Manual merge with Linus.
[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 + 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 ready,
1071                          * retry.
1072                          */
1073                         if (sshdr.asc == 0x04 && sshdr.ascq == 0x01) {
1074                                 scsi_requeue_command(q, cmd);
1075                                 return;
1076                         }
1077                         if (!(req->flags & REQ_QUIET))
1078                                 scmd_printk(KERN_INFO, cmd,
1079                                            "Device not ready.\n");
1080                         scsi_end_request(cmd, 0, this_count, 1);
1081                         return;
1082                 case VOLUME_OVERFLOW:
1083                         if (!(req->flags & REQ_QUIET)) {
1084                                 scmd_printk(KERN_INFO, cmd,
1085                                            "Volume overflow, CDB: ");
1086                                 __scsi_print_command(cmd->data_cmnd);
1087                                 scsi_print_sense("", cmd);
1088                         }
1089                         scsi_end_request(cmd, 0, block_bytes, 1);
1090                         return;
1091                 default:
1092                         break;
1093                 }
1094         }                       /* driver byte != 0 */
1095         if (host_byte(result) == DID_RESET) {
1096                 /*
1097                  * Third party bus reset or reset for error
1098                  * recovery reasons.  Just retry the request
1099                  * and see what happens.  
1100                  */
1101                 scsi_requeue_command(q, cmd);
1102                 return;
1103         }
1104         if (result) {
1105                 if (!(req->flags & REQ_QUIET)) {
1106                         scmd_printk(KERN_INFO, cmd,
1107                                    "SCSI error: return code = 0x%x\n", result);
1108
1109                         if (driver_byte(result) & DRIVER_SENSE)
1110                                 scsi_print_sense("", cmd);
1111                 }
1112                 /*
1113                  * Mark a single buffer as not uptodate.  Queue the remainder.
1114                  * We sometimes get this cruft in the event that a medium error
1115                  * isn't properly reported.
1116                  */
1117                 block_bytes = req->hard_cur_sectors << 9;
1118                 if (!block_bytes)
1119                         block_bytes = req->data_len;
1120                 scsi_end_request(cmd, 0, block_bytes, 1);
1121         }
1122 }
1123 EXPORT_SYMBOL(scsi_io_completion);
1124
1125 /*
1126  * Function:    scsi_init_io()
1127  *
1128  * Purpose:     SCSI I/O initialize function.
1129  *
1130  * Arguments:   cmd   - Command descriptor we wish to initialize
1131  *
1132  * Returns:     0 on success
1133  *              BLKPREP_DEFER if the failure is retryable
1134  *              BLKPREP_KILL if the failure is fatal
1135  */
1136 static int scsi_init_io(struct scsi_cmnd *cmd)
1137 {
1138         struct request     *req = cmd->request;
1139         struct scatterlist *sgpnt;
1140         int                count;
1141
1142         /*
1143          * if this is a rq->data based REQ_BLOCK_PC, setup for a non-sg xfer
1144          */
1145         if ((req->flags & REQ_BLOCK_PC) && !req->bio) {
1146                 cmd->request_bufflen = req->data_len;
1147                 cmd->request_buffer = req->data;
1148                 req->buffer = req->data;
1149                 cmd->use_sg = 0;
1150                 return 0;
1151         }
1152
1153         /*
1154          * we used to not use scatter-gather for single segment request,
1155          * but now we do (it makes highmem I/O easier to support without
1156          * kmapping pages)
1157          */
1158         cmd->use_sg = req->nr_phys_segments;
1159
1160         /*
1161          * if sg table allocation fails, requeue request later.
1162          */
1163         sgpnt = scsi_alloc_sgtable(cmd, GFP_ATOMIC);
1164         if (unlikely(!sgpnt)) {
1165                 scsi_unprep_request(req);
1166                 return BLKPREP_DEFER;
1167         }
1168
1169         cmd->request_buffer = (char *) sgpnt;
1170         cmd->request_bufflen = req->nr_sectors << 9;
1171         if (blk_pc_request(req))
1172                 cmd->request_bufflen = req->data_len;
1173         req->buffer = NULL;
1174
1175         /* 
1176          * Next, walk the list, and fill in the addresses and sizes of
1177          * each segment.
