Merge branch 'percpu-cpumask-x86-for-linus-2' of git://git.kernel.org/pub/scm/linux...
[linux-2.6] / drivers / scsi / scsi_lib.c
1 /*
2  *  scsi_lib.c Copyright (C) 1999 Eric Youngdale
3  *
4  *  SCSI queueing library.
5  *      Initial versions: Eric Youngdale (eric@andante.org).
6  *                        Based upon conversations with large numbers
7  *                        of people at Linux Expo.
8  */
9
10 #include <linux/bio.h>
11 #include <linux/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         req->cmd_flags &= ~REQ_DONTPREP;
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->data_len > 0 && req->data_len <= bufflen))
244                 memset(buffer + (bufflen - req->data_len), 0, req->data_len);
245
246         if (resid)
247                 *resid = req->data_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                 int leftover = (req->hard_nr_sectors << 9);
550
551                 if (blk_pc_request(req))
552                         leftover = req->data_len;
553
554                 /* kill remainder if no retrys */
555                 if (error && scsi_noretry_cmd(cmd))
556                         blk_end_request(req, error, leftover);
557                 else {
558                         if (requeue) {
559                                 /*
560                                  * Bleah.  Leftovers again.  Stick the
561                                  * leftovers in the front of the
562                                  * queue, and goose the queue again.
563                                  */
564                                 scsi_release_buffers(cmd);
565                                 scsi_requeue_command(q, cmd);
566                                 cmd = NULL;
567                         }
568                         return cmd;
569                 }
570         }
571
572         /*
573          * This will goose the queue request function at the end, so we don't
574          * need to worry about launching another command.
575          */
576         __scsi_release_buffers(cmd, 0);
577         scsi_next_command(cmd);
578         return NULL;
579 }
580
581 static inline unsigned int scsi_sgtable_index(unsigned short nents)
582 {
583         unsigned int index;
584
585         BUG_ON(nents > SCSI_MAX_SG_SEGMENTS);
586
587         if (nents <= 8)
588                 index = 0;
589         else
590                 index = get_count_order(nents) - 3;
591
592         return index;
593 }
594
595 static void scsi_sg_free(struct scatterlist *sgl, unsigned int nents)
596 {
597         struct scsi_host_sg_pool *sgp;
598
599         sgp = scsi_sg_pools + scsi_sgtable_index(nents);
600         mempool_free(sgl, sgp->pool);
601 }
602
603 static struct scatterlist *scsi_sg_alloc(unsigned int nents, gfp_t gfp_mask)
604 {
605         struct scsi_host_sg_pool *sgp;
606
607         sgp = scsi_sg_pools + scsi_sgtable_index(nents);
608         return mempool_alloc(sgp->pool, gfp_mask);
609 }
610
611 static int scsi_alloc_sgtable(struct scsi_data_buffer *sdb, int nents,
612                               gfp_t gfp_mask)
613 {
614         int ret;
615
616         BUG_ON(!nents);
617
618         ret = __sg_alloc_table(&sdb->table, nents, SCSI_MAX_SG_SEGMENTS,
619                                gfp_mask, scsi_sg_alloc);
620         if (unlikely(ret))
621                 __sg_free_table(&sdb->table, SCSI_MAX_SG_SEGMENTS,
622                                 scsi_sg_free);
623
624         return ret;
625 }
626
627 static void scsi_free_sgtable(struct scsi_data_buffer *sdb)
628 {
629         __sg_free_table(&sdb->table, SCSI_MAX_SG_SEGMENTS, scsi_sg_free);
630 }
631
632 static void __scsi_release_buffers(struct scsi_cmnd *cmd, int do_bidi_check)
633 {
634
635         if (cmd->sdb.table.nents)
636                 scsi_free_sgtable(&cmd->sdb);
637
638         memset(&cmd->sdb, 0, sizeof(cmd->sdb));
639
640         if (do_bidi_check && scsi_bidi_cmnd(cmd)) {
641                 struct scsi_data_buffer *bidi_sdb =
642                         cmd->request->next_rq->special;
643                 scsi_free_sgtable(bidi_sdb);
644                 kmem_cache_free(scsi_sdb_cache, bidi_sdb);
645                 cmd->request->next_rq->special = NULL;
646         }
647
648         if (scsi_prot_sg_count(cmd))
649                 scsi_free_sgtable(cmd->prot_sdb);
650 }
651
652 /*
653  * Function:    scsi_release_buffers()
654  *
655  * Purpose:     Completion processing for block device I/O requests.
656  *
657  * Arguments:   cmd     - command that we are bailing.
658  *
659  * Lock status: Assumed that no lock is held upon entry.
660  *
661  * Returns:     Nothing
662  *
663  * Notes:       In the event that an upper level driver rejects a
664  *              command, we must release resources allocated during
665  *              the __init_io() function.  Primarily this would involve
666  *              the scatter-gather table, and potentially any bounce
667  *              buffers.
668  */
669 void scsi_release_buffers(struct scsi_cmnd *cmd)
670 {
671         __scsi_release_buffers(cmd, 1);
672 }
673 EXPORT_SYMBOL(scsi_release_buffers);
674
675 /*
676  * Bidi commands Must be complete as a whole, both sides at once.
677  * If part of the bytes were written and lld returned
678  * scsi_in()->resid and/or scsi_out()->resid this information will be left
679  * in req->data_len and req->next_rq->data_len. The upper-layer driver can
680  * decide what to do with this information.
681  */
682 static void scsi_end_bidi_request(struct scsi_cmnd *cmd)
683 {
684         struct request *req = cmd->request;
685         unsigned int dlen = req->data_len;
686         unsigned int next_dlen = req->next_rq->data_len;
687
688         req->data_len = scsi_out(cmd)->resid;
689         req->next_rq->data_len = scsi_in(cmd)->resid;
690
691         /* The req and req->next_rq have not been completed */
692         BUG_ON(blk_end_bidi_request(req, 0, dlen, next_dlen));
693
694         scsi_release_buffers(cmd);
695
696         /*
697          * This will goose the queue request function at the end, so we don't
698          * need to worry about launching another command.
699          */
700         scsi_next_command(cmd);
701 }
702
703 /*
704  * Function:    scsi_io_completion()
705  *
706  * Purpose:     Completion processing for block device I/O requests.
707  *
708  * Arguments:   cmd   - command that is finished.
709  *
710  * Lock status: Assumed that no lock is held upon entry.
711  *
712  * Returns:     Nothing
713  *
714  * Notes:       This function is matched in terms of capabilities to
715  *              the function that created the scatter-gather list.
716  *              In other words, if there are no bounce buffers
717  *              (the normal case for most drivers), we don't need
718  *              the logic to deal with cleaning up afterwards.
719  *
720  *              We must call scsi_end_request().  This will finish off
721  *              the specified number of sectors.  If we are done, the
722  *              command block will be released and the queue function
723  *              will be goosed.  If we are not done then we have to
724  *              figure out what to do next:
725  *
726  *              a) We can call scsi_requeue_command().  The request
727  *                 will be unprepared and put back on the queue.  Then
728  *                 a new command will be created for it.  This should
729  *                 be used if we made forward progress, or if we want
730  *                 to switch from READ(10) to READ(6) for example.
731  *
732  *              b) We can call scsi_queue_insert().  The request will
733  *                 be put back on the queue and retried using the same
734  *                 command as before, possibly after a delay.
735  *
736  *              c) We can call blk_end_request() with -EIO to fail
737  *                 the remainder of the request.
738  */
739 void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes)
740 {
741         int result = cmd->result;
742         int this_count;
743         struct request_queue *q = cmd->device->request_queue;
744         struct request *req = cmd->request;
745         int error = 0;
746         struct scsi_sense_hdr sshdr;
747         int sense_valid = 0;
748         int sense_deferred = 0;
749         enum {ACTION_FAIL, ACTION_REPREP, ACTION_RETRY,
750               ACTION_DELAYED_RETRY} action;
751         char *description = NULL;
752
753         if (result) {
754                 sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
755                 if (sense_valid)
756                         sense_deferred = scsi_sense_is_deferred(&sshdr);
757         }
758
759         if (blk_pc_request(req)) { /* SG_IO ioctl from block level */
760                 req->errors = result;
761                 if (result) {
762                         if (sense_valid && req->sense) {
763                                 /*
764                                  * SG_IO wants current and deferred errors
765                                  */
766                                 int len = 8 + cmd->sense_buffer[7];
767
768                                 if (len > SCSI_SENSE_BUFFERSIZE)
769                                         len = SCSI_SENSE_BUFFERSIZE;
770                                 memcpy(req->sense, cmd->sense_buffer,  len);
771                                 req->sense_len = len;
772                         }
773                         if (!sense_deferred)
774                                 error = -EIO;
775                 }
776                 if (scsi_bidi_cmnd(cmd)) {
777                         /* will also release_buffers */
778                         scsi_end_bidi_request(cmd);
779                         return;
780                 }
781                 req->data_len = scsi_get_resid(cmd);
782         }
783
784         BUG_ON(blk_bidi_rq(req)); /* bidi not support for !blk_pc_request yet */
785
786         /*
787          * Next deal with any sectors which we were able to correctly
788          * handle.
