scsi,block: update SCSI to handle mixed merge failures
[linux-2.6.git] / drivers / scsi / scsi_lib.c
1 /*
2  *  scsi_lib.c Copyright (C) 1999 Eric Youngdale
3  *
4  *  SCSI queueing library.
5  *      Initial versions: Eric Youngdale (eric@andante.org).
6  *                        Based upon conversations with large numbers
7  *                        of people at Linux Expo.
8  */
9
10 #include <linux/bio.h>
11 #include <linux/bitops.h>
12 #include <linux/blkdev.h>
13 #include <linux/completion.h>
14 #include <linux/kernel.h>
15 #include <linux/mempool.h>
16 #include <linux/slab.h>
17 #include <linux/init.h>
18 #include <linux/pci.h>
19 #include <linux/delay.h>
20 #include <linux/hardirq.h>
21 #include <linux/scatterlist.h>
22
23 #include <scsi/scsi.h>
24 #include <scsi/scsi_cmnd.h>
25 #include <scsi/scsi_dbg.h>
26 #include <scsi/scsi_device.h>
27 #include <scsi/scsi_driver.h>
28 #include <scsi/scsi_eh.h>
29 #include <scsi/scsi_host.h>
30
31 #include "scsi_priv.h"
32 #include "scsi_logging.h"
33
34
35 #define SG_MEMPOOL_NR           ARRAY_SIZE(scsi_sg_pools)
36 #define SG_MEMPOOL_SIZE         2
37
38 struct scsi_host_sg_pool {
39         size_t          size;
40         char            *name;
41         struct kmem_cache       *slab;
42         mempool_t       *pool;
43 };
44
45 #define SP(x) { x, "sgpool-" __stringify(x) }
46 #if (SCSI_MAX_SG_SEGMENTS < 32)
47 #error SCSI_MAX_SG_SEGMENTS is too small (must be 32 or greater)
48 #endif
49 static struct scsi_host_sg_pool scsi_sg_pools[] = {
50         SP(8),
51         SP(16),
52 #if (SCSI_MAX_SG_SEGMENTS > 32)
53         SP(32),
54 #if (SCSI_MAX_SG_SEGMENTS > 64)
55         SP(64),
56 #if (SCSI_MAX_SG_SEGMENTS > 128)
57         SP(128),
58 #if (SCSI_MAX_SG_SEGMENTS > 256)
59 #error SCSI_MAX_SG_SEGMENTS is too large (256 MAX)
60 #endif
61 #endif
62 #endif
63 #endif
64         SP(SCSI_MAX_SG_SEGMENTS)
65 };
66 #undef SP
67
68 struct kmem_cache *scsi_sdb_cache;
69
70 static void scsi_run_queue(struct request_queue *q);
71
72 /*
73  * Function:    scsi_unprep_request()
74  *
75  * Purpose:     Remove all preparation done for a request, including its
76  *              associated scsi_cmnd, so that it can be requeued.
77  *
78  * Arguments:   req     - request to unprepare
79  *
80  * Lock status: Assumed that no locks are held upon entry.
81  *
82  * Returns:     Nothing.
83  */
84 static void scsi_unprep_request(struct request *req)
85 {
86         struct scsi_cmnd *cmd = req->special;
87
88         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->resid_len > 0 && req->resid_len <= bufflen))
244                 memset(buffer + (bufflen - req->resid_len), 0, req->resid_len);
245
246         if (resid)
247                 *resid = req->resid_len;
248         ret = req->errors;
249  out:
250         blk_put_request(req);
251
252         return ret;
253 }
254 EXPORT_SYMBOL(scsi_execute);
255
256
257 int scsi_execute_req(struct scsi_device *sdev, const unsigned char *cmd,
258                      int data_direction, void *buffer, unsigned bufflen,
259                      struct scsi_sense_hdr *sshdr, int timeout, int retries,
260                      int *resid)
261 {
262         char *sense = NULL;
263         int result;
264         
265         if (sshdr) {
266                 sense = kzalloc(SCSI_SENSE_BUFFERSIZE, GFP_NOIO);
267                 if (!sense)
268                         return DRIVER_ERROR << 24;
269         }
270         result = scsi_execute(sdev, cmd, data_direction, buffer, bufflen,
271                               sense, timeout, retries, 0, resid);
272         if (sshdr)
273                 scsi_normalize_sense(sense, SCSI_SENSE_BUFFERSIZE, sshdr);
274
275         kfree(sense);
276         return result;
277 }
278 EXPORT_SYMBOL(scsi_execute_req);
279
280 /*
281  * Function:    scsi_init_cmd_errh()
282  *
283  * Purpose:     Initialize cmd fields related to error handling.
284  *
285  * Arguments:   cmd     - command that is ready to be queued.
286  *
287  * Notes:       This function has the job of initializing a number of
288  *              fields related to error handling.   Typically this will
289  *              be called once for each command, as required.
290  */
291 static void scsi_init_cmd_errh(struct scsi_cmnd *cmd)
292 {
293         cmd->serial_number = 0;
294         scsi_set_resid(cmd, 0);
295         memset(cmd->sense_buffer, 0, SCSI_SENSE_BUFFERSIZE);
296         if (cmd->cmd_len == 0)
297                 cmd->cmd_len = scsi_command_size(cmd->cmnd);
298 }
299
300 void scsi_device_unbusy(struct scsi_device *sdev)
301 {
302         struct Scsi_Host *shost = sdev->host;
303         struct scsi_target *starget = scsi_target(sdev);
304         unsigned long flags;
305
306         spin_lock_irqsave(shost->host_lock, flags);
307         shost->host_busy--;
308         starget->target_busy--;
309         if (unlikely(scsi_host_in_recovery(shost) &&
310                      (shost->host_failed || shost->host_eh_scheduled)))
311                 scsi_eh_wakeup(shost);
312         spin_unlock(shost->host_lock);
313         spin_lock(sdev->request_queue->queue_lock);
314         sdev->device_busy--;
315         spin_unlock_irqrestore(sdev->request_queue->queue_lock, flags);
316 }
317
318 /*
319  * Called for single_lun devices on IO completion. Clear starget_sdev_user,
320  * and call blk_run_queue for all the scsi_devices on the target -
321  * including current_sdev first.
322  *
323  * Called with *no* scsi locks held.
324  */
325 static void scsi_single_lun_run(struct scsi_device *current_sdev)
326 {
327         struct Scsi_Host *shost = current_sdev->host;
328         struct scsi_device *sdev, *tmp;
329         struct scsi_target *starget = scsi_target(current_sdev);
330         unsigned long flags;
331
332         spin_lock_irqsave(shost->host_lock, flags);
333         starget->starget_sdev_user = NULL;
334         spin_unlock_irqrestore(shost->host_lock, flags);
335
336         /*
337          * Call blk_run_queue for all LUNs on the target, starting with
338          * current_sdev. We race with others (to set starget_sdev_user),
339          * but in most cases, we will be first. Ideally, each LU on the
340          * target would get some limited time or requests on the target.
341          */
342         blk_run_queue(current_sdev->request_queue);
343
344         spin_lock_irqsave(shost->host_lock, flags);
345         if (starget->starget_sdev_user)
346                 goto out;
347         list_for_each_entry_safe(sdev, tmp, &starget->devices,
348                         same_target_siblings) {
349                 if (sdev == current_sdev)
350                         continue;
351                 if (scsi_device_get(sdev))
352                         continue;
353
354                 spin_unlock_irqrestore(shost->host_lock, flags);
355                 blk_run_queue(sdev->request_queue);
356                 spin_lock_irqsave(shost->host_lock, flags);
357         
358                 scsi_device_put(sdev);
359         }
360  out:
361         spin_unlock_irqrestore(shost->host_lock, flags);
362 }
363
364 static inline int scsi_device_is_busy(struct scsi_device *sdev)
365 {
366         if (sdev->device_busy >= sdev->queue_depth || sdev->device_blocked)
367                 return 1;
368
369         return 0;
370 }
371
372 static inline int scsi_target_is_busy(struct scsi_target *starget)
373 {
374         return ((starget->can_queue > 0 &&
375                  starget->target_busy >= starget->can_queue) ||
376                  starget->target_blocked);
377 }
378
379 static inline int scsi_host_is_busy(struct Scsi_Host *shost)
380 {
381         if ((shost->can_queue > 0 && shost->host_busy >= shost->can_queue) ||
382             shost->host_blocked || shost->host_self_blocked)
383                 return 1;
384
385         return 0;
386 }
387
388 /*
389  * Function:    scsi_run_queue()
390  *
391  * Purpose:     Select a proper request queue to serve next
392  *
393  * Arguments:   q       - last request's queue
394  *
395  * Returns:     Nothing
396  *
397  * Notes:       The previous command was completely finished, start
398  *              a new one if possible.
399  */
400 static void scsi_run_queue(struct request_queue *q)
401 {
402         struct scsi_device *sdev = q->queuedata;
403         struct Scsi_Host *shost = sdev->host;
404         LIST_HEAD(starved_list);
405         unsigned long flags;
406
407         if (scsi_target(sdev)->single_lun)
408                 scsi_single_lun_run(sdev);
409
410         spin_lock_irqsave(shost->host_lock, flags);
411         list_splice_init(&shost->starved_list, &starved_list);
412
413         while (!list_empty(&starved_list)) {
414                 int flagset;
415
416                 /*
417                  * As long as shost is accepting commands and we have
418                  * starved queues, call blk_run_queue. scsi_request_fn
419                  * drops the queue_lock and can add us back to the
420                  * starved_list.
421                  *
422                  * host_lock protects the starved_list and starved_entry.
423                  * scsi_request_fn must get the host_lock before checking
424                  * or modifying starved_list or starved_entry.
425                  */
426                 if (scsi_host_is_busy(shost))
427                         break;
428
429                 sdev = list_entry(starved_list.next,
430                                   struct scsi_device, starved_entry);
431                 list_del_init(&sdev->starved_entry);
432                 if (scsi_target_is_busy(scsi_target(sdev))) {
433                         list_move_tail(&sdev->starved_entry,
434                                        &shost->starved_list);
435                         continue;
436                 }
437
438                 spin_unlock(shost->host_lock);
439
440                 spin_lock(sdev->request_queue->queue_lock);
441                 flagset = test_bit(QUEUE_FLAG_REENTER, &q->queue_flags) &&
442                                 !test_bit(QUEUE_FLAG_REENTER,
443                                         &sdev->request_queue->queue_flags);
444                 if (flagset)
445                         queue_flag_set(QUEUE_FLAG_REENTER, sdev->request_queue);
446                 __blk_run_queue(sdev->request_queue);
447                 if (flagset)
448                         queue_flag_clear(QUEUE_FLAG_REENTER, sdev->request_queue);
449                 spin_unlock(sdev->request_queue->queue_lock);
450
451                 spin_lock(shost->host_lock);
452         }
453         /* put any unprocessed entries back */
454         list_splice(&starved_list, &shost->starved_list);
455         spin_unlock_irqrestore(shost->host_lock, flags);
456
457         blk_run_queue(q);
458 }
459
460 /*
461  * Function:    scsi_requeue_command()
462  *
463  * Purpose:     Handle post-processing of completed commands.
