[SCSI] bnx2fc: Broadcom FCoE offload driver
[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 static int __scsi_error_from_host_byte(struct scsi_cmnd *cmd, int result)
671 {
672         int error = 0;
673
674         switch(host_byte(result)) {
675         case DID_TRANSPORT_FAILFAST:
676                 error = -ENOLINK;
677                 break;
678         case DID_TARGET_FAILURE:
679                 cmd->result |= (DID_OK << 16);
680                 error = -EREMOTEIO;
681                 break;
682         case DID_NEXUS_FAILURE:
683                 cmd->result |= (DID_OK << 16);
684                 error = -EBADE;
685                 break;
686         default:
687                 error = -EIO;
688                 break;
689         }
690
691         return error;
692 }
693
694 /*
695  * Function:    scsi_io_completion()
696  *
697  * Purpose:     Completion processing for block device I/O requests.
698  *
699  * Arguments:   cmd   - command that is finished.
700  *
701  * Lock status: Assumed that no lock is held upon entry.
702  *
703  * Returns:     Nothing
704  *
705  * Notes:       This function is matched in terms of capabilities to
706  *              the function that created the scatter-gather list.
707  *              In other words, if there are no bounce buffers
708  *              (the normal case for most drivers), we don't need
709  *              the logic to deal with cleaning up afterwards.
710  *
711  *              We must call scsi_end_request().  This will finish off
712  *              the specified number of sectors.  If we are done, the
713  *              command block will be released and the queue function
714  *              will be goosed.  If we are not done then we have to
715  *              figure out what to do next:
716  *
717  *              a) We can call scsi_requeue_command().  The request
718  *                 will be unprepared and put back on the queue.  Then
719  *                 a new command will be created for it.  This should
720  *                 be used if we made forward progress, or if we want
721  *                 to switch from READ(10) to READ(6) for example.
722  *
723  *              b) We can call scsi_queue_insert().  The request will
724  *                 be put back on the queue and retried using the same
725  *                 command as before, possibly after a delay.
726  *
727  *              c) We can call blk_end_request() with -EIO to fail
728  *                 the remainder of the request.
729  */
730 void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes)
731 {
732         int result = cmd->result;
733         struct request_queue *q = cmd->device->request_queue;
734         struct request *req = cmd->request;
735         int error = 0;
736         struct scsi_sense_hdr sshdr;
737         int sense_valid = 0;
738         int sense_deferred = 0;
739         enum {ACTION_FAIL, ACTION_REPREP, ACTION_RETRY,
740               ACTION_DELAYED_RETRY} action;
741         char *description = NULL;
742
743         if (result) {
744                 sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
745                 if (sense_valid)
746                         sense_deferred = scsi_sense_is_deferred(&sshdr);
747         }
748
749         if (req->cmd_type == REQ_TYPE_BLOCK_PC) { /* SG_IO ioctl from block level */
750                 req->errors = result;
751                 if (result) {
752                         if (sense_valid && req->sense) {
753                                 /*
754                                  * SG_IO wants current and deferred errors
755                                  */
756                                 int len = 8 + cmd->sense_buffer[7];
757
758                                 if (len > SCSI_SENSE_BUFFERSIZE)
759                                         len = SCSI_SENSE_BUFFERSIZE;
760                                 memcpy(req->sense, cmd->sense_buffer,  len);
761                                 req->sense_len = len;
762                         }
763                         if (!sense_deferred)
764                                 error = __scsi_error_from_host_byte(cmd, result);
765                 }
766
767                 req->resid_len = scsi_get_resid(cmd);
768
769                 if (scsi_bidi_cmnd(cmd)) {
770                         /*
771                          * Bidi commands Must be complete as a whole,
772                          * both sides at once.
773                          */
774                         req->next_rq->resid_len = scsi_in(cmd)->resid;
775
776                         scsi_release_buffers(cmd);
777                         blk_end_request_all(req, 0);
778
779                         scsi_next_command(cmd);
780                         return;
781                 }
782         }
783
784         /* no bidi support for !REQ_TYPE_BLOCK_PC yet */
785         BUG_ON(blk_bidi_rq(req));
786
787         /*
788          * Next deal with any sectors which we were able to correctly
789          * handle.
790          */
791         SCSI_LOG_HLCOMPLETE(1, printk("%u sectors total, "
792                                       "%d bytes done.\n",
793                                       blk_rq_sectors(req), good_bytes));
794
795         /*
796          * Recovered errors need reporting, but they're always treated
797          * as success, so fiddle the result code here.  For BLOCK_PC
798          * we already took a copy of the original into rq->errors which
799          * is what gets returned to the user
800          */
801         if (sense_valid && (sshdr.sense_key == RECOVERED_ERROR)) {
802                 /* if ATA PASS-THROUGH INFORMATION AVAILABLE skip
803                  * print since caller wants ATA registers. Only occurs on
804                  * SCSI ATA PASS_THROUGH commands when CK_COND=1
805                  */
806                 if ((sshdr.asc == 0x0) && (sshdr.ascq == 0x1d))
807                         ;
808                 else if (!(req->cmd_flags & REQ_QUIET))
809                         scsi_print_sense("", cmd);
810                 result = 0;
811                 /* BLOCK_PC may have set error */
812                 error = 0;
813         }
814
815         /*
816          * A number of bytes were successfully read.  If there
817          * are leftovers and there is some kind of error
818          * (result != 0), retry the rest.
819          */
820         if (scsi_end_request(cmd, error, good_bytes, result == 0) == NULL)
821                 return;
822
823         error = __scsi_error_from_host_byte(cmd, result);
824
825         if (host_byte(result) == DID_RESET) {
826                 /* Third party bus reset or reset for error recovery
827                  * reasons.  Just retry the command and see what
828                  * happens.
829                  */
830                 action = ACTION_RETRY;
831         } else if (sense_valid && !sense_deferred) {
832                 switch (sshdr.sense_key) {
833                 case UNIT_ATTENTION:
834                         if (cmd->device->removable) {
835                                 /* Detected disc change.  Set a bit
836                                  * and quietly refuse further access.
837                                  */
838                                 cmd->device->changed = 1;
839                                 description = "Media Changed";
840                                 action = ACTION_FAIL;
841                         } else {
842                                 /* Must have been a power glitch, or a
843                                  * bus reset.  Could not have been a
844                                  * media change, so we just retry the
845                                  * command and see what happens.
846                                  */
847                                 action = ACTION_RETRY;
848                         }
849                         break;
850                 case ILLEGAL_REQUEST:
851                         /* If we had an ILLEGAL REQUEST returned, then
852                          * we may have performed an unsupported
853                          * command.  The only thing this should be
854                          * would be a ten byte read where only a six
855                          * byte read was supported.  Also, on a system
856                          * where READ CAPACITY failed, we may have
857                          * read past the end of the disk.
858                          */
859                         if ((cmd->device->use_10_for_rw &&
860                             sshdr.asc == 0x20 && sshdr.ascq == 0x00) &&
861                             (cmd->cmnd[0] == READ_10 ||
862                              cmd->cmnd[0] == WRITE_10)) {
863                                 /* This will issue a new 6-byte command. */
864                                 cmd->device->use_10_for_rw = 0;
865                                 action = ACTION_REPREP;
866                         } else if (sshdr.asc == 0x10) /* DIX */ {
867                                 description = "Host Data Integrity Failure";
868                                 action = ACTION_FAIL;
869                                 error = -EILSEQ;
870                         } else
871                                 action = ACTION_FAIL;
872                         break;
873                 case ABORTED_COMMAND:
874                         action = ACTION_FAIL;
875                         if (sshdr.asc == 0x10) { /* DIF */
876                                 description = "Target Data Integrity Failure";
877                                 error = -EILSEQ;
878                         }
879                         break;
880                 case NOT_READY:
881                         /* If the device is in the process of becoming
882                          * ready, or has a temporary blockage, retry.