1178          */
1179         count = blk_rq_map_sg(req->q, req, cmd->request_buffer);
1180
1181         /*
1182          * mapped well, send it off
1183          */
1184         if (likely(count <= cmd->use_sg)) {
1185                 cmd->use_sg = count;
1186                 return 0;
1187         }
1188
1189         printk(KERN_ERR "Incorrect number of segments after building list\n");
1190         printk(KERN_ERR "counted %d, received %d\n", count, cmd->use_sg);
1191         printk(KERN_ERR "req nr_sec %lu, cur_nr_sec %u\n", req->nr_sectors,
1192                         req->current_nr_sectors);
1193
1194         /* release the command and kill it */
1195         scsi_release_buffers(cmd);
1196         scsi_put_command(cmd);
1197         return BLKPREP_KILL;
1198 }
1199
1200 static int scsi_issue_flush_fn(request_queue_t *q, struct gendisk *disk,
1201                                sector_t *error_sector)
1202 {
1203         struct scsi_device *sdev = q->queuedata;
1204         struct scsi_driver *drv;
1205
1206         if (sdev->sdev_state != SDEV_RUNNING)
1207                 return -ENXIO;
1208
1209         drv = *(struct scsi_driver **) disk->private_data;
1210         if (drv->issue_flush)
1211                 return drv->issue_flush(&sdev->sdev_gendev, error_sector);
1212
1213         return -EOPNOTSUPP;
1214 }
1215
1216 static void scsi_blk_pc_done(struct scsi_cmnd *cmd)
1217 {
1218         BUG_ON(!blk_pc_request(cmd->request));
1219         /*
1220          * This will complete the whole command with uptodate=1 so
1221          * as far as the block layer is concerned the command completed
1222          * successfully. Since this is a REQ_BLOCK_PC command the
1223          * caller should check the request's errors value
1224          */
1225         scsi_io_completion(cmd, cmd->bufflen, 0);
1226 }
1227
1228 static void scsi_setup_blk_pc_cmnd(struct scsi_cmnd *cmd)
1229 {
1230         struct request *req = cmd->request;
1231
1232         BUG_ON(sizeof(req->cmd) > sizeof(cmd->cmnd));
1233         memcpy(cmd->cmnd, req->cmd, sizeof(cmd->cmnd));
1234         cmd->cmd_len = req->cmd_len;
1235         if (!req->data_len)
1236                 cmd->sc_data_direction = DMA_NONE;
1237         else if (rq_data_dir(req) == WRITE)
1238                 cmd->sc_data_direction = DMA_TO_DEVICE;
1239         else
1240                 cmd->sc_data_direction = DMA_FROM_DEVICE;
1241         
1242         cmd->transfersize = req->data_len;
1243         cmd->allowed = req->retries;
1244         cmd->timeout_per_command = req->timeout;
1245         cmd->done = scsi_blk_pc_done;
1246 }
1247
1248 static int scsi_prep_fn(struct request_queue *q, struct request *req)
1249 {
1250         struct scsi_device *sdev = q->queuedata;
1251         struct scsi_cmnd *cmd;
1252         int specials_only = 0;
1253
1254         /*
1255          * Just check to see if the device is online.  If it isn't, we
1256          * refuse to process any commands.  The device must be brought
1257          * online before trying any recovery commands
1258          */
1259         if (unlikely(!scsi_device_online(sdev))) {
1260                 sdev_printk(KERN_ERR, sdev,
1261                             "rejecting I/O to offline device\n");
1262                 goto kill;
1263         }
1264         if (unlikely(sdev->sdev_state != SDEV_RUNNING)) {
1265                 /* OK, we're not in a running state don't prep
1266                  * user commands */
1267                 if (sdev->sdev_state == SDEV_DEL) {
1268                         /* Device is fully deleted, no commands
1269                          * at all allowed down */
1270                         sdev_printk(KERN_ERR, sdev,
1271                                     "rejecting I/O to dead device\n");
1272                         goto kill;
1273                 }
1274                 /* OK, we only allow special commands (i.e. not
1275                  * user initiated ones */
1276                 specials_only = sdev->sdev_state;
1277         }
1278
1279         /*
1280          * Find the actual device driver associated with this command.
1281          * The SPECIAL requests are things like character device or
1282          * ioctls, which did not originate from ll_rw_blk.  Note that
1283          * the special field is also used to indicate the cmd for
1284          * the remainder of a partially fulfilled request that can 
1285          * come up when there is a medium error.  We have to treat
1286          * these two cases differently.  We differentiate by looking
1287          * at request->cmd, as this tells us the real story.
1288          */
1289         if (req->flags & REQ_SPECIAL && req->special) {
1290                 struct scsi_request *sreq = req->special;
1291
1292                 if (sreq->sr_magic == SCSI_REQ_MAGIC) {
1293                         cmd = scsi_get_command(sreq->sr_device, GFP_ATOMIC);
1294                         if (unlikely(!cmd))
1295                                 goto defer;
1296                         scsi_init_cmd_from_req(cmd, sreq);
1297                 } else
1298                         cmd = req->special;
1299         } else if (req->flags & (REQ_CMD | REQ_BLOCK_PC)) {
1300
1301                 if(unlikely(specials_only) && !(req->flags & REQ_SPECIAL)) {
1302                         if(specials_only == SDEV_QUIESCE ||
1303                                         specials_only == SDEV_BLOCK)
1304                                 goto defer;
1305                         
1306                         sdev_printk(KERN_ERR, sdev,
1307                                     "rejecting I/O to device being removed\n");
1308                         goto kill;
1309                 }
1310                         
1311                         
1312                 /*
1313                  * Now try and find a command block that we can use.
1314                  */
1315                 if (!req->special) {
1316                         cmd = scsi_get_command(sdev, GFP_ATOMIC);
1317                         if (unlikely(!cmd))
1318                                 goto defer;
1319                 } else
1320                         cmd = req->special;
1321                 
1322                 /* pull a tag out of the request if we have one */
1323                 cmd->tag = req->tag;
1324         } else {
1325                 blk_dump_rq_flags(req, "SCSI bad req");
1326                 goto kill;
1327         }
1328         
1329         /* note the overloading of req->special.  When the tag
1330          * is active it always means cmd.  If the tag goes
1331          * back for re-queueing, it may be reset */
1332         req->special = cmd;
1333         cmd->request = req;
1334         
1335         /*
1336          * FIXME: drop the lock here because the functions below
1337          * expect to be called without the queue lock held.  Also,
1338          * previously, we dequeued the request before dropping the
1339          * lock.  We hope REQ_STARTED prevents anything untoward from
1340          * happening now.
1341          */
1342         if (req->flags & (REQ_CMD | REQ_BLOCK_PC)) {
1343                 int ret;
1344
1345                 /*
1346                  * This will do a couple of things:
1347                  *  1) Fill in the actual SCSI command.
1348                  *  2) Fill in any other upper-level specific fields
1349                  * (timeout).
1350                  *
1351                  * If this returns 0, it means that the request failed
1352                  * (reading past end of disk, reading offline device,
1353                  * etc).   This won't actually talk to the device, but
1354                  * some kinds of consistency checking may cause the     
1355                  * request to be rejected immediately.
1356                  */
1357
1358                 /* 
1359                  * This sets up the scatter-gather table (allocating if
1360                  * required).