464  *
465  * Arguments:   q       - queue to operate on
466  *              cmd     - command that may need to be requeued.
467  *
468  * Returns:     Nothing
469  *
470  * Notes:       After command completion, there may be blocks left
471  *              over which weren't finished by the previous command
472  *              this can be for a number of reasons - the main one is
473  *              I/O errors in the middle of the request, in which case
474  *              we need to request the blocks that come after the bad
475  *              sector.
476  * Notes:       Upon return, cmd is a stale pointer.
477  */
478 static void scsi_requeue_command(struct request_queue *q, struct scsi_cmnd *cmd)
479 {
480         struct request *req = cmd->request;
481         unsigned long flags;
482
483         spin_lock_irqsave(q->queue_lock, flags);
484         scsi_unprep_request(req);
485         blk_requeue_request(q, req);
486         spin_unlock_irqrestore(q->queue_lock, flags);
487
488         scsi_run_queue(q);
489 }
490
491 void scsi_next_command(struct scsi_cmnd *cmd)
492 {
493         struct scsi_device *sdev = cmd->device;
494         struct request_queue *q = sdev->request_queue;
495
496         /* need to hold a reference on the device before we let go of the cmd */
497         get_device(&sdev->sdev_gendev);
498
499         scsi_put_command(cmd);
500         scsi_run_queue(q);
501
502         /* ok to remove device now */
503         put_device(&sdev->sdev_gendev);
504 }
505
506 void scsi_run_host_queues(struct Scsi_Host *shost)
507 {
508         struct scsi_device *sdev;
509
510         shost_for_each_device(sdev, shost)
511                 scsi_run_queue(sdev->request_queue);
512 }
513
514 static void __scsi_release_buffers(struct scsi_cmnd *, int);
515
516 /*
517  * Function:    scsi_end_request()
518  *
519  * Purpose:     Post-processing of completed commands (usually invoked at end
520  *              of upper level post-processing and scsi_io_completion).
521  *
522  * Arguments:   cmd      - command that is complete.
523  *              error    - 0 if I/O indicates success, < 0 for I/O error.
524  *              bytes    - number of bytes of completed I/O
525  *              requeue  - indicates whether we should requeue leftovers.
526  *
527  * Lock status: Assumed that lock is not held upon entry.
528  *
529  * Returns:     cmd if requeue required, NULL otherwise.
530  *
531  * Notes:       This is called for block device requests in order to
532  *              mark some number of sectors as complete.
533  * 
534  *              We are guaranteeing that the request queue will be goosed
535  *              at some point during this call.
536  * Notes:       If cmd was requeued, upon return it will be a stale pointer.
537  */
538 static struct scsi_cmnd *scsi_end_request(struct scsi_cmnd *cmd, int error,
539                                           int bytes, int requeue)
540 {
541         struct request_queue *q = cmd->device->request_queue;
542         struct request *req = cmd->request;
543
544         /*
545          * If there are blocks left over at the end, set up the command
546          * to queue the remainder of them.
547          */
548         if (blk_end_request(req, error, bytes)) {
549                 /* kill remainder if no retrys */
550                 if (error && scsi_noretry_cmd(cmd))
551                         blk_end_request_all(req, error);
552                 else {
553                         if (requeue) {
554                                 /*
555                                  * Bleah.  Leftovers again.  Stick the
556                                  * leftovers in the front of the
557                                  * queue, and goose the queue again.
558                                  */
559                                 scsi_release_buffers(cmd);
560                                 scsi_requeue_command(q, cmd);
561                                 cmd = NULL;
562                         }
563                         return cmd;
564                 }
565         }
566
567         /*
568          * This will goose the queue request function at the end, so we don't
569          * need to worry about launching another command.
570          */
571         __scsi_release_buffers(cmd, 0);
572         scsi_next_command(cmd);
573         return NULL;
574 }
575
576 static inline unsigned int scsi_sgtable_index(unsigned short nents)
577 {
578         unsigned int index;
579
580         BUG_ON(nents > SCSI_MAX_SG_SEGMENTS);
581
582         if (nents <= 8)
583                 index = 0;
584         else
585                 index = get_count_order(nents) - 3;
586
587         return index;
588 }
589
590 static void scsi_sg_free(struct scatterlist *sgl, unsigned int nents)
591 {
592         struct scsi_host_sg_pool *sgp;
593
594         sgp = scsi_sg_pools + scsi_sgtable_index(nents);
595         mempool_free(sgl, sgp->pool);
596 }
597
598 static struct scatterlist *scsi_sg_alloc(unsigned int nents, gfp_t gfp_mask)
599 {
600         struct scsi_host_sg_pool *sgp;
601
602         sgp = scsi_sg_pools + scsi_sgtable_index(nents);
603         return mempool_alloc(sgp->pool, gfp_mask);
604 }
605
606 static int scsi_alloc_sgtable(struct scsi_data_buffer *sdb, int nents,
607                               gfp_t gfp_mask)
608 {
609         int ret;
610
611         BUG_ON(!nents);
612
613         ret = __sg_alloc_table(&sdb->table, nents, SCSI_MAX_SG_SEGMENTS,
614                                gfp_mask, scsi_sg_alloc);
615         if (unlikely(ret))
616                 __sg_free_table(&sdb->table, SCSI_MAX_SG_SEGMENTS,
617                                 scsi_sg_free);
618
619         return ret;
620 }
621
622 static void scsi_free_sgtable(struct scsi_data_buffer *sdb)
623 {
624         __sg_free_table(&sdb->table, SCSI_MAX_SG_SEGMENTS, scsi_sg_free);
625 }
626
627 static void __scsi_release_buffers(struct scsi_cmnd *cmd, int do_bidi_check)
628 {
629
630         if (cmd->sdb.table.nents)
631                 scsi_free_sgtable(&cmd->sdb);
632
633         memset(&cmd->sdb, 0, sizeof(cmd->sdb));
634
635         if (do_bidi_check && scsi_bidi_cmnd(cmd)) {
636                 struct scsi_data_buffer *bidi_sdb =
637                         cmd->request->next_rq->special;
638                 scsi_free_sgtable(bidi_sdb);
639                 kmem_cache_free(scsi_sdb_cache, bidi_sdb);
640                 cmd->request->next_rq->special = NULL;
641         }
642
643         if (scsi_prot_sg_count(cmd))
644                 scsi_free_sgtable(cmd->prot_sdb);
645 }
646
647 /*
648  * Function:    scsi_release_buffers()
649  *
650  * Purpose:     Completion processing for block device I/O requests.
651  *
652  * Arguments:   cmd     - command that we are bailing.
653  *
654  * Lock status: Assumed that no lock is held upon entry.
655  *
656  * Returns:     Nothing
657  *
658  * Notes:       In the event that an upper level driver rejects a
659  *              command, we must release resources allocated during
660  *              the __init_io() function.  Primarily this would involve
661  *              the scatter-gather table, and potentially any bounce
662  *              buffers.
663  */
664 void scsi_release_buffers(struct scsi_cmnd *cmd)
665 {
666         __scsi_release_buffers(cmd, 1);
667 }
668 EXPORT_SYMBOL(scsi_release_buffers);
669
670 /*
671  * Function:    scsi_io_completion()
672  *
673  * Purpose:     Completion processing for block device I/O requests.
674  *
675  * Arguments:   cmd   - command that is finished.
676  *
677  * Lock status: Assumed that no lock is held upon entry.
678  *
679  * Returns:     Nothing
680  *
681  * Notes:       This function is matched in terms of capabilities to
682  *              the function that created the scatter-gather list.
683  *              In other words, if there are no bounce buffers
684  *              (the normal case for most drivers), we don't need
685  *              the logic to deal with cleaning up afterwards.
686  *
687  *              We must call scsi_end_request().  This will finish off
688  *              the specified number of sectors.  If we are done, the
689  *              command block will be released and the queue function
690  *              will be goosed.  If we are not done then we have to
691  *              figure out what to do next:
692  *
693  *              a) We can call scsi_requeue_command().  The request
694  *                 will be unprepared and put back on the queue.  Then
695  *                 a new command will be created for it.  This should
696  *                 be used if we made forward progress, or if we want
697  *                 to switch from READ(10) to READ(6) for example.
698  *
699  *              b) We can call scsi_queue_insert().  The request will
700  *                 be put back on the queue and retried using the same
701  *                 command as before, possibly after a delay.
702  *
703  *              c) We can call blk_end_request() with -EIO to fail
704  *                 the remainder of the request.
705  */
706 void scsi_io_completion(struct scsi_cmnd *cmd, unsigned int good_bytes)
707 {
708         int result = cmd->result;
709         struct request_queue *q = cmd->device->request_queue;
710         struct request *req = cmd->request;
711         int error = 0;
712         struct scsi_sense_hdr sshdr;
713         int sense_valid = 0;
714         int sense_deferred = 0;
715         enum {ACTION_FAIL, ACTION_REPREP, ACTION_RETRY,
716               ACTION_DELAYED_RETRY} action;
717         char *description = NULL;
718
719         if (result) {
720                 sense_valid = scsi_command_normalize_sense(cmd, &sshdr);
721                 if (sense_valid)
722                         sense_deferred = scsi_sense_is_deferred(&sshdr);
723         }
724
725         if (blk_pc_request(req)) { /* SG_IO ioctl from block level */
726                 req->errors = result;
727                 if (result) {
728                         if (sense_valid && req->sense) {
729                                 /*
730                                  * SG_IO wants current and deferred errors
731                                  */
732                                 int len = 8 + cmd->sense_buffer[7];
733
734                                 if (len > SCSI_SENSE_BUFFERSIZE)
735                                         len = SCSI_SENSE_BUFFERSIZE;
736                                 memcpy(req->sense, cmd->sense_buffer,  len);
737                                 req->sense_len = len;
738                         }
739                         if (!sense_deferred)
740                                 error = -EIO;
741                 }
742
743                 req->resid_len = scsi_get_resid(cmd);
744
745                 if (scsi_bidi_cmnd(cmd)) {
746                         /*
747                          * Bidi commands Must be complete as a whole,
748                          * both sides at once.
749                          */
750                         req->next_rq->resid_len = scsi_in(cmd)->resid;
751
752                         blk_end_request_all(req, 0);
753
754                         scsi_release_buffers(cmd);
755                         scsi_next_command(cmd);
756                         return;
757                 }
758         }
759
760         BUG_ON(blk_bidi_rq(req)); /* bidi not support for !blk_pc_request yet */
761
762         /*
763          * Next deal with any sectors which we were able to correctly
764          * handle.
