[BLOCK] add @uptodate to end_that_request_last() and @error to rq_end_io_fn()
[linux-2.6.git] / drivers / scsi / scsi_lib.c
1 /*
2  *  scsi_lib.c Copyright (C) 1999 Eric Youngdale
3  *
4  *  SCSI queueing library.
5  *      Initial versions: Eric Youngdale (eric@andante.org).
6  *                        Based upon conversations with large numbers
7  *                        of people at Linux Expo.
8  */
9
10 #include <linux/bio.h>
11 #include <linux/blkdev.h>
12 #include <linux/completion.h>
13 #include <linux/kernel.h>
14 #include <linux/mempool.h>
15 #include <linux/slab.h>
16 #include <linux/init.h>
17 #include <linux/pci.h>
18 #include <linux/delay.h>
19
20 #include <scsi/scsi.h>
21 #include <scsi/scsi_dbg.h>
22 #include <scsi/scsi_device.h>
23 #include <scsi/scsi_driver.h>
24 #include <scsi/scsi_eh.h>
25 #include <scsi/scsi_host.h>
26 #include <scsi/scsi_request.h>
27
28 #include "scsi_priv.h"
29 #include "scsi_logging.h"
30
31
32 #define SG_MEMPOOL_NR           (sizeof(scsi_sg_pools)/sizeof(struct scsi_host_sg_pool))
33 #define SG_MEMPOOL_SIZE         32
34
35 struct scsi_host_sg_pool {
36         size_t          size;
37         char            *name; 
38         kmem_cache_t    *slab;
39         mempool_t       *pool;
40 };
41
42 #if (SCSI_MAX_PHYS_SEGMENTS < 32)
43 #error SCSI_MAX_PHYS_SEGMENTS is too small
44 #endif
45
46 #define SP(x) { x, "sgpool-" #x } 
47 static struct scsi_host_sg_pool scsi_sg_pools[] = {
48         SP(8),
49         SP(16),
50         SP(32),
51 #if (SCSI_MAX_PHYS_SEGMENTS > 32)
52         SP(64),
53 #if (SCSI_MAX_PHYS_SEGMENTS > 64)
54         SP(128),
55 #if (SCSI_MAX_PHYS_SEGMENTS > 128)
56         SP(256),
57 #if (SCSI_MAX_PHYS_SEGMENTS > 256)
58 #error SCSI_MAX_PHYS_SEGMENTS is too large
59 #endif
60 #endif
61 #endif
62 #endif
63 };      
64 #undef SP
65
66 static void scsi_run_queue(struct request_queue *q);
67
68 /*
69  * Function:    scsi_unprep_request()
70  *
71  * Purpose:     Remove all preparation done for a request, including its
72  *              associated scsi_cmnd, so that it can be requeued.
73  *
74  * Arguments:   req     - request to unprepare
75  *
76  * Lock status: Assumed that no locks are held upon entry.
77  *
78  * Returns:     Nothing.
79  */
80 static void scsi_unprep_request(struct request *req)
81 {
82         struct scsi_cmnd *cmd = req->special;
83
84         req->flags &= ~REQ_DONTPREP;
85         req->special = (req->flags & REQ_SPECIAL) ? cmd->sc_request : NULL;
86
87         scsi_put_command(cmd);
88 }
89
90 /*
91  * Function:    scsi_queue_insert()
92  *
93  * Purpose:     Insert a command in the midlevel queue.
94  *
95  * Arguments:   cmd    - command that we are adding to queue.
96  *              reason - why we are inserting command to queue.
97  *
98  * Lock status: Assumed that lock is not held upon entry.
99  *
100  * Returns:     Nothing.
101  *
102  * Notes:       We do this for one of two cases.  Either the host is busy
103  *              and it cannot accept any more commands for the time being,
104  *              or the device returned QUEUE_FULL and can accept no more
105  *              commands.
106  * Notes:       This could be called either from an interrupt context or a
107  *              normal process context.
108  */
109 int scsi_queue_insert(struct scsi_cmnd *cmd, int reason)
110 {
111         struct Scsi_Host *host = cmd->device->host;
112         struct scsi_device *device = cmd->device;
113         struct request_queue *q = device->request_queue;
114         unsigned long flags;
115
116         SCSI_LOG_MLQUEUE(1,
117                  printk("Inserting command %p into mlqueue\n", cmd));
118
119         /*
120          * Set the appropriate busy bit for the device/host.
121          *
122          * If the host/device isn't busy, assume that something actually
123          * completed, and that we should be able to queue a command now.
124          *
125          * Note that the prior mid-layer assumption that any host could
126          * always queue at least one command is now broken.  The mid-layer
127          * will implement a user specifiable stall (see
128          * scsi_host.max_host_blocked and scsi_device.max_device_blocked)
129          * if a command is requeued with no other commands outstanding
130          * either for the device or for the host.
131          */
132         if (reason == SCSI_MLQUEUE_HOST_BUSY)
133                 host->host_blocked = host->max_host_blocked;
134         else if (reason == SCSI_MLQUEUE_DEVICE_BUSY)
135                 device->device_blocked = device->max_device_blocked;
136
137         /*
138          * Decrement the counters, since these commands are no longer
139          * active on the host/device.
140          */
141         scsi_device_unbusy(device);
142
143         /*
144          * Requeue this command.  It will go before all other commands
145          * that are already in the queue.
146          *
147          * NOTE: there is magic here about the way the queue is plugged if
148          * we have no outstanding commands.
149          * 
150          * Although we *don't* plug the queue, we call the request
151          * function.  The SCSI request function detects the blocked condition
152          * and plugs the queue appropriately.
153          */
154         spin_lock_irqsave(q->queue_lock, flags);
155         blk_requeue_request(q, cmd->request);
156         spin_unlock_irqrestore(q->queue_lock, flags);
157
158         scsi_run_queue(q);
159
160         return 0;
161 }
162
163 /*
164  * Function:    scsi_do_req
165  *
166  * Purpose:     Queue a SCSI request
167  *
168  * Arguments:   sreq      - command descriptor.
169  *              cmnd      - actual SCSI command to be performed.
170  *              buffer    - data buffer.
171  *              bufflen   - size of data buffer.
172  *              done      - completion function to be run.
173  *              timeout   - how long to let it run before timeout.
174  *              retries   - number of retries we allow.
175  *
176  * Lock status: No locks held upon entry.
177  *
178  * Returns:     Nothing.
179  *
180  * Notes:       This function is only used for queueing requests for things
181  *              like ioctls and character device requests - this is because
182  *              we essentially just inject a request into the queue for the
183  *              device.
184  *
185  *              In order to support the scsi_device_quiesce function, we
186  *              now inject requests on the *head* of the device queue
187  *              rather than the tail.
188  */
189 void scsi_do_req(struct scsi_request *sreq, const void *cmnd,
190                  void *buffer, unsigned bufflen,
191                  void (*done)(struct scsi_cmnd *),
192                  int timeout, int retries)
193 {
194         /*
195          * If the upper level driver is reusing these things, then
196          * we should release the low-level block now.  Another one will
197          * be allocated later when this request is getting queued.
198          */
199         __scsi_release_request(sreq);
200
201         /*
202          * Our own function scsi_done (which marks the host as not busy,
203          * disables the timeout counter, etc) will be called by us or by the
204          * scsi_hosts[host].queuecommand() function needs to also call
205          * the completion function for the high level driver.
206          */
207         memcpy(sreq->sr_cmnd, cmnd, sizeof(sreq->sr_cmnd));
208         sreq->sr_bufflen = bufflen;
209         sreq->sr_buffer = buffer;
210         sreq->sr_allowed = retries;
211         sreq->sr_done = done;
212         sreq->sr_timeout_per_command = timeout;
213
214         if (sreq->sr_cmd_len == 0)
215                 sreq->sr_cmd_len = COMMAND_SIZE(sreq->sr_cmnd[0]);
216
217         /*
218          * head injection *required* here otherwise quiesce won't work
219          *
220          * Because users of this function are apt to reuse requests with no
221          * modification, we have to sanitise the request flags here
222          */
223         sreq->sr_request->flags &= ~REQ_DONTPREP;
224         blk_insert_request(sreq->sr_device->request_queue, sreq->sr_request,
225                            1, sreq);
226 }
227 EXPORT_SYMBOL(scsi_do_req);
228
229 /**
230  * scsi_execute - insert request and wait for the result
231  * @sdev:       scsi device
232  * @cmd:        scsi command
233  * @data_direction: data direction
234  * @buffer:     data buffer
235  * @bufflen:    len of buffer
236  * @sense:      optional sense buffer
237  * @timeout:    request timeout in seconds
238  * @retries:    number of times to retry request
239  * @flags:      or into request flags;
240  *
241  * returns the req->errors value which is the the scsi_cmnd result
242  * field.