1361                  */
1362                 ret = scsi_init_io(cmd);
1363                 switch(ret) {
1364                         /* For BLKPREP_KILL/DEFER the cmd was released */
1365                 case BLKPREP_KILL:
1366                         goto kill;
1367                 case BLKPREP_DEFER:
1368                         goto defer;
1369                 }
1370                 
1371                 /*
1372                  * Initialize the actual SCSI command for this request.
1373                  */
1374                 if (req->flags & REQ_BLOCK_PC) {
1375                         scsi_setup_blk_pc_cmnd(cmd);
1376                 } else if (req->rq_disk) {
1377                         struct scsi_driver *drv;
1378
1379                         drv = *(struct scsi_driver **)req->rq_disk->private_data;
1380                         if (unlikely(!drv->init_command(cmd))) {
1381                                 scsi_release_buffers(cmd);
1382                                 scsi_put_command(cmd);
1383                                 goto kill;
1384                         }
1385                 }
1386         }
1387
1388         /*
1389          * The request is now prepped, no need to come back here
1390          */
1391         req->flags |= REQ_DONTPREP;
1392         return BLKPREP_OK;
1393
1394  defer:
1395         /* If we defer, the elv_next_request() returns NULL, but the
1396          * queue must be restarted, so we plug here if no returning
1397          * command will automatically do that. */
1398         if (sdev->device_busy == 0)
1399                 blk_plug_device(q);
1400         return BLKPREP_DEFER;
1401  kill:
1402         req->errors = DID_NO_CONNECT << 16;
1403         return BLKPREP_KILL;
1404 }
1405
1406 /*
1407  * scsi_dev_queue_ready: if we can send requests to sdev, return 1 else
1408  * return 0.
1409  *
1410  * Called with the queue_lock held.
1411  */
1412 static inline int scsi_dev_queue_ready(struct request_queue *q,
1413                                   struct scsi_device *sdev)
1414 {
1415         if (sdev->device_busy >= sdev->queue_depth)
1416                 return 0;
1417         if (sdev->device_busy == 0 && sdev->device_blocked) {
1418                 /*
1419                  * unblock after device_blocked iterates to zero
1420                  */
1421                 if (--sdev->device_blocked == 0) {
1422                         SCSI_LOG_MLQUEUE(3,
1423                                    sdev_printk(KERN_INFO, sdev,
1424                                    "unblocking device at zero depth\n"));
1425                 } else {
1426                         blk_plug_device(q);
1427                         return 0;
1428                 }
1429         }
1430         if (sdev->device_blocked)
1431                 return 0;
1432
1433         return 1;
1434 }
1435
1436 /*
1437  * scsi_host_queue_ready: if we can send requests to shost, return 1 else
1438  * return 0. We must end up running the queue again whenever 0 is
1439  * returned, else IO can hang.
1440  *
1441  * Called with host_lock held.
1442  */
1443 static inline int scsi_host_queue_ready(struct request_queue *q,
1444                                    struct Scsi_Host *shost,
1445                                    struct scsi_device *sdev)
1446 {
1447         if (scsi_host_in_recovery(shost))
1448                 return 0;
1449         if (shost->host_busy == 0 && shost->host_blocked) {
1450                 /*
1451                  * unblock after host_blocked iterates to zero
1452                  */
1453                 if (--shost->host_blocked == 0) {
1454                         SCSI_LOG_MLQUEUE(3,
1455                                 printk("scsi%d unblocking host at zero depth\n",
1456                                         shost->host_no));
1457                 } else {
1458                         blk_plug_device(q);
1459                         return 0;
1460                 }
1461         }
1462         if ((shost->can_queue > 0 && shost->host_busy >= shost->can_queue) ||
1463             shost->host_blocked || shost->host_self_blocked) {
1464                 if (list_empty(&sdev->starved_entry))
1465                         list_add_tail(&sdev->starved_entry, &shost->starved_list);
1466                 return 0;
1467         }
1468
1469         /* We're OK to process the command, so we can't be starved */
1470         if (!list_empty(&sdev->starved_entry))
1471                 list_del_init(&sdev->starved_entry);
1472
1473         return 1;
1474 }
1475
1476 /*
1477  * Kill a request for a dead device
1478  */
1479 static void scsi_kill_request(struct request *req, request_queue_t *q)
1480 {
1481         struct scsi_cmnd *cmd = req->special;
1482
1483         blkdev_dequeue_request(req);
1484
1485         if (unlikely(cmd == NULL)) {
1486                 printk(KERN_CRIT "impossible request in %s.\n",
1487                                  __FUNCTION__);
1488                 BUG();
1489         }
1490
1491         scsi_init_cmd_errh(cmd);
1492         cmd->result = DID_NO_CONNECT << 16;
1493         atomic_inc(&cmd->device->iorequest_cnt);
1494         __scsi_done(cmd);
1495 }
1496
1497 static void scsi_softirq_done(struct request *rq)
1498 {
1499         struct scsi_cmnd *cmd = rq->completion_data;
1500         unsigned long wait_for = (cmd->allowed + 1) * cmd->timeout_per_command;
1501         int disposition;
1502
1503         INIT_LIST_HEAD(&cmd->eh_entry);
1504
1505         disposition = scsi_decide_disposition(cmd);
1506         if (disposition != SUCCESS &&
1507             time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) {
1508                 sdev_printk(KERN_ERR, cmd->device,
1509                             "timing out command, waited %lus\n",
1510                             wait_for/HZ);
1511                 disposition = SUCCESS;
1512         }
1513                         
1514         scsi_log_completion(cmd, disposition);
1515
1516         switch (disposition) {
1517                 case SUCCESS:
1518                         scsi_finish_command(cmd);
1519                         break;
1520                 case NEEDS_RETRY:
1521                         scsi_retry_command(cmd);
1522                         break;
1523                 case ADD_TO_MLQUEUE:
1524                         scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY);
1525                         break;
1526                 default:
1527                         if (!scsi_eh_scmd_add(cmd, 0))
1528                                 scsi_finish_command(cmd);
1529         }
1530 }
1531
1532 /*
1533  * Function:    scsi_request_fn()
1534  *
1535  * Purpose:     Main strategy routine for SCSI.
