block: get rid of QUEUE_FLAG_REENTER
[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/bitops.h>
12 #include <linux/blkdev.h>
13 #include <linux/completion.h>
14 #include <linux/kernel.h>
15 #include <linux/mempool.h>
16 #include <linux/slab.h>
17 #include <linux/init.h>
18 #include <linux/pci.h>
19 #include <linux/delay.h>
20 #include <linux/hardirq.h>
21 #include <linux/scatterlist.h>
22
23 #include <scsi/scsi.h>
24 #include <scsi/scsi_cmnd.h>
25 #include <scsi/scsi_dbg.h>
26 #include <scsi/scsi_device.h>
27 #include <scsi/scsi_driver.h>
28 #include <scsi/scsi_eh.h>
29 #include <scsi/scsi_host.h>
30
31 #include "scsi_priv.h"
32 #include "scsi_logging.h"
33
34
35 #define SG_MEMPOOL_NR           ARRAY_SIZE(scsi_sg_pools)
36 #define SG_MEMPOOL_SIZE         2
37
38 struct scsi_host_sg_pool {
39         size_t          size;
40         char            *name;
41         struct kmem_cache       *slab;
42         mempool_t       *pool;
43 };
44
45 #define SP(x) { x, "sgpool-" __stringify(x) }
46 #if (SCSI_MAX_SG_SEGMENTS < 32)
47 #error SCSI_MAX_SG_SEGMENTS is too small (must be 32 or greater)
48 #endif
49 static struct scsi_host_sg_pool scsi_sg_pools[] = {
50         SP(8),
51         SP(16),
52 #if (SCSI_MAX_SG_SEGMENTS > 32)
53         SP(32),
54 #if (SCSI_MAX_SG_SEGMENTS > 64)
55         SP(64),
56 #if (SCSI_MAX_SG_SEGMENTS > 128)
57         SP(128),
58 #if (SCSI_MAX_SG_SEGMENTS > 256)
59 #error SCSI_MAX_SG_SEGMENTS is too large (256 MAX)
60 #endif
61 #endif
62 #endif
63 #endif
64         SP(SCSI_MAX_SG_SEGMENTS)
65 };
66 #undef SP
67
68 struct kmem_cache *scsi_sdb_cache;
69
70 /*
71  * When to reinvoke queueing after a resource shortage. It's 3 msecs to
72  * not change behaviour from the previous unplug mechanism, experimentation
73  * may prove this needs changing.
74  */
75 #define SCSI_QUEUE_DELAY        3
76
77 static void scsi_run_queue(struct request_queue *q);
78
79 /*
80  * Function:    scsi_unprep_request()
81  *
82  * Purpose:     Remove all preparation done for a request, including its
83  *              associated scsi_cmnd, so that it can be requeued.
84  *
85  * Arguments:   req     - request to unprepare
86  *
87  * Lock status: Assumed that no locks are held upon entry.
88  *
89  * Returns:     Nothing.
90  */
91 static void scsi_unprep_request(struct request *req)
92 {
93         struct scsi_cmnd *cmd = req->special;
94
95         blk_unprep_request(req);
96         req->special = NULL;
97
98         scsi_put_command(cmd);
99 }
100
101 /**
102  * __scsi_queue_insert - private queue insertion
103  * @cmd: The SCSI command being requeued
104  * @reason:  The reason for the requeue
105  * @unbusy: Whether the queue should be unbusied
106  *
107  * This is a private queue insertion.  The public interface
108  * scsi_queue_insert() always assumes the queue should be unbusied
109  * because it's always called before the completion.  This function is
110  * for a requeue after completion, which should only occur in this
111  * file.
112  */
113 static int __scsi_queue_insert(struct scsi_cmnd *cmd, int reason, int unbusy)
114 {
115         struct Scsi_Host *host = cmd->device->host;
116         struct scsi_device *device = cmd->device;
117         struct scsi_target *starget = scsi_target(device);
118         struct request_queue *q = device->request_queue;
119         unsigned long flags;
120
121         SCSI_LOG_MLQUEUE(1,
122                  printk("Inserting command %p into mlqueue\n", cmd));
123
124         /*
125          * Set the appropriate busy bit for the device/host.
126          *
127          * If the host/device isn't busy, assume that something actually
128          * completed, and that we should be able to queue a command now.
129          *
130          * Note that the prior mid-layer assumption that any host could
131          * always queue at least one command is now broken.  The mid-layer
132          * will implement a user specifiable stall (see
133          * scsi_host.max_host_blocked and scsi_device.max_device_blocked)
134          * if a command is requeued with no other commands outstanding
135          * either for the device or for the host.
136          */
137         switch (reason) {
138         case SCSI_MLQUEUE_HOST_BUSY:
139                 host->host_blocked = host->max_host_blocked;
140                 break;
141         case SCSI_MLQUEUE_DEVICE_BUSY:
142                 device->device_blocked = device->max_device_blocked;
143                 break;
144         case SCSI_MLQUEUE_TARGET_BUSY:
145                 starget->target_blocked = starget->max_target_blocked;
146                 break;
147         }
148
149         /*
150          * Decrement the counters, since these commands are no longer
151          * active on the host/device.
152          */
153         if (unbusy)
154                 scsi_device_unbusy(device);
155
156         /*
157          * Requeue this command.  It will go before all other commands
158          * that are already in the queue.
159          */
160         spin_lock_irqsave(q->queue_lock, flags);
161         blk_requeue_request(q, cmd->request);
162         spin_unlock_irqrestore(q->queue_lock, flags);
163
164         scsi_run_queue(q);
165
166         return 0;
167 }
168
169 /*
170  * Function:    scsi_queue_insert()
171  *
172  * Purpose:     Insert a command in the midlevel queue.
173  *
174  * Arguments:   cmd    - command that we are adding to queue.
175  *              reason - why we are inserting command to queue.
176  *
177  * Lock status: Assumed that lock is not held upon entry.
178  *
179  * Returns:     Nothing.
180  *
181  * Notes:       We do this for one of two cases.  Either the host is busy
182  *              and it cannot accept any more commands for the time being,
183  *              or the device returned QUEUE_FULL and can accept no more
184  *              commands.
185  * Notes:       This could be called either from an interrupt context or a
186  *              normal process context.
187  */
188 int scsi_queue_insert(struct scsi_cmnd *cmd, int reason)
189 {
190         return __scsi_queue_insert(cmd, reason, 1);
191 }
192 /**
193  * scsi_execute - insert request and wait for the result
194  * @sdev:       scsi device
195  * @cmd:        scsi command
196  * @data_direction: data direction
197  * @buffer:     data buffer
198  * @bufflen:    len of buffer
199  * @sense:      optional sense buffer
200  * @timeout:    request timeout in seconds
201  * @retries:    number of times to retry request
202  * @flags:      or into request flags;
203  * @resid:      optional residual length
204  *
205  * returns the req->errors value which is the scsi_cmnd result
206  * field.
207  */
208 int scsi_execute(struct scsi_device *sdev, const unsigned char *cmd,
209                  int data_direction, void *buffer, unsigned bufflen,
210                  unsigned char *sense, int timeout, int retries, int flags,
211                  int *resid)
212 {
213         struct request *req;
214         int write = (data_direction == DMA_TO_DEVICE);
215         int ret = DRIVER_ERROR << 24;
216
217         req = blk_get_request(sdev->request_queue, write, __GFP_WAIT);
218
219         if (bufflen &&  blk_rq_map_kern(sdev->request_queue, req,
220                                         buffer, bufflen, __GFP_WAIT))
221                 goto out;
222
223         req->cmd_len = COMMAND_SIZE(cmd[0]);
224         memcpy(req->cmd, cmd, req->cmd_len);
225         req->sense = sense;
226         req->sense_len = 0;
227         req->retries = retries;
228         req->timeout = timeout;
229         req->cmd_type = REQ_TYPE_BLOCK_PC;
230         req->cmd_flags |= flags | REQ_QUIET | REQ_PREEMPT;
231
232         /*
233          * head injection *required* here otherwise quiesce won't work
234          */
235         blk_execute_rq(req->q, NULL, req, 1);
236
237         /*
238          * Some devices (USB mass-storage in particular) may transfer
239          * garbage data together with a residue indicating that the data
240          * is invalid.  Prevent the garbage from being misinterpreted
241          * and prevent security leaks by zeroing out the excess data.
242          */
243         if (unlikely(req->resid_len > 0 && req->resid_len <= bufflen))
244                 memset(buffer + (bufflen - req->resid_len), 0, req->resid_len);
245
246         if (resid)
247                 *resid = req->resid_len;
248         ret = req->errors;
249  out:
250         blk_put_request(req);
251
252         return ret;
253 }
254 EXPORT_SYMBOL(scsi_execute);
255
256
257 int scsi_execute_req(struct scsi_device *sdev, const unsigned char *cmd,
258                      int data_direction, void *buffer, unsigned bufflen,
259                      struct scsi_sense_hdr *sshdr, int timeout, int retries,
260                      int *resid)
261 {
262         char *sense = NULL;
263         int result;
264         
265         if (sshdr) {
266                 sense = kzalloc(SCSI_SENSE_BUFFERSIZE, GFP_NOIO);
267                 if (!sense)
268                         return DRIVER_ERROR << 24;
269         }
270         result = scsi_execute(sdev, cmd, data_direction, buffer, bufflen,
271                               sense, timeout, retries, 0, resid);
272         if (sshdr)
273                 scsi_normalize_sense(sense, SCSI_SENSE_BUFFERSIZE, sshdr);
274
275         kfree(sense);
276         return result;
277 }
278 EXPORT_SYMBOL(scsi_execute_req);
279
280 /*
281  * Function:    scsi_init_cmd_errh()
282  *
283  * Purpose:     Initialize cmd fields related to error handling.
284  *
285  * Arguments:   cmd     - command that is ready to be queued.
286  *
287  * Notes:       This function has the job of initializing a number of
288  *              fields related to error handling.   Typically this will
289  *              be called once for each command, as required.
290  */
291 static void scsi_init_cmd_errh(struct scsi_cmnd *cmd)
292 {
293         cmd->serial_number = 0;
294         scsi_set_resid(cmd, 0);
295         memset(cmd->sense_buffer, 0, SCSI_SENSE_BUFFERSIZE);
296         if (cmd->cmd_len == 0)
297                 cmd->cmd_len = scsi_command_size(cmd->cmnd);
298 }
299
300 void scsi_device_unbusy(struct scsi_device *sdev)
301 {
302         struct Scsi_Host *shost = sdev->host;
303         struct scsi_target *starget = scsi_target(sdev);
304         unsigned long flags;
305
306         spin_lock_irqsave(shost->host_lock, flags);
307         shost->host_busy--;
308         starget->target_busy--;
309         if (unlikely(scsi_host_in_recovery(shost) &&
310                      (shost->host_failed || shost->host_eh_scheduled)))
311                 scsi_eh_wakeup(shost);
312         spin_unlock(shost->host_lock);
313         spin_lock(sdev->request_queue->queue_lock);
314         sdev->device_busy--;
315         spin_unlock_irqrestore(sdev->request_queue->queue_lock, flags);
316 }
317
318 /*
319  * Called for single_lun devices on IO completion. Clear starget_sdev_user,
320  * and call blk_run_queue for all the scsi_devices on the target -
321  * including current_sdev first.
322  *
323  * Called with *no* scsi locks held.
