
2043
2044                 if ((scsi_sense_valid(&sshdr)) &&
2045                     ((sshdr.sense_key == UNIT_ATTENTION) ||
2046                      (sshdr.sense_key == NOT_READY))) {
2047                         sdev->changed = 1;
2048                         result = 0;
2049                 }
2050         }
2051         return result;
2052 }
2053 EXPORT_SYMBOL(scsi_test_unit_ready);
2054
2055 /**
2056  *      scsi_device_set_state - Take the given device through the device
2057  *              state model.
2058  *      @sdev:  scsi device to change the state of.
2059  *      @state: state to change to.
2060  *
2061  *      Returns zero if unsuccessful or an error if the requested 
2062  *      transition is illegal.
2063  **/
2064 int
2065 scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
2066 {
2067         enum scsi_device_state oldstate = sdev->sdev_state;
2068
2069         if (state == oldstate)
2070                 return 0;
2071
2072         switch (state) {
2073         case SDEV_CREATED:
2074                 /* There are no legal states that come back to
2075                  * created.  This is the manually initialised start
2076                  * state */
2077                 goto illegal;
2078                         
2079         case SDEV_RUNNING:
2080                 switch (oldstate) {
2081                 case SDEV_CREATED:
2082                 case SDEV_OFFLINE:
2083                 case SDEV_QUIESCE:
2084                 case SDEV_BLOCK:
2085                         break;
2086                 default:
2087                         goto illegal;
2088                 }
2089                 break;
2090
2091         case SDEV_QUIESCE:
2092                 switch (oldstate) {
2093                 case SDEV_RUNNING:
2094                 case SDEV_OFFLINE:
2095                         break;
2096                 default:
2097                         goto illegal;
2098                 }
2099                 break;
2100
2101         case SDEV_OFFLINE:
2102                 switch (oldstate) {
2103                 case SDEV_CREATED:
2104                 case SDEV_RUNNING:
2105                 case SDEV_QUIESCE:
2106                 case SDEV_BLOCK:
2107                         break;
2108                 default:
2109                         goto illegal;
2110                 }
2111                 break;
2112
2113         case SDEV_BLOCK:
2114                 switch (oldstate) {
2115                 case SDEV_CREATED:
2116                 case SDEV_RUNNING:
2117                         break;
2118                 default:
2119                         goto illegal;
2120                 }
2121                 break;
2122
2123         case SDEV_CANCEL:
2124                 switch (oldstate) {
2125                 case SDEV_CREATED:
2126                 case SDEV_RUNNING:
2127                 case SDEV_OFFLINE:
2128                 case SDEV_BLOCK:
2129                         break;
2130                 default:
2131                         goto illegal;
2132                 }
2133                 break;
2134
2135         case SDEV_DEL:
2136                 switch (oldstate) {
2137                 case SDEV_CANCEL:
2138                         break;
2139                 default:
2140                         goto illegal;
2141                 }
2142                 break;
2143
2144         }
2145         sdev->sdev_state = state;
2146         return 0;
2147
2148  illegal:
2149         SCSI_LOG_ERROR_RECOVERY(1, 
2150                                 sdev_printk(KERN_ERR, sdev,
2151                                             "Illegal state transition %s->%s\n",
2152                                             scsi_device_state_name(oldstate),
2153                                             scsi_device_state_name(state))
2154                                 );
2155         return -EINVAL;
2156 }
2157 EXPORT_SYMBOL(scsi_device_set_state);
2158
2159 /**
2160  *      scsi_device_quiesce - Block user issued commands.
2161  *      @sdev:  scsi device to quiesce.
2162  *
2163  *      This works by trying to transition to the SDEV_QUIESCE state
2164  *      (which must be a legal transition).  When the device is in this
2165  *      state, only special requests will be accepted, all others will
2166  *      be deferred.  Since special requests may also be requeued requests,
2167  *      a successful return doesn't guarantee the device will be 
2168  *      totally quiescent.
2169  *
2170  *      Must be called with user context, may sleep.
2171  *
2172  *      Returns zero if unsuccessful or an error if not.
2173  **/
2174 int
2175 scsi_device_quiesce(struct scsi_device *sdev)
2176 {
2177         int err = scsi_device_set_state(sdev, SDEV_QUIESCE);
2178         if (err)
2179                 return err;
2180
2181         scsi_run_queue(sdev->request_queue);
2182         while (sdev->device_busy) {
2183                 msleep_interruptible(200);
2184                 scsi_run_queue(sdev->request_queue);
2185         }
2186         return 0;
2187 }
2188 EXPORT_SYMBOL(scsi_device_quiesce);
2189
2190 /**
2191  *      scsi_device_resume - Restart user issued commands to a quiesced device.
2192  *      @sdev:  scsi device to resume.
2193  *
2194  *      Moves the device from quiesced back to running and restarts the
2195  *      queues.
2196  *
2197  *      Must be called with user context, may sleep.
