Merge branch 'for-2.6.37/core' of git://git.kernel.dk/linux-2.6-block
[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 static void scsi_run_queue(struct request_queue *q);
71
72 /*
73  * Function:    scsi_unprep_request()
74  *
75  * Purpose:     Remove all preparation done for a request, including its
76  *              associated scsi_cmnd, so that it can be requeued.
77  *
78  * Arguments:   req     - request to unprepare
79  *
80  * Lock status: Assumed that no locks are held upon entry.
81  *
82  * Returns:     Nothing.
83  */
84 static void scsi_unprep_request(struct request *req)
85 {
86         struct scsi_cmnd *cmd = req->special;
87
88         blk_unprep_request(req);
89         req->special = NULL;
90
91         scsi_put_command(cmd);
92 }
93
94 /**
95  * __scsi_queue_insert - private queue insertion
96  * @cmd: The SCSI command being requeued
97  * @reason:  The reason for the requeue
98  * @unbusy: Whether the queue should be unbusied
99  *
100  * This is a private queue insertion.  The public interface
101  * scsi_queue_insert() always assumes the queue should be unbusied
102  * because it's always called before the completion.  This function is
103  * for a requeue after completion, which should only occur in this
104  * file.
105  */
106 static int __scsi_queue_insert(struct scsi_cmnd *cmd, int reason, int unbusy)
107 {
108         struct Scsi_Host *host = cmd->device->host;
109         struct scsi_device *device = cmd->device;
110         struct scsi_target *starget = scsi_target(device);
111         struct request_queue *q = device->request_queue;
112         unsigned long flags;
113
114         SCSI_LOG_MLQUEUE(1,
115                  printk("Inserting command %p into mlqueue\n", cmd));
116
117         /*
118          * Set the appropriate busy bit for the device/host.
119          *
120          * If the host/device isn't busy, assume that something actually
121          * completed, and that we should be able to queue a command now.
122          *
123          * Note that the prior mid-layer assumption that any host could
124          * always queue at least one command is now broken.  The mid-layer
125          * will implement a user specifiable stall (see
126          * scsi_host.max_host_blocked and scsi_device.max_device_blocked)
127          * if a command is requeued with no other commands outstanding
128          * either for the device or for the host.
129          */
130         switch (reason) {
131         case SCSI_MLQUEUE_HOST_BUSY:
132                 host->host_blocked = host->max_host_blocked;
133                 break;
134         case SCSI_MLQUEUE_DEVICE_BUSY:
135                 device->device_blocked = device->max_device_blocked;
136                 break;
137         case SCSI_MLQUEUE_TARGET_BUSY:
138                 starget->target_blocked = starget->max_target_blocked;
139                 break;
140         }
141
142         /*
143          * Decrement the counters, since these commands are no longer
144          * active on the host/device.
145          */
146         if (unbusy)
147                 scsi_device_unbusy(device);
148
149         /*
150          * Requeue this command.  It will go before all other commands
151          * that are already in the queue.
152          *
153          * NOTE: there is magic here about the way the queue is plugged if
154          * we have no outstanding commands.
155          * 
156          * Although we *don't* plug the queue, we call the request
157          * function.  The SCSI request function detects the blocked condition
158          * and plugs the queue appropriately.
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                 int flagset;
415
416                 /*
417                  * As long as shost is accepting commands and we have
418                  * starved queues, call blk_run_queue. scsi_request_fn
419                  * drops the queue_lock and can add us back to the
420                  * starved_list.
421                  *
422                  * host_lock protects the starved_list and starved_entry.
423                  * scsi_request_fn must get the host_lock before checking
424                  * or modifying starved_list or starved_entry.
425                  */
426                 if (scsi_host_is_busy(shost))
427                         break;
428
429                 sdev = list_entry(starved_list.next,
430                                   struct scsi_device, starved_entry);
431                 list_del_init(&sdev->starved_entry);
432                 if (scsi_target_is_busy(scsi_target(sdev))) {
433                         list_move_tail(&sdev->starved_entry,
434                                        &shost->starved_list);
435                         continue;
436                 }
437
438                 spin_unlock(shost->host_lock);
439
440                 spin_lock(sdev->request_queue->queue_lock);
441                 flagset = test_bit(QUEUE_FLAG_REENTER, &q->queue_flags) &&
442                                 !test_bit(QUEUE_FLAG_REENTER,
443                                         &sdev->request_queue->queue_flags);
444                 if (flagset)
445                         queue_flag_set(QUEUE_FLAG_REENTER, sdev->request_queue);
446                 __blk_run_queue(sdev->request_queue);
447                 if (flagset)
448                         queue_flag_clear(QUEUE_FLAG_REENTER, sdev->request_queue);
449                 spin_unlock(sdev->request_queue->queue_lock);
450
451                 spin_lock(shost->host_lock);
452         }
453         /* put any unprocessed entries back */
454         list_splice(&starved_list, &shost->starved_list);
455         spin_unlock_irqrestore(shost->host_lock, flags);
456
457         blk_run_queue(q);
458 }
459
460 /*
461  * Function:    scsi_requeue_command()
462  *
463  * Purpose:     Handle post-processing of completed commands.
464  *
465  * Arguments:   q       - queue to operate on
466  *              cmd     - command that may need to be requeued.
467  *
468  * Returns:     Nothing
469  *
470  * Notes:       After command completion, there may be blocks left
471  *              over which weren't finished by the previous command
472  *              this can be for a number of reasons - the main one is
473  *              I/O errors in the middle of the request, in which case
474  *              we need to request the blocks that come after the bad
475  *              sector.
476  * Notes:       Upon return, cmd is a stale pointer.
477  */
478 static void scsi_requeue_command(struct request_queue *q, struct scsi_cmnd *cmd)
479 {
480         struct request *req = cmd->request;
481         unsigned long flags;
482
483         spin_lock_irqsave(q->queue_lock, flags);
484         scsi_unprep_request(req);
485         blk_requeue_request(q, req);
486         spin_unlock_irqrestore(q->queue_lock, flags);
487
488         scsi_run_queue(q);
489 }
490
491 void scsi_next_command(struct scsi_cmnd *cmd)
492 {
493         struct scsi_device *sdev = cmd->device;
494         struct request_queue *q = sdev->request_queue;
495
496         /* need to hold a reference on the device before we let go of the cmd */
497         get_device(&sdev->sdev_gendev);
498
499         scsi_put_command(cmd);
500         scsi_run_queue(q);
501
502         /* ok to remove device now */
503         put_device(&sdev->sdev_gendev);
504 }
505
506 void scsi_run_host_queues(struct Scsi_Host *shost)
507 {
508         struct scsi_device *sdev;
509
510         shost_for_each_device(sdev, shost)
511                 scsi_run_queue(sdev->request_queue);
512 }
513
514 static void __scsi_release_buffers(struct scsi_cmnd *, int);
515
516 /*
517  * Function:    scsi_end_request()
518  *
519  * Purpose:     Post-processing of completed commands (usually invoked at end
520  *              of upper level post-processing and scsi_io_completion).
521  *
522  * Arguments:   cmd      - command that is complete.
523  *              error    - 0 if I/O indicates success, < 0 for I/O error.
524  *              bytes    - number of bytes of completed I/O
525  *              requeue  - indicates whether we should requeue leftovers.
526  *
527  * Lock status: Assumed that lock is not held upon entry.
528  *
529  * Returns:     cmd if requeue required, NULL otherwise.
530  *
531  * Notes:       This is called for block device requests in order to
532  *              mark some number of sectors as complete.
533  * 
534  *              We are guaranteeing that the request queue will be goosed
535  *              at some point during this call.
536  * Notes:       If cmd was requeued, upon return it will be a stale pointer.
537  */
538 static struct scsi_cmnd *scsi_end_request(struct scsi_cmnd *cmd, int error,
539                                           int bytes, int requeue)
540 {
541         struct request_queue *q = cmd->device->request_queue;
542         struct request *req = cmd->request;
543
544         /*
545          * If there are blocks left over at the end, set up the command
546          * to queue the remainder of them.
547          */
548         if (blk_end_request(req, error, bytes)) {
549                 /* kill remainder if no retrys */
550                 if (error && scsi_noretry_cmd(cmd))
551                         blk_end_request_all(req, error);
552                 else {
553                         if (requeue) {
554                                 /*
555                                  * Bleah.  Leftovers again.  Stick the
556                                  * leftovers in the front of the
557                                  * queue, and goose the queue again.
558                                  */
559                                 scsi_release_buffers(cmd);
560                                 scsi_requeue_command(q, cmd);
561                                 cmd = NULL;
562                         }
563                         return cmd;
564                 }
565         }
566
567         /*
568          * This will goose the queue request function at the end, so we don't
569          * need to worry about launching another command.