324  */
325 static void scsi_single_lun_run(struct scsi_device *current_sdev)
326 {
327         struct Scsi_Host *shost = current_sdev->host;
328         struct scsi_device *sdev, *tmp;
329         struct scsi_target *starget = scsi_target(current_sdev);
330         unsigned long flags;
331
332         spin_lock_irqsave(shost->host_lock, flags);
333         starget->starget_sdev_user = NULL;
334         spin_unlock_irqrestore(shost->host_lock, flags);
335
336         /*
337          * Call blk_run_queue for all LUNs on the target, starting with
338          * current_sdev. We race with others (to set starget_sdev_user),
339          * but in most cases, we will be first. Ideally, each LU on the
340          * target would get some limited time or requests on the target.
341          */
342         blk_run_queue(current_sdev->request_queue);
343
344         spin_lock_irqsave(shost->host_lock, flags);
345         if (starget->starget_sdev_user)
346                 goto out;
347         list_for_each_entry_safe(sdev, tmp, &starget->devices,
348                         same_target_siblings) {
349                 if (sdev == current_sdev)
350                         continue;
351                 if (scsi_device_get(sdev))
352                         continue;
353
354                 spin_unlock_irqrestore(shost->host_lock, flags);
355                 blk_run_queue(sdev->request_queue);
356                 spin_lock_irqsave(shost->host_lock, flags);
357         
358                 scsi_device_put(sdev);
359         }
360  out:
361         spin_unlock_irqrestore(shost->host_lock, flags);
362 }
363
364 static inline int scsi_device_is_busy(struct scsi_device *sdev)
365 {
366         if (sdev->device_busy >= sdev->queue_depth || sdev->device_blocked)
367                 return 1;
368
369         return 0;
370 }
371
372 static inline int scsi_target_is_busy(struct scsi_target *starget)
373 {
374         return ((starget->can_queue > 0 &&
375                  starget->target_busy >= starget->can_queue) ||
376                  starget->target_blocked);
377 }
378
379 static inline int scsi_host_is_busy(struct Scsi_Host *shost)
380 {
381         if ((shost->can_queue > 0 && shost->host_busy >= shost->can_queue) ||
382             shost->host_blocked || shost->host_self_blocked)
383                 return 1;
384
385         return 0;
386 }
387
388 /*
389  * Function:    scsi_run_queue()
390  *
391  * Purpose:     Select a proper request queue to serve next
392  *
393  * Arguments:   q       - last request's queue
394  *
395  * Returns:     Nothing
396  *
397  * Notes:       The previous command was completely finished, start
398  *              a new one if possible.
399  */
400 static void scsi_run_queue(struct request_queue *q)
401 {
402         struct scsi_device *sdev = q->queuedata;
403         struct Scsi_Host *shost = sdev->host;
404         LIST_HEAD(starved_list);
405         unsigned long flags;
406
407         if (scsi_target(sdev)->single_lun)
408                 scsi_single_lun_run(sdev);
409
410         spin_lock_irqsave(shost->host_lock, flags);
411         list_splice_init(&shost->starved_list, &starved_list);
412
413         while (!list_empty(&starved_list)) {
414                 /*
415                  * As long as shost is accepting commands and we have
416                  * starved queues, call blk_run_queue. scsi_request_fn
417                  * drops the queue_lock and can add us back to the
418                  * starved_list.
419                  *
420                  * host_lock protects the starved_list and starved_entry.
421                  * scsi_request_fn must get the host_lock before checking
422                  * or modifying starved_list or starved_entry.
423                  */
424                 if (scsi_host_is_busy(shost))
425                         break;
426
427                 sdev = list_entry(starved_list.next,
428                                   struct scsi_device, starved_entry);
429                 list_del_init(&sdev->starved_entry);
430                 if (scsi_target_is_busy(scsi_target(sdev))) {
431                         list_move_tail(&sdev->starved_entry,
432                                        &shost->starved_list);
433                         continue;
434                 }
435
436                 blk_run_queue_async(sdev->request_queue);
437         }
438         /* put any unprocessed entries back */
439         list_splice(&starved_list, &shost->starved_list);
440         spin_unlock_irqrestore(shost->host_lock, flags);
441
442         blk_run_queue(q);
443 }
444
445 /*
446  * Function:    scsi_requeue_command()
447  *
448  * Purpose:     Handle post-processing of completed commands.
449  *
450  * Arguments:   q       - queue to operate on
451  *              cmd     - command that may need to be requeued.
452  *
453  * Returns:     Nothing
454  *
455  * Notes:       After command completion, there may be blocks left
456  *              over which weren't finished by the previous command
457  *              this can be for a number of reasons - the main one is
458  *              I/O errors in the middle of the request, in which case
459  *              we need to request the blocks that come after the bad
460  *              sector.
461  * Notes:       Upon return, cmd is a stale pointer.
462  */
463 static void scsi_requeue_command(struct request_queue *q, struct scsi_cmnd *cmd)
464 {
465         struct request *req = cmd->request;
466         unsigned long flags;
467
468         spin_lock_irqsave(q->queue_lock, flags);
469         scsi_unprep_request(req);
470         blk_requeue_request(q, req);
471         spin_unlock_irqrestore(q->queue_lock, flags);
472
473         scsi_run_queue(q);
474 }
475
476 void scsi_next_command(struct scsi_cmnd *cmd)
477 {
478         struct scsi_device *sdev = cmd->device;
479         struct request_queue *q = sdev->request_queue;
480
481         /* need to hold a reference on the device before we let go of the cmd */
482         get_device(&sdev->sdev_gendev);
483
484         scsi_put_command(cmd);
485         scsi_run_queue(q);
486
487         /* ok to remove device now */
488         put_device(&sdev->sdev_gendev);
489 }
490
491 void scsi_run_host_queues(struct Scsi_Host *shost)
492 {
493         struct scsi_device *sdev;
494
495         shost_for_each_device(sdev, shost)
496                 scsi_run_queue(sdev->request_queue);
497 }
498
499 static void __scsi_release_buffers(struct scsi_cmnd *, int);
500
501 /*
502  * Function:    scsi_end_request()
503  *
504  * Purpose:     Post-processing of completed commands (usually invoked at end
505  *              of upper level post-processing and scsi_io_completion).
506  *
507  * Arguments:   cmd      - command that is complete.
508  *              error    - 0 if I/O indicates success, < 0 for I/O error.
509  *              bytes    - number of bytes of completed I/O
510  *              requeue  - indicates whether we should requeue leftovers.
511  *
512  * Lock status: Assumed that lock is not held upon entry.
513  *
514  * Returns:     cmd if requeue required, NULL otherwise.
515  *
516  * Notes:       This is called for block device requests in order to
517  *              mark some number of sectors as complete.
518  * 
519  *              We are guaranteeing that the request queue will be goosed
520  *              at some point during this call.
521  * Notes:       If cmd was requeued, upon return it will be a stale pointer.
522  */
523 static struct scsi_cmnd *scsi_end_request(struct scsi_cmnd *cmd, int error,
524                                           int bytes, int requeue)
525 {
526         struct request_queue *q = cmd->device->request_queue;
527         struct request *req = cmd->request;
528
529         /*
530          * If there are blocks left over at the end, set up the command
531          * to queue the remainder of them.
532          */
533         if (blk_end_request(req, error, bytes)) {
534                 /* kill remainder if no retrys */
535                 if (error && scsi_noretry_cmd(cmd))
536                         blk_end_request_all(req, error);
537                 else {
538                         if (requeue) {
539                                 /*
540                                  * Bleah.  Leftovers again.  Stick the
541                                  * leftovers in the front of the
542                                  * queue, and goose the queue again.
543                                  */
544                                 scsi_release_buffers(cmd);
545                                 scsi_requeue_command(q, cmd);
546                                 cmd = NULL;
547                         }
548                         return cmd;
549                 }
550         }
551
552         /*
553          * This will goose the queue request function at the end, so we don't
554          * need to worry about launching another command.
555          */
556         __scsi_release_buffers(cmd, 0);
557         scsi_next_command(cmd);
558         return NULL;
559 }
560
561 static inline unsigned int scsi_sgtable_index(unsigned short nents)
562 {
563         unsigned int index;
564
565         BUG_ON(nents > SCSI_MAX_SG_SEGMENTS);
566
567         if (nents <= 8)
568                 index = 0;
569         else
570                 index = get_count_order(nents) - 3;
571
572         return index;
573 }
574
575 static void scsi_sg_free(struct scatterlist *sgl, unsigned int nents)
576 {
577         struct scsi_host_sg_pool *sgp;
578
579         sgp = scsi_sg_pools + scsi_sgtable_index(nents);
580         mempool_free(sgl, sgp->pool);
581 }
582
583 static struct scatterlist *scsi_sg_alloc(unsigned int nents, gfp_t gfp_mask)
584 {
585         struct scsi_host_sg_pool *sgp;
586
587         sgp = scsi_sg_pools + scsi_sgtable_index(nents);
588         return mempool_alloc(sgp->pool, gfp_mask);
589 }
590
591 static int scsi_alloc_sgtable(struct scsi_data_buffer *sdb, int nents,
592                               gfp_t gfp_mask)
593 {
594         int ret;
595
596         BUG_ON(!nents);
597
598         ret = __sg_alloc_table(&sdb->table, nents, SCSI_MAX_SG_SEGMENTS,
599                                gfp_mask, scsi_sg_alloc);
600         if (unlikely(ret))
601                 __sg_free_table(&sdb->table, SCSI_MAX_SG_SEGMENTS,
602                                 scsi_sg_free);
603
604         return ret;
605 }
606
607 static void scsi_free_sgtable(struct scsi_data_buffer *sdb)
608 {
609         __sg_free_table(&sdb->table, SCSI_MAX_SG_SEGMENTS, scsi_sg_free);
610 }
611
612 static void __scsi_release_buffers(struct scsi_cmnd *cmd, int do_bidi_check)
613 {
614
615         if (cmd->sdb.table.nents)
616                 scsi_free_sgtable(&cmd->sdb);
617
618         memset(&cmd->sdb, 0, sizeof(cmd->sdb));
619
620         if (do_bidi_check && scsi_bidi_cmnd(cmd)) {
621                 struct scsi_data_buffer *bidi_sdb =
622                         cmd->request->next_rq->special;
623                 scsi_free_sgtable(bidi_sdb);
624                 kmem_cache_free(scsi_sdb_cache, bidi_sdb);
625                 cmd->request->next_rq->special = NULL;
626         }
627
628         if (scsi_prot_sg_count(cmd))
629                 scsi_free_sgtable(cmd->prot_sdb);
630 }
631
632 /*
633  * Function:    scsi_release_buffers()
634  *
635  * Purpose:     Completion processing for block device I/O requests.
636  *
637  * Arguments:   cmd     - command that we are bailing.
638  *
639  * Lock status: Assumed that no lock is held upon entry.
640  *
641  * Returns:     Nothing
642  *
643  * Notes:       In the event that an upper level driver rejects a
644  *              command, we must release resources allocated during
645  *              the __init_io() function.  Primarily this would involve
646  *              the scatter-gather table, and potentially any bounce
647  *              buffers.