765          */
766         SCSI_LOG_HLCOMPLETE(1, printk("%u sectors total, "
767                                       "%d bytes done.\n",
768                                       blk_rq_sectors(req), good_bytes));
769
770         /*
771          * Recovered errors need reporting, but they're always treated
772          * as success, so fiddle the result code here.  For BLOCK_PC
773          * we already took a copy of the original into rq->errors which
774          * is what gets returned to the user
775          */
776         if (sense_valid && sshdr.sense_key == RECOVERED_ERROR) {
777                 if (!(req->cmd_flags & REQ_QUIET))
778                         scsi_print_sense("", cmd);
779                 result = 0;
780                 /* BLOCK_PC may have set error */
781                 error = 0;
782         }
783
784         /*
785          * A number of bytes were successfully read.  If there
786          * are leftovers and there is some kind of error
787          * (result != 0), retry the rest.
788          */
789         if (scsi_end_request(cmd, error, good_bytes, result == 0) == NULL)
790                 return;
791
792         error = -EIO;
793
794         if (host_byte(result) == DID_RESET) {
795                 /* Third party bus reset or reset for error recovery
796                  * reasons.  Just retry the command and see what
797                  * happens.
798                  */
799                 action = ACTION_RETRY;
800         } else if (sense_valid && !sense_deferred) {
801                 switch (sshdr.sense_key) {
802                 case UNIT_ATTENTION:
803                         if (cmd->device->removable) {
804                                 /* Detected disc change.  Set a bit
805                                  * and quietly refuse further access.
806                                  */
807                                 cmd->device->changed = 1;
808                                 description = "Media Changed";
809                                 action = ACTION_FAIL;
810                         } else {
811                                 /* Must have been a power glitch, or a
812                                  * bus reset.  Could not have been a
813                                  * media change, so we just retry the
814                                  * command and see what happens.
815                                  */
816                                 action = ACTION_RETRY;
817                         }
818                         break;
819                 case ILLEGAL_REQUEST:
820                         /* If we had an ILLEGAL REQUEST returned, then
821                          * we may have performed an unsupported
822                          * command.  The only thing this should be
823                          * would be a ten byte read where only a six
824                          * byte read was supported.  Also, on a system
825                          * where READ CAPACITY failed, we may have
826                          * read past the end of the disk.
827                          */
828                         if ((cmd->device->use_10_for_rw &&
829                             sshdr.asc == 0x20 && sshdr.ascq == 0x00) &&
830                             (cmd->cmnd[0] == READ_10 ||
831                              cmd->cmnd[0] == WRITE_10)) {
832                                 /* This will issue a new 6-byte command. */
833                                 cmd->device->use_10_for_rw = 0;
834                                 action = ACTION_REPREP;
835                         } else if (sshdr.asc == 0x10) /* DIX */ {
836                                 description = "Host Data Integrity Failure";
837                                 action = ACTION_FAIL;
838                                 error = -EILSEQ;
839                         } else
840                                 action = ACTION_FAIL;
841                         break;
842                 case ABORTED_COMMAND:
843                         action = ACTION_FAIL;
844                         if (sshdr.asc == 0x10) { /* DIF */
845                                 description = "Target Data Integrity Failure";
846                                 error = -EILSEQ;
847                         }
848                         break;
849                 case NOT_READY:
850                         /* If the device is in the process of becoming
851                          * ready, or has a temporary blockage, retry.
852                          */
853                         if (sshdr.asc == 0x04) {
854                                 switch (sshdr.ascq) {
855                                 case 0x01: /* becoming ready */
856                                 case 0x04: /* format in progress */
857                                 case 0x05: /* rebuild in progress */
858                                 case 0x06: /* recalculation in progress */
859                                 case 0x07: /* operation in progress */
860                                 case 0x08: /* Long write in progress */
861                                 case 0x09: /* self test in progress */
862                                         action = ACTION_DELAYED_RETRY;
863                                         break;
864                                 default:
865                                         description = "Device not ready";
866                                         action = ACTION_FAIL;
867                                         break;
868                                 }
869                         } else {
870                                 description = "Device not ready";
871                                 action = ACTION_FAIL;
872                         }
873                         break;
874                 case VOLUME_OVERFLOW:
875                         /* See SSC3rXX or current. */
876                         action = ACTION_FAIL;
877                         break;
878                 default:
879                         description = "Unhandled sense code";
880                         action = ACTION_FAIL;
881                         break;
882                 }
883         } else {
884                 description = "Unhandled error code";
885                 action = ACTION_FAIL;
886         }
887
888         switch (action) {
889         case ACTION_FAIL:
890                 /* Give up and fail the remainder of the request */
891                 scsi_release_buffers(cmd);
892                 if (!(req->cmd_flags & REQ_QUIET)) {
893                         if (description)
894                                 scmd_printk(KERN_INFO, cmd, "%s\n",
895                                             description);
896                         scsi_print_result(cmd);
897                         if (driver_byte(result) & DRIVER_SENSE)
898                                 scsi_print_sense("", cmd);
899                 }
900                 if (blk_end_request_err(req, -EIO))
901                         scsi_requeue_command(q, cmd);
902                 else
903                         scsi_next_command(cmd);
904                 break;
905         case ACTION_REPREP:
906                 /* Unprep the request and put it back at the head of the queue.
907                  * A new command will be prepared and issued.
908                  */
909                 scsi_release_buffers(cmd);
910                 scsi_requeue_command(q, cmd);
911                 break;
912         case ACTION_RETRY:
913                 /* Retry the same command immediately */
914                 __scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY, 0);
915                 break;
916         case ACTION_DELAYED_RETRY:
917                 /* Retry the same command after a delay */
918                 __scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY, 0);
919                 break;
920         }
921 }
922
923 static int scsi_init_sgtable(struct request *req, struct scsi_data_buffer *sdb,
924                              gfp_t gfp_mask)
925 {
926         int count;
927
928         /*
929          * If sg table allocation fails, requeue request later.
930          */
931         if (unlikely(scsi_alloc_sgtable(sdb, req->nr_phys_segments,
932                                         gfp_mask))) {
933                 return BLKPREP_DEFER;
934         }
935
936         req->buffer = NULL;
937
938         /* 
939          * Next, walk the list, and fill in the addresses and sizes of
940          * each segment.
941          */
942         count = blk_rq_map_sg(req->q, req, sdb->table.sgl);
943         BUG_ON(count > sdb->table.nents);
944         sdb->table.nents = count;
945         sdb->length = blk_rq_bytes(req);
946         return BLKPREP_OK;
947 }
948
949 /*
950  * Function:    scsi_init_io()
951  *
952  * Purpose:     SCSI I/O initialize function.
953  *
954  * Arguments:   cmd   - Command descriptor we wish to initialize
955  *
956  * Returns:     0 on success
957  *              BLKPREP_DEFER if the failure is retryable
958  *              BLKPREP_KILL if the failure is fatal
959  */
960 int scsi_init_io(struct scsi_cmnd *cmd, gfp_t gfp_mask)
961 {
962         int error = scsi_init_sgtable(cmd->request, &cmd->sdb, gfp_mask);
963         if (error)
964                 goto err_exit;
965
966         if (blk_bidi_rq(cmd->request)) {
967                 struct scsi_data_buffer *bidi_sdb = kmem_cache_zalloc(
968                         scsi_sdb_cache, GFP_ATOMIC);
969                 if (!bidi_sdb) {
970                         error = BLKPREP_DEFER;
971                         goto err_exit;
972                 }
973
974                 cmd->request->next_rq->special = bidi_sdb;
975                 error = scsi_init_sgtable(cmd->request->next_rq, bidi_sdb,
976                                                                     GFP_ATOMIC);
977                 if (error)
978                         goto err_exit;
979         }
980
981         if (blk_integrity_rq(cmd->request)) {
982                 struct scsi_data_buffer *prot_sdb = cmd->prot_sdb;
983                 int ivecs, count;
984
985                 BUG_ON(prot_sdb == NULL);
986                 ivecs = blk_rq_count_integrity_sg(cmd->request);
987
988                 if (scsi_alloc_sgtable(prot_sdb, ivecs, gfp_mask)) {
989                         error = BLKPREP_DEFER;
990                         goto err_exit;
991                 }
992
993                 count = blk_rq_map_integrity_sg(cmd->request,
994                                                 prot_sdb->table.sgl);
995                 BUG_ON(unlikely(count > ivecs));
996
997                 cmd->prot_sdb = prot_sdb;
998                 cmd->prot_sdb->table.nents = count;
999         }
1000
1001         return BLKPREP_OK ;
1002
1003 err_exit:
1004         scsi_release_buffers(cmd);
1005         if (error == BLKPREP_KILL)
1006                 scsi_put_command(cmd);
1007         else /* BLKPREP_DEFER */
1008                 scsi_unprep_request(cmd->request);
1009
1010         return error;
1011 }
1012 EXPORT_SYMBOL(scsi_init_io);
1013
1014 static struct scsi_cmnd *scsi_get_cmd_from_req(struct scsi_device *sdev,
1015                 struct request *req)
1016 {
1017         struct scsi_cmnd *cmd;
1018
1019         if (!req->special) {
1020                 cmd = scsi_get_command(sdev, GFP_ATOMIC);
1021                 if (unlikely(!cmd))
1022                         return NULL;
1023                 req->special = cmd;
1024         } else {
1025                 cmd = req->special;
1026         }
1027
1028         /* pull a tag out of the request if we have one */
1029         cmd->tag = req->tag;
1030         cmd->request = req;
1031
1032         cmd->cmnd = req->cmd;
1033
1034         return cmd;
1035 }
1036
1037 int scsi_setup_blk_pc_cmnd(struct scsi_device *sdev, struct request *req)
1038 {
1039         struct scsi_cmnd *cmd;
1040         int ret = scsi_prep_state_check(sdev, req);
1041
1042         if (ret != BLKPREP_OK)
1043                 return ret;
1044
1045         cmd = scsi_get_cmd_from_req(sdev, req);
1046         if (unlikely(!cmd))
1047                 return BLKPREP_DEFER;
1048
1049         /*
1050          * BLOCK_PC requests may transfer data, in which case they must
1051          * a bio attached to them.  Or they might contain a SCSI command
1052          * that does not transfer data, in which case they may optionally
1053          * submit a request without an attached bio.