883                          */
884                         if (sshdr.asc == 0x04) {
885                                 switch (sshdr.ascq) {
886                                 case 0x01: /* becoming ready */
887                                 case 0x04: /* format in progress */
888                                 case 0x05: /* rebuild in progress */
889                                 case 0x06: /* recalculation in progress */
890                                 case 0x07: /* operation in progress */
891                                 case 0x08: /* Long write in progress */
892                                 case 0x09: /* self test in progress */
893                                 case 0x14: /* space allocation in progress */
894                                         action = ACTION_DELAYED_RETRY;
895                                         break;
896                                 default:
897                                         description = "Device not ready";
898                                         action = ACTION_FAIL;
899                                         break;
900                                 }
901                         } else {
902                                 description = "Device not ready";
903                                 action = ACTION_FAIL;
904                         }
905                         break;
906                 case VOLUME_OVERFLOW:
907                         /* See SSC3rXX or current. */
908                         action = ACTION_FAIL;
909                         break;
910                 default:
911                         description = "Unhandled sense code";
912                         action = ACTION_FAIL;
913                         break;
914                 }
915         } else {
916                 description = "Unhandled error code";
917                 action = ACTION_FAIL;
918         }
919
920         switch (action) {
921         case ACTION_FAIL:
922                 /* Give up and fail the remainder of the request */
923                 scsi_release_buffers(cmd);
924                 if (!(req->cmd_flags & REQ_QUIET)) {
925                         if (description)
926                                 scmd_printk(KERN_INFO, cmd, "%s\n",
927                                             description);
928                         scsi_print_result(cmd);
929                         if (driver_byte(result) & DRIVER_SENSE)
930                                 scsi_print_sense("", cmd);
931                         scsi_print_command(cmd);
932                 }
933                 if (blk_end_request_err(req, error))
934                         scsi_requeue_command(q, cmd);
935                 else
936                         scsi_next_command(cmd);
937                 break;
938         case ACTION_REPREP:
939                 /* Unprep the request and put it back at the head of the queue.
940                  * A new command will be prepared and issued.
941                  */
942                 scsi_release_buffers(cmd);
943                 scsi_requeue_command(q, cmd);
944                 break;
945         case ACTION_RETRY:
946                 /* Retry the same command immediately */
947                 __scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY, 0);
948                 break;
949         case ACTION_DELAYED_RETRY:
950                 /* Retry the same command after a delay */
951                 __scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY, 0);
952                 break;
953         }
954 }
955
956 static int scsi_init_sgtable(struct request *req, struct scsi_data_buffer *sdb,
957                              gfp_t gfp_mask)
958 {
959         int count;
960
961         /*
962          * If sg table allocation fails, requeue request later.
963          */
964         if (unlikely(scsi_alloc_sgtable(sdb, req->nr_phys_segments,
965                                         gfp_mask))) {
966                 return BLKPREP_DEFER;
967         }
968
969         req->buffer = NULL;
970
971         /* 
972          * Next, walk the list, and fill in the addresses and sizes of
973          * each segment.
974          */
975         count = blk_rq_map_sg(req->q, req, sdb->table.sgl);
976         BUG_ON(count > sdb->table.nents);
977         sdb->table.nents = count;
978         sdb->length = blk_rq_bytes(req);
979         return BLKPREP_OK;
980 }
981
982 /*
983  * Function:    scsi_init_io()
984  *
985  * Purpose:     SCSI I/O initialize function.
986  *
987  * Arguments:   cmd   - Command descriptor we wish to initialize
988  *
989  * Returns:     0 on success
990  *              BLKPREP_DEFER if the failure is retryable
991  *              BLKPREP_KILL if the failure is fatal
992  */
993 int scsi_init_io(struct scsi_cmnd *cmd, gfp_t gfp_mask)
994 {
995         struct request *rq = cmd->request;
996
997         int error = scsi_init_sgtable(rq, &cmd->sdb, gfp_mask);
998         if (error)
999                 goto err_exit;
1000
1001         if (blk_bidi_rq(rq)) {
1002                 struct scsi_data_buffer *bidi_sdb = kmem_cache_zalloc(
1003                         scsi_sdb_cache, GFP_ATOMIC);
1004                 if (!bidi_sdb) {
1005                         error = BLKPREP_DEFER;
1006                         goto err_exit;
1007                 }
1008
1009                 rq->next_rq->special = bidi_sdb;
1010                 error = scsi_init_sgtable(rq->next_rq, bidi_sdb, GFP_ATOMIC);
1011                 if (error)
1012                         goto err_exit;
1013         }
1014
1015         if (blk_integrity_rq(rq)) {
1016                 struct scsi_data_buffer *prot_sdb = cmd->prot_sdb;
1017                 int ivecs, count;
1018
1019                 BUG_ON(prot_sdb == NULL);
1020                 ivecs = blk_rq_count_integrity_sg(rq->q, rq->bio);
1021
1022                 if (scsi_alloc_sgtable(prot_sdb, ivecs, gfp_mask)) {
1023                         error = BLKPREP_DEFER;
1024                         goto err_exit;
1025                 }
1026
1027                 count = blk_rq_map_integrity_sg(rq->q, rq->bio,
1028                                                 prot_sdb->table.sgl);
1029                 BUG_ON(unlikely(count > ivecs));
1030                 BUG_ON(unlikely(count > queue_max_integrity_segments(rq->q)));
1031
1032                 cmd->prot_sdb = prot_sdb;
1033                 cmd->prot_sdb->table.nents = count;
1034         }
1035
1036         return BLKPREP_OK ;
1037
1038 err_exit:
1039         scsi_release_buffers(cmd);
1040         cmd->request->special = NULL;
1041         scsi_put_command(cmd);
1042         return error;
1043 }
1044 EXPORT_SYMBOL(scsi_init_io);
1045
1046 static struct scsi_cmnd *scsi_get_cmd_from_req(struct scsi_device *sdev,
1047                 struct request *req)
1048 {
1049         struct scsi_cmnd *cmd;
1050
1051         if (!req->special) {
1052                 cmd = scsi_get_command(sdev, GFP_ATOMIC);
1053                 if (unlikely(!cmd))
1054                         return NULL;
1055                 req->special = cmd;
1056         } else {
1057                 cmd = req->special;
1058         }
1059
1060         /* pull a tag out of the request if we have one */
1061         cmd->tag = req->tag;
1062         cmd->request = req;
1063
1064         cmd->cmnd = req->cmd;
1065
1066         return cmd;
1067 }
1068
1069 int scsi_setup_blk_pc_cmnd(struct scsi_device *sdev, struct request *req)
1070 {
1071         struct scsi_cmnd *cmd;
1072         int ret = scsi_prep_state_check(sdev, req);
1073
1074         if (ret != BLKPREP_OK)
1075                 return ret;
1076
1077         cmd = scsi_get_cmd_from_req(sdev, req);
1078         if (unlikely(!cmd))
1079                 return BLKPREP_DEFER;
1080
1081         /*
1082          * BLOCK_PC requests may transfer data, in which case they must
1083          * a bio attached to them.  Or they might contain a SCSI command
1084          * that does not transfer data, in which case they may optionally
1085          * submit a request without an attached bio.
1086          */
1087         if (req->bio) {
1088                 int ret;
1089
1090                 BUG_ON(!req->nr_phys_segments);
1091
1092                 ret = scsi_init_io(cmd, GFP_ATOMIC);
1093                 if (unlikely(ret))
1094                         return ret;
1095         } else {
1096                 BUG_ON(blk_rq_bytes(req));
1097
1098                 memset(&cmd->sdb, 0, sizeof(cmd->sdb));
1099                 req->buffer = NULL;
1100         }
1101
1102         cmd->cmd_len = req->cmd_len;
1103         if (!blk_rq_bytes(req))
1104                 cmd->sc_data_direction = DMA_NONE;
1105         else if (rq_data_dir(req) == WRITE)
1106                 cmd->sc_data_direction = DMA_TO_DEVICE;
1107         else
1108                 cmd->sc_data_direction = DMA_FROM_DEVICE;
1109         
1110         cmd->transfersize = blk_rq_bytes(req);
1111         cmd->allowed = req->retries;
1112         return BLKPREP_OK;
1113 }
1114 EXPORT_SYMBOL(scsi_setup_blk_pc_cmnd);
1115
1116 /*
1117  * Setup a REQ_TYPE_FS command.  These are simple read/write request
1118  * from filesystems that still need to be translated to SCSI CDBs from
1119  * the ULD.
1120  */
1121 int scsi_setup_fs_cmnd(struct scsi_device *sdev, struct request *req)
1122 {
1123         struct scsi_cmnd *cmd;
1124         int ret = scsi_prep_state_check(sdev, req);
1125
1126         if (ret != BLKPREP_OK)
1127                 return ret;
1128
1129         if (unlikely(sdev->scsi_dh_data && sdev->scsi_dh_data->scsi_dh
1130                          && sdev->scsi_dh_data->scsi_dh->prep_fn)) {
1131                 ret = sdev->scsi_dh_data->scsi_dh->prep_fn(sdev, req);
1132                 if (ret != BLKPREP_OK)
1133                         return ret;
1134         }
1135
1136         /*
1137          * Filesystem requests must transfer data.