570          */
571         __scsi_release_buffers(cmd, 0);
572         scsi_next_command(cmd);
573         return NULL;
574 }
575
576 static inline unsigned int scsi_sgtable_index(unsigned short nents)
577 {
578         unsigned int index;
579
580         BUG_ON(nents > SCSI_MAX_SG_SEGMENTS);
581
582         if (nents <= 8)
583                 index = 0;
584         else
585                 index = get_count_order(nents) - 3;
586
587         return index;
588 }
589
590 static void scsi_sg_free(struct scatterlist *sgl, unsigned int nents)
591 {
592         struct scsi_host_sg_pool *sgp;
593
594         sgp = scsi_sg_pools + scsi_sgtable_index(nents);
595         mempool_free(sgl, sgp->pool);
596 }
597
598 static struct scatterlist *scsi_sg_alloc(unsigned int nents, gfp_t gfp_mask)
599 {
600         struct scsi_host_sg_pool *sgp;
601
602         sgp = scsi_sg_pools + scsi_sgtable_index(nents);
603         return mempool_alloc(sgp->pool, gfp_mask);
604 }
605
606 static int scsi_alloc_sgtable(struct scsi_data_buffer *sdb, int nents,
607                               gfp_t gfp_mask)
608 {
609         int ret;
610
611         BUG_ON(!nents);
612
613         ret = __sg_alloc_table(&sdb->table, nents, SCSI_MAX_SG_SEGMENTS,
614                                gfp_mask, scsi_sg_alloc);
615         if (unlikely(ret))
616                 __sg_free_table(&sdb->table, SCSI_MAX_SG_SEGMENTS,
617                                 scsi_sg_free);
618
619         return ret;
620 }
621
622 static void scsi_free_sgtable(struct scsi_data_buffer *sdb)
623 {
624         __sg_free_table(&sdb->table, SCSI_MAX_SG_SEGMENTS, scsi_sg_free);
625 }
626
627 static void __scsi_release_buffers(struct scsi_cmnd *cmd, int do_bidi_check)
628 {
629
630         if (cmd->sdb.table.nents)
631                 scsi_free_sgtable(&cmd->sdb);
632
633         memset(&cmd->sdb, 0, sizeof(cmd->sdb));
634
635         if (do_bidi_check && scsi_bidi_cmnd(cmd)) {
636                 struct scsi_data_buffer *bidi_sdb =
637                         cmd->request->next_rq->special;
638                 scsi_free_sgtable(bidi_sdb);
639                 kmem_cache_free(scsi_sdb_cache, bidi_sdb);
640                 cmd->request->next_rq->special = NULL;
641         }
642
643         if (scsi_prot_sg_count(cmd))
644                 scsi_free_sgtable(cmd->prot_sdb);
645 }
646
647 /*
648  * Function:    scsi_release_buffers()
649  *
650  * Purpose:     Completion processing for block device I/O requests.
651  *
652  * Arguments:   cmd     - command that we are bailing.
653  *
654  * Lock status: Assumed that no lock is held upon entry.
655  *
656  * Returns:     Nothing
657  *
658  * Notes:       In the event that an upper level driver rejects a
659  *              command, we must release resources allocated during
660  *              the __init_io() function.  Primarily this would involve
661  *              the scatter-gather table, and potentially any bounce
662  *              buffers.
663  */
664 void scsi_release_buffers(struct scsi_cmnd *cmd)
665 {
666         __scsi_release_buffers(cmd, 1);
667 }
668 EXPORT_SYMBOL(scsi_release_buffers);
669
670 /*
671  * Function:    scsi_io_completion()
672  *
673  * Purpose:     Completion processing for block device I/O requests.
674  *
675  * Arguments:   cmd   - command that is finished.
676  *
677  * Lock status: Assumed that no lock is held upon entry.
678  *
679  * Returns:     Nothing
680  *
681  * Notes:       This function is matched in terms of capabilities to
682  *              the function that created the scatter-gather list.
683  *              In other words, if there are no bounce buffers
684  *              (the normal case for most drivers), we don't need
685  *              the logic to deal with cleaning up afterwards.
686  *
687  *              We must call scsi_end_request().  This will finish off
688  *              the specified number of sectors.  If we are done, the
689  *              command block will be released and the queue function
690  *              will be goosed.  If we are not done then we have to
691  *              figure out what to do next:
692  *
693  *              a) We can call scsi_requeue_command().  The request
694  *                 will be unprepared and put back on the queue.  Then
695  *                 a new command will be created for it.  This should
696  *                 be used if we made forward progress, or if we want
697  *                 to switch from READ(10) to READ(6) for example.
698  *
699  *              b) We can call scsi_queue_insert().  The request will
700  *                 be put back on the queue and retried using the same
701  *                 command as before, possibly after a delay.
702  *
703  *              c) We can call blk_end_request() with -EIO to fail
704  *                 the remainder of the request.
705  */
706 void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes)
707 {
708         int result = cmd->result;
709         struct request_queue *q = cmd->device->request_queue;
710         struct request *req = cmd->request;
711         int error = 0;
712         struct scsi_sense_hdr sshdr;
713         int sense_valid = 0;
714         int sense_deferred = 0;
715         enum {ACTION_FAIL, ACTION_REPREP, ACTION_RETRY,
716               ACTION_DELAYED_RETRY} action;
717         char *description = NULL;
718
719         if (result) {
720                 sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
721                 if (sense_valid)
722                         sense_deferred = scsi_sense_is_deferred(&sshdr);
723         }
724
725         if (req->cmd_type == REQ_TYPE_BLOCK_PC) { /* SG_IO ioctl from block level */
726                 req->errors = result;
727                 if (result) {
728                         if (sense_valid && req->sense) {
729                                 /*
730                                  * SG_IO wants current and deferred errors
731                                  */
732                                 int len = 8 + cmd->sense_buffer[7];
733
734                                 if (len > SCSI_SENSE_BUFFERSIZE)
735                                         len = SCSI_SENSE_BUFFERSIZE;
736                                 memcpy(req->sense, cmd->sense_buffer,  len);
737                                 req->sense_len = len;
738                         }
739                         if (!sense_deferred)
740                                 error = -EIO;
741                 }
742
743                 req->resid_len = scsi_get_resid(cmd);
744
745                 if (scsi_bidi_cmnd(cmd)) {
746                         /*
747                          * Bidi commands Must be complete as a whole,
748                          * both sides at once.
749                          */
750                         req->next_rq->resid_len = scsi_in(cmd)->resid;
751
752                         scsi_release_buffers(cmd);
753                         blk_end_request_all(req, 0);
754
755                         scsi_next_command(cmd);
756                         return;
757                 }
758         }
759
760         /* no bidi support for !REQ_TYPE_BLOCK_PC yet */
761         BUG_ON(blk_bidi_rq(req));
762
763         /*
764          * Next deal with any sectors which we were able to correctly
765          * handle.
766          */
767         SCSI_LOG_HLCOMPLETE(1, printk("%u sectors total, "
768                                       "%d bytes done.\n",
769                                       blk_rq_sectors(req), good_bytes));
770
771         /*
772          * Recovered errors need reporting, but they're always treated
773          * as success, so fiddle the result code here.  For BLOCK_PC
774          * we already took a copy of the original into rq->errors which
775          * is what gets returned to the user
776          */
777         if (sense_valid && (sshdr.sense_key == RECOVERED_ERROR)) {
778                 /* if ATA PASS-THROUGH INFORMATION AVAILABLE skip
779                  * print since caller wants ATA registers. Only occurs on
780                  * SCSI ATA PASS_THROUGH commands when CK_COND=1
781                  */
782                 if ((sshdr.asc == 0x0) && (sshdr.ascq == 0x1d))
783                         ;
784                 else if (!(req->cmd_flags & REQ_QUIET))
785                         scsi_print_sense("", cmd);
786                 result = 0;
787                 /* BLOCK_PC may have set error */
788                 error = 0;
789         }
790
791         /*
792          * A number of bytes were successfully read.  If there
793          * are leftovers and there is some kind of error
794          * (result != 0), retry the rest.
795          */
796         if (scsi_end_request(cmd, error, good_bytes, result == 0) == NULL)
797                 return;
798
799         error = -EIO;
800
801         if (host_byte(result) == DID_RESET) {
802                 /* Third party bus reset or reset for error recovery
803                  * reasons.  Just retry the command and see what
804                  * happens.
805                  */
806                 action = ACTION_RETRY;
807         } else if (sense_valid && !sense_deferred) {
808                 switch (sshdr.sense_key) {
809                 case UNIT_ATTENTION:
810                         if (cmd->device->removable) {
811                                 /* Detected disc change.  Set a bit
812                                  * and quietly refuse further access.
813                                  */
814                                 cmd->device->changed = 1;
815                                 description = "Media Changed";
816                                 action = ACTION_FAIL;
817                         } else {
818                                 /* Must have been a power glitch, or a
819                                  * bus reset.  Could not have been a
820                                  * media change, so we just retry the
821                                  * command and see what happens.
822                                  */
823                                 action = ACTION_RETRY;
824                         }
825                         break;
826                 case ILLEGAL_REQUEST:
827                         /* If we had an ILLEGAL REQUEST returned, then
828                          * we may have performed an unsupported
829                          * command.  The only thing this should be
830                          * would be a ten byte read where only a six
831                          * byte read was supported.  Also, on a system
832                          * where READ CAPACITY failed, we may have
833                          * read past the end of the disk.
834                          */
835                         if ((cmd->device->use_10_for_rw &&
836                             sshdr.asc == 0x20 && sshdr.ascq == 0x00) &&
837                             (cmd->cmnd[0] == READ_10 ||
838                              cmd->cmnd[0] == WRITE_10)) {
839                                 /* This will issue a new 6-byte command. */
840                                 cmd->device->use_10_for_rw = 0;
841                                 action = ACTION_REPREP;
842                         } else if (sshdr.asc == 0x10) /* DIX */ {
843                                 description = "Host Data Integrity Failure";
844                                 action = ACTION_FAIL;
845                                 error = -EILSEQ;
846                         } else
847                                 action = ACTION_FAIL;
848                         break;
849                 case ABORTED_COMMAND:
850                         action = ACTION_FAIL;
851                         if (sshdr.asc == 0x10) { /* DIF */
852                                 description = "Target Data Integrity Failure";
853                                 error = -EILSEQ;
854                         }
855                         break;
856                 case NOT_READY:
857                         /* If the device is in the process of becoming
858                          * ready, or has a temporary blockage, retry.