243  **/
244 int scsi_execute(struct scsi_device *sdev, const unsigned char *cmd,
245                  int data_direction, void *buffer, unsigned bufflen,
246                  unsigned char *sense, int timeout, int retries, int flags)
247 {
248         struct request *req;
249         int write = (data_direction == DMA_TO_DEVICE);
250         int ret = DRIVER_ERROR << 24;
251
252         req = blk_get_request(sdev->request_queue, write, __GFP_WAIT);
253
254         if (bufflen &&  blk_rq_map_kern(sdev->request_queue, req,
255                                         buffer, bufflen, __GFP_WAIT))
256                 goto out;
257
258         req->cmd_len = COMMAND_SIZE(cmd[0]);
259         memcpy(req->cmd, cmd, req->cmd_len);
260         req->sense = sense;
261         req->sense_len = 0;
262         req->retries = retries;
263         req->timeout = timeout;
264         req->flags |= flags | REQ_BLOCK_PC | REQ_SPECIAL | REQ_QUIET;
265
266         /*
267          * head injection *required* here otherwise quiesce won't work
268          */
269         blk_execute_rq(req->q, NULL, req, 1);
270
271         ret = req->errors;
272  out:
273         blk_put_request(req);
274
275         return ret;
276 }
277 EXPORT_SYMBOL(scsi_execute);
278
279
280 int scsi_execute_req(struct scsi_device *sdev, const unsigned char *cmd,
281                      int data_direction, void *buffer, unsigned bufflen,
282                      struct scsi_sense_hdr *sshdr, int timeout, int retries)
283 {
284         char *sense = NULL;
285         int result;
286         
287         if (sshdr) {
288                 sense = kmalloc(SCSI_SENSE_BUFFERSIZE, GFP_NOIO);
289                 if (!sense)
290                         return DRIVER_ERROR << 24;
291                 memset(sense, 0, SCSI_SENSE_BUFFERSIZE);
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)
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 + 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  * @data_direction: data direction
440  * @buffer:     data buffer (this can be a kernel buffer or scatterlist)
441  * @bufflen:    len of buffer
442  * @use_sg:     if buffer is a scatterlist this is the number of elements
443  * @timeout:    request timeout in seconds
444  * @retries:    number of times to retry request
445  * @flags:      or into request flags
446  **/
447 int scsi_execute_async(struct scsi_device *sdev, const unsigned char *cmd,
448                        int data_direction, void *buffer, unsigned bufflen,
449                        int use_sg, int timeout, int retries, void *privdata,
450                        void (*done)(void *, char *, int, int), gfp_t gfp)
451 {
452         struct request *req;
453         struct scsi_io_context *sioc;
454         int err = 0;
455         int write = (data_direction == DMA_TO_DEVICE);
456
457         sioc = kmem_cache_alloc(scsi_io_context_cache, gfp);
458         if (!sioc)
459                 return DRIVER_ERROR << 24;
460         memset(sioc, 0, sizeof(*sioc));
461
462         req = blk_get_request(sdev->request_queue, write, gfp);
463         if (!req)
464                 goto free_sense;
465         req->flags |= REQ_BLOCK_PC | REQ_QUIET;
466
467         if (use_sg)
468                 err = scsi_req_map_sg(req, buffer, use_sg, bufflen, gfp);
469         else if (bufflen)
470                 err = blk_rq_map_kern(req->q, req, buffer, bufflen, gfp);
471
472         if (err)
473                 goto free_req;
474
475         req->cmd_len = COMMAND_SIZE(cmd[0]);
476         memcpy(req->cmd, cmd, req->cmd_len);
477         req->sense = sioc->sense;
478         req->sense_len = 0;
479         req->timeout = timeout;
480         req->retries = retries;
481         req->end_io_data = sioc;
482
483         sioc->data = privdata;
484         sioc->done = done;
485
486         blk_execute_rq_nowait(req->q, NULL, req, 1, scsi_end_async);
487         return 0;
488
489 free_req:
490         blk_put_request(req);
491 free_sense:
492         kfree(sioc);
493         return DRIVER_ERROR << 24;
494 }
495 EXPORT_SYMBOL_GPL(scsi_execute_async);
496
497 /*
498  * Function:    scsi_init_cmd_errh()
499  *
500  * Purpose:     Initialize cmd fields related to error handling.
501  *
502  * Arguments:   cmd     - command that is ready to be queued.
503  *
504  * Returns:     Nothing
505  *
506  * Notes:       This function has the job of initializing a number of
507  *              fields related to error handling.   Typically this will
508  *              be called once for each command, as required.
509  */
510 static int scsi_init_cmd_errh(struct scsi_cmnd *cmd)
511 {
512         cmd->serial_number = 0;
513
514         memset(cmd->sense_buffer, 0, sizeof cmd->sense_buffer);
515
516         if (cmd->cmd_len == 0)
517                 cmd->cmd_len = COMMAND_SIZE(cmd->cmnd[0]);
518
519         /*
520          * We need saved copies of a number of fields - this is because
521          * error handling may need to overwrite these with different values
522          * to run different commands, and once error handling is complete,
523          * we will need to restore these values prior to running the actual
524          * command.
525          */
526         cmd->old_use_sg = cmd->use_sg;
527         cmd->old_cmd_len = cmd->cmd_len;
528         cmd->sc_old_data_direction = cmd->sc_data_direction;
529         cmd->old_underflow = cmd->underflow;
530         memcpy(cmd->data_cmnd, cmd->cmnd, sizeof(cmd->cmnd));
531         cmd->buffer = cmd->request_buffer;
532         cmd->bufflen = cmd->request_bufflen;
533
534         return 1;
535 }
536
537 /*
538  * Function:   scsi_setup_cmd_retry()
539  *
540  * Purpose:    Restore the command state for a retry
541  *
542  * Arguments:  cmd      - command to be restored
543  *
544  * Returns:    Nothing
545  *
546  * Notes:      Immediately prior to retrying a command, we need
547  *             to restore certain fields that we saved above.
548  */
549 void scsi_setup_cmd_retry(struct scsi_cmnd *cmd)
550 {
551         memcpy(cmd->cmnd, cmd->data_cmnd, sizeof(cmd->data_cmnd));
552         cmd->request_buffer = cmd->buffer;
553         cmd->request_bufflen = cmd->bufflen;
554         cmd->use_sg = cmd->old_use_sg;
555         cmd->cmd_len = cmd->old_cmd_len;
556         cmd->sc_data_direction = cmd->sc_old_data_direction;
557         cmd->underflow = cmd->old_underflow;
558 }
559
560 void scsi_device_unbusy(struct scsi_device *sdev)
561 {
562         struct Scsi_Host *shost = sdev->host;
563         unsigned long flags;
564
565         spin_lock_irqsave(shost->host_lock, flags);
566         shost->host_busy--;
567         if (unlikely(scsi_host_in_recovery(shost) &&
568                      shost->host_failed))
569                 scsi_eh_wakeup(shost);
570         spin_unlock(shost->host_lock);
571         spin_lock(sdev->request_queue->queue_lock);
572         sdev->device_busy--;
573         spin_unlock_irqrestore(sdev->request_queue->queue_lock, flags);
574 }
575
576 /*
577  * Called for single_lun devices on IO completion. Clear starget_sdev_user,
578  * and call blk_run_queue for all the scsi_devices on the target -
579  * including current_sdev first.
580  *
581  * Called with *no* scsi locks held.
582  */
583 static void scsi_single_lun_run(struct scsi_device *current_sdev)
584 {
585         struct Scsi_Host *shost = current_sdev->host;
586         struct scsi_device *sdev, *tmp;
587         struct scsi_target *starget = scsi_target(current_sdev);
588         unsigned long flags;
589
590         spin_lock_irqsave(shost->host_lock, flags);
591         starget->starget_sdev_user = NULL;
592         spin_unlock_irqrestore(shost->host_lock, flags);
593
594         /*
595          * Call blk_run_queue for all LUNs on the target, starting with
596          * current_sdev. We race with others (to set starget_sdev_user),
597          * but in most cases, we will be first. Ideally, each LU on the
598          * target would get some limited time or requests on the target.
599          */
600         blk_run_queue(current_sdev->request_queue);
601
602         spin_lock_irqsave(shost->host_lock, flags);
603         if (starget->starget_sdev_user)
604                 goto out;
605         list_for_each_entry_safe(sdev, tmp, &starget->devices,
606                         same_target_siblings) {
607                 if (sdev == current_sdev)
608                         continue;
609                 if (scsi_device_get(sdev))
610                         continue;
611
612                 spin_unlock_irqrestore(shost->host_lock, flags);
613                 blk_run_queue(sdev->request_queue);
614                 spin_lock_irqsave(shost->host_lock, flags);
615         
616                 scsi_device_put(sdev);
617         }
618  out:
619         spin_unlock_irqrestore(shost->host_lock, flags);
620 }
621
622 /*
623  * Function:    scsi_run_queue()
624  *
625  * Purpose:     Select a proper request queue to serve next
626  *
627  * Arguments:   q       - last request's queue
628  *
629  * Returns:     Nothing
630  *
631  * Notes:       The previous command was completely finished, start
632  *              a new one if possible.
633  */
634 static void scsi_run_queue(struct request_queue *q)
635 {
636         struct scsi_device *sdev = q->queuedata;
637         struct Scsi_Host *shost = sdev->host;
638         unsigned long flags;
639
640         if (sdev->single_lun)
641                 scsi_single_lun_run(sdev);
642
643         spin_lock_irqsave(shost->host_lock, flags);
644         while (!list_empty(&shost->starved_list) &&
645                !shost->host_blocked && !shost->host_self_blocked &&
646                 !((shost->can_queue > 0) &&
647                   (shost->host_busy >= shost->can_queue))) {
648                 /*
649                  * As long as shost is accepting commands and we have
650                  * starved queues, call blk_run_queue. scsi_request_fn
651                  * drops the queue_lock and can add us back to the
652                  * starved_list.
653                  *
654                  * host_lock protects the starved_list and starved_entry.
655                  * scsi_request_fn must get the host_lock before checking
656                  * or modifying starved_list or starved_entry.