1054          */
1055         if (req->bio) {
1056                 int ret;
1057
1058                 BUG_ON(!req->nr_phys_segments);
1059
1060                 ret = scsi_init_io(cmd, GFP_ATOMIC);
1061                 if (unlikely(ret))
1062                         return ret;
1063         } else {
1064                 BUG_ON(blk_rq_bytes(req));
1065
1066                 memset(&cmd->sdb, 0, sizeof(cmd->sdb));
1067                 req->buffer = NULL;
1068         }
1069
1070         cmd->cmd_len = req->cmd_len;
1071         if (!blk_rq_bytes(req))
1072                 cmd->sc_data_direction = DMA_NONE;
1073         else if (rq_data_dir(req) == WRITE)
1074                 cmd->sc_data_direction = DMA_TO_DEVICE;
1075         else
1076                 cmd->sc_data_direction = DMA_FROM_DEVICE;
1077         
1078         cmd->transfersize = blk_rq_bytes(req);
1079         cmd->allowed = req->retries;
1080         return BLKPREP_OK;
1081 }
1082 EXPORT_SYMBOL(scsi_setup_blk_pc_cmnd);
1083
1084 /*
1085  * Setup a REQ_TYPE_FS command.  These are simple read/write request
1086  * from filesystems that still need to be translated to SCSI CDBs from
1087  * the ULD.
1088  */
1089 int scsi_setup_fs_cmnd(struct scsi_device *sdev, struct request *req)
1090 {
1091         struct scsi_cmnd *cmd;
1092         int ret = scsi_prep_state_check(sdev, req);
1093
1094         if (ret != BLKPREP_OK)
1095                 return ret;
1096
1097         if (unlikely(sdev->scsi_dh_data && sdev->scsi_dh_data->scsi_dh
1098                          && sdev->scsi_dh_data->scsi_dh->prep_fn)) {
1099                 ret = sdev->scsi_dh_data->scsi_dh->prep_fn(sdev, req);
1100                 if (ret != BLKPREP_OK)
1101                         return ret;
1102         }
1103
1104         /*
1105          * Filesystem requests must transfer data.
1106          */
1107         BUG_ON(!req->nr_phys_segments);
1108
1109         cmd = scsi_get_cmd_from_req(sdev, req);
1110         if (unlikely(!cmd))
1111                 return BLKPREP_DEFER;
1112
1113         memset(cmd->cmnd, 0, BLK_MAX_CDB);
1114         return scsi_init_io(cmd, GFP_ATOMIC);
1115 }
1116 EXPORT_SYMBOL(scsi_setup_fs_cmnd);
1117
1118 int scsi_prep_state_check(struct scsi_device *sdev, struct request *req)
1119 {
1120         int ret = BLKPREP_OK;
1121
1122         /*
1123          * If the device is not in running state we will reject some
1124          * or all commands.
1125          */
1126         if (unlikely(sdev->sdev_state != SDEV_RUNNING)) {
1127                 switch (sdev->sdev_state) {
1128                 case SDEV_OFFLINE:
1129                         /*
1130                          * If the device is offline we refuse to process any
1131                          * commands.  The device must be brought online
1132                          * before trying any recovery commands.
1133                          */
1134                         sdev_printk(KERN_ERR, sdev,
1135                                     "rejecting I/O to offline device\n");
1136                         ret = BLKPREP_KILL;
1137                         break;
1138                 case SDEV_DEL:
1139                         /*
1140                          * If the device is fully deleted, we refuse to
1141                          * process any commands as well.
1142                          */
1143                         sdev_printk(KERN_ERR, sdev,
1144                                     "rejecting I/O to dead device\n");
1145                         ret = BLKPREP_KILL;
1146                         break;
1147                 case SDEV_QUIESCE:
1148                 case SDEV_BLOCK:
1149                 case SDEV_CREATED_BLOCK:
1150                         /*
1151                          * If the devices is blocked we defer normal commands.
1152                          */
1153                         if (!(req->cmd_flags & REQ_PREEMPT))
1154                                 ret = BLKPREP_DEFER;
1155                         break;
1156                 default:
1157                         /*
1158                          * For any other not fully online state we only allow
1159                          * special commands.  In particular any user initiated
1160                          * command is not allowed.
1161                          */
1162                         if (!(req->cmd_flags & REQ_PREEMPT))
1163                                 ret = BLKPREP_KILL;
1164                         break;
1165                 }
1166         }
1167         return ret;
1168 }
1169 EXPORT_SYMBOL(scsi_prep_state_check);
1170
1171 int scsi_prep_return(struct request_queue *q, struct request *req, int ret)
1172 {
1173         struct scsi_device *sdev = q->queuedata;
1174
1175         switch (ret) {
1176         case BLKPREP_KILL:
1177                 req->errors = DID_NO_CONNECT << 16;
1178                 /* release the command and kill it */
1179                 if (req->special) {
1180                         struct scsi_cmnd *cmd = req->special;
1181                         scsi_release_buffers(cmd);
1182                         scsi_put_command(cmd);
1183                         req->special = NULL;
1184                 }
1185                 break;
1186         case BLKPREP_DEFER:
1187                 /*
1188                  * If we defer, the blk_peek_request() returns NULL, but the
1189                  * queue must be restarted, so we plug here if no returning
1190                  * command will automatically do that.
1191                  */
1192                 if (sdev->device_busy == 0)
1193                         blk_plug_device(q);
1194                 break;
1195         default:
1196                 req->cmd_flags |= REQ_DONTPREP;
1197         }
1198
1199         return ret;
1200 }
1201 EXPORT_SYMBOL(scsi_prep_return);
1202
1203 int scsi_prep_fn(struct request_queue *q, struct request *req)
1204 {
1205         struct scsi_device *sdev = q->queuedata;
1206         int ret = BLKPREP_KILL;
1207
1208         if (req->cmd_type == REQ_TYPE_BLOCK_PC)
1209                 ret = scsi_setup_blk_pc_cmnd(sdev, req);
1210         return scsi_prep_return(q, req, ret);
1211 }
1212 EXPORT_SYMBOL(scsi_prep_fn);
1213
1214 /*
1215  * scsi_dev_queue_ready: if we can send requests to sdev, return 1 else
1216  * return 0.
1217  *
1218  * Called with the queue_lock held.
1219  */
1220 static inline int scsi_dev_queue_ready(struct request_queue *q,
1221                                   struct scsi_device *sdev)
1222 {
1223         if (sdev->device_busy == 0 && sdev->device_blocked) {
1224                 /*
1225                  * unblock after device_blocked iterates to zero
1226                  */
1227                 if (--sdev->device_blocked == 0) {
1228                         SCSI_LOG_MLQUEUE(3,
1229                                    sdev_printk(KERN_INFO, sdev,
1230                                    "unblocking device at zero depth\n"));
1231                 } else {
1232                         blk_plug_device(q);
1233                         return 0;
1234                 }
1235         }
1236         if (scsi_device_is_busy(sdev))
1237                 return 0;
1238
1239         return 1;
1240 }
1241
1242
1243 /*
1244  * scsi_target_queue_ready: checks if there we can send commands to target
1245  * @sdev: scsi device on starget to check.
1246  *
1247  * Called with the host lock held.
1248  */
1249 static inline int scsi_target_queue_ready(struct Scsi_Host *shost,
1250                                            struct scsi_device *sdev)
1251 {
1252         struct scsi_target *starget = scsi_target(sdev);
1253
1254         if (starget->single_lun) {
1255                 if (starget->starget_sdev_user &&
1256                     starget->starget_sdev_user != sdev)
1257                         return 0;
1258                 starget->starget_sdev_user = sdev;
1259         }
1260
1261         if (starget->target_busy == 0 && starget->target_blocked) {
1262                 /*
1263                  * unblock after target_blocked iterates to zero
1264                  */
1265                 if (--starget->target_blocked == 0) {
1266                         SCSI_LOG_MLQUEUE(3, starget_printk(KERN_INFO, starget,
1267                                          "unblocking target at zero depth\n"));
1268                 } else
1269                         return 0;
1270         }
1271
1272         if (scsi_target_is_busy(starget)) {
1273                 if (list_empty(&sdev->starved_entry)) {
1274                         list_add_tail(&sdev->starved_entry,
1275                                       &shost->starved_list);
1276                         return 0;
1277                 }
1278         }
1279
1280         /* We're OK to process the command, so we can't be starved */
1281         if (!list_empty(&sdev->starved_entry))
1282                 list_del_init(&sdev->starved_entry);
1283         return 1;
1284 }
1285
1286 /*
1287  * scsi_host_queue_ready: if we can send requests to shost, return 1 else
1288  * return 0. We must end up running the queue again whenever 0 is
1289  * returned, else IO can hang.
1290  *
1291  * Called with host_lock held.
1292  */
1293 static inline int scsi_host_queue_ready(struct request_queue *q,
1294                                    struct Scsi_Host *shost,
1295                                    struct scsi_device *sdev)
1296 {
1297         if (scsi_host_in_recovery(shost))
1298                 return 0;
1299         if (shost->host_busy == 0 && shost->host_blocked) {
1300                 /*
1301                  * unblock after host_blocked iterates to zero
1302                  */
1303                 if (--shost->host_blocked == 0) {
1304                         SCSI_LOG_MLQUEUE(3,
1305                                 printk("scsi%d unblocking host at zero depth\n",
1306                                         shost->host_no));
1307                 } else {
1308                         return 0;
1309                 }
1310         }
1311         if (scsi_host_is_busy(shost)) {
1312                 if (list_empty(&sdev->starved_entry))
1313                         list_add_tail(&sdev->starved_entry, &shost->starved_list);
1314                 return 0;
1315         }
1316
1317         /* We're OK to process the command, so we can't be starved */
1318         if (!list_empty(&sdev->starved_entry))
1319                 list_del_init(&sdev->starved_entry);
1320
1321         return 1;
1322 }
1323
1324 /*
1325  * Busy state exporting function for request stacking drivers.
1326  *
1327  * For efficiency, no lock is taken to check the busy state of
1328  * shost/starget/sdev, since the returned value is not guaranteed and
1329  * may be changed after request stacking drivers call the function,
1330  * regardless of taking lock or not.
1331  *
1332  * When scsi can't dispatch I/Os anymore and needs to kill I/Os
1333  * (e.g. !sdev), scsi needs to return 'not busy'.
1334  * Otherwise, request stacking drivers may hold requests forever.
1335  */
1336 static int scsi_lld_busy(struct request_queue *q)
1337 {
1338         struct scsi_device *sdev = q->queuedata;
1339         struct Scsi_Host *shost;
1340         struct scsi_target *starget;
1341
1342         if (!sdev)
1343                 return 0;
1344
1345         shost = sdev->host;
1346         starget = scsi_target(sdev);
1347
1348         if (scsi_host_in_recovery(shost) || scsi_host_is_busy(shost) ||
1349             scsi_target_is_busy(starget) || scsi_device_is_busy(sdev))
1350                 return 1;
1351
1352         return 0;
1353 }
1354
1355 /*
1356  * Kill a request for a dead device
1357  */
1358 static void scsi_kill_request(struct request *req, struct request_queue *q)
1359 {
1360         struct scsi_cmnd *cmd = req->special;
1361         struct scsi_device *sdev = cmd->device;
1362         struct scsi_target *starget = scsi_target(sdev);
1363         struct Scsi_Host *shost = sdev->host;
1364
1365         blk_start_request(req);
1366
1367         if (unlikely(cmd == NULL)) {
1368                 printk(KERN_CRIT "impossible request in %s.\n",
1369                                  __func__);
1370                 BUG();
1371         }
1372
1373         scsi_init_cmd_errh(cmd);
1374         cmd->result = DID_NO_CONNECT << 16;
1375         atomic_inc(&cmd->device->iorequest_cnt);
1376
1377         /*
1378          * SCSI request completion path will do scsi_device_unbusy(),
1379          * bump busy counts.  To bump the counters, we need to dance
1380          * with the locks as normal issue path does.