789          */
790         SCSI_LOG_HLCOMPLETE(1, printk("%ld sectors total, "
791                                       "%d bytes done.\n",
792                                       req->nr_sectors, good_bytes));
793
794         /* A number of bytes were successfully read.  If there
795          * are leftovers and there is some kind of error
796          * (result != 0), retry the rest.
797          */
798         if (scsi_end_request(cmd, error, good_bytes, result == 0) == NULL)
799                 return;
800         this_count = blk_rq_bytes(req);
801
802         error = -EIO;
803
804         if (host_byte(result) == DID_RESET) {
805                 /* Third party bus reset or reset for error recovery
806                  * reasons.  Just retry the command and see what
807                  * happens.
808                  */
809                 action = ACTION_RETRY;
810         } else if (sense_valid && !sense_deferred) {
811                 switch (sshdr.sense_key) {
812                 case UNIT_ATTENTION:
813                         if (cmd->device->removable) {
814                                 /* Detected disc change.  Set a bit
815                                  * and quietly refuse further access.
816                                  */
817                                 cmd->device->changed = 1;
818                                 description = "Media Changed";
819                                 action = ACTION_FAIL;
820                         } else {
821                                 /* Must have been a power glitch, or a
822                                  * bus reset.  Could not have been a
823                                  * media change, so we just retry the
824                                  * command and see what happens.
825                                  */
826                                 action = ACTION_RETRY;
827                         }
828                         break;
829                 case ILLEGAL_REQUEST:
830                         /* If we had an ILLEGAL REQUEST returned, then
831                          * we may have performed an unsupported
832                          * command.  The only thing this should be
833                          * would be a ten byte read where only a six
834                          * byte read was supported.  Also, on a system
835                          * where READ CAPACITY failed, we may have
836                          * read past the end of the disk.
837                          */
838                         if ((cmd->device->use_10_for_rw &&
839                             sshdr.asc == 0x20 && sshdr.ascq == 0x00) &&
840                             (cmd->cmnd[0] == READ_10 ||
841                              cmd->cmnd[0] == WRITE_10)) {
842                                 /* This will issue a new 6-byte command. */
843                                 cmd->device->use_10_for_rw = 0;
844                                 action = ACTION_REPREP;
845                         } else if (sshdr.asc == 0x10) /* DIX */ {
846                                 description = "Host Data Integrity Failure";
847                                 action = ACTION_FAIL;
848                                 error = -EILSEQ;
849                         } else
850                                 action = ACTION_FAIL;
851                         break;
852                 case ABORTED_COMMAND:
853                         action = ACTION_FAIL;
854                         if (sshdr.asc == 0x10) { /* DIF */
855                                 description = "Target Data Integrity Failure";
856                                 error = -EILSEQ;
857                         }
858                         break;
859                 case NOT_READY:
860                         /* If the device is in the process of becoming
861                          * ready, or has a temporary blockage, retry.
862                          */
863                         if (sshdr.asc == 0x04) {
864                                 switch (sshdr.ascq) {
865                                 case 0x01: /* becoming ready */
866                                 case 0x04: /* format in progress */
867                                 case 0x05: /* rebuild in progress */
868                                 case 0x06: /* recalculation in progress */
869                                 case 0x07: /* operation in progress */
870                                 case 0x08: /* Long write in progress */
871                                 case 0x09: /* self test in progress */
872                                         action = ACTION_DELAYED_RETRY;
873                                         break;
874                                 default:
875                                         description = "Device not ready";
876                                         action = ACTION_FAIL;
877                                         break;
878                                 }
879                         } else {
880                                 description = "Device not ready";
881                                 action = ACTION_FAIL;
882                         }
883                         break;
884                 case VOLUME_OVERFLOW:
885                         /* See SSC3rXX or current. */
886                         action = ACTION_FAIL;
887                         break;
888                 default:
889                         description = "Unhandled sense code";
890                         action = ACTION_FAIL;
891                         break;
892                 }
893         } else {
894                 description = "Unhandled error code";
895                 action = ACTION_FAIL;
896         }
897
898         switch (action) {
899         case ACTION_FAIL:
900                 /* Give up and fail the remainder of the request */
901                 scsi_release_buffers(cmd);
902                 if (!(req->cmd_flags & REQ_QUIET)) {
903                         if (description)
904                                 scmd_printk(KERN_INFO, cmd, "%s\n",
905                                             description);
906                         scsi_print_result(cmd);
907                         if (driver_byte(result) & DRIVER_SENSE)
908                                 scsi_print_sense("", cmd);
909                 }
910                 blk_end_request(req, -EIO, blk_rq_bytes(req));
911                 scsi_next_command(cmd);
912                 break;
913         case ACTION_REPREP:
914                 /* Unprep the request and put it back at the head of the queue.
915                  * A new command will be prepared and issued.
916                  */
917                 scsi_release_buffers(cmd);
918                 scsi_requeue_command(q, cmd);
919                 break;
920         case ACTION_RETRY:
921                 /* Retry the same command immediately */
922                 __scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY, 0);
923                 break;
924         case ACTION_DELAYED_RETRY:
925                 /* Retry the same command after a delay */
926                 __scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY, 0);
927                 break;
928         }
929 }
930
931 static int scsi_init_sgtable(struct request *req, struct scsi_data_buffer *sdb,
932                              gfp_t gfp_mask)
933 {
934         int count;
935
936         /*
937          * If sg table allocation fails, requeue request later.
938          */
939         if (unlikely(scsi_alloc_sgtable(sdb, req->nr_phys_segments,
940                                         gfp_mask))) {
941                 return BLKPREP_DEFER;
942         }
943
944         req->buffer = NULL;
945
946         /* 
947          * Next, walk the list, and fill in the addresses and sizes of
948          * each segment.
949          */
950         count = blk_rq_map_sg(req->q, req, sdb->table.sgl);
951         BUG_ON(count > sdb->table.nents);
952         sdb->table.nents = count;
953         if (blk_pc_request(req))
954                 sdb->length = req->data_len;
955         else
956                 sdb->length = req->nr_sectors << 9;
957         return BLKPREP_OK;
958 }
959
960 /*
961  * Function:    scsi_init_io()
962  *
963  * Purpose:     SCSI I/O initialize function.
964  *
965  * Arguments:   cmd   - Command descriptor we wish to initialize
966  *
967  * Returns:     0 on success
968  *              BLKPREP_DEFER if the failure is retryable
969  *              BLKPREP_KILL if the failure is fatal
970  */
971 int scsi_init_io(struct scsi_cmnd *cmd, gfp_t gfp_mask)
972 {
973         int error = scsi_init_sgtable(cmd->request, &cmd->sdb, gfp_mask);
974         if (error)
975                 goto err_exit;
976
977         if (blk_bidi_rq(cmd->request)) {
978                 struct scsi_data_buffer *bidi_sdb = kmem_cache_zalloc(
979                         scsi_sdb_cache, GFP_ATOMIC);
980                 if (!bidi_sdb) {
981                         error = BLKPREP_DEFER;
982                         goto err_exit;
983                 }
984
985                 cmd->request->next_rq->special = bidi_sdb;
986                 error = scsi_init_sgtable(cmd->request->next_rq, bidi_sdb,
987                                                                     GFP_ATOMIC);
988                 if (error)
989                         goto err_exit;
990         }
991
992         if (blk_integrity_rq(cmd->request)) {
993                 struct scsi_data_buffer *prot_sdb = cmd->prot_sdb;
994                 int ivecs, count;
995
996                 BUG_ON(prot_sdb == NULL);
997                 ivecs = blk_rq_count_integrity_sg(cmd->request);
998
999                 if (scsi_alloc_sgtable(prot_sdb, ivecs, gfp_mask)) {
1000                         error = BLKPREP_DEFER;
1001                         goto err_exit;
1002                 }
1003
1004                 count = blk_rq_map_integrity_sg(cmd->request,
1005                                                 prot_sdb->table.sgl);
1006                 BUG_ON(unlikely(count > ivecs));
1007
1008                 cmd->prot_sdb = prot_sdb;
1009                 cmd->prot_sdb->table.nents = count;
1010         }
1011
1012         return BLKPREP_OK ;
1013
1014 err_exit:
1015         scsi_release_buffers(cmd);
1016         if (error == BLKPREP_KILL)
1017                 scsi_put_command(cmd);
1018         else /* BLKPREP_DEFER */
1019                 scsi_unprep_request(cmd->request);
1020
1021         return error;
1022 }
1023 EXPORT_SYMBOL(scsi_init_io);
1024
1025 static struct scsi_cmnd *scsi_get_cmd_from_req(struct scsi_device *sdev,
1026                 struct request *req)
1027 {
1028         struct scsi_cmnd *cmd;
1029
1030         if (!req->special) {
1031                 cmd = scsi_get_command(sdev, GFP_ATOMIC);
1032                 if (unlikely(!cmd))
1033                         return NULL;
1034                 req->special = cmd;
1035         } else {
1036                 cmd = req->special;
1037         }
1038
1039         /* pull a tag out of the request if we have one */
1040         cmd->tag = req->tag;
1041         cmd->request = req;
1042
1043         cmd->cmnd = req->cmd;
1044
1045         return cmd;
1046 }
1047
1048 int scsi_setup_blk_pc_cmnd(struct scsi_device *sdev, struct request *req)
1049 {
1050         struct scsi_cmnd *cmd;
1051         int ret = scsi_prep_state_check(sdev, req);
1052
1053         if (ret != BLKPREP_OK)
1054                 return ret;
1055
1056         cmd = scsi_get_cmd_from_req(sdev, req);
1057         if (unlikely(!cmd))
1058                 return BLKPREP_DEFER;
1059
1060         /*
1061          * BLOCK_PC requests may transfer data, in which case they must
1062          * a bio attached to them.  Or they might contain a SCSI command
1063          * that does not transfer data, in which case they may optionally
1064          * submit a request without an attached bio.