657                  */
658                 sdev = list_entry(shost->starved_list.next,
659                                           struct scsi_device, starved_entry);
660                 list_del_init(&sdev->starved_entry);
661                 spin_unlock_irqrestore(shost->host_lock, flags);
662
663                 blk_run_queue(sdev->request_queue);
664
665                 spin_lock_irqsave(shost->host_lock, flags);
666                 if (unlikely(!list_empty(&sdev->starved_entry)))
667                         /*
668                          * sdev lost a race, and was put back on the
669                          * starved list. This is unlikely but without this
670                          * in theory we could loop forever.
671                          */
672                         break;
673         }
674         spin_unlock_irqrestore(shost->host_lock, flags);
675
676         blk_run_queue(q);
677 }
678
679 /*
680  * Function:    scsi_requeue_command()
681  *
682  * Purpose:     Handle post-processing of completed commands.
683  *
684  * Arguments:   q       - queue to operate on
685  *              cmd     - command that may need to be requeued.
686  *
687  * Returns:     Nothing
688  *
689  * Notes:       After command completion, there may be blocks left
690  *              over which weren't finished by the previous command
691  *              this can be for a number of reasons - the main one is
692  *              I/O errors in the middle of the request, in which case
693  *              we need to request the blocks that come after the bad
694  *              sector.
695  * Notes:       Upon return, cmd is a stale pointer.
696  */
697 static void scsi_requeue_command(struct request_queue *q, struct scsi_cmnd *cmd)
698 {
699         struct request *req = cmd->request;
700         unsigned long flags;
701
702         scsi_unprep_request(req);
703         spin_lock_irqsave(q->queue_lock, flags);
704         blk_requeue_request(q, req);
705         spin_unlock_irqrestore(q->queue_lock, flags);
706
707         scsi_run_queue(q);
708 }
709
710 void scsi_next_command(struct scsi_cmnd *cmd)
711 {
712         struct scsi_device *sdev = cmd->device;
713         struct request_queue *q = sdev->request_queue;
714
715         /* need to hold a reference on the device before we let go of the cmd */
716         get_device(&sdev->sdev_gendev);
717
718         scsi_put_command(cmd);
719         scsi_run_queue(q);
720
721         /* ok to remove device now */
722         put_device(&sdev->sdev_gendev);
723 }
724
725 void scsi_run_host_queues(struct Scsi_Host *shost)
726 {
727         struct scsi_device *sdev;
728
729         shost_for_each_device(sdev, shost)
730                 scsi_run_queue(sdev->request_queue);
731 }
732
733 /*
734  * Function:    scsi_end_request()
735  *
736  * Purpose:     Post-processing of completed commands (usually invoked at end
737  *              of upper level post-processing and scsi_io_completion).
738  *
739  * Arguments:   cmd      - command that is complete.
740  *              uptodate - 1 if I/O indicates success, <= 0 for I/O error.
741  *              bytes    - number of bytes of completed I/O
742  *              requeue  - indicates whether we should requeue leftovers.
743  *
744  * Lock status: Assumed that lock is not held upon entry.
745  *
746  * Returns:     cmd if requeue required, NULL otherwise.
747  *
748  * Notes:       This is called for block device requests in order to
749  *              mark some number of sectors as complete.
750  * 
751  *              We are guaranteeing that the request queue will be goosed
752  *              at some point during this call.
753  * Notes:       If cmd was requeued, upon return it will be a stale pointer.
754  */
755 static struct scsi_cmnd *scsi_end_request(struct scsi_cmnd *cmd, int uptodate,
756                                           int bytes, int requeue)
757 {
758         request_queue_t *q = cmd->device->request_queue;
759         struct request *req = cmd->request;
760         unsigned long flags;
761
762         /*
763          * If there are blocks left over at the end, set up the command
764          * to queue the remainder of them.
765          */
766         if (end_that_request_chunk(req, uptodate, bytes)) {
767                 int leftover = (req->hard_nr_sectors << 9);
768
769                 if (blk_pc_request(req))
770                         leftover = req->data_len;
771
772                 /* kill remainder if no retrys */
773                 if (!uptodate && blk_noretry_request(req))
774                         end_that_request_chunk(req, 0, leftover);
775                 else {
776                         if (requeue) {
777                                 /*
778                                  * Bleah.  Leftovers again.  Stick the
779                                  * leftovers in the front of the
780                                  * queue, and goose the queue again.
781                                  */
782                                 scsi_requeue_command(q, cmd);
783                                 cmd = NULL;
784                         }
785                         return cmd;
786                 }
787         }
788
789         add_disk_randomness(req->rq_disk);
790
791         spin_lock_irqsave(q->queue_lock, flags);
792         if (blk_rq_tagged(req))
793                 blk_queue_end_tag(q, req);
794         end_that_request_last(req, uptodate);
795         spin_unlock_irqrestore(q->queue_lock, flags);
796
797         /*
798          * This will goose the queue request function at the end, so we don't
799          * need to worry about launching another command.
800          */
801         scsi_next_command(cmd);
802         return NULL;
803 }
804
805 static struct scatterlist *scsi_alloc_sgtable(struct scsi_cmnd *cmd, gfp_t gfp_mask)
806 {
807         struct scsi_host_sg_pool *sgp;
808         struct scatterlist *sgl;
809
810         BUG_ON(!cmd->use_sg);
811
812         switch (cmd->use_sg) {
813         case 1 ... 8:
814                 cmd->sglist_len = 0;
815                 break;
816         case 9 ... 16:
817                 cmd->sglist_len = 1;
818                 break;
819         case 17 ... 32:
820                 cmd->sglist_len = 2;
821                 break;
822 #if (SCSI_MAX_PHYS_SEGMENTS > 32)
823         case 33 ... 64:
824                 cmd->sglist_len = 3;
825                 break;
826 #if (SCSI_MAX_PHYS_SEGMENTS > 64)
827         case 65 ... 128:
828                 cmd->sglist_len = 4;
829                 break;
830 #if (SCSI_MAX_PHYS_SEGMENTS  > 128)
831         case 129 ... 256:
832                 cmd->sglist_len = 5;
833                 break;
834 #endif
835 #endif
836 #endif
837         default:
838                 return NULL;
839         }
840
841         sgp = scsi_sg_pools + cmd->sglist_len;
842         sgl = mempool_alloc(sgp->pool, gfp_mask);
843         return sgl;
844 }
845
846 static void scsi_free_sgtable(struct scatterlist *sgl, int index)
847 {
848         struct scsi_host_sg_pool *sgp;
849
850         BUG_ON(index >= SG_MEMPOOL_NR);
851
852         sgp = scsi_sg_pools + index;
853         mempool_free(sgl, sgp->pool);
854 }
855
856 /*
857  * Function:    scsi_release_buffers()
858  *
859  * Purpose:     Completion processing for block device I/O requests.
860  *
861  * Arguments:   cmd     - command that we are bailing.
862  *
863  * Lock status: Assumed that no lock is held upon entry.
864  *
865  * Returns:     Nothing
866  *
867  * Notes:       In the event that an upper level driver rejects a
868  *              command, we must release resources allocated during
869  *              the __init_io() function.  Primarily this would involve
870  *              the scatter-gather table, and potentially any bounce
871  *              buffers.
872  */
873 static void scsi_release_buffers(struct scsi_cmnd *cmd)
874 {
875         struct request *req = cmd->request;
876
877         /*
878          * Free up any indirection buffers we allocated for DMA purposes. 
879          */
880         if (cmd->use_sg)
881                 scsi_free_sgtable(cmd->request_buffer, cmd->sglist_len);
882         else if (cmd->request_buffer != req->buffer)
883                 kfree(cmd->request_buffer);
884
885         /*
886          * Zero these out.  They now point to freed memory, and it is
887          * dangerous to hang onto the pointers.
888          */
889         cmd->buffer  = NULL;
890         cmd->bufflen = 0;
891         cmd->request_buffer = NULL;
892         cmd->request_bufflen = 0;
893 }
894
895 /*
896  * Function:    scsi_io_completion()
897  *
898  * Purpose:     Completion processing for block device I/O requests.
899  *
900  * Arguments:   cmd   - command that is finished.
901  *
902  * Lock status: Assumed that no lock is held upon entry.
903  *
904  * Returns:     Nothing
905  *
906  * Notes:       This function is matched in terms of capabilities to
907  *              the function that created the scatter-gather list.
908  *              In other words, if there are no bounce buffers
909  *              (the normal case for most drivers), we don't need
910  *              the logic to deal with cleaning up afterwards.
911  *
912  *              We must do one of several things here:
913  *
914  *              a) Call scsi_end_request.  This will finish off the
915  *                 specified number of sectors.  If we are done, the
916  *                 command block will be released, and the queue
917  *                 function will be goosed.  If we are not done, then
918  *                 scsi_end_request will directly goose the queue.
919  *
920  *              b) We can just use scsi_requeue_command() here.  This would
921  *                 be used if we just wanted to retry, for example.
922  */
923 void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes,
924                         unsigned int block_bytes)
925 {
926         int result = cmd->result;
927         int this_count = cmd->bufflen;
928         request_queue_t *q = cmd->device->request_queue;
929         struct request *req = cmd->request;
930         int clear_errors = 1;
931         struct scsi_sense_hdr sshdr;
932         int sense_valid = 0;
933         int sense_deferred = 0;
934
935         if (blk_complete_barrier_rq(q, req, good_bytes >> 9))
936                 return;
937
938         /*
939          * Free up any indirection buffers we allocated for DMA purposes. 
940          * For the case of a READ, we need to copy the data out of the
941          * bounce buffer and into the real buffer.