648  */
649 void scsi_release_buffers(struct scsi_cmnd *cmd)
650 {
651         __scsi_release_buffers(cmd, 1);
652 }
653 EXPORT_SYMBOL(scsi_release_buffers);
654
655 static int __scsi_error_from_host_byte(struct scsi_cmnd *cmd, int result)
656 {
657         int error = 0;
658
659         switch(host_byte(result)) {
660         case DID_TRANSPORT_FAILFAST:
661                 error = -ENOLINK;
662                 break;
663         case DID_TARGET_FAILURE:
664                 cmd->result |= (DID_OK << 16);
665                 error = -EREMOTEIO;
666                 break;
667         case DID_NEXUS_FAILURE:
668                 cmd->result |= (DID_OK << 16);
669                 error = -EBADE;
670                 break;
671         default:
672                 error = -EIO;
673                 break;
674         }
675
676         return error;
677 }
678
679 /*
680  * Function:    scsi_io_completion()
681  *
682  * Purpose:     Completion processing for block device I/O requests.
683  *
684  * Arguments:   cmd   - command that is finished.
685  *
686  * Lock status: Assumed that no lock is held upon entry.
687  *
688  * Returns:     Nothing
689  *
690  * Notes:       This function is matched in terms of capabilities to
691  *              the function that created the scatter-gather list.
692  *              In other words, if there are no bounce buffers
693  *              (the normal case for most drivers), we don't need
694  *              the logic to deal with cleaning up afterwards.
695  *
696  *              We must call scsi_end_request().  This will finish off
697  *              the specified number of sectors.  If we are done, the
698  *              command block will be released and the queue function
699  *              will be goosed.  If we are not done then we have to
700  *              figure out what to do next:
701  *
702  *              a) We can call scsi_requeue_command().  The request
703  *                 will be unprepared and put back on the queue.  Then
704  *                 a new command will be created for it.  This should
705  *                 be used if we made forward progress, or if we want
706  *                 to switch from READ(10) to READ(6) for example.
707  *
708  *              b) We can call scsi_queue_insert().  The request will
709  *                 be put back on the queue and retried using the same
710  *                 command as before, possibly after a delay.
711  *
712  *              c) We can call blk_end_request() with -EIO to fail
713  *                 the remainder of the request.
714  */
715 void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes)
716 {
717         int result = cmd->result;
718         struct request_queue *q = cmd->device->request_queue;
719         struct request *req = cmd->request;
720         int error = 0;
721         struct scsi_sense_hdr sshdr;
722         int sense_valid = 0;
723         int sense_deferred = 0;
724         enum {ACTION_FAIL, ACTION_REPREP, ACTION_RETRY,
725               ACTION_DELAYED_RETRY} action;
726         char *description = NULL;
727
728         if (result) {
729                 sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
730                 if (sense_valid)
731                         sense_deferred = scsi_sense_is_deferred(&sshdr);
732         }
733
734         if (req->cmd_type == REQ_TYPE_BLOCK_PC) { /* SG_IO ioctl from block level */
735                 req->errors = result;
736                 if (result) {
737                         if (sense_valid && req->sense) {
738                                 /*
739                                  * SG_IO wants current and deferred errors
740                                  */
741                                 int len = 8 + cmd->sense_buffer[7];
742
743                                 if (len > SCSI_SENSE_BUFFERSIZE)
744                                         len = SCSI_SENSE_BUFFERSIZE;
745                                 memcpy(req->sense, cmd->sense_buffer,  len);
746                                 req->sense_len = len;
747                         }
748                         if (!sense_deferred)
749                                 error = __scsi_error_from_host_byte(cmd, result);
750                 }
751
752                 req->resid_len = scsi_get_resid(cmd);
753
754                 if (scsi_bidi_cmnd(cmd)) {
755                         /*
756                          * Bidi commands Must be complete as a whole,
757                          * both sides at once.
758                          */
759                         req->next_rq->resid_len = scsi_in(cmd)->resid;
760
761                         scsi_release_buffers(cmd);
762                         blk_end_request_all(req, 0);
763
764                         scsi_next_command(cmd);
765                         return;
766                 }
767         }
768
769         /* no bidi support for !REQ_TYPE_BLOCK_PC yet */
770         BUG_ON(blk_bidi_rq(req));
771
772         /*
773          * Next deal with any sectors which we were able to correctly
774          * handle.
775          */
776         SCSI_LOG_HLCOMPLETE(1, printk("%u sectors total, "
777                                       "%d bytes done.\n",
778                                       blk_rq_sectors(req), good_bytes));
779
780         /*
781          * Recovered errors need reporting, but they're always treated
782          * as success, so fiddle the result code here.  For BLOCK_PC
783          * we already took a copy of the original into rq->errors which
784          * is what gets returned to the user
785          */
786         if (sense_valid && (sshdr.sense_key == RECOVERED_ERROR)) {
787                 /* if ATA PASS-THROUGH INFORMATION AVAILABLE skip
788                  * print since caller wants ATA registers. Only occurs on
789                  * SCSI ATA PASS_THROUGH commands when CK_COND=1
790                  */
791                 if ((sshdr.asc == 0x0) && (sshdr.ascq == 0x1d))
792                         ;
793                 else if (!(req->cmd_flags & REQ_QUIET))
794                         scsi_print_sense("", cmd);
795                 result = 0;
796                 /* BLOCK_PC may have set error */
797                 error = 0;
798         }
799
800         /*
801          * A number of bytes were successfully read.  If there
802          * are leftovers and there is some kind of error
803          * (result != 0), retry the rest.
804          */
805         if (scsi_end_request(cmd, error, good_bytes, result == 0) == NULL)
806                 return;
807
808         error = __scsi_error_from_host_byte(cmd, result);
809
810         if (host_byte(result) == DID_RESET) {
811                 /* Third party bus reset or reset for error recovery
812                  * reasons.  Just retry the command and see what
813                  * happens.
814                  */
815                 action = ACTION_RETRY;
816         } else if (sense_valid && !sense_deferred) {
817                 switch (sshdr.sense_key) {
818                 case UNIT_ATTENTION:
819                         if (cmd->device->removable) {
820                                 /* Detected disc change.  Set a bit
821                                  * and quietly refuse further access.
822                                  */
823                                 cmd->device->changed = 1;
824                                 description = "Media Changed";
825                                 action = ACTION_FAIL;
826                         } else {
827                                 /* Must have been a power glitch, or a
828                                  * bus reset.  Could not have been a
829                                  * media change, so we just retry the
830                                  * command and see what happens.
831                                  */
832                                 action = ACTION_RETRY;
833                         }
834                         break;
835                 case ILLEGAL_REQUEST:
836                         /* If we had an ILLEGAL REQUEST returned, then
837                          * we may have performed an unsupported
838                          * command.  The only thing this should be
839                          * would be a ten byte read where only a six
840                          * byte read was supported.  Also, on a system
841                          * where READ CAPACITY failed, we may have
842                          * read past the end of the disk.
843                          */
844                         if ((cmd->device->use_10_for_rw &&
845                             sshdr.asc == 0x20 && sshdr.ascq == 0x00) &&
846                             (cmd->cmnd[0] == READ_10 ||
847                              cmd->cmnd[0] == WRITE_10)) {
848                                 /* This will issue a new 6-byte command. */
849                                 cmd->device->use_10_for_rw = 0;
850                                 action = ACTION_REPREP;
851                         } else if (sshdr.asc == 0x10) /* DIX */ {
852                                 description = "Host Data Integrity Failure";
853                                 action = ACTION_FAIL;
854                                 error = -EILSEQ;
855                         /* INVALID COMMAND OPCODE or INVALID FIELD IN CDB */
856                         } else if ((sshdr.asc == 0x20 || sshdr.asc == 0x24) &&
857                                    (cmd->cmnd[0] == UNMAP ||
858                                     cmd->cmnd[0] == WRITE_SAME_16 ||
859                                     cmd->cmnd[0] == WRITE_SAME)) {
860                                 description = "Discard failure";
861                                 action = ACTION_FAIL;
862                         } else
863                                 action = ACTION_FAIL;
864                         break;
865                 case ABORTED_COMMAND:
866                         action = ACTION_FAIL;
867                         if (sshdr.asc == 0x10) { /* DIF */
868                                 description = "Target Data Integrity Failure";
869                                 error = -EILSEQ;
870                         }
871                         break;
872                 case NOT_READY:
873                         /* If the device is in the process of becoming
874                          * ready, or has a temporary blockage, retry.
875                          */
876                         if (sshdr.asc == 0x04) {
877                                 switch (sshdr.ascq) {
878                                 case 0x01: /* becoming ready */
879                                 case 0x04: /* format in progress */
880                                 case 0x05: /* rebuild in progress */
881                                 case 0x06: /* recalculation in progress */
882                                 case 0x07: /* operation in progress */
883                                 case 0x08: /* Long write in progress */
884                                 case 0x09: /* self test in progress */
885                                 case 0x14: /* space allocation in progress */
886                                         action = ACTION_DELAYED_RETRY;
887                                         break;
888                                 default:
889                                         description = "Device not ready";
890                                         action = ACTION_FAIL;
891                                         break;
892                                 }
893                         } else {
894                                 description = "Device not ready";
895                                 action = ACTION_FAIL;
896                         }
897                         break;
898                 case VOLUME_OVERFLOW:
899                         /* See SSC3rXX or current. */
900                         action = ACTION_FAIL;
901                         break;
902                 default:
903                         description = "Unhandled sense code";
904                         action = ACTION_FAIL;
905                         break;
906                 }
907         } else {
908                 description = "Unhandled error code";
909                 action = ACTION_FAIL;
910         }
911
912         switch (action) {
913         case ACTION_FAIL:
914                 /* Give up and fail the remainder of the request */
915                 scsi_release_buffers(cmd);
916                 if (!(req->cmd_flags & REQ_QUIET)) {
917                         if (description)
918                                 scmd_printk(KERN_INFO, cmd, "%s\n",
919                                             description);
920                         scsi_print_result(cmd);
921                         if (driver_byte(result) & DRIVER_SENSE)
922                                 scsi_print_sense("", cmd);
923                         scsi_print_command(cmd);
924                 }
925                 if (blk_end_request_err(req, error))
926                         scsi_requeue_command(q, cmd);
927                 else
928                         scsi_next_command(cmd);
929                 break;
930         case ACTION_REPREP:
931                 /* Unprep the request and put it back at the head of the queue.
932                  * A new command will be prepared and issued.
933                  */
934                 scsi_release_buffers(cmd);
935                 scsi_requeue_command(q, cmd);
936                 break;
937         case ACTION_RETRY:
938                 /* Retry the same command immediately */
939                 __scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY, 0);
940                 break;
941         case ACTION_DELAYED_RETRY:
942                 /* Retry the same command after a delay */
943                 __scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY, 0);
944                 break;
945         }
946 }
947
948 static int scsi_init_sgtable(struct request *req, struct scsi_data_buffer *sdb,
949                              gfp_t gfp_mask)
950 {
951         int count;
952
953         /*
954          * If sg table allocation fails, requeue request later.
955          */
956         if (unlikely(scsi_alloc_sgtable(sdb, req->nr_phys_segments,
957                                         gfp_mask))) {
958                 return BLKPREP_DEFER;
959         }
960
961         req->buffer = NULL;
962
963         /* 
964          * Next, walk the list, and fill in the addresses and sizes of
965          * each segment.
966          */
967         count = blk_rq_map_sg(req->q, req, sdb->table.sgl);
968         BUG_ON(count > sdb->table.nents);
969         sdb->table.nents = count;
970         sdb->length = blk_rq_bytes(req);
971         return BLKPREP_OK;
972 }
973
974 /*
975  * Function:    scsi_init_io()
976  *
977  * Purpose:     SCSI I/O initialize function.