1065          */
1066         if (req->bio) {
1067                 int ret;
1068
1069                 BUG_ON(!req->nr_phys_segments);
1070
1071                 ret = scsi_init_io(cmd, GFP_ATOMIC);
1072                 if (unlikely(ret))
1073                         return ret;
1074         } else {
1075                 BUG_ON(req->data_len);
1076                 BUG_ON(req->data);
1077
1078                 memset(&cmd->sdb, 0, sizeof(cmd->sdb));
1079                 req->buffer = NULL;
1080         }
1081
1082         cmd->cmd_len = req->cmd_len;
1083         if (!req->data_len)
1084                 cmd->sc_data_direction = DMA_NONE;
1085         else if (rq_data_dir(req) == WRITE)
1086                 cmd->sc_data_direction = DMA_TO_DEVICE;
1087         else
1088                 cmd->sc_data_direction = DMA_FROM_DEVICE;
1089         
1090         cmd->transfersize = req->data_len;
1091         cmd->allowed = req->retries;
1092         return BLKPREP_OK;
1093 }
1094 EXPORT_SYMBOL(scsi_setup_blk_pc_cmnd);
1095
1096 /*
1097  * Setup a REQ_TYPE_FS command.  These are simple read/write request
1098  * from filesystems that still need to be translated to SCSI CDBs from
1099  * the ULD.
1100  */
1101 int scsi_setup_fs_cmnd(struct scsi_device *sdev, struct request *req)
1102 {
1103         struct scsi_cmnd *cmd;
1104         int ret = scsi_prep_state_check(sdev, req);
1105
1106         if (ret != BLKPREP_OK)
1107                 return ret;
1108
1109         if (unlikely(sdev->scsi_dh_data && sdev->scsi_dh_data->scsi_dh
1110                          && sdev->scsi_dh_data->scsi_dh->prep_fn)) {
1111                 ret = sdev->scsi_dh_data->scsi_dh->prep_fn(sdev, req);
1112                 if (ret != BLKPREP_OK)
1113                         return ret;
1114         }
1115
1116         /*
1117          * Filesystem requests must transfer data.
1118          */
1119         BUG_ON(!req->nr_phys_segments);
1120
1121         cmd = scsi_get_cmd_from_req(sdev, req);
1122         if (unlikely(!cmd))
1123                 return BLKPREP_DEFER;
1124
1125         memset(cmd->cmnd, 0, BLK_MAX_CDB);
1126         return scsi_init_io(cmd, GFP_ATOMIC);
1127 }
1128 EXPORT_SYMBOL(scsi_setup_fs_cmnd);
1129
1130 int scsi_prep_state_check(struct scsi_device *sdev, struct request *req)
1131 {
1132         int ret = BLKPREP_OK;
1133
1134         /*
1135          * If the device is not in running state we will reject some
1136          * or all commands.
1137          */
1138         if (unlikely(sdev->sdev_state != SDEV_RUNNING)) {
1139                 switch (sdev->sdev_state) {
1140                 case SDEV_OFFLINE:
1141                         /*
1142                          * If the device is offline we refuse to process any
1143                          * commands.  The device must be brought online
1144                          * before trying any recovery commands.
1145                          */
1146                         sdev_printk(KERN_ERR, sdev,
1147                                     "rejecting I/O to offline device\n");
1148                         ret = BLKPREP_KILL;
1149                         break;
1150                 case SDEV_DEL:
1151                         /*
1152                          * If the device is fully deleted, we refuse to
1153                          * process any commands as well.
1154                          */
1155                         sdev_printk(KERN_ERR, sdev,
1156                                     "rejecting I/O to dead device\n");
1157                         ret = BLKPREP_KILL;
1158                         break;
1159                 case SDEV_QUIESCE:
1160                 case SDEV_BLOCK:
1161                 case SDEV_CREATED_BLOCK:
1162                         /*
1163                          * If the devices is blocked we defer normal commands.
1164                          */
1165                         if (!(req->cmd_flags & REQ_PREEMPT))
1166                                 ret = BLKPREP_DEFER;
1167                         break;
1168                 default:
1169                         /*
1170                          * For any other not fully online state we only allow
1171                          * special commands.  In particular any user initiated
1172                          * command is not allowed.
1173                          */
1174                         if (!(req->cmd_flags & REQ_PREEMPT))
1175                                 ret = BLKPREP_KILL;
1176                         break;
1177                 }
1178         }
1179         return ret;
1180 }
1181 EXPORT_SYMBOL(scsi_prep_state_check);
1182
1183 int scsi_prep_return(struct request_queue *q, struct request *req, int ret)
1184 {
1185         struct scsi_device *sdev = q->queuedata;
1186
1187         switch (ret) {
1188         case BLKPREP_KILL:
1189                 req->errors = DID_NO_CONNECT << 16;
1190                 /* release the command and kill it */
1191                 if (req->special) {
1192                         struct scsi_cmnd *cmd = req->special;
1193                         scsi_release_buffers(cmd);
1194                         scsi_put_command(cmd);
1195                         req->special = NULL;
1196                 }
1197                 break;
1198         case BLKPREP_DEFER:
1199                 /*
1200                  * If we defer, the elv_next_request() returns NULL, but the
1201                  * queue must be restarted, so we plug here if no returning
1202                  * command will automatically do that.
1203                  */
1204                 if (sdev->device_busy == 0)
1205                         blk_plug_device(q);
1206                 break;
1207         default:
1208                 req->cmd_flags |= REQ_DONTPREP;
1209         }
1210
1211         return ret;
1212 }
1213 EXPORT_SYMBOL(scsi_prep_return);
1214
1215 int scsi_prep_fn(struct request_queue *q, struct request *req)
1216 {
1217         struct scsi_device *sdev = q->queuedata;
1218         int ret = BLKPREP_KILL;
1219
1220         if (req->cmd_type == REQ_TYPE_BLOCK_PC)
1221                 ret = scsi_setup_blk_pc_cmnd(sdev, req);
1222         return scsi_prep_return(q, req, ret);
1223 }
1224
1225 /*
1226  * scsi_dev_queue_ready: if we can send requests to sdev, return 1 else
1227  * return 0.
1228  *
1229  * Called with the queue_lock held.
1230  */
1231 static inline int scsi_dev_queue_ready(struct request_queue *q,
1232                                   struct scsi_device *sdev)
1233 {
1234         if (sdev->device_busy == 0 && sdev->device_blocked) {
1235                 /*
1236                  * unblock after device_blocked iterates to zero
1237                  */
1238                 if (--sdev->device_blocked == 0) {
1239                         SCSI_LOG_MLQUEUE(3,
1240                                    sdev_printk(KERN_INFO, sdev,
1241                                    "unblocking device at zero depth\n"));
1242                 } else {
1243                         blk_plug_device(q);
1244                         return 0;
1245                 }
1246         }
1247         if (scsi_device_is_busy(sdev))
1248                 return 0;
1249
1250         return 1;
1251 }
1252
1253
1254 /*
1255  * scsi_target_queue_ready: checks if there we can send commands to target
1256  * @sdev: scsi device on starget to check.
1257  *
1258  * Called with the host lock held.
1259  */
1260 static inline int scsi_target_queue_ready(struct Scsi_Host *shost,
1261                                            struct scsi_device *sdev)
1262 {
1263         struct scsi_target *starget = scsi_target(sdev);
1264
1265         if (starget->single_lun) {
1266                 if (starget->starget_sdev_user &&
1267                     starget->starget_sdev_user != sdev)
1268                         return 0;
1269                 starget->starget_sdev_user = sdev;
1270         }
1271
1272         if (starget->target_busy == 0 && starget->target_blocked) {
1273                 /*
1274                  * unblock after target_blocked iterates to zero
1275                  */
1276                 if (--starget->target_blocked == 0) {
1277                         SCSI_LOG_MLQUEUE(3, starget_printk(KERN_INFO, starget,
1278                                          "unblocking target at zero depth\n"));
1279                 } else {
1280                         blk_plug_device(sdev->request_queue);
1281                         return 0;
1282                 }
1283         }
1284
1285         if (scsi_target_is_busy(starget)) {
1286                 if (list_empty(&sdev->starved_entry)) {
1287                         list_add_tail(&sdev->starved_entry,
1288                                       &shost->starved_list);
1289                         return 0;
1290                 }
1291         }
1292
1293         /* We're OK to process the command, so we can't be starved */
1294         if (!list_empty(&sdev->starved_entry))
1295                 list_del_init(&sdev->starved_entry);
1296         return 1;
1297 }
1298
1299 /*
1300  * scsi_host_queue_ready: if we can send requests to shost, return 1 else
1301  * return 0. We must end up running the queue again whenever 0 is
1302  * returned, else IO can hang.