859                          */
860                         if (sshdr.asc == 0x04) {
861                                 switch (sshdr.ascq) {
862                                 case 0x01: /* becoming ready */
863                                 case 0x04: /* format in progress */
864                                 case 0x05: /* rebuild in progress */
865                                 case 0x06: /* recalculation in progress */
866                                 case 0x07: /* operation in progress */
867                                 case 0x08: /* Long write in progress */
868                                 case 0x09: /* self test in progress */
869                                 case 0x14: /* space allocation in progress */
870                                         action = ACTION_DELAYED_RETRY;
871                                         break;
872                                 default:
873                                         description = "Device not ready";
874                                         action = ACTION_FAIL;
875                                         break;
876                                 }
877                         } else {
878                                 description = "Device not ready";
879                                 action = ACTION_FAIL;
880                         }
881                         break;
882                 case VOLUME_OVERFLOW:
883                         /* See SSC3rXX or current. */
884                         action = ACTION_FAIL;
885                         break;
886                 default:
887                         description = "Unhandled sense code";
888                         action = ACTION_FAIL;
889                         break;
890                 }
891         } else {
892                 description = "Unhandled error code";
893                 action = ACTION_FAIL;
894         }
895
896         switch (action) {
897         case ACTION_FAIL:
898                 /* Give up and fail the remainder of the request */
899                 scsi_release_buffers(cmd);
900                 if (!(req->cmd_flags & REQ_QUIET)) {
901                         if (description)
902                                 scmd_printk(KERN_INFO, cmd, "%s\n",
903                                             description);
904                         scsi_print_result(cmd);
905                         if (driver_byte(result) & DRIVER_SENSE)
906                                 scsi_print_sense("", cmd);
907                         scsi_print_command(cmd);
908                 }
909                 if (blk_end_request_err(req, error))
910                         scsi_requeue_command(q, cmd);
911                 else
912                         scsi_next_command(cmd);
913                 break;
914         case ACTION_REPREP:
915                 /* Unprep the request and put it back at the head of the queue.
916                  * A new command will be prepared and issued.
917                  */
918                 scsi_release_buffers(cmd);
919                 scsi_requeue_command(q, cmd);
920                 break;
921         case ACTION_RETRY:
922                 /* Retry the same command immediately */
923                 __scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY, 0);
924                 break;
925         case ACTION_DELAYED_RETRY:
926                 /* Retry the same command after a delay */
927                 __scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY, 0);
928                 break;
929         }
930 }
931
932 static int scsi_init_sgtable(struct request *req, struct scsi_data_buffer *sdb,
933                              gfp_t gfp_mask)
934 {
935         int count;
936
937         /*
938          * If sg table allocation fails, requeue request later.
939          */
940         if (unlikely(scsi_alloc_sgtable(sdb, req->nr_phys_segments,
941                                         gfp_mask))) {
942                 return BLKPREP_DEFER;
943         }
944
945         req->buffer = NULL;
946
947         /* 
948          * Next, walk the list, and fill in the addresses and sizes of
949          * each segment.
950          */
951         count = blk_rq_map_sg(req->q, req, sdb->table.sgl);
952         BUG_ON(count > sdb->table.nents);
953         sdb->table.nents = count;
954         sdb->length = blk_rq_bytes(req);
955         return BLKPREP_OK;
956 }
957
958 /*
959  * Function:    scsi_init_io()
960  *
961  * Purpose:     SCSI I/O initialize function.
962  *
963  * Arguments:   cmd   - Command descriptor we wish to initialize
964  *
965  * Returns:     0 on success
966  *              BLKPREP_DEFER if the failure is retryable
967  *              BLKPREP_KILL if the failure is fatal
968  */
969 int scsi_init_io(struct scsi_cmnd *cmd, gfp_t gfp_mask)
970 {
971         struct request *rq = cmd->request;
972
973         int error = scsi_init_sgtable(rq, &cmd->sdb, gfp_mask);
974         if (error)
975                 goto err_exit;
976
977         if (blk_bidi_rq(rq)) {
978                 struct scsi_data_buffer *bidi_sdb = kmem_cache_zalloc(
979                         scsi_sdb_cache, GFP_ATOMIC);
980                 if (!bidi_sdb) {
981                         error = BLKPREP_DEFER;
982                         goto err_exit;
983                 }
984
985                 rq->next_rq->special = bidi_sdb;
986                 error = scsi_init_sgtable(rq->next_rq, bidi_sdb, GFP_ATOMIC);
987                 if (error)
988                         goto err_exit;
989         }
990
991         if (blk_integrity_rq(rq)) {
992                 struct scsi_data_buffer *prot_sdb = cmd->prot_sdb;
993                 int ivecs, count;
994
995                 BUG_ON(prot_sdb == NULL);
996                 ivecs = blk_rq_count_integrity_sg(rq->q, rq->bio);
997
998                 if (scsi_alloc_sgtable(prot_sdb, ivecs, gfp_mask)) {
999                         error = BLKPREP_DEFER;
1000                         goto err_exit;
1001                 }
1002
1003                 count = blk_rq_map_integrity_sg(rq->q, rq->bio,
1004                                                 prot_sdb->table.sgl);
1005                 BUG_ON(unlikely(count > ivecs));
1006                 BUG_ON(unlikely(count > queue_max_integrity_segments(rq->q)));
1007
1008                 cmd->prot_sdb = prot_sdb;
1009                 cmd->prot_sdb->table.nents = count;
1010         }
1011
1012         return BLKPREP_OK ;
1013
1014 err_exit:
1015         scsi_release_buffers(cmd);
1016         cmd->request->special = NULL;
1017         scsi_put_command(cmd);
1018         return error;
1019 }
1020 EXPORT_SYMBOL(scsi_init_io);
1021
1022 static struct scsi_cmnd *scsi_get_cmd_from_req(struct scsi_device *sdev,
1023                 struct request *req)
1024 {
1025         struct scsi_cmnd *cmd;
1026
1027         if (!req->special) {
1028                 cmd = scsi_get_command(sdev, GFP_ATOMIC);
1029                 if (unlikely(!cmd))
1030                         return NULL;
1031                 req->special = cmd;
1032         } else {
1033                 cmd = req->special;
1034         }
1035
1036         /* pull a tag out of the request if we have one */
1037         cmd->tag = req->tag;
1038         cmd->request = req;
1039
1040         cmd->cmnd = req->cmd;
1041
1042         return cmd;
1043 }
1044
1045 int scsi_setup_blk_pc_cmnd(struct scsi_device *sdev, struct request *req)
1046 {
1047         struct scsi_cmnd *cmd;
1048         int ret = scsi_prep_state_check(sdev, req);
1049
1050         if (ret != BLKPREP_OK)
1051                 return ret;
1052
1053         cmd = scsi_get_cmd_from_req(sdev, req);
1054         if (unlikely(!cmd))
1055                 return BLKPREP_DEFER;
1056
1057         /*
1058          * BLOCK_PC requests may transfer data, in which case they must
1059          * a bio attached to them.  Or they might contain a SCSI command
1060          * that does not transfer data, in which case they may optionally
1061          * submit a request without an attached bio.
1062          */
1063         if (req->bio) {
1064                 int ret;
1065
1066                 BUG_ON(!req->nr_phys_segments);
1067
1068                 ret = scsi_init_io(cmd, GFP_ATOMIC);
1069                 if (unlikely(ret))
1070                         return ret;
1071         } else {
1072                 BUG_ON(blk_rq_bytes(req));
1073
1074                 memset(&cmd->sdb, 0, sizeof(cmd->sdb));
1075                 req->buffer = NULL;
1076         }
1077
1078         cmd->cmd_len = req->cmd_len;
1079         if (!blk_rq_bytes(req))
1080                 cmd->sc_data_direction = DMA_NONE;
1081         else if (rq_data_dir(req) == WRITE)
1082                 cmd->sc_data_direction = DMA_TO_DEVICE;
1083         else
1084                 cmd->sc_data_direction = DMA_FROM_DEVICE;
1085         
1086         cmd->transfersize = blk_rq_bytes(req);
1087         cmd->allowed = req->retries;
1088         return BLKPREP_OK;
1089 }
1090 EXPORT_SYMBOL(scsi_setup_blk_pc_cmnd);
1091
1092 /*
1093  * Setup a REQ_TYPE_FS command.  These are simple read/write request
1094  * from filesystems that still need to be translated to SCSI CDBs from
1095  * the ULD.
1096  */
1097 int scsi_setup_fs_cmnd(struct scsi_device *sdev, struct request *req)
1098 {
1099         struct scsi_cmnd *cmd;
1100         int ret = scsi_prep_state_check(sdev, req);
1101
1102         if (ret != BLKPREP_OK)
1103                 return ret;
1104
1105         if (unlikely(sdev->scsi_dh_data && sdev->scsi_dh_data->scsi_dh
1106                          && sdev->scsi_dh_data->scsi_dh->prep_fn)) {
1107                 ret = sdev->scsi_dh_data->scsi_dh->prep_fn(sdev, req);
1108                 if (ret != BLKPREP_OK)
1109                         return ret;
1110         }
1111
1112         /*
1113          * Filesystem requests must transfer data.
1114          */
1115         BUG_ON(!req->nr_phys_segments);
1116
1117         cmd = scsi_get_cmd_from_req(sdev, req);
1118         if (unlikely(!cmd))
1119                 return BLKPREP_DEFER;
1120
1121         memset(cmd->cmnd, 0, BLK_MAX_CDB);
1122         return scsi_init_io(cmd, GFP_ATOMIC);
1123 }
1124 EXPORT_SYMBOL(scsi_setup_fs_cmnd);
1125
1126 int scsi_prep_state_check(struct scsi_device *sdev, struct request *req)
1127 {
1128         int ret = BLKPREP_OK;
1129
1130         /*
1131          * If the device is not in running state we will reject some
1132          * or all commands.
1133          */
1134         if (unlikely(sdev->sdev_state != SDEV_RUNNING)) {
1135                 switch (sdev->sdev_state) {
1136                 case SDEV_OFFLINE:
1137                         /*
1138                          * If the device is offline we refuse to process any
1139                          * commands.  The device must be brought online
1140                          * before trying any recovery commands.
1141                          */
1142                         sdev_printk(KERN_ERR, sdev,
1143                                     "rejecting I/O to offline device\n");
1144                         ret = BLKPREP_KILL;
1145                         break;
1146                 case SDEV_DEL:
1147                         /*
1148                          * If the device is fully deleted, we refuse to
1149                          * process any commands as well.
1150                          */
1151                         sdev_printk(KERN_ERR, sdev,
1152                                     "rejecting I/O to dead device\n");
1153                         ret = BLKPREP_KILL;
1154                         break;
1155                 case SDEV_QUIESCE:
1156                 case SDEV_BLOCK:
1157                 case SDEV_CREATED_BLOCK:
1158                         /*
1159                          * If the devices is blocked we defer normal commands.