942          */
943         if (cmd->use_sg)
944                 scsi_free_sgtable(cmd->buffer, cmd->sglist_len);
945         else if (cmd->buffer != req->buffer) {
946                 if (rq_data_dir(req) == READ) {
947                         unsigned long flags;
948                         char *to = bio_kmap_irq(req->bio, &flags);
949                         memcpy(to, cmd->buffer, cmd->bufflen);
950                         bio_kunmap_irq(to, &flags);
951                 }
952                 kfree(cmd->buffer);
953         }
954
955         if (result) {
956                 sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
957                 if (sense_valid)
958                         sense_deferred = scsi_sense_is_deferred(&sshdr);
959         }
960         if (blk_pc_request(req)) { /* SG_IO ioctl from block level */
961                 req->errors = result;
962                 if (result) {
963                         clear_errors = 0;
964                         if (sense_valid && req->sense) {
965                                 /*
966                                  * SG_IO wants current and deferred errors
967                                  */
968                                 int len = 8 + cmd->sense_buffer[7];
969
970                                 if (len > SCSI_SENSE_BUFFERSIZE)
971                                         len = SCSI_SENSE_BUFFERSIZE;
972                                 memcpy(req->sense, cmd->sense_buffer,  len);
973                                 req->sense_len = len;
974                         }
975                 } else
976                         req->data_len = cmd->resid;
977         }
978
979         /*
980          * Zero these out.  They now point to freed memory, and it is
981          * dangerous to hang onto the pointers.
982          */
983         cmd->buffer  = NULL;
984         cmd->bufflen = 0;
985         cmd->request_buffer = NULL;
986         cmd->request_bufflen = 0;
987
988         /*
989          * Next deal with any sectors which we were able to correctly
990          * handle.
991          */
992         if (good_bytes >= 0) {
993                 SCSI_LOG_HLCOMPLETE(1, printk("%ld sectors total, %d bytes done.\n",
994                                               req->nr_sectors, good_bytes));
995                 SCSI_LOG_HLCOMPLETE(1, printk("use_sg is %d\n", cmd->use_sg));
996
997                 if (clear_errors)
998                         req->errors = 0;
999                 /*
1000                  * If multiple sectors are requested in one buffer, then
1001                  * they will have been finished off by the first command.
1002                  * If not, then we have a multi-buffer command.
1003                  *
1004                  * If block_bytes != 0, it means we had a medium error
1005                  * of some sort, and that we want to mark some number of
1006                  * sectors as not uptodate.  Thus we want to inhibit
1007                  * requeueing right here - we will requeue down below
1008                  * when we handle the bad sectors.
1009                  */
1010
1011                 /*
1012                  * If the command completed without error, then either
1013                  * finish off the rest of the command, or start a new one.
1014                  */
1015                 if (scsi_end_request(cmd, 1, good_bytes, result == 0) == NULL)
1016                         return;
1017         }
1018         /*
1019          * Now, if we were good little boys and girls, Santa left us a request
1020          * sense buffer.  We can extract information from this, so we
1021          * can choose a block to remap, etc.
1022          */
1023         if (sense_valid && !sense_deferred) {
1024                 switch (sshdr.sense_key) {
1025                 case UNIT_ATTENTION:
1026                         if (cmd->device->removable) {
1027                                 /* detected disc change.  set a bit 
1028                                  * and quietly refuse further access.
1029                                  */
1030                                 cmd->device->changed = 1;
1031                                 scsi_end_request(cmd, 0,
1032                                                 this_count, 1);
1033                                 return;
1034                         } else {
1035                                 /*
1036                                 * Must have been a power glitch, or a
1037                                 * bus reset.  Could not have been a
1038                                 * media change, so we just retry the
1039                                 * request and see what happens.  
1040                                 */
1041                                 scsi_requeue_command(q, cmd);
1042                                 return;
1043                         }
1044                         break;
1045                 case ILLEGAL_REQUEST:
1046                         /*
1047                         * If we had an ILLEGAL REQUEST returned, then we may
1048                         * have performed an unsupported command.  The only
1049                         * thing this should be would be a ten byte read where
1050                         * only a six byte read was supported.  Also, on a
1051                         * system where READ CAPACITY failed, we may have read
1052                         * past the end of the disk.
1053                         */
1054                         if ((cmd->device->use_10_for_rw &&
1055                             sshdr.asc == 0x20 && sshdr.ascq == 0x00) &&
1056                             (cmd->cmnd[0] == READ_10 ||
1057                              cmd->cmnd[0] == WRITE_10)) {
1058                                 cmd->device->use_10_for_rw = 0;
1059                                 /*
1060                                  * This will cause a retry with a 6-byte
1061                                  * command.
1062                                  */
1063                                 scsi_requeue_command(q, cmd);
1064                                 result = 0;
1065                         } else {
1066                                 scsi_end_request(cmd, 0, this_count, 1);
1067                                 return;
1068                         }
1069                         break;
1070                 case NOT_READY:
1071                         /*
1072                          * If the device is in the process of becoming ready,
1073                          * retry.
1074                          */
1075                         if (sshdr.asc == 0x04 && sshdr.ascq == 0x01) {
1076                                 scsi_requeue_command(q, cmd);
1077                                 return;
1078                         }
1079                         if (!(req->flags & REQ_QUIET))
1080                                 scmd_printk(KERN_INFO, cmd,
1081                                            "Device not ready.\n");
1082                         scsi_end_request(cmd, 0, this_count, 1);
1083                         return;
1084                 case VOLUME_OVERFLOW:
1085                         if (!(req->flags & REQ_QUIET)) {
1086                                 scmd_printk(KERN_INFO, cmd,
1087                                            "Volume overflow, CDB: ");
1088                                 __scsi_print_command(cmd->data_cmnd);
1089                                 scsi_print_sense("", cmd);
1090                         }
1091                         scsi_end_request(cmd, 0, block_bytes, 1);
1092                         return;
1093                 default:
1094                         break;
1095                 }
1096         }                       /* driver byte != 0 */
1097         if (host_byte(result) == DID_RESET) {
1098                 /*
1099                  * Third party bus reset or reset for error
1100                  * recovery reasons.  Just retry the request
1101                  * and see what happens.  
1102                  */
1103                 scsi_requeue_command(q, cmd);
1104                 return;
1105         }
1106         if (result) {
1107                 if (!(req->flags & REQ_QUIET)) {
1108                         scmd_printk(KERN_INFO, cmd,
1109                                    "SCSI error: return code = 0x%x\n", result);
1110
1111                         if (driver_byte(result) & DRIVER_SENSE)
1112                                 scsi_print_sense("", cmd);
1113                 }
1114                 /*
1115                  * Mark a single buffer as not uptodate.  Queue the remainder.
1116                  * We sometimes get this cruft in the event that a medium error
1117                  * isn't properly reported.
1118                  */
1119                 block_bytes = req->hard_cur_sectors << 9;
1120                 if (!block_bytes)
1121                         block_bytes = req->data_len;
1122                 scsi_end_request(cmd, 0, block_bytes, 1);
1123         }
1124 }
1125 EXPORT_SYMBOL(scsi_io_completion);
1126
1127 /*
1128  * Function:    scsi_init_io()
1129  *
1130  * Purpose:     SCSI I/O initialize function.
1131  *
1132  * Arguments:   cmd   - Command descriptor we wish to initialize
1133  *
1134  * Returns:     0 on success
1135  *              BLKPREP_DEFER if the failure is retryable
1136  *              BLKPREP_KILL if the failure is fatal
1137  */
1138 static int scsi_init_io(struct scsi_cmnd *cmd)
1139 {
1140         struct request     *req = cmd->request;
1141         struct scatterlist *sgpnt;
1142         int                count;
1143
1144         /*
1145          * if this is a rq->data based REQ_BLOCK_PC, setup for a non-sg xfer
1146          */
1147         if ((req->flags & REQ_BLOCK_PC) && !req->bio) {
1148                 cmd->request_bufflen = req->data_len;
1149                 cmd->request_buffer = req->data;
1150                 req->buffer = req->data;
1151                 cmd->use_sg = 0;
1152                 return 0;
1153         }
1154
1155         /*
1156          * we used to not use scatter-gather for single segment request,
1157          * but now we do (it makes highmem I/O easier to support without
1158          * kmapping pages)
1159          */
1160         cmd->use_sg = req->nr_phys_segments;
1161
1162         /*
1163          * if sg table allocation fails, requeue request later.
1164          */
1165         sgpnt = scsi_alloc_sgtable(cmd, GFP_ATOMIC);
1166         if (unlikely(!sgpnt)) {
1167                 scsi_unprep_request(req);
1168                 return BLKPREP_DEFER;
1169         }
1170
1171         cmd->request_buffer = (char *) sgpnt;
1172         cmd->request_bufflen = req->nr_sectors << 9;
1173         if (blk_pc_request(req))
1174                 cmd->request_bufflen = req->data_len;
1175         req->buffer = NULL;
1176
1177         /* 
1178          * Next, walk the list, and fill in the addresses and sizes of
1179          * each segment.