1138          */
1139         BUG_ON(!req->nr_phys_segments);
1140
1141         cmd = scsi_get_cmd_from_req(sdev, req);
1142         if (unlikely(!cmd))
1143                 return BLKPREP_DEFER;
1144
1145         memset(cmd->cmnd, 0, BLK_MAX_CDB);
1146         return scsi_init_io(cmd, GFP_ATOMIC);
1147 }
1148 EXPORT_SYMBOL(scsi_setup_fs_cmnd);
1149
1150 int scsi_prep_state_check(struct scsi_device *sdev, struct request *req)
1151 {
1152         int ret = BLKPREP_OK;
1153
1154         /*
1155          * If the device is not in running state we will reject some
1156          * or all commands.
1157          */
1158         if (unlikely(sdev->sdev_state != SDEV_RUNNING)) {
1159                 switch (sdev->sdev_state) {
1160                 case SDEV_OFFLINE:
1161                         /*
1162                          * If the device is offline we refuse to process any
1163                          * commands.  The device must be brought online
1164                          * before trying any recovery commands.
1165                          */
1166                         sdev_printk(KERN_ERR, sdev,
1167                                     "rejecting I/O to offline device\n");
1168                         ret = BLKPREP_KILL;
1169                         break;
1170                 case SDEV_DEL:
1171                         /*
1172                          * If the device is fully deleted, we refuse to
1173                          * process any commands as well.
1174                          */
1175                         sdev_printk(KERN_ERR, sdev,
1176                                     "rejecting I/O to dead device\n");
1177                         ret = BLKPREP_KILL;
1178                         break;
1179                 case SDEV_QUIESCE:
1180                 case SDEV_BLOCK:
1181                 case SDEV_CREATED_BLOCK:
1182                         /*
1183                          * If the devices is blocked we defer normal commands.
1184                          */
1185                         if (!(req->cmd_flags & REQ_PREEMPT))
1186                                 ret = BLKPREP_DEFER;
1187                         break;
1188                 default:
1189                         /*
1190                          * For any other not fully online state we only allow
1191                          * special commands.  In particular any user initiated
1192                          * command is not allowed.
1193                          */
1194                         if (!(req->cmd_flags & REQ_PREEMPT))
1195                                 ret = BLKPREP_KILL;
1196                         break;
1197                 }
1198         }
1199         return ret;
1200 }
1201 EXPORT_SYMBOL(scsi_prep_state_check);
1202
1203 int scsi_prep_return(struct request_queue *q, struct request *req, int ret)
1204 {
1205         struct scsi_device *sdev = q->queuedata;
1206
1207         switch (ret) {
1208         case BLKPREP_KILL:
1209                 req->errors = DID_NO_CONNECT << 16;
1210                 /* release the command and kill it */
1211                 if (req->special) {
1212                         struct scsi_cmnd *cmd = req->special;
1213                         scsi_release_buffers(cmd);
1214                         scsi_put_command(cmd);
1215                         req->special = NULL;
1216                 }
1217                 break;
1218         case BLKPREP_DEFER:
1219                 /*
1220                  * If we defer, the blk_peek_request() returns NULL, but the
1221                  * queue must be restarted, so we plug here if no returning
1222                  * command will automatically do that.
1223                  */
1224                 if (sdev->device_busy == 0)
1225                         blk_plug_device(q);
1226                 break;
1227         default:
1228                 req->cmd_flags |= REQ_DONTPREP;
1229         }
1230
1231         return ret;
1232 }
1233 EXPORT_SYMBOL(scsi_prep_return);
1234
1235 int scsi_prep_fn(struct request_queue *q, struct request *req)
1236 {
1237         struct scsi_device *sdev = q->queuedata;
1238         int ret = BLKPREP_KILL;
1239
1240         if (req->cmd_type == REQ_TYPE_BLOCK_PC)
1241                 ret = scsi_setup_blk_pc_cmnd(sdev, req);
1242         return scsi_prep_return(q, req, ret);
1243 }
1244 EXPORT_SYMBOL(scsi_prep_fn);
1245
1246 /*
1247  * scsi_dev_queue_ready: if we can send requests to sdev, return 1 else
1248  * return 0.
1249  *
1250  * Called with the queue_lock held.
1251  */
1252 static inline int scsi_dev_queue_ready(struct request_queue *q,
1253                                   struct scsi_device *sdev)
1254 {
1255         if (sdev->device_busy == 0 && sdev->device_blocked) {
1256                 /*
1257                  * unblock after device_blocked iterates to zero
1258                  */
1259                 if (--sdev->device_blocked == 0) {
1260                         SCSI_LOG_MLQUEUE(3,
1261                                    sdev_printk(KERN_INFO, sdev,
1262                                    "unblocking device at zero depth\n"));
1263                 } else {
1264                         blk_plug_device(q);
1265                         return 0;
1266                 }
1267         }
1268         if (scsi_device_is_busy(sdev))
1269                 return 0;
1270
1271         return 1;
1272 }
1273
1274
1275 /*
1276  * scsi_target_queue_ready: checks if there we can send commands to target
1277  * @sdev: scsi device on starget to check.
1278  *
1279  * Called with the host lock held.
1280  */
1281 static inline int scsi_target_queue_ready(struct Scsi_Host *shost,
1282                                            struct scsi_device *sdev)
1283 {
1284         struct scsi_target *starget = scsi_target(sdev);
1285
1286         if (starget->single_lun) {
1287                 if (starget->starget_sdev_user &&
1288                     starget->starget_sdev_user != sdev)
1289                         return 0;
1290                 starget->starget_sdev_user = sdev;
1291         }
1292
1293         if (starget->target_busy == 0 && starget->target_blocked) {
1294                 /*
1295                  * unblock after target_blocked iterates to zero
1296                  */
1297                 if (--starget->target_blocked == 0) {
1298                         SCSI_LOG_MLQUEUE(3, starget_printk(KERN_INFO, starget,
1299                                          "unblocking target at zero depth\n"));
1300                 } else
1301                         return 0;
1302         }
1303
1304         if (scsi_target_is_busy(starget)) {
1305                 if (list_empty(&sdev->starved_entry))
1306                         list_add_tail(&sdev->starved_entry,
1307                                       &shost->starved_list);
1308                 return 0;
1309         }
1310
1311         /* We're OK to process the command, so we can't be starved */
1312         if (!list_empty(&sdev->starved_entry))
1313                 list_del_init(&sdev->starved_entry);
1314         return 1;
1315 }
1316
1317 /*
1318  * scsi_host_queue_ready: if we can send requests to shost, return 1 else
1319  * return 0. We must end up running the queue again whenever 0 is
1320  * returned, else IO can hang.
1321  *
1322  * Called with host_lock held.
1323  */
1324 static inline int scsi_host_queue_ready(struct request_queue *q,
1325                                    struct Scsi_Host *shost,
1326                                    struct scsi_device *sdev)
1327 {
1328         if (scsi_host_in_recovery(shost))
1329                 return 0;
1330         if (shost->host_busy == 0 && shost->host_blocked) {
1331                 /*
1332                  * unblock after host_blocked iterates to zero
1333                  */
1334                 if (--shost->host_blocked == 0) {
1335                         SCSI_LOG_MLQUEUE(3,
1336                                 printk("scsi%d unblocking host at zero depth\n",
1337                                         shost->host_no));
1338                 } else {
1339                         return 0;
1340                 }
1341         }
1342         if (scsi_host_is_busy(shost)) {
1343                 if (list_empty(&sdev->starved_entry))
1344                         list_add_tail(&sdev->starved_entry, &shost->starved_list);
1345                 return 0;
1346         }
1347
1348         /* We're OK to process the command, so we can't be starved */
1349         if (!list_empty(&sdev->starved_entry))
1350                 list_del_init(&sdev->starved_entry);
1351
1352         return 1;
1353 }
1354
1355 /*
1356  * Busy state exporting function for request stacking drivers.
1357  *
1358  * For efficiency, no lock is taken to check the busy state of
1359  * shost/starget/sdev, since the returned value is not guaranteed and
1360  * may be changed after request stacking drivers call the function,
1361  * regardless of taking lock or not.
1362  *
1363  * When scsi can't dispatch I/Os anymore and needs to kill I/Os
1364  * (e.g. !sdev), scsi needs to return 'not busy'.
1365  * Otherwise, request stacking drivers may hold requests forever.