978  *
979  * Arguments:   cmd   - Command descriptor we wish to initialize
980  *
981  * Returns:     0 on success
982  *              BLKPREP_DEFER if the failure is retryable
983  *              BLKPREP_KILL if the failure is fatal
984  */
985 int scsi_init_io(struct scsi_cmnd *cmd, gfp_t gfp_mask)
986 {
987         struct request *rq = cmd->request;
988
989         int error = scsi_init_sgtable(rq, &cmd->sdb, gfp_mask);
990         if (error)
991                 goto err_exit;
992
993         if (blk_bidi_rq(rq)) {
994                 struct scsi_data_buffer *bidi_sdb = kmem_cache_zalloc(
995                         scsi_sdb_cache, GFP_ATOMIC);
996                 if (!bidi_sdb) {
997                         error = BLKPREP_DEFER;
998                         goto err_exit;
999                 }
1000
1001                 rq->next_rq->special = bidi_sdb;
1002                 error = scsi_init_sgtable(rq->next_rq, bidi_sdb, GFP_ATOMIC);
1003                 if (error)
1004                         goto err_exit;
1005         }
1006
1007         if (blk_integrity_rq(rq)) {
1008                 struct scsi_data_buffer *prot_sdb = cmd->prot_sdb;
1009                 int ivecs, count;
1010
1011                 BUG_ON(prot_sdb == NULL);
1012                 ivecs = blk_rq_count_integrity_sg(rq->q, rq->bio);
1013
1014                 if (scsi_alloc_sgtable(prot_sdb, ivecs, gfp_mask)) {
1015                         error = BLKPREP_DEFER;
1016                         goto err_exit;
1017                 }
1018
1019                 count = blk_rq_map_integrity_sg(rq->q, rq->bio,
1020                                                 prot_sdb->table.sgl);
1021                 BUG_ON(unlikely(count > ivecs));
1022                 BUG_ON(unlikely(count > queue_max_integrity_segments(rq->q)));
1023
1024                 cmd->prot_sdb = prot_sdb;
1025                 cmd->prot_sdb->table.nents = count;
1026         }
1027
1028         return BLKPREP_OK ;
1029
1030 err_exit:
1031         scsi_release_buffers(cmd);
1032         cmd->request->special = NULL;
1033         scsi_put_command(cmd);
1034         return error;
1035 }
1036 EXPORT_SYMBOL(scsi_init_io);
1037
1038 static struct scsi_cmnd *scsi_get_cmd_from_req(struct scsi_device *sdev,
1039                 struct request *req)
1040 {
1041         struct scsi_cmnd *cmd;
1042
1043         if (!req->special) {
1044                 cmd = scsi_get_command(sdev, GFP_ATOMIC);
1045                 if (unlikely(!cmd))
1046                         return NULL;
1047                 req->special = cmd;
1048         } else {
1049                 cmd = req->special;
1050         }
1051
1052         /* pull a tag out of the request if we have one */
1053         cmd->tag = req->tag;
1054         cmd->request = req;
1055
1056         cmd->cmnd = req->cmd;
1057         cmd->prot_op = SCSI_PROT_NORMAL;
1058
1059         return cmd;
1060 }
1061
1062 int scsi_setup_blk_pc_cmnd(struct scsi_device *sdev, struct request *req)
1063 {
1064         struct scsi_cmnd *cmd;
1065         int ret = scsi_prep_state_check(sdev, req);
1066
1067         if (ret != BLKPREP_OK)
1068                 return ret;
1069
1070         cmd = scsi_get_cmd_from_req(sdev, req);
1071         if (unlikely(!cmd))
1072                 return BLKPREP_DEFER;
1073
1074         /*
1075          * BLOCK_PC requests may transfer data, in which case they must
1076          * a bio attached to them.  Or they might contain a SCSI command
1077          * that does not transfer data, in which case they may optionally
1078          * submit a request without an attached bio.
1079          */
1080         if (req->bio) {
1081                 int ret;
1082
1083                 BUG_ON(!req->nr_phys_segments);
1084
1085                 ret = scsi_init_io(cmd, GFP_ATOMIC);
1086                 if (unlikely(ret))
1087                         return ret;
1088         } else {
1089                 BUG_ON(blk_rq_bytes(req));
1090
1091                 memset(&cmd->sdb, 0, sizeof(cmd->sdb));
1092                 req->buffer = NULL;
1093         }
1094
1095         cmd->cmd_len = req->cmd_len;
1096         if (!blk_rq_bytes(req))
1097                 cmd->sc_data_direction = DMA_NONE;
1098         else if (rq_data_dir(req) == WRITE)
1099                 cmd->sc_data_direction = DMA_TO_DEVICE;
1100         else
1101                 cmd->sc_data_direction = DMA_FROM_DEVICE;
1102         
1103         cmd->transfersize = blk_rq_bytes(req);
1104         cmd->allowed = req->retries;
1105         return BLKPREP_OK;
1106 }
1107 EXPORT_SYMBOL(scsi_setup_blk_pc_cmnd);
1108
1109 /*
1110  * Setup a REQ_TYPE_FS command.  These are simple read/write request
1111  * from filesystems that still need to be translated to SCSI CDBs from
1112  * the ULD.
1113  */
1114 int scsi_setup_fs_cmnd(struct scsi_device *sdev, struct request *req)
1115 {
1116         struct scsi_cmnd *cmd;
1117         int ret = scsi_prep_state_check(sdev, req);
1118
1119         if (ret != BLKPREP_OK)
1120                 return ret;
1121
1122         if (unlikely(sdev->scsi_dh_data && sdev->scsi_dh_data->scsi_dh
1123                          && sdev->scsi_dh_data->scsi_dh->prep_fn)) {
1124                 ret = sdev->scsi_dh_data->scsi_dh->prep_fn(sdev, req);
1125                 if (ret != BLKPREP_OK)
1126                         return ret;
1127         }
1128
1129         /*
1130          * Filesystem requests must transfer data.
1131          */
1132         BUG_ON(!req->nr_phys_segments);
1133
1134         cmd = scsi_get_cmd_from_req(sdev, req);
1135         if (unlikely(!cmd))
1136                 return BLKPREP_DEFER;
1137
1138         memset(cmd->cmnd, 0, BLK_MAX_CDB);
1139         return scsi_init_io(cmd, GFP_ATOMIC);
1140 }
1141 EXPORT_SYMBOL(scsi_setup_fs_cmnd);
1142
1143 int scsi_prep_state_check(struct scsi_device *sdev, struct request *req)
1144 {
1145         int ret = BLKPREP_OK;
1146
1147         /*
1148          * If the device is not in running state we will reject some
1149          * or all commands.
1150          */
1151         if (unlikely(sdev->sdev_state != SDEV_RUNNING)) {
1152                 switch (sdev->sdev_state) {
1153                 case SDEV_OFFLINE:
1154                         /*
1155                          * If the device is offline we refuse to process any
1156                          * commands.  The device must be brought online
1157                          * before trying any recovery commands.
1158                          */
1159                         sdev_printk(KERN_ERR, sdev,
1160                                     "rejecting I/O to offline device\n");
1161                         ret = BLKPREP_KILL;
1162                         break;
1163                 case SDEV_DEL:
1164                         /*
1165                          * If the device is fully deleted, we refuse to
1166                          * process any commands as well.
1167                          */
1168                         sdev_printk(KERN_ERR, sdev,
1169                                     "rejecting I/O to dead device\n");
1170                         ret = BLKPREP_KILL;
1171                         break;
1172                 case SDEV_QUIESCE:
1173                 case SDEV_BLOCK:
1174                 case SDEV_CREATED_BLOCK:
1175                         /*
1176                          * If the devices is blocked we defer normal commands.
1177                          */
1178                         if (!(req->cmd_flags & REQ_PREEMPT))
1179                                 ret = BLKPREP_DEFER;
1180                         break;
1181                 default:
1182                         /*
1183                          * For any other not fully online state we only allow
1184                          * special commands.  In particular any user initiated
1185                          * command is not allowed.
1186                          */
1187                         if (!(req->cmd_flags & REQ_PREEMPT))
1188                                 ret = BLKPREP_KILL;
1189                         break;
1190                 }
1191         }
1192         return ret;
1193 }
1194 EXPORT_SYMBOL(scsi_prep_state_check);
1195
1196 int scsi_prep_return(struct request_queue *q, struct request *req, int ret)
1197 {
1198         struct scsi_device *sdev = q->queuedata;
1199
1200         switch (ret) {
1201         case BLKPREP_KILL:
1202                 req->errors = DID_NO_CONNECT << 16;
1203                 /* release the command and kill it */
1204                 if (req->special) {
1205                         struct scsi_cmnd *cmd = req->special;
1206                         scsi_release_buffers(cmd);
1207                         scsi_put_command(cmd);
1208                         req->special = NULL;
1209                 }
1210                 break;
1211         case BLKPREP_DEFER:
1212                 /*
1213                  * If we defer, the blk_peek_request() returns NULL, but the
1214                  * queue must be restarted, so we schedule a callback to happen
1215                  * shortly.
1216                  */
1217                 if (sdev->device_busy == 0)
1218                         blk_delay_queue(q, SCSI_QUEUE_DELAY);
1219                 break;
1220         default:
1221                 req->cmd_flags |= REQ_DONTPREP;
1222         }
1223
1224         return ret;
1225 }
1226 EXPORT_SYMBOL(scsi_prep_return);
1227
1228 int scsi_prep_fn(struct request_queue *q, struct request *req)
1229 {
1230         struct scsi_device *sdev = q->queuedata;
1231         int ret = BLKPREP_KILL;
1232
1233         if (req->cmd_type == REQ_TYPE_BLOCK_PC)
1234                 ret = scsi_setup_blk_pc_cmnd(sdev, req);
1235         return scsi_prep_return(q, req, ret);
1236 }
1237 EXPORT_SYMBOL(scsi_prep_fn);
1238
1239 /*
1240  * scsi_dev_queue_ready: if we can send requests to sdev, return 1 else
1241  * return 0.
1242  *
1243  * Called with the queue_lock held.
1244  */
1245 static inline int scsi_dev_queue_ready(struct request_queue *q,
1246                                   struct scsi_device *sdev)
1247 {
1248         if (sdev->device_busy == 0 && sdev->device_blocked) {
1249                 /*
1250                  * unblock after device_blocked iterates to zero
1251                  */
1252                 if (--sdev->device_blocked == 0) {
1253                         SCSI_LOG_MLQUEUE(3,
1254                                    sdev_printk(KERN_INFO, sdev,
1255                                    "unblocking device at zero depth\n"));
1256                 } else {
1257                         blk_delay_queue(q, SCSI_QUEUE_DELAY);
1258                         return 0;
1259                 }
1260         }
1261         if (scsi_device_is_busy(sdev))
1262                 return 0;
1263
1264         return 1;
1265 }
1266
1267
1268 /*
1269  * scsi_target_queue_ready: checks if there we can send commands to target
1270  * @sdev: scsi device on starget to check.
1271  *
1272  * Called with the host lock held.