1180          */
1181         count = blk_rq_map_sg(req->q, req, cmd->request_buffer);
1182
1183         /*
1184          * mapped well, send it off
1185          */
1186         if (likely(count <= cmd->use_sg)) {
1187                 cmd->use_sg = count;
1188                 return 0;
1189         }
1190
1191         printk(KERN_ERR "Incorrect number of segments after building list\n");
1192         printk(KERN_ERR "counted %d, received %d\n", count, cmd->use_sg);
1193         printk(KERN_ERR "req nr_sec %lu, cur_nr_sec %u\n", req->nr_sectors,
1194                         req->current_nr_sectors);
1195
1196         /* release the command and kill it */
1197         scsi_release_buffers(cmd);
1198         scsi_put_command(cmd);
1199         return BLKPREP_KILL;
1200 }
1201
1202 static int scsi_prepare_flush_fn(request_queue_t *q, struct request *rq)
1203 {
1204         struct scsi_device *sdev = q->queuedata;
1205         struct scsi_driver *drv;
1206
1207         if (sdev->sdev_state == SDEV_RUNNING) {
1208                 drv = *(struct scsi_driver **) rq->rq_disk->private_data;
1209
1210                 if (drv->prepare_flush)
1211                         return drv->prepare_flush(q, rq);
1212         }
1213
1214         return 0;
1215 }
1216
1217 static void scsi_end_flush_fn(request_queue_t *q, struct request *rq)
1218 {
1219         struct scsi_device *sdev = q->queuedata;
1220         struct request *flush_rq = rq->end_io_data;
1221         struct scsi_driver *drv;
1222
1223         if (flush_rq->errors) {
1224                 printk("scsi: barrier error, disabling flush support\n");
1225                 blk_queue_ordered(q, QUEUE_ORDERED_NONE);
1226         }
1227
1228         if (sdev->sdev_state == SDEV_RUNNING) {
1229                 drv = *(struct scsi_driver **) rq->rq_disk->private_data;
1230                 drv->end_flush(q, rq);
1231         }
1232 }
1233
1234 static int scsi_issue_flush_fn(request_queue_t *q, struct gendisk *disk,
1235                                sector_t *error_sector)
1236 {
1237         struct scsi_device *sdev = q->queuedata;
1238         struct scsi_driver *drv;
1239
1240         if (sdev->sdev_state != SDEV_RUNNING)
1241                 return -ENXIO;
1242
1243         drv = *(struct scsi_driver **) disk->private_data;
1244         if (drv->issue_flush)
1245                 return drv->issue_flush(&sdev->sdev_gendev, error_sector);
1246
1247         return -EOPNOTSUPP;
1248 }
1249
1250 static void scsi_generic_done(struct scsi_cmnd *cmd)
1251 {
1252         BUG_ON(!blk_pc_request(cmd->request));
1253         /*
1254          * This will complete the whole command with uptodate=1 so
1255          * as far as the block layer is concerned the command completed
1256          * successfully. Since this is a REQ_BLOCK_PC command the
1257          * caller should check the request's errors value
1258          */
1259         scsi_io_completion(cmd, cmd->bufflen, 0);
1260 }
1261
1262 void scsi_setup_blk_pc_cmnd(struct scsi_cmnd *cmd)
1263 {
1264         struct request *req = cmd->request;
1265
1266         BUG_ON(sizeof(req->cmd) > sizeof(cmd->cmnd));
1267         memcpy(cmd->cmnd, req->cmd, sizeof(cmd->cmnd));
1268         cmd->cmd_len = req->cmd_len;
1269         if (!req->data_len)
1270                 cmd->sc_data_direction = DMA_NONE;
1271         else if (rq_data_dir(req) == WRITE)
1272                 cmd->sc_data_direction = DMA_TO_DEVICE;
1273         else
1274                 cmd->sc_data_direction = DMA_FROM_DEVICE;
1275         
1276         cmd->transfersize = req->data_len;
1277         cmd->allowed = req->retries;
1278         cmd->timeout_per_command = req->timeout;
1279 }
1280 EXPORT_SYMBOL_GPL(scsi_setup_blk_pc_cmnd);
1281
1282 static int scsi_prep_fn(struct request_queue *q, struct request *req)
1283 {
1284         struct scsi_device *sdev = q->queuedata;
1285         struct scsi_cmnd *cmd;
1286         int specials_only = 0;
1287
1288         /*
1289          * Just check to see if the device is online.  If it isn't, we
1290          * refuse to process any commands.  The device must be brought
1291          * online before trying any recovery commands
1292          */
1293         if (unlikely(!scsi_device_online(sdev))) {
1294                 sdev_printk(KERN_ERR, sdev,
1295                             "rejecting I/O to offline device\n");
1296                 goto kill;
1297         }
1298         if (unlikely(sdev->sdev_state != SDEV_RUNNING)) {
1299                 /* OK, we're not in a running state don't prep
1300                  * user commands */
1301                 if (sdev->sdev_state == SDEV_DEL) {
1302                         /* Device is fully deleted, no commands
1303                          * at all allowed down */
1304                         sdev_printk(KERN_ERR, sdev,
1305                                     "rejecting I/O to dead device\n");
1306                         goto kill;
1307                 }
1308                 /* OK, we only allow special commands (i.e. not
1309                  * user initiated ones */
1310                 specials_only = sdev->sdev_state;
1311         }
1312
1313         /*
1314          * Find the actual device driver associated with this command.
1315          * The SPECIAL requests are things like character device or
1316          * ioctls, which did not originate from ll_rw_blk.  Note that
1317          * the special field is also used to indicate the cmd for
1318          * the remainder of a partially fulfilled request that can 
1319          * come up when there is a medium error.  We have to treat
1320          * these two cases differently.  We differentiate by looking
1321          * at request->cmd, as this tells us the real story.
1322          */
1323         if (req->flags & REQ_SPECIAL && req->special) {
1324                 struct scsi_request *sreq = req->special;
1325
1326                 if (sreq->sr_magic == SCSI_REQ_MAGIC) {
1327                         cmd = scsi_get_command(sreq->sr_device, GFP_ATOMIC);
1328                         if (unlikely(!cmd))
1329                                 goto defer;
1330                         scsi_init_cmd_from_req(cmd, sreq);
1331                 } else
1332                         cmd = req->special;
1333         } else if (req->flags & (REQ_CMD | REQ_BLOCK_PC)) {
1334
1335                 if(unlikely(specials_only) && !(req->flags & REQ_SPECIAL)) {
1336                         if(specials_only == SDEV_QUIESCE ||
1337                                         specials_only == SDEV_BLOCK)
1338                                 goto defer;
1339                         
1340                         sdev_printk(KERN_ERR, sdev,
1341                                     "rejecting I/O to device being removed\n");
1342                         goto kill;
1343                 }
1344                         
1345                         
1346                 /*
1347                  * Now try and find a command block that we can use.
1348                  */
1349                 if (!req->special) {
1350                         cmd = scsi_get_command(sdev, GFP_ATOMIC);
1351                         if (unlikely(!cmd))
1352                                 goto defer;
1353                 } else
1354                         cmd = req->special;
1355                 
1356                 /* pull a tag out of the request if we have one */
1357                 cmd->tag = req->tag;
1358         } else {
1359                 blk_dump_rq_flags(req, "SCSI bad req");
1360                 goto kill;
1361         }
1362         
1363         /* note the overloading of req->special.  When the tag
1364          * is active it always means cmd.  If the tag goes
1365          * back for re-queueing, it may be reset */
1366         req->special = cmd;
1367         cmd->request = req;
1368         
1369         /*
1370          * FIXME: drop the lock here because the functions below
1371          * expect to be called without the queue lock held.  Also,
1372          * previously, we dequeued the request before dropping the
1373          * lock.  We hope REQ_STARTED prevents anything untoward from
1374          * happening now.
1375          */
1376         if (req->flags & (REQ_CMD | REQ_BLOCK_PC)) {
1377                 struct scsi_driver *drv;
1378                 int ret;
1379
1380                 /*
1381                  * This will do a couple of things:
1382                  *  1) Fill in the actual SCSI command.
1383                  *  2) Fill in any other upper-level specific fields
1384                  * (timeout).
1385                  *
1386                  * If this returns 0, it means that the request failed
1387                  * (reading past end of disk, reading offline device,
1388                  * etc).   This won't actually talk to the device, but
1389                  * some kinds of consistency checking may cause the     
1390                  * request to be rejected immediately.
1391                  */
1392
1393                 /* 
1394                  * This sets up the scatter-gather table (allocating if
1395                  * required).
1396                  */
1397                 ret = scsi_init_io(cmd);
1398                 switch(ret) {
1399                         /* For BLKPREP_KILL/DEFER the cmd was released */
1400                 case BLKPREP_KILL:
1401                         goto kill;
1402                 case BLKPREP_DEFER:
1403                         goto defer;
1404                 }
1405                 
1406                 /*
1407                  * Initialize the actual SCSI command for this request.
1408                  */
1409                 if (req->rq_disk) {
1410                         drv = *(struct scsi_driver **)req->rq_disk->private_data;
1411                         if (unlikely(!drv->init_command(cmd))) {
1412                                 scsi_release_buffers(cmd);
1413                                 scsi_put_command(cmd);
1414                                 goto kill;
1415                         }
1416                 } else {
1417                         scsi_setup_blk_pc_cmnd(cmd);
1418                         cmd->done = scsi_generic_done;
1419                 }
1420         }
1421
1422         /*
1423          * The request is now prepped, no need to come back here
1424          */
1425         req->flags |= REQ_DONTPREP;
1426         return BLKPREP_OK;
1427
1428  defer:
1429         /* If we defer, the elv_next_request() returns NULL, but the
1430          * queue must be restarted, so we plug here if no returning
1431          * command will automatically do that. */
1432         if (sdev->device_busy == 0)
1433                 blk_plug_device(q);
1434         return BLKPREP_DEFER;
1435  kill:
1436         req->errors = DID_NO_CONNECT << 16;
1437         return BLKPREP_KILL;
1438 }
1439
1440 /*
1441  * scsi_dev_queue_ready: if we can send requests to sdev, return 1 else
1442  * return 0.