2198  **/
2199 void
2200 scsi_device_resume(struct scsi_device *sdev)
2201 {
2202         if(scsi_device_set_state(sdev, SDEV_RUNNING))
2203                 return;
2204         scsi_run_queue(sdev->request_queue);
2205 }
2206 EXPORT_SYMBOL(scsi_device_resume);
2207
2208 static void
2209 device_quiesce_fn(struct scsi_device *sdev, void *data)
2210 {
2211         scsi_device_quiesce(sdev);
2212 }
2213
2214 void
2215 scsi_target_quiesce(struct scsi_target *starget)
2216 {
2217         starget_for_each_device(starget, NULL, device_quiesce_fn);
2218 }
2219 EXPORT_SYMBOL(scsi_target_quiesce);
2220
2221 static void
2222 device_resume_fn(struct scsi_device *sdev, void *data)
2223 {
2224         scsi_device_resume(sdev);
2225 }
2226
2227 void
2228 scsi_target_resume(struct scsi_target *starget)
2229 {
2230         starget_for_each_device(starget, NULL, device_resume_fn);
2231 }
2232 EXPORT_SYMBOL(scsi_target_resume);
2233
2234 /**
2235  * scsi_internal_device_block - internal function to put a device
2236  *                              temporarily into the SDEV_BLOCK state
2237  * @sdev:       device to block
2238  *
2239  * Block request made by scsi lld's to temporarily stop all
2240  * scsi commands on the specified device.  Called from interrupt
2241  * or normal process context.
2242  *
2243  * Returns zero if successful or error if not
2244  *
2245  * Notes:       
2246  *      This routine transitions the device to the SDEV_BLOCK state
2247  *      (which must be a legal transition).  When the device is in this
2248  *      state, all commands are deferred until the scsi lld reenables
2249  *      the device with scsi_device_unblock or device_block_tmo fires.
2250  *      This routine assumes the host_lock is held on entry.
2251  **/
2252 int
2253 scsi_internal_device_block(struct scsi_device *sdev)
2254 {
2255         request_queue_t *q = sdev->request_queue;
2256         unsigned long flags;
2257         int err = 0;
2258
2259         err = scsi_device_set_state(sdev, SDEV_BLOCK);
2260         if (err)
2261                 return err;
2262
2263         /* 
2264          * The device has transitioned to SDEV_BLOCK.  Stop the
2265          * block layer from calling the midlayer with this device's
2266          * request queue. 
2267          */
2268         spin_lock_irqsave(q->queue_lock, flags);
2269         blk_stop_queue(q);
2270         spin_unlock_irqrestore(q->queue_lock, flags);
2271
2272         return 0;
2273 }
2274 EXPORT_SYMBOL_GPL(scsi_internal_device_block);
2275  
2276 /**
2277  * scsi_internal_device_unblock - resume a device after a block request
2278  * @sdev:       device to resume
2279  *
2280  * Called by scsi lld's or the midlayer to restart the device queue
2281  * for the previously suspended scsi device.  Called from interrupt or
2282  * normal process context.
2283  *
2284  * Returns zero if successful or error if not.
2285  *
2286  * Notes:       
2287  *      This routine transitions the device to the SDEV_RUNNING state
2288  *      (which must be a legal transition) allowing the midlayer to
2289  *      goose the queue for this device.  This routine assumes the 
2290  *      host_lock is held upon entry.
2291  **/
2292 int
2293 scsi_internal_device_unblock(struct scsi_device *sdev)
2294 {
2295         request_queue_t *q = sdev->request_queue; 
2296         int err;
2297         unsigned long flags;
2298         
2299         /* 
2300          * Try to transition the scsi device to SDEV_RUNNING
2301          * and goose the device queue if successful.  
2302          */
2303         err = scsi_device_set_state(sdev, SDEV_RUNNING);
2304         if (err)
2305                 return err;
2306
2307         spin_lock_irqsave(q->queue_lock, flags);
2308         blk_start_queue(q);
2309         spin_unlock_irqrestore(q->queue_lock, flags);
2310
2311         return 0;
2312 }
2313 EXPORT_SYMBOL_GPL(scsi_internal_device_unblock);
2314
2315 static void
2316 device_block(struct scsi_device *sdev, void *data)
2317 {
2318         scsi_internal_device_block(sdev);
2319 }
2320
2321 static int
2322 target_block(struct device *dev, void *data)
2323 {
2324         if (scsi_is_target_device(dev))
2325                 starget_for_each_device(to_scsi_target(dev), NULL,
2326                                         device_block);
2327         return 0;
2328 }
2329
2330 void
2331 scsi_target_block(struct device *dev)
2332 {
2333         if (scsi_is_target_device(dev))
2334                 starget_for_each_device(to_scsi_target(dev), NULL,
2335                                         device_block);
2336         else
2337                 device_for_each_child(dev, NULL, target_block);
2338 }
2339 EXPORT_SYMBOL_GPL(scsi_target_block);
2340
2341 static void
2342 device_unblock(struct scsi_device *sdev, void *data)
2343 {
2344         scsi_internal_device_unblock(sdev);
2345 }
2346
2347 static int
2348 target_unblock(struct device *dev, void *data)
2349 {
2350         if (scsi_is_target_device(dev))
2351                 starget_for_each_device(to_scsi_target(dev), NULL,
2352                                         device_unblock);
2353         return 0;
2354 }
2355
2356 void
2357 scsi_target_unblock(struct device *dev)
2358 {
2359         if (scsi_is_target_device(dev))
2360                 starget_for_each_device(to_scsi_target(dev), NULL,
2361                                         device_unblock);
2362         else
2363                 device_for_each_child(dev, NULL, target_unblock);
2364 }
2365 EXPORT_SYMBOL_GPL(scsi_target_unblock);