1536  *
1537  * Arguments:   q       - Pointer to actual queue.
1538  *
1539  * Returns:     Nothing
1540  *
1541  * Lock status: IO request lock assumed to be held when called.
1542  */
1543 static void scsi_request_fn(struct request_queue *q)
1544 {
1545         struct scsi_device *sdev = q->queuedata;
1546         struct Scsi_Host *shost;
1547         struct scsi_cmnd *cmd;
1548         struct request *req;
1549
1550         if (!sdev) {
1551                 printk("scsi: killing requests for dead queue\n");
1552                 while ((req = elv_next_request(q)) != NULL)
1553                         scsi_kill_request(req, q);
1554                 return;
1555         }
1556
1557         if(!get_device(&sdev->sdev_gendev))
1558                 /* We must be tearing the block queue down already */
1559                 return;
1560
1561         /*
1562          * To start with, we keep looping until the queue is empty, or until
1563          * the host is no longer able to accept any more requests.
1564          */
1565         shost = sdev->host;
1566         while (!blk_queue_plugged(q)) {
1567                 int rtn;
1568                 /*
1569                  * get next queueable request.  We do this early to make sure
1570                  * that the request is fully prepared even if we cannot 
1571                  * accept it.
1572                  */
1573                 req = elv_next_request(q);
1574                 if (!req || !scsi_dev_queue_ready(q, sdev))
1575                         break;
1576
1577                 if (unlikely(!scsi_device_online(sdev))) {
1578                         sdev_printk(KERN_ERR, sdev,
1579                                     "rejecting I/O to offline device\n");
1580                         scsi_kill_request(req, q);
1581                         continue;
1582                 }
1583
1584
1585                 /*
1586                  * Remove the request from the request list.
1587                  */
1588                 if (!(blk_queue_tagged(q) && !blk_queue_start_tag(q, req)))
1589                         blkdev_dequeue_request(req);
1590                 sdev->device_busy++;
1591
1592                 spin_unlock(q->queue_lock);
1593                 cmd = req->special;
1594                 if (unlikely(cmd == NULL)) {
1595                         printk(KERN_CRIT "impossible request in %s.\n"
1596                                          "please mail a stack trace to "
1597                                          "linux-scsi@vger.kernel.org",
1598                                          __FUNCTION__);
1599                         BUG();
1600                 }
1601                 spin_lock(shost->host_lock);
1602
1603                 if (!scsi_host_queue_ready(q, shost, sdev))
1604                         goto not_ready;
1605                 if (sdev->single_lun) {
1606                         if (scsi_target(sdev)->starget_sdev_user &&
1607                             scsi_target(sdev)->starget_sdev_user != sdev)
1608                                 goto not_ready;
1609                         scsi_target(sdev)->starget_sdev_user = sdev;
1610                 }
1611                 shost->host_busy++;
1612
1613                 /*
1614                  * XXX(hch): This is rather suboptimal, scsi_dispatch_cmd will
1615                  *              take the lock again.
1616                  */
1617                 spin_unlock_irq(shost->host_lock);
1618
1619                 /*
1620                  * Finally, initialize any error handling parameters, and set up
1621                  * the timers for timeouts.
1622                  */
1623                 scsi_init_cmd_errh(cmd);
1624
1625                 /*
1626                  * Dispatch the command to the low-level driver.
1627                  */
1628                 rtn = scsi_dispatch_cmd(cmd);
1629                 spin_lock_irq(q->queue_lock);
1630                 if(rtn) {
1631                         /* we're refusing the command; because of
1632                          * the way locks get dropped, we need to 
1633                          * check here if plugging is required */
1634                         if(sdev->device_busy == 0)
1635                                 blk_plug_device(q);
1636
1637                         break;
1638                 }
1639         }
1640
1641         goto out;
1642
1643  not_ready:
1644         spin_unlock_irq(shost->host_lock);
1645
1646         /*
1647          * lock q, handle tag, requeue req, and decrement device_busy. We
1648          * must return with queue_lock held.
1649          *
1650          * Decrementing device_busy without checking it is OK, as all such
1651          * cases (host limits or settings) should run the queue at some
1652          * later time.
1653          */
1654         spin_lock_irq(q->queue_lock);
1655         blk_requeue_request(q, req);
1656         sdev->device_busy--;
1657         if(sdev->device_busy == 0)
1658                 blk_plug_device(q);
1659  out:
1660         /* must be careful here...if we trigger the ->remove() function
1661          * we cannot be holding the q lock */
1662         spin_unlock_irq(q->queue_lock);
1663         put_device(&sdev->sdev_gendev);
1664         spin_lock_irq(q->queue_lock);
1665 }
1666
1667 u64 scsi_calculate_bounce_limit(struct Scsi_Host *shost)
1668 {
1669         struct device *host_dev;
1670         u64 bounce_limit = 0xffffffff;
1671
1672         if (shost->unchecked_isa_dma)
1673                 return BLK_BOUNCE_ISA;
1674         /*
1675          * Platforms with virtual-DMA translation
1676          * hardware have no practical limit.