1366  */
1367 static int scsi_lld_busy(struct request_queue *q)
1368 {
1369         struct scsi_device *sdev = q->queuedata;
1370         struct Scsi_Host *shost;
1371         struct scsi_target *starget;
1372
1373         if (!sdev)
1374                 return 0;
1375
1376         shost = sdev->host;
1377         starget = scsi_target(sdev);
1378
1379         if (scsi_host_in_recovery(shost) || scsi_host_is_busy(shost) ||
1380             scsi_target_is_busy(starget) || scsi_device_is_busy(sdev))
1381                 return 1;
1382
1383         return 0;
1384 }
1385
1386 /*
1387  * Kill a request for a dead device
1388  */
1389 static void scsi_kill_request(struct request *req, struct request_queue *q)
1390 {
1391         struct scsi_cmnd *cmd = req->special;
1392         struct scsi_device *sdev;
1393         struct scsi_target *starget;
1394         struct Scsi_Host *shost;
1395
1396         blk_start_request(req);
1397
1398         sdev = cmd->device;
1399         starget = scsi_target(sdev);
1400         shost = sdev->host;
1401         scsi_init_cmd_errh(cmd);
1402         cmd->result = DID_NO_CONNECT << 16;
1403         atomic_inc(&cmd->device->iorequest_cnt);
1404
1405         /*
1406          * SCSI request completion path will do scsi_device_unbusy(),
1407          * bump busy counts.  To bump the counters, we need to dance
1408          * with the locks as normal issue path does.
1409          */
1410         sdev->device_busy++;
1411         spin_unlock(sdev->request_queue->queue_lock);
1412         spin_lock(shost->host_lock);
1413         shost->host_busy++;
1414         starget->target_busy++;
1415         spin_unlock(shost->host_lock);
1416         spin_lock(sdev->request_queue->queue_lock);
1417
1418         blk_complete_request(req);
1419 }
1420
1421 static void scsi_softirq_done(struct request *rq)
1422 {
1423         struct scsi_cmnd *cmd = rq->special;
1424         unsigned long wait_for = (cmd->allowed + 1) * rq->timeout;
1425         int disposition;
1426
1427         INIT_LIST_HEAD(&cmd->eh_entry);
1428
1429         atomic_inc(&cmd->device->iodone_cnt);
1430         if (cmd->result)
1431                 atomic_inc(&cmd->device->ioerr_cnt);
1432
1433         disposition = scsi_decide_disposition(cmd);
1434         if (disposition != SUCCESS &&
1435             time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) {
1436                 sdev_printk(KERN_ERR, cmd->device,
1437                             "timing out command, waited %lus\n",
1438                             wait_for/HZ);
1439                 disposition = SUCCESS;
1440         }
1441                         
1442         scsi_log_completion(cmd, disposition);
1443
1444         switch (disposition) {
1445                 case SUCCESS:
1446                         scsi_finish_command(cmd);
1447                         break;
1448                 case NEEDS_RETRY:
1449                         scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY);
1450                         break;
1451                 case ADD_TO_MLQUEUE:
1452                         scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY);
1453                         break;
1454                 default:
1455                         if (!scsi_eh_scmd_add(cmd, 0))
1456                                 scsi_finish_command(cmd);
1457         }
1458 }
1459
1460 /*
1461  * Function:    scsi_request_fn()
1462  *
1463  * Purpose:     Main strategy routine for SCSI.
1464  *
1465  * Arguments:   q       - Pointer to actual queue.
1466  *
1467  * Returns:     Nothing
1468  *
1469  * Lock status: IO request lock assumed to be held when called.
1470  */
1471 static void scsi_request_fn(struct request_queue *q)
1472 {
1473         struct scsi_device *sdev = q->queuedata;
1474         struct Scsi_Host *shost;
1475         struct scsi_cmnd *cmd;
1476         struct request *req;
1477
1478         if (!sdev) {
1479                 printk("scsi: killing requests for dead queue\n");
1480                 while ((req = blk_peek_request(q)) != NULL)
1481                         scsi_kill_request(req, q);
1482                 return;
1483         }
1484
1485         if(!get_device(&sdev->sdev_gendev))
1486                 /* We must be tearing the block queue down already */
1487                 return;
1488
1489         /*
1490          * To start with, we keep looping until the queue is empty, or until
1491          * the host is no longer able to accept any more requests.
1492          */
1493         shost = sdev->host;
1494         while (!blk_queue_plugged(q)) {
1495                 int rtn;
1496                 /*
1497                  * get next queueable request.  We do this early to make sure
1498                  * that the request is fully prepared even if we cannot 
1499                  * accept it.
1500                  */
1501                 req = blk_peek_request(q);
1502                 if (!req || !scsi_dev_queue_ready(q, sdev))
1503                         break;
1504
1505                 if (unlikely(!scsi_device_online(sdev))) {
1506                         sdev_printk(KERN_ERR, sdev,
1507                                     "rejecting I/O to offline device\n");
1508                         scsi_kill_request(req, q);
1509                         continue;
1510                 }
1511
1512
1513                 /*
1514                  * Remove the request from the request list.
1515                  */
1516                 if (!(blk_queue_tagged(q) && !blk_queue_start_tag(q, req)))
1517                         blk_start_request(req);
1518                 sdev->device_busy++;
1519
1520                 spin_unlock(q->queue_lock);
1521                 cmd = req->special;
1522                 if (unlikely(cmd == NULL)) {
1523                         printk(KERN_CRIT "impossible request in %s.\n"
1524                                          "please mail a stack trace to "
1525                                          "linux-scsi@vger.kernel.org\n",
1526                                          __func__);
1527                         blk_dump_rq_flags(req, "foo");
1528                         BUG();
1529                 }
1530                 spin_lock(shost->host_lock);
1531
1532                 /*
1533                  * We hit this when the driver is using a host wide
1534                  * tag map. For device level tag maps the queue_depth check
1535                  * in the device ready fn would prevent us from trying
1536                  * to allocate a tag. Since the map is a shared host resource
1537                  * we add the dev to the starved list so it eventually gets
1538                  * a run when a tag is freed.
1539                  */
1540                 if (blk_queue_tagged(q) && !blk_rq_tagged(req)) {
1541                         if (list_empty(&sdev->starved_entry))
1542                                 list_add_tail(&sdev->starved_entry,
1543                                               &shost->starved_list);
1544                         goto not_ready;
1545                 }
1546
1547                 if (!scsi_target_queue_ready(shost, sdev))
1548                         goto not_ready;
1549
1550                 if (!scsi_host_queue_ready(q, shost, sdev))
1551                         goto not_ready;
1552
1553                 scsi_target(sdev)->target_busy++;
1554                 shost->host_busy++;
1555
1556                 /*
1557                  * XXX(hch): This is rather suboptimal, scsi_dispatch_cmd will
1558                  *              take the lock again.
1559                  */
1560                 spin_unlock_irq(shost->host_lock);
1561
1562                 /*
1563                  * Finally, initialize any error handling parameters, and set up
1564                  * the timers for timeouts.
1565                  */
1566                 scsi_init_cmd_errh(cmd);
1567
1568                 /*
1569                  * Dispatch the command to the low-level driver.
1570                  */
1571                 rtn = scsi_dispatch_cmd(cmd);
1572                 spin_lock_irq(q->queue_lock);
1573                 if(rtn) {
1574                         /* we're refusing the command; because of
1575                          * the way locks get dropped, we need to 
1576                          * check here if plugging is required */
1577                         if(sdev->device_busy == 0)
1578                                 blk_plug_device(q);
1579
1580                         break;
1581                 }
1582         }
1583
1584         goto out;
1585
1586  not_ready:
1587         spin_unlock_irq(shost->host_lock);
1588
1589         /*
1590          * lock q, handle tag, requeue req, and decrement device_busy. We
1591          * must return with queue_lock held.
1592          *
1593          * Decrementing device_busy without checking it is OK, as all such
1594          * cases (host limits or settings) should run the queue at some
1595          * later time.
1596          */
1597         spin_lock_irq(q->queue_lock);
1598         blk_requeue_request(q, req);
1599         sdev->device_busy--;
1600         if(sdev->device_busy == 0)
1601                 blk_plug_device(q);
1602  out:
1603         /* must be careful here...if we trigger the ->remove() function
1604          * we cannot be holding the q lock */
1605         spin_unlock_irq(q->queue_lock);
1606         put_device(&sdev->sdev_gendev);
1607         spin_lock_irq(q->queue_lock);
1608 }
1609
1610 u64 scsi_calculate_bounce_limit(struct Scsi_Host *shost)
1611 {
1612         struct device *host_dev;
1613         u64 bounce_limit = 0xffffffff;
1614
1615         if (shost->unchecked_isa_dma)
1616                 return BLK_BOUNCE_ISA;
1617         /*
1618          * Platforms with virtual-DMA translation
1619          * hardware have no practical limit.