1160                          */
1161                         if (!(req->cmd_flags & REQ_PREEMPT))
1162                                 ret = BLKPREP_DEFER;
1163                         break;
1164                 default:
1165                         /*
1166                          * For any other not fully online state we only allow
1167                          * special commands.  In particular any user initiated
1168                          * command is not allowed.
1169                          */
1170                         if (!(req->cmd_flags & REQ_PREEMPT))
1171                                 ret = BLKPREP_KILL;
1172                         break;
1173                 }
1174         }
1175         return ret;
1176 }
1177 EXPORT_SYMBOL(scsi_prep_state_check);
1178
1179 int scsi_prep_return(struct request_queue *q, struct request *req, int ret)
1180 {
1181         struct scsi_device *sdev = q->queuedata;
1182
1183         switch (ret) {
1184         case BLKPREP_KILL:
1185                 req->errors = DID_NO_CONNECT << 16;
1186                 /* release the command and kill it */
1187                 if (req->special) {
1188                         struct scsi_cmnd *cmd = req->special;
1189                         scsi_release_buffers(cmd);
1190                         scsi_put_command(cmd);
1191                         req->special = NULL;
1192                 }
1193                 break;
1194         case BLKPREP_DEFER:
1195                 /*
1196                  * If we defer, the blk_peek_request() returns NULL, but the
1197                  * queue must be restarted, so we plug here if no returning
1198                  * command will automatically do that.
1199                  */
1200                 if (sdev->device_busy == 0)
1201                         blk_plug_device(q);
1202                 break;
1203         default:
1204                 req->cmd_flags |= REQ_DONTPREP;
1205         }
1206
1207         return ret;
1208 }
1209 EXPORT_SYMBOL(scsi_prep_return);
1210
1211 int scsi_prep_fn(struct request_queue *q, struct request *req)
1212 {
1213         struct scsi_device *sdev = q->queuedata;
1214         int ret = BLKPREP_KILL;
1215
1216         if (req->cmd_type == REQ_TYPE_BLOCK_PC)
1217                 ret = scsi_setup_blk_pc_cmnd(sdev, req);
1218         return scsi_prep_return(q, req, ret);
1219 }
1220 EXPORT_SYMBOL(scsi_prep_fn);
1221
1222 /*
1223  * scsi_dev_queue_ready: if we can send requests to sdev, return 1 else
1224  * return 0.
1225  *
1226  * Called with the queue_lock held.
1227  */
1228 static inline int scsi_dev_queue_ready(struct request_queue *q,
1229                                   struct scsi_device *sdev)
1230 {
1231         if (sdev->device_busy == 0 && sdev->device_blocked) {
1232                 /*
1233                  * unblock after device_blocked iterates to zero
1234                  */
1235                 if (--sdev->device_blocked == 0) {
1236                         SCSI_LOG_MLQUEUE(3,
1237                                    sdev_printk(KERN_INFO, sdev,
1238                                    "unblocking device at zero depth\n"));
1239                 } else {
1240                         blk_plug_device(q);
1241                         return 0;
1242                 }
1243         }
1244         if (scsi_device_is_busy(sdev))
1245                 return 0;
1246
1247         return 1;
1248 }
1249
1250
1251 /*
1252  * scsi_target_queue_ready: checks if there we can send commands to target
1253  * @sdev: scsi device on starget to check.
1254  *
1255  * Called with the host lock held.
1256  */
1257 static inline int scsi_target_queue_ready(struct Scsi_Host *shost,
1258                                            struct scsi_device *sdev)
1259 {
1260         struct scsi_target *starget = scsi_target(sdev);
1261
1262         if (starget->single_lun) {
1263                 if (starget->starget_sdev_user &&
1264                     starget->starget_sdev_user != sdev)
1265                         return 0;
1266                 starget->starget_sdev_user = sdev;
1267         }
1268
1269         if (starget->target_busy == 0 && starget->target_blocked) {
1270                 /*
1271                  * unblock after target_blocked iterates to zero
1272                  */
1273                 if (--starget->target_blocked == 0) {
1274                         SCSI_LOG_MLQUEUE(3, starget_printk(KERN_INFO, starget,
1275                                          "unblocking target at zero depth\n"));
1276                 } else
1277                         return 0;
1278         }
1279
1280         if (scsi_target_is_busy(starget)) {
1281                 if (list_empty(&sdev->starved_entry)) {
1282                         list_add_tail(&sdev->starved_entry,
1283                                       &shost->starved_list);
1284                         return 0;
1285                 }
1286         }
1287
1288         /* We're OK to process the command, so we can't be starved */
1289         if (!list_empty(&sdev->starved_entry))
1290                 list_del_init(&sdev->starved_entry);
1291         return 1;
1292 }
1293
1294 /*
1295  * scsi_host_queue_ready: if we can send requests to shost, return 1 else
1296  * return 0. We must end up running the queue again whenever 0 is
1297  * returned, else IO can hang.
1298  *
1299  * Called with host_lock held.
1300  */
1301 static inline int scsi_host_queue_ready(struct request_queue *q,
1302                                    struct Scsi_Host *shost,
1303                                    struct scsi_device *sdev)
1304 {
1305         if (scsi_host_in_recovery(shost))
1306                 return 0;
1307         if (shost->host_busy == 0 && shost->host_blocked) {
1308                 /*
1309                  * unblock after host_blocked iterates to zero
1310                  */
1311                 if (--shost->host_blocked == 0) {
1312                         SCSI_LOG_MLQUEUE(3,
1313                                 printk("scsi%d unblocking host at zero depth\n",
1314                                         shost->host_no));
1315                 } else {
1316                         return 0;
1317                 }
1318         }
1319         if (scsi_host_is_busy(shost)) {
1320                 if (list_empty(&sdev->starved_entry))
1321                         list_add_tail(&sdev->starved_entry, &shost->starved_list);
1322                 return 0;
1323         }
1324
1325         /* We're OK to process the command, so we can't be starved */
1326         if (!list_empty(&sdev->starved_entry))
1327                 list_del_init(&sdev->starved_entry);
1328
1329         return 1;
1330 }
1331
1332 /*
1333  * Busy state exporting function for request stacking drivers.
1334  *
1335  * For efficiency, no lock is taken to check the busy state of
1336  * shost/starget/sdev, since the returned value is not guaranteed and
1337  * may be changed after request stacking drivers call the function,
1338  * regardless of taking lock or not.
1339  *
1340  * When scsi can't dispatch I/Os anymore and needs to kill I/Os
1341  * (e.g. !sdev), scsi needs to return 'not busy'.
1342  * Otherwise, request stacking drivers may hold requests forever.
1343  */
1344 static int scsi_lld_busy(struct request_queue *q)
1345 {
1346         struct scsi_device *sdev = q->queuedata;
1347         struct Scsi_Host *shost;
1348         struct scsi_target *starget;
1349
1350         if (!sdev)
1351                 return 0;
1352
1353         shost = sdev->host;
1354         starget = scsi_target(sdev);
1355
1356         if (scsi_host_in_recovery(shost) || scsi_host_is_busy(shost) ||
1357             scsi_target_is_busy(starget) || scsi_device_is_busy(sdev))
1358                 return 1;
1359
1360         return 0;
1361 }
1362
1363 /*
1364  * Kill a request for a dead device
1365  */
1366 static void scsi_kill_request(struct request *req, struct request_queue *q)
1367 {
1368         struct scsi_cmnd *cmd = req->special;
1369         struct scsi_device *sdev;
1370         struct scsi_target *starget;
1371         struct Scsi_Host *shost;
1372
1373         blk_start_request(req);
1374
1375         sdev = cmd->device;
1376         starget = scsi_target(sdev);
1377         shost = sdev->host;
1378         scsi_init_cmd_errh(cmd);
1379         cmd->result = DID_NO_CONNECT << 16;
1380         atomic_inc(&cmd->device->iorequest_cnt);
1381
1382         /*
1383          * SCSI request completion path will do scsi_device_unbusy(),
1384          * bump busy counts.  To bump the counters, we need to dance
1385          * with the locks as normal issue path does.
1386          */
1387         sdev->device_busy++;
1388         spin_unlock(sdev->request_queue->queue_lock);
1389         spin_lock(shost->host_lock);
1390         shost->host_busy++;
1391         starget->target_busy++;
1392         spin_unlock(shost->host_lock);
1393         spin_lock(sdev->request_queue->queue_lock);
1394
1395         blk_complete_request(req);
1396 }
1397
1398 static void scsi_softirq_done(struct request *rq)
1399 {
1400         struct scsi_cmnd *cmd = rq->special;
1401         unsigned long wait_for = (cmd->allowed + 1) * rq->timeout;
1402         int disposition;
1403
1404         INIT_LIST_HEAD(&cmd->eh_entry);
1405
1406         /*
1407          * Set the serial numbers back to zero
1408          */
1409         cmd->serial_number = 0;
1410
1411         atomic_inc(&cmd->device->iodone_cnt);
1412         if (cmd->result)
1413                 atomic_inc(&cmd->device->ioerr_cnt);
1414
1415         disposition = scsi_decide_disposition(cmd);
1416         if (disposition != SUCCESS &&
1417             time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) {
1418                 sdev_printk(KERN_ERR, cmd->device,
1419                             "timing out command, waited %lus\n",
1420                             wait_for/HZ);
1421                 disposition = SUCCESS;
1422         }
1423                         
1424         scsi_log_completion(cmd, disposition);
1425
1426         switch (disposition) {
1427                 case SUCCESS:
1428                         scsi_finish_command(cmd);
1429                         break;
1430                 case NEEDS_RETRY:
1431                         scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY);
1432                         break;
1433                 case ADD_TO_MLQUEUE:
1434                         scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY);
1435                         break;
1436                 default:
1437                         if (!scsi_eh_scmd_add(cmd, 0))
1438                                 scsi_finish_command(cmd);
1439         }
1440 }
1441
1442 /*
1443  * Function:    scsi_request_fn()
1444  *
1445  * Purpose:     Main strategy routine for SCSI.
1446  *
1447  * Arguments:   q       - Pointer to actual queue.