1381          */
1382         sdev->device_busy++;
1383         spin_unlock(sdev->request_queue->queue_lock);
1384         spin_lock(shost->host_lock);
1385         shost->host_busy++;
1386         starget->target_busy++;
1387         spin_unlock(shost->host_lock);
1388         spin_lock(sdev->request_queue->queue_lock);
1389
1390         blk_complete_request(req);
1391 }
1392
1393 static void scsi_softirq_done(struct request *rq)
1394 {
1395         struct scsi_cmnd *cmd = rq->special;
1396         unsigned long wait_for = (cmd->allowed + 1) * rq->timeout;
1397         int disposition;
1398
1399         INIT_LIST_HEAD(&cmd->eh_entry);
1400
1401         /*
1402          * Set the serial numbers back to zero
1403          */
1404         cmd->serial_number = 0;
1405
1406         atomic_inc(&cmd->device->iodone_cnt);
1407         if (cmd->result)
1408                 atomic_inc(&cmd->device->ioerr_cnt);
1409
1410         disposition = scsi_decide_disposition(cmd);
1411         if (disposition != SUCCESS &&
1412             time_before(cmd->jiffies_at_alloc + wait_for, jiffies)) {
1413                 sdev_printk(KERN_ERR, cmd->device,
1414                             "timing out command, waited %lus\n",
1415                             wait_for/HZ);
1416                 disposition = SUCCESS;
1417         }
1418                         
1419         scsi_log_completion(cmd, disposition);
1420
1421         switch (disposition) {
1422                 case SUCCESS:
1423                         scsi_finish_command(cmd);
1424                         break;
1425                 case NEEDS_RETRY:
1426                         scsi_queue_insert(cmd, SCSI_MLQUEUE_EH_RETRY);
1427                         break;
1428                 case ADD_TO_MLQUEUE:
1429                         scsi_queue_insert(cmd, SCSI_MLQUEUE_DEVICE_BUSY);
1430                         break;
1431                 default:
1432                         if (!scsi_eh_scmd_add(cmd, 0))
1433                                 scsi_finish_command(cmd);
1434         }
1435 }
1436
1437 /*
1438  * Function:    scsi_request_fn()
1439  *
1440  * Purpose:     Main strategy routine for SCSI.
1441  *
1442  * Arguments:   q       - Pointer to actual queue.
1443  *
1444  * Returns:     Nothing
1445  *
1446  * Lock status: IO request lock assumed to be held when called.
1447  */
1448 static void scsi_request_fn(struct request_queue *q)
1449 {
1450         struct scsi_device *sdev = q->queuedata;
1451         struct Scsi_Host *shost;
1452         struct scsi_cmnd *cmd;
1453         struct request *req;
1454
1455         if (!sdev) {
1456                 printk("scsi: killing requests for dead queue\n");
1457                 while ((req = blk_peek_request(q)) != NULL)
1458                         scsi_kill_request(req, q);
1459                 return;
1460         }
1461
1462         if(!get_device(&sdev->sdev_gendev))
1463                 /* We must be tearing the block queue down already */
1464                 return;
1465
1466         /*
1467          * To start with, we keep looping until the queue is empty, or until
1468          * the host is no longer able to accept any more requests.
1469          */
1470         shost = sdev->host;
1471         while (!blk_queue_plugged(q)) {
1472                 int rtn;
1473                 /*
1474                  * get next queueable request.  We do this early to make sure
1475                  * that the request is fully prepared even if we cannot 
1476                  * accept it.
1477                  */
1478                 req = blk_peek_request(q);
1479                 if (!req || !scsi_dev_queue_ready(q, sdev))
1480                         break;
1481
1482                 if (unlikely(!scsi_device_online(sdev))) {
1483                         sdev_printk(KERN_ERR, sdev,
1484                                     "rejecting I/O to offline device\n");
1485                         scsi_kill_request(req, q);
1486                         continue;
1487                 }
1488
1489
1490                 /*
1491                  * Remove the request from the request list.
1492                  */
1493                 if (!(blk_queue_tagged(q) && !blk_queue_start_tag(q, req)))
1494                         blk_start_request(req);
1495                 sdev->device_busy++;
1496
1497                 spin_unlock(q->queue_lock);
1498                 cmd = req->special;
1499                 if (unlikely(cmd == NULL)) {
1500                         printk(KERN_CRIT "impossible request in %s.\n"
1501                                          "please mail a stack trace to "
1502                                          "linux-scsi@vger.kernel.org\n",
1503                                          __func__);
1504                         blk_dump_rq_flags(req, "foo");
1505                         BUG();
1506                 }
1507                 spin_lock(shost->host_lock);
1508
1509                 /*
1510                  * We hit this when the driver is using a host wide
1511                  * tag map. For device level tag maps the queue_depth check
1512                  * in the device ready fn would prevent us from trying
1513                  * to allocate a tag. Since the map is a shared host resource
1514                  * we add the dev to the starved list so it eventually gets
1515                  * a run when a tag is freed.
1516                  */
1517                 if (blk_queue_tagged(q) && !blk_rq_tagged(req)) {
1518                         if (list_empty(&sdev->starved_entry))
1519                                 list_add_tail(&sdev->starved_entry,
1520                                               &shost->starved_list);
1521                         goto not_ready;
1522                 }
1523
1524                 if (!scsi_target_queue_ready(shost, sdev))
1525                         goto not_ready;
1526
1527                 if (!scsi_host_queue_ready(q, shost, sdev))
1528                         goto not_ready;
1529
1530                 scsi_target(sdev)->target_busy++;
1531                 shost->host_busy++;
1532
1533                 /*
1534                  * XXX(hch): This is rather suboptimal, scsi_dispatch_cmd will
1535                  *              take the lock again.
1536                  */
1537                 spin_unlock_irq(shost->host_lock);
1538
1539                 /*
1540                  * Finally, initialize any error handling parameters, and set up
1541                  * the timers for timeouts.
1542                  */
1543                 scsi_init_cmd_errh(cmd);
1544
1545                 /*
1546                  * Dispatch the command to the low-level driver.
1547                  */
1548                 rtn = scsi_dispatch_cmd(cmd);
1549                 spin_lock_irq(q->queue_lock);
1550                 if(rtn) {
1551                         /* we're refusing the command; because of
1552                          * the way locks get dropped, we need to 
1553                          * check here if plugging is required */
1554                         if(sdev->device_busy == 0)
1555                                 blk_plug_device(q);
1556
1557                         break;
1558                 }
1559         }
1560
1561         goto out;
1562
1563  not_ready:
1564         spin_unlock_irq(shost->host_lock);
1565
1566         /*
1567          * lock q, handle tag, requeue req, and decrement device_busy. We
1568          * must return with queue_lock held.
1569          *
1570          * Decrementing device_busy without checking it is OK, as all such
1571          * cases (host limits or settings) should run the queue at some
1572          * later time.
1573          */
1574         spin_lock_irq(q->queue_lock);
1575         blk_requeue_request(q, req);
1576         sdev->device_busy--;
1577         if(sdev->device_busy == 0)
1578                 blk_plug_device(q);
1579  out:
1580         /* must be careful here...if we trigger the ->remove() function
1581          * we cannot be holding the q lock */
1582         spin_unlock_irq(q->queue_lock);
1583         put_device(&sdev->sdev_gendev);
1584         spin_lock_irq(q->queue_lock);
1585 }
1586
1587 u64 scsi_calculate_bounce_limit(struct Scsi_Host *shost)
1588 {
1589         struct device *host_dev;
1590         u64 bounce_limit = 0xffffffff;
1591
1592         if (shost->unchecked_isa_dma)
1593                 return BLK_BOUNCE_ISA;
1594         /*
1595          * Platforms with virtual-DMA translation
1596          * hardware have no practical limit.
1597          */
1598         if (!PCI_DMA_BUS_IS_PHYS)
1599                 return BLK_BOUNCE_ANY;
1600
1601         host_dev = scsi_get_device(shost);
1602         if (host_dev && host_dev->dma_mask)
1603                 bounce_limit = *host_dev->dma_mask;
1604
1605         return bounce_limit;
1606 }
1607 EXPORT_SYMBOL(scsi_calculate_bounce_limit);
1608
1609 struct request_queue *__scsi_alloc_queue(struct Scsi_Host *shost,
1610                                          request_fn_proc *request_fn)
1611 {
1612         struct request_queue *q;
1613         struct device *dev = shost->shost_gendev.parent;
1614
1615         q = blk_init_queue(request_fn, NULL);
1616         if (!q)
1617                 return NULL;
1618
1619         /*
1620          * this limit is imposed by hardware restrictions
1621          */
1622         blk_queue_max_hw_segments(q, shost->sg_tablesize);
1623         blk_queue_max_phys_segments(q, SCSI_MAX_SG_CHAIN_SEGMENTS);
1624
1625         blk_queue_max_sectors(q, shost->max_sectors);
1626         blk_queue_bounce_limit(q, scsi_calculate_bounce_limit(shost));
1627         blk_queue_segment_boundary(q, shost->dma_boundary);
1628         dma_set_seg_boundary(dev, shost->dma_boundary);
1629
1630         blk_queue_max_segment_size(q, dma_get_max_seg_size(dev));
1631
1632         /* New queue, no concurrency on queue_flags */
1633         if (!shost->use_clustering)
1634                 queue_flag_clear_unlocked(QUEUE_FLAG_CLUSTER, q);
1635
1636         /*
1637          * set a reasonable default alignment on word boundaries: the
1638          * host and device may alter it using
1639          * blk_queue_update_dma_alignment() later.
1640          */
1641         blk_queue_dma_alignment(q, 0x03);
1642
1643         return q;
1644 }
1645 EXPORT_SYMBOL(__scsi_alloc_queue);
1646
1647 struct request_queue *scsi_alloc_queue(struct scsi_device *sdev)
1648 {
1649         struct request_queue *q;
1650
1651         q = __scsi_alloc_queue(sdev->host, scsi_request_fn);
1652         if (!q)
1653                 return NULL;
1654
1655         blk_queue_prep_rq(q, scsi_prep_fn);
1656         blk_queue_softirq_done(q, scsi_softirq_done);
1657         blk_queue_rq_timed_out(q, scsi_times_out);
1658         blk_queue_lld_busy(q, scsi_lld_busy);
1659         return q;
1660 }
1661
1662 void scsi_free_queue(struct request_queue *q)
1663 {
1664         blk_cleanup_queue(q);
1665 }
1666
1667 /*
1668  * Function:    scsi_block_requests()
1669  *
1670  * Purpose:     Utility function used by low-level drivers to prevent further
1671  *              commands from being queued to the device.