1303  *
1304  * Called with host_lock held.
1305  */
1306 static inline int scsi_host_queue_ready(struct request_queue *q,
1307                                    struct Scsi_Host *shost,
1308                                    struct scsi_device *sdev)
1309 {
1310         if (scsi_host_in_recovery(shost))
1311                 return 0;
1312         if (shost->host_busy == 0 && shost->host_blocked) {
1313                 /*
1314                  * unblock after host_blocked iterates to zero
1315                  */
1316                 if (--shost->host_blocked == 0) {
1317                         SCSI_LOG_MLQUEUE(3,
1318                                 printk("scsi%d unblocking host at zero depth\n",
1319                                         shost->host_no));
1320                 } else {
1321                         return 0;
1322                 }
1323         }
1324         if (scsi_host_is_busy(shost)) {
1325                 if (list_empty(&sdev->starved_entry))
1326                         list_add_tail(&sdev->starved_entry, &shost->starved_list);
1327                 return 0;
1328         }
1329
1330         /* We're OK to process the command, so we can't be starved */
1331         if (!list_empty(&sdev->starved_entry))
1332                 list_del_init(&sdev->starved_entry);
1333
1334         return 1;
1335 }
1336
1337 /*
1338  * Busy state exporting function for request stacking drivers.
1339  *
1340  * For efficiency, no lock is taken to check the busy state of
1341  * shost/starget/sdev, since the returned value is not guaranteed and
1342  * may be changed after request stacking drivers call the function,
1343  * regardless of taking lock or not.
1344  *
1345  * When scsi can't dispatch I/Os anymore and needs to kill I/Os
1346  * (e.g. !sdev), scsi needs to return 'not busy'.
1347  * Otherwise, request stacking drivers may hold requests forever.
1348  */
1349 static int scsi_lld_busy(struct request_queue *q)
1350 {
1351         struct scsi_device *sdev = q->queuedata;
1352         struct Scsi_Host *shost;
1353         struct scsi_target *starget;
1354
1355         if (!sdev)
1356                 return 0;
1357
1358         shost = sdev->host;
1359         starget = scsi_target(sdev);
1360
1361         if (scsi_host_in_recovery(shost) || scsi_host_is_busy(shost) ||
1362             scsi_target_is_busy(starget) || scsi_device_is_busy(sdev))
1363                 return 1;
1364
1365         return 0;
1366 }
1367
1368 /*
1369  * Kill a request for a dead device
1370  */
1371 static void scsi_kill_request(struct request *req, struct request_queue *q)
1372 {
1373         struct scsi_cmnd *cmd = req->special;
1374         struct scsi_device *sdev = cmd->device;
1375         struct scsi_target *starget = scsi_target(sdev);
1376         struct Scsi_Host *shost = sdev->host;
1377
1378         blkdev_dequeue_request(req);
1379
1380         if (unlikely(cmd == NULL)) {
1381                 printk(KERN_CRIT "impossible request in %s.\n",
1382                                  __func__);
1383                 BUG();
1384         }
1385
1386         scsi_init_cmd_errh(cmd);
1387         cmd->result = DID_NO_CONNECT << 16;
1388         atomic_inc(&cmd->device->iorequest_cnt);
1389
1390         /*
1391          * SCSI request completion path will do scsi_device_unbusy(),
1392          * bump busy counts.  To bump the counters, we need to dance
1393          * with the locks as normal issue path does.
1394          */
1395         sdev->device_busy++;
1396         spin_unlock(sdev->request_queue->queue_lock);
1397         spin_lock(shost->host_lock);
1398         shost->host_busy++;
1399         starget->target_busy++;
1400         spin_unlock(shost->host_lock);
1401         spin_lock(sdev->request_queue->queue_lock);
1402
1403         blk_complete_request(req);
1404 }
1405
1406 static void scsi_softirq_done(struct request *rq)
1407 {
1408         struct scsi_cmnd *cmd = rq->special;
1409         unsigned long wait_for = (cmd->allowed + 1) * rq->timeout;
1410         int disposition;
1411
1412         INIT_LIST_HEAD(&cmd->eh_entry);
1413
1414         /*
1415          * Set the serial numbers back to zero
1416          */
1417         cmd->serial_number = 0;
1418
1419         atomic_inc(&cmd->device->iodone_cnt);
1420         if (cmd->result)
1421                 atomic_inc(&cmd->device->ioerr_cnt);
1422
1423         disposition = scsi_decide_disposition(cmd);
1424         if (disposition != SUCCESS &&
1425             time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) {
1426                 sdev_printk(KERN_ERR, cmd->device,
1427                             "timing out command, waited %lus\n",
1428                             wait_for/HZ);
1429                 disposition = SUCCESS;
1430         }
1431                         
1432         scsi_log_completion(cmd, disposition);
1433
1434         switch (disposition) {
1435                 case SUCCESS:
1436                         scsi_finish_command(cmd);
1437                         break;
1438                 case NEEDS_RETRY:
1439                         scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY);
1440                         break;
1441                 case ADD_TO_MLQUEUE:
1442                         scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY);
1443                         break;
1444                 default:
1445                         if (!scsi_eh_scmd_add(cmd, 0))
1446                                 scsi_finish_command(cmd);
1447         }
1448 }
1449
1450 /*
1451  * Function:    scsi_request_fn()
1452  *
1453  * Purpose:     Main strategy routine for SCSI.
1454  *
1455  * Arguments:   q       - Pointer to actual queue.
1456  *
1457  * Returns:     Nothing
1458  *
1459  * Lock status: IO request lock assumed to be held when called.
1460  */
1461 static void scsi_request_fn(struct request_queue *q)
1462 {
1463         struct scsi_device *sdev = q->queuedata;
1464         struct Scsi_Host *shost;
1465         struct scsi_cmnd *cmd;
1466         struct request *req;
1467
1468         if (!sdev) {
1469                 printk("scsi: killing requests for dead queue\n");
1470                 while ((req = elv_next_request(q)) != NULL)
1471                         scsi_kill_request(req, q);
1472                 return;
1473         }
1474
1475         if(!get_device(&sdev->sdev_gendev))
1476                 /* We must be tearing the block queue down already */
1477                 return;
1478
1479         /*
1480          * To start with, we keep looping until the queue is empty, or until
1481          * the host is no longer able to accept any more requests.
1482          */
1483         shost = sdev->host;
1484         while (!blk_queue_plugged(q)) {
1485                 int rtn;
1486                 /*
1487                  * get next queueable request.  We do this early to make sure
1488                  * that the request is fully prepared even if we cannot 
1489                  * accept it.
1490                  */
1491                 req = elv_next_request(q);
1492                 if (!req || !scsi_dev_queue_ready(q, sdev))
1493                         break;
1494
1495                 if (unlikely(!scsi_device_online(sdev))) {
1496                         sdev_printk(KERN_ERR, sdev,
1497                                     "rejecting I/O to offline device\n");
1498                         scsi_kill_request(req, q);
1499                         continue;
1500                 }
1501
1502
1503                 /*
1504                  * Remove the request from the request list.
1505                  */
1506                 if (!(blk_queue_tagged(q) && !blk_queue_start_tag(q, req)))
1507                         blkdev_dequeue_request(req);
1508                 sdev->device_busy++;
1509
1510                 spin_unlock(q->queue_lock);
1511                 cmd = req->special;
1512                 if (unlikely(cmd == NULL)) {
1513                         printk(KERN_CRIT "impossible request in %s.\n"
1514                                          "please mail a stack trace to "
1515                                          "linux-scsi@vger.kernel.org\n",
1516                                          __func__);
1517                         blk_dump_rq_flags(req, "foo");
1518                         BUG();
1519                 }
1520                 spin_lock(shost->host_lock);
1521
1522                 /*
1523                  * We hit this when the driver is using a host wide
1524                  * tag map. For device level tag maps the queue_depth check
1525                  * in the device ready fn would prevent us from trying
1526                  * to allocate a tag. Since the map is a shared host resource
1527                  * we add the dev to the starved list so it eventually gets
1528                  * a run when a tag is freed.
1529                  */
1530                 if (blk_queue_tagged(q) && !blk_rq_tagged(req)) {
1531                         if (list_empty(&sdev->starved_entry))
1532                                 list_add_tail(&sdev->starved_entry,
1533                                               &shost->starved_list);
1534                         goto not_ready;
1535                 }
1536
1537                 if (!scsi_target_queue_ready(shost, sdev))
1538                         goto not_ready;
1539
1540                 if (!scsi_host_queue_ready(q, shost, sdev))
1541                         goto not_ready;
1542
1543                 scsi_target(sdev)->target_busy++;
1544                 shost->host_busy++;
1545
1546                 /*
1547                  * XXX(hch): This is rather suboptimal, scsi_dispatch_cmd will
1548                  *              take the lock again.