1273  */
1274 static inline int scsi_target_queue_ready(struct Scsi_Host *shost,
1275                                            struct scsi_device *sdev)
1276 {
1277         struct scsi_target *starget = scsi_target(sdev);
1278
1279         if (starget->single_lun) {
1280                 if (starget->starget_sdev_user &&
1281                     starget->starget_sdev_user != sdev)
1282                         return 0;
1283                 starget->starget_sdev_user = sdev;
1284         }
1285
1286         if (starget->target_busy == 0 && starget->target_blocked) {
1287                 /*
1288                  * unblock after target_blocked iterates to zero
1289                  */
1290                 if (--starget->target_blocked == 0) {
1291                         SCSI_LOG_MLQUEUE(3, starget_printk(KERN_INFO, starget,
1292                                          "unblocking target at zero depth\n"));
1293                 } else
1294                         return 0;
1295         }
1296
1297         if (scsi_target_is_busy(starget)) {
1298                 if (list_empty(&sdev->starved_entry))
1299                         list_add_tail(&sdev->starved_entry,
1300                                       &shost->starved_list);
1301                 return 0;
1302         }
1303
1304         /* We're OK to process the command, so we can't be starved */
1305         if (!list_empty(&sdev->starved_entry))
1306                 list_del_init(&sdev->starved_entry);
1307         return 1;
1308 }
1309
1310 /*
1311  * scsi_host_queue_ready: if we can send requests to shost, return 1 else
1312  * return 0. We must end up running the queue again whenever 0 is
1313  * returned, else IO can hang.
1314  *
1315  * Called with host_lock held.
1316  */
1317 static inline int scsi_host_queue_ready(struct request_queue *q,
1318                                    struct Scsi_Host *shost,
1319                                    struct scsi_device *sdev)
1320 {
1321         if (scsi_host_in_recovery(shost))
1322                 return 0;
1323         if (shost->host_busy == 0 && shost->host_blocked) {
1324                 /*
1325                  * unblock after host_blocked iterates to zero
1326                  */
1327                 if (--shost->host_blocked == 0) {
1328                         SCSI_LOG_MLQUEUE(3,
1329                                 printk("scsi%d unblocking host at zero depth\n",
1330                                         shost->host_no));
1331                 } else {
1332                         return 0;
1333                 }
1334         }
1335         if (scsi_host_is_busy(shost)) {
1336                 if (list_empty(&sdev->starved_entry))
1337                         list_add_tail(&sdev->starved_entry, &shost->starved_list);
1338                 return 0;
1339         }
1340
1341         /* We're OK to process the command, so we can't be starved */
1342         if (!list_empty(&sdev->starved_entry))
1343                 list_del_init(&sdev->starved_entry);
1344
1345         return 1;
1346 }
1347
1348 /*
1349  * Busy state exporting function for request stacking drivers.
1350  *
1351  * For efficiency, no lock is taken to check the busy state of
1352  * shost/starget/sdev, since the returned value is not guaranteed and
1353  * may be changed after request stacking drivers call the function,
1354  * regardless of taking lock or not.
1355  *
1356  * When scsi can't dispatch I/Os anymore and needs to kill I/Os
1357  * (e.g. !sdev), scsi needs to return 'not busy'.
1358  * Otherwise, request stacking drivers may hold requests forever.
1359  */
1360 static int scsi_lld_busy(struct request_queue *q)
1361 {
1362         struct scsi_device *sdev = q->queuedata;
1363         struct Scsi_Host *shost;
1364         struct scsi_target *starget;
1365
1366         if (!sdev)
1367                 return 0;
1368
1369         shost = sdev->host;
1370         starget = scsi_target(sdev);
1371
1372         if (scsi_host_in_recovery(shost) || scsi_host_is_busy(shost) ||
1373             scsi_target_is_busy(starget) || scsi_device_is_busy(sdev))
1374                 return 1;
1375
1376         return 0;
1377 }
1378
1379 /*
1380  * Kill a request for a dead device
1381  */
1382 static void scsi_kill_request(struct request *req, struct request_queue *q)
1383 {
1384         struct scsi_cmnd *cmd = req->special;
1385         struct scsi_device *sdev;
1386         struct scsi_target *starget;
1387         struct Scsi_Host *shost;
1388
1389         blk_start_request(req);
1390
1391         sdev = cmd->device;
1392         starget = scsi_target(sdev);
1393         shost = sdev->host;
1394         scsi_init_cmd_errh(cmd);
1395         cmd->result = DID_NO_CONNECT << 16;
1396         atomic_inc(&cmd->device->iorequest_cnt);
1397
1398         /*
1399          * SCSI request completion path will do scsi_device_unbusy(),
1400          * bump busy counts.  To bump the counters, we need to dance
1401          * with the locks as normal issue path does.
1402          */
1403         sdev->device_busy++;
1404         spin_unlock(sdev->request_queue->queue_lock);
1405         spin_lock(shost->host_lock);
1406         shost->host_busy++;
1407         starget->target_busy++;
1408         spin_unlock(shost->host_lock);
1409         spin_lock(sdev->request_queue->queue_lock);
1410
1411         blk_complete_request(req);
1412 }
1413
1414 static void scsi_softirq_done(struct request *rq)
1415 {
1416         struct scsi_cmnd *cmd = rq->special;
1417         unsigned long wait_for = (cmd->allowed + 1) * rq->timeout;
1418         int disposition;
1419
1420         INIT_LIST_HEAD(&cmd->eh_entry);
1421
1422         atomic_inc(&cmd->device->iodone_cnt);
1423         if (cmd->result)
1424                 atomic_inc(&cmd->device->ioerr_cnt);
1425
1426         disposition = scsi_decide_disposition(cmd);
1427         if (disposition != SUCCESS &&
1428             time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) {
1429                 sdev_printk(KERN_ERR, cmd->device,
1430                             "timing out command, waited %lus\n",
1431                             wait_for/HZ);
1432                 disposition = SUCCESS;
1433         }
1434                         
1435         scsi_log_completion(cmd, disposition);
1436
1437         switch (disposition) {
1438                 case SUCCESS:
1439                         scsi_finish_command(cmd);
1440                         break;
1441                 case NEEDS_RETRY:
1442                         scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY);
1443                         break;
1444                 case ADD_TO_MLQUEUE:
1445                         scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY);
1446                         break;
1447                 default:
1448                         if (!scsi_eh_scmd_add(cmd, 0))
1449                                 scsi_finish_command(cmd);
1450         }
1451 }
1452
1453 /*
1454  * Function:    scsi_request_fn()
1455  *
1456  * Purpose:     Main strategy routine for SCSI.
1457  *
1458  * Arguments:   q       - Pointer to actual queue.
1459  *
1460  * Returns:     Nothing
1461  *
1462  * Lock status: IO request lock assumed to be held when called.
1463  */
1464 static void scsi_request_fn(struct request_queue *q)
1465 {
1466         struct scsi_device *sdev = q->queuedata;
1467         struct Scsi_Host *shost;
1468         struct scsi_cmnd *cmd;
1469         struct request *req;
1470
1471         if (!sdev) {
1472                 printk("scsi: killing requests for dead queue\n");
1473                 while ((req = blk_peek_request(q)) != NULL)
1474                         scsi_kill_request(req, q);
1475                 return;
1476         }
1477
1478         if(!get_device(&sdev->sdev_gendev))
1479                 /* We must be tearing the block queue down already */
1480                 return;
1481
1482         /*
1483          * To start with, we keep looping until the queue is empty, or until
1484          * the host is no longer able to accept any more requests.
1485          */
1486         shost = sdev->host;
1487         for (;;) {
1488                 int rtn;
1489                 /*
1490                  * get next queueable request.  We do this early to make sure
1491                  * that the request is fully prepared even if we cannot 
1492                  * accept it.
1493                  */
1494                 req = blk_peek_request(q);
1495                 if (!req || !scsi_dev_queue_ready(q, sdev))
1496                         break;
1497
1498                 if (unlikely(!scsi_device_online(sdev))) {
1499                         sdev_printk(KERN_ERR, sdev,
1500                                     "rejecting I/O to offline device\n");
1501                         scsi_kill_request(req, q);
1502                         continue;
1503                 }
1504
1505
1506                 /*
1507                  * Remove the request from the request list.
1508                  */
1509                 if (!(blk_queue_tagged(q) && !blk_queue_start_tag(q, req)))
1510                         blk_start_request(req);
1511                 sdev->device_busy++;
1512
1513                 spin_unlock(q->queue_lock);
1514                 cmd = req->special;
1515                 if (unlikely(cmd == NULL)) {
1516                         printk(KERN_CRIT "impossible request in %s.\n"
1517                                          "please mail a stack trace to "
1518                                          "linux-scsi@vger.kernel.org\n",
1519                                          __func__);
1520                         blk_dump_rq_flags(req, "foo");
1521                         BUG();
1522                 }
1523                 spin_lock(shost->host_lock);
1524
1525                 /*
1526                  * We hit this when the driver is using a host wide
1527                  * tag map. For device level tag maps the queue_depth check
1528                  * in the device ready fn would prevent us from trying
1529                  * to allocate a tag. Since the map is a shared host resource
1530                  * we add the dev to the starved list so it eventually gets
1531                  * a run when a tag is freed.
1532                  */
1533                 if (blk_queue_tagged(q) && !blk_rq_tagged(req)) {
1534                         if (list_empty(&sdev->starved_entry))
1535                                 list_add_tail(&sdev->starved_entry,
1536                                               &shost->starved_list);
1537                         goto not_ready;
1538                 }
1539
1540                 if (!scsi_target_queue_ready(shost, sdev))
1541                         goto not_ready;
1542
1543                 if (!scsi_host_queue_ready(q, shost, sdev))
1544                         goto not_ready;
1545
1546                 scsi_target(sdev)->target_busy++;
1547                 shost->host_busy++;
1548
1549                 /*
1550                  * XXX(hch): This is rather suboptimal, scsi_dispatch_cmd will
1551                  *              take the lock again.
1552                  */
1553                 spin_unlock_irq(shost->host_lock);
1554
1555                 /*
1556                  * Finally, initialize any error handling parameters, and set up
1557                  * the timers for timeouts.
1558                  */
1559                 scsi_init_cmd_errh(cmd);
1560
1561                 /*
1562                  * Dispatch the command to the low-level driver.
1563                  */
1564                 rtn = scsi_dispatch_cmd(cmd);
1565                 spin_lock_irq(q->queue_lock);
1566                 if (rtn)
1567                         goto out_delay;
1568         }
1569
1570         goto out;
1571
1572  not_ready:
1573         spin_unlock_irq(shost->host_lock);
1574
1575         /*
1576          * lock q, handle tag, requeue req, and decrement device_busy. We
1577          * must return with queue_lock held.
1578          *
1579          * Decrementing device_busy without checking it is OK, as all such
1580          * cases (host limits or settings) should run the queue at some
1581          * later time.
1582          */
1583         spin_lock_irq(q->queue_lock);
1584         blk_requeue_request(q, req);
1585         sdev->device_busy--;
1586 out_delay:
1587         if (sdev->device_busy == 0)
1588                 blk_delay_queue(q, SCSI_QUEUE_DELAY);
1589 out:
1590         /* must be careful here...if we trigger the ->remove() function
1591          * we cannot be holding the q lock */
1592         spin_unlock_irq(q->queue_lock);
1593         put_device(&sdev->sdev_gendev);
1594         spin_lock_irq(q->queue_lock);
1595 }
1596
1597 u64 scsi_calculate_bounce_limit(struct Scsi_Host *shost)
1598 {
1599         struct device *host_dev;
1600         u64 bounce_limit = 0xffffffff;
1601
1602         if (shost->unchecked_isa_dma)
1603                 return BLK_BOUNCE_ISA;
1604         /*
1605          * Platforms with virtual-DMA translation
1606          * hardware have no practical limit.