1672  *
1673  * Arguments:   shost       - Host in question
1674  *
1675  * Returns:     Nothing
1676  *
1677  * Lock status: No locks are assumed held.
1678  *
1679  * Notes:       There is no timer nor any other means by which the requests
1680  *              get unblocked other than the low-level driver calling
1681  *              scsi_unblock_requests().
1682  */
1683 void scsi_block_requests(struct Scsi_Host *shost)
1684 {
1685         shost->host_self_blocked = 1;
1686 }
1687 EXPORT_SYMBOL(scsi_block_requests);
1688
1689 /*
1690  * Function:    scsi_unblock_requests()
1691  *
1692  * Purpose:     Utility function used by low-level drivers to allow further
1693  *              commands from being queued to the device.
1694  *
1695  * Arguments:   shost       - Host in question
1696  *
1697  * Returns:     Nothing
1698  *
1699  * Lock status: No locks are assumed held.
1700  *
1701  * Notes:       There is no timer nor any other means by which the requests
1702  *              get unblocked other than the low-level driver calling
1703  *              scsi_unblock_requests().
1704  *
1705  *              This is done as an API function so that changes to the
1706  *              internals of the scsi mid-layer won't require wholesale
1707  *              changes to drivers that use this feature.
1708  */
1709 void scsi_unblock_requests(struct Scsi_Host *shost)
1710 {
1711         shost->host_self_blocked = 0;
1712         scsi_run_host_queues(shost);
1713 }
1714 EXPORT_SYMBOL(scsi_unblock_requests);
1715
1716 int __init scsi_init_queue(void)
1717 {
1718         int i;
1719
1720         scsi_sdb_cache = kmem_cache_create("scsi_data_buffer",
1721                                            sizeof(struct scsi_data_buffer),
1722                                            0, 0, NULL);
1723         if (!scsi_sdb_cache) {
1724                 printk(KERN_ERR "SCSI: can't init scsi sdb cache\n");
1725                 return -ENOMEM;
1726         }
1727
1728         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1729                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1730                 int size = sgp->size * sizeof(struct scatterlist);
1731
1732                 sgp->slab = kmem_cache_create(sgp->name, size, 0,
1733                                 SLAB_HWCACHE_ALIGN, NULL);
1734                 if (!sgp->slab) {
1735                         printk(KERN_ERR "SCSI: can't init sg slab %s\n",
1736                                         sgp->name);
1737                         goto cleanup_sdb;
1738                 }
1739
1740                 sgp->pool = mempool_create_slab_pool(SG_MEMPOOL_SIZE,
1741                                                      sgp->slab);
1742                 if (!sgp->pool) {
1743                         printk(KERN_ERR "SCSI: can't init sg mempool %s\n",
1744                                         sgp->name);
1745                         goto cleanup_sdb;
1746                 }
1747         }
1748
1749         return 0;
1750
1751 cleanup_sdb:
1752         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1753                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1754                 if (sgp->pool)
1755                         mempool_destroy(sgp->pool);
1756                 if (sgp->slab)
1757                         kmem_cache_destroy(sgp->slab);
1758         }
1759         kmem_cache_destroy(scsi_sdb_cache);
1760
1761         return -ENOMEM;
1762 }
1763
1764 void scsi_exit_queue(void)
1765 {
1766         int i;
1767
1768         kmem_cache_destroy(scsi_sdb_cache);
1769
1770         for (i = 0; i < SG_MEMPOOL_NR; i++) {
1771                 struct scsi_host_sg_pool *sgp = scsi_sg_pools + i;
1772                 mempool_destroy(sgp->pool);
1773                 kmem_cache_destroy(sgp->slab);
1774         }
1775 }
1776
1777 /**
1778  *      scsi_mode_select - issue a mode select
1779  *      @sdev:  SCSI device to be queried
1780  *      @pf:    Page format bit (1 == standard, 0 == vendor specific)
1781  *      @sp:    Save page bit (0 == don't save, 1 == save)
1782  *      @modepage: mode page being requested
1783  *      @buffer: request buffer (may not be smaller than eight bytes)
1784  *      @len:   length of request buffer.
1785  *      @timeout: command timeout
1786  *      @retries: number of retries before failing
1787  *      @data: returns a structure abstracting the mode header data
1788  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
1789  *              must be SCSI_SENSE_BUFFERSIZE big.
1790  *
1791  *      Returns zero if successful; negative error number or scsi
1792  *      status on error
1793  *
1794  */
1795 int
1796 scsi_mode_select(struct scsi_device *sdev, int pf, int sp, int modepage,
1797                  unsigned char *buffer, int len, int timeout, int retries,
1798                  struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1799 {
1800         unsigned char cmd[10];
1801         unsigned char *real_buffer;
1802         int ret;
1803
1804         memset(cmd, 0, sizeof(cmd));
1805         cmd[1] = (pf ? 0x10 : 0) | (sp ? 0x01 : 0);
1806
1807         if (sdev->use_10_for_ms) {
1808                 if (len > 65535)
1809                         return -EINVAL;
1810                 real_buffer = kmalloc(8 + len, GFP_KERNEL);
1811                 if (!real_buffer)
1812                         return -ENOMEM;
1813                 memcpy(real_buffer + 8, buffer, len);
1814                 len += 8;
1815                 real_buffer[0] = 0;
1816                 real_buffer[1] = 0;
1817                 real_buffer[2] = data->medium_type;
1818                 real_buffer[3] = data->device_specific;
1819                 real_buffer[4] = data->longlba ? 0x01 : 0;
1820                 real_buffer[5] = 0;
1821                 real_buffer[6] = data->block_descriptor_length >> 8;
1822                 real_buffer[7] = data->block_descriptor_length;
1823
1824                 cmd[0] = MODE_SELECT_10;
1825                 cmd[7] = len >> 8;
1826                 cmd[8] = len;
1827         } else {
1828                 if (len > 255 || data->block_descriptor_length > 255 ||
1829                     data->longlba)
1830                         return -EINVAL;
1831
1832                 real_buffer = kmalloc(4 + len, GFP_KERNEL);
1833                 if (!real_buffer)
1834                         return -ENOMEM;
1835                 memcpy(real_buffer + 4, buffer, len);
1836                 len += 4;
1837                 real_buffer[0] = 0;
1838                 real_buffer[1] = data->medium_type;
1839                 real_buffer[2] = data->device_specific;
1840                 real_buffer[3] = data->block_descriptor_length;
1841                 
1842
1843                 cmd[0] = MODE_SELECT;
1844                 cmd[4] = len;
1845         }
1846
1847         ret = scsi_execute_req(sdev, cmd, DMA_TO_DEVICE, real_buffer, len,
1848                                sshdr, timeout, retries, NULL);
1849         kfree(real_buffer);
1850         return ret;
1851 }
1852 EXPORT_SYMBOL_GPL(scsi_mode_select);
1853
1854 /**
1855  *      scsi_mode_sense - issue a mode sense, falling back from 10 to six bytes if necessary.
1856  *      @sdev:  SCSI device to be queried
1857  *      @dbd:   set if mode sense will allow block descriptors to be returned
1858  *      @modepage: mode page being requested
1859  *      @buffer: request buffer (may not be smaller than eight bytes)
1860  *      @len:   length of request buffer.
1861  *      @timeout: command timeout
1862  *      @retries: number of retries before failing
1863  *      @data: returns a structure abstracting the mode header data
1864  *      @sshdr: place to put sense data (or NULL if no sense to be collected).
1865  *              must be SCSI_SENSE_BUFFERSIZE big.
1866  *
1867  *      Returns zero if unsuccessful, or the header offset (either 4
1868  *      or 8 depending on whether a six or ten byte command was
1869  *      issued) if successful.
1870  */
1871 int
1872 scsi_mode_sense(struct scsi_device *sdev, int dbd, int modepage,
1873                   unsigned char *buffer, int len, int timeout, int retries,
1874                   struct scsi_mode_data *data, struct scsi_sense_hdr *sshdr)
1875 {
1876         unsigned char cmd[12];
1877         int use_10_for_ms;
1878         int header_length;
1879         int result;
1880         struct scsi_sense_hdr my_sshdr;
1881
1882         memset(data, 0, sizeof(*data));
1883         memset(&cmd[0], 0, 12);
1884         cmd[1] = dbd & 0x18;    /* allows DBD and LLBA bits */
1885         cmd[2] = modepage;
1886
1887         /* caller might not be interested in sense, but we need it */
1888         if (!sshdr)
1889                 sshdr = &my_sshdr;
1890
1891  retry:
1892         use_10_for_ms = sdev->use_10_for_ms;
1893
1894         if (use_10_for_ms) {
1895                 if (len < 8)
1896                         len = 8;
1897
1898                 cmd[0] = MODE_SENSE_10;
1899                 cmd[8] = len;
1900                 header_length = 8;
1901         } else {
1902                 if (len < 4)
1903                         len = 4;
1904
1905                 cmd[0] = MODE_SENSE;
1906                 cmd[4] = len;
1907                 header_length = 4;
1908         }
1909
1910         memset(buffer, 0, len);
1911
1912         result = scsi_execute_req(sdev, cmd, DMA_FROM_DEVICE, buffer, len,
1913                                   sshdr, timeout, retries, NULL);
1914
1915         /* This code looks awful: what it's doing is making sure an
1916          * ILLEGAL REQUEST sense return identifies the actual command
1917          * byte as the problem.  MODE_SENSE commands can return
1918          * ILLEGAL REQUEST if the code page isn't supported */
1919
1920         if (use_10_for_ms && !scsi_status_is_good(result) &&
1921             (driver_byte(result) & DRIVER_SENSE)) {
1922                 if (scsi_sense_valid(sshdr)) {
1923                         if ((sshdr->sense_key == ILLEGAL_REQUEST) &&
1924                             (sshdr->asc == 0x20) && (sshdr->ascq == 0)) {
1925                                 /* 
1926                                  * Invalid command operation code
1927                                  */
1928                                 sdev->use_10_for_ms = 0;
1929                                 goto retry;
1930                         }
1931                 }
1932         }
1933
1934         if(scsi_status_is_good(result)) {
1935                 if (unlikely(buffer[0] == 0x86 && buffer[1] == 0x0b &&
1936                              (modepage == 6 || modepage == 8))) {
1937                         /* Initio breakage? */
1938                         header_length = 0;
1939                         data->length = 13;
1940                         data->medium_type = 0;
1941                         data->device_specific = 0;
1942                         data->longlba = 0;
1943                         data->block_descriptor_length = 0;
1944                 } else if(use_10_for_ms) {
1945                         data->length = buffer[0]*256 + buffer[1] + 2;
1946                         data->medium_type = buffer[2];
1947                         data->device_specific = buffer[3];
1948                         data->longlba = buffer[4] & 0x01;
1949                         data->block_descriptor_length = buffer[6]*256
1950                                 + buffer[7];
1951                 } else {
1952                         data->length = buffer[0] + 1;
1953                         data->medium_type = buffer[1];
1954                         data->device_specific = buffer[2];
1955                         data->block_descriptor_length = buffer[3];
1956                 }
1957                 data->header_length = header_length;
1958         }
1959
1960         return result;
1961 }
1962 EXPORT_SYMBOL(scsi_mode_sense);
1963
1964 /**
1965  *      scsi_test_unit_ready - test if unit is ready
1966  *      @sdev:  scsi device to change the state of.