1448  *
1449  * Returns:     Nothing
1450  *
1451  * Lock status: IO request lock assumed to be held when called.
1452  */
1453 static void scsi_request_fn(struct request_queue *q)
1454 {
1455         struct scsi_device *sdev = q->queuedata;
1456         struct Scsi_Host *shost;
1457         struct scsi_cmnd *cmd;
1458         struct request *req;
1459
1460         if (!sdev) {
1461                 printk("scsi: killing requests for dead queue\n");
1462                 while ((req = blk_peek_request(q)) != NULL)
1463                         scsi_kill_request(req, q);
1464                 return;
1465         }
1466
1467         if(!get_device(&sdev->sdev_gendev))
1468                 /* We must be tearing the block queue down already */
1469                 return;
1470
1471         /*
1472          * To start with, we keep looping until the queue is empty, or until
1473          * the host is no longer able to accept any more requests.
1474          */
1475         shost = sdev->host;
1476         while (!blk_queue_plugged(q)) {
1477                 int rtn;
1478                 /*
1479                  * get next queueable request.  We do this early to make sure
1480                  * that the request is fully prepared even if we cannot 
1481                  * accept it.
1482                  */
1483                 req = blk_peek_request(q);
1484                 if (!req || !scsi_dev_queue_ready(q, sdev))
1485                         break;
1486
1487                 if (unlikely(!scsi_device_online(sdev))) {
1488                         sdev_printk(KERN_ERR, sdev,
1489                                     "rejecting I/O to offline device\n");
1490                         scsi_kill_request(req, q);
1491                         continue;
1492                 }
1493
1494
1495                 /*
1496                  * Remove the request from the request list.
1497                  */
1498                 if (!(blk_queue_tagged(q) && !blk_queue_start_tag(q, req)))
1499                         blk_start_request(req);
1500                 sdev->device_busy++;
1501
1502                 spin_unlock(q->queue_lock);
1503                 cmd = req->special;
1504                 if (unlikely(cmd == NULL)) {
1505                         printk(KERN_CRIT "impossible request in %s.\n"
1506                                          "please mail a stack trace to "
1507                                          "linux-scsi@vger.kernel.org\n",
1508                                          __func__);
1509                         blk_dump_rq_flags(req, "foo");
1510                         BUG();
1511                 }
1512                 spin_lock(shost->host_lock);
1513
1514                 /*
1515                  * We hit this when the driver is using a host wide
1516                  * tag map. For device level tag maps the queue_depth check
1517                  * in the device ready fn would prevent us from trying
1518                  * to allocate a tag. Since the map is a shared host resource
1519                  * we add the dev to the starved list so it eventually gets
1520                  * a run when a tag is freed.
1521                  */
1522                 if (blk_queue_tagged(q) && !blk_rq_tagged(req)) {
1523                         if (list_empty(&sdev->starved_entry))
1524                                 list_add_tail(&sdev->starved_entry,
1525                                               &shost->starved_list);
1526                         goto not_ready;
1527                 }
1528
1529                 if (!scsi_target_queue_ready(shost, sdev))
1530                         goto not_ready;
1531
1532                 if (!scsi_host_queue_ready(q, shost, sdev))
1533                         goto not_ready;
1534
1535                 scsi_target(sdev)->target_busy++;
1536                 shost->host_busy++;
1537
1538                 /*
1539                  * XXX(hch): This is rather suboptimal, scsi_dispatch_cmd will
1540                  *              take the lock again.
1541                  */
1542                 spin_unlock_irq(shost->host_lock);
1543
1544                 /*
1545                  * Finally, initialize any error handling parameters, and set up
1546                  * the timers for timeouts.
1547                  */
1548                 scsi_init_cmd_errh(cmd);
1549
1550                 /*
1551                  * Dispatch the command to the low-level driver.
1552                  */
1553                 rtn = scsi_dispatch_cmd(cmd);
1554                 spin_lock_irq(q->queue_lock);
1555                 if(rtn) {
1556                         /* we're refusing the command; because of
1557                          * the way locks get dropped, we need to 
1558                          * check here if plugging is required */
1559                         if(sdev->device_busy == 0)
1560                                 blk_plug_device(q);
1561
1562                         break;
1563                 }
1564         }
1565
1566         goto out;
1567
1568  not_ready:
1569         spin_unlock_irq(shost->host_lock);
1570
1571         /*
1572          * lock q, handle tag, requeue req, and decrement device_busy. We
1573          * must return with queue_lock held.
1574          *
1575          * Decrementing device_busy without checking it is OK, as all such
1576          * cases (host limits or settings) should run the queue at some
1577          * later time.
1578          */
1579         spin_lock_irq(q->queue_lock);
1580         blk_requeue_request(q, req);
1581         sdev->device_busy--;
1582         if(sdev->device_busy == 0)
1583                 blk_plug_device(q);
1584  out:
1585         /* must be careful here...if we trigger the ->remove() function
1586          * we cannot be holding the q lock */
1587         spin_unlock_irq(q->queue_lock);
1588         put_device(&sdev->sdev_gendev);
1589         spin_lock_irq(q->queue_lock);
1590 }
1591
1592 u64 scsi_calculate_bounce_limit(struct Scsi_Host *shost)
1593 {
1594         struct device *host_dev;
1595         u64 bounce_limit = 0xffffffff;
1596
1597         if (shost->unchecked_isa_dma)
1598                 return BLK_BOUNCE_ISA;
1599         /*
1600          * Platforms with virtual-DMA translation
1601          * hardware have no practical limit.
1602          */
1603         if (!PCI_DMA_BUS_IS_PHYS)
1604                 return BLK_BOUNCE_ANY;
1605
1606         host_dev = scsi_get_device(shost);
1607         if (host_dev && host_dev->dma_mask)
1608                 bounce_limit = *host_dev->dma_mask;
1609
1610         return bounce_limit;
1611 }
1612 EXPORT_SYMBOL(scsi_calculate_bounce_limit);
1613
1614 struct request_queue *__scsi_alloc_queue(struct Scsi_Host *shost,
1615                                          request_fn_proc *request_fn)
1616 {
1617         struct request_queue *q;
1618         struct device *dev = shost->shost_gendev.parent;
1619
1620         q = blk_init_queue(request_fn, NULL);
1621         if (!q)
1622                 return NULL;
1623
1624         /*
1625          * this limit is imposed by hardware restrictions
1626          */
1627         blk_queue_max_segments(q, min_t(unsigned short, shost->sg_tablesize,
1628                                         SCSI_MAX_SG_CHAIN_SEGMENTS));
1629
1630         if (scsi_host_prot_dma(shost)) {
1631                 shost->sg_prot_tablesize =
1632                         min_not_zero(shost->sg_prot_tablesize,
1633                                      (unsigned short)SCSI_MAX_PROT_SG_SEGMENTS);
1634                 BUG_ON(shost->sg_prot_tablesize < shost->sg_tablesize);
1635                 blk_queue_max_integrity_segments(q, shost->sg_prot_tablesize);
1636         }
1637
1638         blk_queue_max_hw_sectors(q, shost->max_sectors);
1639         blk_queue_bounce_limit(q, scsi_calculate_bounce_limit(shost));
1640         blk_queue_segment_boundary(q, shost->dma_boundary);
1641         dma_set_seg_boundary(dev, shost->dma_boundary);
1642
1643         blk_queue_max_segment_size(q, dma_get_max_seg_size(dev));
1644
1645         /* New queue, no concurrency on queue_flags */
1646         if (!shost->use_clustering)
1647                 queue_flag_clear_unlocked(QUEUE_FLAG_CLUSTER, q);
1648
1649         /*
1650          * set a reasonable default alignment on word boundaries: the
1651          * host and device may alter it using
1652          * blk_queue_update_dma_alignment() later.
1653          */
1654         blk_queue_dma_alignment(q, 0x03);
1655
1656         return q;
1657 }
1658 EXPORT_SYMBOL(__scsi_alloc_queue);
1659
1660 struct request_queue *scsi_alloc_queue(struct scsi_device *sdev)
1661 {
1662         struct request_queue *q;
1663
1664         q = __scsi_alloc_queue(sdev->host, scsi_request_fn);
1665         if (!q)
1666                 return NULL;
1667
1668         blk_queue_prep_rq(q, scsi_prep_fn);
1669         blk_queue_softirq_done(q, scsi_softirq_done);
1670         blk_queue_rq_timed_out(q, scsi_times_out);
1671         blk_queue_lld_busy(q, scsi_lld_busy);
1672         return q;
1673 }
1674
1675 void scsi_free_queue(struct request_queue *q)
1676 {
1677         blk_cleanup_queue(q);
1678 }
1679
1680 /*
1681  * Function:    scsi_block_requests()
1682  *
1683  * Purpose:     Utility function used by low-level drivers to prevent further
1684  *              commands from being queued to the device.
1685  *
1686  * Arguments:   shost       - Host in question
1687  *
1688  * Returns:     Nothing
1689  *
1690  * Lock status: No locks are assumed held.
1691  *
1692  * Notes:       There is no timer nor any other means by which the requests
1693  *              get unblocked other than the low-level driver calling
1694  *              scsi_unblock_requests().
1695  */
1696 void scsi_block_requests(struct Scsi_Host *shost)
1697 {
1698         shost->host_self_blocked = 1;
1699 }
1700 EXPORT_SYMBOL(scsi_block_requests);
1701
1702 /*
1703  * Function:    scsi_unblock_requests()
1704  *
1705  * Purpose:     Utility function used by low-level drivers to allow further
1706  *              commands from being queued to the device.
1707  *
1708  * Arguments:   shost       - Host in question
1709  *
1710  * Returns:     Nothing
1711  *
1712  * Lock status: No locks are assumed held.
1713  *
1714  * Notes:       There is no timer nor any other means by which the requests
1715  *              get unblocked other than the low-level driver calling
1716  *              scsi_unblock_requests().
1717  *
1718  *              This is done as an API function so that changes to the
1719  *              internals of the scsi mid-layer won't require wholesale
1720  *              changes to drivers that use this feature.