1607          */
1608         if (!PCI_DMA_BUS_IS_PHYS)
1609                 return BLK_BOUNCE_ANY;
1610
1611         host_dev = scsi_get_device(shost);
1612         if (host_dev && host_dev->dma_mask)
1613                 bounce_limit = *host_dev->dma_mask;
1614
1615         return bounce_limit;
1616 }
1617 EXPORT_SYMBOL(scsi_calculate_bounce_limit);
1618
1619 struct request_queue *__scsi_alloc_queue(struct Scsi_Host *shost,
1620                                          request_fn_proc *request_fn)
1621 {
1622         struct request_queue *q;
1623         struct device *dev = shost->shost_gendev.parent;
1624
1625         q = blk_init_queue(request_fn, NULL);
1626         if (!q)
1627                 return NULL;
1628
1629         /*
1630          * this limit is imposed by hardware restrictions
1631          */
1632         blk_queue_max_segments(q, min_t(unsigned short, shost->sg_tablesize,
1633                                         SCSI_MAX_SG_CHAIN_SEGMENTS));
1634
1635         if (scsi_host_prot_dma(shost)) {
1636                 shost->sg_prot_tablesize =
1637                         min_not_zero(shost->sg_prot_tablesize,
1638                                      (unsigned short)SCSI_MAX_PROT_SG_SEGMENTS);
1639                 BUG_ON(shost->sg_prot_tablesize < shost->sg_tablesize);
1640                 blk_queue_max_integrity_segments(q, shost->sg_prot_tablesize);
1641         }
1642
1643         blk_queue_max_hw_sectors(q, shost->max_sectors);
1644         blk_queue_bounce_limit(q, scsi_calculate_bounce_limit(shost));
1645         blk_queue_segment_boundary(q, shost->dma_boundary);
1646         dma_set_seg_boundary(dev, shost->dma_boundary);
1647
1648         blk_queue_max_segment_size(q, dma_get_max_seg_size(dev));
1649
1650         if (!shost->use_clustering)
1651                 q->limits.cluster = 0;
1652
1653         /*
1654          * set a reasonable default alignment on word boundaries: the
1655          * host and device may alter it using
1656          * blk_queue_update_dma_alignment() later.
1657          */
1658         blk_queue_dma_alignment(q, 0x03);
1659
1660         return q;
1661 }
1662 EXPORT_SYMBOL(__scsi_alloc_queue);
1663
1664 struct request_queue *scsi_alloc_queue(struct scsi_device *sdev)
1665 {
1666         struct request_queue *q;
1667
1668         q = __scsi_alloc_queue(sdev->host, scsi_request_fn);
1669         if (!q)
1670                 return NULL;
1671
1672         blk_queue_prep_rq(q, scsi_prep_fn);
1673         blk_queue_softirq_done(q, scsi_softirq_done);
1674         blk_queue_rq_timed_out(q, scsi_times_out);
1675         blk_queue_lld_busy(q, scsi_lld_busy);
1676         return q;
1677 }
1678
1679 void scsi_free_queue(struct request_queue *q)
1680 {
1681         blk_cleanup_queue(q);
1682 }
1683
1684 /*
1685  * Function:    scsi_block_requests()
1686  *
1687  * Purpose:     Utility function used by low-level drivers to prevent further
1688  *              commands from being queued to the device.
1689  *
1690  * Arguments:   shost       - Host in question
1691  *
1692  * Returns:     Nothing
1693  *
1694  * Lock status: No locks are assumed held.
1695  *
1696  * Notes:       There is no timer nor any other means by which the requests
1697  *              get unblocked other than the low-level driver calling
1698  *              scsi_unblock_requests().
1699  */
1700 void scsi_block_requests(struct Scsi_Host *shost)
1701 {
1702         shost->host_self_blocked = 1;
1703 }
1704 EXPORT_SYMBOL(scsi_block_requests);
1705
1706 /*
1707  * Function:    scsi_unblock_requests()
1708  *
1709  * Purpose:     Utility function used by low-level drivers to allow further
1710  *              commands from being queued to the device.
1711  *
1712  * Arguments:   shost       - Host in question
1713  *
1714  * Returns:     Nothing
1715  *
1716  * Lock status: No locks are assumed held.
1717  *
1718  * Notes:       There is no timer nor any other means by which the requests
1719  *              get unblocked other than the low-level driver calling
1720  *              scsi_unblock_requests().
1721  *
1722  *              This is done as an API function so that changes to the
1723  *              internals of the scsi mid-layer won't require wholesale
1724  *              changes to drivers that use this feature.
1725  */
1726 void scsi_unblock_requests(struct Scsi_Host *shost)
1727 {
1728         shost->host_self_blocked = 0;
1729         scsi_run_host_queues(shost);
1730 }
1731 EXPORT_SYMBOL(scsi_unblock_requests);
1732
1733 int __init scsi_init_queue(void)
1734 {
1735         int i;
1736
1737         scsi_sdb_cache = kmem_cache_create("scsi_data_buffer",
1738                                            sizeof(struct scsi_data_buffer),
1739                                            0, 0, NULL);
1740         if (!scsi_sdb_cache) {
1741                 printk(KERN_ERR "SCSI: can't init scsi sdb cache\n");
1742                 return -ENOMEM;
1743         }
1744
1745         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1746                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1747                 int size = sgp->size * sizeof(struct scatterlist);
1748
1749                 sgp->slab = kmem_cache_create(sgp->name, size, 0,
1750                                 SLAB_HWCACHE_ALIGN, NULL);
1751                 if (!sgp->slab) {
1752                         printk(KERN_ERR "SCSI: can't init sg slab %s\n",
1753                                         sgp->name);
1754                         goto cleanup_sdb;
1755                 }
1756
1757                 sgp->pool = mempool_create_slab_pool(SG_MEMPOOL_SIZE,
1758                                                      sgp->slab);
1759                 if (!sgp->pool) {
1760                         printk(KERN_ERR "SCSI: can't init sg mempool %s\n",
1761                                         sgp->name);
1762                         goto cleanup_sdb;
1763                 }
1764         }
1765
1766         return 0;
1767
1768 cleanup_sdb:
1769         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1770                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1771                 if (sgp->pool)
1772                         mempool_destroy(sgp->pool);
1773                 if (sgp->slab)
1774                         kmem_cache_destroy(sgp->slab);
1775         }
1776         kmem_cache_destroy(scsi_sdb_cache);
1777
1778         return -ENOMEM;
1779 }
1780
1781 void scsi_exit_queue(void)
1782 {
1783         int i;
1784
1785         kmem_cache_destroy(scsi_sdb_cache);
1786
1787         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1788                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1789                 mempool_destroy(sgp->pool);
1790                 kmem_cache_destroy(sgp->slab);
1791         }
1792 }
1793
1794 /**
1795  *      scsi_mode_select - issue a mode select
1796  *      @sdev:  SCSI device to be queried
1797  *      @pf:    Page format bit (1 == standard, 0 == vendor specific)
1798  *      @sp:    Save page bit (0 == don't save, 1 == save)
1799  *      @modepage: mode page being requested
1800  *      @buffer: request buffer (may not be smaller than eight bytes)
1801  *      @len:   length of request buffer.
1802  *      @timeout: command timeout
1803  *      @retries: number of retries before failing
1804  *      @data: returns a structure abstracting the mode header data
1805  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
1806  *              must be SCSI_SENSE_BUFFERSIZE big.
1807  *
1808  *      Returns zero if successful; negative error number or scsi
1809  *      status on error
1810  *
1811  */
1812 int
1813 scsi_mode_select(struct scsi_device *sdev, int pf, int sp, int modepage,
1814                  unsigned char *buffer, int len, int timeout, int retries,
1815                  struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1816 {
1817         unsigned char cmd[10];
1818         unsigned char *real_buffer;
1819         int ret;
1820
1821         memset(cmd, 0, sizeof(cmd));
1822         cmd[1] = (pf ? 0x10 : 0) | (sp ? 0x01 : 0);
1823
1824         if (sdev->use_10_for_ms) {
1825                 if (len > 65535)
1826                         return -EINVAL;
1827                 real_buffer = kmalloc(8 + len, GFP_KERNEL);
1828                 if (!real_buffer)
1829                         return -ENOMEM;
1830                 memcpy(real_buffer + 8, buffer, len);
1831                 len += 8;
1832                 real_buffer[0] = 0;
1833                 real_buffer[1] = 0;
1834                 real_buffer[2] = data->medium_type;
1835                 real_buffer[3] = data->device_specific;
1836                 real_buffer[4] = data->longlba ? 0x01 : 0;
1837                 real_buffer[5] = 0;
1838                 real_buffer[6] = data->block_descriptor_length >> 8;
1839                 real_buffer[7] = data->block_descriptor_length;
1840
1841                 cmd[0] = MODE_SELECT_10;
1842                 cmd[7] = len >> 8;
1843                 cmd[8] = len;
1844         } else {
1845                 if (len > 255 || data->block_descriptor_length > 255 ||
1846                     data->longlba)
1847                         return -EINVAL;
1848
1849                 real_buffer = kmalloc(4 + len, GFP_KERNEL);
1850                 if (!real_buffer)
1851                         return -ENOMEM;
1852                 memcpy(real_buffer + 4, buffer, len);
1853                 len += 4;
1854                 real_buffer[0] = 0;
1855                 real_buffer[1] = data->medium_type;
1856                 real_buffer[2] = data->device_specific;
1857                 real_buffer[3] = data->block_descriptor_length;
1858                 
1859
1860                 cmd[0] = MODE_SELECT;
1861                 cmd[4] = len;
1862         }
1863
1864         ret = scsi_execute_req(sdev, cmd, DMA_TO_DEVICE, real_buffer, len,
1865                                sshdr, timeout, retries, NULL);
1866         kfree(real_buffer);
1867         return ret;
1868 }
1869 EXPORT_SYMBOL_GPL(scsi_mode_select);
1870
1871 /**
1872  *      scsi_mode_sense - issue a mode sense, falling back from 10 to six bytes if necessary.
1873  *      @sdev:  SCSI device to be queried
1874  *      @dbd:   set if mode sense will allow block descriptors to be returned
1875  *      @modepage: mode page being requested
1876  *      @buffer: request buffer (may not be smaller than eight bytes)
1877  *      @len:   length of request buffer.
1878  *      @timeout: command timeout
1879  *      @retries: number of retries before failing
1880  *      @data: returns a structure abstracting the mode header data
1881  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
1882  *              must be SCSI_SENSE_BUFFERSIZE big.
1883  *
1884  *      Returns zero if unsuccessful, or the header offset (either 4
1885  *      or 8 depending on whether a six or ten byte command was
1886  *      issued) if successful.