1967  *      @timeout: command timeout
1968  *      @retries: number of retries before failing
1969  *      @sshdr_external: Optional pointer to struct scsi_sense_hdr for
1970  *              returning sense. Make sure that this is cleared before passing
1971  *              in.
1972  *
1973  *      Returns zero if unsuccessful or an error if TUR failed.  For
1974  *      removable media, a return of NOT_READY or UNIT_ATTENTION is
1975  *      translated to success, with the ->changed flag updated.
1976  **/
1977 int
1978 scsi_test_unit_ready(struct scsi_device *sdev, int timeout, int retries,
1979                      struct scsi_sense_hdr *sshdr_external)
1980 {
1981         char cmd[] = {
1982                 TEST_UNIT_READY, 0, 0, 0, 0, 0,
1983         };
1984         struct scsi_sense_hdr *sshdr;
1985         int result;
1986
1987         if (!sshdr_external)
1988                 sshdr = kzalloc(sizeof(*sshdr), GFP_KERNEL);
1989         else
1990                 sshdr = sshdr_external;
1991
1992         /* try to eat the UNIT_ATTENTION if there are enough retries */
1993         do {
1994                 result = scsi_execute_req(sdev, cmd, DMA_NONE, NULL, 0, sshdr,
1995                                           timeout, retries, NULL);
1996                 if (sdev->removable && scsi_sense_valid(sshdr) &&
1997                     sshdr->sense_key == UNIT_ATTENTION)
1998                         sdev->changed = 1;
1999         } while (scsi_sense_valid(sshdr) &&
2000                  sshdr->sense_key == UNIT_ATTENTION && --retries);
2001
2002         if (!sshdr)
2003                 /* could not allocate sense buffer, so can't process it */
2004                 return result;
2005
2006         if (sdev->removable && scsi_sense_valid(sshdr) &&
2007             (sshdr->sense_key == UNIT_ATTENTION ||
2008              sshdr->sense_key == NOT_READY)) {
2009                 sdev->changed = 1;
2010                 result = 0;
2011         }
2012         if (!sshdr_external)
2013                 kfree(sshdr);
2014         return result;
2015 }
2016 EXPORT_SYMBOL(scsi_test_unit_ready);
2017
2018 /**
2019  *      scsi_device_set_state - Take the given device through the device state model.
2020  *      @sdev:  scsi device to change the state of.
2021  *      @state: state to change to.
2022  *
2023  *      Returns zero if unsuccessful or an error if the requested 
2024  *      transition is illegal.
2025  */
2026 int
2027 scsi_device_set_state(struct scsi_device *sdev, enum scsi_device_state state)
2028 {
2029         enum scsi_device_state oldstate = sdev->sdev_state;
2030
2031         if (state == oldstate)
2032                 return 0;
2033
2034         switch (state) {
2035         case SDEV_CREATED:
2036                 switch (oldstate) {
2037                 case SDEV_CREATED_BLOCK:
2038                         break;
2039                 default:
2040                         goto illegal;
2041                 }
2042                 break;
2043                         
2044         case SDEV_RUNNING:
2045                 switch (oldstate) {
2046                 case SDEV_CREATED:
2047                 case SDEV_OFFLINE:
2048                 case SDEV_QUIESCE:
2049                 case SDEV_BLOCK:
2050                         break;
2051                 default:
2052                         goto illegal;
2053                 }
2054                 break;
2055
2056         case SDEV_QUIESCE:
2057                 switch (oldstate) {
2058                 case SDEV_RUNNING:
2059                 case SDEV_OFFLINE:
2060                         break;
2061                 default:
2062                         goto illegal;
2063                 }
2064                 break;
2065
2066         case SDEV_OFFLINE:
2067                 switch (oldstate) {
2068                 case SDEV_CREATED:
2069                 case SDEV_RUNNING:
2070                 case SDEV_QUIESCE:
2071                 case SDEV_BLOCK:
2072                         break;
2073                 default:
2074                         goto illegal;
2075                 }
2076                 break;
2077
2078         case SDEV_BLOCK:
2079                 switch (oldstate) {
2080                 case SDEV_RUNNING:
2081                 case SDEV_CREATED_BLOCK:
2082                         break;
2083                 default:
2084                         goto illegal;
2085                 }
2086                 break;
2087
2088         case SDEV_CREATED_BLOCK:
2089                 switch (oldstate) {
2090                 case SDEV_CREATED:
2091                         break;
2092                 default:
2093                         goto illegal;
2094                 }
2095                 break;
2096
2097         case SDEV_CANCEL:
2098                 switch (oldstate) {
2099                 case SDEV_CREATED:
2100                 case SDEV_RUNNING:
2101                 case SDEV_QUIESCE:
2102                 case SDEV_OFFLINE:
2103                 case SDEV_BLOCK:
2104                         break;
2105                 default:
2106                         goto illegal;
2107                 }
2108                 break;
2109
2110         case SDEV_DEL:
2111                 switch (oldstate) {
2112                 case SDEV_CREATED:
2113                 case SDEV_RUNNING:
2114                 case SDEV_OFFLINE:
2115                 case SDEV_CANCEL:
2116                         break;
2117                 default:
2118                         goto illegal;
2119                 }
2120                 break;
2121
2122         }
2123         sdev->sdev_state = state;
2124         return 0;
2125
2126  illegal:
2127         SCSI_LOG_ERROR_RECOVERY(1, 
2128                                 sdev_printk(KERN_ERR, sdev,
2129                                             "Illegal state transition %s->%s\n",
2130                                             scsi_device_state_name(oldstate),
2131                                             scsi_device_state_name(state))
2132                                 );
2133         return -EINVAL;
2134 }
2135 EXPORT_SYMBOL(scsi_device_set_state);
2136
2137 /**
2138  *      sdev_evt_emit - emit a single SCSI device uevent
2139  *      @sdev: associated SCSI device
2140  *      @evt: event to emit
2141  *
2142  *      Send a single uevent (scsi_event) to the associated scsi_device.
2143  */
2144 static void scsi_evt_emit(struct scsi_device *sdev, struct scsi_event *evt)
2145 {
2146         int idx = 0;
2147         char *envp[3];
2148
2149         switch (evt->evt_type) {
2150         case SDEV_EVT_MEDIA_CHANGE:
2151                 envp[idx++] = "SDEV_MEDIA_CHANGE=1";
2152                 break;
2153
2154         default:
2155                 /* do nothing */
2156                 break;
2157         }
2158
2159         envp[idx++] = NULL;
2160
2161         kobject_uevent_env(&sdev->sdev_gendev.kobj, KOBJ_CHANGE, envp);
2162 }
2163
2164 /**
2165  *      sdev_evt_thread - send a uevent for each scsi event
2166  *      @work: work struct for scsi_device
2167  *
2168  *      Dispatch queued events to their associated scsi_device kobjects
2169  *      as uevents.
2170  */
2171 void scsi_evt_thread(struct work_struct *work)
2172 {
2173         struct scsi_device *sdev;
2174         LIST_HEAD(event_list);
2175
2176         sdev = container_of(work, struct scsi_device, event_work);
2177
2178         while (1) {
2179                 struct scsi_event *evt;
2180                 struct list_head *this, *tmp;
2181                 unsigned long flags;
2182
2183                 spin_lock_irqsave(&sdev->list_lock, flags);
2184                 list_splice_init(&sdev->event_list, &event_list);
2185                 spin_unlock_irqrestore(&sdev->list_lock, flags);
2186
2187                 if (list_empty(&event_list))
2188                         break;
2189
2190                 list_for_each_safe(this, tmp, &event_list) {
2191                         evt = list_entry(this, struct scsi_event, node);
2192                         list_del(&evt->node);
2193                         scsi_evt_emit(sdev, evt);
2194                         kfree(evt);
2195                 }
2196         }
2197 }
2198
2199 /**
2200  *      sdev_evt_send - send asserted event to uevent thread
2201  *      @sdev: scsi_device event occurred on
2202  *      @evt: event to send
2203  *
2204  *      Assert scsi device event asynchronously.
2205  */
2206 void sdev_evt_send(struct scsi_device *sdev, struct scsi_event *evt)
2207 {
2208         unsigned long flags;
2209
2210 #if 0
2211         /* FIXME: currently this check eliminates all media change events
2212          * for polled devices.  Need to update to discriminate between AN
2213          * and polled events */
2214         if (!test_bit(evt->evt_type, sdev->supported_events)) {
2215                 kfree(evt);
2216                 return;
2217         }
2218 #endif
2219
2220         spin_lock_irqsave(&sdev->list_lock, flags);
2221         list_add_tail(&evt->node, &sdev->event_list);
2222         schedule_work(&sdev->event_work);
2223         spin_unlock_irqrestore(&sdev->list_lock, flags);
2224 }
2225 EXPORT_SYMBOL_GPL(sdev_evt_send);
2226
2227 /**
2228  *      sdev_evt_alloc - allocate a new scsi event
2229  *      @evt_type: type of event to allocate
2230  *      @gfpflags: GFP flags for allocation
2231  *
2232  *      Allocates and returns a new scsi_event.
2233  */
2234 struct scsi_event *sdev_evt_alloc(enum scsi_device_event evt_type,
2235                                   gfp_t gfpflags)
2236 {
2237         struct scsi_event *evt = kzalloc(sizeof(struct scsi_event), gfpflags);
2238         if (!evt)
2239                 return NULL;
2240
2241         evt->evt_type = evt_type;
2242         INIT_LIST_HEAD(&evt->node);
2243
2244         /* evt_type-specific initialization, if any */
2245         switch (evt_type) {
2246         case SDEV_EVT_MEDIA_CHANGE:
2247         default:
2248                 /* do nothing */
2249                 break;
2250         }
2251
2252         return evt;
2253 }
2254 EXPORT_SYMBOL_GPL(sdev_evt_alloc);
2255
2256 /**
2257  *      sdev_evt_send_simple - send asserted event to uevent thread
2258  *      @sdev: scsi_device event occurred on
2259  *      @evt_type: type of event to send
2260  *      @gfpflags: GFP flags for allocation
2261  *
2262  *      Assert scsi device event asynchronously, given an event type.