1677          */
1678         if (!PCI_DMA_BUS_IS_PHYS)
1679                 return BLK_BOUNCE_ANY;
1680
1681         host_dev = scsi_get_device(shost);
1682         if (host_dev && host_dev->dma_mask)
1683                 bounce_limit = *host_dev->dma_mask;
1684
1685         return bounce_limit;
1686 }
1687 EXPORT_SYMBOL(scsi_calculate_bounce_limit);
1688
1689 struct request_queue *scsi_alloc_queue(struct scsi_device *sdev)
1690 {
1691         struct Scsi_Host *shost = sdev->host;
1692         struct request_queue *q;
1693
1694         q = blk_init_queue(scsi_request_fn, NULL);
1695         if (!q)
1696                 return NULL;
1697
1698         blk_queue_prep_rq(q, scsi_prep_fn);
1699
1700         blk_queue_max_hw_segments(q, shost->sg_tablesize);
1701         blk_queue_max_phys_segments(q, SCSI_MAX_PHYS_SEGMENTS);
1702         blk_queue_max_sectors(q, shost->max_sectors);
1703         blk_queue_bounce_limit(q, scsi_calculate_bounce_limit(shost));
1704         blk_queue_segment_boundary(q, shost->dma_boundary);
1705         blk_queue_issue_flush_fn(q, scsi_issue_flush_fn);
1706         blk_queue_softirq_done(q, scsi_softirq_done);
1707
1708         if (!shost->use_clustering)
1709                 clear_bit(QUEUE_FLAG_CLUSTER, &q->queue_flags);
1710         return q;
1711 }
1712
1713 void scsi_free_queue(struct request_queue *q)
1714 {
1715         blk_cleanup_queue(q);
1716 }
1717
1718 /*
1719  * Function:    scsi_block_requests()
1720  *
1721  * Purpose:     Utility function used by low-level drivers to prevent further
1722  *              commands from being queued to the device.
1723  *
1724  * Arguments:   shost       - Host in question
1725  *
1726  * Returns:     Nothing
1727  *
1728  * Lock status: No locks are assumed held.
1729  *
1730  * Notes:       There is no timer nor any other means by which the requests
1731  *              get unblocked other than the low-level driver calling
1732  *              scsi_unblock_requests().
1733  */
1734 void scsi_block_requests(struct Scsi_Host *shost)
1735 {
1736         shost->host_self_blocked = 1;
1737 }
1738 EXPORT_SYMBOL(scsi_block_requests);
1739
1740 /*
1741  * Function:    scsi_unblock_requests()
1742  *
1743  * Purpose:     Utility function used by low-level drivers to allow further
1744  *              commands from being queued to the device.
1745  *
1746  * Arguments:   shost       - Host in question
1747  *
1748  * Returns:     Nothing
1749  *
1750  * Lock status: No locks are assumed held.
1751  *
1752  * Notes:       There is no timer nor any other means by which the requests
1753  *              get unblocked other than the low-level driver calling
1754  *              scsi_unblock_requests().
1755  *
1756  *              This is done as an API function so that changes to the
1757  *              internals of the scsi mid-layer won't require wholesale
1758  *              changes to drivers that use this feature.
1759  */
1760 void scsi_unblock_requests(struct Scsi_Host *shost)
1761 {
1762         shost->host_self_blocked = 0;
1763         scsi_run_host_queues(shost);
1764 }
1765 EXPORT_SYMBOL(scsi_unblock_requests);
1766
1767 int __init scsi_init_queue(void)
1768 {
1769         int i;
1770
1771         scsi_io_context_cache = kmem_cache_create("scsi_io_context",
1772                                         sizeof(struct scsi_io_context),
1773                                         0, 0, NULL, NULL);
1774         if (!scsi_io_context_cache) {
1775                 printk(KERN_ERR "SCSI: can't init scsi io context cache\n");
1776                 return -ENOMEM;
1777         }
1778
1779         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1780                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1781                 int size = sgp->size * sizeof(struct scatterlist);
1782
1783                 sgp->slab = kmem_cache_create(sgp->name, size, 0,
1784                                 SLAB_HWCACHE_ALIGN, NULL, NULL);
1785                 if (!sgp->slab) {
1786                         printk(KERN_ERR "SCSI: can't init sg slab %s\n",
1787                                         sgp->name);
1788                 }
1789
1790                 sgp->pool = mempool_create_slab_pool(SG_MEMPOOL_SIZE,
1791                                                      sgp->slab);
1792                 if (!sgp->pool) {
1793                         printk(KERN_ERR "SCSI: can't init sg mempool %s\n",
1794                                         sgp->name);
1795                 }
1796         }
1797
1798         return 0;
1799 }
1800
1801 void scsi_exit_queue(void)
1802 {
1803         int i;
1804
1805         kmem_cache_destroy(scsi_io_context_cache);
1806
1807         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1808                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1809                 mempool_destroy(sgp->pool);
1810                 kmem_cache_destroy(sgp->slab);
1811         }
1812 }
1813
1814 /**
1815  *      scsi_mode_select - issue a mode select
1816  *      @sdev:  SCSI device to be queried
1817  *      @pf:    Page format bit (1 == standard, 0 == vendor specific)
1818  *      @sp:    Save page bit (0 == don't save, 1 == save)
1819  *      @modepage: mode page being requested
1820  *      @buffer: request buffer (may not be smaller than eight bytes)
1821  *      @len:   length of request buffer.
1822  *      @timeout: command timeout
1823  *      @retries: number of retries before failing
1824  *      @data: returns a structure abstracting the mode header data
1825  *      @sense: place to put sense data (or NULL if no sense to be collected).
1826  *              must be SCSI_SENSE_BUFFERSIZE big.