1721  */
1722 void scsi_unblock_requests(struct Scsi_Host *shost)
1723 {
1724         shost->host_self_blocked = 0;
1725         scsi_run_host_queues(shost);
1726 }
1727 EXPORT_SYMBOL(scsi_unblock_requests);
1728
1729 int __init scsi_init_queue(void)
1730 {
1731         int i;
1732
1733         scsi_sdb_cache = kmem_cache_create("scsi_data_buffer",
1734                                            sizeof(struct scsi_data_buffer),
1735                                            0, 0, NULL);
1736         if (!scsi_sdb_cache) {
1737                 printk(KERN_ERR "SCSI: can't init scsi sdb cache\n");
1738                 return -ENOMEM;
1739         }
1740
1741         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1742                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1743                 int size = sgp->size * sizeof(struct scatterlist);
1744
1745                 sgp->slab = kmem_cache_create(sgp->name, size, 0,
1746                                 SLAB_HWCACHE_ALIGN, NULL);
1747                 if (!sgp->slab) {
1748                         printk(KERN_ERR "SCSI: can't init sg slab %s\n",
1749                                         sgp->name);
1750                         goto cleanup_sdb;
1751                 }
1752
1753                 sgp->pool = mempool_create_slab_pool(SG_MEMPOOL_SIZE,
1754                                                      sgp->slab);
1755                 if (!sgp->pool) {
1756                         printk(KERN_ERR "SCSI: can't init sg mempool %s\n",
1757                                         sgp->name);
1758                         goto cleanup_sdb;
1759                 }
1760         }
1761
1762         return 0;
1763
1764 cleanup_sdb:
1765         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1766                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1767                 if (sgp->pool)
1768                         mempool_destroy(sgp->pool);
1769                 if (sgp->slab)
1770                         kmem_cache_destroy(sgp->slab);
1771         }
1772         kmem_cache_destroy(scsi_sdb_cache);
1773
1774         return -ENOMEM;
1775 }
1776
1777 void scsi_exit_queue(void)
1778 {
1779         int i;
1780
1781         kmem_cache_destroy(scsi_sdb_cache);
1782
1783         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1784                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1785                 mempool_destroy(sgp->pool);
1786                 kmem_cache_destroy(sgp->slab);
1787         }
1788 }
1789
1790 /**
1791  *      scsi_mode_select - issue a mode select
1792  *      @sdev:  SCSI device to be queried
1793  *      @pf:    Page format bit (1 == standard, 0 == vendor specific)
1794  *      @sp:    Save page bit (0 == don't save, 1 == save)
1795  *      @modepage: mode page being requested
1796  *      @buffer: request buffer (may not be smaller than eight bytes)
1797  *      @len:   length of request buffer.
1798  *      @timeout: command timeout
1799  *      @retries: number of retries before failing
1800  *      @data: returns a structure abstracting the mode header data
1801  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
1802  *              must be SCSI_SENSE_BUFFERSIZE big.
1803  *
1804  *      Returns zero if successful; negative error number or scsi
1805  *      status on error
1806  *
1807  */
1808 int
1809 scsi_mode_select(struct scsi_device *sdev, int pf, int sp, int modepage,
1810                  unsigned char *buffer, int len, int timeout, int retries,
1811                  struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1812 {
1813         unsigned char cmd[10];
1814         unsigned char *real_buffer;
1815         int ret;
1816
1817         memset(cmd, 0, sizeof(cmd));
1818         cmd[1] = (pf ? 0x10 : 0) | (sp ? 0x01 : 0);
1819
1820         if (sdev->use_10_for_ms) {
1821                 if (len > 65535)
1822                         return -EINVAL;
1823                 real_buffer = kmalloc(8 + len, GFP_KERNEL);
1824                 if (!real_buffer)
1825                         return -ENOMEM;
1826                 memcpy(real_buffer + 8, buffer, len);
1827                 len += 8;
1828                 real_buffer[0] = 0;
1829                 real_buffer[1] = 0;
1830                 real_buffer[2] = data->medium_type;
1831                 real_buffer[3] = data->device_specific;
1832                 real_buffer[4] = data->longlba ? 0x01 : 0;
1833                 real_buffer[5] = 0;
1834                 real_buffer[6] = data->block_descriptor_length >> 8;
1835                 real_buffer[7] = data->block_descriptor_length;
1836
1837                 cmd[0] = MODE_SELECT_10;
1838                 cmd[7] = len >> 8;
1839                 cmd[8] = len;
1840         } else {
1841                 if (len > 255 || data->block_descriptor_length > 255 ||
1842                     data->longlba)
1843                         return -EINVAL;
1844
1845                 real_buffer = kmalloc(4 + len, GFP_KERNEL);
1846                 if (!real_buffer)
1847                         return -ENOMEM;
1848                 memcpy(real_buffer + 4, buffer, len);
1849                 len += 4;
1850                 real_buffer[0] = 0;
1851                 real_buffer[1] = data->medium_type;
1852                 real_buffer[2] = data->device_specific;
1853                 real_buffer[3] = data->block_descriptor_length;
1854                 
1855
1856                 cmd[0] = MODE_SELECT;
1857                 cmd[4] = len;
1858         }
1859
1860         ret = scsi_execute_req(sdev, cmd, DMA_TO_DEVICE, real_buffer, len,
1861                                sshdr, timeout, retries, NULL);
1862         kfree(real_buffer);
1863         return ret;
1864 }
1865 EXPORT_SYMBOL_GPL(scsi_mode_select);
1866
1867 /**
1868  *      scsi_mode_sense - issue a mode sense, falling back from 10 to six bytes if necessary.
1869  *      @sdev:  SCSI device to be queried
1870  *      @dbd:   set if mode sense will allow block descriptors to be returned
1871  *      @modepage: mode page being requested
1872  *      @buffer: request buffer (may not be smaller than eight bytes)
1873  *      @len:   length of request buffer.
1874  *      @timeout: command timeout
1875  *      @retries: number of retries before failing
1876  *      @data: returns a structure abstracting the mode header data
1877  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
1878  *              must be SCSI_SENSE_BUFFERSIZE big.
1879  *
1880  *      Returns zero if unsuccessful, or the header offset (either 4
1881  *      or 8 depending on whether a six or ten byte command was
1882  *      issued) if successful.
1883  */
1884 int
1885 scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
1886                   unsigned char *buffer, int len, int timeout, int retries,
1887                   struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1888 {
1889         unsigned char cmd[12];
1890         int use_10_for_ms;
1891         int header_length;
1892         int result;
1893         struct scsi_sense_hdr my_sshdr;
1894
1895         memset(data, 0, sizeof(*data));
1896         memset(&cmd[0], 0, 12);
1897         cmd[1] = dbd & 0x18;    /* allows DBD and LLBA bits */
1898         cmd[2] = modepage;
1899
1900         /* caller might not be interested in sense, but we need it */
1901         if (!sshdr)
1902                 sshdr = &my_sshdr;
1903
1904  retry:
1905         use_10_for_ms = sdev->use_10_for_ms;
1906
1907         if (use_10_for_ms) {
1908                 if (len < 8)
1909                         len = 8;
1910
1911                 cmd[0] = MODE_SENSE_10;
1912                 cmd[8] = len;
1913                 header_length = 8;
1914         } else {
1915                 if (len < 4)
1916                         len = 4;
1917
1918                 cmd[0] = MODE_SENSE;
1919                 cmd[4] = len;
1920                 header_length = 4;
1921         }
1922
1923         memset(buffer, 0, len);
1924
1925         result = scsi_execute_req(sdev, cmd, DMA_FROM_DEVICE, buffer, len,
1926                                   sshdr, timeout, retries, NULL);
1927
1928         /* This code looks awful: what it's doing is making sure an
1929          * ILLEGAL REQUEST sense return identifies the actual command
1930          * byte as the problem.  MODE_SENSE commands can return
1931          * ILLEGAL REQUEST if the code page isn't supported */
1932
1933         if (use_10_for_ms && !scsi_status_is_good(result) &&
1934             (driver_byte(result) & DRIVER_SENSE)) {
1935                 if (scsi_sense_valid(sshdr)) {
1936                         if ((sshdr->sense_key == ILLEGAL_REQUEST) &&
1937                             (sshdr->asc == 0x20) && (sshdr->ascq == 0)) {
1938                                 /* 
1939                                  * Invalid command operation code
1940                                  */
1941                                 sdev->use_10_for_ms = 0;
1942                                 goto retry;
1943                         }
1944                 }
1945         }
1946
1947         if(scsi_status_is_good(result)) {
1948                 if (unlikely(buffer[0] == 0x86 && buffer[1] == 0x0b &&
1949                              (modepage == 6 || modepage == 8))) {
1950                         /* Initio breakage? */
1951                         header_length = 0;
1952                         data->length = 13;
1953                         data->medium_type = 0;
1954                         data->device_specific = 0;
1955                         data->longlba = 0;
1956                         data->block_descriptor_length = 0;
1957                 } else if(use_10_for_ms) {
1958                         data->length = buffer[0]*256 + buffer[1] + 2;
1959                         data->medium_type = buffer[2];
1960                         data->device_specific = buffer[3];
1961                         data->longlba = buffer[4] & 0x01;
1962                         data->block_descriptor_length = buffer[6]*256
1963                                 + buffer[7];
1964                 } else {
1965                         data->length = buffer[0] + 1;
1966                         data->medium_type = buffer[1];
1967                         data->device_specific = buffer[2];
1968                         data->block_descriptor_length = buffer[3];
1969                 }
1970                 data->header_length = header_length;
1971         }
1972
1973         return result;
1974 }
1975 EXPORT_SYMBOL(scsi_mode_sense);
1976
1977 /**
1978  *      scsi_test_unit_ready - test if unit is ready
1979  *      @sdev:  scsi device to change the state of.
1980  *      @timeout: command timeout
1981  *      @retries: number of retries before failing
1982  *      @sshdr_external: Optional pointer to struct scsi_sense_hdr for
1983  *              returning sense. Make sure that this is cleared before passing
1984  *              in.
1985  *
1986  *      Returns zero if unsuccessful or an error if TUR failed.  For
1987  *      removable media, a return of NOT_READY or UNIT_ATTENTION is
1988  *      translated to success, with the ->changed flag updated.