1887  */
1888 int
1889 scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
1890                   unsigned char *buffer, int len, int timeout, int retries,
1891                   struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1892 {
1893         unsigned char cmd[12];
1894         int use_10_for_ms;
1895         int header_length;
1896         int result;
1897         struct scsi_sense_hdr my_sshdr;
1898
1899         memset(data, 0, sizeof(*data));
1900         memset(&cmd[0], 0, 12);
1901         cmd[1] = dbd & 0x18;    /* allows DBD and LLBA bits */
1902         cmd[2] = modepage;
1903
1904         /* caller might not be interested in sense, but we need it */
1905         if (!sshdr)
1906                 sshdr = &my_sshdr;
1907
1908  retry:
1909         use_10_for_ms = sdev->use_10_for_ms;
1910
1911         if (use_10_for_ms) {
1912                 if (len < 8)
1913                         len = 8;
1914
1915                 cmd[0] = MODE_SENSE_10;
1916                 cmd[8] = len;
1917                 header_length = 8;
1918         } else {
1919                 if (len < 4)
1920                         len = 4;
1921
1922                 cmd[0] = MODE_SENSE;
1923                 cmd[4] = len;
1924                 header_length = 4;
1925         }
1926
1927         memset(buffer, 0, len);
1928
1929         result = scsi_execute_req(sdev, cmd, DMA_FROM_DEVICE, buffer, len,
1930                                   sshdr, timeout, retries, NULL);
1931
1932         /* This code looks awful: what it's doing is making sure an
1933          * ILLEGAL REQUEST sense return identifies the actual command
1934          * byte as the problem.  MODE_SENSE commands can return
1935          * ILLEGAL REQUEST if the code page isn't supported */
1936
1937         if (use_10_for_ms && !scsi_status_is_good(result) &&
1938             (driver_byte(result) & DRIVER_SENSE)) {
1939                 if (scsi_sense_valid(sshdr)) {
1940                         if ((sshdr->sense_key == ILLEGAL_REQUEST) &&
1941                             (sshdr->asc == 0x20) && (sshdr->ascq == 0)) {
1942                                 /* 
1943                                  * Invalid command operation code
1944                                  */
1945                                 sdev->use_10_for_ms = 0;
1946                                 goto retry;
1947                         }
1948                 }
1949         }
1950
1951         if(scsi_status_is_good(result)) {
1952                 if (unlikely(buffer[0] == 0x86 && buffer[1] == 0x0b &&
1953                              (modepage == 6 || modepage == 8))) {
1954                         /* Initio breakage? */
1955                         header_length = 0;
1956                         data->length = 13;
1957                         data->medium_type = 0;
1958                         data->device_specific = 0;
1959                         data->longlba = 0;
1960                         data->block_descriptor_length = 0;
1961                 } else if(use_10_for_ms) {
1962                         data->length = buffer[0]*256 + buffer[1] + 2;
1963                         data->medium_type = buffer[2];
1964                         data->device_specific = buffer[3];
1965                         data->longlba = buffer[4] & 0x01;
1966                         data->block_descriptor_length = buffer[6]*256
1967                                 + buffer[7];
1968                 } else {
1969                         data->length = buffer[0] + 1;
1970                         data->medium_type = buffer[1];
1971                         data->device_specific = buffer[2];
1972                         data->block_descriptor_length = buffer[3];
1973                 }
1974                 data->header_length = header_length;
1975         }
1976
1977         return result;
1978 }
1979 EXPORT_SYMBOL(scsi_mode_sense);
1980
1981 /**
1982  *      scsi_test_unit_ready - test if unit is ready
1983  *      @sdev:  scsi device to change the state of.
1984  *      @timeout: command timeout
1985  *      @retries: number of retries before failing
1986  *      @sshdr_external: Optional pointer to struct scsi_sense_hdr for
1987  *              returning sense. Make sure that this is cleared before passing
1988  *              in.
1989  *
1990  *      Returns zero if unsuccessful or an error if TUR failed.  For
1991  *      removable media, UNIT_ATTENTION sets ->changed flag.
1992  **/
1993 int
1994 scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries,
1995                      struct scsi_sense_hdr *sshdr_external)
1996 {
1997         char cmd[] = {
1998                 TEST_UNIT_READY, 0, 0, 0, 0, 0,
1999         };
2000         struct scsi_sense_hdr *sshdr;
2001         int result;
2002
2003         if (!sshdr_external)
2004                 sshdr = kzalloc(sizeof(*sshdr), GFP_KERNEL);
2005         else
2006                 sshdr = sshdr_external;
2007
2008         /* try to eat the UNIT_ATTENTION if there are enough retries */
2009         do {
2010                 result = scsi_execute_req(sdev, cmd, DMA_NONE, NULL, 0, sshdr,
2011                                           timeout, retries, NULL);
2012                 if (sdev->removable && scsi_sense_valid(sshdr) &&
2013                     sshdr->sense_key == UNIT_ATTENTION)
2014                         sdev->changed = 1;
2015         } while (scsi_sense_valid(sshdr) &&
2016                  sshdr->sense_key == UNIT_ATTENTION && --retries);
2017
2018         if (!sshdr_external)
2019                 kfree(sshdr);
2020         return result;
2021 }
2022 EXPORT_SYMBOL(scsi_test_unit_ready);
2023
2024 /**
2025  *      scsi_device_set_state - Take the given device through the device state model.
2026  *      @sdev:  scsi device to change the state of.
2027  *      @state: state to change to.
2028  *
2029  *      Returns zero if unsuccessful or an error if the requested 
2030  *      transition is illegal.
2031  */
2032 int
2033 scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
2034 {
2035         enum scsi_device_state oldstate = sdev->sdev_state;
2036
2037         if (state == oldstate)
2038                 return 0;
2039
2040         switch (state) {
2041         case SDEV_CREATED:
2042                 switch (oldstate) {
2043                 case SDEV_CREATED_BLOCK:
2044                         break;
2045                 default:
2046                         goto illegal;
2047                 }
2048                 break;
2049                         
2050         case SDEV_RUNNING:
2051                 switch (oldstate) {
2052                 case SDEV_CREATED:
2053                 case SDEV_OFFLINE:
2054                 case SDEV_QUIESCE:
2055                 case SDEV_BLOCK:
2056                         break;
2057                 default:
2058                         goto illegal;
2059                 }
2060                 break;
2061
2062         case SDEV_QUIESCE:
2063                 switch (oldstate) {
2064                 case SDEV_RUNNING:
2065                 case SDEV_OFFLINE:
2066                         break;
2067                 default:
2068                         goto illegal;
2069                 }
2070                 break;
2071
2072         case SDEV_OFFLINE:
2073                 switch (oldstate) {
2074                 case SDEV_CREATED:
2075                 case SDEV_RUNNING:
2076                 case SDEV_QUIESCE:
2077                 case SDEV_BLOCK:
2078                         break;
2079                 default:
2080                         goto illegal;
2081                 }
2082                 break;
2083
2084         case SDEV_BLOCK:
2085                 switch (oldstate) {
2086                 case SDEV_RUNNING:
2087                 case SDEV_CREATED_BLOCK:
2088                         break;
2089                 default:
2090                         goto illegal;
2091                 }
2092                 break;
2093
2094         case SDEV_CREATED_BLOCK:
2095                 switch (oldstate) {
2096                 case SDEV_CREATED:
2097                         break;
2098                 default:
2099                         goto illegal;
2100                 }
2101                 break;
2102
2103         case SDEV_CANCEL:
2104                 switch (oldstate) {
2105                 case SDEV_CREATED:
2106                 case SDEV_RUNNING:
2107                 case SDEV_QUIESCE:
2108                 case SDEV_OFFLINE:
2109                 case SDEV_BLOCK:
2110                         break;
2111                 default:
2112                         goto illegal;
2113                 }
2114                 break;
2115
2116         case SDEV_DEL:
2117                 switch (oldstate) {
2118                 case SDEV_CREATED:
2119                 case SDEV_RUNNING:
2120                 case SDEV_OFFLINE:
2121                 case SDEV_CANCEL:
2122                         break;
2123                 default:
2124                         goto illegal;
2125                 }
2126                 break;
2127
2128         }
2129         sdev->sdev_state = state;
2130         return 0;
2131
2132  illegal:
2133         SCSI_LOG_ERROR_RECOVERY(1, 
2134                                 sdev_printk(KERN_ERR, sdev,
2135                                             "Illegal state transition %s->%s\n",
2136                                             scsi_device_state_name(oldstate),
2137                                             scsi_device_state_name(state))
2138                                 );
2139         return -EINVAL;
2140 }
2141 EXPORT_SYMBOL(scsi_device_set_state);
2142
2143 /**
2144  *      sdev_evt_emit - emit a single SCSI device uevent
2145  *      @sdev: associated SCSI device
2146  *      @evt: event to emit
2147  *
2148  *      Send a single uevent (scsi_event) to the associated scsi_device.
2149  */
2150 static void scsi_evt_emit(struct scsi_device *sdev, struct scsi_event *evt)
2151 {
2152         int idx = 0;
2153         char *envp[3];
2154
2155         switch (evt->evt_type) {
2156         case SDEV_EVT_MEDIA_CHANGE:
2157                 envp[idx++] = "SDEV_MEDIA_CHANGE=1";
2158                 break;
2159
2160         default:
2161                 /* do nothing */
2162                 break;
2163         }
2164
2165         envp[idx++] = NULL;
2166
2167         kobject_uevent_env(&sdev->sdev_gendev.kobj, KOBJ_CHANGE, envp);
2168 }
2169
2170 /**
2171  *      sdev_evt_thread - send a uevent for each scsi event
2172  *      @work: work struct for scsi_device
2173  *
2174  *      Dispatch queued events to their associated scsi_device kobjects
2175  *      as uevents.
2176  */
2177 void scsi_evt_thread(struct work_struct *work)
2178 {
2179         struct scsi_device *sdev;
2180         LIST_HEAD(event_list);
2181
2182         sdev = container_of(work, struct scsi_device, event_work);
2183
2184         while (1) {
2185                 struct scsi_event *evt;
2186                 struct list_head *this, *tmp;
2187                 unsigned long flags;
2188
2189                 spin_lock_irqsave(&sdev->list_lock, flags);
2190                 list_splice_init(&sdev->event_list, &event_list);
2191                 spin_unlock_irqrestore(&sdev->list_lock, flags);
2192
2193                 if (list_empty(&event_list))
2194                         break;
2195
2196                 list_for_each_safe(this, tmp, &event_list) {
2197                         evt = list_entry(this, struct scsi_event, node);
2198                         list_del(&evt->node);
2199                         scsi_evt_emit(sdev, evt);
2200                         kfree(evt);
2201                 }
2202         }
2203 }
2204
2205 /**
2206  *      sdev_evt_send - send asserted event to uevent thread
2207  *      @sdev: scsi_device event occurred on
2208  *      @evt: event to send
2209  *
2210  *      Assert scsi device event asynchronously.
2211  */
2212 void sdev_evt_send(struct scsi_device *sdev, struct scsi_event *evt)
2213 {
2214         unsigned long flags;
2215
2216 #if 0
2217         /* FIXME: currently this check eliminates all media change events
2218          * for polled devices.  Need to update to discriminate between AN
2219          * and polled events */
2220         if (!test_bit(evt->evt_type, sdev->supported_events)) {
2221                 kfree(evt);
2222                 return;
2223         }
2224 #endif
2225
2226         spin_lock_irqsave(&sdev->list_lock, flags);
2227         list_add_tail(&evt->node, &sdev->event_list);
2228         schedule_work(&sdev->event_work);
2229         spin_unlock_irqrestore(&sdev->list_lock, flags);
2230 }
2231 EXPORT_SYMBOL_GPL(sdev_evt_send);
2232
2233 /**
2234  *      sdev_evt_alloc - allocate a new scsi event
2235  *      @evt_type: type of event to allocate
2236  *      @gfpflags: GFP flags for allocation
2237  *
2238  *      Allocates and returns a new scsi_event.