1443  *
1444  * Called with the queue_lock held.
1445  */
1446 static inline int scsi_dev_queue_ready(struct request_queue *q,
1447                                   struct scsi_device *sdev)
1448 {
1449         if (sdev->device_busy >= sdev->queue_depth)
1450                 return 0;
1451         if (sdev->device_busy == 0 && sdev->device_blocked) {
1452                 /*
1453                  * unblock after device_blocked iterates to zero
1454                  */
1455                 if (--sdev->device_blocked == 0) {
1456                         SCSI_LOG_MLQUEUE(3,
1457                                    sdev_printk(KERN_INFO, sdev,
1458                                    "unblocking device at zero depth\n"));
1459                 } else {
1460                         blk_plug_device(q);
1461                         return 0;
1462                 }
1463         }
1464         if (sdev->device_blocked)
1465                 return 0;
1466
1467         return 1;
1468 }
1469
1470 /*
1471  * scsi_host_queue_ready: if we can send requests to shost, return 1 else
1472  * return 0. We must end up running the queue again whenever 0 is
1473  * returned, else IO can hang.
1474  *
1475  * Called with host_lock held.
1476  */
1477 static inline int scsi_host_queue_ready(struct request_queue *q,
1478                                    struct Scsi_Host *shost,
1479                                    struct scsi_device *sdev)
1480 {
1481         if (scsi_host_in_recovery(shost))
1482                 return 0;
1483         if (shost->host_busy == 0 && shost->host_blocked) {
1484                 /*
1485                  * unblock after host_blocked iterates to zero
1486                  */
1487                 if (--shost->host_blocked == 0) {
1488                         SCSI_LOG_MLQUEUE(3,
1489                                 printk("scsi%d unblocking host at zero depth\n",
1490                                         shost->host_no));
1491                 } else {
1492                         blk_plug_device(q);
1493                         return 0;
1494                 }
1495         }
1496         if ((shost->can_queue > 0 && shost->host_busy >= shost->can_queue) ||
1497             shost->host_blocked || shost->host_self_blocked) {
1498                 if (list_empty(&sdev->starved_entry))
1499                         list_add_tail(&sdev->starved_entry, &shost->starved_list);
1500                 return 0;
1501         }
1502
1503         /* We're OK to process the command, so we can't be starved */
1504         if (!list_empty(&sdev->starved_entry))
1505                 list_del_init(&sdev->starved_entry);
1506
1507         return 1;
1508 }
1509
1510 /*
1511  * Kill a request for a dead device
1512  */
1513 static void scsi_kill_request(struct request *req, request_queue_t *q)
1514 {
1515         struct scsi_cmnd *cmd = req->special;
1516
1517         blkdev_dequeue_request(req);
1518
1519         if (unlikely(cmd == NULL)) {
1520                 printk(KERN_CRIT "impossible request in %s.\n",
1521                                  __FUNCTION__);
1522                 BUG();
1523         }
1524
1525         scsi_init_cmd_errh(cmd);
1526         cmd->result = DID_NO_CONNECT << 16;
1527         atomic_inc(&cmd->device->iorequest_cnt);
1528         __scsi_done(cmd);
1529 }
1530
1531 /*
1532  * Function:    scsi_request_fn()
1533  *
1534  * Purpose:     Main strategy routine for SCSI.
1535  *
1536  * Arguments:   q       - Pointer to actual queue.
1537  *
1538  * Returns:     Nothing
1539  *
1540  * Lock status: IO request lock assumed to be held when called.
1541  */
1542 static void scsi_request_fn(struct request_queue *q)
1543 {
1544         struct scsi_device *sdev = q->queuedata;
1545         struct Scsi_Host *shost;
1546         struct scsi_cmnd *cmd;
1547         struct request *req;
1548
1549         if (!sdev) {
1550                 printk("scsi: killing requests for dead queue\n");
1551                 while ((req = elv_next_request(q)) != NULL)
1552                         scsi_kill_request(req, q);
1553                 return;
1554         }
1555
1556         if(!get_device(&sdev->sdev_gendev))
1557                 /* We must be tearing the block queue down already */
1558                 return;
1559
1560         /*
1561          * To start with, we keep looping until the queue is empty, or until
1562          * the host is no longer able to accept any more requests.
1563          */
1564         shost = sdev->host;
1565         while (!blk_queue_plugged(q)) {
1566                 int rtn;
1567                 /*
1568                  * get next queueable request.  We do this early to make sure
1569                  * that the request is fully prepared even if we cannot 
1570                  * accept it.
1571                  */
1572                 req = elv_next_request(q);
1573                 if (!req || !scsi_dev_queue_ready(q, sdev))
1574                         break;
1575
1576                 if (unlikely(!scsi_device_online(sdev))) {
1577                         sdev_printk(KERN_ERR, sdev,
1578                                     "rejecting I/O to offline device\n");
1579                         scsi_kill_request(req, q);
1580                         continue;
1581                 }
1582
1583
1584                 /*
1585                  * Remove the request from the request list.
1586                  */
1587                 if (!(blk_queue_tagged(q) && !blk_queue_start_tag(q, req)))
1588                         blkdev_dequeue_request(req);
1589                 sdev->device_busy++;
1590
1591                 spin_unlock(q->queue_lock);
1592                 cmd = req->special;
1593                 if (unlikely(cmd == NULL)) {
1594                         printk(KERN_CRIT "impossible request in %s.\n"
1595                                          "please mail a stack trace to "
1596                                          "linux-scsi@vger.kernel.org",
1597                                          __FUNCTION__);
1598                         BUG();
1599                 }
1600                 spin_lock(shost->host_lock);
1601
1602                 if (!scsi_host_queue_ready(q, shost, sdev))
1603                         goto not_ready;
1604                 if (sdev->single_lun) {
1605                         if (scsi_target(sdev)->starget_sdev_user &&
1606                             scsi_target(sdev)->starget_sdev_user != sdev)
1607                                 goto not_ready;
1608                         scsi_target(sdev)->starget_sdev_user = sdev;
1609                 }
1610                 shost->host_busy++;
1611
1612                 /*
1613                  * XXX(hch): This is rather suboptimal, scsi_dispatch_cmd will
1614                  *              take the lock again.
1615                  */
1616                 spin_unlock_irq(shost->host_lock);
1617
1618                 /*
1619                  * Finally, initialize any error handling parameters, and set up
1620                  * the timers for timeouts.
1621                  */
1622                 scsi_init_cmd_errh(cmd);
1623
1624                 /*
1625                  * Dispatch the command to the low-level driver.
1626                  */
1627                 rtn = scsi_dispatch_cmd(cmd);
1628                 spin_lock_irq(q->queue_lock);
1629                 if(rtn) {
1630                         /* we're refusing the command; because of
1631                          * the way locks get dropped, we need to 
1632                          * check here if plugging is required */
1633                         if(sdev->device_busy == 0)
1634                                 blk_plug_device(q);
1635
1636                         break;
1637                 }
1638         }
1639
1640         goto out;
1641
1642  not_ready:
1643         spin_unlock_irq(shost->host_lock);
1644
1645         /*
1646          * lock q, handle tag, requeue req, and decrement device_busy. We
1647          * must return with queue_lock held.
1648          *
1649          * Decrementing device_busy without checking it is OK, as all such
1650          * cases (host limits or settings) should run the queue at some
1651          * later time.
1652          */
1653         spin_lock_irq(q->queue_lock);
1654         blk_requeue_request(q, req);
1655         sdev->device_busy--;
1656         if(sdev->device_busy == 0)
1657                 blk_plug_device(q);
1658  out:
1659         /* must be careful here...if we trigger the ->remove() function
1660          * we cannot be holding the q lock */
1661         spin_unlock_irq(q->queue_lock);
1662         put_device(&sdev->sdev_gendev);
1663         spin_lock_irq(q->queue_lock);
1664 }
1665
1666 u64 scsi_calculate_bounce_limit(struct Scsi_Host *shost)
1667 {
1668         struct device *host_dev;
1669         u64 bounce_limit = 0xffffffff;
1670
1671         if (shost->unchecked_isa_dma)
1672                 return BLK_BOUNCE_ISA;
1673         /*
1674          * Platforms with virtual-DMA translation
1675          * hardware have no practical limit.
1676          */
1677         if (!PCI_DMA_BUS_IS_PHYS)
1678                 return BLK_BOUNCE_ANY;
1679
1680         host_dev = scsi_get_device(shost);
1681         if (host_dev && host_dev->dma_mask)
1682                 bounce_limit = *host_dev->dma_mask;
1683
1684         return bounce_limit;
1685 }
1686 EXPORT_SYMBOL(scsi_calculate_bounce_limit);
1687
1688 struct request_queue *scsi_alloc_queue(struct scsi_device *sdev)
1689 {
1690         struct Scsi_Host *shost = sdev->host;
1691         struct request_queue *q;
1692
1693         q = blk_init_queue(scsi_request_fn, NULL);
1694         if (!q)
1695                 return NULL;
1696
1697         blk_queue_prep_rq(q, scsi_prep_fn);
1698
1699         blk_queue_max_hw_segments(q, shost->sg_tablesize);
1700         blk_queue_max_phys_segments(q, SCSI_MAX_PHYS_SEGMENTS);
1701         blk_queue_max_sectors(q, shost->max_sectors);
1702         blk_queue_bounce_limit(q, scsi_calculate_bounce_limit(shost));
1703         blk_queue_segment_boundary(q, shost->dma_boundary);
1704         blk_queue_issue_flush_fn(q, scsi_issue_flush_fn);
1705
1706         /*
1707          * ordered tags are superior to flush ordering
1708          */
1709         if (shost->ordered_tag)
1710                 blk_queue_ordered(q, QUEUE_ORDERED_TAG);
1711         else if (shost->ordered_flush) {
1712                 blk_queue_ordered(q, QUEUE_ORDERED_FLUSH);
1713                 q->prepare_flush_fn = scsi_prepare_flush_fn;
1714                 q->end_flush_fn = scsi_end_flush_fn;
1715         }
1716
1717         if (!shost->use_clustering)
1718                 clear_bit(QUEUE_FLAG_CLUSTER, &q->queue_flags);
1719         return q;
1720 }
1721
1722 void scsi_free_queue(struct request_queue *q)
1723 {
1724         blk_cleanup_queue(q);
1725 }
1726
1727 /*
1728  * Function:    scsi_block_requests()
1729  *
1730  * Purpose:     Utility function used by low-level drivers to prevent further
1731  *              commands from being queued to the device.