1989  **/
1990 int
1991 scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries,
1992                      struct scsi_sense_hdr *sshdr_external)
1993 {
1994         char cmd[] = {
1995                 TEST_UNIT_READY, 0, 0, 0, 0, 0,
1996         };
1997         struct scsi_sense_hdr *sshdr;
1998         int result;
1999
2000         if (!sshdr_external)
2001                 sshdr = kzalloc(sizeof(*sshdr), GFP_KERNEL);
2002         else
2003                 sshdr = sshdr_external;
2004
2005         /* try to eat the UNIT_ATTENTION if there are enough retries */
2006         do {
2007                 result = scsi_execute_req(sdev, cmd, DMA_NONE, NULL, 0, sshdr,
2008                                           timeout, retries, NULL);
2009                 if (sdev->removable && scsi_sense_valid(sshdr) &&
2010                     sshdr->sense_key == UNIT_ATTENTION)
2011                         sdev->changed = 1;
2012         } while (scsi_sense_valid(sshdr) &&
2013                  sshdr->sense_key == UNIT_ATTENTION && --retries);
2014
2015         if (!sshdr)
2016                 /* could not allocate sense buffer, so can't process it */
2017                 return result;
2018
2019         if (sdev->removable && scsi_sense_valid(sshdr) &&
2020             (sshdr->sense_key == UNIT_ATTENTION ||
2021              sshdr->sense_key == NOT_READY)) {
2022                 sdev->changed = 1;
2023                 result = 0;
2024         }
2025         if (!sshdr_external)
2026                 kfree(sshdr);
2027         return result;
2028 }
2029 EXPORT_SYMBOL(scsi_test_unit_ready);
2030
2031 /**
2032  *      scsi_device_set_state - Take the given device through the device state model.
2033  *      @sdev:  scsi device to change the state of.
2034  *      @state: state to change to.
2035  *
2036  *      Returns zero if unsuccessful or an error if the requested 
2037  *      transition is illegal.
2038  */
2039 int
2040 scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
2041 {
2042         enum scsi_device_state oldstate = sdev->sdev_state;
2043
2044         if (state == oldstate)
2045                 return 0;
2046
2047         switch (state) {
2048         case SDEV_CREATED:
2049                 switch (oldstate) {
2050                 case SDEV_CREATED_BLOCK:
2051                         break;
2052                 default:
2053                         goto illegal;
2054                 }
2055                 break;
2056                         
2057         case SDEV_RUNNING:
2058                 switch (oldstate) {
2059                 case SDEV_CREATED:
2060                 case SDEV_OFFLINE:
2061                 case SDEV_QUIESCE:
2062                 case SDEV_BLOCK:
2063                         break;
2064                 default:
2065                         goto illegal;
2066                 }
2067                 break;
2068
2069         case SDEV_QUIESCE:
2070                 switch (oldstate) {
2071                 case SDEV_RUNNING:
2072                 case SDEV_OFFLINE:
2073                         break;
2074                 default:
2075                         goto illegal;
2076                 }
2077                 break;
2078
2079         case SDEV_OFFLINE:
2080                 switch (oldstate) {
2081                 case SDEV_CREATED:
2082                 case SDEV_RUNNING:
2083                 case SDEV_QUIESCE:
2084                 case SDEV_BLOCK:
2085                         break;
2086                 default:
2087                         goto illegal;
2088                 }
2089                 break;
2090
2091         case SDEV_BLOCK:
2092                 switch (oldstate) {
2093                 case SDEV_RUNNING:
2094                 case SDEV_CREATED_BLOCK:
2095                         break;
2096                 default:
2097                         goto illegal;
2098                 }
2099                 break;
2100
2101         case SDEV_CREATED_BLOCK:
2102                 switch (oldstate) {
2103                 case SDEV_CREATED:
2104                         break;
2105                 default:
2106                         goto illegal;
2107                 }
2108                 break;
2109
2110         case SDEV_CANCEL:
2111                 switch (oldstate) {
2112                 case SDEV_CREATED:
2113                 case SDEV_RUNNING:
2114                 case SDEV_QUIESCE:
2115                 case SDEV_OFFLINE:
2116                 case SDEV_BLOCK:
2117                         break;
2118                 default:
2119                         goto illegal;
2120                 }
2121                 break;
2122
2123         case SDEV_DEL:
2124                 switch (oldstate) {
2125                 case SDEV_CREATED:
2126                 case SDEV_RUNNING:
2127                 case SDEV_OFFLINE:
2128                 case SDEV_CANCEL:
2129                         break;
2130                 default:
2131                         goto illegal;
2132                 }
2133                 break;
2134
2135         }
2136         sdev->sdev_state = state;
2137         return 0;
2138
2139  illegal:
2140         SCSI_LOG_ERROR_RECOVERY(1, 
2141                                 sdev_printk(KERN_ERR, sdev,
2142                                             "Illegal state transition %s->%s\n",
2143                                             scsi_device_state_name(oldstate),
2144                                             scsi_device_state_name(state))
2145                                 );
2146         return -EINVAL;
2147 }
2148 EXPORT_SYMBOL(scsi_device_set_state);
2149
2150 /**
2151  *      sdev_evt_emit - emit a single SCSI device uevent
2152  *      @sdev: associated SCSI device
2153  *      @evt: event to emit
2154  *
2155  *      Send a single uevent (scsi_event) to the associated scsi_device.
2156  */
2157 static void scsi_evt_emit(struct scsi_device *sdev, struct scsi_event *evt)
2158 {
2159         int idx = 0;
2160         char *envp[3];
2161
2162         switch (evt->evt_type) {
2163         case SDEV_EVT_MEDIA_CHANGE:
2164                 envp[idx++] = "SDEV_MEDIA_CHANGE=1";
2165                 break;
2166
2167         default:
2168                 /* do nothing */
2169                 break;
2170         }
2171
2172         envp[idx++] = NULL;
2173
2174         kobject_uevent_env(&sdev->sdev_gendev.kobj, KOBJ_CHANGE, envp);
2175 }
2176
2177 /**
2178  *      sdev_evt_thread - send a uevent for each scsi event
2179  *      @work: work struct for scsi_device
2180  *
2181  *      Dispatch queued events to their associated scsi_device kobjects
2182  *      as uevents.
2183  */
2184 void scsi_evt_thread(struct work_struct *work)
2185 {
2186         struct scsi_device *sdev;
2187         LIST_HEAD(event_list);
2188
2189         sdev = container_of(work, struct scsi_device, event_work);
2190
2191         while (1) {
2192                 struct scsi_event *evt;
2193                 struct list_head *this, *tmp;
2194                 unsigned long flags;
2195
2196                 spin_lock_irqsave(&sdev->list_lock, flags);
2197                 list_splice_init(&sdev->event_list, &event_list);
2198                 spin_unlock_irqrestore(&sdev->list_lock, flags);
2199
2200                 if (list_empty(&event_list))
2201                         break;
2202
2203                 list_for_each_safe(this, tmp, &event_list) {
2204                         evt = list_entry(this, struct scsi_event, node);
2205                         list_del(&evt->node);
2206                         scsi_evt_emit(sdev, evt);
2207                         kfree(evt);
2208                 }
2209         }
2210 }
2211
2212 /**
2213  *      sdev_evt_send - send asserted event to uevent thread
2214  *      @sdev: scsi_device event occurred on
2215  *      @evt: event to send
2216  *
2217  *      Assert scsi device event asynchronously.
2218  */
2219 void sdev_evt_send(struct scsi_device *sdev, struct scsi_event *evt)
2220 {
2221         unsigned long flags;
2222
2223 #if 0
2224         /* FIXME: currently this check eliminates all media change events
2225          * for polled devices.  Need to update to discriminate between AN
2226          * and polled events */
2227         if (!test_bit(evt->evt_type, sdev->supported_events)) {
2228                 kfree(evt);
2229                 return;
2230         }
2231 #endif
2232
2233         spin_lock_irqsave(&sdev->list_lock, flags);
2234         list_add_tail(&evt->node, &sdev->event_list);
2235         schedule_work(&sdev->event_work);
2236         spin_unlock_irqrestore(&sdev->list_lock, flags);
2237 }
2238 EXPORT_SYMBOL_GPL(sdev_evt_send);
2239
2240 /**
2241  *      sdev_evt_alloc - allocate a new scsi event
2242  *      @evt_type: type of event to allocate
2243  *      @gfpflags: GFP flags for allocation
2244  *
2245  *      Allocates and returns a new scsi_event.
2246  */
2247 struct scsi_event *sdev_evt_alloc(enum scsi_device_event evt_type,
2248                                   gfp_t gfpflags)
2249 {
2250         struct scsi_event *evt = kzalloc(sizeof(struct scsi_event), gfpflags);
2251         if (!evt)
2252                 return NULL;
2253
2254         evt->evt_type = evt_type;
2255         INIT_LIST_HEAD(&evt->node);
2256
2257         /* evt_type-specific initialization, if any */
2258         switch (evt_type) {
2259         case SDEV_EVT_MEDIA_CHANGE:
2260         default:
2261                 /* do nothing */
2262                 break;
2263         }
2264
2265         return evt;
2266 }
2267 EXPORT_SYMBOL_GPL(sdev_evt_alloc);
2268
2269 /**
2270  *      sdev_evt_send_simple - send asserted event to uevent thread
2271  *      @sdev: scsi_device event occurred on
2272  *      @evt_type: type of event to send
2273  *      @gfpflags: GFP flags for allocation
2274  *
2275  *      Assert scsi device event asynchronously, given an event type.
2276  */
2277 void sdev_evt_send_simple(struct scsi_device *sdev,
2278                           enum scsi_device_event evt_type, gfp_t gfpflags)
2279 {
2280         struct scsi_event *evt = sdev_evt_alloc(evt_type, gfpflags);
2281         if (!evt) {
2282                 sdev_printk(KERN_ERR, sdev, "event %d eaten due to OOM\n",
2283                             evt_type);
2284                 return;
2285         }
2286
2287         sdev_evt_send(sdev, evt);
2288 }
2289 EXPORT_SYMBOL_GPL(sdev_evt_send_simple);
2290
2291 /**
2292  *      scsi_device_quiesce - Block user issued commands.