2239  */
2240 struct scsi_event *sdev_evt_alloc(enum scsi_device_event evt_type,
2241                                   gfp_t gfpflags)
2242 {
2243         struct scsi_event *evt = kzalloc(sizeof(struct scsi_event), gfpflags);
2244         if (!evt)
2245                 return NULL;
2246
2247         evt->evt_type = evt_type;
2248         INIT_LIST_HEAD(&evt->node);
2249
2250         /* evt_type-specific initialization, if any */
2251         switch (evt_type) {
2252         case SDEV_EVT_MEDIA_CHANGE:
2253         default:
2254                 /* do nothing */
2255                 break;
2256         }
2257
2258         return evt;
2259 }
2260 EXPORT_SYMBOL_GPL(sdev_evt_alloc);
2261
2262 /**
2263  *      sdev_evt_send_simple - send asserted event to uevent thread
2264  *      @sdev: scsi_device event occurred on
2265  *      @evt_type: type of event to send
2266  *      @gfpflags: GFP flags for allocation
2267  *
2268  *      Assert scsi device event asynchronously, given an event type.
2269  */
2270 void sdev_evt_send_simple(struct scsi_device *sdev,
2271                           enum scsi_device_event evt_type, gfp_t gfpflags)
2272 {
2273         struct scsi_event *evt = sdev_evt_alloc(evt_type, gfpflags);
2274         if (!evt) {
2275                 sdev_printk(KERN_ERR, sdev, "event %d eaten due to OOM\n",
2276                             evt_type);
2277                 return;
2278         }
2279
2280         sdev_evt_send(sdev, evt);
2281 }
2282 EXPORT_SYMBOL_GPL(sdev_evt_send_simple);
2283
2284 /**
2285  *      scsi_device_quiesce - Block user issued commands.
2286  *      @sdev:  scsi device to quiesce.
2287  *
2288  *      This works by trying to transition to the SDEV_QUIESCE state
2289  *      (which must be a legal transition).  When the device is in this
2290  *      state, only special requests will be accepted, all others will
2291  *      be deferred.  Since special requests may also be requeued requests,
2292  *      a successful return doesn't guarantee the device will be 
2293  *      totally quiescent.
2294  *
2295  *      Must be called with user context, may sleep.
2296  *
2297  *      Returns zero if unsuccessful or an error if not.
2298  */
2299 int
2300 scsi_device_quiesce(struct scsi_device *sdev)
2301 {
2302         int err = scsi_device_set_state(sdev, SDEV_QUIESCE);
2303         if (err)
2304                 return err;
2305
2306         scsi_run_queue(sdev->request_queue);
2307         while (sdev->device_busy) {
2308                 msleep_interruptible(200);
2309                 scsi_run_queue(sdev->request_queue);
2310         }
2311         return 0;
2312 }
2313 EXPORT_SYMBOL(scsi_device_quiesce);
2314
2315 /**
2316  *      scsi_device_resume - Restart user issued commands to a quiesced device.
2317  *      @sdev:  scsi device to resume.
2318  *
2319  *      Moves the device from quiesced back to running and restarts the
2320  *      queues.
2321  *
2322  *      Must be called with user context, may sleep.
2323  */
2324 void
2325 scsi_device_resume(struct scsi_device *sdev)
2326 {
2327         if(scsi_device_set_state(sdev, SDEV_RUNNING))
2328                 return;
2329         scsi_run_queue(sdev->request_queue);
2330 }
2331 EXPORT_SYMBOL(scsi_device_resume);
2332
2333 static void
2334 device_quiesce_fn(struct scsi_device *sdev, void *data)
2335 {
2336         scsi_device_quiesce(sdev);
2337 }
2338
2339 void
2340 scsi_target_quiesce(struct scsi_target *starget)
2341 {
2342         starget_for_each_device(starget, NULL, device_quiesce_fn);
2343 }
2344 EXPORT_SYMBOL(scsi_target_quiesce);
2345
2346 static void
2347 device_resume_fn(struct scsi_device *sdev, void *data)
2348 {
2349         scsi_device_resume(sdev);
2350 }
2351
2352 void
2353 scsi_target_resume(struct scsi_target *starget)
2354 {
2355         starget_for_each_device(starget, NULL, device_resume_fn);
2356 }
2357 EXPORT_SYMBOL(scsi_target_resume);
2358
2359 /**
2360  * scsi_internal_device_block - internal function to put a device temporarily into the SDEV_BLOCK state
2361  * @sdev:       device to block
2362  *
2363  * Block request made by scsi lld's to temporarily stop all
2364  * scsi commands on the specified device.  Called from interrupt
2365  * or normal process context.
2366  *
2367  * Returns zero if successful or error if not
2368  *
2369  * Notes:       
2370  *      This routine transitions the device to the SDEV_BLOCK state
2371  *      (which must be a legal transition).  When the device is in this
2372  *      state, all commands are deferred until the scsi lld reenables
2373  *      the device with scsi_device_unblock or device_block_tmo fires.
2374  *      This routine assumes the host_lock is held on entry.
2375  */
2376 int
2377 scsi_internal_device_block(struct scsi_device *sdev)
2378 {
2379         struct request_queue *q = sdev->request_queue;
2380         unsigned long flags;
2381         int err = 0;
2382
2383         err = scsi_device_set_state(sdev, SDEV_BLOCK);
2384         if (err) {
2385                 err = scsi_device_set_state(sdev, SDEV_CREATED_BLOCK);
2386
2387                 if (err)
2388                         return err;
2389         }
2390
2391         /* 
2392          * The device has transitioned to SDEV_BLOCK.  Stop the
2393          * block layer from calling the midlayer with this device's
2394          * request queue. 
2395          */
2396         spin_lock_irqsave(q->queue_lock, flags);
2397         blk_stop_queue(q);
2398         spin_unlock_irqrestore(q->queue_lock, flags);
2399
2400         return 0;
2401 }
2402 EXPORT_SYMBOL_GPL(scsi_internal_device_block);
2403  
2404 /**
2405  * scsi_internal_device_unblock - resume a device after a block request
2406  * @sdev:       device to resume
2407  *
2408  * Called by scsi lld's or the midlayer to restart the device queue
2409  * for the previously suspended scsi device.  Called from interrupt or
2410  * normal process context.
2411  *
2412  * Returns zero if successful or error if not.
2413  *
2414  * Notes:       
2415  *      This routine transitions the device to the SDEV_RUNNING state
2416  *      (which must be a legal transition) allowing the midlayer to
2417  *      goose the queue for this device.  This routine assumes the 
2418  *      host_lock is held upon entry.
2419  */
2420 int
2421 scsi_internal_device_unblock(struct scsi_device *sdev)
2422 {
2423         struct request_queue *q = sdev->request_queue; 
2424         unsigned long flags;
2425         
2426         /* 
2427          * Try to transition the scsi device to SDEV_RUNNING
2428          * and goose the device queue if successful.  
2429          */
2430         if (sdev->sdev_state == SDEV_BLOCK)
2431                 sdev->sdev_state = SDEV_RUNNING;
2432         else if (sdev->sdev_state == SDEV_CREATED_BLOCK)
2433                 sdev->sdev_state = SDEV_CREATED;
2434         else if (sdev->sdev_state != SDEV_CANCEL &&
2435                  sdev->sdev_state != SDEV_OFFLINE)
2436                 return -EINVAL;
2437
2438         spin_lock_irqsave(q->queue_lock, flags);
2439         blk_start_queue(q);
2440         spin_unlock_irqrestore(q->queue_lock, flags);
2441
2442         return 0;
2443 }
2444 EXPORT_SYMBOL_GPL(scsi_internal_device_unblock);
2445
2446 static void
2447 device_block(struct scsi_device *sdev, void *data)
2448 {
2449         scsi_internal_device_block(sdev);
2450 }
2451
2452 static int
2453 target_block(struct device *dev, void *data)
2454 {
2455         if (scsi_is_target_device(dev))
2456                 starget_for_each_device(to_scsi_target(dev), NULL,
2457                                         device_block);
2458         return 0;
2459 }
2460
2461 void
2462 scsi_target_block(struct device *dev)
2463 {
2464         if (scsi_is_target_device(dev))
2465                 starget_for_each_device(to_scsi_target(dev), NULL,
2466                                         device_block);
2467         else
2468                 device_for_each_child(dev, NULL, target_block);
2469 }
2470 EXPORT_SYMBOL_GPL(scsi_target_block);
2471
2472 static void
2473 device_unblock(struct scsi_device *sdev, void *data)
2474 {
2475         scsi_internal_device_unblock(sdev);
2476 }
2477
2478 static int
2479 target_unblock(struct device *dev, void *data)
2480 {
2481         if (scsi_is_target_device(dev))
2482                 starget_for_each_device(to_scsi_target(dev), NULL,
2483                                         device_unblock);
2484         return 0;
2485 }
2486
2487 void
2488 scsi_target_unblock(struct device *dev)
2489 {
2490         if (scsi_is_target_device(dev))
2491                 starget_for_each_device(to_scsi_target(dev), NULL,
2492                                         device_unblock);
2493         else
2494                 device_for_each_child(dev, NULL, target_unblock);
2495 }
2496 EXPORT_SYMBOL_GPL(scsi_target_unblock);
2497
2498 /**
2499  * scsi_kmap_atomic_sg - find and atomically map an sg-elemnt
2500  * @sgl:        scatter-gather list
2501  * @sg_count:   number of segments in sg
2502  * @offset:     offset in bytes into sg, on return offset into the mapped area
2503  * @len:        bytes to map, on return number of bytes mapped
2504  *
2505  * Returns virtual address of the start of the mapped page
2506  */
2507 void *scsi_kmap_atomic_sg(struct scatterlist *sgl, int sg_count,
2508                           size_t *offset, size_t *len)
2509 {
2510         int i;
2511         size_t sg_len = 0, len_complete = 0;
2512         struct scatterlist *sg;
2513         struct page *page;
2514
2515         WARN_ON(!irqs_disabled());
2516
2517         for_each_sg(sgl, sg, sg_count, i) {
2518                 len_complete = sg_len; /* Complete sg-entries */
2519                 sg_len += sg->length;
2520                 if (sg_len > *offset)
2521                         break;
2522         }
2523
2524         if (unlikely(i == sg_count)) {
2525                 printk(KERN_ERR "%s: Bytes in sg: %zu, requested offset %zu, "
2526                         "elements %d\n",
2527                        __func__, sg_len, *offset, sg_count);
2528                 WARN_ON(1);
2529                 return NULL;
2530         }
2531
2532         /* Offset starting from the beginning of first page in this sg-entry */
2533         *offset = *offset - len_complete + sg->offset;
2534
2535         /* Assumption: contiguous pages can be accessed as "page + i" */
2536         page = nth_page(sg_page(sg), (*offset >> PAGE_SHIFT));
2537         *offset &= ~PAGE_MASK;
2538
2539         /* Bytes in this sg-entry from *offset to the end of the page */
2540         sg_len = PAGE_SIZE - *offset;
2541         if (*len > sg_len)
2542                 *len = sg_len;
2543
2544         return kmap_atomic(page, KM_BIO_SRC_IRQ);
2545 }
2546 EXPORT_SYMBOL(scsi_kmap_atomic_sg);
2547
2548 /**
2549  * scsi_kunmap_atomic_sg - atomically unmap a virtual address, previously mapped with scsi_kmap_atomic_sg
2550  * @virt:       virtual address to be unmapped
2551  */
2552 void scsi_kunmap_atomic_sg(void *virt)
2553 {
2554         kunmap_atomic(virt, KM_BIO_SRC_IRQ);
2555 }
2556 EXPORT_SYMBOL(scsi_kunmap_atomic_sg);