1732  *
1733  * Arguments:   shost       - Host in question
1734  *
1735  * Returns:     Nothing
1736  *
1737  * Lock status: No locks are assumed held.
1738  *
1739  * Notes:       There is no timer nor any other means by which the requests
1740  *              get unblocked other than the low-level driver calling
1741  *              scsi_unblock_requests().
1742  */
1743 void scsi_block_requests(struct Scsi_Host *shost)
1744 {
1745         shost->host_self_blocked = 1;
1746 }
1747 EXPORT_SYMBOL(scsi_block_requests);
1748
1749 /*
1750  * Function:    scsi_unblock_requests()
1751  *
1752  * Purpose:     Utility function used by low-level drivers to allow further
1753  *              commands from being queued to the device.
1754  *
1755  * Arguments:   shost       - Host in question
1756  *
1757  * Returns:     Nothing
1758  *
1759  * Lock status: No locks are assumed held.
1760  *
1761  * Notes:       There is no timer nor any other means by which the requests
1762  *              get unblocked other than the low-level driver calling
1763  *              scsi_unblock_requests().
1764  *
1765  *              This is done as an API function so that changes to the
1766  *              internals of the scsi mid-layer won't require wholesale
1767  *              changes to drivers that use this feature.
1768  */
1769 void scsi_unblock_requests(struct Scsi_Host *shost)
1770 {
1771         shost->host_self_blocked = 0;
1772         scsi_run_host_queues(shost);
1773 }
1774 EXPORT_SYMBOL(scsi_unblock_requests);
1775
1776 int __init scsi_init_queue(void)
1777 {
1778         int i;
1779
1780         scsi_io_context_cache = kmem_cache_create("scsi_io_context",
1781                                         sizeof(struct scsi_io_context),
1782                                         0, 0, NULL, NULL);
1783         if (!scsi_io_context_cache) {
1784                 printk(KERN_ERR "SCSI: can't init scsi io context cache\n");
1785                 return -ENOMEM;
1786         }
1787
1788         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1789                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1790                 int size = sgp->size * sizeof(struct scatterlist);
1791
1792                 sgp->slab = kmem_cache_create(sgp->name, size, 0,
1793                                 SLAB_HWCACHE_ALIGN, NULL, NULL);
1794                 if (!sgp->slab) {
1795                         printk(KERN_ERR "SCSI: can't init sg slab %s\n",
1796                                         sgp->name);
1797                 }
1798
1799                 sgp->pool = mempool_create(SG_MEMPOOL_SIZE,
1800                                 mempool_alloc_slab, mempool_free_slab,
1801                                 sgp->slab);
1802                 if (!sgp->pool) {
1803                         printk(KERN_ERR "SCSI: can't init sg mempool %s\n",
1804                                         sgp->name);
1805                 }
1806         }
1807
1808         return 0;
1809 }
1810
1811 void scsi_exit_queue(void)
1812 {
1813         int i;
1814
1815         kmem_cache_destroy(scsi_io_context_cache);
1816
1817         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1818                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1819                 mempool_destroy(sgp->pool);
1820                 kmem_cache_destroy(sgp->slab);
1821         }
1822 }
1823 /**
1824  *      scsi_mode_sense - issue a mode sense, falling back from 10 to 
1825  *              six bytes if necessary.
1826  *      @sdev:  SCSI device to be queried
1827  *      @dbd:   set if mode sense will allow block descriptors to be returned
1828  *      @modepage: mode page being requested
1829  *      @buffer: request buffer (may not be smaller than eight bytes)
1830  *      @len:   length of request buffer.
1831  *      @timeout: command timeout
1832  *      @retries: number of retries before failing
1833  *      @data: returns a structure abstracting the mode header data
1834  *      @sense: place to put sense data (or NULL if no sense to be collected).
1835  *              must be SCSI_SENSE_BUFFERSIZE big.
1836  *
1837  *      Returns zero if unsuccessful, or the header offset (either 4
1838  *      or 8 depending on whether a six or ten byte command was
1839  *      issued) if successful.
1840  **/
1841 int
1842 scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
1843                   unsigned char *buffer, int len, int timeout, int retries,
1844                   struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr) {
1845         unsigned char cmd[12];
1846         int use_10_for_ms;
1847         int header_length;
1848         int result;
1849         struct scsi_sense_hdr my_sshdr;
1850
1851         memset(data, 0, sizeof(*data));
1852         memset(&cmd[0], 0, 12);
1853         cmd[1] = dbd & 0x18;    /* allows DBD and LLBA bits */
1854         cmd[2] = modepage;
1855
1856         /* caller might not be interested in sense, but we need it */
1857         if (!sshdr)
1858                 sshdr = &my_sshdr;
1859
1860  retry:
1861         use_10_for_ms = sdev->use_10_for_ms;
1862
1863         if (use_10_for_ms) {
1864                 if (len < 8)
1865                         len = 8;
1866
1867                 cmd[0] = MODE_SENSE_10;
1868                 cmd[8] = len;
1869                 header_length = 8;
1870         } else {
1871                 if (len < 4)
1872                         len = 4;
1873
1874                 cmd[0] = MODE_SENSE;
1875                 cmd[4] = len;
1876                 header_length = 4;
1877         }
1878
1879         memset(buffer, 0, len);
1880
1881         result = scsi_execute_req(sdev, cmd, DMA_FROM_DEVICE, buffer, len,
1882                                   sshdr, timeout, retries);
1883
1884         /* This code looks awful: what it's doing is making sure an
1885          * ILLEGAL REQUEST sense return identifies the actual command
1886          * byte as the problem.  MODE_SENSE commands can return
1887          * ILLEGAL REQUEST if the code page isn't supported */
1888
1889         if (use_10_for_ms && !scsi_status_is_good(result) &&
1890             (driver_byte(result) & DRIVER_SENSE)) {
1891                 if (scsi_sense_valid(sshdr)) {
1892                         if ((sshdr->sense_key == ILLEGAL_REQUEST) &&
1893                             (sshdr->asc == 0x20) && (sshdr->ascq == 0)) {
1894                                 /* 
1895                                  * Invalid command operation code
1896                                  */
1897                                 sdev->use_10_for_ms = 0;
1898                                 goto retry;
1899                         }
1900                 }
1901         }
1902
1903         if(scsi_status_is_good(result)) {
1904                 data->header_length = header_length;
1905                 if(use_10_for_ms) {
1906                         data->length = buffer[0]*256 + buffer[1] + 2;
1907                         data->medium_type = buffer[2];
1908                         data->device_specific = buffer[3];
1909                         data->longlba = buffer[4] & 0x01;
1910                         data->block_descriptor_length = buffer[6]*256
1911                                 + buffer[7];
1912                 } else {
1913                         data->length = buffer[0] + 1;
1914                         data->medium_type = buffer[1];
1915                         data->device_specific = buffer[2];
1916                         data->block_descriptor_length = buffer[3];
1917                 }
1918         }
1919
1920         return result;
1921 }
1922 EXPORT_SYMBOL(scsi_mode_sense);
1923
1924 int
1925 scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries)
1926 {
1927         char cmd[] = {
1928                 TEST_UNIT_READY, 0, 0, 0, 0, 0,
1929         };
1930         struct scsi_sense_hdr sshdr;
1931         int result;
1932         
1933         result = scsi_execute_req(sdev, cmd, DMA_NONE, NULL, 0, &sshdr,
1934                                   timeout, retries);
1935
1936         if ((driver_byte(result) & DRIVER_SENSE) && sdev->removable) {
1937
1938                 if ((scsi_sense_valid(&sshdr)) &&
1939                     ((sshdr.sense_key == UNIT_ATTENTION) ||
1940                      (sshdr.sense_key == NOT_READY))) {
1941                         sdev->changed = 1;
1942                         result = 0;
1943                 }
1944         }
1945         return result;
1946 }
1947 EXPORT_SYMBOL(scsi_test_unit_ready);
1948
1949 /**
1950  *      scsi_device_set_state - Take the given device through the device
1951  *              state model.
1952  *      @sdev:  scsi device to change the state of.
1953  *      @state: state to change to.
1954  *
1955  *      Returns zero if unsuccessful or an error if the requested 
1956  *      transition is illegal.