1620          */
1621         if (!PCI_DMA_BUS_IS_PHYS)
1622                 return BLK_BOUNCE_ANY;
1623
1624         host_dev = scsi_get_device(shost);
1625         if (host_dev && host_dev->dma_mask)
1626                 bounce_limit = *host_dev->dma_mask;
1627
1628         return bounce_limit;
1629 }
1630 EXPORT_SYMBOL(scsi_calculate_bounce_limit);
1631
1632 struct request_queue *__scsi_alloc_queue(struct Scsi_Host *shost,
1633                                          request_fn_proc *request_fn)
1634 {
1635         struct request_queue *q;
1636         struct device *dev = shost->shost_gendev.parent;
1637
1638         q = blk_init_queue(request_fn, NULL);
1639         if (!q)
1640                 return NULL;
1641
1642         /*
1643          * this limit is imposed by hardware restrictions
1644          */
1645         blk_queue_max_segments(q, min_t(unsigned short, shost->sg_tablesize,
1646                                         SCSI_MAX_SG_CHAIN_SEGMENTS));
1647
1648         if (scsi_host_prot_dma(shost)) {
1649                 shost->sg_prot_tablesize =
1650                         min_not_zero(shost->sg_prot_tablesize,
1651                                      (unsigned short)SCSI_MAX_PROT_SG_SEGMENTS);
1652                 BUG_ON(shost->sg_prot_tablesize < shost->sg_tablesize);
1653                 blk_queue_max_integrity_segments(q, shost->sg_prot_tablesize);
1654         }
1655
1656         blk_queue_max_hw_sectors(q, shost->max_sectors);
1657         blk_queue_bounce_limit(q, scsi_calculate_bounce_limit(shost));
1658         blk_queue_segment_boundary(q, shost->dma_boundary);
1659         dma_set_seg_boundary(dev, shost->dma_boundary);
1660
1661         blk_queue_max_segment_size(q, dma_get_max_seg_size(dev));
1662
1663         if (!shost->use_clustering)
1664                 q->limits.cluster = 0;
1665
1666         /*
1667          * set a reasonable default alignment on word boundaries: the
1668          * host and device may alter it using
1669          * blk_queue_update_dma_alignment() later.
1670          */
1671         blk_queue_dma_alignment(q, 0x03);
1672
1673         return q;
1674 }
1675 EXPORT_SYMBOL(__scsi_alloc_queue);
1676
1677 struct request_queue *scsi_alloc_queue(struct scsi_device *sdev)
1678 {
1679         struct request_queue *q;
1680
1681         q = __scsi_alloc_queue(sdev->host, scsi_request_fn);
1682         if (!q)
1683                 return NULL;
1684
1685         blk_queue_prep_rq(q, scsi_prep_fn);
1686         blk_queue_softirq_done(q, scsi_softirq_done);
1687         blk_queue_rq_timed_out(q, scsi_times_out);
1688         blk_queue_lld_busy(q, scsi_lld_busy);
1689         return q;
1690 }
1691
1692 void scsi_free_queue(struct request_queue *q)
1693 {
1694         blk_cleanup_queue(q);
1695 }
1696
1697 /*
1698  * Function:    scsi_block_requests()
1699  *
1700  * Purpose:     Utility function used by low-level drivers to prevent further
1701  *              commands from being queued to the device.
1702  *
1703  * Arguments:   shost       - Host in question
1704  *
1705  * Returns:     Nothing
1706  *
1707  * Lock status: No locks are assumed held.
1708  *
1709  * Notes:       There is no timer nor any other means by which the requests
1710  *              get unblocked other than the low-level driver calling
1711  *              scsi_unblock_requests().
1712  */
1713 void scsi_block_requests(struct Scsi_Host *shost)
1714 {
1715         shost->host_self_blocked = 1;
1716 }
1717 EXPORT_SYMBOL(scsi_block_requests);
1718
1719 /*
1720  * Function:    scsi_unblock_requests()
1721  *
1722  * Purpose:     Utility function used by low-level drivers to allow further
1723  *              commands from being queued to the device.
1724  *
1725  * Arguments:   shost       - Host in question
1726  *
1727  * Returns:     Nothing
1728  *
1729  * Lock status: No locks are assumed held.
1730  *
1731  * Notes:       There is no timer nor any other means by which the requests
1732  *              get unblocked other than the low-level driver calling
1733  *              scsi_unblock_requests().
1734  *
1735  *              This is done as an API function so that changes to the
1736  *              internals of the scsi mid-layer won't require wholesale
1737  *              changes to drivers that use this feature.
1738  */
1739 void scsi_unblock_requests(struct Scsi_Host *shost)
1740 {
1741         shost->host_self_blocked = 0;
1742         scsi_run_host_queues(shost);
1743 }
1744 EXPORT_SYMBOL(scsi_unblock_requests);
1745
1746 int __init scsi_init_queue(void)
1747 {
1748         int i;
1749
1750         scsi_sdb_cache = kmem_cache_create("scsi_data_buffer",
1751                                            sizeof(struct scsi_data_buffer),
1752                                            0, 0, NULL);
1753         if (!scsi_sdb_cache) {
1754                 printk(KERN_ERR "SCSI: can't init scsi sdb cache\n");
1755                 return -ENOMEM;
1756         }
1757
1758         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1759                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1760                 int size = sgp->size * sizeof(struct scatterlist);
1761
1762                 sgp->slab = kmem_cache_create(sgp->name, size, 0,
1763                                 SLAB_HWCACHE_ALIGN, NULL);
1764                 if (!sgp->slab) {
1765                         printk(KERN_ERR "SCSI: can't init sg slab %s\n",
1766                                         sgp->name);
1767                         goto cleanup_sdb;
1768                 }
1769
1770                 sgp->pool = mempool_create_slab_pool(SG_MEMPOOL_SIZE,
1771                                                      sgp->slab);
1772                 if (!sgp->pool) {
1773                         printk(KERN_ERR "SCSI: can't init sg mempool %s\n",
1774                                         sgp->name);
1775                         goto cleanup_sdb;
1776                 }
1777         }
1778
1779         return 0;
1780
1781 cleanup_sdb:
1782         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1783                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1784                 if (sgp->pool)
1785                         mempool_destroy(sgp->pool);
1786                 if (sgp->slab)
1787                         kmem_cache_destroy(sgp->slab);
1788         }
1789         kmem_cache_destroy(scsi_sdb_cache);
1790
1791         return -ENOMEM;
1792 }
1793
1794 void scsi_exit_queue(void)
1795 {
1796         int i;
1797
1798         kmem_cache_destroy(scsi_sdb_cache);
1799
1800         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1801                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1802                 mempool_destroy(sgp->pool);
1803                 kmem_cache_destroy(sgp->slab);
1804         }
1805 }
1806
1807 /**
1808  *      scsi_mode_select - issue a mode select
1809  *      @sdev:  SCSI device to be queried
1810  *      @pf:    Page format bit (1 == standard, 0 == vendor specific)
1811  *      @sp:    Save page bit (0 == don't save, 1 == save)
1812  *      @modepage: mode page being requested
1813  *      @buffer: request buffer (may not be smaller than eight bytes)
1814  *      @len:   length of request buffer.
1815  *      @timeout: command timeout
1816  *      @retries: number of retries before failing
1817  *      @data: returns a structure abstracting the mode header data
1818  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
1819  *              must be SCSI_SENSE_BUFFERSIZE big.
1820  *
1821  *      Returns zero if successful; negative error number or scsi
1822  *      status on error
1823  *
1824  */
1825 int
1826 scsi_mode_select(struct scsi_device *sdev, int pf, int sp, int modepage,
1827                  unsigned char *buffer, int len, int timeout, int retries,
1828                  struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1829 {
1830         unsigned char cmd[10];
1831         unsigned char *real_buffer;
1832         int ret;
1833
1834         memset(cmd, 0, sizeof(cmd));
1835         cmd[1] = (pf ? 0x10 : 0) | (sp ? 0x01 : 0);
1836
1837         if (sdev->use_10_for_ms) {
1838                 if (len > 65535)
1839                         return -EINVAL;
1840                 real_buffer = kmalloc(8 + len, GFP_KERNEL);
1841                 if (!real_buffer)
1842                         return -ENOMEM;
1843                 memcpy(real_buffer + 8, buffer, len);
1844                 len += 8;
1845                 real_buffer[0] = 0;
1846                 real_buffer[1] = 0;
1847                 real_buffer[2] = data->medium_type;
1848                 real_buffer[3] = data->device_specific;
1849                 real_buffer[4] = data->longlba ? 0x01 : 0;
1850                 real_buffer[5] = 0;
1851                 real_buffer[6] = data->block_descriptor_length >> 8;
1852                 real_buffer[7] = data->block_descriptor_length;
1853
1854                 cmd[0] = MODE_SELECT_10;
1855                 cmd[7] = len >> 8;
1856                 cmd[8] = len;
1857         } else {
1858                 if (len > 255 || data->block_descriptor_length > 255 ||
1859                     data->longlba)
1860                         return -EINVAL;
1861
1862                 real_buffer = kmalloc(4 + len, GFP_KERNEL);
1863                 if (!real_buffer)
1864                         return -ENOMEM;
1865                 memcpy(real_buffer + 4, buffer, len);
1866                 len += 4;
1867                 real_buffer[0] = 0;
1868                 real_buffer[1] = data->medium_type;
1869                 real_buffer[2] = data->device_specific;
1870                 real_buffer[3] = data->block_descriptor_length;
1871                 
1872
1873                 cmd[0] = MODE_SELECT;
1874                 cmd[4] = len;
1875         }
1876
1877         ret = scsi_execute_req(sdev, cmd, DMA_TO_DEVICE, real_buffer, len,
1878                                sshdr, timeout, retries, NULL);
1879         kfree(real_buffer);
1880         return ret;
1881 }
1882 EXPORT_SYMBOL_GPL(scsi_mode_select);
1883
1884 /**
1885  *      scsi_mode_sense - issue a mode sense, falling back from 10 to six bytes if necessary.