1549                  */
1550                 spin_unlock_irq(shost->host_lock);
1551
1552                 /*
1553                  * Finally, initialize any error handling parameters, and set up
1554                  * the timers for timeouts.
1555                  */
1556                 scsi_init_cmd_errh(cmd);
1557
1558                 /*
1559                  * Dispatch the command to the low-level driver.
1560                  */
1561                 rtn = scsi_dispatch_cmd(cmd);
1562                 spin_lock_irq(q->queue_lock);
1563                 if(rtn) {
1564                         /* we're refusing the command; because of
1565                          * the way locks get dropped, we need to 
1566                          * check here if plugging is required */
1567                         if(sdev->device_busy == 0)
1568                                 blk_plug_device(q);
1569
1570                         break;
1571                 }
1572         }
1573
1574         goto out;
1575
1576  not_ready:
1577         spin_unlock_irq(shost->host_lock);
1578
1579         /*
1580          * lock q, handle tag, requeue req, and decrement device_busy. We
1581          * must return with queue_lock held.
1582          *
1583          * Decrementing device_busy without checking it is OK, as all such
1584          * cases (host limits or settings) should run the queue at some
1585          * later time.
1586          */
1587         spin_lock_irq(q->queue_lock);
1588         blk_requeue_request(q, req);
1589         sdev->device_busy--;
1590         if(sdev->device_busy == 0)
1591                 blk_plug_device(q);
1592  out:
1593         /* must be careful here...if we trigger the ->remove() function
1594          * we cannot be holding the q lock */
1595         spin_unlock_irq(q->queue_lock);
1596         put_device(&sdev->sdev_gendev);
1597         spin_lock_irq(q->queue_lock);
1598 }
1599
1600 u64 scsi_calculate_bounce_limit(struct Scsi_Host *shost)
1601 {
1602         struct device *host_dev;
1603         u64 bounce_limit = 0xffffffff;
1604
1605         if (shost->unchecked_isa_dma)
1606                 return BLK_BOUNCE_ISA;
1607         /*
1608          * Platforms with virtual-DMA translation
1609          * hardware have no practical limit.
1610          */
1611         if (!PCI_DMA_BUS_IS_PHYS)
1612                 return BLK_BOUNCE_ANY;
1613
1614         host_dev = scsi_get_device(shost);
1615         if (host_dev && host_dev->dma_mask)
1616                 bounce_limit = *host_dev->dma_mask;
1617
1618         return bounce_limit;
1619 }
1620 EXPORT_SYMBOL(scsi_calculate_bounce_limit);
1621
1622 struct request_queue *__scsi_alloc_queue(struct Scsi_Host *shost,
1623                                          request_fn_proc *request_fn)
1624 {
1625         struct request_queue *q;
1626         struct device *dev = shost->shost_gendev.parent;
1627
1628         q = blk_init_queue(request_fn, NULL);
1629         if (!q)
1630                 return NULL;
1631
1632         /*
1633          * this limit is imposed by hardware restrictions
1634          */
1635         blk_queue_max_hw_segments(q, shost->sg_tablesize);
1636         blk_queue_max_phys_segments(q, SCSI_MAX_SG_CHAIN_SEGMENTS);
1637
1638         blk_queue_max_sectors(q, shost->max_sectors);
1639         blk_queue_bounce_limit(q, scsi_calculate_bounce_limit(shost));
1640         blk_queue_segment_boundary(q, shost->dma_boundary);
1641         dma_set_seg_boundary(dev, shost->dma_boundary);
1642
1643         blk_queue_max_segment_size(q, dma_get_max_seg_size(dev));
1644
1645         /* New queue, no concurrency on queue_flags */
1646         if (!shost->use_clustering)
1647                 queue_flag_clear_unlocked(QUEUE_FLAG_CLUSTER, q);
1648
1649         /*
1650          * set a reasonable default alignment on word boundaries: the
1651          * host and device may alter it using
1652          * blk_queue_update_dma_alignment() later.
1653          */
1654         blk_queue_dma_alignment(q, 0x03);
1655
1656         return q;
1657 }
1658 EXPORT_SYMBOL(__scsi_alloc_queue);
1659
1660 struct request_queue *scsi_alloc_queue(struct scsi_device *sdev)
1661 {
1662         struct request_queue *q;
1663
1664         q = __scsi_alloc_queue(sdev->host, scsi_request_fn);
1665         if (!q)
1666                 return NULL;
1667
1668         blk_queue_prep_rq(q, scsi_prep_fn);
1669         blk_queue_softirq_done(q, scsi_softirq_done);
1670         blk_queue_rq_timed_out(q, scsi_times_out);
1671         blk_queue_lld_busy(q, scsi_lld_busy);
1672         return q;
1673 }
1674
1675 void scsi_free_queue(struct request_queue *q)
1676 {
1677         blk_cleanup_queue(q);
1678 }
1679
1680 /*
1681  * Function:    scsi_block_requests()
1682  *
1683  * Purpose:     Utility function used by low-level drivers to prevent further
1684  *              commands from being queued to the device.
1685  *
1686  * Arguments:   shost       - Host in question
1687  *
1688  * Returns:     Nothing
1689  *
1690  * Lock status: No locks are assumed held.
1691  *
1692  * Notes:       There is no timer nor any other means by which the requests
1693  *              get unblocked other than the low-level driver calling
1694  *              scsi_unblock_requests().
1695  */
1696 void scsi_block_requests(struct Scsi_Host *shost)
1697 {
1698         shost->host_self_blocked = 1;
1699 }
1700 EXPORT_SYMBOL(scsi_block_requests);
1701
1702 /*
1703  * Function:    scsi_unblock_requests()
1704  *
1705  * Purpose:     Utility function used by low-level drivers to allow further
1706  *              commands from being queued to the device.
1707  *
1708  * Arguments:   shost       - Host in question
1709  *
1710  * Returns:     Nothing
1711  *
1712  * Lock status: No locks are assumed held.
1713  *
1714  * Notes:       There is no timer nor any other means by which the requests
1715  *              get unblocked other than the low-level driver calling
1716  *              scsi_unblock_requests().
1717  *
1718  *              This is done as an API function so that changes to the
1719  *              internals of the scsi mid-layer won't require wholesale
1720  *              changes to drivers that use this feature.
1721  */
1722 void scsi_unblock_requests(struct Scsi_Host *shost)
1723 {
1724         shost->host_self_blocked = 0;
1725         scsi_run_host_queues(shost);
1726 }
1727 EXPORT_SYMBOL(scsi_unblock_requests);
1728
1729 int __init scsi_init_queue(void)
1730 {
1731         int i;
1732
1733         scsi_sdb_cache = kmem_cache_create("scsi_data_buffer",
1734                                            sizeof(struct scsi_data_buffer),
1735                                            0, 0, NULL);
1736         if (!scsi_sdb_cache) {
1737                 printk(KERN_ERR "SCSI: can't init scsi sdb cache\n");
1738                 return -ENOMEM;
1739         }
1740
1741         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1742                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1743                 int size = sgp->size * sizeof(struct scatterlist);
1744
1745                 sgp->slab = kmem_cache_create(sgp->name, size, 0,
1746                                 SLAB_HWCACHE_ALIGN, NULL);
1747                 if (!sgp->slab) {
1748                         printk(KERN_ERR "SCSI: can't init sg slab %s\n",
1749                                         sgp->name);
1750                         goto cleanup_sdb;
1751                 }
1752
1753                 sgp->pool = mempool_create_slab_pool(SG_MEMPOOL_SIZE,
1754                                                      sgp->slab);
1755                 if (!sgp->pool) {
1756                         printk(KERN_ERR "SCSI: can't init sg mempool %s\n",
1757                                         sgp->name);
1758                         goto cleanup_sdb;
1759                 }
1760         }
1761
1762         return 0;
1763
1764 cleanup_sdb:
1765         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1766                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1767                 if (sgp->pool)
1768                         mempool_destroy(sgp->pool);
1769                 if (sgp->slab)
1770                         kmem_cache_destroy(sgp->slab);
1771         }
1772         kmem_cache_destroy(scsi_sdb_cache);
1773
1774         return -ENOMEM;
1775 }
1776
1777 void scsi_exit_queue(void)
1778 {
1779         int i;
1780
1781         kmem_cache_destroy(scsi_sdb_cache);
1782
1783         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1784                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1785                 mempool_destroy(sgp->pool);
1786                 kmem_cache_destroy(sgp->slab);
1787         }
1788 }
1789
1790 /**
1791  *      scsi_mode_select - issue a mode select
1792  *      @sdev:  SCSI device to be queried
1793  *      @pf:    Page format bit (1 == standard, 0 == vendor specific)
1794  *      @sp:    Save page bit (0 == don't save, 1 == save)
1795  *      @modepage: mode page being requested
1796  *      @buffer: request buffer (may not be smaller than eight bytes)
1797  *      @len:   length of request buffer.
1798  *      @timeout: command timeout
1799  *      @retries: number of retries before failing
1800  *      @data: returns a structure abstracting the mode header data
1801  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
1802  *              must be SCSI_SENSE_BUFFERSIZE big.