1957  **/
1958 int
1959 scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
1960 {
1961         enum scsi_device_state oldstate = sdev->sdev_state;
1962
1963         if (state == oldstate)
1964                 return 0;
1965
1966         switch (state) {
1967         case SDEV_CREATED:
1968                 /* There are no legal states that come back to
1969                  * created.  This is the manually initialised start
1970                  * state */
1971                 goto illegal;
1972                         
1973         case SDEV_RUNNING:
1974                 switch (oldstate) {
1975                 case SDEV_CREATED:
1976                 case SDEV_OFFLINE:
1977                 case SDEV_QUIESCE:
1978                 case SDEV_BLOCK:
1979                         break;
1980                 default:
1981                         goto illegal;
1982                 }
1983                 break;
1984
1985         case SDEV_QUIESCE:
1986                 switch (oldstate) {
1987                 case SDEV_RUNNING:
1988                 case SDEV_OFFLINE:
1989                         break;
1990                 default:
1991                         goto illegal;
1992                 }
1993                 break;
1994
1995         case SDEV_OFFLINE:
1996                 switch (oldstate) {
1997                 case SDEV_CREATED:
1998                 case SDEV_RUNNING:
1999                 case SDEV_QUIESCE:
2000                 case SDEV_BLOCK:
2001                         break;
2002                 default:
2003                         goto illegal;
2004                 }
2005                 break;
2006
2007         case SDEV_BLOCK:
2008                 switch (oldstate) {
2009                 case SDEV_CREATED:
2010                 case SDEV_RUNNING:
2011                         break;
2012                 default:
2013                         goto illegal;
2014                 }
2015                 break;
2016
2017         case SDEV_CANCEL:
2018                 switch (oldstate) {
2019                 case SDEV_CREATED:
2020                 case SDEV_RUNNING:
2021                 case SDEV_OFFLINE:
2022                 case SDEV_BLOCK:
2023                         break;
2024                 default:
2025                         goto illegal;
2026                 }
2027                 break;
2028
2029         case SDEV_DEL:
2030                 switch (oldstate) {
2031                 case SDEV_CANCEL:
2032                         break;
2033                 default:
2034                         goto illegal;
2035                 }
2036                 break;
2037
2038         }
2039         sdev->sdev_state = state;
2040         return 0;
2041
2042  illegal:
2043         SCSI_LOG_ERROR_RECOVERY(1, 
2044                                 sdev_printk(KERN_ERR, sdev,
2045                                             "Illegal state transition %s->%s\n",
2046                                             scsi_device_state_name(oldstate),
2047                                             scsi_device_state_name(state))
2048                                 );
2049         return -EINVAL;
2050 }
2051 EXPORT_SYMBOL(scsi_device_set_state);
2052
2053 /**
2054  *      scsi_device_quiesce - Block user issued commands.
2055  *      @sdev:  scsi device to quiesce.
2056  *
2057  *      This works by trying to transition to the SDEV_QUIESCE state
2058  *      (which must be a legal transition).  When the device is in this
2059  *      state, only special requests will be accepted, all others will
2060  *      be deferred.  Since special requests may also be requeued requests,
2061  *      a successful return doesn't guarantee the device will be 
2062  *      totally quiescent.
2063  *
2064  *      Must be called with user context, may sleep.
2065  *
2066  *      Returns zero if unsuccessful or an error if not.
2067  **/
2068 int
2069 scsi_device_quiesce(struct scsi_device *sdev)
2070 {
2071         int err = scsi_device_set_state(sdev, SDEV_QUIESCE);
2072         if (err)
2073                 return err;
2074
2075         scsi_run_queue(sdev->request_queue);
2076         while (sdev->device_busy) {
2077                 msleep_interruptible(200);
2078                 scsi_run_queue(sdev->request_queue);
2079         }
2080         return 0;
2081 }
2082 EXPORT_SYMBOL(scsi_device_quiesce);
2083
2084 /**
2085  *      scsi_device_resume - Restart user issued commands to a quiesced device.
2086  *      @sdev:  scsi device to resume.
2087  *
2088  *      Moves the device from quiesced back to running and restarts the
2089  *      queues.
2090  *
2091  *      Must be called with user context, may sleep.
2092  **/
2093 void
2094 scsi_device_resume(struct scsi_device *sdev)
2095 {
2096         if(scsi_device_set_state(sdev, SDEV_RUNNING))
2097                 return;
2098         scsi_run_queue(sdev->request_queue);
2099 }
2100 EXPORT_SYMBOL(scsi_device_resume);
2101
2102 static void
2103 device_quiesce_fn(struct scsi_device *sdev, void *data)
2104 {
2105         scsi_device_quiesce(sdev);
2106 }
2107
2108 void
2109 scsi_target_quiesce(struct scsi_target *starget)
2110 {
2111         starget_for_each_device(starget, NULL, device_quiesce_fn);
2112 }
2113 EXPORT_SYMBOL(scsi_target_quiesce);
2114
2115 static void
2116 device_resume_fn(struct scsi_device *sdev, void *data)
2117 {
2118         scsi_device_resume(sdev);
2119 }
2120
2121 void
2122 scsi_target_resume(struct scsi_target *starget)
2123 {
2124         starget_for_each_device(starget, NULL, device_resume_fn);
2125 }
2126 EXPORT_SYMBOL(scsi_target_resume);
2127
2128 /**
2129  * scsi_internal_device_block - internal function to put a device
2130  *                              temporarily into the SDEV_BLOCK state
2131  * @sdev:       device to block
2132  *
2133  * Block request made by scsi lld's to temporarily stop all
2134  * scsi commands on the specified device.  Called from interrupt
2135  * or normal process context.
2136  *
2137  * Returns zero if successful or error if not
2138  *
2139  * Notes:       
2140  *      This routine transitions the device to the SDEV_BLOCK state
2141  *      (which must be a legal transition).  When the device is in this
2142  *      state, all commands are deferred until the scsi lld reenables
2143  *      the device with scsi_device_unblock or device_block_tmo fires.
2144  *      This routine assumes the host_lock is held on entry.
2145  **/
2146 int
2147 scsi_internal_device_block(struct scsi_device *sdev)
2148 {
2149         request_queue_t *q = sdev->request_queue;
2150         unsigned long flags;
2151         int err = 0;
2152
2153         err = scsi_device_set_state(sdev, SDEV_BLOCK);
2154         if (err)
2155                 return err;
2156
2157         /* 
2158          * The device has transitioned to SDEV_BLOCK.  Stop the
2159          * block layer from calling the midlayer with this device's
2160          * request queue. 
2161          */
2162         spin_lock_irqsave(q->queue_lock, flags);
2163         blk_stop_queue(q);
2164         spin_unlock_irqrestore(q->queue_lock, flags);
2165
2166         return 0;
2167 }
2168 EXPORT_SYMBOL_GPL(scsi_internal_device_block);
2169  
2170 /**
2171  * scsi_internal_device_unblock - resume a device after a block request
2172  * @sdev:       device to resume
2173  *
2174  * Called by scsi lld's or the midlayer to restart the device queue
2175  * for the previously suspended scsi device.  Called from interrupt or
2176  * normal process context.
2177  *
2178  * Returns zero if successful or error if not.
2179  *
2180  * Notes:       
2181  *      This routine transitions the device to the SDEV_RUNNING state
2182  *      (which must be a legal transition) allowing the midlayer to
2183  *      goose the queue for this device.  This routine assumes the 
2184  *      host_lock is held upon entry.
2185  **/
2186 int
2187 scsi_internal_device_unblock(struct scsi_device *sdev)
2188 {
2189         request_queue_t *q = sdev->request_queue; 
2190         int err;
2191         unsigned long flags;
2192         
2193         /* 
2194          * Try to transition the scsi device to SDEV_RUNNING
2195          * and goose the device queue if successful.  
2196          */
2197         err = scsi_device_set_state(sdev, SDEV_RUNNING);
2198         if (err)
2199                 return err;
2200
2201         spin_lock_irqsave(q->queue_lock, flags);
2202         blk_start_queue(q);
2203         spin_unlock_irqrestore(q->queue_lock, flags);
2204
2205         return 0;
2206 }
2207 EXPORT_SYMBOL_GPL(scsi_internal_device_unblock);
2208
2209 static void
2210 device_block(struct scsi_device *sdev, void *data)
2211 {
2212         scsi_internal_device_block(sdev);
2213 }
2214
2215 static int
2216 target_block(struct device *dev, void *data)
2217 {
2218         if (scsi_is_target_device(dev))
2219                 starget_for_each_device(to_scsi_target(dev), NULL,
2220                                         device_block);
2221         return 0;
2222 }
2223
2224 void
2225 scsi_target_block(struct device *dev)
2226 {
2227         if (scsi_is_target_device(dev))
2228                 starget_for_each_device(to_scsi_target(dev), NULL,
2229                                         device_block);
2230         else
2231                 device_for_each_child(dev, NULL, target_block);
2232 }
2233 EXPORT_SYMBOL_GPL(scsi_target_block);
2234
2235 static void
2236 device_unblock(struct scsi_device *sdev, void *data)
2237 {
2238         scsi_internal_device_unblock(sdev);
2239 }
2240
2241 static int
2242 target_unblock(struct device *dev, void *data)
2243 {
2244         if (scsi_is_target_device(dev))
2245                 starget_for_each_device(to_scsi_target(dev), NULL,
2246                                         device_unblock);
2247         return 0;
2248 }
2249
2250 void
2251 scsi_target_unblock(struct device *dev)
2252 {
2253         if (scsi_is_target_device(dev))
2254                 starget_for_each_device(to_scsi_target(dev), NULL,
2255                                         device_unblock);
2256         else
2257                 device_for_each_child(dev, NULL, target_unblock);
2258 }
2259 EXPORT_SYMBOL_GPL(scsi_target_unblock);