1886  *      @sdev:  SCSI device to be queried
1887  *      @dbd:   set if mode sense will allow block descriptors to be returned
1888  *      @modepage: mode page being requested
1889  *      @buffer: request buffer (may not be smaller than eight bytes)
1890  *      @len:   length of request buffer.
1891  *      @timeout: command timeout
1892  *      @retries: number of retries before failing
1893  *      @data: returns a structure abstracting the mode header data
1894  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
1895  *              must be SCSI_SENSE_BUFFERSIZE big.
1896  *
1897  *      Returns zero if unsuccessful, or the header offset (either 4
1898  *      or 8 depending on whether a six or ten byte command was
1899  *      issued) if successful.
1900  */
1901 int
1902 scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
1903                   unsigned char *buffer, int len, int timeout, int retries,
1904                   struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1905 {
1906         unsigned char cmd[12];
1907         int use_10_for_ms;
1908         int header_length;
1909         int result;
1910         struct scsi_sense_hdr my_sshdr;
1911
1912         memset(data, 0, sizeof(*data));
1913         memset(&cmd[0], 0, 12);
1914         cmd[1] = dbd & 0x18;    /* allows DBD and LLBA bits */
1915         cmd[2] = modepage;
1916
1917         /* caller might not be interested in sense, but we need it */
1918         if (!sshdr)
1919                 sshdr = &my_sshdr;
1920
1921  retry:
1922         use_10_for_ms = sdev->use_10_for_ms;
1923
1924         if (use_10_for_ms) {
1925                 if (len < 8)
1926                         len = 8;
1927
1928                 cmd[0] = MODE_SENSE_10;
1929                 cmd[8] = len;
1930                 header_length = 8;
1931         } else {
1932                 if (len < 4)
1933                         len = 4;
1934
1935                 cmd[0] = MODE_SENSE;
1936                 cmd[4] = len;
1937                 header_length = 4;
1938         }
1939
1940         memset(buffer, 0, len);
1941
1942         result = scsi_execute_req(sdev, cmd, DMA_FROM_DEVICE, buffer, len,
1943                                   sshdr, timeout, retries, NULL);
1944
1945         /* This code looks awful: what it's doing is making sure an
1946          * ILLEGAL REQUEST sense return identifies the actual command
1947          * byte as the problem.  MODE_SENSE commands can return
1948          * ILLEGAL REQUEST if the code page isn't supported */
1949
1950         if (use_10_for_ms && !scsi_status_is_good(result) &&
1951             (driver_byte(result) & DRIVER_SENSE)) {
1952                 if (scsi_sense_valid(sshdr)) {
1953                         if ((sshdr->sense_key == ILLEGAL_REQUEST) &&
1954                             (sshdr->asc == 0x20) && (sshdr->ascq == 0)) {
1955                                 /* 
1956                                  * Invalid command operation code
1957                                  */
1958                                 sdev->use_10_for_ms = 0;
1959                                 goto retry;
1960                         }
1961                 }
1962         }
1963
1964         if(scsi_status_is_good(result)) {
1965                 if (unlikely(buffer[0] == 0x86 && buffer[1] == 0x0b &&
1966                              (modepage == 6 || modepage == 8))) {
1967                         /* Initio breakage? */
1968                         header_length = 0;
1969                         data->length = 13;
1970                         data->medium_type = 0;
1971                         data->device_specific = 0;
1972                         data->longlba = 0;
1973                         data->block_descriptor_length = 0;
1974                 } else if(use_10_for_ms) {
1975                         data->length = buffer[0]*256 + buffer[1] + 2;
1976                         data->medium_type = buffer[2];
1977                         data->device_specific = buffer[3];
1978                         data->longlba = buffer[4] & 0x01;
1979                         data->block_descriptor_length = buffer[6]*256
1980                                 + buffer[7];
1981                 } else {
1982                         data->length = buffer[0] + 1;
1983                         data->medium_type = buffer[1];
1984                         data->device_specific = buffer[2];
1985                         data->block_descriptor_length = buffer[3];
1986                 }
1987                 data->header_length = header_length;
1988         }
1989
1990         return result;
1991 }
1992 EXPORT_SYMBOL(scsi_mode_sense);
1993
1994 /**
1995  *      scsi_test_unit_ready - test if unit is ready
1996  *      @sdev:  scsi device to change the state of.
1997  *      @timeout: command timeout
1998  *      @retries: number of retries before failing
1999  *      @sshdr_external: Optional pointer to struct scsi_sense_hdr for
2000  *              returning sense. Make sure that this is cleared before passing
2001  *              in.
2002  *
2003  *      Returns zero if unsuccessful or an error if TUR failed.  For
2004  *      removable media, UNIT_ATTENTION sets ->changed flag.
2005  **/
2006 int
2007 scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries,
2008                      struct scsi_sense_hdr *sshdr_external)
2009 {
2010         char cmd[] = {
2011                 TEST_UNIT_READY, 0, 0, 0, 0, 0,
2012         };
2013         struct scsi_sense_hdr *sshdr;
2014         int result;
2015
2016         if (!sshdr_external)
2017                 sshdr = kzalloc(sizeof(*sshdr), GFP_KERNEL);
2018         else
2019                 sshdr = sshdr_external;
2020
2021         /* try to eat the UNIT_ATTENTION if there are enough retries */
2022         do {
2023                 result = scsi_execute_req(sdev, cmd, DMA_NONE, NULL, 0, sshdr,
2024                                           timeout, retries, NULL);
2025                 if (sdev->removable && scsi_sense_valid(sshdr) &&
2026                     sshdr->sense_key == UNIT_ATTENTION)
2027                         sdev->changed = 1;
2028         } while (scsi_sense_valid(sshdr) &&
2029                  sshdr->sense_key == UNIT_ATTENTION && --retries);
2030
2031         if (!sshdr_external)
2032                 kfree(sshdr);
2033         return result;
2034 }
2035 EXPORT_SYMBOL(scsi_test_unit_ready);
2036
2037 /**
2038  *      scsi_device_set_state - Take the given device through the device state model.
2039  *      @sdev:  scsi device to change the state of.
2040  *      @state: state to change to.
2041  *
2042  *      Returns zero if unsuccessful or an error if the requested 
2043  *      transition is illegal.