2293  *      @sdev:  scsi device to quiesce.
2294  *
2295  *      This works by trying to transition to the SDEV_QUIESCE state
2296  *      (which must be a legal transition).  When the device is in this
2297  *      state, only special requests will be accepted, all others will
2298  *      be deferred.  Since special requests may also be requeued requests,
2299  *      a successful return doesn't guarantee the device will be 
2300  *      totally quiescent.
2301  *
2302  *      Must be called with user context, may sleep.
2303  *
2304  *      Returns zero if unsuccessful or an error if not.
2305  */
2306 int
2307 scsi_device_quiesce(struct scsi_device *sdev)
2308 {
2309         int err = scsi_device_set_state(sdev, SDEV_QUIESCE);
2310         if (err)
2311                 return err;
2312
2313         scsi_run_queue(sdev->request_queue);
2314         while (sdev->device_busy) {
2315                 msleep_interruptible(200);
2316                 scsi_run_queue(sdev->request_queue);
2317         }
2318         return 0;
2319 }
2320 EXPORT_SYMBOL(scsi_device_quiesce);
2321
2322 /**
2323  *      scsi_device_resume - Restart user issued commands to a quiesced device.
2324  *      @sdev:  scsi device to resume.
2325  *
2326  *      Moves the device from quiesced back to running and restarts the
2327  *      queues.
2328  *
2329  *      Must be called with user context, may sleep.
2330  */
2331 void
2332 scsi_device_resume(struct scsi_device *sdev)
2333 {
2334         if(scsi_device_set_state(sdev, SDEV_RUNNING))
2335                 return;
2336         scsi_run_queue(sdev->request_queue);
2337 }
2338 EXPORT_SYMBOL(scsi_device_resume);
2339
2340 static void
2341 device_quiesce_fn(struct scsi_device *sdev, void *data)
2342 {
2343         scsi_device_quiesce(sdev);
2344 }
2345
2346 void
2347 scsi_target_quiesce(struct scsi_target *starget)
2348 {
2349         starget_for_each_device(starget, NULL, device_quiesce_fn);
2350 }
2351 EXPORT_SYMBOL(scsi_target_quiesce);
2352
2353 static void
2354 device_resume_fn(struct scsi_device *sdev, void *data)
2355 {
2356         scsi_device_resume(sdev);
2357 }
2358
2359 void
2360 scsi_target_resume(struct scsi_target *starget)
2361 {
2362         starget_for_each_device(starget, NULL, device_resume_fn);
2363 }
2364 EXPORT_SYMBOL(scsi_target_resume);
2365
2366 /**
2367  * scsi_internal_device_block - internal function to put a device temporarily into the SDEV_BLOCK state
2368  * @sdev:       device to block
2369  *
2370  * Block request made by scsi lld's to temporarily stop all
2371  * scsi commands on the specified device.  Called from interrupt
2372  * or normal process context.
2373  *
2374  * Returns zero if successful or error if not
2375  *
2376  * Notes:       
2377  *      This routine transitions the device to the SDEV_BLOCK state
2378  *      (which must be a legal transition).  When the device is in this
2379  *      state, all commands are deferred until the scsi lld reenables
2380  *      the device with scsi_device_unblock or device_block_tmo fires.
2381  *      This routine assumes the host_lock is held on entry.
2382  */
2383 int
2384 scsi_internal_device_block(struct scsi_device *sdev)
2385 {
2386         struct request_queue *q = sdev->request_queue;
2387         unsigned long flags;
2388         int err = 0;
2389
2390         err = scsi_device_set_state(sdev, SDEV_BLOCK);
2391         if (err) {
2392                 err = scsi_device_set_state(sdev, SDEV_CREATED_BLOCK);
2393
2394                 if (err)
2395                         return err;
2396         }
2397
2398         /* 
2399          * The device has transitioned to SDEV_BLOCK.  Stop the
2400          * block layer from calling the midlayer with this device's
2401          * request queue. 
2402          */
2403         spin_lock_irqsave(q->queue_lock, flags);
2404         blk_stop_queue(q);
2405         spin_unlock_irqrestore(q->queue_lock, flags);
2406
2407         return 0;
2408 }
2409 EXPORT_SYMBOL_GPL(scsi_internal_device_block);
2410  
2411 /**
2412  * scsi_internal_device_unblock - resume a device after a block request
2413  * @sdev:       device to resume
2414  *
2415  * Called by scsi lld's or the midlayer to restart the device queue
2416  * for the previously suspended scsi device.  Called from interrupt or
2417  * normal process context.
2418  *
2419  * Returns zero if successful or error if not.
2420  *
2421  * Notes:       
2422  *      This routine transitions the device to the SDEV_RUNNING state
2423  *      (which must be a legal transition) allowing the midlayer to
2424  *      goose the queue for this device.  This routine assumes the 
2425  *      host_lock is held upon entry.
2426  */
2427 int
2428 scsi_internal_device_unblock(struct scsi_device *sdev)
2429 {
2430         struct request_queue *q = sdev->request_queue; 
2431         unsigned long flags;
2432         
2433         /* 
2434          * Try to transition the scsi device to SDEV_RUNNING
2435          * and goose the device queue if successful.  
2436          */
2437         if (sdev->sdev_state == SDEV_BLOCK)
2438                 sdev->sdev_state = SDEV_RUNNING;
2439         else if (sdev->sdev_state == SDEV_CREATED_BLOCK)
2440                 sdev->sdev_state = SDEV_CREATED;
2441         else
2442                 return -EINVAL;
2443
2444         spin_lock_irqsave(q->queue_lock, flags);
2445         blk_start_queue(q);
2446         spin_unlock_irqrestore(q->queue_lock, flags);
2447
2448         return 0;
2449 }
2450 EXPORT_SYMBOL_GPL(scsi_internal_device_unblock);
2451
2452 static void
2453 device_block(struct scsi_device *sdev, void *data)
2454 {
2455         scsi_internal_device_block(sdev);
2456 }
2457
2458 static int
2459 target_block(struct device *dev, void *data)
2460 {
2461         if (scsi_is_target_device(dev))
2462                 starget_for_each_device(to_scsi_target(dev), NULL,
2463                                         device_block);
2464         return 0;
2465 }
2466
2467 void
2468 scsi_target_block(struct device *dev)
2469 {
2470         if (scsi_is_target_device(dev))
2471                 starget_for_each_device(to_scsi_target(dev), NULL,
2472                                         device_block);
2473         else
2474                 device_for_each_child(dev, NULL, target_block);
2475 }
2476 EXPORT_SYMBOL_GPL(scsi_target_block);
2477
2478 static void
2479 device_unblock(struct scsi_device *sdev, void *data)
2480 {
2481         scsi_internal_device_unblock(sdev);
2482 }
2483
2484 static int
2485 target_unblock(struct device *dev, void *data)
2486 {
2487         if (scsi_is_target_device(dev))
2488                 starget_for_each_device(to_scsi_target(dev), NULL,
2489                                         device_unblock);
2490         return 0;
2491 }
2492
2493 void
2494 scsi_target_unblock(struct device *dev)
2495 {
2496         if (scsi_is_target_device(dev))
2497                 starget_for_each_device(to_scsi_target(dev), NULL,
2498                                         device_unblock);
2499         else
2500                 device_for_each_child(dev, NULL, target_unblock);
2501 }
2502 EXPORT_SYMBOL_GPL(scsi_target_unblock);
2503
2504 /**
2505  * scsi_kmap_atomic_sg - find and atomically map an sg-elemnt
2506  * @sgl:        scatter-gather list
2507  * @sg_count:   number of segments in sg
2508  * @offset:     offset in bytes into sg, on return offset into the mapped area
2509  * @len:        bytes to map, on return number of bytes mapped
2510  *
2511  * Returns virtual address of the start of the mapped page
2512  */
2513 void *scsi_kmap_atomic_sg(struct scatterlist *sgl, int sg_count,
2514                           size_t *offset, size_t *len)
2515 {
2516         int i;
2517         size_t sg_len = 0, len_complete = 0;
2518         struct scatterlist *sg;
2519         struct page *page;
2520
2521         WARN_ON(!irqs_disabled());
2522
2523         for_each_sg(sgl, sg, sg_count, i) {
2524                 len_complete = sg_len; /* Complete sg-entries */
2525                 sg_len += sg->length;
2526                 if (sg_len > *offset)
2527                         break;
2528         }
2529
2530         if (unlikely(i == sg_count)) {
2531                 printk(KERN_ERR "%s: Bytes in sg: %zu, requested offset %zu, "
2532                         "elements %d\n",
2533                        __func__, sg_len, *offset, sg_count);
2534                 WARN_ON(1);
2535                 return NULL;
2536         }
2537
2538         /* Offset starting from the beginning of first page in this sg-entry */
2539         *offset = *offset - len_complete + sg->offset;
2540
2541         /* Assumption: contiguous pages can be accessed as "page + i" */
2542         page = nth_page(sg_page(sg), (*offset >> PAGE_SHIFT));
2543         *offset &= ~PAGE_MASK;
2544
2545         /* Bytes in this sg-entry from *offset to the end of the page */
2546         sg_len = PAGE_SIZE - *offset;
2547         if (*len > sg_len)
2548                 *len = sg_len;
2549
2550         return kmap_atomic(page, KM_BIO_SRC_IRQ);
2551 }
2552 EXPORT_SYMBOL(scsi_kmap_atomic_sg);
2553
2554 /**
2555  * scsi_kunmap_atomic_sg - atomically unmap a virtual address, previously mapped with scsi_kmap_atomic_sg
2556  * @virt:       virtual address to be unmapped
2557  */
2558 void scsi_kunmap_atomic_sg(void *virt)
2559 {
2560         kunmap_atomic(virt, KM_BIO_SRC_IRQ);
2561 }
2562 EXPORT_SYMBOL(scsi_kunmap_atomic_sg);