2263  */
2264 void sdev_evt_send_simple(struct scsi_device *sdev,
2265                           enum scsi_device_event evt_type, gfp_t gfpflags)
2266 {
2267         struct scsi_event *evt = sdev_evt_alloc(evt_type, gfpflags);
2268         if (!evt) {
2269                 sdev_printk(KERN_ERR, sdev, "event %d eaten due to OOM\n",
2270                             evt_type);
2271                 return;
2272         }
2273
2274         sdev_evt_send(sdev, evt);
2275 }
2276 EXPORT_SYMBOL_GPL(sdev_evt_send_simple);
2277
2278 /**
2279  *      scsi_device_quiesce - Block user issued commands.
2280  *      @sdev:  scsi device to quiesce.
2281  *
2282  *      This works by trying to transition to the SDEV_QUIESCE state
2283  *      (which must be a legal transition).  When the device is in this
2284  *      state, only special requests will be accepted, all others will
2285  *      be deferred.  Since special requests may also be requeued requests,
2286  *      a successful return doesn't guarantee the device will be 
2287  *      totally quiescent.
2288  *
2289  *      Must be called with user context, may sleep.
2290  *
2291  *      Returns zero if unsuccessful or an error if not.
2292  */
2293 int
2294 scsi_device_quiesce(struct scsi_device *sdev)
2295 {
2296         int err = scsi_device_set_state(sdev, SDEV_QUIESCE);
2297         if (err)
2298                 return err;
2299
2300         scsi_run_queue(sdev->request_queue);
2301         while (sdev->device_busy) {
2302                 msleep_interruptible(200);
2303                 scsi_run_queue(sdev->request_queue);
2304         }
2305         return 0;
2306 }
2307 EXPORT_SYMBOL(scsi_device_quiesce);
2308
2309 /**
2310  *      scsi_device_resume - Restart user issued commands to a quiesced device.
2311  *      @sdev:  scsi device to resume.
2312  *
2313  *      Moves the device from quiesced back to running and restarts the
2314  *      queues.
2315  *
2316  *      Must be called with user context, may sleep.
2317  */
2318 void
2319 scsi_device_resume(struct scsi_device *sdev)
2320 {
2321         if(scsi_device_set_state(sdev, SDEV_RUNNING))
2322                 return;
2323         scsi_run_queue(sdev->request_queue);
2324 }
2325 EXPORT_SYMBOL(scsi_device_resume);
2326
2327 static void
2328 device_quiesce_fn(struct scsi_device *sdev, void *data)
2329 {
2330         scsi_device_quiesce(sdev);
2331 }
2332
2333 void
2334 scsi_target_quiesce(struct scsi_target *starget)
2335 {
2336         starget_for_each_device(starget, NULL, device_quiesce_fn);
2337 }
2338 EXPORT_SYMBOL(scsi_target_quiesce);
2339
2340 static void
2341 device_resume_fn(struct scsi_device *sdev, void *data)
2342 {
2343         scsi_device_resume(sdev);
2344 }
2345
2346 void
2347 scsi_target_resume(struct scsi_target *starget)
2348 {
2349         starget_for_each_device(starget, NULL, device_resume_fn);
2350 }
2351 EXPORT_SYMBOL(scsi_target_resume);
2352
2353 /**
2354  * scsi_internal_device_block - internal function to put a device temporarily into the SDEV_BLOCK state
2355  * @sdev:       device to block
2356  *
2357  * Block request made by scsi lld's to temporarily stop all
2358  * scsi commands on the specified device.  Called from interrupt
2359  * or normal process context.
2360  *
2361  * Returns zero if successful or error if not
2362  *
2363  * Notes:       
2364  *      This routine transitions the device to the SDEV_BLOCK state
2365  *      (which must be a legal transition).  When the device is in this
2366  *      state, all commands are deferred until the scsi lld reenables
2367  *      the device with scsi_device_unblock or device_block_tmo fires.
2368  *      This routine assumes the host_lock is held on entry.
2369  */
2370 int
2371 scsi_internal_device_block(struct scsi_device *sdev)
2372 {
2373         struct request_queue *q = sdev->request_queue;
2374         unsigned long flags;
2375         int err = 0;
2376
2377         err = scsi_device_set_state(sdev, SDEV_BLOCK);
2378         if (err) {
2379                 err = scsi_device_set_state(sdev, SDEV_CREATED_BLOCK);
2380
2381                 if (err)
2382                         return err;
2383         }
2384
2385         /* 
2386          * The device has transitioned to SDEV_BLOCK.  Stop the
2387          * block layer from calling the midlayer with this device's
2388          * request queue. 
2389          */
2390         spin_lock_irqsave(q->queue_lock, flags);
2391         blk_stop_queue(q);
2392         spin_unlock_irqrestore(q->queue_lock, flags);
2393
2394         return 0;
2395 }
2396 EXPORT_SYMBOL_GPL(scsi_internal_device_block);
2397  
2398 /**
2399  * scsi_internal_device_unblock - resume a device after a block request
2400  * @sdev:       device to resume
2401  *
2402  * Called by scsi lld's or the midlayer to restart the device queue
2403  * for the previously suspended scsi device.  Called from interrupt or
2404  * normal process context.
2405  *
2406  * Returns zero if successful or error if not.
2407  *
2408  * Notes:       
2409  *      This routine transitions the device to the SDEV_RUNNING state
2410  *      (which must be a legal transition) allowing the midlayer to
2411  *      goose the queue for this device.  This routine assumes the 
2412  *      host_lock is held upon entry.
2413  */
2414 int
2415 scsi_internal_device_unblock(struct scsi_device *sdev)
2416 {
2417         struct request_queue *q = sdev->request_queue; 
2418         unsigned long flags;
2419         
2420         /* 
2421          * Try to transition the scsi device to SDEV_RUNNING
2422          * and goose the device queue if successful.  
2423          */
2424         if (sdev->sdev_state == SDEV_BLOCK)
2425                 sdev->sdev_state = SDEV_RUNNING;
2426         else if (sdev->sdev_state == SDEV_CREATED_BLOCK)
2427                 sdev->sdev_state = SDEV_CREATED;
2428         else
2429                 return -EINVAL;
2430
2431         spin_lock_irqsave(q->queue_lock, flags);
2432         blk_start_queue(q);
2433         spin_unlock_irqrestore(q->queue_lock, flags);
2434
2435         return 0;
2436 }
2437 EXPORT_SYMBOL_GPL(scsi_internal_device_unblock);
2438
2439 static void
2440 device_block(struct scsi_device *sdev, void *data)
2441 {
2442         scsi_internal_device_block(sdev);
2443 }
2444
2445 static int
2446 target_block(struct device *dev, void *data)
2447 {
2448         if (scsi_is_target_device(dev))
2449                 starget_for_each_device(to_scsi_target(dev), NULL,
2450                                         device_block);
2451         return 0;
2452 }
2453
2454 void
2455 scsi_target_block(struct device *dev)
2456 {
2457         if (scsi_is_target_device(dev))
2458                 starget_for_each_device(to_scsi_target(dev), NULL,
2459                                         device_block);
2460         else
2461                 device_for_each_child(dev, NULL, target_block);
2462 }
2463 EXPORT_SYMBOL_GPL(scsi_target_block);
2464
2465 static void
2466 device_unblock(struct scsi_device *sdev, void *data)
2467 {
2468         scsi_internal_device_unblock(sdev);
2469 }
2470
2471 static int
2472 target_unblock(struct device *dev, void *data)
2473 {
2474         if (scsi_is_target_device(dev))
2475                 starget_for_each_device(to_scsi_target(dev), NULL,
2476                                         device_unblock);
2477         return 0;
2478 }
2479
2480 void
2481 scsi_target_unblock(struct device *dev)
2482 {
2483         if (scsi_is_target_device(dev))
2484                 starget_for_each_device(to_scsi_target(dev), NULL,
2485                                         device_unblock);
2486         else
2487                 device_for_each_child(dev, NULL, target_unblock);
2488 }
2489 EXPORT_SYMBOL_GPL(scsi_target_unblock);
2490
2491 /**
2492  * scsi_kmap_atomic_sg - find and atomically map an sg-elemnt
2493  * @sgl:        scatter-gather list
2494  * @sg_count:   number of segments in sg
2495  * @offset:     offset in bytes into sg, on return offset into the mapped area
2496  * @len:        bytes to map, on return number of bytes mapped
2497  *
2498  * Returns virtual address of the start of the mapped page
2499  */
2500 void *scsi_kmap_atomic_sg(struct scatterlist *sgl, int sg_count,
2501                           size_t *offset, size_t *len)
2502 {
2503         int i;
2504         size_t sg_len = 0, len_complete = 0;
2505         struct scatterlist *sg;
2506         struct page *page;
2507
2508         WARN_ON(!irqs_disabled());
2509
2510         for_each_sg(sgl, sg, sg_count, i) {
2511                 len_complete = sg_len; /* Complete sg-entries */
2512                 sg_len += sg->length;
2513                 if (sg_len > *offset)
2514                         break;
2515         }
2516
2517         if (unlikely(i == sg_count)) {
2518                 printk(KERN_ERR "%s: Bytes in sg: %zu, requested offset %zu, "
2519                         "elements %d\n",
2520                        __func__, sg_len, *offset, sg_count);
2521                 WARN_ON(1);
2522                 return NULL;
2523         }
2524
2525         /* Offset starting from the beginning of first page in this sg-entry */
2526         *offset = *offset - len_complete + sg->offset;
2527
2528         /* Assumption: contiguous pages can be accessed as "page + i" */
2529         page = nth_page(sg_page(sg), (*offset >> PAGE_SHIFT));
2530         *offset &= ~PAGE_MASK;
2531
2532         /* Bytes in this sg-entry from *offset to the end of the page */
2533         sg_len = PAGE_SIZE - *offset;
2534         if (*len > sg_len)
2535                 *len = sg_len;
2536
2537         return kmap_atomic(page, KM_BIO_SRC_IRQ);
2538 }
2539 EXPORT_SYMBOL(scsi_kmap_atomic_sg);
2540
2541 /**
2542  * scsi_kunmap_atomic_sg - atomically unmap a virtual address, previously mapped with scsi_kmap_atomic_sg
2543  * @virt:       virtual address to be unmapped
2544  */
2545 void scsi_kunmap_atomic_sg(void *virt)
2546 {
2547         kunmap_atomic(virt, KM_BIO_SRC_IRQ);
2548 }
2549 EXPORT_SYMBOL(scsi_kunmap_atomic_sg);