2044  */
2045 int
2046 scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
2047 {
2048         enum scsi_device_state oldstate = sdev->sdev_state;
2049
2050         if (state == oldstate)
2051                 return 0;
2052
2053         switch (state) {
2054         case SDEV_CREATED:
2055                 switch (oldstate) {
2056                 case SDEV_CREATED_BLOCK:
2057                         break;
2058                 default:
2059                         goto illegal;
2060                 }
2061                 break;
2062                         
2063         case SDEV_RUNNING:
2064                 switch (oldstate) {
2065                 case SDEV_CREATED:
2066                 case SDEV_OFFLINE:
2067                 case SDEV_QUIESCE:
2068                 case SDEV_BLOCK:
2069                         break;
2070                 default:
2071                         goto illegal;
2072                 }
2073                 break;
2074
2075         case SDEV_QUIESCE:
2076                 switch (oldstate) {
2077                 case SDEV_RUNNING:
2078                 case SDEV_OFFLINE:
2079                         break;
2080                 default:
2081                         goto illegal;
2082                 }
2083                 break;
2084
2085         case SDEV_OFFLINE:
2086                 switch (oldstate) {
2087                 case SDEV_CREATED:
2088                 case SDEV_RUNNING:
2089                 case SDEV_QUIESCE:
2090                 case SDEV_BLOCK:
2091                         break;
2092                 default:
2093                         goto illegal;
2094                 }
2095                 break;
2096
2097         case SDEV_BLOCK:
2098                 switch (oldstate) {
2099                 case SDEV_RUNNING:
2100                 case SDEV_CREATED_BLOCK:
2101                         break;
2102                 default:
2103                         goto illegal;
2104                 }
2105                 break;
2106
2107         case SDEV_CREATED_BLOCK:
2108                 switch (oldstate) {
2109                 case SDEV_CREATED:
2110                         break;
2111                 default:
2112                         goto illegal;
2113                 }
2114                 break;
2115
2116         case SDEV_CANCEL:
2117                 switch (oldstate) {
2118                 case SDEV_CREATED:
2119                 case SDEV_RUNNING:
2120                 case SDEV_QUIESCE:
2121                 case SDEV_OFFLINE:
2122                 case SDEV_BLOCK:
2123                         break;
2124                 default:
2125                         goto illegal;
2126                 }
2127                 break;
2128
2129         case SDEV_DEL:
2130                 switch (oldstate) {
2131                 case SDEV_CREATED:
2132                 case SDEV_RUNNING:
2133                 case SDEV_OFFLINE:
2134                 case SDEV_CANCEL:
2135                         break;
2136                 default:
2137                         goto illegal;
2138                 }
2139                 break;
2140
2141         }
2142         sdev->sdev_state = state;
2143         return 0;
2144
2145  illegal:
2146         SCSI_LOG_ERROR_RECOVERY(1, 
2147                                 sdev_printk(KERN_ERR, sdev,
2148                                             "Illegal state transition %s->%s\n",
2149                                             scsi_device_state_name(oldstate),
2150                                             scsi_device_state_name(state))
2151                                 );
2152         return -EINVAL;
2153 }
2154 EXPORT_SYMBOL(scsi_device_set_state);
2155
2156 /**
2157  *      sdev_evt_emit - emit a single SCSI device uevent
2158  *      @sdev: associated SCSI device
2159  *      @evt: event to emit
2160  *
2161  *      Send a single uevent (scsi_event) to the associated scsi_device.
2162  */
2163 static void scsi_evt_emit(struct scsi_device *sdev, struct scsi_event *evt)
2164 {
2165         int idx = 0;
2166         char *envp[3];
2167
2168         switch (evt->evt_type) {
2169         case SDEV_EVT_MEDIA_CHANGE:
2170                 envp[idx++] = "SDEV_MEDIA_CHANGE=1";
2171                 break;
2172
2173         default:
2174                 /* do nothing */
2175                 break;
2176         }
2177
2178         envp[idx++] = NULL;
2179
2180         kobject_uevent_env(&sdev->sdev_gendev.kobj, KOBJ_CHANGE, envp);
2181 }
2182
2183 /**
2184  *      sdev_evt_thread - send a uevent for each scsi event
2185  *      @work: work struct for scsi_device
2186  *
2187  *      Dispatch queued events to their associated scsi_device kobjects
2188  *      as uevents.
2189  */
2190 void scsi_evt_thread(struct work_struct *work)
2191 {
2192         struct scsi_device *sdev;
2193         LIST_HEAD(event_list);
2194
2195         sdev = container_of(work, struct scsi_device, event_work);
2196
2197         while (1) {
2198                 struct scsi_event *evt;
2199                 struct list_head *this, *tmp;
2200                 unsigned long flags;
2201
2202                 spin_lock_irqsave(&sdev->list_lock, flags);
2203                 list_splice_init(&sdev->event_list, &event_list);
2204                 spin_unlock_irqrestore(&sdev->list_lock, flags);
2205
2206                 if (list_empty(&event_list))
2207                         break;
2208
2209                 list_for_each_safe(this, tmp, &event_list) {
2210                         evt = list_entry(this, struct scsi_event, node);
2211                         list_del(&evt->node);
2212                         scsi_evt_emit(sdev, evt);
2213                         kfree(evt);
2214                 }
2215         }
2216 }
2217
2218 /**
2219  *      sdev_evt_send - send asserted event to uevent thread
2220  *      @sdev: scsi_device event occurred on
2221  *      @evt: event to send
2222  *
2223  *      Assert scsi device event asynchronously.
2224  */
2225 void sdev_evt_send(struct scsi_device *sdev, struct scsi_event *evt)
2226 {
2227         unsigned long flags;
2228
2229 #if 0
2230         /* FIXME: currently this check eliminates all media change events
2231          * for polled devices.  Need to update to discriminate between AN
2232          * and polled events */
2233         if (!test_bit(evt->evt_type, sdev->supported_events)) {
2234                 kfree(evt);
2235                 return;
2236         }
2237 #endif
2238
2239         spin_lock_irqsave(&sdev->list_lock, flags);
2240         list_add_tail(&evt->node, &sdev->event_list);
2241         schedule_work(&sdev->event_work);
2242         spin_unlock_irqrestore(&sdev->list_lock, flags);
2243 }
2244 EXPORT_SYMBOL_GPL(sdev_evt_send);
2245
2246 /**
2247  *      sdev_evt_alloc - allocate a new scsi event
2248  *      @evt_type: type of event to allocate
2249  *      @gfpflags: GFP flags for allocation
2250  *
2251  *      Allocates and returns a new scsi_event.
2252  */
2253 struct scsi_event *sdev_evt_alloc(enum scsi_device_event evt_type,
2254                                   gfp_t gfpflags)
2255 {
2256         struct scsi_event *evt = kzalloc(sizeof(struct scsi_event), gfpflags);
2257         if (!evt)
2258                 return NULL;
2259
2260         evt->evt_type = evt_type;
2261         INIT_LIST_HEAD(&evt->node);
2262
2263         /* evt_type-specific initialization, if any */
2264         switch (evt_type) {
2265         case SDEV_EVT_MEDIA_CHANGE:
2266         default:
2267                 /* do nothing */
2268                 break;
2269         }
2270
2271         return evt;
2272 }
2273 EXPORT_SYMBOL_GPL(sdev_evt_alloc);
2274
2275 /**
2276  *      sdev_evt_send_simple - send asserted event to uevent thread
2277  *      @sdev: scsi_device event occurred on
2278  *      @evt_type: type of event to send
2279  *      @gfpflags: GFP flags for allocation
2280  *
2281  *      Assert scsi device event asynchronously, given an event type.
2282  */
2283 void sdev_evt_send_simple(struct scsi_device *sdev,
2284                           enum scsi_device_event evt_type, gfp_t gfpflags)
2285 {
2286         struct scsi_event *evt = sdev_evt_alloc(evt_type, gfpflags);
2287         if (!evt) {
2288                 sdev_printk(KERN_ERR, sdev, "event %d eaten due to OOM\n",
2289                             evt_type);
2290                 return;
2291         }
2292
2293         sdev_evt_send(sdev, evt);
2294 }
2295 EXPORT_SYMBOL_GPL(sdev_evt_send_simple);
2296
2297 /**
2298  *      scsi_device_quiesce - Block user issued commands.
2299  *      @sdev:  scsi device to quiesce.
2300  *
2301  *      This works by trying to transition to the SDEV_QUIESCE state
2302  *      (which must be a legal transition).  When the device is in this
2303  *      state, only special requests will be accepted, all others will
2304  *      be deferred.  Since special requests may also be requeued requests,
2305  *      a successful return doesn't guarantee the device will be 
2306  *      totally quiescent.
2307  *
2308  *      Must be called with user context, may sleep.
2309  *
2310  *      Returns zero if unsuccessful or an error if not.
2311  */
2312 int
2313 scsi_device_quiesce(struct scsi_device *sdev)
2314 {
2315         int err = scsi_device_set_state(sdev, SDEV_QUIESCE);
2316         if (err)
2317                 return err;
2318
2319         scsi_run_queue(sdev->request_queue);
2320         while (sdev->device_busy) {
2321                 msleep_interruptible(200);
2322                 scsi_run_queue(sdev->request_queue);
2323         }
2324         return 0;
2325 }
2326 EXPORT_SYMBOL(scsi_device_quiesce);
2327
2328 /**
2329  *      scsi_device_resume - Restart user issued commands to a quiesced device.
2330  *      @sdev:  scsi device to resume.
2331  *
2332  *      Moves the device from quiesced back to running and restarts the
2333  *      queues.
2334  *
2335  *      Must be called with user context, may sleep.