1803  *
1804  *      Returns zero if successful; negative error number or scsi
1805  *      status on error
1806  *
1807  */
1808 int
1809 scsi_mode_select(struct scsi_device *sdev, int pf, int sp, int modepage,
1810                  unsigned char *buffer, int len, int timeout, int retries,
1811                  struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1812 {
1813         unsigned char cmd[10];
1814         unsigned char *real_buffer;
1815         int ret;
1816
1817         memset(cmd, 0, sizeof(cmd));
1818         cmd[1] = (pf ? 0x10 : 0) | (sp ? 0x01 : 0);
1819
1820         if (sdev->use_10_for_ms) {
1821                 if (len > 65535)
1822                         return -EINVAL;
1823                 real_buffer = kmalloc(8 + len, GFP_KERNEL);
1824                 if (!real_buffer)
1825                         return -ENOMEM;
1826                 memcpy(real_buffer + 8, buffer, len);
1827                 len += 8;
1828                 real_buffer[0] = 0;
1829                 real_buffer[1] = 0;
1830                 real_buffer[2] = data->medium_type;
1831                 real_buffer[3] = data->device_specific;
1832                 real_buffer[4] = data->longlba ? 0x01 : 0;
1833                 real_buffer[5] = 0;
1834                 real_buffer[6] = data->block_descriptor_length >> 8;
1835                 real_buffer[7] = data->block_descriptor_length;
1836
1837                 cmd[0] = MODE_SELECT_10;
1838                 cmd[7] = len >> 8;
1839                 cmd[8] = len;
1840         } else {
1841                 if (len > 255 || data->block_descriptor_length > 255 ||
1842                     data->longlba)
1843                         return -EINVAL;
1844
1845                 real_buffer = kmalloc(4 + len, GFP_KERNEL);
1846                 if (!real_buffer)
1847                         return -ENOMEM;
1848                 memcpy(real_buffer + 4, buffer, len);
1849                 len += 4;
1850                 real_buffer[0] = 0;
1851                 real_buffer[1] = data->medium_type;
1852                 real_buffer[2] = data->device_specific;
1853                 real_buffer[3] = data->block_descriptor_length;
1854                 
1855
1856                 cmd[0] = MODE_SELECT;
1857                 cmd[4] = len;
1858         }
1859
1860         ret = scsi_execute_req(sdev, cmd, DMA_TO_DEVICE, real_buffer, len,
1861                                sshdr, timeout, retries, NULL);
1862         kfree(real_buffer);
1863         return ret;
1864 }
1865 EXPORT_SYMBOL_GPL(scsi_mode_select);
1866
1867 /**
1868  *      scsi_mode_sense - issue a mode sense, falling back from 10 to six bytes if necessary.
1869  *      @sdev:  SCSI device to be queried
1870  *      @dbd:   set if mode sense will allow block descriptors to be returned
1871  *      @modepage: mode page being requested
1872  *      @buffer: request buffer (may not be smaller than eight bytes)
1873  *      @len:   length of request buffer.
1874  *      @timeout: command timeout
1875  *      @retries: number of retries before failing
1876  *      @data: returns a structure abstracting the mode header data
1877  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
1878  *              must be SCSI_SENSE_BUFFERSIZE big.
1879  *
1880  *      Returns zero if unsuccessful, or the header offset (either 4
1881  *      or 8 depending on whether a six or ten byte command was
1882  *      issued) if successful.
1883  */
1884 int
1885 scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
1886                   unsigned char *buffer, int len, int timeout, int retries,
1887                   struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1888 {
1889         unsigned char cmd[12];
1890         int use_10_for_ms;
1891         int header_length;
1892         int result;
1893         struct scsi_sense_hdr my_sshdr;
1894
1895         memset(data, 0, sizeof(*data));
1896         memset(&cmd[0], 0, 12);
1897         cmd[1] = dbd & 0x18;    /* allows DBD and LLBA bits */
1898         cmd[2] = modepage;
1899
1900         /* caller might not be interested in sense, but we need it */
1901         if (!sshdr)
1902                 sshdr = &my_sshdr;
1903
1904  retry:
1905         use_10_for_ms = sdev->use_10_for_ms;
1906
1907         if (use_10_for_ms) {
1908                 if (len < 8)
1909                         len = 8;
1910
1911                 cmd[0] = MODE_SENSE_10;
1912                 cmd[8] = len;
1913                 header_length = 8;
1914         } else {
1915                 if (len < 4)
1916                         len = 4;
1917
1918                 cmd[0] = MODE_SENSE;
1919                 cmd[4] = len;
1920                 header_length = 4;
1921         }
1922
1923         memset(buffer, 0, len);
1924
1925         result = scsi_execute_req(sdev, cmd, DMA_FROM_DEVICE, buffer, len,
1926                                   sshdr, timeout, retries, NULL);
1927
1928         /* This code looks awful: what it's doing is making sure an
1929          * ILLEGAL REQUEST sense return identifies the actual command
1930          * byte as the problem.  MODE_SENSE commands can return
1931          * ILLEGAL REQUEST if the code page isn't supported */
1932
1933         if (use_10_for_ms && !scsi_status_is_good(result) &&
1934             (driver_byte(result) & DRIVER_SENSE)) {
1935                 if (scsi_sense_valid(sshdr)) {
1936                         if ((sshdr->sense_key == ILLEGAL_REQUEST) &&
1937                             (sshdr->asc == 0x20) && (sshdr->ascq == 0)) {
1938                                 /* 
1939                                  * Invalid command operation code
1940                                  */
1941                                 sdev->use_10_for_ms = 0;
1942                                 goto retry;
1943                         }
1944                 }
1945         }
1946
1947         if(scsi_status_is_good(result)) {
1948                 if (unlikely(buffer[0] == 0x86 && buffer[1] == 0x0b &&
1949                              (modepage == 6 || modepage == 8))) {
1950                         /* Initio breakage? */
1951                         header_length = 0;
1952                         data->length = 13;
1953                         data->medium_type = 0;
1954                         data->device_specific = 0;
1955                         data->longlba = 0;
1956                         data->block_descriptor_length = 0;
1957                 } else if(use_10_for_ms) {
1958                         data->length = buffer[0]*256 + buffer[1] + 2;
1959                         data->medium_type = buffer[2];
1960                         data->device_specific = buffer[3];
1961                         data->longlba = buffer[4] & 0x01;
1962                         data->block_descriptor_length = buffer[6]*256
1963                                 + buffer[7];
1964                 } else {
1965                         data->length = buffer[0] + 1;
1966                         data->medium_type = buffer[1];
1967                         data->device_specific = buffer[2];
1968                         data->block_descriptor_length = buffer[3];
1969                 }
1970                 data->header_length = header_length;
1971         }
1972
1973         return result;
1974 }
1975 EXPORT_SYMBOL(scsi_mode_sense);
1976
1977 /**
1978  *      scsi_test_unit_ready - test if unit is ready
1979  *      @sdev:  scsi device to change the state of.
1980  *      @timeout: command timeout
1981  *      @retries: number of retries before failing
1982  *      @sshdr_external: Optional pointer to struct scsi_sense_hdr for
1983  *              returning sense. Make sure that this is cleared before passing
1984  *              in.
1985  *
1986  *      Returns zero if unsuccessful or an error if TUR failed.  For
1987  *      removable media, a return of NOT_READY or UNIT_ATTENTION is
1988  *      translated to success, with the ->changed flag updated.
1989  **/
1990 int
1991 scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries,
1992                      struct scsi_sense_hdr *sshdr_external)
1993 {
1994         char cmd[] = {
1995                 TEST_UNIT_READY, 0, 0, 0, 0, 0,
1996         };
1997         struct scsi_sense_hdr *sshdr;
1998         int result;
1999
2000         if (!sshdr_external)
2001                 sshdr = kzalloc(sizeof(*sshdr), GFP_KERNEL);
2002         else
2003                 sshdr = sshdr_external;
2004
2005         /* try to eat the UNIT_ATTENTION if there are enough retries */
2006         do {
2007                 result = scsi_execute_req(sdev, cmd, DMA_NONE, NULL, 0, sshdr,
2008                                           timeout, retries, NULL);
2009                 if (sdev->removable && scsi_sense_valid(sshdr) &&
2010                     sshdr->sense_key == UNIT_ATTENTION)
2011                         sdev->changed = 1;
2012         } while (scsi_sense_valid(sshdr) &&
2013                  sshdr->sense_key == UNIT_ATTENTION && --retries);
2014
2015         if (!sshdr)
2016                 /* could not allocate sense buffer, so can't process it */
2017                 return result;
2018
2019         if (sdev->removable && scsi_sense_valid(sshdr) &&
2020             (sshdr->sense_key == UNIT_ATTENTION ||
2021              sshdr->sense_key == NOT_READY)) {
2022                 sdev->changed = 1;
2023                 result = 0;
2024         }
2025         if (!sshdr_external)
2026                 kfree(sshdr);
2027         return result;
2028 }
2029 EXPORT_SYMBOL(scsi_test_unit_ready);
2030
2031 /**
2032  *      scsi_device_set_state - Take the given device through the device state model.