1827  *
1828  *      Returns zero if successful; negative error number or scsi
1829  *      status on error
1830  *
1831  */
1832 int
1833 scsi_mode_select(struct scsi_device *sdev, int pf, int sp, int modepage,
1834                  unsigned char *buffer, int len, int timeout, int retries,
1835                  struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1836 {
1837         unsigned char cmd[10];
1838         unsigned char *real_buffer;
1839         int ret;
1840
1841         memset(cmd, 0, sizeof(cmd));
1842         cmd[1] = (pf ? 0x10 : 0) | (sp ? 0x01 : 0);
1843
1844         if (sdev->use_10_for_ms) {
1845                 if (len > 65535)
1846                         return -EINVAL;
1847                 real_buffer = kmalloc(8 + len, GFP_KERNEL);
1848                 if (!real_buffer)
1849                         return -ENOMEM;
1850                 memcpy(real_buffer + 8, buffer, len);
1851                 len += 8;
1852                 real_buffer[0] = 0;
1853                 real_buffer[1] = 0;
1854                 real_buffer[2] = data->medium_type;
1855                 real_buffer[3] = data->device_specific;
1856                 real_buffer[4] = data->longlba ? 0x01 : 0;
1857                 real_buffer[5] = 0;
1858                 real_buffer[6] = data->block_descriptor_length >> 8;
1859                 real_buffer[7] = data->block_descriptor_length;
1860
1861                 cmd[0] = MODE_SELECT_10;
1862                 cmd[7] = len >> 8;
1863                 cmd[8] = len;
1864         } else {
1865                 if (len > 255 || data->block_descriptor_length > 255 ||
1866                     data->longlba)
1867                         return -EINVAL;
1868
1869                 real_buffer = kmalloc(4 + len, GFP_KERNEL);
1870                 if (!real_buffer)
1871                         return -ENOMEM;
1872                 memcpy(real_buffer + 4, buffer, len);
1873                 len += 4;
1874                 real_buffer[0] = 0;
1875                 real_buffer[1] = data->medium_type;
1876                 real_buffer[2] = data->device_specific;
1877                 real_buffer[3] = data->block_descriptor_length;
1878                 
1879
1880                 cmd[0] = MODE_SELECT;
1881                 cmd[4] = len;
1882         }
1883
1884         ret = scsi_execute_req(sdev, cmd, DMA_TO_DEVICE, real_buffer, len,
1885                                sshdr, timeout, retries);
1886         kfree(real_buffer);
1887         return ret;
1888 }
1889 EXPORT_SYMBOL_GPL(scsi_mode_select);
1890
1891 /**
1892  *      scsi_mode_sense - issue a mode sense, falling back from 10 to 
1893  *              six bytes if necessary.
1894  *      @sdev:  SCSI device to be queried
1895  *      @dbd:   set if mode sense will allow block descriptors to be returned
1896  *      @modepage: mode page being requested
1897  *      @buffer: request buffer (may not be smaller than eight bytes)
1898  *      @len:   length of request buffer.
1899  *      @timeout: command timeout
1900  *      @retries: number of retries before failing
1901  *      @data: returns a structure abstracting the mode header data
1902  *      @sense: place to put sense data (or NULL if no sense to be collected).
1903  *              must be SCSI_SENSE_BUFFERSIZE big.
1904  *
1905  *      Returns zero if unsuccessful, or the header offset (either 4
1906  *      or 8 depending on whether a six or ten byte command was
1907  *      issued) if successful.
1908  **/
1909 int
1910 scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
1911                   unsigned char *buffer, int len, int timeout, int retries,
1912                   struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1913 {
1914         unsigned char cmd[12];
1915         int use_10_for_ms;
1916         int header_length;
1917         int result;
1918         struct scsi_sense_hdr my_sshdr;
1919
1920         memset(data, 0, sizeof(*data));
1921         memset(&cmd[0], 0, 12);
1922         cmd[1] = dbd & 0x18;    /* allows DBD and LLBA bits */
1923         cmd[2] = modepage;
1924
1925         /* caller might not be interested in sense, but we need it */
1926         if (!sshdr)
1927                 sshdr = &my_sshdr;
1928
1929  retry:
1930         use_10_for_ms = sdev->use_10_for_ms;
1931
1932         if (use_10_for_ms) {
1933                 if (len < 8)
1934                         len = 8;
1935
1936                 cmd[0] = MODE_SENSE_10;
1937                 cmd[8] = len;
1938                 header_length = 8;
1939         } else {
1940                 if (len < 4)
1941                         len = 4;
1942
1943                 cmd[0] = MODE_SENSE;
1944                 cmd[4] = len;
1945                 header_length = 4;
1946         }
1947
1948         memset(buffer, 0, len);
1949
1950         result = scsi_execute_req(sdev, cmd, DMA_FROM_DEVICE, buffer, len,
1951                                   sshdr, timeout, retries);
1952
1953         /* This code looks awful: what it's doing is making sure an
1954          * ILLEGAL REQUEST sense return identifies the actual command
1955          * byte as the problem.  MODE_SENSE commands can return
1956          * ILLEGAL REQUEST if the code page isn't supported */
1957
1958         if (use_10_for_ms && !scsi_status_is_good(result) &&
1959             (driver_byte(result) & DRIVER_SENSE)) {
1960                 if (scsi_sense_valid(sshdr)) {
1961                         if ((sshdr->sense_key == ILLEGAL_REQUEST) &&
1962                             (sshdr->asc == 0x20) && (sshdr->ascq == 0)) {
1963                                 /* 
1964                                  * Invalid command operation code
1965                                  */
1966                                 sdev->use_10_for_ms = 0;
1967                                 goto retry;
1968                         }
1969                 }
1970         }
1971
1972         if(scsi_status_is_good(result)) {
1973                 if (unlikely(buffer[0] == 0x86 && buffer[1] == 0x0b &&
1974                              (modepage == 6 || modepage == 8))) {
1975                         /* Initio breakage? */
1976                         header_length = 0;
1977                         data->length = 13;
1978                         data->medium_type = 0;
1979                         data->device_specific = 0;
1980                         data->longlba = 0;
1981                         data->block_descriptor_length = 0;
1982                 } else if(use_10_for_ms) {
1983                         data->length = buffer[0]*256 + buffer[1] + 2;
1984                         data->medium_type = buffer[2];
1985                         data->device_specific = buffer[3];
1986                         data->longlba = buffer[4] & 0x01;
1987                         data->block_descriptor_length = buffer[6]*256
1988                                 + buffer[7];
1989                 } else {
1990                         data->length = buffer[0] + 1;
1991                         data->medium_type = buffer[1];
1992                         data->device_specific = buffer[2];
1993                         data->block_descriptor_length = buffer[3];
1994                 }
1995                 data->header_length = header_length;
1996         }
1997
1998         return result;
1999 }
2000 EXPORT_SYMBOL(scsi_mode_sense);
2001
2002 int
2003 scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries)
2004 {
2005         char cmd[] = {
2006                 TEST_UNIT_READY, 0, 0, 0, 0, 0,
2007         };
2008         struct scsi_sense_hdr sshdr;
2009         int result;
2010         
2011         result = scsi_execute_req(sdev, cmd, DMA_NONE, NULL, 0, &sshdr,
2012                                   timeout, retries);
2013
2014         if ((driver_byte(result) & DRIVER_SENSE) && sdev->removable) {
2015
2016                 if ((scsi_sense_valid(&sshdr)) &&
2017                     ((sshdr.sense_key == UNIT_ATTENTION) ||
2018                      (sshdr.sense_key == NOT_READY))) {
2019                         sdev->changed = 1;
2020                         result = 0;
2021                 }
2022         }
2023         return result;
2024 }
2025 EXPORT_SYMBOL(scsi_test_unit_ready);
2026
2027 /**
2028  *      scsi_device_set_state - Take the given device through the device
2029  *              state model.