2336  */
2337 void
2338 scsi_device_resume(struct scsi_device *sdev)
2339 {
2340         if(scsi_device_set_state(sdev, SDEV_RUNNING))
2341                 return;
2342         scsi_run_queue(sdev->request_queue);
2343 }
2344 EXPORT_SYMBOL(scsi_device_resume);
2345
2346 static void
2347 device_quiesce_fn(struct scsi_device *sdev, void *data)
2348 {
2349         scsi_device_quiesce(sdev);
2350 }
2351
2352 void
2353 scsi_target_quiesce(struct scsi_target *starget)
2354 {
2355         starget_for_each_device(starget, NULL, device_quiesce_fn);
2356 }
2357 EXPORT_SYMBOL(scsi_target_quiesce);
2358
2359 static void
2360 device_resume_fn(struct scsi_device *sdev, void *data)
2361 {
2362         scsi_device_resume(sdev);
2363 }
2364
2365 void
2366 scsi_target_resume(struct scsi_target *starget)
2367 {
2368         starget_for_each_device(starget, NULL, device_resume_fn);
2369 }
2370 EXPORT_SYMBOL(scsi_target_resume);
2371
2372 /**
2373  * scsi_internal_device_block - internal function to put a device temporarily into the SDEV_BLOCK state
2374  * @sdev:       device to block
2375  *
2376  * Block request made by scsi lld's to temporarily stop all
2377  * scsi commands on the specified device.  Called from interrupt
2378  * or normal process context.
2379  *
2380  * Returns zero if successful or error if not
2381  *
2382  * Notes:       
2383  *      This routine transitions the device to the SDEV_BLOCK state
2384  *      (which must be a legal transition).  When the device is in this
2385  *      state, all commands are deferred until the scsi lld reenables
2386  *      the device with scsi_device_unblock or device_block_tmo fires.
2387  *      This routine assumes the host_lock is held on entry.
2388  */
2389 int
2390 scsi_internal_device_block(struct scsi_device *sdev)
2391 {
2392         struct request_queue *q = sdev->request_queue;
2393         unsigned long flags;
2394         int err = 0;
2395
2396         err = scsi_device_set_state(sdev, SDEV_BLOCK);
2397         if (err) {
2398                 err = scsi_device_set_state(sdev, SDEV_CREATED_BLOCK);
2399
2400                 if (err)
2401                         return err;
2402         }
2403
2404         /* 
2405          * The device has transitioned to SDEV_BLOCK.  Stop the
2406          * block layer from calling the midlayer with this device's
2407          * request queue. 
2408          */
2409         spin_lock_irqsave(q->queue_lock, flags);
2410         blk_stop_queue(q);
2411         spin_unlock_irqrestore(q->queue_lock, flags);
2412
2413         return 0;
2414 }
2415 EXPORT_SYMBOL_GPL(scsi_internal_device_block);
2416  
2417 /**
2418  * scsi_internal_device_unblock - resume a device after a block request
2419  * @sdev:       device to resume
2420  *
2421  * Called by scsi lld's or the midlayer to restart the device queue
2422  * for the previously suspended scsi device.  Called from interrupt or
2423  * normal process context.
2424  *
2425  * Returns zero if successful or error if not.
2426  *
2427  * Notes:       
2428  *      This routine transitions the device to the SDEV_RUNNING state
2429  *      (which must be a legal transition) allowing the midlayer to
2430  *      goose the queue for this device.  This routine assumes the 
2431  *      host_lock is held upon entry.
2432  */
2433 int
2434 scsi_internal_device_unblock(struct scsi_device *sdev)
2435 {
2436         struct request_queue *q = sdev->request_queue; 
2437         unsigned long flags;
2438         
2439         /* 
2440          * Try to transition the scsi device to SDEV_RUNNING
2441          * and goose the device queue if successful.  
2442          */
2443         if (sdev->sdev_state == SDEV_BLOCK)
2444                 sdev->sdev_state = SDEV_RUNNING;
2445         else if (sdev->sdev_state == SDEV_CREATED_BLOCK)
2446                 sdev->sdev_state = SDEV_CREATED;
2447         else if (sdev->sdev_state != SDEV_CANCEL &&
2448                  sdev->sdev_state != SDEV_OFFLINE)
2449                 return -EINVAL;
2450
2451         spin_lock_irqsave(q->queue_lock, flags);
2452         blk_start_queue(q);
2453         spin_unlock_irqrestore(q->queue_lock, flags);
2454
2455         return 0;
2456 }
2457 EXPORT_SYMBOL_GPL(scsi_internal_device_unblock);
2458
2459 static void
2460 device_block(struct scsi_device *sdev, void *data)
2461 {
2462         scsi_internal_device_block(sdev);
2463 }
2464
2465 static int
2466 target_block(struct device *dev, void *data)
2467 {
2468         if (scsi_is_target_device(dev))
2469                 starget_for_each_device(to_scsi_target(dev), NULL,
2470                                         device_block);
2471         return 0;
2472 }
2473
2474 void
2475 scsi_target_block(struct device *dev)
2476 {
2477         if (scsi_is_target_device(dev))
2478                 starget_for_each_device(to_scsi_target(dev), NULL,
2479                                         device_block);
2480         else
2481                 device_for_each_child(dev, NULL, target_block);
2482 }
2483 EXPORT_SYMBOL_GPL(scsi_target_block);
2484
2485 static void
2486 device_unblock(struct scsi_device *sdev, void *data)
2487 {
2488         scsi_internal_device_unblock(sdev);
2489 }
2490
2491 static int
2492 target_unblock(struct device *dev, void *data)
2493 {
2494         if (scsi_is_target_device(dev))
2495                 starget_for_each_device(to_scsi_target(dev), NULL,
2496                                         device_unblock);
2497         return 0;
2498 }
2499
2500 void
2501 scsi_target_unblock(struct device *dev)
2502 {
2503         if (scsi_is_target_device(dev))
2504                 starget_for_each_device(to_scsi_target(dev), NULL,
2505                                         device_unblock);
2506         else
2507                 device_for_each_child(dev, NULL, target_unblock);
2508 }
2509 EXPORT_SYMBOL_GPL(scsi_target_unblock);
2510
2511 /**
2512  * scsi_kmap_atomic_sg - find and atomically map an sg-elemnt
2513  * @sgl:        scatter-gather list
2514  * @sg_count:   number of segments in sg
2515  * @offset:     offset in bytes into sg, on return offset into the mapped area
2516  * @len:        bytes to map, on return number of bytes mapped
2517  *
2518  * Returns virtual address of the start of the mapped page
2519  */
2520 void *scsi_kmap_atomic_sg(struct scatterlist *sgl, int sg_count,
2521                           size_t *offset, size_t *len)
2522 {
2523         int i;
2524         size_t sg_len = 0, len_complete = 0;
2525         struct scatterlist *sg;
2526         struct page *page;
2527
2528         WARN_ON(!irqs_disabled());
2529
2530         for_each_sg(sgl, sg, sg_count, i) {
2531                 len_complete = sg_len; /* Complete sg-entries */
2532                 sg_len += sg->length;
2533                 if (sg_len > *offset)
2534                         break;
2535         }
2536
2537         if (unlikely(i == sg_count)) {
2538                 printk(KERN_ERR "%s: Bytes in sg: %zu, requested offset %zu, "
2539                         "elements %d\n",
2540                        __func__, sg_len, *offset, sg_count);
2541                 WARN_ON(1);
2542                 return NULL;
2543         }
2544
2545         /* Offset starting from the beginning of first page in this sg-entry */
2546         *offset = *offset - len_complete + sg->offset;
2547
2548         /* Assumption: contiguous pages can be accessed as "page + i" */
2549         page = nth_page(sg_page(sg), (*offset >> PAGE_SHIFT));
2550         *offset &= ~PAGE_MASK;
2551
2552         /* Bytes in this sg-entry from *offset to the end of the page */
2553         sg_len = PAGE_SIZE - *offset;
2554         if (*len > sg_len)
2555                 *len = sg_len;
2556
2557         return kmap_atomic(page, KM_BIO_SRC_IRQ);
2558 }
2559 EXPORT_SYMBOL(scsi_kmap_atomic_sg);
2560
2561 /**
2562  * scsi_kunmap_atomic_sg - atomically unmap a virtual address, previously mapped with scsi_kmap_atomic_sg
2563  * @virt:       virtual address to be unmapped
2564  */
2565 void scsi_kunmap_atomic_sg(void *virt)
2566 {
2567         kunmap_atomic(virt, KM_BIO_SRC_IRQ);
2568 }
2569 EXPORT_SYMBOL(scsi_kunmap_atomic_sg);