2033  *      @sdev:  scsi device to change the state of.
2034  *      @state: state to change to.
2035  *
2036  *      Returns zero if unsuccessful or an error if the requested 
2037  *      transition is illegal.
2038  */
2039 int
2040 scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
2041 {
2042         enum scsi_device_state oldstate = sdev->sdev_state;
2043
2044         if (state == oldstate)
2045                 return 0;
2046
2047         switch (state) {
2048         case SDEV_CREATED:
2049                 switch (oldstate) {
2050                 case SDEV_CREATED_BLOCK:
2051                         break;
2052                 default:
2053                         goto illegal;
2054                 }
2055                 break;
2056                         
2057         case SDEV_RUNNING:
2058                 switch (oldstate) {
2059                 case SDEV_CREATED:
2060                 case SDEV_OFFLINE:
2061                 case SDEV_QUIESCE:
2062                 case SDEV_BLOCK:
2063                         break;
2064                 default:
2065                         goto illegal;
2066                 }
2067                 break;
2068
2069         case SDEV_QUIESCE:
2070                 switch (oldstate) {
2071                 case SDEV_RUNNING:
2072                 case SDEV_OFFLINE:
2073                         break;
2074                 default:
2075                         goto illegal;
2076                 }
2077                 break;
2078
2079         case SDEV_OFFLINE:
2080                 switch (oldstate) {
2081                 case SDEV_CREATED:
2082                 case SDEV_RUNNING:
2083                 case SDEV_QUIESCE:
2084                 case SDEV_BLOCK:
2085                         break;
2086                 default:
2087                         goto illegal;
2088                 }
2089                 break;
2090
2091         case SDEV_BLOCK:
2092                 switch (oldstate) {
2093                 case SDEV_RUNNING:
2094                 case SDEV_CREATED_BLOCK:
2095                         break;
2096                 default:
2097                         goto illegal;
2098                 }
2099                 break;
2100
2101         case SDEV_CREATED_BLOCK:
2102                 switch (oldstate) {
2103                 case SDEV_CREATED:
2104                         break;
2105                 default:
2106                         goto illegal;
2107                 }
2108                 break;
2109
2110         case SDEV_CANCEL:
2111                 switch (oldstate) {
2112                 case SDEV_CREATED:
2113                 case SDEV_RUNNING:
2114                 case SDEV_QUIESCE:
2115                 case SDEV_OFFLINE:
2116                 case SDEV_BLOCK:
2117                         break;
2118                 default:
2119                         goto illegal;
2120                 }
2121                 break;
2122
2123         case SDEV_DEL:
2124                 switch (oldstate) {
2125                 case SDEV_CREATED:
2126                 case SDEV_RUNNING:
2127                 case SDEV_OFFLINE:
2128                 case SDEV_CANCEL:
2129                         break;
2130                 default:
2131                         goto illegal;
2132                 }
2133                 break;
2134
2135         }
2136         sdev->sdev_state = state;
2137         return 0;
2138
2139  illegal:
2140         SCSI_LOG_ERROR_RECOVERY(1, 
2141                                 sdev_printk(KERN_ERR, sdev,
2142                                             "Illegal state transition %s->%s\n",
2143                                             scsi_device_state_name(oldstate),
2144                                             scsi_device_state_name(state))
2145                                 );
2146         return -EINVAL;
2147 }
2148 EXPORT_SYMBOL(scsi_device_set_state);
2149
2150 /**
2151  *      sdev_evt_emit - emit a single SCSI device uevent
2152  *      @sdev: associated SCSI device
2153  *      @evt: event to emit
2154  *
2155  *      Send a single uevent (scsi_event) to the associated scsi_device.
2156  */
2157 static void scsi_evt_emit(struct scsi_device *sdev, struct scsi_event *evt)
2158 {
2159         int idx = 0;
2160         char *envp[3];
2161
2162         switch (evt->evt_type) {
2163         case SDEV_EVT_MEDIA_CHANGE:
2164                 envp[idx++] = "SDEV_MEDIA_CHANGE=1";
2165                 break;
2166
2167         default:
2168                 /* do nothing */
2169                 break;
2170         }
2171
2172         envp[idx++] = NULL;
2173
2174         kobject_uevent_env(&sdev->sdev_gendev.kobj, KOBJ_CHANGE, envp);
2175 }
2176
2177 /**
2178  *      sdev_evt_thread - send a uevent for each scsi event
2179  *      @work: work struct for scsi_device
2180  *
2181  *      Dispatch queued events to their associated scsi_device kobjects
2182  *      as uevents.
2183  */
2184 void scsi_evt_thread(struct work_struct *work)
2185 {
2186         struct scsi_device *sdev;
2187         LIST_HEAD(event_list);
2188
2189         sdev = container_of(work, struct scsi_device, event_work);
2190
2191         while (1) {
2192                 struct scsi_event *evt;
2193                 struct list_head *this, *tmp;
2194                 unsigned long flags;
2195
2196                 spin_lock_irqsave(&sdev->list_lock, flags);
2197                 list_splice_init(&sdev->event_list, &event_list);
2198                 spin_unlock_irqrestore(&sdev->list_lock, flags);
2199
2200                 if (list_empty(&event_list))
2201                         break;
2202
2203                 list_for_each_safe(this, tmp, &event_list) {
2204                         evt = list_entry(this, struct scsi_event, node);
2205                         list_del(&evt->node);
2206                         scsi_evt_emit(sdev, evt);
2207                         kfree(evt);
2208                 }
2209         }
2210 }
2211
2212 /**
2213  *      sdev_evt_send - send asserted event to uevent thread
2214  *      @sdev: scsi_device event occurred on
2215  *      @evt: event to send
2216  *
2217  *      Assert scsi device event asynchronously.
2218  */
2219 void sdev_evt_send(struct scsi_device *sdev, struct scsi_event *evt)
2220 {
2221         unsigned long flags;
2222
2223 #if 0
2224         /* FIXME: currently this check eliminates all media change events
2225          * for polled devices.  Need to update to discriminate between AN
2226          * and polled events */
2227         if (!test_bit(evt->evt_type, sdev->supported_events)) {
2228                 kfree(evt);
2229                 return;
2230         }
2231 #endif
2232
2233         spin_lock_irqsave(&sdev->list_lock, flags);
2234         list_add_tail(&evt->node, &sdev->event_list);
2235         schedule_work(&sdev->event_work);
2236         spin_unlock_irqrestore(&sdev->list_lock, flags);
2237 }
2238 EXPORT_SYMBOL_GPL(sdev_evt_send);
2239
2240 /**
2241  *      sdev_evt_alloc - allocate a new scsi event
2242  *      @evt_type: type of event to allocate
2243  *      @gfpflags: GFP flags for allocation
2244  *
2245  *      Allocates and returns a new scsi_event.
2246  */
2247 struct scsi_event *sdev_evt_alloc(enum scsi_device_event evt_type,
2248                                   gfp_t gfpflags)
2249 {
2250         struct scsi_event *evt = kzalloc(sizeof(struct scsi_event), gfpflags);
2251         if (!evt)
2252                 return NULL;
2253
2254         evt->evt_type = evt_type;
2255         INIT_LIST_HEAD(&evt->node);
2256
2257         /* evt_type-specific initialization, if any */
2258         switch (evt_type) {
2259         case SDEV_EVT_MEDIA_CHANGE:
2260         default:
2261                 /* do nothing */
2262                 break;
2263         }
2264
2265         return evt;
2266 }
2267 EXPORT_SYMBOL_GPL(sdev_evt_alloc);
2268
2269 /**
2270  *      sdev_evt_send_simple - send asserted event to uevent thread
2271  *      @sdev: scsi_device event occurred on
2272  *      @evt_type: type of event to send
2273  *      @gfpflags: GFP flags for allocation
2274  *
2275  *      Assert scsi device event asynchronously, given an event type.
2276  */
2277 void sdev_evt_send_simple(struct scsi_device *sdev,
2278                           enum scsi_device_event evt_type, gfp_t gfpflags)
2279 {
2280         struct scsi_event *evt = sdev_evt_alloc(evt_type, gfpflags);
2281         if (!evt) {
2282                 sdev_printk(KERN_ERR, sdev, "event %d eaten due to OOM\n",
2283                             evt_type);
2284                 return;
2285         }
2286
2287         sdev_evt_send(sdev, evt);
2288 }
2289 EXPORT_SYMBOL_GPL(sdev_evt_send_simple);
2290
2291 /**
2292  *      scsi_device_quiesce - Block user issued commands.
2293  *      @sdev:  scsi device to quiesce.
2294  *
2295  *      This works by trying to transition to the SDEV_QUIESCE state
2296  *      (which must be a legal transition).  When the device is in this
2297  *      state, only special requests will be accepted, all others will
2298  *      be deferred.  Since special requests may also be requeued requests,
2299  *      a successful return doesn't guarantee the device will be 
2300  *      totally quiescent.
2301  *
2302  *      Must be called with user context, may sleep.
2303  *
2304  *      Returns zero if unsuccessful or an error if not.