2030  *      @sdev:  scsi device to change the state of.
2031  *      @state: state to change to.
2032  *
2033  *      Returns zero if unsuccessful or an error if the requested 
2034  *      transition is illegal.
2035  **/
2036 int
2037 scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
2038 {
2039         enum scsi_device_state oldstate = sdev->sdev_state;
2040
2041         if (state == oldstate)
2042                 return 0;
2043
2044         switch (state) {
2045         case SDEV_CREATED:
2046                 /* There are no legal states that come back to
2047                  * created.  This is the manually initialised start
2048                  * state */
2049                 goto illegal;
2050                         
2051         case SDEV_RUNNING:
2052                 switch (oldstate) {
2053                 case SDEV_CREATED:
2054                 case SDEV_OFFLINE:
2055                 case SDEV_QUIESCE:
2056                 case SDEV_BLOCK:
2057                         break;
2058                 default:
2059                         goto illegal;
2060                 }
2061                 break;
2062
2063         case SDEV_QUIESCE:
2064                 switch (oldstate) {
2065                 case SDEV_RUNNING:
2066                 case SDEV_OFFLINE:
2067                         break;
2068                 default:
2069                         goto illegal;
2070                 }
2071                 break;
2072
2073         case SDEV_OFFLINE:
2074                 switch (oldstate) {
2075                 case SDEV_CREATED:
2076                 case SDEV_RUNNING:
2077                 case SDEV_QUIESCE:
2078                 case SDEV_BLOCK:
2079                         break;
2080                 default:
2081                         goto illegal;
2082                 }
2083                 break;
2084
2085         case SDEV_BLOCK:
2086                 switch (oldstate) {
2087                 case SDEV_CREATED:
2088                 case SDEV_RUNNING:
2089                         break;
2090                 default:
2091                         goto illegal;
2092                 }
2093                 break;
2094
2095         case SDEV_CANCEL:
2096                 switch (oldstate) {
2097                 case SDEV_CREATED:
2098                 case SDEV_RUNNING:
2099                 case SDEV_OFFLINE:
2100                 case SDEV_BLOCK:
2101                         break;
2102                 default:
2103                         goto illegal;
2104                 }
2105                 break;
2106
2107         case SDEV_DEL:
2108                 switch (oldstate) {
2109                 case SDEV_CANCEL:
2110                         break;
2111                 default:
2112                         goto illegal;
2113                 }
2114                 break;
2115
2116         }
2117         sdev->sdev_state = state;
2118         return 0;
2119
2120  illegal:
2121         SCSI_LOG_ERROR_RECOVERY(1, 
2122                                 sdev_printk(KERN_ERR, sdev,
2123                                             "Illegal state transition %s->%s\n",
2124                                             scsi_device_state_name(oldstate),
2125                                             scsi_device_state_name(state))
2126                                 );
2127         return -EINVAL;
2128 }
2129 EXPORT_SYMBOL(scsi_device_set_state);
2130
2131 /**
2132  *      scsi_device_quiesce - Block user issued commands.
2133  *      @sdev:  scsi device to quiesce.
2134  *
2135  *      This works by trying to transition to the SDEV_QUIESCE state
2136  *      (which must be a legal transition).  When the device is in this
2137  *      state, only special requests will be accepted, all others will
2138  *      be deferred.  Since special requests may also be requeued requests,
2139  *      a successful return doesn't guarantee the device will be 
2140  *      totally quiescent.
2141  *
2142  *      Must be called with user context, may sleep.
2143  *
2144  *      Returns zero if unsuccessful or an error if not.
2145  **/
2146 int
2147 scsi_device_quiesce(struct scsi_device *sdev)
2148 {
2149         int err = scsi_device_set_state(sdev, SDEV_QUIESCE);
2150         if (err)
2151                 return err;
2152
2153         scsi_run_queue(sdev->request_queue);
2154         while (sdev->device_busy) {
2155                 msleep_interruptible(200);
2156                 scsi_run_queue(sdev->request_queue);
2157         }
2158         return 0;
2159 }
2160 EXPORT_SYMBOL(scsi_device_quiesce);
2161
2162 /**
2163  *      scsi_device_resume - Restart user issued commands to a quiesced device.
2164  *      @sdev:  scsi device to resume.
2165  *
2166  *      Moves the device from quiesced back to running and restarts the
2167  *      queues.
2168  *
2169  *      Must be called with user context, may sleep.
2170  **/
2171 void
2172 scsi_device_resume(struct scsi_device *sdev)
2173 {
2174         if(scsi_device_set_state(sdev, SDEV_RUNNING))
2175                 return;
2176         scsi_run_queue(sdev->request_queue);
2177 }
2178 EXPORT_SYMBOL(scsi_device_resume);
2179
2180 static void
2181 device_quiesce_fn(struct scsi_device *sdev, void *data)
2182 {
2183         scsi_device_quiesce(sdev);
2184 }
2185
2186 void
2187 scsi_target_quiesce(struct scsi_target *starget)
2188 {
2189         starget_for_each_device(starget, NULL, device_quiesce_fn);
2190 }
2191 EXPORT_SYMBOL(scsi_target_quiesce);
2192
2193 static void
2194 device_resume_fn(struct scsi_device *sdev, void *data)
2195 {
2196         scsi_device_resume(sdev);
2197 }
2198
2199 void
2200 scsi_target_resume(struct scsi_target *starget)
2201 {
2202         starget_for_each_device(starget, NULL, device_resume_fn);
2203 }
2204 EXPORT_SYMBOL(scsi_target_resume);
2205
2206 /**
2207  * scsi_internal_device_block - internal function to put a device
2208  *                              temporarily into the SDEV_BLOCK state
2209  * @sdev:       device to block
2210  *
2211  * Block request made by scsi lld's to temporarily stop all
2212  * scsi commands on the specified device.  Called from interrupt
2213  * or normal process context.
2214  *
2215  * Returns zero if successful or error if not
2216  *
2217  * Notes:       
2218  *      This routine transitions the device to the SDEV_BLOCK state
2219  *      (which must be a legal transition).  When the device is in this
2220  *      state, all commands are deferred until the scsi lld reenables
2221  *      the device with scsi_device_unblock or device_block_tmo fires.
2222  *      This routine assumes the host_lock is held on entry.
2223  **/
2224 int
2225 scsi_internal_device_block(struct scsi_device *sdev)
2226 {
2227         request_queue_t *q = sdev->request_queue;
2228         unsigned long flags;
2229         int err = 0;
2230
2231         err = scsi_device_set_state(sdev, SDEV_BLOCK);
2232         if (err)
2233                 return err;
2234
2235         /* 
2236          * The device has transitioned to SDEV_BLOCK.  Stop the
2237          * block layer from calling the midlayer with this device's
2238          * request queue. 
2239          */
2240         spin_lock_irqsave(q->queue_lock, flags);
2241         blk_stop_queue(q);
2242         spin_unlock_irqrestore(q->queue_lock, flags);
2243
2244         return 0;
2245 }
2246 EXPORT_SYMBOL_GPL(scsi_internal_device_block);
2247  
2248 /**
2249  * scsi_internal_device_unblock - resume a device after a block request
2250  * @sdev:       device to resume
2251  *
2252  * Called by scsi lld's or the midlayer to restart the device queue
2253  * for the previously suspended scsi device.  Called from interrupt or
2254  * normal process context.
2255  *
2256  * Returns zero if successful or error if not.
2257  *
2258  * Notes:       
2259  *      This routine transitions the device to the SDEV_RUNNING state
2260  *      (which must be a legal transition) allowing the midlayer to
2261  *      goose the queue for this device.  This routine assumes the 
2262  *      host_lock is held upon entry.
2263  **/
2264 int
2265 scsi_internal_device_unblock(struct scsi_device *sdev)
2266 {
2267         request_queue_t *q = sdev->request_queue; 
2268         int err;
2269         unsigned long flags;
2270         
2271         /* 
2272          * Try to transition the scsi device to SDEV_RUNNING
2273          * and goose the device queue if successful.  
2274          */
2275         err = scsi_device_set_state(sdev, SDEV_RUNNING);
2276         if (err)
2277                 return err;
2278
2279         spin_lock_irqsave(q->queue_lock, flags);
2280         blk_start_queue(q);
2281         spin_unlock_irqrestore(q->queue_lock, flags);
2282
2283         return 0;
2284 }
2285 EXPORT_SYMBOL_GPL(scsi_internal_device_unblock);
2286
2287 static void
2288 device_block(struct scsi_device *sdev, void *data)
2289 {
2290         scsi_internal_device_block(sdev);
2291 }
2292
2293 static int
2294 target_block(struct device *dev, void *data)
2295 {
2296         if (scsi_is_target_device(dev))
2297                 starget_for_each_device(to_scsi_target(dev), NULL,
2298                                         device_block);
2299         return 0;
2300 }
2301
2302 void
2303 scsi_target_block(struct device *dev)
2304 {
2305         if (scsi_is_target_device(dev))
2306                 starget_for_each_device(to_scsi_target(dev), NULL,
2307                                         device_block);
2308         else
2309                 device_for_each_child(dev, NULL, target_block);
2310 }
2311 EXPORT_SYMBOL_GPL(scsi_target_block);
2312
2313 static void
2314 device_unblock(struct scsi_device *sdev, void *data)
2315 {
2316         scsi_internal_device_unblock(sdev);
2317 }
2318
2319 static int
2320 target_unblock(struct device *dev, void *data)
2321 {
2322         if (scsi_is_target_device(dev))
2323                 starget_for_each_device(to_scsi_target(dev), NULL,
2324                                         device_unblock);
2325         return 0;
2326 }
2327
2328 void
2329 scsi_target_unblock(struct device *dev)
2330 {
2331         if (scsi_is_target_device(dev))
2332                 starget_for_each_device(to_scsi_target(dev), NULL,
2333                                         device_unblock);
2334         else
2335                 device_for_each_child(dev, NULL, target_unblock);
2336 }
2337 EXPORT_SYMBOL_GPL(scsi_target_unblock);