2305  */
2306 int
2307 scsi_device_quiesce(struct scsi_device *sdev)
2308 {
2309         int err = scsi_device_set_state(sdev, SDEV_QUIESCE);
2310         if (err)
2311                 return err;
2312
2313         scsi_run_queue(sdev->request_queue);
2314         while (sdev->device_busy) {
2315                 msleep_interruptible(200);
2316                 scsi_run_queue(sdev->request_queue);
2317         }
2318         return 0;
2319 }
2320 EXPORT_SYMBOL(scsi_device_quiesce);
2321
2322 /**
2323  *      scsi_device_resume - Restart user issued commands to a quiesced device.
2324  *      @sdev:  scsi device to resume.
2325  *
2326  *      Moves the device from quiesced back to running and restarts the
2327  *      queues.
2328  *
2329  *      Must be called with user context, may sleep.
2330  */
2331 void
2332 scsi_device_resume(struct scsi_device *sdev)
2333 {
2334         if(scsi_device_set_state(sdev, SDEV_RUNNING))
2335                 return;
2336         scsi_run_queue(sdev->request_queue);
2337 }
2338 EXPORT_SYMBOL(scsi_device_resume);
2339
2340 static void
2341 device_quiesce_fn(struct scsi_device *sdev, void *data)
2342 {
2343         scsi_device_quiesce(sdev);
2344 }
2345
2346 void
2347 scsi_target_quiesce(struct scsi_target *starget)
2348 {
2349         starget_for_each_device(starget, NULL, device_quiesce_fn);
2350 }
2351 EXPORT_SYMBOL(scsi_target_quiesce);
2352
2353 static void
2354 device_resume_fn(struct scsi_device *sdev, void *data)
2355 {
2356         scsi_device_resume(sdev);
2357 }
2358
2359 void
2360 scsi_target_resume(struct scsi_target *starget)
2361 {
2362         starget_for_each_device(starget, NULL, device_resume_fn);
2363 }
2364 EXPORT_SYMBOL(scsi_target_resume);
2365
2366 /**
2367  * scsi_internal_device_block - internal function to put a device temporarily into the SDEV_BLOCK state
2368  * @sdev:       device to block
2369  *
2370  * Block request made by scsi lld's to temporarily stop all
2371  * scsi commands on the specified device.  Called from interrupt
2372  * or normal process context.
2373  *
2374  * Returns zero if successful or error if not
2375  *
2376  * Notes:       
2377  *      This routine transitions the device to the SDEV_BLOCK state
2378  *      (which must be a legal transition).  When the device is in this
2379  *      state, all commands are deferred until the scsi lld reenables
2380  *      the device with scsi_device_unblock or device_block_tmo fires.
2381  *      This routine assumes the host_lock is held on entry.
2382  */
2383 int
2384 scsi_internal_device_block(struct scsi_device *sdev)
2385 {
2386         struct request_queue *q = sdev->request_queue;
2387         unsigned long flags;
2388         int err = 0;
2389
2390         err = scsi_device_set_state(sdev, SDEV_BLOCK);
2391         if (err) {
2392                 err = scsi_device_set_state(sdev, SDEV_CREATED_BLOCK);
2393
2394                 if (err)
2395                         return err;
2396         }
2397
2398         /* 
2399          * The device has transitioned to SDEV_BLOCK.  Stop the
2400          * block layer from calling the midlayer with this device's
2401          * request queue. 
2402          */
2403         spin_lock_irqsave(q->queue_lock, flags);
2404         blk_stop_queue(q);
2405         spin_unlock_irqrestore(q->queue_lock, flags);
2406
2407         return 0;
2408 }
2409 EXPORT_SYMBOL_GPL(scsi_internal_device_block);
2410  
2411 /**
2412  * scsi_internal_device_unblock - resume a device after a block request
2413  * @sdev:       device to resume
2414  *
2415  * Called by scsi lld's or the midlayer to restart the device queue
2416  * for the previously suspended scsi device.  Called from interrupt or
2417  * normal process context.
2418  *
2419  * Returns zero if successful or error if not.
2420  *
2421  * Notes:       
2422  *      This routine transitions the device to the SDEV_RUNNING state
2423  *      (which must be a legal transition) allowing the midlayer to
2424  *      goose the queue for this device.  This routine assumes the 
2425  *      host_lock is held upon entry.
2426  */
2427 int
2428 scsi_internal_device_unblock(struct scsi_device *sdev)
2429 {
2430         struct request_queue *q = sdev->request_queue; 
2431         int err;
2432         unsigned long flags;
2433         
2434         /* 
2435          * Try to transition the scsi device to SDEV_RUNNING
2436          * and goose the device queue if successful.  
2437          */
2438         err = scsi_device_set_state(sdev, SDEV_RUNNING);
2439         if (err) {
2440                 err = scsi_device_set_state(sdev, SDEV_CREATED);
2441
2442                 if (err)
2443                         return err;
2444         }
2445
2446         spin_lock_irqsave(q->queue_lock, flags);
2447         blk_start_queue(q);
2448         spin_unlock_irqrestore(q->queue_lock, flags);
2449
2450         return 0;
2451 }
2452 EXPORT_SYMBOL_GPL(scsi_internal_device_unblock);
2453
2454 static void
2455 device_block(struct scsi_device *sdev, void *data)
2456 {
2457         scsi_internal_device_block(sdev);
2458 }
2459
2460 static int
2461 target_block(struct device *dev, void *data)
2462 {
2463         if (scsi_is_target_device(dev))
2464                 starget_for_each_device(to_scsi_target(dev), NULL,
2465                                         device_block);
2466         return 0;
2467 }
2468
2469 void
2470 scsi_target_block(struct device *dev)
2471 {
2472         if (scsi_is_target_device(dev))
2473                 starget_for_each_device(to_scsi_target(dev), NULL,
2474                                         device_block);
2475         else
2476                 device_for_each_child(dev, NULL, target_block);
2477 }
2478 EXPORT_SYMBOL_GPL(scsi_target_block);
2479
2480 static void
2481 device_unblock(struct scsi_device *sdev, void *data)
2482 {
2483         scsi_internal_device_unblock(sdev);
2484 }
2485
2486 static int
2487 target_unblock(struct device *dev, void *data)
2488 {
2489         if (scsi_is_target_device(dev))
2490                 starget_for_each_device(to_scsi_target(dev), NULL,
2491                                         device_unblock);
2492         return 0;
2493 }
2494
2495 void
2496 scsi_target_unblock(struct device *dev)
2497 {
2498         if (scsi_is_target_device(dev))
2499                 starget_for_each_device(to_scsi_target(dev), NULL,
2500                                         device_unblock);
2501         else
2502                 device_for_each_child(dev, NULL, target_unblock);
2503 }
2504 EXPORT_SYMBOL_GPL(scsi_target_unblock);
2505
2506 /**
2507  * scsi_kmap_atomic_sg - find and atomically map an sg-elemnt
2508  * @sgl:        scatter-gather list
2509  * @sg_count:   number of segments in sg
2510  * @offset:     offset in bytes into sg, on return offset into the mapped area
2511  * @len:        bytes to map, on return number of bytes mapped
2512  *
2513  * Returns virtual address of the start of the mapped page
2514  */
2515 void *scsi_kmap_atomic_sg(struct scatterlist *sgl, int sg_count,
2516                           size_t *offset, size_t *len)
2517 {
2518         int i;
2519         size_t sg_len = 0, len_complete = 0;
2520         struct scatterlist *sg;
2521         struct page *page;
2522
2523         WARN_ON(!irqs_disabled());
2524
2525         for_each_sg(sgl, sg, sg_count, i) {
2526                 len_complete = sg_len; /* Complete sg-entries */
2527                 sg_len += sg->length;
2528                 if (sg_len > *offset)
2529                         break;
2530         }
2531
2532         if (unlikely(i == sg_count)) {
2533                 printk(KERN_ERR "%s: Bytes in sg: %zu, requested offset %zu, "
2534                         "elements %d\n",
2535                        __func__, sg_len, *offset, sg_count);
2536                 WARN_ON(1);
2537                 return NULL;
2538         }
2539
2540         /* Offset starting from the beginning of first page in this sg-entry */
2541         *offset = *offset - len_complete + sg->offset;
2542
2543         /* Assumption: contiguous pages can be accessed as "page + i" */
2544         page = nth_page(sg_page(sg), (*offset >> PAGE_SHIFT));
2545         *offset &= ~PAGE_MASK;
2546
2547         /* Bytes in this sg-entry from *offset to the end of the page */
2548         sg_len = PAGE_SIZE - *offset;
2549         if (*len > sg_len)
2550                 *len = sg_len;
2551
2552         return kmap_atomic(page, KM_BIO_SRC_IRQ);
2553 }
2554 EXPORT_SYMBOL(scsi_kmap_atomic_sg);
2555
2556 /**
2557  * scsi_kunmap_atomic_sg - atomically unmap a virtual address, previously mapped with scsi_kmap_atomic_sg
2558  * @virt:       virtual address to be unmapped
2559  */
2560 void scsi_kunmap_atomic_sg(void *virt)
2561 {
2562         kunmap_atomic(virt, KM_BIO_SRC_IRQ);
2563 }
2564 EXPORT_SYMBOL(scsi_kunmap_atomic_sg);